1 /*
   2  * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "classfile/javaClasses.inline.hpp"
  26 #include "classfile/symbolTable.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "classfile/vmClasses.hpp"
  29 #include "code/codeCache.hpp"
  30 #include "code/debugInfoRec.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "code/pcDesc.hpp"
  33 #include "code/scopeDesc.hpp"
  34 #include "compiler/compilationPolicy.hpp"
  35 #include "compiler/compilerDefinitions.inline.hpp"
  36 #include "gc/shared/collectedHeap.hpp"
  37 #include "gc/shared/memAllocator.hpp"
  38 #include "interpreter/bytecode.hpp"
  39 #include "interpreter/bytecode.inline.hpp"
  40 #include "interpreter/bytecodeStream.hpp"
  41 #include "interpreter/interpreter.hpp"
  42 #include "interpreter/oopMapCache.hpp"
  43 #include "jvm.h"
  44 #include "logging/log.hpp"
  45 #include "logging/logLevel.hpp"
  46 #include "logging/logMessage.hpp"
  47 #include "logging/logStream.hpp"
  48 #include "memory/allocation.inline.hpp"
  49 #include "memory/oopFactory.hpp"
  50 #include "memory/resourceArea.hpp"
  51 #include "memory/universe.hpp"
  52 #include "oops/constantPool.hpp"
  53 #include "oops/fieldStreams.inline.hpp"
  54 #include "oops/method.hpp"
  55 #include "oops/objArrayKlass.hpp"
  56 #include "oops/objArrayOop.inline.hpp"
  57 #include "oops/oop.inline.hpp"
  58 #include "oops/typeArrayOop.inline.hpp"
  59 #include "oops/verifyOopClosure.hpp"
  60 #include "prims/jvmtiDeferredUpdates.hpp"
  61 #include "prims/jvmtiExport.hpp"
  62 #include "prims/jvmtiThreadState.hpp"
  63 #include "prims/methodHandles.hpp"
  64 #include "prims/vectorSupport.hpp"
  65 #include "runtime/atomic.hpp"
  66 #include "runtime/basicLock.inline.hpp"
  67 #include "runtime/continuation.hpp"
  68 #include "runtime/continuationEntry.inline.hpp"
  69 #include "runtime/deoptimization.hpp"
  70 #include "runtime/escapeBarrier.hpp"
  71 #include "runtime/fieldDescriptor.hpp"
  72 #include "runtime/fieldDescriptor.inline.hpp"
  73 #include "runtime/frame.inline.hpp"
  74 #include "runtime/handles.inline.hpp"
  75 #include "runtime/interfaceSupport.inline.hpp"
  76 #include "runtime/javaThread.hpp"
  77 #include "runtime/jniHandles.inline.hpp"
  78 #include "runtime/keepStackGCProcessed.hpp"
  79 #include "runtime/lightweightSynchronizer.hpp"
  80 #include "runtime/lockStack.inline.hpp"
  81 #include "runtime/objectMonitor.inline.hpp"
  82 #include "runtime/osThread.hpp"
  83 #include "runtime/safepointVerifiers.hpp"
  84 #include "runtime/sharedRuntime.hpp"
  85 #include "runtime/signature.hpp"
  86 #include "runtime/stackFrameStream.inline.hpp"
  87 #include "runtime/stackValue.hpp"
  88 #include "runtime/stackWatermarkSet.hpp"
  89 #include "runtime/stubRoutines.hpp"
  90 #include "runtime/synchronizer.inline.hpp"
  91 #include "runtime/threadSMR.hpp"
  92 #include "runtime/threadWXSetters.inline.hpp"
  93 #include "runtime/vframe.hpp"
  94 #include "runtime/vframeArray.hpp"
  95 #include "runtime/vframe_hp.hpp"
  96 #include "runtime/vmOperations.hpp"
  97 #include "utilities/checkedCast.hpp"
  98 #include "utilities/events.hpp"
  99 #include "utilities/growableArray.hpp"
 100 #include "utilities/macros.hpp"
 101 #include "utilities/preserveException.hpp"
 102 #include "utilities/xmlstream.hpp"
 103 #if INCLUDE_JFR
 104 #include "jfr/jfr.inline.hpp"
 105 #include "jfr/jfrEvents.hpp"
 106 #include "jfr/metadata/jfrSerializer.hpp"
 107 #endif
 108 
 109 uint64_t DeoptimizationScope::_committed_deopt_gen = 0;
 110 uint64_t DeoptimizationScope::_active_deopt_gen    = 1;
 111 bool     DeoptimizationScope::_committing_in_progress = false;
 112 
 113 DeoptimizationScope::DeoptimizationScope() : _required_gen(0) {
 114   DEBUG_ONLY(_deopted = false;)
 115 
 116   MutexLocker ml(NMethodState_lock, Mutex::_no_safepoint_check_flag);
 117   // If there is nothing to deopt _required_gen is the same as comitted.
 118   _required_gen = DeoptimizationScope::_committed_deopt_gen;
 119 }
 120 
 121 DeoptimizationScope::~DeoptimizationScope() {
 122   assert(_deopted, "Deopt not executed");
 123 }
 124 
 125 void DeoptimizationScope::mark(nmethod* nm, bool inc_recompile_counts) {
 126   if (!nm->can_be_deoptimized()) {
 127     return;
 128   }
 129 
 130   ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
 131 
 132   // If it's already marked but we still need it to be deopted.
 133   if (nm->is_marked_for_deoptimization()) {
 134     dependent(nm);
 135     return;
 136   }
 137 
 138   nmethod::DeoptimizationStatus status =
 139     inc_recompile_counts ? nmethod::deoptimize : nmethod::deoptimize_noupdate;
 140   Atomic::store(&nm->_deoptimization_status, status);
 141 
 142   // Make sure active is not committed
 143   assert(DeoptimizationScope::_committed_deopt_gen < DeoptimizationScope::_active_deopt_gen, "Must be");
 144   assert(nm->_deoptimization_generation == 0, "Is already marked");
 145 
 146   nm->_deoptimization_generation = DeoptimizationScope::_active_deopt_gen;
 147   _required_gen                  = DeoptimizationScope::_active_deopt_gen;
 148 }
 149 
 150 void DeoptimizationScope::dependent(nmethod* nm) {
 151   ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
 152 
 153   // A method marked by someone else may have a _required_gen lower than what we marked with.
 154   // Therefore only store it if it's higher than _required_gen.
 155   if (_required_gen < nm->_deoptimization_generation) {
 156     _required_gen = nm->_deoptimization_generation;
 157   }
 158 }
 159 
 160 void DeoptimizationScope::deoptimize_marked() {
 161   assert(!_deopted, "Already deopted");
 162 
 163   // We are not alive yet.
 164   if (!Universe::is_fully_initialized()) {
 165     DEBUG_ONLY(_deopted = true;)
 166     return;
 167   }
 168 
 169   // Safepoints are a special case, handled here.
 170   if (SafepointSynchronize::is_at_safepoint()) {
 171     DeoptimizationScope::_committed_deopt_gen = DeoptimizationScope::_active_deopt_gen;
 172     DeoptimizationScope::_active_deopt_gen++;
 173     Deoptimization::deoptimize_all_marked();
 174     DEBUG_ONLY(_deopted = true;)
 175     return;
 176   }
 177 
 178   uint64_t comitting = 0;
 179   bool wait = false;
 180   while (true) {
 181     {
 182       ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
 183 
 184       // First we check if we or someone else already deopted the gen we want.
 185       if (DeoptimizationScope::_committed_deopt_gen >= _required_gen) {
 186         DEBUG_ONLY(_deopted = true;)
 187         return;
 188       }
 189       if (!_committing_in_progress) {
 190         // The version we are about to commit.
 191         comitting = DeoptimizationScope::_active_deopt_gen;
 192         // Make sure new marks use a higher gen.
 193         DeoptimizationScope::_active_deopt_gen++;
 194         _committing_in_progress = true;
 195         wait = false;
 196       } else {
 197         // Another thread is handshaking and committing a gen.
 198         wait = true;
 199       }
 200     }
 201     if (wait) {
 202       // Wait and let the concurrent handshake be performed.
 203       ThreadBlockInVM tbivm(JavaThread::current());
 204       os::naked_yield();
 205     } else {
 206       // Performs the handshake.
 207       Deoptimization::deoptimize_all_marked(); // May safepoint and an additional deopt may have occurred.
 208       DEBUG_ONLY(_deopted = true;)
 209       {
 210         ConditionalMutexLocker ml(NMethodState_lock, !NMethodState_lock->owned_by_self(), Mutex::_no_safepoint_check_flag);
 211 
 212         // Make sure that committed doesn't go backwards.
 213         // Should only happen if we did a deopt during a safepoint above.
 214         if (DeoptimizationScope::_committed_deopt_gen < comitting) {
 215           DeoptimizationScope::_committed_deopt_gen = comitting;
 216         }
 217         _committing_in_progress = false;
 218 
 219         assert(DeoptimizationScope::_committed_deopt_gen >= _required_gen, "Must be");
 220 
 221         return;
 222       }
 223     }
 224   }
 225 }
 226 
 227 Deoptimization::UnrollBlock::UnrollBlock(int  size_of_deoptimized_frame,
 228                                          int  caller_adjustment,
 229                                          int  caller_actual_parameters,
 230                                          int  number_of_frames,
 231                                          intptr_t* frame_sizes,
 232                                          address* frame_pcs,
 233                                          BasicType return_type,
 234                                          int exec_mode) {
 235   _size_of_deoptimized_frame = size_of_deoptimized_frame;
 236   _caller_adjustment         = caller_adjustment;
 237   _caller_actual_parameters  = caller_actual_parameters;
 238   _number_of_frames          = number_of_frames;
 239   _frame_sizes               = frame_sizes;
 240   _frame_pcs                 = frame_pcs;
 241   _register_block            = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
 242   _return_type               = return_type;
 243   _initial_info              = 0;
 244   // PD (x86 only)
 245   _counter_temp              = 0;
 246   _unpack_kind               = exec_mode;
 247   _sender_sp_temp            = 0;
 248 
 249   _total_frame_sizes         = size_of_frames();
 250   assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
 251 }
 252 
 253 Deoptimization::UnrollBlock::~UnrollBlock() {
 254   FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
 255   FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
 256   FREE_C_HEAP_ARRAY(intptr_t, _register_block);
 257 }
 258 
 259 int Deoptimization::UnrollBlock::size_of_frames() const {
 260   // Account first for the adjustment of the initial frame
 261   intptr_t result = _caller_adjustment;
 262   for (int index = 0; index < number_of_frames(); index++) {
 263     result += frame_sizes()[index];
 264   }
 265   return checked_cast<int>(result);
 266 }
 267 
 268 void Deoptimization::UnrollBlock::print() {
 269   ResourceMark rm;
 270   stringStream st;
 271   st.print_cr("UnrollBlock");
 272   st.print_cr("  size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
 273   st.print(   "  frame_sizes: ");
 274   for (int index = 0; index < number_of_frames(); index++) {
 275     st.print("%zd ", frame_sizes()[index]);
 276   }
 277   st.cr();
 278   tty->print_raw(st.freeze());
 279 }
 280 
 281 // In order to make fetch_unroll_info work properly with escape
 282 // analysis, the method was changed from JRT_LEAF to JRT_BLOCK_ENTRY.
 283 // The actual reallocation of previously eliminated objects occurs in realloc_objects,
 284 // which is called from the method fetch_unroll_info_helper below.
 285 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* current, int exec_mode))
 286   // fetch_unroll_info() is called at the beginning of the deoptimization
 287   // handler. Note this fact before we start generating temporary frames
 288   // that can confuse an asynchronous stack walker. This counter is
 289   // decremented at the end of unpack_frames().
 290   current->inc_in_deopt_handler();
 291 
 292   if (exec_mode == Unpack_exception) {
 293     // When we get here, a callee has thrown an exception into a deoptimized
 294     // frame. That throw might have deferred stack watermark checking until
 295     // after unwinding. So we deal with such deferred requests here.
 296     StackWatermarkSet::after_unwind(current);
 297   }
 298 
 299   return fetch_unroll_info_helper(current, exec_mode);
 300 JRT_END
 301 
 302 #if COMPILER2_OR_JVMCI
 303 // print information about reallocated objects
 304 static void print_objects(JavaThread* deoptee_thread,
 305                           GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
 306   ResourceMark rm;
 307   stringStream st;  // change to logStream with logging
 308   st.print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
 309   fieldDescriptor fd;
 310 
 311   for (int i = 0; i < objects->length(); i++) {
 312     ObjectValue* sv = (ObjectValue*) objects->at(i);
 313     Handle obj = sv->value();
 314 
 315     if (obj.is_null()) {
 316       st.print_cr("     nullptr");
 317       continue;
 318     }
 319 
 320     Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
 321 
 322     st.print("     object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
 323     k->print_value_on(&st);
 324     st.print_cr(" allocated (%zu bytes)", obj->size() * HeapWordSize);
 325 
 326     if (Verbose && k != nullptr) {
 327       k->oop_print_on(obj(), &st);
 328     }
 329   }
 330   tty->print_raw(st.freeze());
 331 }
 332 
 333 static bool rematerialize_objects(JavaThread* thread, int exec_mode, nmethod* compiled_method,
 334                                   frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
 335                                   bool& deoptimized_objects) {
 336   bool realloc_failures = false;
 337   assert (chunk->at(0)->scope() != nullptr,"expect only compiled java frames");
 338 
 339   JavaThread* deoptee_thread = chunk->at(0)->thread();
 340   assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread),
 341          "a frame can only be deoptimized by the owner thread");
 342 
 343   GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects_to_rematerialize(deoptee, map);
 344 
 345   // The flag return_oop() indicates call sites which return oop
 346   // in compiled code. Such sites include java method calls,
 347   // runtime calls (for example, used to allocate new objects/arrays
 348   // on slow code path) and any other calls generated in compiled code.
 349   // It is not guaranteed that we can get such information here only
 350   // by analyzing bytecode in deoptimized frames. This is why this flag
 351   // is set during method compilation (see Compile::Process_OopMap_Node()).
 352   // If the previous frame was popped or if we are dispatching an exception,
 353   // we don't have an oop result.
 354   bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
 355   Handle return_value;
 356   if (save_oop_result) {
 357     // Reallocation may trigger GC. If deoptimization happened on return from
 358     // call which returns oop we need to save it since it is not in oopmap.
 359     oop result = deoptee.saved_oop_result(&map);
 360     assert(oopDesc::is_oop_or_null(result), "must be oop");
 361     return_value = Handle(thread, result);
 362     assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
 363     if (TraceDeoptimization) {
 364       tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
 365       tty->cr();
 366     }
 367   }
 368   if (objects != nullptr) {
 369     if (exec_mode == Deoptimization::Unpack_none) {
 370       assert(thread->thread_state() == _thread_in_vm, "assumption");
 371       JavaThread* THREAD = thread; // For exception macros.
 372       // Clear pending OOM if reallocation fails and return true indicating allocation failure
 373       realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
 374       deoptimized_objects = true;
 375     } else {
 376       JavaThread* current = thread; // For JRT_BLOCK
 377       JRT_BLOCK
 378       realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
 379       JRT_END
 380     }
 381     guarantee(compiled_method != nullptr, "deopt must be associated with an nmethod");
 382     bool is_jvmci = compiled_method->is_compiled_by_jvmci();
 383     Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, is_jvmci);
 384     if (TraceDeoptimization) {
 385       print_objects(deoptee_thread, objects, realloc_failures);
 386     }
 387   }
 388   if (save_oop_result) {
 389     // Restore result.
 390     deoptee.set_saved_oop_result(&map, return_value());
 391   }
 392   return realloc_failures;
 393 }
 394 
 395 static void restore_eliminated_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
 396                                      frame& deoptee, int exec_mode, bool& deoptimized_objects) {
 397   JavaThread* deoptee_thread = chunk->at(0)->thread();
 398   assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
 399   assert(thread == Thread::current(), "should be");
 400   HandleMark hm(thread);
 401 #ifndef PRODUCT
 402   bool first = true;
 403 #endif // !PRODUCT
 404   // Start locking from outermost/oldest frame
 405   for (int i = (chunk->length() - 1); i >= 0; i--) {
 406     compiledVFrame* cvf = chunk->at(i);
 407     assert (cvf->scope() != nullptr,"expect only compiled java frames");
 408     GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
 409     if (monitors->is_nonempty()) {
 410       bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee,
 411                                                      exec_mode, realloc_failures);
 412       deoptimized_objects = deoptimized_objects || relocked;
 413 #ifndef PRODUCT
 414       if (PrintDeoptimizationDetails) {
 415         ResourceMark rm;
 416         stringStream st;
 417         for (int j = 0; j < monitors->length(); j++) {
 418           MonitorInfo* mi = monitors->at(j);
 419           if (mi->eliminated()) {
 420             if (first) {
 421               first = false;
 422               st.print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
 423             }
 424             if (exec_mode == Deoptimization::Unpack_none) {
 425               ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
 426               if (monitor != nullptr && monitor->object() == mi->owner()) {
 427                 st.print_cr("     object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
 428                 continue;
 429               }
 430             }
 431             if (mi->owner_is_scalar_replaced()) {
 432               Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
 433               st.print_cr("     failed reallocation for klass %s", k->external_name());
 434             } else {
 435               st.print_cr("     object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
 436             }
 437           }
 438         }
 439         tty->print_raw(st.freeze());
 440       }
 441 #endif // !PRODUCT
 442     }
 443   }
 444 }
 445 
 446 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
 447 // The given vframes cover one physical frame.
 448 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk,
 449                                                  bool& realloc_failures) {
 450   frame deoptee = chunk->at(0)->fr();
 451   JavaThread* deoptee_thread = chunk->at(0)->thread();
 452   nmethod* nm = deoptee.cb()->as_nmethod_or_null();
 453   RegisterMap map(chunk->at(0)->register_map());
 454   bool deoptimized_objects = false;
 455 
 456   bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
 457 
 458   // Reallocate the non-escaping objects and restore their fields.
 459   if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations)
 460                                       || EliminateAutoBox || EnableVectorAggressiveReboxing)) {
 461     realloc_failures = rematerialize_objects(thread, Unpack_none, nm, deoptee, map, chunk, deoptimized_objects);
 462   }
 463 
 464   // MonitorInfo structures used in eliminate_locks are not GC safe.
 465   NoSafepointVerifier no_safepoint;
 466 
 467   // Now relock objects if synchronization on them was eliminated.
 468   if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
 469     restore_eliminated_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
 470   }
 471   return deoptimized_objects;
 472 }
 473 #endif // COMPILER2_OR_JVMCI
 474 
 475 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
 476 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* current, int exec_mode) {
 477   JFR_ONLY(Jfr::check_and_process_sample_request(current);)
 478   // When we get here we are about to unwind the deoptee frame. In order to
 479   // catch not yet safe to use frames, the following stack watermark barrier
 480   // poll will make such frames safe to use.
 481   StackWatermarkSet::before_unwind(current);
 482 
 483   // Note: there is a safepoint safety issue here. No matter whether we enter
 484   // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
 485   // the vframeArray is created.
 486   //
 487 
 488   // Allocate our special deoptimization ResourceMark
 489   DeoptResourceMark* dmark = new DeoptResourceMark(current);
 490   assert(current->deopt_mark() == nullptr, "Pending deopt!");
 491   current->set_deopt_mark(dmark);
 492 
 493   frame stub_frame = current->last_frame(); // Makes stack walkable as side effect
 494   RegisterMap map(current,
 495                   RegisterMap::UpdateMap::include,
 496                   RegisterMap::ProcessFrames::include,
 497                   RegisterMap::WalkContinuation::skip);
 498   RegisterMap dummy_map(current,
 499                         RegisterMap::UpdateMap::skip,
 500                         RegisterMap::ProcessFrames::include,
 501                         RegisterMap::WalkContinuation::skip);
 502   // Now get the deoptee with a valid map
 503   frame deoptee = stub_frame.sender(&map);
 504   // Set the deoptee nmethod
 505   assert(current->deopt_compiled_method() == nullptr, "Pending deopt!");
 506   nmethod* nm = deoptee.cb()->as_nmethod_or_null();
 507   current->set_deopt_compiled_method(nm);
 508 
 509   if (VerifyStack) {
 510     current->validate_frame_layout();
 511   }
 512 
 513   // Create a growable array of VFrames where each VFrame represents an inlined
 514   // Java frame.  This storage is allocated with the usual system arena.
 515   assert(deoptee.is_compiled_frame(), "Wrong frame type");
 516   GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
 517   vframe* vf = vframe::new_vframe(&deoptee, &map, current);
 518   while (!vf->is_top()) {
 519     assert(vf->is_compiled_frame(), "Wrong frame type");
 520     chunk->push(compiledVFrame::cast(vf));
 521     vf = vf->sender();
 522   }
 523   assert(vf->is_compiled_frame(), "Wrong frame type");
 524   chunk->push(compiledVFrame::cast(vf));
 525 
 526   bool realloc_failures = false;
 527 
 528 #if COMPILER2_OR_JVMCI
 529   bool const jvmci_enabled = JVMCI_ONLY(EnableJVMCI) NOT_JVMCI(false);
 530 
 531   // Reallocate the non-escaping objects and restore their fields. Then
 532   // relock objects if synchronization on them was eliminated.
 533   if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations)
 534                                        || EliminateAutoBox || EnableVectorAggressiveReboxing )) {
 535     bool unused;
 536     realloc_failures = rematerialize_objects(current, exec_mode, nm, deoptee, map, chunk, unused);
 537   }
 538 #endif // COMPILER2_OR_JVMCI
 539 
 540   // Ensure that no safepoint is taken after pointers have been stored
 541   // in fields of rematerialized objects.  If a safepoint occurs from here on
 542   // out the java state residing in the vframeArray will be missed.
 543   // Locks may be rebaised in a safepoint.
 544   NoSafepointVerifier no_safepoint;
 545 
 546 #if COMPILER2_OR_JVMCI
 547   if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
 548       && !EscapeBarrier::objs_are_deoptimized(current, deoptee.id())) {
 549     bool unused = false;
 550     restore_eliminated_locks(current, chunk, realloc_failures, deoptee, exec_mode, unused);
 551   }
 552 #endif // COMPILER2_OR_JVMCI
 553 
 554   ScopeDesc* trap_scope = chunk->at(0)->scope();
 555   Handle exceptionObject;
 556   if (trap_scope->rethrow_exception()) {
 557 #ifndef PRODUCT
 558     if (PrintDeoptimizationDetails) {
 559       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());
 560     }
 561 #endif // !PRODUCT
 562 
 563     GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
 564     guarantee(expressions != nullptr && expressions->length() == 1, "should have only exception on stack");
 565     guarantee(exec_mode != Unpack_exception, "rethrow_exception set with Unpack_exception");
 566     ScopeValue* topOfStack = expressions->top();
 567     exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
 568     guarantee(exceptionObject() != nullptr, "exception oop can not be null");
 569   }
 570 
 571   vframeArray* array = create_vframeArray(current, deoptee, &map, chunk, realloc_failures);
 572 #if COMPILER2_OR_JVMCI
 573   if (realloc_failures) {
 574     // This destroys all ScopedValue bindings.
 575     current->clear_scopedValueBindings();
 576     pop_frames_failed_reallocs(current, array);
 577   }
 578 #endif
 579 
 580   assert(current->vframe_array_head() == nullptr, "Pending deopt!");
 581   current->set_vframe_array_head(array);
 582 
 583   // Now that the vframeArray has been created if we have any deferred local writes
 584   // added by jvmti then we can free up that structure as the data is now in the
 585   // vframeArray
 586 
 587   JvmtiDeferredUpdates::delete_updates_for_frame(current, array->original().id());
 588 
 589   // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
 590   CodeBlob* cb = stub_frame.cb();
 591   // Verify we have the right vframeArray
 592   assert(cb->frame_size() >= 0, "Unexpected frame size");
 593   intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
 594 
 595   // If the deopt call site is a MethodHandle invoke call site we have
 596   // to adjust the unpack_sp.
 597   nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
 598   if (deoptee_nm != nullptr && deoptee_nm->is_method_handle_return(deoptee.pc()))
 599     unpack_sp = deoptee.unextended_sp();
 600 
 601 #ifdef ASSERT
 602   assert(cb->is_deoptimization_stub() ||
 603          cb->is_uncommon_trap_stub() ||
 604          strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
 605          strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
 606          "unexpected code blob: %s", cb->name());
 607 #endif
 608 
 609   // This is a guarantee instead of an assert because if vframe doesn't match
 610   // we will unpack the wrong deoptimized frame and wind up in strange places
 611   // where it will be very difficult to figure out what went wrong. Better
 612   // to die an early death here than some very obscure death later when the
 613   // trail is cold.
 614   guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
 615 
 616   int number_of_frames = array->frames();
 617 
 618   // Compute the vframes' sizes.  Note that frame_sizes[] entries are ordered from outermost to innermost
 619   // virtual activation, which is the reverse of the elements in the vframes array.
 620   intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
 621   // +1 because we always have an interpreter return address for the final slot.
 622   address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
 623   int popframe_extra_args = 0;
 624   // Create an interpreter return address for the stub to use as its return
 625   // address so the skeletal frames are perfectly walkable
 626   frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
 627 
 628   // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
 629   // activation be put back on the expression stack of the caller for reexecution
 630   if (JvmtiExport::can_pop_frame() && current->popframe_forcing_deopt_reexecution()) {
 631     popframe_extra_args = in_words(current->popframe_preserved_args_size_in_words());
 632   }
 633 
 634   // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
 635   // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
 636   // than simply use array->sender.pc(). This requires us to walk the current set of frames
 637   //
 638   frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
 639   deopt_sender = deopt_sender.sender(&dummy_map);     // Now deoptee caller
 640 
 641   // It's possible that the number of parameters at the call site is
 642   // different than number of arguments in the callee when method
 643   // handles are used.  If the caller is interpreted get the real
 644   // value so that the proper amount of space can be added to it's
 645   // frame.
 646   bool caller_was_method_handle = false;
 647   if (deopt_sender.is_interpreted_frame()) {
 648     methodHandle method(current, deopt_sender.interpreter_frame_method());
 649     Bytecode_invoke cur(method, deopt_sender.interpreter_frame_bci());
 650     if (cur.has_member_arg()) {
 651       // This should cover all real-world cases.  One exception is a pathological chain of
 652       // MH.linkToXXX() linker calls, which only trusted code could do anyway.  To handle that case, we
 653       // would need to get the size from the resolved method entry.  Another exception would
 654       // be an invokedynamic with an adapter that is really a MethodHandle linker.
 655       caller_was_method_handle = true;
 656     }
 657   }
 658 
 659   //
 660   // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
 661   // frame_sizes/frame_pcs[1] next oldest frame (int)
 662   // frame_sizes/frame_pcs[n] youngest frame (int)
 663   //
 664   // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
 665   // owns the space for the return address to it's caller).  Confusing ain't it.
 666   //
 667   // The vframe array can address vframes with indices running from
 668   // 0.._frames-1. Index  0 is the youngest frame and _frame - 1 is the oldest (root) frame.
 669   // When we create the skeletal frames we need the oldest frame to be in the zero slot
 670   // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
 671   // so things look a little strange in this loop.
 672   //
 673   int callee_parameters = 0;
 674   int callee_locals = 0;
 675   for (int index = 0; index < array->frames(); index++ ) {
 676     // frame[number_of_frames - 1 ] = on_stack_size(youngest)
 677     // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
 678     // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
 679     frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
 680                                                                                                     callee_locals,
 681                                                                                                     index == 0,
 682                                                                                                     popframe_extra_args);
 683     // This pc doesn't have to be perfect just good enough to identify the frame
 684     // as interpreted so the skeleton frame will be walkable
 685     // The correct pc will be set when the skeleton frame is completely filled out
 686     // The final pc we store in the loop is wrong and will be overwritten below
 687     frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
 688 
 689     callee_parameters = array->element(index)->method()->size_of_parameters();
 690     callee_locals = array->element(index)->method()->max_locals();
 691     popframe_extra_args = 0;
 692   }
 693 
 694   // Compute whether the root vframe returns a float or double value.
 695   BasicType return_type;
 696   {
 697     methodHandle method(current, array->element(0)->method());
 698     Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
 699     return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
 700   }
 701 
 702   // Compute information for handling adapters and adjusting the frame size of the caller.
 703   int caller_adjustment = 0;
 704 
 705   // Compute the amount the oldest interpreter frame will have to adjust
 706   // its caller's stack by. If the caller is a compiled frame then
 707   // we pretend that the callee has no parameters so that the
 708   // extension counts for the full amount of locals and not just
 709   // locals-parms. This is because without a c2i adapter the parm
 710   // area as created by the compiled frame will not be usable by
 711   // the interpreter. (Depending on the calling convention there
 712   // may not even be enough space).
 713 
 714   // QQQ I'd rather see this pushed down into last_frame_adjust
 715   // and have it take the sender (aka caller).
 716 
 717   if (!deopt_sender.is_interpreted_frame() || caller_was_method_handle) {
 718     caller_adjustment = last_frame_adjust(0, callee_locals);
 719   } else if (callee_locals > callee_parameters) {
 720     // The caller frame may need extending to accommodate
 721     // non-parameter locals of the first unpacked interpreted frame.
 722     // Compute that adjustment.
 723     caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
 724   }
 725 
 726   // If the sender is deoptimized the we must retrieve the address of the handler
 727   // since the frame will "magically" show the original pc before the deopt
 728   // and we'd undo the deopt.
 729 
 730   frame_pcs[0] = Continuation::is_cont_barrier_frame(deoptee) ? StubRoutines::cont_returnBarrier() : deopt_sender.raw_pc();
 731   if (Continuation::is_continuation_enterSpecial(deopt_sender)) {
 732     ContinuationEntry::from_frame(deopt_sender)->set_argsize(0);
 733   }
 734 
 735   assert(CodeCache::find_blob(frame_pcs[0]) != nullptr, "bad pc");
 736 
 737 #if INCLUDE_JVMCI
 738   if (exceptionObject() != nullptr) {
 739     current->set_exception_oop(exceptionObject());
 740     exec_mode = Unpack_exception;
 741     assert(array->element(0)->rethrow_exception(), "must be");
 742   }
 743 #endif
 744 
 745   if (current->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
 746     assert(current->has_pending_exception(), "should have thrown OOME");
 747     current->set_exception_oop(current->pending_exception());
 748     current->clear_pending_exception();
 749     exec_mode = Unpack_exception;
 750   }
 751 
 752 #if INCLUDE_JVMCI
 753   if (current->frames_to_pop_failed_realloc() > 0) {
 754     current->set_pending_monitorenter(false);
 755   }
 756 #endif
 757 
 758   int caller_actual_parameters = -1; // value not used except for interpreted frames, see below
 759   if (deopt_sender.is_interpreted_frame()) {
 760     caller_actual_parameters = callee_parameters + (caller_was_method_handle ? 1 : 0);
 761   }
 762 
 763   UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
 764                                       caller_adjustment * BytesPerWord,
 765                                       caller_actual_parameters,
 766                                       number_of_frames,
 767                                       frame_sizes,
 768                                       frame_pcs,
 769                                       return_type,
 770                                       exec_mode);
 771   // On some platforms, we need a way to pass some platform dependent
 772   // information to the unpacking code so the skeletal frames come out
 773   // correct (initial fp value, unextended sp, ...)
 774   info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
 775 
 776   if (array->frames() > 1) {
 777     if (VerifyStack && TraceDeoptimization) {
 778       tty->print_cr("Deoptimizing method containing inlining");
 779     }
 780   }
 781 
 782   array->set_unroll_block(info);
 783   return info;
 784 }
 785 
 786 // Called to cleanup deoptimization data structures in normal case
 787 // after unpacking to stack and when stack overflow error occurs
 788 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
 789                                         vframeArray *array) {
 790 
 791   // Get array if coming from exception
 792   if (array == nullptr) {
 793     array = thread->vframe_array_head();
 794   }
 795   thread->set_vframe_array_head(nullptr);
 796 
 797   // Free the previous UnrollBlock
 798   vframeArray* old_array = thread->vframe_array_last();
 799   thread->set_vframe_array_last(array);
 800 
 801   if (old_array != nullptr) {
 802     UnrollBlock* old_info = old_array->unroll_block();
 803     old_array->set_unroll_block(nullptr);
 804     delete old_info;
 805     delete old_array;
 806   }
 807 
 808   // Deallocate any resource creating in this routine and any ResourceObjs allocated
 809   // inside the vframeArray (StackValueCollections)
 810 
 811   delete thread->deopt_mark();
 812   thread->set_deopt_mark(nullptr);
 813   thread->set_deopt_compiled_method(nullptr);
 814 
 815 
 816   if (JvmtiExport::can_pop_frame()) {
 817     // Regardless of whether we entered this routine with the pending
 818     // popframe condition bit set, we should always clear it now
 819     thread->clear_popframe_condition();
 820   }
 821 
 822   // unpack_frames() is called at the end of the deoptimization handler
 823   // and (in C2) at the end of the uncommon trap handler. Note this fact
 824   // so that an asynchronous stack walker can work again. This counter is
 825   // incremented at the beginning of fetch_unroll_info() and (in C2) at
 826   // the beginning of uncommon_trap().
 827   thread->dec_in_deopt_handler();
 828 }
 829 
 830 // Moved from cpu directories because none of the cpus has callee save values.
 831 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
 832 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
 833 
 834   // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
 835   // the days we had adapter frames. When we deoptimize a situation where a
 836   // compiled caller calls a compiled caller will have registers it expects
 837   // to survive the call to the callee. If we deoptimize the callee the only
 838   // way we can restore these registers is to have the oldest interpreter
 839   // frame that we create restore these values. That is what this routine
 840   // will accomplish.
 841 
 842   // At the moment we have modified c2 to not have any callee save registers
 843   // so this problem does not exist and this routine is just a place holder.
 844 
 845   assert(f->is_interpreted_frame(), "must be interpreted");
 846 }
 847 
 848 #ifndef PRODUCT
 849 #ifdef ASSERT
 850 // Return true if the execution after the provided bytecode continues at the
 851 // next bytecode in the code. This is not the case for gotos, returns, and
 852 // throws.
 853 static bool falls_through(Bytecodes::Code bc) {
 854   switch (bc) {
 855     case Bytecodes::_goto:
 856     case Bytecodes::_goto_w:
 857     case Bytecodes::_athrow:
 858     case Bytecodes::_areturn:
 859     case Bytecodes::_dreturn:
 860     case Bytecodes::_freturn:
 861     case Bytecodes::_ireturn:
 862     case Bytecodes::_lreturn:
 863     case Bytecodes::_jsr:
 864     case Bytecodes::_ret:
 865     case Bytecodes::_return:
 866     case Bytecodes::_lookupswitch:
 867     case Bytecodes::_tableswitch:
 868       return false;
 869     default:
 870       return true;
 871   }
 872 }
 873 #endif
 874 #endif
 875 
 876 // Return BasicType of value being returned
 877 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
 878   assert(thread == JavaThread::current(), "pre-condition");
 879 
 880   // We are already active in the special DeoptResourceMark any ResourceObj's we
 881   // allocate will be freed at the end of the routine.
 882 
 883   // JRT_LEAF methods don't normally allocate handles and there is a
 884   // NoHandleMark to enforce that. It is actually safe to use Handles
 885   // in a JRT_LEAF method, and sometimes desirable, but to do so we
 886   // must use ResetNoHandleMark to bypass the NoHandleMark, and
 887   // then use a HandleMark to ensure any Handles we do create are
 888   // cleaned up in this scope.
 889   ResetNoHandleMark rnhm;
 890   HandleMark hm(thread);
 891 
 892   frame stub_frame = thread->last_frame();
 893 
 894   Continuation::notify_deopt(thread, stub_frame.sp());
 895 
 896   // Since the frame to unpack is the top frame of this thread, the vframe_array_head
 897   // must point to the vframeArray for the unpack frame.
 898   vframeArray* array = thread->vframe_array_head();
 899   UnrollBlock* info = array->unroll_block();
 900 
 901   // We set the last_Java frame. But the stack isn't really parsable here. So we
 902   // clear it to make sure JFR understands not to try and walk stacks from events
 903   // in here.
 904   intptr_t* sp = thread->frame_anchor()->last_Java_sp();
 905   thread->frame_anchor()->set_last_Java_sp(nullptr);
 906 
 907   // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
 908   array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
 909 
 910   thread->frame_anchor()->set_last_Java_sp(sp);
 911 
 912   BasicType bt = info->return_type();
 913 
 914   // If we have an exception pending, claim that the return type is an oop
 915   // so the deopt_blob does not overwrite the exception_oop.
 916 
 917   if (exec_mode == Unpack_exception)
 918     bt = T_OBJECT;
 919 
 920   // Cleanup thread deopt data
 921   cleanup_deopt_info(thread, array);
 922 
 923 #ifndef PRODUCT
 924   if (VerifyStack) {
 925     ResourceMark res_mark;
 926     // Clear pending exception to not break verification code (restored afterwards)
 927     PreserveExceptionMark pm(thread);
 928 
 929     thread->validate_frame_layout();
 930 
 931     // Verify that the just-unpacked frames match the interpreter's
 932     // notions of expression stack and locals
 933     vframeArray* cur_array = thread->vframe_array_last();
 934     RegisterMap rm(thread,
 935                    RegisterMap::UpdateMap::skip,
 936                    RegisterMap::ProcessFrames::include,
 937                    RegisterMap::WalkContinuation::skip);
 938     rm.set_include_argument_oops(false);
 939     int callee_size_of_parameters = 0;
 940     for (int frame_idx = 0; frame_idx < cur_array->frames(); frame_idx++) {
 941       bool is_top_frame = (frame_idx == 0);
 942       vframeArrayElement* el = cur_array->element(frame_idx);
 943       frame* iframe = el->iframe();
 944       guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
 945       methodHandle mh(thread, iframe->interpreter_frame_method());
 946       bool reexecute = el->should_reexecute();
 947 
 948       int cur_invoke_parameter_size = 0;
 949       int top_frame_expression_stack_adjustment = 0;
 950       int max_bci = mh->code_size();
 951       BytecodeStream str(mh, iframe->interpreter_frame_bci());
 952       assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
 953       Bytecodes::Code cur_code = str.next();
 954 
 955       if (!reexecute && !Bytecodes::is_invoke(cur_code)) {
 956         // We can only compute OopMaps for the before state, so we need to roll forward
 957         // to the next bytecode.
 958         assert(is_top_frame, "must be");
 959         assert(falls_through(cur_code), "must be");
 960         assert(cur_code != Bytecodes::_illegal, "illegal bytecode");
 961         assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
 962 
 963         // Need to subtract off the size of the result type of
 964         // the bytecode because this is not described in the
 965         // debug info but returned to the interpreter in the TOS
 966         // caching register
 967         BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
 968         if (bytecode_result_type != T_ILLEGAL) {
 969           top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
 970         }
 971         assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
 972 
 973         cur_code = str.next();
 974         // Reflect the fact that we have rolled forward and now need
 975         // top_frame_expression_stack_adjustment
 976         reexecute = true;
 977       }
 978 
 979       assert(cur_code != Bytecodes::_illegal, "illegal bytecode");
 980       assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
 981 
 982       // Get the oop map for this bci
 983       InterpreterOopMap mask;
 984       OopMapCache::compute_one_oop_map(mh, str.bci(), &mask);
 985       // Check to see if we can grab the number of outgoing arguments
 986       // at an uncommon trap for an invoke (where the compiler
 987       // generates debug info before the invoke has executed)
 988       if (Bytecodes::is_invoke(cur_code)) {
 989         Bytecode_invoke invoke(mh, str.bci());
 990         cur_invoke_parameter_size = invoke.size_of_parameters();
 991         if (!is_top_frame && invoke.has_member_arg()) {
 992           callee_size_of_parameters++;
 993         }
 994       }
 995 
 996       // Verify stack depth and oops in frame
 997       auto match = [&]() {
 998         int iframe_expr_ssize = iframe->interpreter_frame_expression_stack_size();
 999 #if INCLUDE_JVMCI
1000         if (is_top_frame && el->rethrow_exception()) {
1001           return iframe_expr_ssize == 1;
1002         }
1003 #endif
1004         // This should only be needed for C1
1005         if (is_top_frame && exec_mode == Unpack_exception && iframe_expr_ssize == 0) {
1006           return true;
1007         }
1008         if (reexecute) {
1009           int expr_ssize_before = iframe_expr_ssize + top_frame_expression_stack_adjustment;
1010           int oopmap_expr_invoke_ssize = mask.expression_stack_size() + cur_invoke_parameter_size;
1011           return expr_ssize_before == oopmap_expr_invoke_ssize;
1012         } else {
1013           int oopmap_expr_callee_ssize = mask.expression_stack_size() + callee_size_of_parameters;
1014           return iframe_expr_ssize == oopmap_expr_callee_ssize;
1015         }
1016       };
1017       if (!match()) {
1018         // Print out some information that will help us debug the problem
1019         tty->print_cr("Wrong number of expression stack elements during deoptimization");
1020         tty->print_cr("  Error occurred while verifying frame %d (0..%d, 0 is topmost)", frame_idx, cur_array->frames() - 1);
1021         tty->print_cr("  Current code %s", Bytecodes::name(cur_code));
1022         tty->print_cr("  Fabricated interpreter frame had %d expression stack elements",
1023                       iframe->interpreter_frame_expression_stack_size());
1024         tty->print_cr("  Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
1025         tty->print_cr("  callee_size_of_parameters = %d", callee_size_of_parameters);
1026         tty->print_cr("  top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
1027         tty->print_cr("  exec_mode = %d", exec_mode);
1028         tty->print_cr("  original should_reexecute = %s", el->should_reexecute() ? "true" : "false");
1029         tty->print_cr("  reexecute = %s%s", reexecute ? "true" : "false",
1030                       (reexecute != el->should_reexecute()) ? " (changed)" : "");
1031 #if INCLUDE_JVMCI
1032         tty->print_cr("  rethrow_exception = %s", el->rethrow_exception() ? "true" : "false");
1033 #endif
1034         tty->print_cr("  cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
1035         tty->print_cr("  Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
1036         tty->print_cr("  Interpreted frames:");
1037         for (int k = 0; k < cur_array->frames(); k++) {
1038           vframeArrayElement* el = cur_array->element(k);
1039           tty->print_cr("    %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
1040         }
1041         cur_array->print_on_2(tty);
1042         guarantee(false, "wrong number of expression stack elements during deopt");
1043       }
1044       VerifyOopClosure verify;
1045       iframe->oops_interpreted_do(&verify, &rm, false);
1046       callee_size_of_parameters = mh->size_of_parameters();
1047     }
1048   }
1049 #endif // !PRODUCT
1050 
1051   return bt;
1052 JRT_END
1053 
1054 class DeoptimizeMarkedHandshakeClosure : public HandshakeClosure {
1055  public:
1056   DeoptimizeMarkedHandshakeClosure() : HandshakeClosure("Deoptimize") {}
1057   void do_thread(Thread* thread) {
1058     JavaThread* jt = JavaThread::cast(thread);
1059     jt->deoptimize_marked_methods();
1060   }
1061 };
1062 
1063 void Deoptimization::deoptimize_all_marked() {
1064   ResourceMark rm;
1065 
1066   // Make the dependent methods not entrant
1067   CodeCache::make_marked_nmethods_deoptimized();
1068 
1069   DeoptimizeMarkedHandshakeClosure deopt;
1070   if (SafepointSynchronize::is_at_safepoint()) {
1071     Threads::java_threads_do(&deopt);
1072   } else {
1073     Handshake::execute(&deopt);
1074   }
1075 }
1076 
1077 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1078   = Deoptimization::Action_reinterpret;
1079 
1080 #if INCLUDE_JVMCI
1081 template<typename CacheType>
1082 class BoxCacheBase : public CHeapObj<mtCompiler> {
1083 protected:
1084   static InstanceKlass* find_cache_klass(Thread* thread, Symbol* klass_name) {
1085     ResourceMark rm(thread);
1086     char* klass_name_str = klass_name->as_C_string();
1087     InstanceKlass* ik = SystemDictionary::find_instance_klass(thread, klass_name, Handle());
1088     guarantee(ik != nullptr, "%s must be loaded", klass_name_str);
1089     if (!ik->is_in_error_state()) {
1090       guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
1091       CacheType::compute_offsets(ik);
1092     }
1093     return ik;
1094   }
1095 };
1096 
1097 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache  : public BoxCacheBase<CacheType> {
1098   PrimitiveType _low;
1099   PrimitiveType _high;
1100   jobject _cache;
1101 protected:
1102   static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
1103   BoxCache(Thread* thread) {
1104     InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(thread, CacheType::symbol());
1105     if (ik->is_in_error_state()) {
1106       _low = 1;
1107       _high = 0;
1108       _cache = nullptr;
1109     } else {
1110       objArrayOop cache = CacheType::cache(ik);
1111       assert(cache->length() > 0, "Empty cache");
1112       _low = BoxType::value(cache->obj_at(0));
1113       _high = checked_cast<PrimitiveType>(_low + cache->length() - 1);
1114       _cache = JNIHandles::make_global(Handle(thread, cache));
1115     }
1116   }
1117   ~BoxCache() {
1118     JNIHandles::destroy_global(_cache);
1119   }
1120 public:
1121   static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
1122     if (_singleton == nullptr) {
1123       BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
1124       if (!Atomic::replace_if_null(&_singleton, s)) {
1125         delete s;
1126       }
1127     }
1128     return _singleton;
1129   }
1130   oop lookup(PrimitiveType value) {
1131     if (_low <= value && value <= _high) {
1132       int offset = checked_cast<int>(value - _low);
1133       return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
1134     }
1135     return nullptr;
1136   }
1137   oop lookup_raw(intptr_t raw_value, bool& cache_init_error) {
1138     if (_cache == nullptr) {
1139       cache_init_error = true;
1140       return nullptr;
1141     }
1142     // Have to cast to avoid little/big-endian problems.
1143     if (sizeof(PrimitiveType) > sizeof(jint)) {
1144       jlong value = (jlong)raw_value;
1145       return lookup(value);
1146     }
1147     PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
1148     return lookup(value);
1149   }
1150 };
1151 
1152 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
1153 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
1154 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
1155 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
1156 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
1157 
1158 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = nullptr;
1159 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = nullptr;
1160 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = nullptr;
1161 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = nullptr;
1162 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = nullptr;
1163 
1164 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
1165   jobject _true_cache;
1166   jobject _false_cache;
1167 protected:
1168   static BooleanBoxCache *_singleton;
1169   BooleanBoxCache(Thread *thread) {
1170     InstanceKlass* ik = find_cache_klass(thread, java_lang_Boolean::symbol());
1171     if (ik->is_in_error_state()) {
1172       _true_cache = nullptr;
1173       _false_cache = nullptr;
1174     } else {
1175       _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
1176       _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
1177     }
1178   }
1179   ~BooleanBoxCache() {
1180     JNIHandles::destroy_global(_true_cache);
1181     JNIHandles::destroy_global(_false_cache);
1182   }
1183 public:
1184   static BooleanBoxCache* singleton(Thread* thread) {
1185     if (_singleton == nullptr) {
1186       BooleanBoxCache* s = new BooleanBoxCache(thread);
1187       if (!Atomic::replace_if_null(&_singleton, s)) {
1188         delete s;
1189       }
1190     }
1191     return _singleton;
1192   }
1193   oop lookup_raw(intptr_t raw_value, bool& cache_in_error) {
1194     if (_true_cache == nullptr) {
1195       cache_in_error = true;
1196       return nullptr;
1197     }
1198     // Have to cast to avoid little/big-endian problems.
1199     jboolean value = (jboolean)*((jint*)&raw_value);
1200     return lookup(value);
1201   }
1202   oop lookup(jboolean value) {
1203     if (value != 0) {
1204       return JNIHandles::resolve_non_null(_true_cache);
1205     }
1206     return JNIHandles::resolve_non_null(_false_cache);
1207   }
1208 };
1209 
1210 BooleanBoxCache* BooleanBoxCache::_singleton = nullptr;
1211 
1212 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, bool& cache_init_error, TRAPS) {
1213    Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1214    BasicType box_type = vmClasses::box_klass_type(k);
1215    if (box_type != T_OBJECT) {
1216      StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1217      switch(box_type) {
1218        case T_INT:     return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1219        case T_CHAR:    return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1220        case T_SHORT:   return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1221        case T_BYTE:    return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1222        case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1223        case T_LONG:    return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_intptr(), cache_init_error);
1224        default:;
1225      }
1226    }
1227    return nullptr;
1228 }
1229 #endif // INCLUDE_JVMCI
1230 
1231 #if COMPILER2_OR_JVMCI
1232 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1233   Handle pending_exception(THREAD, thread->pending_exception());
1234   const char* exception_file = thread->exception_file();
1235   int exception_line = thread->exception_line();
1236   thread->clear_pending_exception();
1237 
1238   bool failures = false;
1239 
1240   for (int i = 0; i < objects->length(); i++) {
1241     assert(objects->at(i)->is_object(), "invalid debug information");
1242     ObjectValue* sv = (ObjectValue*) objects->at(i);
1243 
1244     Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1245     oop obj = nullptr;
1246 
1247     bool cache_init_error = false;
1248     if (k->is_instance_klass()) {
1249 #if INCLUDE_JVMCI
1250       nmethod* nm = fr->cb()->as_nmethod_or_null();
1251       if (nm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1252         AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1253         obj = get_cached_box(abv, fr, reg_map, cache_init_error, THREAD);
1254         if (obj != nullptr) {
1255           // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1256           abv->set_cached(true);
1257         } else if (cache_init_error) {
1258           // Results in an OOME which is valid (as opposed to a class initialization error)
1259           // and is fine for the rare case a cache initialization failing.
1260           failures = true;
1261         }
1262       }
1263 #endif // INCLUDE_JVMCI
1264 
1265       InstanceKlass* ik = InstanceKlass::cast(k);
1266       if (obj == nullptr && !cache_init_error) {
1267         InternalOOMEMark iom(THREAD);
1268         if (EnableVectorSupport && VectorSupport::is_vector(ik)) {
1269           obj = VectorSupport::allocate_vector(ik, fr, reg_map, sv, THREAD);
1270         } else {
1271           obj = ik->allocate_instance(THREAD);
1272         }
1273       }
1274     } else if (k->is_typeArray_klass()) {
1275       TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1276       assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1277       int len = sv->field_size() / type2size[ak->element_type()];
1278       InternalOOMEMark iom(THREAD);
1279       obj = ak->allocate_instance(len, THREAD);
1280     } else if (k->is_objArray_klass()) {
1281       ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1282       InternalOOMEMark iom(THREAD);
1283       obj = ak->allocate_instance(sv->field_size(), THREAD);
1284     }
1285 
1286     if (obj == nullptr) {
1287       failures = true;
1288     }
1289 
1290     assert(sv->value().is_null(), "redundant reallocation");
1291     assert(obj != nullptr || HAS_PENDING_EXCEPTION || cache_init_error, "allocation should succeed or we should get an exception");
1292     CLEAR_PENDING_EXCEPTION;
1293     sv->set_value(obj);
1294   }
1295 
1296   if (failures) {
1297     THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1298   } else if (pending_exception.not_null()) {
1299     thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1300   }
1301 
1302   return failures;
1303 }
1304 
1305 #if INCLUDE_JVMCI
1306 /**
1307  * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1308  * we need to somehow be able to recover the actual kind to be able to write the correct
1309  * amount of bytes.
1310  * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1311  * the entries at index n + 1 to n + i are 'markers'.
1312  * For example, if we were writing a short at index 4 of a byte array of size 8, the
1313  * expected form of the array would be:
1314  *
1315  * {b0, b1, b2, b3, INT, marker, b6, b7}
1316  *
1317  * Thus, in order to get back the size of the entry, we simply need to count the number
1318  * of marked entries
1319  *
1320  * @param virtualArray the virtualized byte array
1321  * @param i index of the virtual entry we are recovering
1322  * @return The number of bytes the entry spans
1323  */
1324 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1325   int index = i;
1326   while (++index < virtualArray->field_size() &&
1327            virtualArray->field_at(index)->is_marker()) {}
1328   return index - i;
1329 }
1330 
1331 /**
1332  * If there was a guarantee for byte array to always start aligned to a long, we could
1333  * do a simple check on the parity of the index. Unfortunately, that is not always the
1334  * case. Thus, we check alignment of the actual address we are writing to.
1335  * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1336  * write a long to index 3.
1337  */
1338 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1339     jbyte* res = obj->byte_at_addr(index);
1340     assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1341     return res;
1342 }
1343 
1344 static void byte_array_put(typeArrayOop obj, StackValue* value, int index, int byte_count) {
1345   switch (byte_count) {
1346     case 1:
1347       obj->byte_at_put(index, (jbyte) value->get_jint());
1348       break;
1349     case 2:
1350       *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) value->get_jint();
1351       break;
1352     case 4:
1353       *((jint *) check_alignment_get_addr(obj, index, 4)) = value->get_jint();
1354       break;
1355     case 8:
1356       *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) value->get_intptr();
1357       break;
1358     default:
1359       ShouldNotReachHere();
1360   }
1361 }
1362 #endif // INCLUDE_JVMCI
1363 
1364 
1365 // restore elements of an eliminated type array
1366 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1367   int index = 0;
1368 
1369   for (int i = 0; i < sv->field_size(); i++) {
1370     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1371     switch(type) {
1372     case T_LONG: case T_DOUBLE: {
1373       assert(value->type() == T_INT, "Agreement.");
1374       StackValue* low =
1375         StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1376 #ifdef _LP64
1377       jlong res = (jlong)low->get_intptr();
1378 #else
1379       jlong res = jlong_from(value->get_jint(), low->get_jint());
1380 #endif
1381       obj->long_at_put(index, res);
1382       break;
1383     }
1384 
1385     case T_INT: case T_FLOAT: { // 4 bytes.
1386       assert(value->type() == T_INT, "Agreement.");
1387       bool big_value = false;
1388       if (i + 1 < sv->field_size() && type == T_INT) {
1389         if (sv->field_at(i)->is_location()) {
1390           Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1391           if (type == Location::dbl || type == Location::lng) {
1392             big_value = true;
1393           }
1394         } else if (sv->field_at(i)->is_constant_int()) {
1395           ScopeValue* next_scope_field = sv->field_at(i + 1);
1396           if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1397             big_value = true;
1398           }
1399         }
1400       }
1401 
1402       if (big_value) {
1403         StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1404   #ifdef _LP64
1405         jlong res = (jlong)low->get_intptr();
1406   #else
1407         jlong res = jlong_from(value->get_jint(), low->get_jint());
1408   #endif
1409         obj->int_at_put(index, *(jint*)&res);
1410         obj->int_at_put(++index, *((jint*)&res + 1));
1411       } else {
1412         obj->int_at_put(index, value->get_jint());
1413       }
1414       break;
1415     }
1416 
1417     case T_SHORT:
1418       assert(value->type() == T_INT, "Agreement.");
1419       obj->short_at_put(index, (jshort)value->get_jint());
1420       break;
1421 
1422     case T_CHAR:
1423       assert(value->type() == T_INT, "Agreement.");
1424       obj->char_at_put(index, (jchar)value->get_jint());
1425       break;
1426 
1427     case T_BYTE: {
1428       assert(value->type() == T_INT, "Agreement.");
1429 #if INCLUDE_JVMCI
1430       // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1431       int byte_count = count_number_of_bytes_for_entry(sv, i);
1432       byte_array_put(obj, value, index, byte_count);
1433       // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1434       i += byte_count - 1; // Balance the loop counter.
1435       index += byte_count;
1436       // index has been updated so continue at top of loop
1437       continue;
1438 #else
1439       obj->byte_at_put(index, (jbyte)value->get_jint());
1440       break;
1441 #endif // INCLUDE_JVMCI
1442     }
1443 
1444     case T_BOOLEAN: {
1445       assert(value->type() == T_INT, "Agreement.");
1446       obj->bool_at_put(index, (jboolean)value->get_jint());
1447       break;
1448     }
1449 
1450       default:
1451         ShouldNotReachHere();
1452     }
1453     index++;
1454   }
1455 }
1456 
1457 // restore fields of an eliminated object array
1458 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1459   for (int i = 0; i < sv->field_size(); i++) {
1460     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1461     assert(value->type() == T_OBJECT, "object element expected");
1462     obj->obj_at_put(i, value->get_obj()());
1463   }
1464 }
1465 
1466 class ReassignedField {
1467 public:
1468   int _offset;
1469   BasicType _type;
1470 public:
1471   ReassignedField() {
1472     _offset = 0;
1473     _type = T_ILLEGAL;
1474   }
1475 };
1476 
1477 // Gets the fields of `klass` that are eliminated by escape analysis and need to be reassigned
1478 static GrowableArray<ReassignedField>* get_reassigned_fields(InstanceKlass* klass, GrowableArray<ReassignedField>* fields, bool is_jvmci) {
1479   InstanceKlass* super = klass->superklass();
1480   if (super != nullptr) {
1481     get_reassigned_fields(super, fields, is_jvmci);
1482   }
1483   for (AllFieldStream fs(klass); !fs.done(); fs.next()) {
1484     if (!fs.access_flags().is_static() && (is_jvmci || !fs.field_flags().is_injected())) {
1485       ReassignedField field;
1486       field._offset = fs.offset();
1487       field._type = Signature::basic_type(fs.signature());
1488       fields->append(field);
1489     }
1490   }
1491   return fields;
1492 }
1493 
1494 // Restore fields of an eliminated instance object employing the same field order used by the compiler.
1495 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool is_jvmci) {
1496   GrowableArray<ReassignedField>* fields = get_reassigned_fields(klass, new GrowableArray<ReassignedField>(), is_jvmci);
1497   for (int i = 0; i < fields->length(); i++) {
1498     ScopeValue* scope_field = sv->field_at(svIndex);
1499     StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1500     int offset = fields->at(i)._offset;
1501     BasicType type = fields->at(i)._type;
1502     switch (type) {
1503       case T_OBJECT: case T_ARRAY:
1504         assert(value->type() == T_OBJECT, "Agreement.");
1505         obj->obj_field_put(offset, value->get_obj()());
1506         break;
1507 
1508       case T_INT: case T_FLOAT: { // 4 bytes.
1509         assert(value->type() == T_INT, "Agreement.");
1510         bool big_value = false;
1511         if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1512           if (scope_field->is_location()) {
1513             Location::Type type = ((LocationValue*) scope_field)->location().type();
1514             if (type == Location::dbl || type == Location::lng) {
1515               big_value = true;
1516             }
1517           }
1518           if (scope_field->is_constant_int()) {
1519             ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1520             if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1521               big_value = true;
1522             }
1523           }
1524         }
1525 
1526         if (big_value) {
1527           i++;
1528           assert(i < fields->length(), "second T_INT field needed");
1529           assert(fields->at(i)._type == T_INT, "T_INT field needed");
1530         } else {
1531           obj->int_field_put(offset, value->get_jint());
1532           break;
1533         }
1534       }
1535         /* no break */
1536 
1537       case T_LONG: case T_DOUBLE: {
1538         assert(value->type() == T_INT, "Agreement.");
1539         StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1540 #ifdef _LP64
1541         jlong res = (jlong)low->get_intptr();
1542 #else
1543         jlong res = jlong_from(value->get_jint(), low->get_jint());
1544 #endif
1545         obj->long_field_put(offset, res);
1546         break;
1547       }
1548 
1549       case T_SHORT:
1550         assert(value->type() == T_INT, "Agreement.");
1551         obj->short_field_put(offset, (jshort)value->get_jint());
1552         break;
1553 
1554       case T_CHAR:
1555         assert(value->type() == T_INT, "Agreement.");
1556         obj->char_field_put(offset, (jchar)value->get_jint());
1557         break;
1558 
1559       case T_BYTE:
1560         assert(value->type() == T_INT, "Agreement.");
1561         obj->byte_field_put(offset, (jbyte)value->get_jint());
1562         break;
1563 
1564       case T_BOOLEAN:
1565         assert(value->type() == T_INT, "Agreement.");
1566         obj->bool_field_put(offset, (jboolean)value->get_jint());
1567         break;
1568 
1569       default:
1570         ShouldNotReachHere();
1571     }
1572     svIndex++;
1573   }
1574   return svIndex;
1575 }
1576 
1577 // restore fields of all eliminated objects and arrays
1578 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool is_jvmci) {
1579   for (int i = 0; i < objects->length(); i++) {
1580     assert(objects->at(i)->is_object(), "invalid debug information");
1581     ObjectValue* sv = (ObjectValue*) objects->at(i);
1582     Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1583     Handle obj = sv->value();
1584     assert(obj.not_null() || realloc_failures, "reallocation was missed");
1585 #ifndef PRODUCT
1586     if (PrintDeoptimizationDetails) {
1587       tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1588     }
1589 #endif // !PRODUCT
1590 
1591     if (obj.is_null()) {
1592       continue;
1593     }
1594 
1595 #if INCLUDE_JVMCI
1596     // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1597     if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1598       continue;
1599     }
1600 #endif // INCLUDE_JVMCI
1601     if (EnableVectorSupport && VectorSupport::is_vector(k)) {
1602       assert(sv->field_size() == 1, "%s not a vector", k->name()->as_C_string());
1603       ScopeValue* payload = sv->field_at(0);
1604       if (payload->is_location() &&
1605           payload->as_LocationValue()->location().type() == Location::vector) {
1606 #ifndef PRODUCT
1607         if (PrintDeoptimizationDetails) {
1608           tty->print_cr("skip field reassignment for this vector - it should be assigned already");
1609           if (Verbose) {
1610             Handle obj = sv->value();
1611             k->oop_print_on(obj(), tty);
1612           }
1613         }
1614 #endif // !PRODUCT
1615         continue; // Such vector's value was already restored in VectorSupport::allocate_vector().
1616       }
1617       // Else fall-through to do assignment for scalar-replaced boxed vector representation
1618       // which could be restored after vector object allocation.
1619     }
1620     if (k->is_instance_klass()) {
1621       InstanceKlass* ik = InstanceKlass::cast(k);
1622       reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), is_jvmci);
1623     } else if (k->is_typeArray_klass()) {
1624       TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1625       reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1626     } else if (k->is_objArray_klass()) {
1627       reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1628     }
1629   }
1630   // These objects may escape when we return to Interpreter after deoptimization.
1631   // We need barrier so that stores that initialize these objects can't be reordered
1632   // with subsequent stores that make these objects accessible by other threads.
1633   OrderAccess::storestore();
1634 }
1635 
1636 
1637 // relock objects for which synchronization was eliminated
1638 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1639                                     JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1640   bool relocked_objects = false;
1641   for (int i = 0; i < monitors->length(); i++) {
1642     MonitorInfo* mon_info = monitors->at(i);
1643     if (mon_info->eliminated()) {
1644       assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1645       relocked_objects = true;
1646       if (!mon_info->owner_is_scalar_replaced()) {
1647         Handle obj(thread, mon_info->owner());
1648         markWord mark = obj->mark();
1649         if (exec_mode == Unpack_none) {
1650           if (LockingMode == LM_LEGACY && mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1651             // With exec_mode == Unpack_none obj may be thread local and locked in
1652             // a callee frame. Make the lock in the callee a recursive lock and restore the displaced header.
1653             markWord dmw = mark.displaced_mark_helper();
1654             mark.locker()->set_displaced_header(markWord::encode((BasicLock*) nullptr));
1655             obj->set_mark(dmw);
1656           }
1657           if (mark.has_monitor()) {
1658             // defer relocking if the deoptee thread is currently waiting for obj
1659             ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1660             if (waiting_monitor != nullptr && waiting_monitor->object() == obj()) {
1661               assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1662               if (LockingMode == LM_LEGACY) {
1663                 mon_info->lock()->set_displaced_header(markWord::unused_mark());
1664               } else if (UseObjectMonitorTable) {
1665                 mon_info->lock()->clear_object_monitor_cache();
1666               }
1667 #ifdef ASSERT
1668               else {
1669                 assert(LockingMode == LM_MONITOR || !UseObjectMonitorTable, "must be");
1670                 mon_info->lock()->set_bad_metadata_deopt();
1671               }
1672 #endif
1673               JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1674               continue;
1675             }
1676           }
1677         }
1678         BasicLock* lock = mon_info->lock();
1679         if (LockingMode == LM_LIGHTWEIGHT) {
1680           // We have lost information about the correct state of the lock stack.
1681           // Entering may create an invalid lock stack. Inflate the lock if it
1682           // was fast_locked to restore the valid lock stack.
1683           if (UseObjectMonitorTable) {
1684             // UseObjectMonitorTable expects the BasicLock cache to be either a
1685             // valid ObjectMonitor* or nullptr. Right now it is garbage, set it
1686             // to nullptr.
1687             lock->clear_object_monitor_cache();
1688           }
1689           ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1690           if (deoptee_thread->lock_stack().contains(obj())) {
1691             LightweightSynchronizer::inflate_fast_locked_object(obj(), ObjectSynchronizer::InflateCause::inflate_cause_vm_internal,
1692                                                                 deoptee_thread, thread);
1693           }
1694           assert(mon_info->owner()->is_locked(), "object must be locked now");
1695           assert(obj->mark().has_monitor(), "must be");
1696           assert(!deoptee_thread->lock_stack().contains(obj()), "must be");
1697           assert(ObjectSynchronizer::read_monitor(thread, obj(), obj->mark())->has_owner(deoptee_thread), "must be");
1698         } else {
1699           ObjectSynchronizer::enter_for(obj, lock, deoptee_thread);
1700           assert(mon_info->owner()->is_locked(), "object must be locked now");
1701         }
1702       }
1703     }
1704   }
1705   return relocked_objects;
1706 }
1707 #endif // COMPILER2_OR_JVMCI
1708 
1709 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1710   Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1711 
1712   // Register map for next frame (used for stack crawl).  We capture
1713   // the state of the deopt'ing frame's caller.  Thus if we need to
1714   // stuff a C2I adapter we can properly fill in the callee-save
1715   // register locations.
1716   frame caller = fr.sender(reg_map);
1717   int frame_size = pointer_delta_as_int(caller.sp(), fr.sp());
1718 
1719   frame sender = caller;
1720 
1721   // Since the Java thread being deoptimized will eventually adjust it's own stack,
1722   // the vframeArray containing the unpacking information is allocated in the C heap.
1723   // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1724   vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1725 
1726   // Compare the vframeArray to the collected vframes
1727   assert(array->structural_compare(thread, chunk), "just checking");
1728 
1729   if (TraceDeoptimization) {
1730     ResourceMark rm;
1731     stringStream st;
1732     st.print_cr("DEOPT PACKING thread=" INTPTR_FORMAT " vframeArray=" INTPTR_FORMAT, p2i(thread), p2i(array));
1733     st.print("   ");
1734     fr.print_on(&st);
1735     st.print_cr("   Virtual frames (innermost/newest first):");
1736     for (int index = 0; index < chunk->length(); index++) {
1737       compiledVFrame* vf = chunk->at(index);
1738       int bci = vf->raw_bci();
1739       const char* code_name;
1740       if (bci == SynchronizationEntryBCI) {
1741         code_name = "sync entry";
1742       } else {
1743         Bytecodes::Code code = vf->method()->code_at(bci);
1744         code_name = Bytecodes::name(code);
1745       }
1746 
1747       st.print("      VFrame %d (" INTPTR_FORMAT ")", index, p2i(vf));
1748       st.print(" - %s", vf->method()->name_and_sig_as_C_string());
1749       st.print(" - %s", code_name);
1750       st.print_cr(" @ bci=%d ", bci);
1751     }
1752     tty->print_raw(st.freeze());
1753     tty->cr();
1754   }
1755 
1756   return array;
1757 }
1758 
1759 #if COMPILER2_OR_JVMCI
1760 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1761   // Reallocation of some scalar replaced objects failed. Record
1762   // that we need to pop all the interpreter frames for the
1763   // deoptimized compiled frame.
1764   assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1765   thread->set_frames_to_pop_failed_realloc(array->frames());
1766   // Unlock all monitors here otherwise the interpreter will see a
1767   // mix of locked and unlocked monitors (because of failed
1768   // reallocations of synchronized objects) and be confused.
1769   for (int i = 0; i < array->frames(); i++) {
1770     MonitorChunk* monitors = array->element(i)->monitors();
1771     if (monitors != nullptr) {
1772       // Unlock in reverse order starting from most nested monitor.
1773       for (int j = (monitors->number_of_monitors() - 1); j >= 0; j--) {
1774         BasicObjectLock* src = monitors->at(j);
1775         if (src->obj() != nullptr) {
1776           ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1777         }
1778       }
1779       array->element(i)->free_monitors();
1780 #ifdef ASSERT
1781       array->element(i)->set_removed_monitors();
1782 #endif
1783     }
1784   }
1785 }
1786 #endif
1787 
1788 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1789   assert(fr.can_be_deoptimized(), "checking frame type");
1790 
1791   gather_statistics(reason, Action_none, Bytecodes::_illegal);
1792 
1793   if (LogCompilation && xtty != nullptr) {
1794     nmethod* nm = fr.cb()->as_nmethod_or_null();
1795     assert(nm != nullptr, "only compiled methods can deopt");
1796 
1797     ttyLocker ttyl;
1798     xtty->begin_head("deoptimized thread='%zu' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1799     nm->log_identity(xtty);
1800     xtty->end_head();
1801     for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1802       xtty->begin_elem("jvms bci='%d'", sd->bci());
1803       xtty->method(sd->method());
1804       xtty->end_elem();
1805       if (sd->is_top())  break;
1806     }
1807     xtty->tail("deoptimized");
1808   }
1809 
1810   Continuation::notify_deopt(thread, fr.sp());
1811 
1812   // Patch the compiled method so that when execution returns to it we will
1813   // deopt the execution state and return to the interpreter.
1814   fr.deoptimize(thread);
1815 }
1816 
1817 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1818   // Deoptimize only if the frame comes from compile code.
1819   // Do not deoptimize the frame which is already patched
1820   // during the execution of the loops below.
1821   if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1822     return;
1823   }
1824   ResourceMark rm;
1825   deoptimize_single_frame(thread, fr, reason);
1826 }
1827 
1828 #if INCLUDE_JVMCI
1829 address Deoptimization::deoptimize_for_missing_exception_handler(nmethod* nm) {
1830   // there is no exception handler for this pc => deoptimize
1831   nm->make_not_entrant(nmethod::InvalidationReason::MISSING_EXCEPTION_HANDLER);
1832 
1833   // Use Deoptimization::deoptimize for all of its side-effects:
1834   // gathering traps statistics, logging...
1835   // it also patches the return pc but we do not care about that
1836   // since we return a continuation to the deopt_blob below.
1837   JavaThread* thread = JavaThread::current();
1838   RegisterMap reg_map(thread,
1839                       RegisterMap::UpdateMap::skip,
1840                       RegisterMap::ProcessFrames::include,
1841                       RegisterMap::WalkContinuation::skip);
1842   frame runtime_frame = thread->last_frame();
1843   frame caller_frame = runtime_frame.sender(&reg_map);
1844   assert(caller_frame.cb()->as_nmethod_or_null() == nm, "expect top frame compiled method");
1845   vframe* vf = vframe::new_vframe(&caller_frame, &reg_map, thread);
1846   compiledVFrame* cvf = compiledVFrame::cast(vf);
1847   ScopeDesc* imm_scope = cvf->scope();
1848   MethodData* imm_mdo = get_method_data(thread, methodHandle(thread, imm_scope->method()), true);
1849   if (imm_mdo != nullptr) {
1850     // Lock to read ProfileData, and ensure lock is not broken by a safepoint
1851     MutexLocker ml(imm_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
1852 
1853     ProfileData* pdata = imm_mdo->allocate_bci_to_data(imm_scope->bci(), nullptr);
1854     if (pdata != nullptr && pdata->is_BitData()) {
1855       BitData* bit_data = (BitData*) pdata;
1856       bit_data->set_exception_seen();
1857     }
1858   }
1859 
1860   Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1861 
1862   MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, nm->method()), true);
1863   if (trap_mdo != nullptr) {
1864     trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1865   }
1866 
1867   return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1868 }
1869 #endif
1870 
1871 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1872   assert(thread == Thread::current() ||
1873          thread->is_handshake_safe_for(Thread::current()) ||
1874          SafepointSynchronize::is_at_safepoint(),
1875          "can only deoptimize other thread at a safepoint/handshake");
1876   // Compute frame and register map based on thread and sp.
1877   RegisterMap reg_map(thread,
1878                       RegisterMap::UpdateMap::skip,
1879                       RegisterMap::ProcessFrames::include,
1880                       RegisterMap::WalkContinuation::skip);
1881   frame fr = thread->last_frame();
1882   while (fr.id() != id) {
1883     fr = fr.sender(&reg_map);
1884   }
1885   deoptimize(thread, fr, reason);
1886 }
1887 
1888 
1889 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1890   Thread* current = Thread::current();
1891   if (thread == current || thread->is_handshake_safe_for(current)) {
1892     Deoptimization::deoptimize_frame_internal(thread, id, reason);
1893   } else {
1894     VM_DeoptimizeFrame deopt(thread, id, reason);
1895     VMThread::execute(&deopt);
1896   }
1897 }
1898 
1899 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1900   deoptimize_frame(thread, id, Reason_constraint);
1901 }
1902 
1903 // JVMTI PopFrame support
1904 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1905 {
1906   assert(thread == JavaThread::current(), "pre-condition");
1907   thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1908 }
1909 JRT_END
1910 
1911 MethodData*
1912 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1913                                 bool create_if_missing) {
1914   JavaThread* THREAD = thread; // For exception macros.
1915   MethodData* mdo = m()->method_data();
1916   if (mdo == nullptr && create_if_missing && !HAS_PENDING_EXCEPTION) {
1917     // Build an MDO.  Ignore errors like OutOfMemory;
1918     // that simply means we won't have an MDO to update.
1919     Method::build_profiling_method_data(m, THREAD);
1920     if (HAS_PENDING_EXCEPTION) {
1921       // Only metaspace OOM is expected. No Java code executed.
1922       assert((PENDING_EXCEPTION->is_a(vmClasses::OutOfMemoryError_klass())), "we expect only an OOM error here");
1923       CLEAR_PENDING_EXCEPTION;
1924     }
1925     mdo = m()->method_data();
1926   }
1927   return mdo;
1928 }
1929 
1930 #if COMPILER2_OR_JVMCI
1931 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1932   // In case of an unresolved klass entry, load the class.
1933   // This path is exercised from case _ldc in Parse::do_one_bytecode,
1934   // and probably nowhere else.
1935   // Even that case would benefit from simply re-interpreting the
1936   // bytecode, without paying special attention to the class index.
1937   // So this whole "class index" feature should probably be removed.
1938 
1939   if (constant_pool->tag_at(index).is_unresolved_klass()) {
1940     Klass* tk = constant_pool->klass_at(index, THREAD);
1941     if (HAS_PENDING_EXCEPTION) {
1942       // Exception happened during classloading. We ignore the exception here, since it
1943       // is going to be rethrown since the current activation is going to be deoptimized and
1944       // the interpreter will re-execute the bytecode.
1945       // Do not clear probable Async Exceptions.
1946       CLEAR_PENDING_NONASYNC_EXCEPTION;
1947       // Class loading called java code which may have caused a stack
1948       // overflow. If the exception was thrown right before the return
1949       // to the runtime the stack is no longer guarded. Reguard the
1950       // stack otherwise if we return to the uncommon trap blob and the
1951       // stack bang causes a stack overflow we crash.
1952       JavaThread* jt = THREAD;
1953       bool guard_pages_enabled = jt->stack_overflow_state()->reguard_stack_if_needed();
1954       assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1955     }
1956     return;
1957   }
1958 
1959   assert(!constant_pool->tag_at(index).is_symbol(),
1960          "no symbolic names here, please");
1961 }
1962 
1963 #if INCLUDE_JFR
1964 
1965 class DeoptReasonSerializer : public JfrSerializer {
1966  public:
1967   void serialize(JfrCheckpointWriter& writer) {
1968     writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1969     for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1970       writer.write_key((u8)i);
1971       writer.write(Deoptimization::trap_reason_name(i));
1972     }
1973   }
1974 };
1975 
1976 class DeoptActionSerializer : public JfrSerializer {
1977  public:
1978   void serialize(JfrCheckpointWriter& writer) {
1979     static const u4 nof_actions = Deoptimization::Action_LIMIT;
1980     writer.write_count(nof_actions);
1981     for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1982       writer.write_key(i);
1983       writer.write(Deoptimization::trap_action_name((int)i));
1984     }
1985   }
1986 };
1987 
1988 static void register_serializers() {
1989   static int critical_section = 0;
1990   if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1991     return;
1992   }
1993   JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1994   JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1995 }
1996 
1997 static void post_deoptimization_event(nmethod* nm,
1998                                       const Method* method,
1999                                       int trap_bci,
2000                                       int instruction,
2001                                       Deoptimization::DeoptReason reason,
2002                                       Deoptimization::DeoptAction action) {
2003   assert(nm != nullptr, "invariant");
2004   assert(method != nullptr, "invariant");
2005   if (EventDeoptimization::is_enabled()) {
2006     static bool serializers_registered = false;
2007     if (!serializers_registered) {
2008       register_serializers();
2009       serializers_registered = true;
2010     }
2011     EventDeoptimization event;
2012     event.set_compileId(nm->compile_id());
2013     event.set_compiler(nm->compiler_type());
2014     event.set_method(method);
2015     event.set_lineNumber(method->line_number_from_bci(trap_bci));
2016     event.set_bci(trap_bci);
2017     event.set_instruction(instruction);
2018     event.set_reason(reason);
2019     event.set_action(action);
2020     event.commit();
2021   }
2022 }
2023 
2024 #endif // INCLUDE_JFR
2025 
2026 static void log_deopt(nmethod* nm, Method* tm, intptr_t pc, frame& fr, int trap_bci,
2027                               const char* reason_name, const char* reason_action) {
2028   LogTarget(Debug, deoptimization) lt;
2029   if (lt.is_enabled()) {
2030     LogStream ls(lt);
2031     bool is_osr = nm->is_osr_method();
2032     ls.print("cid=%4d %s level=%d",
2033              nm->compile_id(), (is_osr ? "osr" : "   "), nm->comp_level());
2034     ls.print(" %s", tm->name_and_sig_as_C_string());
2035     ls.print(" trap_bci=%d ", trap_bci);
2036     if (is_osr) {
2037       ls.print("osr_bci=%d ", nm->osr_entry_bci());
2038     }
2039     ls.print("%s ", reason_name);
2040     ls.print("%s ", reason_action);
2041     ls.print_cr("pc=" INTPTR_FORMAT " relative_pc=" INTPTR_FORMAT,
2042              pc, fr.pc() - nm->code_begin());
2043   }
2044 }
2045 
2046 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* current, jint trap_request)) {
2047   HandleMark hm(current);
2048 
2049   // uncommon_trap() is called at the beginning of the uncommon trap
2050   // handler. Note this fact before we start generating temporary frames
2051   // that can confuse an asynchronous stack walker. This counter is
2052   // decremented at the end of unpack_frames().
2053 
2054   current->inc_in_deopt_handler();
2055 
2056 #if INCLUDE_JVMCI
2057   // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
2058   RegisterMap reg_map(current,
2059                       RegisterMap::UpdateMap::include,
2060                       RegisterMap::ProcessFrames::include,
2061                       RegisterMap::WalkContinuation::skip);
2062 #else
2063   RegisterMap reg_map(current,
2064                       RegisterMap::UpdateMap::skip,
2065                       RegisterMap::ProcessFrames::include,
2066                       RegisterMap::WalkContinuation::skip);
2067 #endif
2068   frame stub_frame = current->last_frame();
2069   frame fr = stub_frame.sender(&reg_map);
2070 
2071   // Log a message
2072   Events::log_deopt_message(current, "Uncommon trap: trap_request=" INT32_FORMAT_X_0 " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
2073               trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
2074 
2075   {
2076     ResourceMark rm;
2077 
2078     DeoptReason reason = trap_request_reason(trap_request);
2079     DeoptAction action = trap_request_action(trap_request);
2080 #if INCLUDE_JVMCI
2081     int debug_id = trap_request_debug_id(trap_request);
2082 #endif
2083     jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
2084 
2085     vframe*  vf  = vframe::new_vframe(&fr, &reg_map, current);
2086     compiledVFrame* cvf = compiledVFrame::cast(vf);
2087 
2088     nmethod* nm = cvf->code();
2089 
2090     ScopeDesc*      trap_scope  = cvf->scope();
2091 
2092     bool is_receiver_constraint_failure = COMPILER2_PRESENT(VerifyReceiverTypes &&) (reason == Deoptimization::Reason_receiver_constraint);
2093 
2094     if (is_receiver_constraint_failure) {
2095       tty->print_cr("  bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"),
2096                     trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
2097                     JVMCI_ONLY(COMMA debug_id));
2098     }
2099 
2100     methodHandle    trap_method(current, trap_scope->method());
2101     int             trap_bci    = trap_scope->bci();
2102 #if INCLUDE_JVMCI
2103     jlong           speculation = current->pending_failed_speculation();
2104     if (nm->is_compiled_by_jvmci()) {
2105       nm->update_speculation(current);
2106     } else {
2107       assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
2108     }
2109 
2110     if (trap_bci == SynchronizationEntryBCI) {
2111       trap_bci = 0;
2112       current->set_pending_monitorenter(true);
2113     }
2114 
2115     if (reason == Deoptimization::Reason_transfer_to_interpreter) {
2116       current->set_pending_transfer_to_interpreter(true);
2117     }
2118 #endif
2119 
2120     Bytecodes::Code trap_bc     = trap_method->java_code_at(trap_bci);
2121     // Record this event in the histogram.
2122     gather_statistics(reason, action, trap_bc);
2123 
2124     // Ensure that we can record deopt. history:
2125     bool create_if_missing = ProfileTraps;
2126 
2127     methodHandle profiled_method;
2128 #if INCLUDE_JVMCI
2129     if (nm->is_compiled_by_jvmci()) {
2130       profiled_method = methodHandle(current, nm->method());
2131     } else {
2132       profiled_method = trap_method;
2133     }
2134 #else
2135     profiled_method = trap_method;
2136 #endif
2137 
2138     MethodData* trap_mdo =
2139       get_method_data(current, profiled_method, create_if_missing);
2140 
2141     { // Log Deoptimization event for JFR, UL and event system
2142       Method* tm = trap_method();
2143       const char* reason_name = trap_reason_name(reason);
2144       const char* reason_action = trap_action_name(action);
2145       intptr_t pc = p2i(fr.pc());
2146 
2147       JFR_ONLY(post_deoptimization_event(nm, tm, trap_bci, trap_bc, reason, action);)
2148       log_deopt(nm, tm, pc, fr, trap_bci, reason_name, reason_action);
2149       Events::log_deopt_message(current, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
2150                                 reason_name, reason_action, pc,
2151                                 tm->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
2152     }
2153 
2154     // Print a bunch of diagnostics, if requested.
2155     if (TraceDeoptimization || LogCompilation || is_receiver_constraint_failure) {
2156       ResourceMark rm;
2157 
2158       // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2159       // We must do this already now, since we cannot acquire this lock while
2160       // holding the tty lock (lock ordering by rank).
2161       MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2162 
2163       ttyLocker ttyl;
2164 
2165       char buf[100];
2166       if (xtty != nullptr) {
2167         xtty->begin_head("uncommon_trap thread='%zu' %s",
2168                          os::current_thread_id(),
2169                          format_trap_request(buf, sizeof(buf), trap_request));
2170 #if INCLUDE_JVMCI
2171         if (speculation != 0) {
2172           xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
2173         }
2174 #endif
2175         nm->log_identity(xtty);
2176       }
2177       Symbol* class_name = nullptr;
2178       bool unresolved = false;
2179       if (unloaded_class_index >= 0) {
2180         constantPoolHandle constants (current, trap_method->constants());
2181         if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2182           class_name = constants->klass_name_at(unloaded_class_index);
2183           unresolved = true;
2184           if (xtty != nullptr)
2185             xtty->print(" unresolved='1'");
2186         } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2187           class_name = constants->symbol_at(unloaded_class_index);
2188         }
2189         if (xtty != nullptr)
2190           xtty->name(class_name);
2191       }
2192       if (xtty != nullptr && trap_mdo != nullptr && (int)reason < (int)MethodData::_trap_hist_limit) {
2193         // Dump the relevant MDO state.
2194         // This is the deopt count for the current reason, any previous
2195         // reasons or recompiles seen at this point.
2196         int dcnt = trap_mdo->trap_count(reason);
2197         if (dcnt != 0)
2198           xtty->print(" count='%d'", dcnt);
2199 
2200         // We need to lock to read the ProfileData. But to keep the locks ordered, we need to
2201         // lock extra_data_lock before the tty lock.
2202         ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2203         int dos = (pdata == nullptr)? 0: pdata->trap_state();
2204         if (dos != 0) {
2205           xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2206           if (trap_state_is_recompiled(dos)) {
2207             int recnt2 = trap_mdo->overflow_recompile_count();
2208             if (recnt2 != 0)
2209               xtty->print(" recompiles2='%d'", recnt2);
2210           }
2211         }
2212       }
2213       if (xtty != nullptr) {
2214         xtty->stamp();
2215         xtty->end_head();
2216       }
2217       if (TraceDeoptimization) {  // make noise on the tty
2218         stringStream st;
2219         st.print("UNCOMMON TRAP method=%s", trap_scope->method()->name_and_sig_as_C_string());
2220         st.print("  bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT JVMCI_ONLY(", debug_id=%d"),
2221                  trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin() JVMCI_ONLY(COMMA debug_id));
2222         st.print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2223 #if INCLUDE_JVMCI
2224         if (nm->is_compiled_by_jvmci()) {
2225           const char* installed_code_name = nm->jvmci_name();
2226           if (installed_code_name != nullptr) {
2227             st.print(" (JVMCI: installed code name=%s) ", installed_code_name);
2228           }
2229         }
2230 #endif
2231         st.print(" (@" INTPTR_FORMAT ") thread=%zu reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2232                    p2i(fr.pc()),
2233                    os::current_thread_id(),
2234                    trap_reason_name(reason),
2235                    trap_action_name(action),
2236                    unloaded_class_index
2237 #if INCLUDE_JVMCI
2238                    , debug_id
2239 #endif
2240                    );
2241         if (class_name != nullptr) {
2242           st.print(unresolved ? " unresolved class: " : " symbol: ");
2243           class_name->print_symbol_on(&st);
2244         }
2245         st.cr();
2246         tty->print_raw(st.freeze());
2247       }
2248       if (xtty != nullptr) {
2249         // Log the precise location of the trap.
2250         for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2251           xtty->begin_elem("jvms bci='%d'", sd->bci());
2252           xtty->method(sd->method());
2253           xtty->end_elem();
2254           if (sd->is_top())  break;
2255         }
2256         xtty->tail("uncommon_trap");
2257       }
2258     }
2259     // (End diagnostic printout.)
2260 
2261     if (is_receiver_constraint_failure) {
2262       fatal("missing receiver type check");
2263     }
2264 
2265     // Load class if necessary
2266     if (unloaded_class_index >= 0) {
2267       constantPoolHandle constants(current, trap_method->constants());
2268       load_class_by_index(constants, unloaded_class_index, THREAD);
2269     }
2270 
2271     // Flush the nmethod if necessary and desirable.
2272     //
2273     // We need to avoid situations where we are re-flushing the nmethod
2274     // because of a hot deoptimization site.  Repeated flushes at the same
2275     // point need to be detected by the compiler and avoided.  If the compiler
2276     // cannot avoid them (or has a bug and "refuses" to avoid them), this
2277     // module must take measures to avoid an infinite cycle of recompilation
2278     // and deoptimization.  There are several such measures:
2279     //
2280     //   1. If a recompilation is ordered a second time at some site X
2281     //   and for the same reason R, the action is adjusted to 'reinterpret',
2282     //   to give the interpreter time to exercise the method more thoroughly.
2283     //   If this happens, the method's overflow_recompile_count is incremented.
2284     //
2285     //   2. If the compiler fails to reduce the deoptimization rate, then
2286     //   the method's overflow_recompile_count will begin to exceed the set
2287     //   limit PerBytecodeRecompilationCutoff.  If this happens, the action
2288     //   is adjusted to 'make_not_compilable', and the method is abandoned
2289     //   to the interpreter.  This is a performance hit for hot methods,
2290     //   but is better than a disastrous infinite cycle of recompilations.
2291     //   (Actually, only the method containing the site X is abandoned.)
2292     //
2293     //   3. In parallel with the previous measures, if the total number of
2294     //   recompilations of a method exceeds the much larger set limit
2295     //   PerMethodRecompilationCutoff, the method is abandoned.
2296     //   This should only happen if the method is very large and has
2297     //   many "lukewarm" deoptimizations.  The code which enforces this
2298     //   limit is elsewhere (class nmethod, class Method).
2299     //
2300     // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2301     // to recompile at each bytecode independently of the per-BCI cutoff.
2302     //
2303     // The decision to update code is up to the compiler, and is encoded
2304     // in the Action_xxx code.  If the compiler requests Action_none
2305     // no trap state is changed, no compiled code is changed, and the
2306     // computation suffers along in the interpreter.
2307     //
2308     // The other action codes specify various tactics for decompilation
2309     // and recompilation.  Action_maybe_recompile is the loosest, and
2310     // allows the compiled code to stay around until enough traps are seen,
2311     // and until the compiler gets around to recompiling the trapping method.
2312     //
2313     // The other actions cause immediate removal of the present code.
2314 
2315     // Traps caused by injected profile shouldn't pollute trap counts.
2316     bool injected_profile_trap = trap_method->has_injected_profile() &&
2317                                  (reason == Reason_intrinsic || reason == Reason_unreached);
2318 
2319     bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2320     bool make_not_entrant = false;
2321     bool make_not_compilable = false;
2322     bool reprofile = false;
2323     switch (action) {
2324     case Action_none:
2325       // Keep the old code.
2326       update_trap_state = false;
2327       break;
2328     case Action_maybe_recompile:
2329       // Do not need to invalidate the present code, but we can
2330       // initiate another
2331       // Start compiler without (necessarily) invalidating the nmethod.
2332       // The system will tolerate the old code, but new code should be
2333       // generated when possible.
2334       break;
2335     case Action_reinterpret:
2336       // Go back into the interpreter for a while, and then consider
2337       // recompiling form scratch.
2338       make_not_entrant = true;
2339       // Reset invocation counter for outer most method.
2340       // This will allow the interpreter to exercise the bytecodes
2341       // for a while before recompiling.
2342       // By contrast, Action_make_not_entrant is immediate.
2343       //
2344       // Note that the compiler will track null_check, null_assert,
2345       // range_check, and class_check events and log them as if they
2346       // had been traps taken from compiled code.  This will update
2347       // the MDO trap history so that the next compilation will
2348       // properly detect hot trap sites.
2349       reprofile = true;
2350       break;
2351     case Action_make_not_entrant:
2352       // Request immediate recompilation, and get rid of the old code.
2353       // Make them not entrant, so next time they are called they get
2354       // recompiled.  Unloaded classes are loaded now so recompile before next
2355       // time they are called.  Same for uninitialized.  The interpreter will
2356       // link the missing class, if any.
2357       make_not_entrant = true;
2358       break;
2359     case Action_make_not_compilable:
2360       // Give up on compiling this method at all.
2361       make_not_entrant = true;
2362       make_not_compilable = true;
2363       break;
2364     default:
2365       ShouldNotReachHere();
2366     }
2367 
2368 #if INCLUDE_JVMCI
2369     // Deoptimization count is used by the CompileBroker to reason about compilations
2370     // it requests so do not pollute the count for deoptimizations in non-default (i.e.
2371     // non-CompilerBroker) compilations.
2372     if (nm->jvmci_skip_profile_deopt()) {
2373       update_trap_state = false;
2374     }
2375 #endif
2376     // Setting +ProfileTraps fixes the following, on all platforms:
2377     // The result is infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2378     // recompile relies on a MethodData* to record heroic opt failures.
2379 
2380     // Whether the interpreter is producing MDO data or not, we also need
2381     // to use the MDO to detect hot deoptimization points and control
2382     // aggressive optimization.
2383     bool inc_recompile_count = false;
2384 
2385     // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2386     ConditionalMutexLocker ml((trap_mdo != nullptr) ? trap_mdo->extra_data_lock() : nullptr,
2387                               (trap_mdo != nullptr),
2388                               Mutex::_no_safepoint_check_flag);
2389     ProfileData* pdata = nullptr;
2390     if (ProfileTraps && CompilerConfig::is_c2_or_jvmci_compiler_enabled() && update_trap_state && trap_mdo != nullptr) {
2391       assert(trap_mdo == get_method_data(current, profiled_method, false), "sanity");
2392       uint this_trap_count = 0;
2393       bool maybe_prior_trap = false;
2394       bool maybe_prior_recompile = false;
2395 
2396       pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2397 #if INCLUDE_JVMCI
2398                                    nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2399 #endif
2400                                    nm->method(),
2401                                    //outputs:
2402                                    this_trap_count,
2403                                    maybe_prior_trap,
2404                                    maybe_prior_recompile);
2405       // Because the interpreter also counts null, div0, range, and class
2406       // checks, these traps from compiled code are double-counted.
2407       // This is harmless; it just means that the PerXTrapLimit values
2408       // are in effect a little smaller than they look.
2409 
2410       DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2411       if (per_bc_reason != Reason_none) {
2412         // Now take action based on the partially known per-BCI history.
2413         if (maybe_prior_trap
2414             && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2415           // If there are too many traps at this BCI, force a recompile.
2416           // This will allow the compiler to see the limit overflow, and
2417           // take corrective action, if possible.  The compiler generally
2418           // does not use the exact PerBytecodeTrapLimit value, but instead
2419           // changes its tactics if it sees any traps at all.  This provides
2420           // a little hysteresis, delaying a recompile until a trap happens
2421           // several times.
2422           //
2423           // Actually, since there is only one bit of counter per BCI,
2424           // the possible per-BCI counts are {0,1,(per-method count)}.
2425           // This produces accurate results if in fact there is only
2426           // one hot trap site, but begins to get fuzzy if there are
2427           // many sites.  For example, if there are ten sites each
2428           // trapping two or more times, they each get the blame for
2429           // all of their traps.
2430           make_not_entrant = true;
2431         }
2432 
2433         // Detect repeated recompilation at the same BCI, and enforce a limit.
2434         if (make_not_entrant && maybe_prior_recompile) {
2435           // More than one recompile at this point.
2436           inc_recompile_count = maybe_prior_trap;
2437         }
2438       } else {
2439         // For reasons which are not recorded per-bytecode, we simply
2440         // force recompiles unconditionally.
2441         // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2442         make_not_entrant = true;
2443       }
2444 
2445       // Go back to the compiler if there are too many traps in this method.
2446       if (this_trap_count >= per_method_trap_limit(reason)) {
2447         // If there are too many traps in this method, force a recompile.
2448         // This will allow the compiler to see the limit overflow, and
2449         // take corrective action, if possible.
2450         // (This condition is an unlikely backstop only, because the
2451         // PerBytecodeTrapLimit is more likely to take effect first,
2452         // if it is applicable.)
2453         make_not_entrant = true;
2454       }
2455 
2456       // Here's more hysteresis:  If there has been a recompile at
2457       // this trap point already, run the method in the interpreter
2458       // for a while to exercise it more thoroughly.
2459       if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2460         reprofile = true;
2461       }
2462     }
2463 
2464     // Take requested actions on the method:
2465 
2466     // Recompile
2467     if (make_not_entrant) {
2468       if (!nm->make_not_entrant(nmethod::InvalidationReason::UNCOMMON_TRAP)) {
2469         return; // the call did not change nmethod's state
2470       }
2471 
2472       if (pdata != nullptr) {
2473         // Record the recompilation event, if any.
2474         int tstate0 = pdata->trap_state();
2475         int tstate1 = trap_state_set_recompiled(tstate0, true);
2476         if (tstate1 != tstate0)
2477           pdata->set_trap_state(tstate1);
2478       }
2479 
2480       // For code aging we count traps separately here, using make_not_entrant()
2481       // as a guard against simultaneous deopts in multiple threads.
2482       if (reason == Reason_tenured && trap_mdo != nullptr) {
2483         trap_mdo->inc_tenure_traps();
2484       }
2485     }
2486     if (inc_recompile_count) {
2487       trap_mdo->inc_overflow_recompile_count();
2488       if ((uint)trap_mdo->overflow_recompile_count() >
2489           (uint)PerBytecodeRecompilationCutoff) {
2490         // Give up on the method containing the bad BCI.
2491         if (trap_method() == nm->method()) {
2492           make_not_compilable = true;
2493         } else {
2494           trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2495           // But give grace to the enclosing nm->method().
2496         }
2497       }
2498     }
2499 
2500     // Reprofile
2501     if (reprofile) {
2502       CompilationPolicy::reprofile(trap_scope, nm->is_osr_method());
2503     }
2504 
2505     // Give up compiling
2506     if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2507       assert(make_not_entrant, "consistent");
2508       nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2509     }
2510 
2511     if (ProfileExceptionHandlers && trap_mdo != nullptr) {
2512       BitData* exception_handler_data = trap_mdo->exception_handler_bci_to_data_or_null(trap_bci);
2513       if (exception_handler_data != nullptr) {
2514         // uncommon trap at the start of an exception handler.
2515         // C2 generates these for un-entered exception handlers.
2516         // mark the handler as entered to avoid generating
2517         // another uncommon trap the next time the handler is compiled
2518         exception_handler_data->set_exception_handler_entered();
2519       }
2520     }
2521 
2522   } // Free marked resources
2523 
2524 }
2525 JRT_END
2526 
2527 ProfileData*
2528 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2529                                          int trap_bci,
2530                                          Deoptimization::DeoptReason reason,
2531                                          bool update_total_trap_count,
2532 #if INCLUDE_JVMCI
2533                                          bool is_osr,
2534 #endif
2535                                          Method* compiled_method,
2536                                          //outputs:
2537                                          uint& ret_this_trap_count,
2538                                          bool& ret_maybe_prior_trap,
2539                                          bool& ret_maybe_prior_recompile) {
2540   trap_mdo->check_extra_data_locked();
2541 
2542   bool maybe_prior_trap = false;
2543   bool maybe_prior_recompile = false;
2544   uint this_trap_count = 0;
2545   if (update_total_trap_count) {
2546     uint idx = reason;
2547 #if INCLUDE_JVMCI
2548     if (is_osr) {
2549       // Upper half of history array used for traps in OSR compilations
2550       idx += Reason_TRAP_HISTORY_LENGTH;
2551     }
2552 #endif
2553     uint prior_trap_count = trap_mdo->trap_count(idx);
2554     this_trap_count  = trap_mdo->inc_trap_count(idx);
2555 
2556     // If the runtime cannot find a place to store trap history,
2557     // it is estimated based on the general condition of the method.
2558     // If the method has ever been recompiled, or has ever incurred
2559     // a trap with the present reason , then this BCI is assumed
2560     // (pessimistically) to be the culprit.
2561     maybe_prior_trap      = (prior_trap_count != 0);
2562     maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2563   }
2564   ProfileData* pdata = nullptr;
2565 
2566 
2567   // For reasons which are recorded per bytecode, we check per-BCI data.
2568   DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2569   assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2570   if (per_bc_reason != Reason_none) {
2571     // Find the profile data for this BCI.  If there isn't one,
2572     // try to allocate one from the MDO's set of spares.
2573     // This will let us detect a repeated trap at this point.
2574     pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : nullptr);
2575 
2576     if (pdata != nullptr) {
2577       if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2578         if (LogCompilation && xtty != nullptr) {
2579           ttyLocker ttyl;
2580           // no more room for speculative traps in this MDO
2581           xtty->elem("speculative_traps_oom");
2582         }
2583       }
2584       // Query the trap state of this profile datum.
2585       int tstate0 = pdata->trap_state();
2586       if (!trap_state_has_reason(tstate0, per_bc_reason))
2587         maybe_prior_trap = false;
2588       if (!trap_state_is_recompiled(tstate0))
2589         maybe_prior_recompile = false;
2590 
2591       // Update the trap state of this profile datum.
2592       int tstate1 = tstate0;
2593       // Record the reason.
2594       tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2595       // Store the updated state on the MDO, for next time.
2596       if (tstate1 != tstate0)
2597         pdata->set_trap_state(tstate1);
2598     } else {
2599       if (LogCompilation && xtty != nullptr) {
2600         ttyLocker ttyl;
2601         // Missing MDP?  Leave a small complaint in the log.
2602         xtty->elem("missing_mdp bci='%d'", trap_bci);
2603       }
2604     }
2605   }
2606 
2607   // Return results:
2608   ret_this_trap_count = this_trap_count;
2609   ret_maybe_prior_trap = maybe_prior_trap;
2610   ret_maybe_prior_recompile = maybe_prior_recompile;
2611   return pdata;
2612 }
2613 
2614 void
2615 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2616   ResourceMark rm;
2617   // Ignored outputs:
2618   uint ignore_this_trap_count;
2619   bool ignore_maybe_prior_trap;
2620   bool ignore_maybe_prior_recompile;
2621   assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2622   // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2623   bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2624 
2625   // Lock to read ProfileData, and ensure lock is not broken by a safepoint
2626   MutexLocker ml(trap_mdo->extra_data_lock(), Mutex::_no_safepoint_check_flag);
2627 
2628   query_update_method_data(trap_mdo, trap_bci,
2629                            (DeoptReason)reason,
2630                            update_total_counts,
2631 #if INCLUDE_JVMCI
2632                            false,
2633 #endif
2634                            nullptr,
2635                            ignore_this_trap_count,
2636                            ignore_maybe_prior_trap,
2637                            ignore_maybe_prior_recompile);
2638 }
2639 
2640 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* current, jint trap_request, jint exec_mode) {
2641   // Enable WXWrite: current function is called from methods compiled by C2 directly
2642   MACOS_AARCH64_ONLY(ThreadWXEnable wx(WXWrite, current));
2643 
2644   // Still in Java no safepoints
2645   {
2646     // This enters VM and may safepoint
2647     uncommon_trap_inner(current, trap_request);
2648   }
2649   HandleMark hm(current);
2650   return fetch_unroll_info_helper(current, exec_mode);
2651 }
2652 
2653 // Local derived constants.
2654 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2655 const int DS_REASON_MASK   = ((uint)DataLayout::trap_mask) >> 1;
2656 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2657 
2658 //---------------------------trap_state_reason---------------------------------
2659 Deoptimization::DeoptReason
2660 Deoptimization::trap_state_reason(int trap_state) {
2661   // This assert provides the link between the width of DataLayout::trap_bits
2662   // and the encoding of "recorded" reasons.  It ensures there are enough
2663   // bits to store all needed reasons in the per-BCI MDO profile.
2664   assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2665   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2666   trap_state -= recompile_bit;
2667   if (trap_state == DS_REASON_MASK) {
2668     return Reason_many;
2669   } else {
2670     assert((int)Reason_none == 0, "state=0 => Reason_none");
2671     return (DeoptReason)trap_state;
2672   }
2673 }
2674 //-------------------------trap_state_has_reason-------------------------------
2675 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2676   assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2677   assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2678   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2679   trap_state -= recompile_bit;
2680   if (trap_state == DS_REASON_MASK) {
2681     return -1;  // true, unspecifically (bottom of state lattice)
2682   } else if (trap_state == reason) {
2683     return 1;   // true, definitely
2684   } else if (trap_state == 0) {
2685     return 0;   // false, definitely (top of state lattice)
2686   } else {
2687     return 0;   // false, definitely
2688   }
2689 }
2690 //-------------------------trap_state_add_reason-------------------------------
2691 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2692   assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2693   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2694   trap_state -= recompile_bit;
2695   if (trap_state == DS_REASON_MASK) {
2696     return trap_state + recompile_bit;     // already at state lattice bottom
2697   } else if (trap_state == reason) {
2698     return trap_state + recompile_bit;     // the condition is already true
2699   } else if (trap_state == 0) {
2700     return reason + recompile_bit;          // no condition has yet been true
2701   } else {
2702     return DS_REASON_MASK + recompile_bit;  // fall to state lattice bottom
2703   }
2704 }
2705 //-----------------------trap_state_is_recompiled------------------------------
2706 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2707   return (trap_state & DS_RECOMPILE_BIT) != 0;
2708 }
2709 //-----------------------trap_state_set_recompiled-----------------------------
2710 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2711   if (z)  return trap_state |  DS_RECOMPILE_BIT;
2712   else    return trap_state & ~DS_RECOMPILE_BIT;
2713 }
2714 //---------------------------format_trap_state---------------------------------
2715 // This is used for debugging and diagnostics, including LogFile output.
2716 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2717                                               int trap_state) {
2718   assert(buflen > 0, "sanity");
2719   DeoptReason reason      = trap_state_reason(trap_state);
2720   bool        recomp_flag = trap_state_is_recompiled(trap_state);
2721   // Re-encode the state from its decoded components.
2722   int decoded_state = 0;
2723   if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2724     decoded_state = trap_state_add_reason(decoded_state, reason);
2725   if (recomp_flag)
2726     decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2727   // If the state re-encodes properly, format it symbolically.
2728   // Because this routine is used for debugging and diagnostics,
2729   // be robust even if the state is a strange value.
2730   size_t len;
2731   if (decoded_state != trap_state) {
2732     // Random buggy state that doesn't decode??
2733     len = jio_snprintf(buf, buflen, "#%d", trap_state);
2734   } else {
2735     len = jio_snprintf(buf, buflen, "%s%s",
2736                        trap_reason_name(reason),
2737                        recomp_flag ? " recompiled" : "");
2738   }
2739   return buf;
2740 }
2741 
2742 
2743 //--------------------------------statics--------------------------------------
2744 const char* Deoptimization::_trap_reason_name[] = {
2745   // Note:  Keep this in sync. with enum DeoptReason.
2746   "none",
2747   "null_check",
2748   "null_assert" JVMCI_ONLY("_or_unreached0"),
2749   "range_check",
2750   "class_check",
2751   "array_check",
2752   "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2753   "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2754   "profile_predicate",
2755   "auto_vectorization_check",
2756   "unloaded",
2757   "uninitialized",
2758   "initialized",
2759   "unreached",
2760   "unhandled",
2761   "constraint",
2762   "div0_check",
2763   "age",
2764   "predicate",
2765   "loop_limit_check",
2766   "speculate_class_check",
2767   "speculate_null_check",
2768   "speculate_null_assert",
2769   "unstable_if",
2770   "unstable_fused_if",
2771   "receiver_constraint",
2772   "short_running_loop" JVMCI_ONLY("_or_aliasing"),
2773 #if INCLUDE_JVMCI
2774   "transfer_to_interpreter",
2775   "not_compiled_exception_handler",
2776   "unresolved",
2777   "jsr_mismatch",
2778 #endif
2779   "tenured"
2780 };
2781 const char* Deoptimization::_trap_action_name[] = {
2782   // Note:  Keep this in sync. with enum DeoptAction.
2783   "none",
2784   "maybe_recompile",
2785   "reinterpret",
2786   "make_not_entrant",
2787   "make_not_compilable"
2788 };
2789 
2790 const char* Deoptimization::trap_reason_name(int reason) {
2791   // Check that every reason has a name
2792   STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2793 
2794   if (reason == Reason_many)  return "many";
2795   if ((uint)reason < Reason_LIMIT)
2796     return _trap_reason_name[reason];
2797   static char buf[20];
2798   os::snprintf_checked(buf, sizeof(buf), "reason%d", reason);
2799   return buf;
2800 }
2801 const char* Deoptimization::trap_action_name(int action) {
2802   // Check that every action has a name
2803   STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2804 
2805   if ((uint)action < Action_LIMIT)
2806     return _trap_action_name[action];
2807   static char buf[20];
2808   os::snprintf_checked(buf, sizeof(buf), "action%d", action);
2809   return buf;
2810 }
2811 
2812 // This is used for debugging and diagnostics, including LogFile output.
2813 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2814                                                 int trap_request) {
2815   jint unloaded_class_index = trap_request_index(trap_request);
2816   const char* reason = trap_reason_name(trap_request_reason(trap_request));
2817   const char* action = trap_action_name(trap_request_action(trap_request));
2818 #if INCLUDE_JVMCI
2819   int debug_id = trap_request_debug_id(trap_request);
2820 #endif
2821   size_t len;
2822   if (unloaded_class_index < 0) {
2823     len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2824                        reason, action
2825 #if INCLUDE_JVMCI
2826                        ,debug_id
2827 #endif
2828                        );
2829   } else {
2830     len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2831                        reason, action, unloaded_class_index
2832 #if INCLUDE_JVMCI
2833                        ,debug_id
2834 #endif
2835                        );
2836   }
2837   return buf;
2838 }
2839 
2840 juint Deoptimization::_deoptimization_hist
2841         [Deoptimization::Reason_LIMIT]
2842     [1 + Deoptimization::Action_LIMIT]
2843         [Deoptimization::BC_CASE_LIMIT]
2844   = {0};
2845 
2846 enum {
2847   LSB_BITS = 8,
2848   LSB_MASK = right_n_bits(LSB_BITS)
2849 };
2850 
2851 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2852                                        Bytecodes::Code bc) {
2853   assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2854   assert(action >= 0 && action < Action_LIMIT, "oob");
2855   _deoptimization_hist[Reason_none][0][0] += 1;  // total
2856   _deoptimization_hist[reason][0][0]      += 1;  // per-reason total
2857   juint* cases = _deoptimization_hist[reason][1+action];
2858   juint* bc_counter_addr = nullptr;
2859   juint  bc_counter      = 0;
2860   // Look for an unused counter, or an exact match to this BC.
2861   if (bc != Bytecodes::_illegal) {
2862     for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2863       juint* counter_addr = &cases[bc_case];
2864       juint  counter = *counter_addr;
2865       if ((counter == 0 && bc_counter_addr == nullptr)
2866           || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2867         // this counter is either free or is already devoted to this BC
2868         bc_counter_addr = counter_addr;
2869         bc_counter = counter | bc;
2870       }
2871     }
2872   }
2873   if (bc_counter_addr == nullptr) {
2874     // Overflow, or no given bytecode.
2875     bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2876     bc_counter = (*bc_counter_addr & ~LSB_MASK);  // clear LSB
2877   }
2878   *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2879 }
2880 
2881 jint Deoptimization::total_deoptimization_count() {
2882   return _deoptimization_hist[Reason_none][0][0];
2883 }
2884 
2885 // Get the deopt count for a specific reason and a specific action. If either
2886 // one of 'reason' or 'action' is null, the method returns the sum of all
2887 // deoptimizations with the specific 'action' or 'reason' respectively.
2888 // If both arguments are null, the method returns the total deopt count.
2889 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2890   if (reason_str == nullptr && action_str == nullptr) {
2891     return total_deoptimization_count();
2892   }
2893   juint counter = 0;
2894   for (int reason = 0; reason < Reason_LIMIT; reason++) {
2895     if (reason_str == nullptr || !strcmp(reason_str, trap_reason_name(reason))) {
2896       for (int action = 0; action < Action_LIMIT; action++) {
2897         if (action_str == nullptr || !strcmp(action_str, trap_action_name(action))) {
2898           juint* cases = _deoptimization_hist[reason][1+action];
2899           for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2900             counter += cases[bc_case] >> LSB_BITS;
2901           }
2902         }
2903       }
2904     }
2905   }
2906   return counter;
2907 }
2908 
2909 void Deoptimization::print_statistics() {
2910   juint total = total_deoptimization_count();
2911   juint account = total;
2912   if (total != 0) {
2913     ttyLocker ttyl;
2914     if (xtty != nullptr)  xtty->head("statistics type='deoptimization'");
2915     tty->print_cr("Deoptimization traps recorded:");
2916     #define PRINT_STAT_LINE(name, r) \
2917       tty->print_cr("  %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2918     PRINT_STAT_LINE("total", total);
2919     // For each non-zero entry in the histogram, print the reason,
2920     // the action, and (if specifically known) the type of bytecode.
2921     for (int reason = 0; reason < Reason_LIMIT; reason++) {
2922       for (int action = 0; action < Action_LIMIT; action++) {
2923         juint* cases = _deoptimization_hist[reason][1+action];
2924         for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2925           juint counter = cases[bc_case];
2926           if (counter != 0) {
2927             char name[1*K];
2928             Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2929             if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2930               bc = Bytecodes::_illegal;
2931             os::snprintf_checked(name, sizeof(name), "%s/%s/%s",
2932                     trap_reason_name(reason),
2933                     trap_action_name(action),
2934                     Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2935             juint r = counter >> LSB_BITS;
2936             tty->print_cr("  %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2937             account -= r;
2938           }
2939         }
2940       }
2941     }
2942     if (account != 0) {
2943       PRINT_STAT_LINE("unaccounted", account);
2944     }
2945     #undef PRINT_STAT_LINE
2946     if (xtty != nullptr)  xtty->tail("statistics");
2947   }
2948 }
2949 
2950 #else // COMPILER2_OR_JVMCI
2951 
2952 
2953 // Stubs for C1 only system.
2954 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2955   return false;
2956 }
2957 
2958 const char* Deoptimization::trap_reason_name(int reason) {
2959   return "unknown";
2960 }
2961 
2962 jint Deoptimization::total_deoptimization_count() {
2963   return 0;
2964 }
2965 
2966 jint Deoptimization::deoptimization_count(const char *reason_str, const char *action_str) {
2967   return 0;
2968 }
2969 
2970 void Deoptimization::print_statistics() {
2971   // no output
2972 }
2973 
2974 void
2975 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2976   // no update
2977 }
2978 
2979 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2980   return 0;
2981 }
2982 
2983 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2984                                        Bytecodes::Code bc) {
2985   // no update
2986 }
2987 
2988 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2989                                               int trap_state) {
2990   jio_snprintf(buf, buflen, "#%d", trap_state);
2991   return buf;
2992 }
2993 
2994 #endif // COMPILER2_OR_JVMCI