1 /*
   2  * Copyright (c) 1997, 2016, 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 "precompiled.hpp"
  26 #include "classfile/systemDictionary.hpp"
  27 #include "code/debugInfoRec.hpp"
  28 #include "code/nmethod.hpp"
  29 #include "code/pcDesc.hpp"
  30 #include "code/scopeDesc.hpp"
  31 #include "interpreter/bytecode.hpp"
  32 #include "interpreter/interpreter.hpp"
  33 #include "interpreter/oopMapCache.hpp"
  34 #include "memory/allocation.inline.hpp"
  35 #include "memory/oopFactory.hpp"
  36 #include "memory/resourceArea.hpp"
  37 #include "oops/method.hpp"
  38 #include "oops/oop.inline.hpp"
  39 #include "prims/jvmtiThreadState.hpp"
  40 #include "runtime/biasedLocking.hpp"
  41 #include "runtime/compilationPolicy.hpp"
  42 #include "runtime/deoptimization.hpp"
  43 #include "runtime/interfaceSupport.hpp"
  44 #include "runtime/sharedRuntime.hpp"
  45 #include "runtime/signature.hpp"
  46 #include "runtime/stubRoutines.hpp"
  47 #include "runtime/thread.hpp"
  48 #include "runtime/vframe.hpp"
  49 #include "runtime/vframeArray.hpp"
  50 #include "runtime/vframe_hp.hpp"
  51 #include "utilities/events.hpp"
  52 #include "utilities/xmlstream.hpp"
  53 #ifdef TARGET_ARCH_x86
  54 # include "vmreg_x86.inline.hpp"
  55 #endif
  56 #ifdef TARGET_ARCH_sparc
  57 # include "vmreg_sparc.inline.hpp"
  58 #endif
  59 #ifdef TARGET_ARCH_zero
  60 # include "vmreg_zero.inline.hpp"
  61 #endif
  62 #ifdef TARGET_ARCH_arm
  63 # include "vmreg_arm.inline.hpp"
  64 #endif
  65 #ifdef TARGET_ARCH_ppc
  66 # include "vmreg_ppc.inline.hpp"
  67 #endif
  68 #ifdef COMPILER2
  69 #if defined AD_MD_HPP
  70 # include AD_MD_HPP
  71 #elif defined TARGET_ARCH_MODEL_x86_32
  72 # include "adfiles/ad_x86_32.hpp"
  73 #elif defined TARGET_ARCH_MODEL_x86_64
  74 # include "adfiles/ad_x86_64.hpp"
  75 #elif defined TARGET_ARCH_MODEL_sparc
  76 # include "adfiles/ad_sparc.hpp"
  77 #elif defined TARGET_ARCH_MODEL_zero
  78 # include "adfiles/ad_zero.hpp"
  79 #elif defined TARGET_ARCH_MODEL_ppc_64
  80 # include "adfiles/ad_ppc_64.hpp"
  81 #endif
  82 #endif // COMPILER2
  83 
  84 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
  85 
  86 bool DeoptimizationMarker::_is_active = false;
  87 
  88 Deoptimization::UnrollBlock::UnrollBlock(int  size_of_deoptimized_frame,
  89                                          int  caller_adjustment,
  90                                          int  caller_actual_parameters,
  91                                          int  number_of_frames,
  92                                          intptr_t* frame_sizes,
  93                                          address* frame_pcs,
  94                                          BasicType return_type) {
  95   _size_of_deoptimized_frame = size_of_deoptimized_frame;
  96   _caller_adjustment         = caller_adjustment;
  97   _caller_actual_parameters  = caller_actual_parameters;
  98   _number_of_frames          = number_of_frames;
  99   _frame_sizes               = frame_sizes;
 100   _frame_pcs                 = frame_pcs;
 101   _register_block            = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
 102   _return_type               = return_type;
 103   _initial_info              = 0;
 104   // PD (x86 only)
 105   _counter_temp              = 0;
 106   _unpack_kind               = 0;
 107   _sender_sp_temp            = 0;
 108 
 109   _total_frame_sizes         = size_of_frames();
 110 }
 111 
 112 
 113 Deoptimization::UnrollBlock::~UnrollBlock() {
 114   FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler);
 115   FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler);
 116   FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler);
 117 }
 118 
 119 
 120 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
 121   assert(register_number < RegisterMap::reg_count, "checking register number");
 122   return &_register_block[register_number * 2];
 123 }
 124 
 125 
 126 
 127 int Deoptimization::UnrollBlock::size_of_frames() const {
 128   // Acount first for the adjustment of the initial frame
 129   int result = _caller_adjustment;
 130   for (int index = 0; index < number_of_frames(); index++) {
 131     result += frame_sizes()[index];
 132   }
 133   return result;
 134 }
 135 
 136 
 137 void Deoptimization::UnrollBlock::print() {
 138   ttyLocker ttyl;
 139   tty->print_cr("UnrollBlock");
 140   tty->print_cr("  size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
 141   tty->print(   "  frame_sizes: ");
 142   for (int index = 0; index < number_of_frames(); index++) {
 143     tty->print("%d ", frame_sizes()[index]);
 144   }
 145   tty->cr();
 146 }
 147 
 148 
 149 // In order to make fetch_unroll_info work properly with escape
 150 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
 151 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
 152 // of previously eliminated objects occurs in realloc_objects, which is
 153 // called from the method fetch_unroll_info_helper below.
 154 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread))
 155   // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
 156   // but makes the entry a little slower. There is however a little dance we have to
 157   // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
 158 
 159   // fetch_unroll_info() is called at the beginning of the deoptimization
 160   // handler. Note this fact before we start generating temporary frames
 161   // that can confuse an asynchronous stack walker. This counter is
 162   // decremented at the end of unpack_frames().
 163   thread->inc_in_deopt_handler();
 164 
 165   return fetch_unroll_info_helper(thread);
 166 JRT_END
 167 
 168 
 169 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
 170 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) {
 171 
 172   // Note: there is a safepoint safety issue here. No matter whether we enter
 173   // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
 174   // the vframeArray is created.
 175   //
 176 
 177   // Allocate our special deoptimization ResourceMark
 178   DeoptResourceMark* dmark = new DeoptResourceMark(thread);
 179   assert(thread->deopt_mark() == NULL, "Pending deopt!");
 180   thread->set_deopt_mark(dmark);
 181 
 182   frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
 183   RegisterMap map(thread, true);
 184   RegisterMap dummy_map(thread, false);
 185   // Now get the deoptee with a valid map
 186   frame deoptee = stub_frame.sender(&map);
 187   // Set the deoptee nmethod
 188   assert(thread->deopt_nmethod() == NULL, "Pending deopt!");
 189   thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null());
 190 
 191   if (VerifyStack) {
 192     thread->validate_frame_layout();
 193   }
 194 
 195   // Create a growable array of VFrames where each VFrame represents an inlined
 196   // Java frame.  This storage is allocated with the usual system arena.
 197   assert(deoptee.is_compiled_frame(), "Wrong frame type");
 198   GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
 199   vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
 200   while (!vf->is_top()) {
 201     assert(vf->is_compiled_frame(), "Wrong frame type");
 202     chunk->push(compiledVFrame::cast(vf));
 203     vf = vf->sender();
 204   }
 205   assert(vf->is_compiled_frame(), "Wrong frame type");
 206   chunk->push(compiledVFrame::cast(vf));
 207 
 208   bool realloc_failures = false;
 209 
 210 #ifdef COMPILER2
 211   // Reallocate the non-escaping objects and restore their fields. Then
 212   // relock objects if synchronization on them was eliminated.
 213   if (DoEscapeAnalysis || EliminateNestedLocks) {
 214     if (EliminateAllocations) {
 215       assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
 216       GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
 217 
 218       // The flag return_oop() indicates call sites which return oop
 219       // in compiled code. Such sites include java method calls,
 220       // runtime calls (for example, used to allocate new objects/arrays
 221       // on slow code path) and any other calls generated in compiled code.
 222       // It is not guaranteed that we can get such information here only
 223       // by analyzing bytecode in deoptimized frames. This is why this flag
 224       // is set during method compilation (see Compile::Process_OopMap_Node()).
 225       // If the previous frame was popped, we don't have a result.
 226       bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution();
 227       Handle return_value;
 228       if (save_oop_result) {
 229         // Reallocation may trigger GC. If deoptimization happened on return from
 230         // call which returns oop we need to save it since it is not in oopmap.
 231         oop result = deoptee.saved_oop_result(&map);
 232         assert(result == NULL || result->is_oop(), "must be oop");
 233         return_value = Handle(thread, result);
 234         assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
 235         if (TraceDeoptimization) {
 236           ttyLocker ttyl;
 237           tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread);
 238         }
 239       }
 240       if (objects != NULL) {
 241         JRT_BLOCK
 242           realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD);
 243         JRT_END
 244         reassign_fields(&deoptee, &map, objects, realloc_failures);
 245 #ifndef PRODUCT
 246         if (TraceDeoptimization) {
 247           ttyLocker ttyl;
 248           tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread);
 249           print_objects(objects, realloc_failures);
 250         }
 251 #endif
 252       }
 253       if (save_oop_result) {
 254         // Restore result.
 255         deoptee.set_saved_oop_result(&map, return_value());
 256       }
 257     }
 258     if (EliminateLocks) {
 259 #ifndef PRODUCT
 260       bool first = true;
 261 #endif
 262       for (int i = 0; i < chunk->length(); i++) {
 263         compiledVFrame* cvf = chunk->at(i);
 264         assert (cvf->scope() != NULL,"expect only compiled java frames");
 265         GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
 266         if (monitors->is_nonempty()) {
 267           relock_objects(monitors, thread, realloc_failures);
 268 #ifndef PRODUCT
 269           if (TraceDeoptimization) {
 270             ttyLocker ttyl;
 271             for (int j = 0; j < monitors->length(); j++) {
 272               MonitorInfo* mi = monitors->at(j);
 273               if (mi->eliminated()) {
 274                 if (first) {
 275                   first = false;
 276                   tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread);
 277                 }
 278                 if (mi->owner_is_scalar_replaced()) {
 279                   Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
 280                   tty->print_cr("     failed reallocation for klass %s", k->external_name());
 281                 } else {
 282                   tty->print_cr("     object <" INTPTR_FORMAT "> locked", (void *)mi->owner());
 283                 }
 284               }
 285             }
 286           }
 287 #endif
 288         }
 289       }
 290     }
 291   }
 292 #endif // COMPILER2
 293   // Ensure that no safepoint is taken after pointers have been stored
 294   // in fields of rematerialized objects.  If a safepoint occurs from here on
 295   // out the java state residing in the vframeArray will be missed.
 296   No_Safepoint_Verifier no_safepoint;
 297 
 298   vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
 299 #ifdef COMPILER2
 300   if (realloc_failures) {
 301     pop_frames_failed_reallocs(thread, array);
 302   }
 303 #endif
 304 
 305   assert(thread->vframe_array_head() == NULL, "Pending deopt!");
 306   thread->set_vframe_array_head(array);
 307 
 308   // Now that the vframeArray has been created if we have any deferred local writes
 309   // added by jvmti then we can free up that structure as the data is now in the
 310   // vframeArray
 311 
 312   if (thread->deferred_locals() != NULL) {
 313     GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
 314     int i = 0;
 315     do {
 316       // Because of inlining we could have multiple vframes for a single frame
 317       // and several of the vframes could have deferred writes. Find them all.
 318       if (list->at(i)->id() == array->original().id()) {
 319         jvmtiDeferredLocalVariableSet* dlv = list->at(i);
 320         list->remove_at(i);
 321         // individual jvmtiDeferredLocalVariableSet are CHeapObj's
 322         delete dlv;
 323       } else {
 324         i++;
 325       }
 326     } while ( i < list->length() );
 327     if (list->length() == 0) {
 328       thread->set_deferred_locals(NULL);
 329       // free the list and elements back to C heap.
 330       delete list;
 331     }
 332 
 333   }
 334 
 335 #ifndef SHARK
 336   // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
 337   CodeBlob* cb = stub_frame.cb();
 338   // Verify we have the right vframeArray
 339   assert(cb->frame_size() >= 0, "Unexpected frame size");
 340   intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
 341 
 342   // If the deopt call site is a MethodHandle invoke call site we have
 343   // to adjust the unpack_sp.
 344   nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
 345   if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
 346     unpack_sp = deoptee.unextended_sp();
 347 
 348 #ifdef ASSERT
 349   assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking");
 350 #endif
 351 #else
 352   intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp();
 353 #endif // !SHARK
 354 
 355   // This is a guarantee instead of an assert because if vframe doesn't match
 356   // we will unpack the wrong deoptimized frame and wind up in strange places
 357   // where it will be very difficult to figure out what went wrong. Better
 358   // to die an early death here than some very obscure death later when the
 359   // trail is cold.
 360   // Note: on ia64 this guarantee can be fooled by frames with no memory stack
 361   // in that it will fail to detect a problem when there is one. This needs
 362   // more work in tiger timeframe.
 363   guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
 364 
 365   int number_of_frames = array->frames();
 366 
 367   // Compute the vframes' sizes.  Note that frame_sizes[] entries are ordered from outermost to innermost
 368   // virtual activation, which is the reverse of the elements in the vframes array.
 369   intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
 370   // +1 because we always have an interpreter return address for the final slot.
 371   address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
 372   int popframe_extra_args = 0;
 373   // Create an interpreter return address for the stub to use as its return
 374   // address so the skeletal frames are perfectly walkable
 375   frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
 376 
 377   // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
 378   // activation be put back on the expression stack of the caller for reexecution
 379   if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
 380     popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
 381   }
 382 
 383   // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
 384   // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
 385   // than simply use array->sender.pc(). This requires us to walk the current set of frames
 386   //
 387   frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
 388   deopt_sender = deopt_sender.sender(&dummy_map);     // Now deoptee caller
 389 
 390   // It's possible that the number of paramters at the call site is
 391   // different than number of arguments in the callee when method
 392   // handles are used.  If the caller is interpreted get the real
 393   // value so that the proper amount of space can be added to it's
 394   // frame.
 395   bool caller_was_method_handle = false;
 396   if (deopt_sender.is_interpreted_frame()) {
 397     methodHandle method = deopt_sender.interpreter_frame_method();
 398     Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
 399     if (cur.is_invokedynamic() || cur.is_invokehandle()) {
 400       // Method handle invokes may involve fairly arbitrary chains of
 401       // calls so it's impossible to know how much actual space the
 402       // caller has for locals.
 403       caller_was_method_handle = true;
 404     }
 405   }
 406 
 407   //
 408   // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
 409   // frame_sizes/frame_pcs[1] next oldest frame (int)
 410   // frame_sizes/frame_pcs[n] youngest frame (int)
 411   //
 412   // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
 413   // owns the space for the return address to it's caller).  Confusing ain't it.
 414   //
 415   // The vframe array can address vframes with indices running from
 416   // 0.._frames-1. Index  0 is the youngest frame and _frame - 1 is the oldest (root) frame.
 417   // When we create the skeletal frames we need the oldest frame to be in the zero slot
 418   // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
 419   // so things look a little strange in this loop.
 420   //
 421   int callee_parameters = 0;
 422   int callee_locals = 0;
 423   for (int index = 0; index < array->frames(); index++ ) {
 424     // frame[number_of_frames - 1 ] = on_stack_size(youngest)
 425     // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
 426     // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
 427     frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
 428                                                                                                     callee_locals,
 429                                                                                                     index == 0,
 430                                                                                                     popframe_extra_args);
 431     // This pc doesn't have to be perfect just good enough to identify the frame
 432     // as interpreted so the skeleton frame will be walkable
 433     // The correct pc will be set when the skeleton frame is completely filled out
 434     // The final pc we store in the loop is wrong and will be overwritten below
 435     frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
 436 
 437     callee_parameters = array->element(index)->method()->size_of_parameters();
 438     callee_locals = array->element(index)->method()->max_locals();
 439     popframe_extra_args = 0;
 440   }
 441 
 442   // Compute whether the root vframe returns a float or double value.
 443   BasicType return_type;
 444   {
 445     HandleMark hm;
 446     methodHandle method(thread, array->element(0)->method());
 447     Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
 448     return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
 449   }
 450 
 451   // Compute information for handling adapters and adjusting the frame size of the caller.
 452   int caller_adjustment = 0;
 453 
 454   // Compute the amount the oldest interpreter frame will have to adjust
 455   // its caller's stack by. If the caller is a compiled frame then
 456   // we pretend that the callee has no parameters so that the
 457   // extension counts for the full amount of locals and not just
 458   // locals-parms. This is because without a c2i adapter the parm
 459   // area as created by the compiled frame will not be usable by
 460   // the interpreter. (Depending on the calling convention there
 461   // may not even be enough space).
 462 
 463   // QQQ I'd rather see this pushed down into last_frame_adjust
 464   // and have it take the sender (aka caller).
 465 
 466   if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
 467     caller_adjustment = last_frame_adjust(0, callee_locals);
 468   } else if (callee_locals > callee_parameters) {
 469     // The caller frame may need extending to accommodate
 470     // non-parameter locals of the first unpacked interpreted frame.
 471     // Compute that adjustment.
 472     caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
 473   }
 474 
 475   // If the sender is deoptimized the we must retrieve the address of the handler
 476   // since the frame will "magically" show the original pc before the deopt
 477   // and we'd undo the deopt.
 478 
 479   frame_pcs[0] = deopt_sender.raw_pc();
 480 
 481 #ifndef SHARK
 482   assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
 483 #endif // SHARK
 484 
 485   UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
 486                                       caller_adjustment * BytesPerWord,
 487                                       caller_was_method_handle ? 0 : callee_parameters,
 488                                       number_of_frames,
 489                                       frame_sizes,
 490                                       frame_pcs,
 491                                       return_type);
 492   // On some platforms, we need a way to pass some platform dependent
 493   // information to the unpacking code so the skeletal frames come out
 494   // correct (initial fp value, unextended sp, ...)
 495   info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
 496 
 497   if (array->frames() > 1) {
 498     if (VerifyStack && TraceDeoptimization) {
 499       ttyLocker ttyl;
 500       tty->print_cr("Deoptimizing method containing inlining");
 501     }
 502   }
 503 
 504   array->set_unroll_block(info);
 505   return info;
 506 }
 507 
 508 // Called to cleanup deoptimization data structures in normal case
 509 // after unpacking to stack and when stack overflow error occurs
 510 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
 511                                         vframeArray *array) {
 512 
 513   // Get array if coming from exception
 514   if (array == NULL) {
 515     array = thread->vframe_array_head();
 516   }
 517   thread->set_vframe_array_head(NULL);
 518 
 519   // Free the previous UnrollBlock
 520   vframeArray* old_array = thread->vframe_array_last();
 521   thread->set_vframe_array_last(array);
 522 
 523   if (old_array != NULL) {
 524     UnrollBlock* old_info = old_array->unroll_block();
 525     old_array->set_unroll_block(NULL);
 526     delete old_info;
 527     delete old_array;
 528   }
 529 
 530   // Deallocate any resource creating in this routine and any ResourceObjs allocated
 531   // inside the vframeArray (StackValueCollections)
 532 
 533   delete thread->deopt_mark();
 534   thread->set_deopt_mark(NULL);
 535   thread->set_deopt_nmethod(NULL);
 536 
 537 
 538   if (JvmtiExport::can_pop_frame()) {
 539 #ifndef CC_INTERP
 540     // Regardless of whether we entered this routine with the pending
 541     // popframe condition bit set, we should always clear it now
 542     thread->clear_popframe_condition();
 543 #else
 544     // C++ interpeter will clear has_pending_popframe when it enters
 545     // with method_resume. For deopt_resume2 we clear it now.
 546     if (thread->popframe_forcing_deopt_reexecution())
 547         thread->clear_popframe_condition();
 548 #endif /* CC_INTERP */
 549   }
 550 
 551   // unpack_frames() is called at the end of the deoptimization handler
 552   // and (in C2) at the end of the uncommon trap handler. Note this fact
 553   // so that an asynchronous stack walker can work again. This counter is
 554   // incremented at the beginning of fetch_unroll_info() and (in C2) at
 555   // the beginning of uncommon_trap().
 556   thread->dec_in_deopt_handler();
 557 }
 558 
 559 
 560 // Return BasicType of value being returned
 561 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
 562 
 563   // We are already active int he special DeoptResourceMark any ResourceObj's we
 564   // allocate will be freed at the end of the routine.
 565 
 566   // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
 567   // but makes the entry a little slower. There is however a little dance we have to
 568   // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
 569   ResetNoHandleMark rnhm; // No-op in release/product versions
 570   HandleMark hm;
 571 
 572   frame stub_frame = thread->last_frame();
 573 
 574   // Since the frame to unpack is the top frame of this thread, the vframe_array_head
 575   // must point to the vframeArray for the unpack frame.
 576   vframeArray* array = thread->vframe_array_head();
 577 
 578 #ifndef PRODUCT
 579   if (TraceDeoptimization) {
 580     ttyLocker ttyl;
 581     tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode);
 582   }
 583 #endif
 584   Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
 585               stub_frame.pc(), stub_frame.sp(), exec_mode);
 586 
 587   UnrollBlock* info = array->unroll_block();
 588 
 589   // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
 590   array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
 591 
 592   BasicType bt = info->return_type();
 593 
 594   // If we have an exception pending, claim that the return type is an oop
 595   // so the deopt_blob does not overwrite the exception_oop.
 596 
 597   if (exec_mode == Unpack_exception)
 598     bt = T_OBJECT;
 599 
 600   // Cleanup thread deopt data
 601   cleanup_deopt_info(thread, array);
 602 
 603 #ifndef PRODUCT
 604   if (VerifyStack) {
 605     ResourceMark res_mark;
 606 
 607     thread->validate_frame_layout();
 608 
 609     // Verify that the just-unpacked frames match the interpreter's
 610     // notions of expression stack and locals
 611     vframeArray* cur_array = thread->vframe_array_last();
 612     RegisterMap rm(thread, false);
 613     rm.set_include_argument_oops(false);
 614     bool is_top_frame = true;
 615     int callee_size_of_parameters = 0;
 616     int callee_max_locals = 0;
 617     for (int i = 0; i < cur_array->frames(); i++) {
 618       vframeArrayElement* el = cur_array->element(i);
 619       frame* iframe = el->iframe();
 620       guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
 621 
 622       // Get the oop map for this bci
 623       InterpreterOopMap mask;
 624       int cur_invoke_parameter_size = 0;
 625       bool try_next_mask = false;
 626       int next_mask_expression_stack_size = -1;
 627       int top_frame_expression_stack_adjustment = 0;
 628       methodHandle mh(thread, iframe->interpreter_frame_method());
 629       OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
 630       BytecodeStream str(mh);
 631       str.set_start(iframe->interpreter_frame_bci());
 632       int max_bci = mh->code_size();
 633       // Get to the next bytecode if possible
 634       assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
 635       // Check to see if we can grab the number of outgoing arguments
 636       // at an uncommon trap for an invoke (where the compiler
 637       // generates debug info before the invoke has executed)
 638       Bytecodes::Code cur_code = str.next();
 639       if (cur_code == Bytecodes::_invokevirtual   ||
 640           cur_code == Bytecodes::_invokespecial   ||
 641           cur_code == Bytecodes::_invokestatic    ||
 642           cur_code == Bytecodes::_invokeinterface ||
 643           cur_code == Bytecodes::_invokedynamic) {
 644         Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
 645         Symbol* signature = invoke.signature();
 646         ArgumentSizeComputer asc(signature);
 647         cur_invoke_parameter_size = asc.size();
 648         if (invoke.has_receiver()) {
 649           // Add in receiver
 650           ++cur_invoke_parameter_size;
 651         }
 652         if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
 653           callee_size_of_parameters++;
 654         }
 655       }
 656       if (str.bci() < max_bci) {
 657         Bytecodes::Code bc = str.next();
 658         if (bc >= 0) {
 659           // The interpreter oop map generator reports results before
 660           // the current bytecode has executed except in the case of
 661           // calls. It seems to be hard to tell whether the compiler
 662           // has emitted debug information matching the "state before"
 663           // a given bytecode or the state after, so we try both
 664           switch (cur_code) {
 665             case Bytecodes::_invokevirtual:
 666             case Bytecodes::_invokespecial:
 667             case Bytecodes::_invokestatic:
 668             case Bytecodes::_invokeinterface:
 669             case Bytecodes::_invokedynamic:
 670             case Bytecodes::_athrow:
 671               break;
 672             default: {
 673               InterpreterOopMap next_mask;
 674               OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
 675               next_mask_expression_stack_size = next_mask.expression_stack_size();
 676               // Need to subtract off the size of the result type of
 677               // the bytecode because this is not described in the
 678               // debug info but returned to the interpreter in the TOS
 679               // caching register
 680               BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
 681               if (bytecode_result_type != T_ILLEGAL) {
 682                 top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
 683               }
 684               assert(top_frame_expression_stack_adjustment >= 0, "");
 685               try_next_mask = true;
 686               break;
 687             }
 688           }
 689         }
 690       }
 691 
 692       // Verify stack depth and oops in frame
 693       // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
 694       if (!(
 695             /* SPARC */
 696             (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
 697             /* x86 */
 698             (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
 699             (try_next_mask &&
 700              (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
 701                                                                     top_frame_expression_stack_adjustment))) ||
 702             (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
 703             (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
 704              (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
 705             )) {
 706         ttyLocker ttyl;
 707 
 708         // Print out some information that will help us debug the problem
 709         tty->print_cr("Wrong number of expression stack elements during deoptimization");
 710         tty->print_cr("  Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
 711         tty->print_cr("  Fabricated interpreter frame had %d expression stack elements",
 712                       iframe->interpreter_frame_expression_stack_size());
 713         tty->print_cr("  Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
 714         tty->print_cr("  try_next_mask = %d", try_next_mask);
 715         tty->print_cr("  next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
 716         tty->print_cr("  callee_size_of_parameters = %d", callee_size_of_parameters);
 717         tty->print_cr("  callee_max_locals = %d", callee_max_locals);
 718         tty->print_cr("  top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
 719         tty->print_cr("  exec_mode = %d", exec_mode);
 720         tty->print_cr("  cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
 721         tty->print_cr("  Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id());
 722         tty->print_cr("  Interpreted frames:");
 723         for (int k = 0; k < cur_array->frames(); k++) {
 724           vframeArrayElement* el = cur_array->element(k);
 725           tty->print_cr("    %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
 726         }
 727         cur_array->print_on_2(tty);
 728         guarantee(false, "wrong number of expression stack elements during deopt");
 729       }
 730       VerifyOopClosure verify;
 731       iframe->oops_interpreted_do(&verify, NULL, &rm, false);
 732       callee_size_of_parameters = mh->size_of_parameters();
 733       callee_max_locals = mh->max_locals();
 734       is_top_frame = false;
 735     }
 736   }
 737 #endif /* !PRODUCT */
 738 
 739 
 740   return bt;
 741 JRT_END
 742 
 743 
 744 int Deoptimization::deoptimize_dependents() {
 745   Threads::deoptimized_wrt_marked_nmethods();
 746   return 0;
 747 }
 748 
 749 
 750 #ifdef COMPILER2
 751 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) {
 752   Handle pending_exception(thread->pending_exception());
 753   const char* exception_file = thread->exception_file();
 754   int exception_line = thread->exception_line();
 755   thread->clear_pending_exception();
 756 
 757   bool failures = false;
 758 
 759   for (int i = 0; i < objects->length(); i++) {
 760     assert(objects->at(i)->is_object(), "invalid debug information");
 761     ObjectValue* sv = (ObjectValue*) objects->at(i);
 762 
 763     KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
 764     oop obj = NULL;
 765 
 766     if (k->oop_is_instance()) {
 767       InstanceKlass* ik = InstanceKlass::cast(k());
 768       obj = ik->allocate_instance(THREAD);
 769     } else if (k->oop_is_typeArray()) {
 770       TypeArrayKlass* ak = TypeArrayKlass::cast(k());
 771       assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
 772       int len = sv->field_size() / type2size[ak->element_type()];
 773       obj = ak->allocate(len, THREAD);
 774     } else if (k->oop_is_objArray()) {
 775       ObjArrayKlass* ak = ObjArrayKlass::cast(k());
 776       obj = ak->allocate(sv->field_size(), THREAD);
 777     }
 778 
 779     if (obj == NULL) {
 780       failures = true;
 781     }
 782 
 783     assert(sv->value().is_null(), "redundant reallocation");
 784     assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
 785     CLEAR_PENDING_EXCEPTION;
 786     sv->set_value(obj);
 787   }
 788 
 789   if (failures) {
 790     THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
 791   } else if (pending_exception.not_null()) {
 792     thread->set_pending_exception(pending_exception(), exception_file, exception_line);
 793   }
 794 
 795   return failures;
 796 }
 797 
 798 // This assumes that the fields are stored in ObjectValue in the same order
 799 // they are yielded by do_nonstatic_fields.
 800 class FieldReassigner: public FieldClosure {
 801   frame* _fr;
 802   RegisterMap* _reg_map;
 803   ObjectValue* _sv;
 804   InstanceKlass* _ik;
 805   oop _obj;
 806 
 807   int _i;
 808 public:
 809   FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) :
 810     _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {}
 811 
 812   int i() const { return _i; }
 813 
 814 
 815   void do_field(fieldDescriptor* fd) {
 816     intptr_t val;
 817     StackValue* value =
 818       StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i()));
 819     int offset = fd->offset();
 820     switch (fd->field_type()) {
 821     case T_OBJECT: case T_ARRAY:
 822       assert(value->type() == T_OBJECT, "Agreement.");
 823       _obj->obj_field_put(offset, value->get_obj()());
 824       break;
 825 
 826     case T_LONG: case T_DOUBLE: {
 827       assert(value->type() == T_INT, "Agreement.");
 828       StackValue* low =
 829         StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i));
 830 #ifdef _LP64
 831       jlong res = (jlong)low->get_int();
 832 #else
 833 #ifdef SPARC
 834       // For SPARC we have to swap high and low words.
 835       jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
 836 #else
 837       jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
 838 #endif //SPARC
 839 #endif
 840       _obj->long_field_put(offset, res);
 841       break;
 842     }
 843     // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
 844     case T_INT: case T_FLOAT: // 4 bytes.
 845       assert(value->type() == T_INT, "Agreement.");
 846       val = value->get_int();
 847       _obj->int_field_put(offset, (jint)*((jint*)&val));
 848       break;
 849 
 850     case T_SHORT:
 851       assert(value->type() == T_INT, "Agreement.");
 852       val = value->get_int();
 853       _obj->short_field_put(offset, (jshort)*((jint*)&val));
 854       break;
 855 
 856     case T_CHAR:
 857       assert(value->type() == T_INT, "Agreement.");
 858       val = value->get_int();
 859       _obj->char_field_put(offset, (jchar)*((jint*)&val));
 860       break;
 861 
 862     case T_BYTE:
 863       assert(value->type() == T_INT, "Agreement.");
 864       val = value->get_int();
 865       _obj->byte_field_put(offset, (jbyte)*((jint*)&val));
 866       break;
 867 
 868     case T_BOOLEAN:
 869       assert(value->type() == T_INT, "Agreement.");
 870       val = value->get_int();
 871       _obj->bool_field_put(offset, (jboolean)*((jint*)&val));
 872       break;
 873 
 874     default:
 875       ShouldNotReachHere();
 876     }
 877     _i++;
 878   }
 879 };
 880 
 881 // restore elements of an eliminated type array
 882 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
 883   int index = 0;
 884   intptr_t val;
 885 
 886   for (int i = 0; i < sv->field_size(); i++) {
 887     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
 888     switch(type) {
 889     case T_LONG: case T_DOUBLE: {
 890       assert(value->type() == T_INT, "Agreement.");
 891       StackValue* low =
 892         StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
 893 #ifdef _LP64
 894       jlong res = (jlong)low->get_int();
 895 #else
 896 #ifdef SPARC
 897       // For SPARC we have to swap high and low words.
 898       jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int());
 899 #else
 900       jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
 901 #endif //SPARC
 902 #endif
 903       obj->long_at_put(index, res);
 904       break;
 905     }
 906 
 907     // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
 908     case T_INT: case T_FLOAT: // 4 bytes.
 909       assert(value->type() == T_INT, "Agreement.");
 910       val = value->get_int();
 911       obj->int_at_put(index, (jint)*((jint*)&val));
 912       break;
 913 
 914     case T_SHORT:
 915       assert(value->type() == T_INT, "Agreement.");
 916       val = value->get_int();
 917       obj->short_at_put(index, (jshort)*((jint*)&val));
 918       break;
 919 
 920     case T_CHAR:
 921       assert(value->type() == T_INT, "Agreement.");
 922       val = value->get_int();
 923       obj->char_at_put(index, (jchar)*((jint*)&val));
 924       break;
 925 
 926     case T_BYTE:
 927       assert(value->type() == T_INT, "Agreement.");
 928       val = value->get_int();
 929       obj->byte_at_put(index, (jbyte)*((jint*)&val));
 930       break;
 931 
 932     case T_BOOLEAN:
 933       assert(value->type() == T_INT, "Agreement.");
 934       val = value->get_int();
 935       obj->bool_at_put(index, (jboolean)*((jint*)&val));
 936       break;
 937 
 938       default:
 939         ShouldNotReachHere();
 940     }
 941     index++;
 942   }
 943 }
 944 
 945 
 946 // restore fields of an eliminated object array
 947 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
 948   for (int i = 0; i < sv->field_size(); i++) {
 949     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
 950     assert(value->type() == T_OBJECT, "object element expected");
 951     obj->obj_at_put(i, value->get_obj()());
 952   }
 953 }
 954 
 955 
 956 // restore fields of all eliminated objects and arrays
 957 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
 958   for (int i = 0; i < objects->length(); i++) {
 959     ObjectValue* sv = (ObjectValue*) objects->at(i);
 960     KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
 961     Handle obj = sv->value();
 962     assert(obj.not_null() || realloc_failures, "reallocation was missed");
 963     if (obj.is_null()) {
 964       continue;
 965     }
 966 
 967     if (k->oop_is_instance()) {
 968       InstanceKlass* ik = InstanceKlass::cast(k());
 969       FieldReassigner reassign(fr, reg_map, sv, obj());
 970       ik->do_nonstatic_fields(&reassign);
 971     } else if (k->oop_is_typeArray()) {
 972       TypeArrayKlass* ak = TypeArrayKlass::cast(k());
 973       reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
 974     } else if (k->oop_is_objArray()) {
 975       reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
 976     }
 977   }
 978 }
 979 
 980 
 981 // relock objects for which synchronization was eliminated
 982 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) {
 983   for (int i = 0; i < monitors->length(); i++) {
 984     MonitorInfo* mon_info = monitors->at(i);
 985     if (mon_info->eliminated()) {
 986       assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
 987       if (!mon_info->owner_is_scalar_replaced()) {
 988         Handle obj = Handle(mon_info->owner());
 989         markOop mark = obj->mark();
 990         if (UseBiasedLocking && mark->has_bias_pattern()) {
 991           // New allocated objects may have the mark set to anonymously biased.
 992           // Also the deoptimized method may called methods with synchronization
 993           // where the thread-local object is bias locked to the current thread.
 994           assert(mark->is_biased_anonymously() ||
 995                  mark->biased_locker() == thread, "should be locked to current thread");
 996           // Reset mark word to unbiased prototype.
 997           markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age());
 998           obj->set_mark(unbiased_prototype);
 999         }
1000         BasicLock* lock = mon_info->lock();
1001         ObjectSynchronizer::slow_enter(obj, lock, thread);
1002         assert(mon_info->owner()->is_locked(), "object must be locked now");
1003       }
1004     }
1005   }
1006 }
1007 
1008 
1009 #ifndef PRODUCT
1010 // print information about reallocated objects
1011 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1012   fieldDescriptor fd;
1013 
1014   for (int i = 0; i < objects->length(); i++) {
1015     ObjectValue* sv = (ObjectValue*) objects->at(i);
1016     KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()));
1017     Handle obj = sv->value();
1018 
1019     tty->print("     object <" INTPTR_FORMAT "> of type ", (void *)sv->value()());
1020     k->print_value();
1021     assert(obj.not_null() || realloc_failures, "reallocation was missed");
1022     if (obj.is_null()) {
1023       tty->print(" allocation failed");
1024     } else {
1025       tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1026     }
1027     tty->cr();
1028 
1029     if (Verbose && !obj.is_null()) {
1030       k->oop_print_on(obj(), tty);
1031     }
1032   }
1033 }
1034 #endif
1035 #endif // COMPILER2
1036 
1037 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1038   Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp());
1039 
1040 #ifndef PRODUCT
1041   if (TraceDeoptimization) {
1042     ttyLocker ttyl;
1043     tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread);
1044     fr.print_on(tty);
1045     tty->print_cr("     Virtual frames (innermost first):");
1046     for (int index = 0; index < chunk->length(); index++) {
1047       compiledVFrame* vf = chunk->at(index);
1048       tty->print("       %2d - ", index);
1049       vf->print_value();
1050       int bci = chunk->at(index)->raw_bci();
1051       const char* code_name;
1052       if (bci == SynchronizationEntryBCI) {
1053         code_name = "sync entry";
1054       } else {
1055         Bytecodes::Code code = vf->method()->code_at(bci);
1056         code_name = Bytecodes::name(code);
1057       }
1058       tty->print(" - %s", code_name);
1059       tty->print_cr(" @ bci %d ", bci);
1060       if (Verbose) {
1061         vf->print();
1062         tty->cr();
1063       }
1064     }
1065   }
1066 #endif
1067 
1068   // Register map for next frame (used for stack crawl).  We capture
1069   // the state of the deopt'ing frame's caller.  Thus if we need to
1070   // stuff a C2I adapter we can properly fill in the callee-save
1071   // register locations.
1072   frame caller = fr.sender(reg_map);
1073   int frame_size = caller.sp() - fr.sp();
1074 
1075   frame sender = caller;
1076 
1077   // Since the Java thread being deoptimized will eventually adjust it's own stack,
1078   // the vframeArray containing the unpacking information is allocated in the C heap.
1079   // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1080   vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1081 
1082   // Compare the vframeArray to the collected vframes
1083   assert(array->structural_compare(thread, chunk), "just checking");
1084 
1085 #ifndef PRODUCT
1086   if (TraceDeoptimization) {
1087     ttyLocker ttyl;
1088     tty->print_cr("     Created vframeArray " INTPTR_FORMAT, array);
1089   }
1090 #endif // PRODUCT
1091 
1092   return array;
1093 }
1094 
1095 #ifdef COMPILER2
1096 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1097   // Reallocation of some scalar replaced objects failed. Record
1098   // that we need to pop all the interpreter frames for the
1099   // deoptimized compiled frame.
1100   assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1101   thread->set_frames_to_pop_failed_realloc(array->frames());
1102   // Unlock all monitors here otherwise the interpreter will see a
1103   // mix of locked and unlocked monitors (because of failed
1104   // reallocations of synchronized objects) and be confused.
1105   for (int i = 0; i < array->frames(); i++) {
1106     MonitorChunk* monitors = array->element(i)->monitors();
1107     if (monitors != NULL) {
1108       for (int j = 0; j < monitors->number_of_monitors(); j++) {
1109         BasicObjectLock* src = monitors->at(j);
1110         if (src->obj() != NULL) {
1111           ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread);
1112         }
1113       }
1114       array->element(i)->free_monitors(thread);
1115 #ifdef ASSERT
1116       array->element(i)->set_removed_monitors();
1117 #endif
1118     }
1119   }
1120 }
1121 #endif
1122 
1123 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) {
1124   GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1125   for (int i = 0; i < monitors->length(); i++) {
1126     MonitorInfo* mon_info = monitors->at(i);
1127     if (!mon_info->eliminated() && mon_info->owner() != NULL) {
1128       objects_to_revoke->append(Handle(mon_info->owner()));
1129     }
1130   }
1131 }
1132 
1133 
1134 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) {
1135   if (!UseBiasedLocking) {
1136     return;
1137   }
1138 
1139   GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1140 
1141   // Unfortunately we don't have a RegisterMap available in most of
1142   // the places we want to call this routine so we need to walk the
1143   // stack again to update the register map.
1144   if (map == NULL || !map->update_map()) {
1145     StackFrameStream sfs(thread, true);
1146     bool found = false;
1147     while (!found && !sfs.is_done()) {
1148       frame* cur = sfs.current();
1149       sfs.next();
1150       found = cur->id() == fr.id();
1151     }
1152     assert(found, "frame to be deoptimized not found on target thread's stack");
1153     map = sfs.register_map();
1154   }
1155 
1156   vframe* vf = vframe::new_vframe(&fr, map, thread);
1157   compiledVFrame* cvf = compiledVFrame::cast(vf);
1158   // Revoke monitors' biases in all scopes
1159   while (!cvf->is_top()) {
1160     collect_monitors(cvf, objects_to_revoke);
1161     cvf = compiledVFrame::cast(cvf->sender());
1162   }
1163   collect_monitors(cvf, objects_to_revoke);
1164 
1165   if (SafepointSynchronize::is_at_safepoint()) {
1166     BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1167   } else {
1168     BiasedLocking::revoke(objects_to_revoke);
1169   }
1170 }
1171 
1172 
1173 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) {
1174   if (!UseBiasedLocking) {
1175     return;
1176   }
1177 
1178   assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint");
1179   GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1180   for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) {
1181     if (jt->has_last_Java_frame()) {
1182       StackFrameStream sfs(jt, true);
1183       while (!sfs.is_done()) {
1184         frame* cur = sfs.current();
1185         if (cb->contains(cur->pc())) {
1186           vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt);
1187           compiledVFrame* cvf = compiledVFrame::cast(vf);
1188           // Revoke monitors' biases in all scopes
1189           while (!cvf->is_top()) {
1190             collect_monitors(cvf, objects_to_revoke);
1191             cvf = compiledVFrame::cast(cvf->sender());
1192           }
1193           collect_monitors(cvf, objects_to_revoke);
1194         }
1195         sfs.next();
1196       }
1197     }
1198   }
1199   BiasedLocking::revoke_at_safepoint(objects_to_revoke);
1200 }
1201 
1202 
1203 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) {
1204   assert(fr.can_be_deoptimized(), "checking frame type");
1205 
1206   gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal);
1207 
1208   // Patch the nmethod so that when execution returns to it we will
1209   // deopt the execution state and return to the interpreter.
1210   fr.deoptimize(thread);
1211 }
1212 
1213 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) {
1214   // Deoptimize only if the frame comes from compile code.
1215   // Do not deoptimize the frame which is already patched
1216   // during the execution of the loops below.
1217   if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1218     return;
1219   }
1220   ResourceMark rm;
1221   DeoptimizationMarker dm;
1222   if (UseBiasedLocking) {
1223     revoke_biases_of_monitors(thread, fr, map);
1224   }
1225   deoptimize_single_frame(thread, fr);
1226 
1227 }
1228 
1229 
1230 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) {
1231   assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1232          "can only deoptimize other thread at a safepoint");
1233   // Compute frame and register map based on thread and sp.
1234   RegisterMap reg_map(thread, UseBiasedLocking);
1235   frame fr = thread->last_frame();
1236   while (fr.id() != id) {
1237     fr = fr.sender(&reg_map);
1238   }
1239   deoptimize(thread, fr, &reg_map);
1240 }
1241 
1242 
1243 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1244   if (thread == Thread::current()) {
1245     Deoptimization::deoptimize_frame_internal(thread, id);
1246   } else {
1247     VM_DeoptimizeFrame deopt(thread, id);
1248     VMThread::execute(&deopt);
1249   }
1250 }
1251 
1252 
1253 // JVMTI PopFrame support
1254 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1255 {
1256   thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1257 }
1258 JRT_END
1259 
1260 
1261 #if defined(COMPILER2) || defined(SHARK)
1262 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) {
1263   // in case of an unresolved klass entry, load the class.
1264   if (constant_pool->tag_at(index).is_unresolved_klass()) {
1265     Klass* tk = constant_pool->klass_at(index, CHECK);
1266     return;
1267   }
1268 
1269   if (!constant_pool->tag_at(index).is_symbol()) return;
1270 
1271   Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader());
1272   Symbol*  symbol  = constant_pool->symbol_at(index);
1273 
1274   // class name?
1275   if (symbol->byte_at(0) != '(') {
1276     Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1277     SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK);
1278     return;
1279   }
1280 
1281   // then it must be a signature!
1282   ResourceMark rm(THREAD);
1283   for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) {
1284     if (ss.is_object()) {
1285       Symbol* class_name = ss.as_symbol(CHECK);
1286       Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain());
1287       SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK);
1288     }
1289   }
1290 }
1291 
1292 
1293 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) {
1294   EXCEPTION_MARK;
1295   load_class_by_index(constant_pool, index, THREAD);
1296   if (HAS_PENDING_EXCEPTION) {
1297     // Exception happened during classloading. We ignore the exception here, since it
1298     // is going to be rethrown since the current activation is going to be deoptimized and
1299     // the interpreter will re-execute the bytecode.
1300     CLEAR_PENDING_EXCEPTION;
1301     // Class loading called java code which may have caused a stack
1302     // overflow. If the exception was thrown right before the return
1303     // to the runtime the stack is no longer guarded. Reguard the
1304     // stack otherwise if we return to the uncommon trap blob and the
1305     // stack bang causes a stack overflow we crash.
1306     assert(THREAD->is_Java_thread(), "only a java thread can be here");
1307     JavaThread* thread = (JavaThread*)THREAD;
1308     bool guard_pages_enabled = thread->stack_yellow_zone_enabled();
1309     if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1310     assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1311   }
1312 }
1313 
1314 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1315   HandleMark hm;
1316 
1317   // uncommon_trap() is called at the beginning of the uncommon trap
1318   // handler. Note this fact before we start generating temporary frames
1319   // that can confuse an asynchronous stack walker. This counter is
1320   // decremented at the end of unpack_frames().
1321   thread->inc_in_deopt_handler();
1322 
1323   // We need to update the map if we have biased locking.
1324   RegisterMap reg_map(thread, UseBiasedLocking);
1325   frame stub_frame = thread->last_frame();
1326   frame fr = stub_frame.sender(&reg_map);
1327   // Make sure the calling nmethod is not getting deoptimized and removed
1328   // before we are done with it.
1329   nmethodLocker nl(fr.pc());
1330 
1331   // Log a message
1332   Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT,
1333               trap_request, fr.pc());
1334 
1335   {
1336     ResourceMark rm;
1337 
1338     // Revoke biases of any monitors in the frame to ensure we can migrate them
1339     revoke_biases_of_monitors(thread, fr, &reg_map);
1340 
1341     DeoptReason reason = trap_request_reason(trap_request);
1342     DeoptAction action = trap_request_action(trap_request);
1343     jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1344 
1345     vframe*  vf  = vframe::new_vframe(&fr, &reg_map, thread);
1346     compiledVFrame* cvf = compiledVFrame::cast(vf);
1347 
1348     nmethod* nm = cvf->code();
1349 
1350     ScopeDesc*      trap_scope  = cvf->scope();
1351     methodHandle    trap_method = trap_scope->method();
1352     int             trap_bci    = trap_scope->bci();
1353     Bytecodes::Code trap_bc     = trap_method->java_code_at(trap_bci);
1354 
1355     // Record this event in the histogram.
1356     gather_statistics(reason, action, trap_bc);
1357 
1358     // Ensure that we can record deopt. history:
1359     // Need MDO to record RTM code generation state.
1360     bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking );
1361 
1362     MethodData* trap_mdo =
1363       get_method_data(thread, trap_method, create_if_missing);
1364 
1365     // Log a message
1366     Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d",
1367                               trap_reason_name(reason), trap_action_name(action), fr.pc(),
1368                               trap_method->name_and_sig_as_C_string(), trap_bci);
1369 
1370     // Print a bunch of diagnostics, if requested.
1371     if (TraceDeoptimization || LogCompilation) {
1372       ResourceMark rm;
1373       ttyLocker ttyl;
1374       char buf[100];
1375       if (xtty != NULL) {
1376         xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1377                          os::current_thread_id(),
1378                          format_trap_request(buf, sizeof(buf), trap_request));
1379         nm->log_identity(xtty);
1380       }
1381       Symbol* class_name = NULL;
1382       bool unresolved = false;
1383       if (unloaded_class_index >= 0) {
1384         constantPoolHandle constants (THREAD, trap_method->constants());
1385         if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
1386           class_name = constants->klass_name_at(unloaded_class_index);
1387           unresolved = true;
1388           if (xtty != NULL)
1389             xtty->print(" unresolved='1'");
1390         } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
1391           class_name = constants->symbol_at(unloaded_class_index);
1392         }
1393         if (xtty != NULL)
1394           xtty->name(class_name);
1395       }
1396       if (xtty != NULL && trap_mdo != NULL) {
1397         // Dump the relevant MDO state.
1398         // This is the deopt count for the current reason, any previous
1399         // reasons or recompiles seen at this point.
1400         int dcnt = trap_mdo->trap_count(reason);
1401         if (dcnt != 0)
1402           xtty->print(" count='%d'", dcnt);
1403         ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
1404         int dos = (pdata == NULL)? 0: pdata->trap_state();
1405         if (dos != 0) {
1406           xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
1407           if (trap_state_is_recompiled(dos)) {
1408             int recnt2 = trap_mdo->overflow_recompile_count();
1409             if (recnt2 != 0)
1410               xtty->print(" recompiles2='%d'", recnt2);
1411           }
1412         }
1413       }
1414       if (xtty != NULL) {
1415         xtty->stamp();
1416         xtty->end_head();
1417       }
1418       if (TraceDeoptimization) {  // make noise on the tty
1419         tty->print("Uncommon trap occurred in");
1420         nm->method()->print_short_name(tty);
1421         tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d",
1422                    fr.pc(),
1423                    os::current_thread_id(),
1424                    trap_reason_name(reason),
1425                    trap_action_name(action),
1426                    unloaded_class_index);
1427         if (class_name != NULL) {
1428           tty->print(unresolved ? " unresolved class: " : " symbol: ");
1429           class_name->print_symbol_on(tty);
1430         }
1431         tty->cr();
1432       }
1433       if (xtty != NULL) {
1434         // Log the precise location of the trap.
1435         for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
1436           xtty->begin_elem("jvms bci='%d'", sd->bci());
1437           xtty->method(sd->method());
1438           xtty->end_elem();
1439           if (sd->is_top())  break;
1440         }
1441         xtty->tail("uncommon_trap");
1442       }
1443     }
1444     // (End diagnostic printout.)
1445 
1446     // Load class if necessary
1447     if (unloaded_class_index >= 0) {
1448       constantPoolHandle constants(THREAD, trap_method->constants());
1449       load_class_by_index(constants, unloaded_class_index);
1450     }
1451 
1452     // Flush the nmethod if necessary and desirable.
1453     //
1454     // We need to avoid situations where we are re-flushing the nmethod
1455     // because of a hot deoptimization site.  Repeated flushes at the same
1456     // point need to be detected by the compiler and avoided.  If the compiler
1457     // cannot avoid them (or has a bug and "refuses" to avoid them), this
1458     // module must take measures to avoid an infinite cycle of recompilation
1459     // and deoptimization.  There are several such measures:
1460     //
1461     //   1. If a recompilation is ordered a second time at some site X
1462     //   and for the same reason R, the action is adjusted to 'reinterpret',
1463     //   to give the interpreter time to exercise the method more thoroughly.
1464     //   If this happens, the method's overflow_recompile_count is incremented.
1465     //
1466     //   2. If the compiler fails to reduce the deoptimization rate, then
1467     //   the method's overflow_recompile_count will begin to exceed the set
1468     //   limit PerBytecodeRecompilationCutoff.  If this happens, the action
1469     //   is adjusted to 'make_not_compilable', and the method is abandoned
1470     //   to the interpreter.  This is a performance hit for hot methods,
1471     //   but is better than a disastrous infinite cycle of recompilations.
1472     //   (Actually, only the method containing the site X is abandoned.)
1473     //
1474     //   3. In parallel with the previous measures, if the total number of
1475     //   recompilations of a method exceeds the much larger set limit
1476     //   PerMethodRecompilationCutoff, the method is abandoned.
1477     //   This should only happen if the method is very large and has
1478     //   many "lukewarm" deoptimizations.  The code which enforces this
1479     //   limit is elsewhere (class nmethod, class Method).
1480     //
1481     // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
1482     // to recompile at each bytecode independently of the per-BCI cutoff.
1483     //
1484     // The decision to update code is up to the compiler, and is encoded
1485     // in the Action_xxx code.  If the compiler requests Action_none
1486     // no trap state is changed, no compiled code is changed, and the
1487     // computation suffers along in the interpreter.
1488     //
1489     // The other action codes specify various tactics for decompilation
1490     // and recompilation.  Action_maybe_recompile is the loosest, and
1491     // allows the compiled code to stay around until enough traps are seen,
1492     // and until the compiler gets around to recompiling the trapping method.
1493     //
1494     // The other actions cause immediate removal of the present code.
1495 
1496     // Traps caused by injected profile shouldn't pollute trap counts.
1497     bool injected_profile_trap = trap_method->has_injected_profile() &&
1498                                  (reason == Reason_intrinsic || reason == Reason_unreached);
1499     bool update_trap_state = !injected_profile_trap;
1500     bool make_not_entrant = false;
1501     bool make_not_compilable = false;
1502     bool reprofile = false;
1503     switch (action) {
1504     case Action_none:
1505       // Keep the old code.
1506       update_trap_state = false;
1507       break;
1508     case Action_maybe_recompile:
1509       // Do not need to invalidate the present code, but we can
1510       // initiate another
1511       // Start compiler without (necessarily) invalidating the nmethod.
1512       // The system will tolerate the old code, but new code should be
1513       // generated when possible.
1514       break;
1515     case Action_reinterpret:
1516       // Go back into the interpreter for a while, and then consider
1517       // recompiling form scratch.
1518       make_not_entrant = true;
1519       // Reset invocation counter for outer most method.
1520       // This will allow the interpreter to exercise the bytecodes
1521       // for a while before recompiling.
1522       // By contrast, Action_make_not_entrant is immediate.
1523       //
1524       // Note that the compiler will track null_check, null_assert,
1525       // range_check, and class_check events and log them as if they
1526       // had been traps taken from compiled code.  This will update
1527       // the MDO trap history so that the next compilation will
1528       // properly detect hot trap sites.
1529       reprofile = true;
1530       break;
1531     case Action_make_not_entrant:
1532       // Request immediate recompilation, and get rid of the old code.
1533       // Make them not entrant, so next time they are called they get
1534       // recompiled.  Unloaded classes are loaded now so recompile before next
1535       // time they are called.  Same for uninitialized.  The interpreter will
1536       // link the missing class, if any.
1537       make_not_entrant = true;
1538       break;
1539     case Action_make_not_compilable:
1540       // Give up on compiling this method at all.
1541       make_not_entrant = true;
1542       make_not_compilable = true;
1543       break;
1544     default:
1545       ShouldNotReachHere();
1546     }
1547 
1548     // Setting +ProfileTraps fixes the following, on all platforms:
1549     // 4852688: ProfileInterpreter is off by default for ia64.  The result is
1550     // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
1551     // recompile relies on a MethodData* to record heroic opt failures.
1552 
1553     // Whether the interpreter is producing MDO data or not, we also need
1554     // to use the MDO to detect hot deoptimization points and control
1555     // aggressive optimization.
1556     bool inc_recompile_count = false;
1557     ProfileData* pdata = NULL;
1558     if (ProfileTraps && update_trap_state && trap_mdo != NULL) {
1559       assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity");
1560       uint this_trap_count = 0;
1561       bool maybe_prior_trap = false;
1562       bool maybe_prior_recompile = false;
1563       pdata = query_update_method_data(trap_mdo, trap_bci, reason,
1564                                    nm->method(),
1565                                    //outputs:
1566                                    this_trap_count,
1567                                    maybe_prior_trap,
1568                                    maybe_prior_recompile);
1569       // Because the interpreter also counts null, div0, range, and class
1570       // checks, these traps from compiled code are double-counted.
1571       // This is harmless; it just means that the PerXTrapLimit values
1572       // are in effect a little smaller than they look.
1573 
1574       DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1575       if (per_bc_reason != Reason_none) {
1576         // Now take action based on the partially known per-BCI history.
1577         if (maybe_prior_trap
1578             && this_trap_count >= (uint)PerBytecodeTrapLimit) {
1579           // If there are too many traps at this BCI, force a recompile.
1580           // This will allow the compiler to see the limit overflow, and
1581           // take corrective action, if possible.  The compiler generally
1582           // does not use the exact PerBytecodeTrapLimit value, but instead
1583           // changes its tactics if it sees any traps at all.  This provides
1584           // a little hysteresis, delaying a recompile until a trap happens
1585           // several times.
1586           //
1587           // Actually, since there is only one bit of counter per BCI,
1588           // the possible per-BCI counts are {0,1,(per-method count)}.
1589           // This produces accurate results if in fact there is only
1590           // one hot trap site, but begins to get fuzzy if there are
1591           // many sites.  For example, if there are ten sites each
1592           // trapping two or more times, they each get the blame for
1593           // all of their traps.
1594           make_not_entrant = true;
1595         }
1596 
1597         // Detect repeated recompilation at the same BCI, and enforce a limit.
1598         if (make_not_entrant && maybe_prior_recompile) {
1599           // More than one recompile at this point.
1600           inc_recompile_count = maybe_prior_trap;
1601         }
1602       } else {
1603         // For reasons which are not recorded per-bytecode, we simply
1604         // force recompiles unconditionally.
1605         // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
1606         make_not_entrant = true;
1607       }
1608 
1609       // Go back to the compiler if there are too many traps in this method.
1610       if (this_trap_count >= per_method_trap_limit(reason)) {
1611         // If there are too many traps in this method, force a recompile.
1612         // This will allow the compiler to see the limit overflow, and
1613         // take corrective action, if possible.
1614         // (This condition is an unlikely backstop only, because the
1615         // PerBytecodeTrapLimit is more likely to take effect first,
1616         // if it is applicable.)
1617         make_not_entrant = true;
1618       }
1619 
1620       // Here's more hysteresis:  If there has been a recompile at
1621       // this trap point already, run the method in the interpreter
1622       // for a while to exercise it more thoroughly.
1623       if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
1624         reprofile = true;
1625       }
1626 
1627     }
1628 
1629     // Take requested actions on the method:
1630 
1631     // Recompile
1632     if (make_not_entrant) {
1633       if (!nm->make_not_entrant()) {
1634         return; // the call did not change nmethod's state
1635       }
1636 
1637       if (pdata != NULL) {
1638         // Record the recompilation event, if any.
1639         int tstate0 = pdata->trap_state();
1640         int tstate1 = trap_state_set_recompiled(tstate0, true);
1641         if (tstate1 != tstate0)
1642           pdata->set_trap_state(tstate1);
1643       }
1644 
1645 #if INCLUDE_RTM_OPT
1646       // Restart collecting RTM locking abort statistic if the method
1647       // is recompiled for a reason other than RTM state change.
1648       // Assume that in new recompiled code the statistic could be different,
1649       // for example, due to different inlining.
1650       if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
1651           UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) {
1652         trap_mdo->atomic_set_rtm_state(ProfileRTM);
1653       }
1654 #endif
1655     }
1656 
1657     if (inc_recompile_count) {
1658       trap_mdo->inc_overflow_recompile_count();
1659       if ((uint)trap_mdo->overflow_recompile_count() >
1660           (uint)PerBytecodeRecompilationCutoff) {
1661         // Give up on the method containing the bad BCI.
1662         if (trap_method() == nm->method()) {
1663           make_not_compilable = true;
1664         } else {
1665           trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff");
1666           // But give grace to the enclosing nm->method().
1667         }
1668       }
1669     }
1670 
1671     // Reprofile
1672     if (reprofile) {
1673       CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
1674     }
1675 
1676     // Give up compiling
1677     if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
1678       assert(make_not_entrant, "consistent");
1679       nm->method()->set_not_compilable(CompLevel_full_optimization);
1680     }
1681 
1682   } // Free marked resources
1683 
1684 }
1685 JRT_END
1686 
1687 MethodData*
1688 Deoptimization::get_method_data(JavaThread* thread, methodHandle m,
1689                                 bool create_if_missing) {
1690   Thread* THREAD = thread;
1691   MethodData* mdo = m()->method_data();
1692   if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1693     // Build an MDO.  Ignore errors like OutOfMemory;
1694     // that simply means we won't have an MDO to update.
1695     Method::build_interpreter_method_data(m, THREAD);
1696     if (HAS_PENDING_EXCEPTION) {
1697       assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1698       CLEAR_PENDING_EXCEPTION;
1699     }
1700     mdo = m()->method_data();
1701   }
1702   return mdo;
1703 }
1704 
1705 ProfileData*
1706 Deoptimization::query_update_method_data(MethodData* trap_mdo,
1707                                          int trap_bci,
1708                                          Deoptimization::DeoptReason reason,
1709                                          Method* compiled_method,
1710                                          //outputs:
1711                                          uint& ret_this_trap_count,
1712                                          bool& ret_maybe_prior_trap,
1713                                          bool& ret_maybe_prior_recompile) {
1714   uint prior_trap_count = trap_mdo->trap_count(reason);
1715   uint this_trap_count  = trap_mdo->inc_trap_count(reason);
1716 
1717   // If the runtime cannot find a place to store trap history,
1718   // it is estimated based on the general condition of the method.
1719   // If the method has ever been recompiled, or has ever incurred
1720   // a trap with the present reason , then this BCI is assumed
1721   // (pessimistically) to be the culprit.
1722   bool maybe_prior_trap      = (prior_trap_count != 0);
1723   bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
1724   ProfileData* pdata = NULL;
1725 
1726 
1727   // For reasons which are recorded per bytecode, we check per-BCI data.
1728   DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
1729   if (per_bc_reason != Reason_none) {
1730     // Find the profile data for this BCI.  If there isn't one,
1731     // try to allocate one from the MDO's set of spares.
1732     // This will let us detect a repeated trap at this point.
1733     pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
1734 
1735     if (pdata != NULL) {
1736       if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
1737         if (LogCompilation && xtty != NULL) {
1738           ttyLocker ttyl;
1739           // no more room for speculative traps in this MDO
1740           xtty->elem("speculative_traps_oom");
1741         }
1742       }
1743       // Query the trap state of this profile datum.
1744       int tstate0 = pdata->trap_state();
1745       if (!trap_state_has_reason(tstate0, per_bc_reason))
1746         maybe_prior_trap = false;
1747       if (!trap_state_is_recompiled(tstate0))
1748         maybe_prior_recompile = false;
1749 
1750       // Update the trap state of this profile datum.
1751       int tstate1 = tstate0;
1752       // Record the reason.
1753       tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
1754       // Store the updated state on the MDO, for next time.
1755       if (tstate1 != tstate0)
1756         pdata->set_trap_state(tstate1);
1757     } else {
1758       if (LogCompilation && xtty != NULL) {
1759         ttyLocker ttyl;
1760         // Missing MDP?  Leave a small complaint in the log.
1761         xtty->elem("missing_mdp bci='%d'", trap_bci);
1762       }
1763     }
1764   }
1765 
1766   // Return results:
1767   ret_this_trap_count = this_trap_count;
1768   ret_maybe_prior_trap = maybe_prior_trap;
1769   ret_maybe_prior_recompile = maybe_prior_recompile;
1770   return pdata;
1771 }
1772 
1773 void
1774 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
1775   ResourceMark rm;
1776   // Ignored outputs:
1777   uint ignore_this_trap_count;
1778   bool ignore_maybe_prior_trap;
1779   bool ignore_maybe_prior_recompile;
1780   assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
1781   query_update_method_data(trap_mdo, trap_bci,
1782                            (DeoptReason)reason,
1783                            NULL,
1784                            ignore_this_trap_count,
1785                            ignore_maybe_prior_trap,
1786                            ignore_maybe_prior_recompile);
1787 }
1788 
1789 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) {
1790 
1791   // Still in Java no safepoints
1792   {
1793     // This enters VM and may safepoint
1794     uncommon_trap_inner(thread, trap_request);
1795   }
1796   return fetch_unroll_info_helper(thread);
1797 }
1798 
1799 // Local derived constants.
1800 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
1801 const int DS_REASON_MASK   = DataLayout::trap_mask >> 1;
1802 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
1803 
1804 //---------------------------trap_state_reason---------------------------------
1805 Deoptimization::DeoptReason
1806 Deoptimization::trap_state_reason(int trap_state) {
1807   // This assert provides the link between the width of DataLayout::trap_bits
1808   // and the encoding of "recorded" reasons.  It ensures there are enough
1809   // bits to store all needed reasons in the per-BCI MDO profile.
1810   assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1811   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1812   trap_state -= recompile_bit;
1813   if (trap_state == DS_REASON_MASK) {
1814     return Reason_many;
1815   } else {
1816     assert((int)Reason_none == 0, "state=0 => Reason_none");
1817     return (DeoptReason)trap_state;
1818   }
1819 }
1820 //-------------------------trap_state_has_reason-------------------------------
1821 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
1822   assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
1823   assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
1824   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1825   trap_state -= recompile_bit;
1826   if (trap_state == DS_REASON_MASK) {
1827     return -1;  // true, unspecifically (bottom of state lattice)
1828   } else if (trap_state == reason) {
1829     return 1;   // true, definitely
1830   } else if (trap_state == 0) {
1831     return 0;   // false, definitely (top of state lattice)
1832   } else {
1833     return 0;   // false, definitely
1834   }
1835 }
1836 //-------------------------trap_state_add_reason-------------------------------
1837 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
1838   assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
1839   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
1840   trap_state -= recompile_bit;
1841   if (trap_state == DS_REASON_MASK) {
1842     return trap_state + recompile_bit;     // already at state lattice bottom
1843   } else if (trap_state == reason) {
1844     return trap_state + recompile_bit;     // the condition is already true
1845   } else if (trap_state == 0) {
1846     return reason + recompile_bit;          // no condition has yet been true
1847   } else {
1848     return DS_REASON_MASK + recompile_bit;  // fall to state lattice bottom
1849   }
1850 }
1851 //-----------------------trap_state_is_recompiled------------------------------
1852 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
1853   return (trap_state & DS_RECOMPILE_BIT) != 0;
1854 }
1855 //-----------------------trap_state_set_recompiled-----------------------------
1856 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
1857   if (z)  return trap_state |  DS_RECOMPILE_BIT;
1858   else    return trap_state & ~DS_RECOMPILE_BIT;
1859 }
1860 //---------------------------format_trap_state---------------------------------
1861 // This is used for debugging and diagnostics, including LogFile output.
1862 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
1863                                               int trap_state) {
1864   DeoptReason reason      = trap_state_reason(trap_state);
1865   bool        recomp_flag = trap_state_is_recompiled(trap_state);
1866   // Re-encode the state from its decoded components.
1867   int decoded_state = 0;
1868   if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
1869     decoded_state = trap_state_add_reason(decoded_state, reason);
1870   if (recomp_flag)
1871     decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
1872   // If the state re-encodes properly, format it symbolically.
1873   // Because this routine is used for debugging and diagnostics,
1874   // be robust even if the state is a strange value.
1875   size_t len;
1876   if (decoded_state != trap_state) {
1877     // Random buggy state that doesn't decode??
1878     len = jio_snprintf(buf, buflen, "#%d", trap_state);
1879   } else {
1880     len = jio_snprintf(buf, buflen, "%s%s",
1881                        trap_reason_name(reason),
1882                        recomp_flag ? " recompiled" : "");
1883   }
1884   return buf;
1885 }
1886 
1887 
1888 //--------------------------------statics--------------------------------------
1889 Deoptimization::DeoptAction Deoptimization::_unloaded_action
1890   = Deoptimization::Action_reinterpret;
1891 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = {
1892   // Note:  Keep this in sync. with enum DeoptReason.
1893   "none",
1894   "null_check",
1895   "null_assert",
1896   "range_check",
1897   "class_check",
1898   "array_check",
1899   "intrinsic",
1900   "bimorphic",
1901   "unloaded",
1902   "uninitialized",
1903   "unreached",
1904   "unhandled",
1905   "constraint",
1906   "div0_check",
1907   "age",
1908   "predicate",
1909   "loop_limit_check",
1910   "speculate_class_check",
1911   "rtm_state_change",
1912   "unstable_if"
1913 };
1914 const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
1915   // Note:  Keep this in sync. with enum DeoptAction.
1916   "none",
1917   "maybe_recompile",
1918   "reinterpret",
1919   "make_not_entrant",
1920   "make_not_compilable"
1921 };
1922 
1923 const char* Deoptimization::trap_reason_name(int reason) {
1924   if (reason == Reason_many)  return "many";
1925   if ((uint)reason < Reason_LIMIT)
1926     return _trap_reason_name[reason];
1927   static char buf[20];
1928   sprintf(buf, "reason%d", reason);
1929   return buf;
1930 }
1931 const char* Deoptimization::trap_action_name(int action) {
1932   if ((uint)action < Action_LIMIT)
1933     return _trap_action_name[action];
1934   static char buf[20];
1935   sprintf(buf, "action%d", action);
1936   return buf;
1937 }
1938 
1939 // This is used for debugging and diagnostics, including LogFile output.
1940 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
1941                                                 int trap_request) {
1942   jint unloaded_class_index = trap_request_index(trap_request);
1943   const char* reason = trap_reason_name(trap_request_reason(trap_request));
1944   const char* action = trap_action_name(trap_request_action(trap_request));
1945   size_t len;
1946   if (unloaded_class_index < 0) {
1947     len = jio_snprintf(buf, buflen, "reason='%s' action='%s'",
1948                        reason, action);
1949   } else {
1950     len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'",
1951                        reason, action, unloaded_class_index);
1952   }
1953   return buf;
1954 }
1955 
1956 juint Deoptimization::_deoptimization_hist
1957         [Deoptimization::Reason_LIMIT]
1958     [1 + Deoptimization::Action_LIMIT]
1959         [Deoptimization::BC_CASE_LIMIT]
1960   = {0};
1961 
1962 enum {
1963   LSB_BITS = 8,
1964   LSB_MASK = right_n_bits(LSB_BITS)
1965 };
1966 
1967 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
1968                                        Bytecodes::Code bc) {
1969   assert(reason >= 0 && reason < Reason_LIMIT, "oob");
1970   assert(action >= 0 && action < Action_LIMIT, "oob");
1971   _deoptimization_hist[Reason_none][0][0] += 1;  // total
1972   _deoptimization_hist[reason][0][0]      += 1;  // per-reason total
1973   juint* cases = _deoptimization_hist[reason][1+action];
1974   juint* bc_counter_addr = NULL;
1975   juint  bc_counter      = 0;
1976   // Look for an unused counter, or an exact match to this BC.
1977   if (bc != Bytecodes::_illegal) {
1978     for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
1979       juint* counter_addr = &cases[bc_case];
1980       juint  counter = *counter_addr;
1981       if ((counter == 0 && bc_counter_addr == NULL)
1982           || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
1983         // this counter is either free or is already devoted to this BC
1984         bc_counter_addr = counter_addr;
1985         bc_counter = counter | bc;
1986       }
1987     }
1988   }
1989   if (bc_counter_addr == NULL) {
1990     // Overflow, or no given bytecode.
1991     bc_counter_addr = &cases[BC_CASE_LIMIT-1];
1992     bc_counter = (*bc_counter_addr & ~LSB_MASK);  // clear LSB
1993   }
1994   *bc_counter_addr = bc_counter + (1 << LSB_BITS);
1995 }
1996 
1997 jint Deoptimization::total_deoptimization_count() {
1998   return _deoptimization_hist[Reason_none][0][0];
1999 }
2000 
2001 jint Deoptimization::deoptimization_count(DeoptReason reason) {
2002   assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2003   return _deoptimization_hist[reason][0][0];
2004 }
2005 
2006 void Deoptimization::print_statistics() {
2007   juint total = total_deoptimization_count();
2008   juint account = total;
2009   if (total != 0) {
2010     ttyLocker ttyl;
2011     if (xtty != NULL)  xtty->head("statistics type='deoptimization'");
2012     tty->print_cr("Deoptimization traps recorded:");
2013     #define PRINT_STAT_LINE(name, r) \
2014       tty->print_cr("  %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2015     PRINT_STAT_LINE("total", total);
2016     // For each non-zero entry in the histogram, print the reason,
2017     // the action, and (if specifically known) the type of bytecode.
2018     for (int reason = 0; reason < Reason_LIMIT; reason++) {
2019       for (int action = 0; action < Action_LIMIT; action++) {
2020         juint* cases = _deoptimization_hist[reason][1+action];
2021         for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2022           juint counter = cases[bc_case];
2023           if (counter != 0) {
2024             char name[1*K];
2025             Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2026             if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2027               bc = Bytecodes::_illegal;
2028             sprintf(name, "%s/%s/%s",
2029                     trap_reason_name(reason),
2030                     trap_action_name(action),
2031                     Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2032             juint r = counter >> LSB_BITS;
2033             tty->print_cr("  %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2034             account -= r;
2035           }
2036         }
2037       }
2038     }
2039     if (account != 0) {
2040       PRINT_STAT_LINE("unaccounted", account);
2041     }
2042     #undef PRINT_STAT_LINE
2043     if (xtty != NULL)  xtty->tail("statistics");
2044   }
2045 }
2046 #else // COMPILER2 || SHARK
2047 
2048 
2049 // Stubs for C1 only system.
2050 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2051   return false;
2052 }
2053 
2054 const char* Deoptimization::trap_reason_name(int reason) {
2055   return "unknown";
2056 }
2057 
2058 void Deoptimization::print_statistics() {
2059   // no output
2060 }
2061 
2062 void
2063 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2064   // no udpate
2065 }
2066 
2067 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2068   return 0;
2069 }
2070 
2071 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2072                                        Bytecodes::Code bc) {
2073   // no update
2074 }
2075 
2076 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2077                                               int trap_state) {
2078   jio_snprintf(buf, buflen, "#%d", trap_state);
2079   return buf;
2080 }
2081 
2082 #endif // COMPILER2 || SHARK