1 2 3 /* 4 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved. 5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 6 * 7 * This code is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License version 2 only, as 9 * published by the Free Software Foundation. 10 * 11 * This code is distributed in the hope that it will be useful, but WITHOUT 12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 14 * version 2 for more details (a copy is included in the LICENSE file that 15 * accompanied this code). 16 * 17 * You should have received a copy of the GNU General Public License version 18 * 2 along with this work; if not, write to the Free Software Foundation, 19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 20 * 21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 22 * or visit www.oracle.com if you need additional information or have any 23 * questions. 24 * 25 */ 26 27 #include "precompiled.hpp" 28 #include "jvm.h" 29 #include "classfile/symbolTable.hpp" 30 #include "classfile/systemDictionary.hpp" 31 #include "code/codeCache.hpp" 32 #include "code/debugInfoRec.hpp" 33 #include "code/nmethod.hpp" 34 #include "code/pcDesc.hpp" 35 #include "code/scopeDesc.hpp" 36 #include "compiler/compilationPolicy.hpp" 37 #include "interpreter/bytecode.hpp" 38 #include "interpreter/interpreter.hpp" 39 #include "interpreter/oopMapCache.hpp" 40 #include "memory/allocation.inline.hpp" 41 #include "memory/oopFactory.hpp" 42 #include "memory/resourceArea.hpp" 43 #include "memory/universe.hpp" 44 #include "oops/constantPool.hpp" 45 #include "oops/method.hpp" 46 #include "oops/objArrayKlass.hpp" 47 #include "oops/objArrayOop.inline.hpp" 48 #include "oops/oop.inline.hpp" 49 #include "oops/fieldStreams.inline.hpp" 50 #include "oops/typeArrayOop.inline.hpp" 51 #include "oops/verifyOopClosure.hpp" 52 #include "prims/jvmtiThreadState.hpp" 53 #include "runtime/atomic.hpp" 54 #include "runtime/biasedLocking.hpp" 55 #include "runtime/deoptimization.hpp" 56 #include "runtime/fieldDescriptor.hpp" 57 #include "runtime/fieldDescriptor.inline.hpp" 58 #include "runtime/frame.inline.hpp" 59 #include "runtime/jniHandles.inline.hpp" 60 #include "runtime/handles.inline.hpp" 61 #include "runtime/interfaceSupport.inline.hpp" 62 #include "runtime/safepointVerifiers.hpp" 63 #include "runtime/sharedRuntime.hpp" 64 #include "runtime/signature.hpp" 65 #include "runtime/stubRoutines.hpp" 66 #include "runtime/thread.hpp" 67 #include "runtime/threadSMR.hpp" 68 #include "runtime/vframe.hpp" 69 #include "runtime/vframeArray.hpp" 70 #include "runtime/vframe_hp.hpp" 71 #include "utilities/events.hpp" 72 #include "utilities/preserveException.hpp" 73 #include "utilities/xmlstream.hpp" 74 75 76 bool DeoptimizationMarker::_is_active = false; 77 78 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 79 int caller_adjustment, 80 int caller_actual_parameters, 81 int number_of_frames, 82 intptr_t* frame_sizes, 83 address* frame_pcs, 84 BasicType return_type, 85 int exec_mode) { 86 _size_of_deoptimized_frame = size_of_deoptimized_frame; 87 _caller_adjustment = caller_adjustment; 88 _caller_actual_parameters = caller_actual_parameters; 89 _number_of_frames = number_of_frames; 90 _frame_sizes = frame_sizes; 91 _frame_pcs = frame_pcs; 92 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 93 _return_type = return_type; 94 _initial_info = 0; 95 // PD (x86 only) 96 _counter_temp = 0; 97 _unpack_kind = exec_mode; 98 _sender_sp_temp = 0; 99 100 _total_frame_sizes = size_of_frames(); 101 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 102 } 103 104 105 Deoptimization::UnrollBlock::~UnrollBlock() { 106 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 107 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 108 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 109 } 110 111 112 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 113 assert(register_number < RegisterMap::reg_count, "checking register number"); 114 return &_register_block[register_number * 2]; 115 } 116 117 118 119 int Deoptimization::UnrollBlock::size_of_frames() const { 120 // Acount first for the adjustment of the initial frame 121 int result = _caller_adjustment; 122 for (int index = 0; index < number_of_frames(); index++) { 123 result += frame_sizes()[index]; 124 } 125 return result; 126 } 127 128 129 void Deoptimization::UnrollBlock::print() { 130 ttyLocker ttyl; 131 tty->print_cr("UnrollBlock"); 132 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 133 tty->print( " frame_sizes: "); 134 for (int index = 0; index < number_of_frames(); index++) { 135 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 136 } 137 tty->cr(); 138 } 139 140 141 // In order to make fetch_unroll_info work properly with escape 142 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 143 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 144 // of previously eliminated objects occurs in realloc_objects, which is 145 // called from the method fetch_unroll_info_helper below. 146 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 147 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 148 // but makes the entry a little slower. There is however a little dance we have to 149 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 150 151 // fetch_unroll_info() is called at the beginning of the deoptimization 152 // handler. Note this fact before we start generating temporary frames 153 // that can confuse an asynchronous stack walker. This counter is 154 // decremented at the end of unpack_frames(). 155 if (TraceDeoptimization) { 156 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 157 } 158 thread->inc_in_deopt_handler(); 159 160 return fetch_unroll_info_helper(thread, exec_mode); 161 JRT_END 162 163 #if COMPILER2_OR_JVMCI 164 static bool eliminate_allocations(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method, 165 frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk) { 166 bool realloc_failures = false; 167 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 168 169 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 170 171 // The flag return_oop() indicates call sites which return oop 172 // in compiled code. Such sites include java method calls, 173 // runtime calls (for example, used to allocate new objects/arrays 174 // on slow code path) and any other calls generated in compiled code. 175 // It is not guaranteed that we can get such information here only 176 // by analyzing bytecode in deoptimized frames. This is why this flag 177 // is set during method compilation (see Compile::Process_OopMap_Node()). 178 // If the previous frame was popped or if we are dispatching an exception, 179 // we don't have an oop result. 180 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt); 181 Handle return_value; 182 if (save_oop_result) { 183 // Reallocation may trigger GC. If deoptimization happened on return from 184 // call which returns oop we need to save it since it is not in oopmap. 185 oop result = deoptee.saved_oop_result(&map); 186 assert(oopDesc::is_oop_or_null(result), "must be oop"); 187 return_value = Handle(thread, result); 188 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 189 if (TraceDeoptimization) { 190 ttyLocker ttyl; 191 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 192 } 193 } 194 if (objects != NULL) { 195 JRT_BLOCK 196 realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD); 197 JRT_END 198 bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci(); 199 Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal); 200 #ifndef PRODUCT 201 if (TraceDeoptimization) { 202 ttyLocker ttyl; 203 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 204 Deoptimization::print_objects(objects, realloc_failures); 205 } 206 #endif 207 } 208 if (save_oop_result) { 209 // Restore result. 210 deoptee.set_saved_oop_result(&map, return_value()); 211 } 212 return realloc_failures; 213 } 214 215 static void eliminate_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 216 #ifndef PRODUCT 217 bool first = true; 218 #endif 219 for (int i = 0; i < chunk->length(); i++) { 220 compiledVFrame* cvf = chunk->at(i); 221 assert (cvf->scope() != NULL,"expect only compiled java frames"); 222 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 223 if (monitors->is_nonempty()) { 224 Deoptimization::relock_objects(monitors, thread, realloc_failures); 225 #ifndef PRODUCT 226 if (PrintDeoptimizationDetails) { 227 ttyLocker ttyl; 228 for (int j = 0; j < monitors->length(); j++) { 229 MonitorInfo* mi = monitors->at(j); 230 if (mi->eliminated()) { 231 if (first) { 232 first = false; 233 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 234 } 235 if (mi->owner_is_scalar_replaced()) { 236 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 237 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 238 } else { 239 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 240 } 241 } 242 } 243 } 244 #endif // !PRODUCT 245 } 246 } 247 } 248 #endif // COMPILER2_OR_JVMCI 249 250 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 251 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 252 253 // Note: there is a safepoint safety issue here. No matter whether we enter 254 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 255 // the vframeArray is created. 256 // 257 258 // Allocate our special deoptimization ResourceMark 259 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 260 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 261 thread->set_deopt_mark(dmark); 262 263 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 264 RegisterMap map(thread, true); 265 RegisterMap dummy_map(thread, false); 266 // Now get the deoptee with a valid map 267 frame deoptee = stub_frame.sender(&map); 268 // Set the deoptee nmethod 269 assert(thread->deopt_compiled_method() == NULL, "Pending deopt!"); 270 CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null(); 271 thread->set_deopt_compiled_method(cm); 272 273 if (VerifyStack) { 274 thread->validate_frame_layout(); 275 } 276 277 // Create a growable array of VFrames where each VFrame represents an inlined 278 // Java frame. This storage is allocated with the usual system arena. 279 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 280 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 281 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 282 while (!vf->is_top()) { 283 assert(vf->is_compiled_frame(), "Wrong frame type"); 284 chunk->push(compiledVFrame::cast(vf)); 285 vf = vf->sender(); 286 } 287 assert(vf->is_compiled_frame(), "Wrong frame type"); 288 chunk->push(compiledVFrame::cast(vf)); 289 290 bool realloc_failures = false; 291 292 #if COMPILER2_OR_JVMCI 293 #if INCLUDE_JVMCI 294 bool jvmci_enabled = true; 295 #else 296 bool jvmci_enabled = false; 297 #endif 298 299 // Reallocate the non-escaping objects and restore their fields. Then 300 // relock objects if synchronization on them was eliminated. 301 if (jvmci_enabled || (DoEscapeAnalysis && EliminateAllocations)) { 302 realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk); 303 } 304 #endif // COMPILER2_OR_JVMCI 305 306 // Revoke biases, done with in java state. 307 // No safepoints allowed after this 308 revoke_from_deopt_handler(thread, deoptee, &map); 309 310 // Ensure that no safepoint is taken after pointers have been stored 311 // in fields of rematerialized objects. If a safepoint occurs from here on 312 // out the java state residing in the vframeArray will be missed. 313 // Locks may be rebaised in a safepoint. 314 NoSafepointVerifier no_safepoint; 315 316 #if COMPILER2_OR_JVMCI 317 if (jvmci_enabled || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks)) { 318 eliminate_locks(thread, chunk, realloc_failures); 319 } 320 #endif // COMPILER2_OR_JVMCI 321 322 ScopeDesc* trap_scope = chunk->at(0)->scope(); 323 Handle exceptionObject; 324 if (trap_scope->rethrow_exception()) { 325 if (PrintDeoptimizationDetails) { 326 tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d", trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci()); 327 } 328 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 329 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 330 ScopeValue* topOfStack = expressions->top(); 331 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 332 guarantee(exceptionObject() != NULL, "exception oop can not be null"); 333 } 334 335 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 336 #if COMPILER2_OR_JVMCI 337 if (realloc_failures) { 338 pop_frames_failed_reallocs(thread, array); 339 } 340 #endif 341 342 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 343 thread->set_vframe_array_head(array); 344 345 // Now that the vframeArray has been created if we have any deferred local writes 346 // added by jvmti then we can free up that structure as the data is now in the 347 // vframeArray 348 349 if (thread->deferred_locals() != NULL) { 350 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 351 int i = 0; 352 do { 353 // Because of inlining we could have multiple vframes for a single frame 354 // and several of the vframes could have deferred writes. Find them all. 355 if (list->at(i)->id() == array->original().id()) { 356 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 357 list->remove_at(i); 358 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 359 delete dlv; 360 } else { 361 i++; 362 } 363 } while ( i < list->length() ); 364 if (list->length() == 0) { 365 thread->set_deferred_locals(NULL); 366 // free the list and elements back to C heap. 367 delete list; 368 } 369 370 } 371 372 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 373 CodeBlob* cb = stub_frame.cb(); 374 // Verify we have the right vframeArray 375 assert(cb->frame_size() >= 0, "Unexpected frame size"); 376 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 377 378 // If the deopt call site is a MethodHandle invoke call site we have 379 // to adjust the unpack_sp. 380 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 381 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 382 unpack_sp = deoptee.unextended_sp(); 383 384 #ifdef ASSERT 385 assert(cb->is_deoptimization_stub() || 386 cb->is_uncommon_trap_stub() || 387 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 388 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 389 "unexpected code blob: %s", cb->name()); 390 #endif 391 392 // This is a guarantee instead of an assert because if vframe doesn't match 393 // we will unpack the wrong deoptimized frame and wind up in strange places 394 // where it will be very difficult to figure out what went wrong. Better 395 // to die an early death here than some very obscure death later when the 396 // trail is cold. 397 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 398 // in that it will fail to detect a problem when there is one. This needs 399 // more work in tiger timeframe. 400 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 401 402 int number_of_frames = array->frames(); 403 404 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 405 // virtual activation, which is the reverse of the elements in the vframes array. 406 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 407 // +1 because we always have an interpreter return address for the final slot. 408 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 409 int popframe_extra_args = 0; 410 // Create an interpreter return address for the stub to use as its return 411 // address so the skeletal frames are perfectly walkable 412 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 413 414 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 415 // activation be put back on the expression stack of the caller for reexecution 416 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 417 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 418 } 419 420 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 421 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 422 // than simply use array->sender.pc(). This requires us to walk the current set of frames 423 // 424 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 425 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 426 427 // It's possible that the number of parameters at the call site is 428 // different than number of arguments in the callee when method 429 // handles are used. If the caller is interpreted get the real 430 // value so that the proper amount of space can be added to it's 431 // frame. 432 bool caller_was_method_handle = false; 433 if (deopt_sender.is_interpreted_frame()) { 434 methodHandle method(thread, deopt_sender.interpreter_frame_method()); 435 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 436 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 437 // Method handle invokes may involve fairly arbitrary chains of 438 // calls so it's impossible to know how much actual space the 439 // caller has for locals. 440 caller_was_method_handle = true; 441 } 442 } 443 444 // 445 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 446 // frame_sizes/frame_pcs[1] next oldest frame (int) 447 // frame_sizes/frame_pcs[n] youngest frame (int) 448 // 449 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 450 // owns the space for the return address to it's caller). Confusing ain't it. 451 // 452 // The vframe array can address vframes with indices running from 453 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 454 // When we create the skeletal frames we need the oldest frame to be in the zero slot 455 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 456 // so things look a little strange in this loop. 457 // 458 int callee_parameters = 0; 459 int callee_locals = 0; 460 for (int index = 0; index < array->frames(); index++ ) { 461 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 462 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 463 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 464 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 465 callee_locals, 466 index == 0, 467 popframe_extra_args); 468 // This pc doesn't have to be perfect just good enough to identify the frame 469 // as interpreted so the skeleton frame will be walkable 470 // The correct pc will be set when the skeleton frame is completely filled out 471 // The final pc we store in the loop is wrong and will be overwritten below 472 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 473 474 callee_parameters = array->element(index)->method()->size_of_parameters(); 475 callee_locals = array->element(index)->method()->max_locals(); 476 popframe_extra_args = 0; 477 } 478 479 // Compute whether the root vframe returns a float or double value. 480 BasicType return_type; 481 { 482 methodHandle method(thread, array->element(0)->method()); 483 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 484 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 485 } 486 487 // Compute information for handling adapters and adjusting the frame size of the caller. 488 int caller_adjustment = 0; 489 490 // Compute the amount the oldest interpreter frame will have to adjust 491 // its caller's stack by. If the caller is a compiled frame then 492 // we pretend that the callee has no parameters so that the 493 // extension counts for the full amount of locals and not just 494 // locals-parms. This is because without a c2i adapter the parm 495 // area as created by the compiled frame will not be usable by 496 // the interpreter. (Depending on the calling convention there 497 // may not even be enough space). 498 499 // QQQ I'd rather see this pushed down into last_frame_adjust 500 // and have it take the sender (aka caller). 501 502 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 503 caller_adjustment = last_frame_adjust(0, callee_locals); 504 } else if (callee_locals > callee_parameters) { 505 // The caller frame may need extending to accommodate 506 // non-parameter locals of the first unpacked interpreted frame. 507 // Compute that adjustment. 508 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 509 } 510 511 // If the sender is deoptimized the we must retrieve the address of the handler 512 // since the frame will "magically" show the original pc before the deopt 513 // and we'd undo the deopt. 514 515 frame_pcs[0] = deopt_sender.raw_pc(); 516 517 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 518 519 #if INCLUDE_JVMCI 520 if (exceptionObject() != NULL) { 521 thread->set_exception_oop(exceptionObject()); 522 exec_mode = Unpack_exception; 523 } 524 #endif 525 526 if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) { 527 assert(thread->has_pending_exception(), "should have thrown OOME"); 528 thread->set_exception_oop(thread->pending_exception()); 529 thread->clear_pending_exception(); 530 exec_mode = Unpack_exception; 531 } 532 533 #if INCLUDE_JVMCI 534 if (thread->frames_to_pop_failed_realloc() > 0) { 535 thread->set_pending_monitorenter(false); 536 } 537 #endif 538 539 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 540 caller_adjustment * BytesPerWord, 541 caller_was_method_handle ? 0 : callee_parameters, 542 number_of_frames, 543 frame_sizes, 544 frame_pcs, 545 return_type, 546 exec_mode); 547 // On some platforms, we need a way to pass some platform dependent 548 // information to the unpacking code so the skeletal frames come out 549 // correct (initial fp value, unextended sp, ...) 550 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 551 552 if (array->frames() > 1) { 553 if (VerifyStack && TraceDeoptimization) { 554 ttyLocker ttyl; 555 tty->print_cr("Deoptimizing method containing inlining"); 556 } 557 } 558 559 array->set_unroll_block(info); 560 return info; 561 } 562 563 // Called to cleanup deoptimization data structures in normal case 564 // after unpacking to stack and when stack overflow error occurs 565 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 566 vframeArray *array) { 567 568 // Get array if coming from exception 569 if (array == NULL) { 570 array = thread->vframe_array_head(); 571 } 572 thread->set_vframe_array_head(NULL); 573 574 // Free the previous UnrollBlock 575 vframeArray* old_array = thread->vframe_array_last(); 576 thread->set_vframe_array_last(array); 577 578 if (old_array != NULL) { 579 UnrollBlock* old_info = old_array->unroll_block(); 580 old_array->set_unroll_block(NULL); 581 delete old_info; 582 delete old_array; 583 } 584 585 // Deallocate any resource creating in this routine and any ResourceObjs allocated 586 // inside the vframeArray (StackValueCollections) 587 588 delete thread->deopt_mark(); 589 thread->set_deopt_mark(NULL); 590 thread->set_deopt_compiled_method(NULL); 591 592 593 if (JvmtiExport::can_pop_frame()) { 594 #ifndef CC_INTERP 595 // Regardless of whether we entered this routine with the pending 596 // popframe condition bit set, we should always clear it now 597 thread->clear_popframe_condition(); 598 #else 599 // C++ interpreter will clear has_pending_popframe when it enters 600 // with method_resume. For deopt_resume2 we clear it now. 601 if (thread->popframe_forcing_deopt_reexecution()) 602 thread->clear_popframe_condition(); 603 #endif /* CC_INTERP */ 604 } 605 606 // unpack_frames() is called at the end of the deoptimization handler 607 // and (in C2) at the end of the uncommon trap handler. Note this fact 608 // so that an asynchronous stack walker can work again. This counter is 609 // incremented at the beginning of fetch_unroll_info() and (in C2) at 610 // the beginning of uncommon_trap(). 611 thread->dec_in_deopt_handler(); 612 } 613 614 // Moved from cpu directories because none of the cpus has callee save values. 615 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 616 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 617 618 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 619 // the days we had adapter frames. When we deoptimize a situation where a 620 // compiled caller calls a compiled caller will have registers it expects 621 // to survive the call to the callee. If we deoptimize the callee the only 622 // way we can restore these registers is to have the oldest interpreter 623 // frame that we create restore these values. That is what this routine 624 // will accomplish. 625 626 // At the moment we have modified c2 to not have any callee save registers 627 // so this problem does not exist and this routine is just a place holder. 628 629 assert(f->is_interpreted_frame(), "must be interpreted"); 630 } 631 632 // Return BasicType of value being returned 633 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 634 635 // We are already active in the special DeoptResourceMark any ResourceObj's we 636 // allocate will be freed at the end of the routine. 637 638 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 639 // but makes the entry a little slower. There is however a little dance we have to 640 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 641 ResetNoHandleMark rnhm; // No-op in release/product versions 642 HandleMark hm; 643 644 frame stub_frame = thread->last_frame(); 645 646 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 647 // must point to the vframeArray for the unpack frame. 648 vframeArray* array = thread->vframe_array_head(); 649 650 #ifndef PRODUCT 651 if (TraceDeoptimization) { 652 ttyLocker ttyl; 653 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 654 p2i(thread), p2i(array), exec_mode); 655 } 656 #endif 657 Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 658 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 659 660 UnrollBlock* info = array->unroll_block(); 661 662 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 663 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 664 665 BasicType bt = info->return_type(); 666 667 // If we have an exception pending, claim that the return type is an oop 668 // so the deopt_blob does not overwrite the exception_oop. 669 670 if (exec_mode == Unpack_exception) 671 bt = T_OBJECT; 672 673 // Cleanup thread deopt data 674 cleanup_deopt_info(thread, array); 675 676 #ifndef PRODUCT 677 if (VerifyStack) { 678 ResourceMark res_mark; 679 // Clear pending exception to not break verification code (restored afterwards) 680 PRESERVE_EXCEPTION_MARK; 681 682 thread->validate_frame_layout(); 683 684 // Verify that the just-unpacked frames match the interpreter's 685 // notions of expression stack and locals 686 vframeArray* cur_array = thread->vframe_array_last(); 687 RegisterMap rm(thread, false); 688 rm.set_include_argument_oops(false); 689 bool is_top_frame = true; 690 int callee_size_of_parameters = 0; 691 int callee_max_locals = 0; 692 for (int i = 0; i < cur_array->frames(); i++) { 693 vframeArrayElement* el = cur_array->element(i); 694 frame* iframe = el->iframe(); 695 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 696 697 // Get the oop map for this bci 698 InterpreterOopMap mask; 699 int cur_invoke_parameter_size = 0; 700 bool try_next_mask = false; 701 int next_mask_expression_stack_size = -1; 702 int top_frame_expression_stack_adjustment = 0; 703 methodHandle mh(thread, iframe->interpreter_frame_method()); 704 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 705 BytecodeStream str(mh, iframe->interpreter_frame_bci()); 706 int max_bci = mh->code_size(); 707 // Get to the next bytecode if possible 708 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 709 // Check to see if we can grab the number of outgoing arguments 710 // at an uncommon trap for an invoke (where the compiler 711 // generates debug info before the invoke has executed) 712 Bytecodes::Code cur_code = str.next(); 713 if (Bytecodes::is_invoke(cur_code)) { 714 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 715 cur_invoke_parameter_size = invoke.size_of_parameters(); 716 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 717 callee_size_of_parameters++; 718 } 719 } 720 if (str.bci() < max_bci) { 721 Bytecodes::Code next_code = str.next(); 722 if (next_code >= 0) { 723 // The interpreter oop map generator reports results before 724 // the current bytecode has executed except in the case of 725 // calls. It seems to be hard to tell whether the compiler 726 // has emitted debug information matching the "state before" 727 // a given bytecode or the state after, so we try both 728 if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) { 729 // Get expression stack size for the next bytecode 730 InterpreterOopMap next_mask; 731 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 732 next_mask_expression_stack_size = next_mask.expression_stack_size(); 733 if (Bytecodes::is_invoke(next_code)) { 734 Bytecode_invoke invoke(mh, str.bci()); 735 next_mask_expression_stack_size += invoke.size_of_parameters(); 736 } 737 // Need to subtract off the size of the result type of 738 // the bytecode because this is not described in the 739 // debug info but returned to the interpreter in the TOS 740 // caching register 741 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 742 if (bytecode_result_type != T_ILLEGAL) { 743 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 744 } 745 assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive"); 746 try_next_mask = true; 747 } 748 } 749 } 750 751 // Verify stack depth and oops in frame 752 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 753 if (!( 754 /* SPARC */ 755 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 756 /* x86 */ 757 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 758 (try_next_mask && 759 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 760 top_frame_expression_stack_adjustment))) || 761 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 762 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 763 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 764 )) { 765 { 766 ttyLocker ttyl; 767 768 // Print out some information that will help us debug the problem 769 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 770 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 771 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 772 iframe->interpreter_frame_expression_stack_size()); 773 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 774 tty->print_cr(" try_next_mask = %d", try_next_mask); 775 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 776 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 777 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 778 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 779 tty->print_cr(" exec_mode = %d", exec_mode); 780 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 781 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 782 tty->print_cr(" Interpreted frames:"); 783 for (int k = 0; k < cur_array->frames(); k++) { 784 vframeArrayElement* el = cur_array->element(k); 785 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 786 } 787 cur_array->print_on_2(tty); 788 } // release tty lock before calling guarantee 789 guarantee(false, "wrong number of expression stack elements during deopt"); 790 } 791 VerifyOopClosure verify; 792 iframe->oops_interpreted_do(&verify, &rm, false); 793 callee_size_of_parameters = mh->size_of_parameters(); 794 callee_max_locals = mh->max_locals(); 795 is_top_frame = false; 796 } 797 } 798 #endif /* !PRODUCT */ 799 800 801 return bt; 802 JRT_END 803 804 class DeoptimizeMarkedClosure : public HandshakeClosure { 805 public: 806 DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {} 807 void do_thread(Thread* thread) { 808 JavaThread* jt = (JavaThread*)thread; 809 jt->deoptimize_marked_methods(); 810 } 811 }; 812 813 void Deoptimization::deoptimize_all_marked() { 814 ResourceMark rm; 815 DeoptimizationMarker dm; 816 817 if (SafepointSynchronize::is_at_safepoint()) { 818 DeoptimizeMarkedClosure deopt; 819 // Make the dependent methods not entrant 820 CodeCache::make_marked_nmethods_not_entrant(); 821 Threads::java_threads_do(&deopt); 822 } else { 823 // Make the dependent methods not entrant 824 { 825 MutexLocker mu(CodeCache_lock, Mutex::_no_safepoint_check_flag); 826 CodeCache::make_marked_nmethods_not_entrant(); 827 } 828 DeoptimizeMarkedClosure deopt; 829 Handshake::execute(&deopt); 830 } 831 } 832 833 Deoptimization::DeoptAction Deoptimization::_unloaded_action 834 = Deoptimization::Action_reinterpret; 835 836 837 838 #if INCLUDE_JVMCI || INCLUDE_AOT 839 template<typename CacheType> 840 class BoxCacheBase : public CHeapObj<mtCompiler> { 841 protected: 842 static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) { 843 ResourceMark rm; 844 char* klass_name_str = klass_name->as_C_string(); 845 Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD); 846 guarantee(k != NULL, "%s must be loaded", klass_name_str); 847 InstanceKlass* ik = InstanceKlass::cast(k); 848 guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str); 849 CacheType::compute_offsets(ik); 850 return ik; 851 } 852 }; 853 854 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache : public BoxCacheBase<CacheType> { 855 PrimitiveType _low; 856 PrimitiveType _high; 857 jobject _cache; 858 protected: 859 static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton; 860 BoxCache(Thread* thread) { 861 InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread); 862 objArrayOop cache = CacheType::cache(ik); 863 assert(cache->length() > 0, "Empty cache"); 864 _low = BoxType::value(cache->obj_at(0)); 865 _high = _low + cache->length() - 1; 866 _cache = JNIHandles::make_global(Handle(thread, cache)); 867 } 868 ~BoxCache() { 869 JNIHandles::destroy_global(_cache); 870 } 871 public: 872 static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) { 873 if (_singleton == NULL) { 874 BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread); 875 if (!Atomic::replace_if_null(&_singleton, s)) { 876 delete s; 877 } 878 } 879 return _singleton; 880 } 881 oop lookup(PrimitiveType value) { 882 if (_low <= value && value <= _high) { 883 int offset = value - _low; 884 return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset); 885 } 886 return NULL; 887 } 888 oop lookup_raw(intptr_t raw_value) { 889 // Have to cast to avoid little/big-endian problems. 890 if (sizeof(PrimitiveType) > sizeof(jint)) { 891 jlong value = (jlong)raw_value; 892 return lookup(value); 893 } 894 PrimitiveType value = (PrimitiveType)*((jint*)&raw_value); 895 return lookup(value); 896 } 897 }; 898 899 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache; 900 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache; 901 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache; 902 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache; 903 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache; 904 905 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL; 906 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL; 907 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL; 908 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL; 909 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL; 910 911 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> { 912 jobject _true_cache; 913 jobject _false_cache; 914 protected: 915 static BooleanBoxCache *_singleton; 916 BooleanBoxCache(Thread *thread) { 917 InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread); 918 _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik))); 919 _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik))); 920 } 921 ~BooleanBoxCache() { 922 JNIHandles::destroy_global(_true_cache); 923 JNIHandles::destroy_global(_false_cache); 924 } 925 public: 926 static BooleanBoxCache* singleton(Thread* thread) { 927 if (_singleton == NULL) { 928 BooleanBoxCache* s = new BooleanBoxCache(thread); 929 if (!Atomic::replace_if_null(&_singleton, s)) { 930 delete s; 931 } 932 } 933 return _singleton; 934 } 935 oop lookup_raw(intptr_t raw_value) { 936 // Have to cast to avoid little/big-endian problems. 937 jboolean value = (jboolean)*((jint*)&raw_value); 938 return lookup(value); 939 } 940 oop lookup(jboolean value) { 941 if (value != 0) { 942 return JNIHandles::resolve_non_null(_true_cache); 943 } 944 return JNIHandles::resolve_non_null(_false_cache); 945 } 946 }; 947 948 BooleanBoxCache* BooleanBoxCache::_singleton = NULL; 949 950 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) { 951 Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()()); 952 BasicType box_type = SystemDictionary::box_klass_type(k); 953 if (box_type != T_OBJECT) { 954 StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0)); 955 switch(box_type) { 956 case T_INT: return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 957 case T_CHAR: return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 958 case T_SHORT: return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 959 case T_BYTE: return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 960 case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 961 case T_LONG: return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int()); 962 default:; 963 } 964 } 965 return NULL; 966 } 967 #endif // INCLUDE_JVMCI || INCLUDE_AOT 968 969 #if COMPILER2_OR_JVMCI 970 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) { 971 Handle pending_exception(THREAD, thread->pending_exception()); 972 const char* exception_file = thread->exception_file(); 973 int exception_line = thread->exception_line(); 974 thread->clear_pending_exception(); 975 976 bool failures = false; 977 978 for (int i = 0; i < objects->length(); i++) { 979 assert(objects->at(i)->is_object(), "invalid debug information"); 980 ObjectValue* sv = (ObjectValue*) objects->at(i); 981 982 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 983 oop obj = NULL; 984 985 if (k->is_instance_klass()) { 986 #if INCLUDE_JVMCI || INCLUDE_AOT 987 CompiledMethod* cm = fr->cb()->as_compiled_method_or_null(); 988 if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) { 989 AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv; 990 obj = get_cached_box(abv, fr, reg_map, THREAD); 991 if (obj != NULL) { 992 // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it. 993 abv->set_cached(true); 994 } 995 } 996 #endif // INCLUDE_JVMCI || INCLUDE_AOT 997 InstanceKlass* ik = InstanceKlass::cast(k); 998 if (obj == NULL) { 999 obj = ik->allocate_instance(THREAD); 1000 } 1001 } else if (k->is_typeArray_klass()) { 1002 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1003 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 1004 int len = sv->field_size() / type2size[ak->element_type()]; 1005 obj = ak->allocate(len, THREAD); 1006 } else if (k->is_objArray_klass()) { 1007 ObjArrayKlass* ak = ObjArrayKlass::cast(k); 1008 obj = ak->allocate(sv->field_size(), THREAD); 1009 } 1010 1011 if (obj == NULL) { 1012 failures = true; 1013 } 1014 1015 assert(sv->value().is_null(), "redundant reallocation"); 1016 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 1017 CLEAR_PENDING_EXCEPTION; 1018 sv->set_value(obj); 1019 } 1020 1021 if (failures) { 1022 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 1023 } else if (pending_exception.not_null()) { 1024 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 1025 } 1026 1027 return failures; 1028 } 1029 1030 // restore elements of an eliminated type array 1031 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 1032 int index = 0; 1033 intptr_t val; 1034 1035 for (int i = 0; i < sv->field_size(); i++) { 1036 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1037 switch(type) { 1038 case T_LONG: case T_DOUBLE: { 1039 assert(value->type() == T_INT, "Agreement."); 1040 StackValue* low = 1041 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1042 #ifdef _LP64 1043 jlong res = (jlong)low->get_int(); 1044 #else 1045 #ifdef SPARC 1046 // For SPARC we have to swap high and low words. 1047 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1048 #else 1049 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1050 #endif //SPARC 1051 #endif 1052 obj->long_at_put(index, res); 1053 break; 1054 } 1055 1056 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1057 case T_INT: case T_FLOAT: { // 4 bytes. 1058 assert(value->type() == T_INT, "Agreement."); 1059 bool big_value = false; 1060 if (i + 1 < sv->field_size() && type == T_INT) { 1061 if (sv->field_at(i)->is_location()) { 1062 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 1063 if (type == Location::dbl || type == Location::lng) { 1064 big_value = true; 1065 } 1066 } else if (sv->field_at(i)->is_constant_int()) { 1067 ScopeValue* next_scope_field = sv->field_at(i + 1); 1068 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1069 big_value = true; 1070 } 1071 } 1072 } 1073 1074 if (big_value) { 1075 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 1076 #ifdef _LP64 1077 jlong res = (jlong)low->get_int(); 1078 #else 1079 #ifdef SPARC 1080 // For SPARC we have to swap high and low words. 1081 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1082 #else 1083 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1084 #endif //SPARC 1085 #endif 1086 obj->int_at_put(index, (jint)*((jint*)&res)); 1087 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 1088 } else { 1089 val = value->get_int(); 1090 obj->int_at_put(index, (jint)*((jint*)&val)); 1091 } 1092 break; 1093 } 1094 1095 case T_SHORT: 1096 assert(value->type() == T_INT, "Agreement."); 1097 val = value->get_int(); 1098 obj->short_at_put(index, (jshort)*((jint*)&val)); 1099 break; 1100 1101 case T_CHAR: 1102 assert(value->type() == T_INT, "Agreement."); 1103 val = value->get_int(); 1104 obj->char_at_put(index, (jchar)*((jint*)&val)); 1105 break; 1106 1107 case T_BYTE: 1108 assert(value->type() == T_INT, "Agreement."); 1109 val = value->get_int(); 1110 obj->byte_at_put(index, (jbyte)*((jint*)&val)); 1111 break; 1112 1113 case T_BOOLEAN: 1114 assert(value->type() == T_INT, "Agreement."); 1115 val = value->get_int(); 1116 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 1117 break; 1118 1119 default: 1120 ShouldNotReachHere(); 1121 } 1122 index++; 1123 } 1124 } 1125 1126 1127 // restore fields of an eliminated object array 1128 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 1129 for (int i = 0; i < sv->field_size(); i++) { 1130 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 1131 assert(value->type() == T_OBJECT, "object element expected"); 1132 obj->obj_at_put(i, value->get_obj()()); 1133 } 1134 } 1135 1136 class ReassignedField { 1137 public: 1138 int _offset; 1139 BasicType _type; 1140 public: 1141 ReassignedField() { 1142 _offset = 0; 1143 _type = T_ILLEGAL; 1144 } 1145 }; 1146 1147 int compare(ReassignedField* left, ReassignedField* right) { 1148 return left->_offset - right->_offset; 1149 } 1150 1151 // Restore fields of an eliminated instance object using the same field order 1152 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 1153 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) { 1154 if (klass->superklass() != NULL) { 1155 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal); 1156 } 1157 1158 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 1159 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 1160 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 1161 ReassignedField field; 1162 field._offset = fs.offset(); 1163 field._type = FieldType::basic_type(fs.signature()); 1164 fields->append(field); 1165 } 1166 } 1167 fields->sort(compare); 1168 for (int i = 0; i < fields->length(); i++) { 1169 intptr_t val; 1170 ScopeValue* scope_field = sv->field_at(svIndex); 1171 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 1172 int offset = fields->at(i)._offset; 1173 BasicType type = fields->at(i)._type; 1174 switch (type) { 1175 case T_OBJECT: case T_ARRAY: 1176 assert(value->type() == T_OBJECT, "Agreement."); 1177 obj->obj_field_put(offset, value->get_obj()()); 1178 break; 1179 1180 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 1181 case T_INT: case T_FLOAT: { // 4 bytes. 1182 assert(value->type() == T_INT, "Agreement."); 1183 bool big_value = false; 1184 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 1185 if (scope_field->is_location()) { 1186 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1187 if (type == Location::dbl || type == Location::lng) { 1188 big_value = true; 1189 } 1190 } 1191 if (scope_field->is_constant_int()) { 1192 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1193 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1194 big_value = true; 1195 } 1196 } 1197 } 1198 1199 if (big_value) { 1200 i++; 1201 assert(i < fields->length(), "second T_INT field needed"); 1202 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1203 } else { 1204 val = value->get_int(); 1205 obj->int_field_put(offset, (jint)*((jint*)&val)); 1206 break; 1207 } 1208 } 1209 /* no break */ 1210 1211 case T_LONG: case T_DOUBLE: { 1212 assert(value->type() == T_INT, "Agreement."); 1213 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1214 #ifdef _LP64 1215 jlong res = (jlong)low->get_int(); 1216 #else 1217 #ifdef SPARC 1218 // For SPARC we have to swap high and low words. 1219 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1220 #else 1221 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1222 #endif //SPARC 1223 #endif 1224 obj->long_field_put(offset, res); 1225 break; 1226 } 1227 1228 case T_SHORT: 1229 assert(value->type() == T_INT, "Agreement."); 1230 val = value->get_int(); 1231 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1232 break; 1233 1234 case T_CHAR: 1235 assert(value->type() == T_INT, "Agreement."); 1236 val = value->get_int(); 1237 obj->char_field_put(offset, (jchar)*((jint*)&val)); 1238 break; 1239 1240 case T_BYTE: 1241 assert(value->type() == T_INT, "Agreement."); 1242 val = value->get_int(); 1243 obj->byte_field_put(offset, (jbyte)*((jint*)&val)); 1244 break; 1245 1246 case T_BOOLEAN: 1247 assert(value->type() == T_INT, "Agreement."); 1248 val = value->get_int(); 1249 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1250 break; 1251 1252 default: 1253 ShouldNotReachHere(); 1254 } 1255 svIndex++; 1256 } 1257 return svIndex; 1258 } 1259 1260 // restore fields of all eliminated objects and arrays 1261 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) { 1262 for (int i = 0; i < objects->length(); i++) { 1263 ObjectValue* sv = (ObjectValue*) objects->at(i); 1264 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1265 Handle obj = sv->value(); 1266 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1267 if (PrintDeoptimizationDetails) { 1268 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1269 } 1270 if (obj.is_null()) { 1271 continue; 1272 } 1273 #if INCLUDE_JVMCI || INCLUDE_AOT 1274 // Don't reassign fields of boxes that came from a cache. Caches may be in CDS. 1275 if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) { 1276 continue; 1277 } 1278 #endif // INCLUDE_JVMCI || INCLUDE_AOT 1279 if (k->is_instance_klass()) { 1280 InstanceKlass* ik = InstanceKlass::cast(k); 1281 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal); 1282 } else if (k->is_typeArray_klass()) { 1283 TypeArrayKlass* ak = TypeArrayKlass::cast(k); 1284 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1285 } else if (k->is_objArray_klass()) { 1286 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1287 } 1288 } 1289 } 1290 1291 1292 // relock objects for which synchronization was eliminated 1293 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1294 for (int i = 0; i < monitors->length(); i++) { 1295 MonitorInfo* mon_info = monitors->at(i); 1296 if (mon_info->eliminated()) { 1297 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1298 if (!mon_info->owner_is_scalar_replaced()) { 1299 Handle obj(thread, mon_info->owner()); 1300 markWord mark = obj->mark(); 1301 if (UseBiasedLocking && mark.has_bias_pattern()) { 1302 // New allocated objects may have the mark set to anonymously biased. 1303 // Also the deoptimized method may called methods with synchronization 1304 // where the thread-local object is bias locked to the current thread. 1305 assert(mark.is_biased_anonymously() || 1306 mark.biased_locker() == thread, "should be locked to current thread"); 1307 // Reset mark word to unbiased prototype. 1308 markWord unbiased_prototype = markWord::prototype().set_age(mark.age()); 1309 obj->set_mark(unbiased_prototype); 1310 } 1311 BasicLock* lock = mon_info->lock(); 1312 ObjectSynchronizer::enter(obj, lock, thread); 1313 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1314 } 1315 } 1316 } 1317 } 1318 1319 1320 #ifndef PRODUCT 1321 // print information about reallocated objects 1322 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1323 fieldDescriptor fd; 1324 1325 for (int i = 0; i < objects->length(); i++) { 1326 ObjectValue* sv = (ObjectValue*) objects->at(i); 1327 Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()()); 1328 Handle obj = sv->value(); 1329 1330 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1331 k->print_value(); 1332 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1333 if (obj.is_null()) { 1334 tty->print(" allocation failed"); 1335 } else { 1336 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1337 } 1338 tty->cr(); 1339 1340 if (Verbose && !obj.is_null()) { 1341 k->oop_print_on(obj(), tty); 1342 } 1343 } 1344 } 1345 #endif 1346 #endif // COMPILER2_OR_JVMCI 1347 1348 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1349 Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1350 1351 #ifndef PRODUCT 1352 if (PrintDeoptimizationDetails) { 1353 ttyLocker ttyl; 1354 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1355 fr.print_on(tty); 1356 tty->print_cr(" Virtual frames (innermost first):"); 1357 for (int index = 0; index < chunk->length(); index++) { 1358 compiledVFrame* vf = chunk->at(index); 1359 tty->print(" %2d - ", index); 1360 vf->print_value(); 1361 int bci = chunk->at(index)->raw_bci(); 1362 const char* code_name; 1363 if (bci == SynchronizationEntryBCI) { 1364 code_name = "sync entry"; 1365 } else { 1366 Bytecodes::Code code = vf->method()->code_at(bci); 1367 code_name = Bytecodes::name(code); 1368 } 1369 tty->print(" - %s", code_name); 1370 tty->print_cr(" @ bci %d ", bci); 1371 if (Verbose) { 1372 vf->print(); 1373 tty->cr(); 1374 } 1375 } 1376 } 1377 #endif 1378 1379 // Register map for next frame (used for stack crawl). We capture 1380 // the state of the deopt'ing frame's caller. Thus if we need to 1381 // stuff a C2I adapter we can properly fill in the callee-save 1382 // register locations. 1383 frame caller = fr.sender(reg_map); 1384 int frame_size = caller.sp() - fr.sp(); 1385 1386 frame sender = caller; 1387 1388 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1389 // the vframeArray containing the unpacking information is allocated in the C heap. 1390 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1391 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1392 1393 // Compare the vframeArray to the collected vframes 1394 assert(array->structural_compare(thread, chunk), "just checking"); 1395 1396 #ifndef PRODUCT 1397 if (PrintDeoptimizationDetails) { 1398 ttyLocker ttyl; 1399 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1400 } 1401 #endif // PRODUCT 1402 1403 return array; 1404 } 1405 1406 #if COMPILER2_OR_JVMCI 1407 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1408 // Reallocation of some scalar replaced objects failed. Record 1409 // that we need to pop all the interpreter frames for the 1410 // deoptimized compiled frame. 1411 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1412 thread->set_frames_to_pop_failed_realloc(array->frames()); 1413 // Unlock all monitors here otherwise the interpreter will see a 1414 // mix of locked and unlocked monitors (because of failed 1415 // reallocations of synchronized objects) and be confused. 1416 for (int i = 0; i < array->frames(); i++) { 1417 MonitorChunk* monitors = array->element(i)->monitors(); 1418 if (monitors != NULL) { 1419 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1420 BasicObjectLock* src = monitors->at(j); 1421 if (src->obj() != NULL) { 1422 ObjectSynchronizer::exit(src->obj(), src->lock(), thread); 1423 } 1424 } 1425 array->element(i)->free_monitors(thread); 1426 #ifdef ASSERT 1427 array->element(i)->set_removed_monitors(); 1428 #endif 1429 } 1430 } 1431 } 1432 #endif 1433 1434 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1435 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1436 Thread* thread = Thread::current(); 1437 for (int i = 0; i < monitors->length(); i++) { 1438 MonitorInfo* mon_info = monitors->at(i); 1439 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1440 objects_to_revoke->append(Handle(thread, mon_info->owner())); 1441 } 1442 } 1443 } 1444 1445 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map) { 1446 // Unfortunately we don't have a RegisterMap available in most of 1447 // the places we want to call this routine so we need to walk the 1448 // stack again to update the register map. 1449 if (map == NULL || !map->update_map()) { 1450 StackFrameStream sfs(thread, true); 1451 bool found = false; 1452 while (!found && !sfs.is_done()) { 1453 frame* cur = sfs.current(); 1454 sfs.next(); 1455 found = cur->id() == fr.id(); 1456 } 1457 assert(found, "frame to be deoptimized not found on target thread's stack"); 1458 map = sfs.register_map(); 1459 } 1460 1461 vframe* vf = vframe::new_vframe(&fr, map, thread); 1462 compiledVFrame* cvf = compiledVFrame::cast(vf); 1463 // Revoke monitors' biases in all scopes 1464 while (!cvf->is_top()) { 1465 collect_monitors(cvf, objects_to_revoke); 1466 cvf = compiledVFrame::cast(cvf->sender()); 1467 } 1468 collect_monitors(cvf, objects_to_revoke); 1469 } 1470 1471 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) { 1472 if (!UseBiasedLocking) { 1473 return; 1474 } 1475 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1476 get_monitors_from_stack(objects_to_revoke, thread, fr, map); 1477 1478 int len = objects_to_revoke->length(); 1479 for (int i = 0; i < len; i++) { 1480 oop obj = (objects_to_revoke->at(i))(); 1481 BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread); 1482 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1483 } 1484 } 1485 1486 1487 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1488 assert(fr.can_be_deoptimized(), "checking frame type"); 1489 1490 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1491 1492 if (LogCompilation && xtty != NULL) { 1493 CompiledMethod* cm = fr.cb()->as_compiled_method_or_null(); 1494 assert(cm != NULL, "only compiled methods can deopt"); 1495 1496 ttyLocker ttyl; 1497 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc())); 1498 cm->log_identity(xtty); 1499 xtty->end_head(); 1500 for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1501 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1502 xtty->method(sd->method()); 1503 xtty->end_elem(); 1504 if (sd->is_top()) break; 1505 } 1506 xtty->tail("deoptimized"); 1507 } 1508 1509 // Patch the compiled method so that when execution returns to it we will 1510 // deopt the execution state and return to the interpreter. 1511 fr.deoptimize(thread); 1512 } 1513 1514 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1515 // Deoptimize only if the frame comes from compile code. 1516 // Do not deoptimize the frame which is already patched 1517 // during the execution of the loops below. 1518 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1519 return; 1520 } 1521 ResourceMark rm; 1522 DeoptimizationMarker dm; 1523 deoptimize_single_frame(thread, fr, reason); 1524 } 1525 1526 #if INCLUDE_JVMCI 1527 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) { 1528 // there is no exception handler for this pc => deoptimize 1529 cm->make_not_entrant(); 1530 1531 // Use Deoptimization::deoptimize for all of its side-effects: 1532 // revoking biases of monitors, gathering traps statistics, logging... 1533 // it also patches the return pc but we do not care about that 1534 // since we return a continuation to the deopt_blob below. 1535 JavaThread* thread = JavaThread::current(); 1536 RegisterMap reg_map(thread, UseBiasedLocking); 1537 frame runtime_frame = thread->last_frame(); 1538 frame caller_frame = runtime_frame.sender(®_map); 1539 assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method"); 1540 Deoptimization::deoptimize(thread, caller_frame, ®_map, Deoptimization::Reason_not_compiled_exception_handler); 1541 1542 MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true); 1543 if (trap_mdo != NULL) { 1544 trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler); 1545 } 1546 1547 return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls(); 1548 } 1549 #endif 1550 1551 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1552 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1553 "can only deoptimize other thread at a safepoint"); 1554 // Compute frame and register map based on thread and sp. 1555 RegisterMap reg_map(thread, UseBiasedLocking); 1556 frame fr = thread->last_frame(); 1557 while (fr.id() != id) { 1558 fr = fr.sender(®_map); 1559 } 1560 deoptimize(thread, fr, ®_map, reason); 1561 } 1562 1563 1564 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1565 if (thread == Thread::current()) { 1566 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1567 } else { 1568 VM_DeoptimizeFrame deopt(thread, id, reason); 1569 VMThread::execute(&deopt); 1570 } 1571 } 1572 1573 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1574 deoptimize_frame(thread, id, Reason_constraint); 1575 } 1576 1577 // JVMTI PopFrame support 1578 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1579 { 1580 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1581 } 1582 JRT_END 1583 1584 MethodData* 1585 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m, 1586 bool create_if_missing) { 1587 Thread* THREAD = thread; 1588 MethodData* mdo = m()->method_data(); 1589 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1590 // Build an MDO. Ignore errors like OutOfMemory; 1591 // that simply means we won't have an MDO to update. 1592 Method::build_interpreter_method_data(m, THREAD); 1593 if (HAS_PENDING_EXCEPTION) { 1594 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1595 CLEAR_PENDING_EXCEPTION; 1596 } 1597 mdo = m()->method_data(); 1598 } 1599 return mdo; 1600 } 1601 1602 #if COMPILER2_OR_JVMCI 1603 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1604 // in case of an unresolved klass entry, load the class. 1605 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1606 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1607 return; 1608 } 1609 1610 if (!constant_pool->tag_at(index).is_symbol()) return; 1611 1612 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1613 Symbol* symbol = constant_pool->symbol_at(index); 1614 1615 // class name? 1616 if (symbol->char_at(0) != '(') { 1617 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1618 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1619 return; 1620 } 1621 1622 // then it must be a signature! 1623 ResourceMark rm(THREAD); 1624 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1625 if (ss.is_object()) { 1626 Symbol* class_name = ss.as_symbol(); 1627 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1628 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1629 } 1630 } 1631 } 1632 1633 1634 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1635 EXCEPTION_MARK; 1636 load_class_by_index(constant_pool, index, THREAD); 1637 if (HAS_PENDING_EXCEPTION) { 1638 // Exception happened during classloading. We ignore the exception here, since it 1639 // is going to be rethrown since the current activation is going to be deoptimized and 1640 // the interpreter will re-execute the bytecode. 1641 CLEAR_PENDING_EXCEPTION; 1642 // Class loading called java code which may have caused a stack 1643 // overflow. If the exception was thrown right before the return 1644 // to the runtime the stack is no longer guarded. Reguard the 1645 // stack otherwise if we return to the uncommon trap blob and the 1646 // stack bang causes a stack overflow we crash. 1647 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1648 JavaThread* thread = (JavaThread*)THREAD; 1649 bool guard_pages_enabled = thread->stack_guards_enabled(); 1650 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1651 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1652 } 1653 } 1654 1655 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1656 HandleMark hm; 1657 1658 // uncommon_trap() is called at the beginning of the uncommon trap 1659 // handler. Note this fact before we start generating temporary frames 1660 // that can confuse an asynchronous stack walker. This counter is 1661 // decremented at the end of unpack_frames(). 1662 thread->inc_in_deopt_handler(); 1663 1664 // We need to update the map if we have biased locking. 1665 #if INCLUDE_JVMCI 1666 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1667 RegisterMap reg_map(thread, true); 1668 #else 1669 RegisterMap reg_map(thread, UseBiasedLocking); 1670 #endif 1671 frame stub_frame = thread->last_frame(); 1672 frame fr = stub_frame.sender(®_map); 1673 // Make sure the calling nmethod is not getting deoptimized and removed 1674 // before we are done with it. 1675 nmethodLocker nl(fr.pc()); 1676 1677 // Log a message 1678 Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1679 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1680 1681 { 1682 ResourceMark rm; 1683 1684 DeoptReason reason = trap_request_reason(trap_request); 1685 DeoptAction action = trap_request_action(trap_request); 1686 #if INCLUDE_JVMCI 1687 int debug_id = trap_request_debug_id(trap_request); 1688 #endif 1689 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1690 1691 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1692 compiledVFrame* cvf = compiledVFrame::cast(vf); 1693 1694 CompiledMethod* nm = cvf->code(); 1695 1696 ScopeDesc* trap_scope = cvf->scope(); 1697 1698 if (TraceDeoptimization) { 1699 ttyLocker ttyl; 1700 tty->print_cr(" bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string() 1701 #if INCLUDE_JVMCI 1702 , debug_id 1703 #endif 1704 ); 1705 } 1706 1707 methodHandle trap_method(THREAD, trap_scope->method()); 1708 int trap_bci = trap_scope->bci(); 1709 #if INCLUDE_JVMCI 1710 jlong speculation = thread->pending_failed_speculation(); 1711 if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods 1712 nm->as_nmethod()->update_speculation(thread); 1713 } else { 1714 assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers"); 1715 } 1716 1717 if (trap_bci == SynchronizationEntryBCI) { 1718 trap_bci = 0; 1719 thread->set_pending_monitorenter(true); 1720 } 1721 1722 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1723 thread->set_pending_transfer_to_interpreter(true); 1724 } 1725 #endif 1726 1727 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1728 // Record this event in the histogram. 1729 gather_statistics(reason, action, trap_bc); 1730 1731 // Ensure that we can record deopt. history: 1732 // Need MDO to record RTM code generation state. 1733 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1734 1735 methodHandle profiled_method; 1736 #if INCLUDE_JVMCI 1737 if (nm->is_compiled_by_jvmci()) { 1738 profiled_method = methodHandle(THREAD, nm->method()); 1739 } else { 1740 profiled_method = trap_method; 1741 } 1742 #else 1743 profiled_method = trap_method; 1744 #endif 1745 1746 MethodData* trap_mdo = 1747 get_method_data(thread, profiled_method, create_if_missing); 1748 1749 // Log a message 1750 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s", 1751 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1752 trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name()); 1753 1754 // Print a bunch of diagnostics, if requested. 1755 if (TraceDeoptimization || LogCompilation) { 1756 ResourceMark rm; 1757 ttyLocker ttyl; 1758 char buf[100]; 1759 if (xtty != NULL) { 1760 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1761 os::current_thread_id(), 1762 format_trap_request(buf, sizeof(buf), trap_request)); 1763 #if INCLUDE_JVMCI 1764 if (speculation != 0) { 1765 xtty->print(" speculation='" JLONG_FORMAT "'", speculation); 1766 } 1767 #endif 1768 nm->log_identity(xtty); 1769 } 1770 Symbol* class_name = NULL; 1771 bool unresolved = false; 1772 if (unloaded_class_index >= 0) { 1773 constantPoolHandle constants (THREAD, trap_method->constants()); 1774 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1775 class_name = constants->klass_name_at(unloaded_class_index); 1776 unresolved = true; 1777 if (xtty != NULL) 1778 xtty->print(" unresolved='1'"); 1779 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1780 class_name = constants->symbol_at(unloaded_class_index); 1781 } 1782 if (xtty != NULL) 1783 xtty->name(class_name); 1784 } 1785 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1786 // Dump the relevant MDO state. 1787 // This is the deopt count for the current reason, any previous 1788 // reasons or recompiles seen at this point. 1789 int dcnt = trap_mdo->trap_count(reason); 1790 if (dcnt != 0) 1791 xtty->print(" count='%d'", dcnt); 1792 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1793 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1794 if (dos != 0) { 1795 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1796 if (trap_state_is_recompiled(dos)) { 1797 int recnt2 = trap_mdo->overflow_recompile_count(); 1798 if (recnt2 != 0) 1799 xtty->print(" recompiles2='%d'", recnt2); 1800 } 1801 } 1802 } 1803 if (xtty != NULL) { 1804 xtty->stamp(); 1805 xtty->end_head(); 1806 } 1807 if (TraceDeoptimization) { // make noise on the tty 1808 tty->print("Uncommon trap occurred in"); 1809 nm->method()->print_short_name(tty); 1810 tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id()); 1811 #if INCLUDE_JVMCI 1812 if (nm->is_nmethod()) { 1813 const char* installed_code_name = nm->as_nmethod()->jvmci_name(); 1814 if (installed_code_name != NULL) { 1815 tty->print(" (JVMCI: installed code name=%s) ", installed_code_name); 1816 } 1817 } 1818 #endif 1819 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1820 p2i(fr.pc()), 1821 os::current_thread_id(), 1822 trap_reason_name(reason), 1823 trap_action_name(action), 1824 unloaded_class_index 1825 #if INCLUDE_JVMCI 1826 , debug_id 1827 #endif 1828 ); 1829 if (class_name != NULL) { 1830 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1831 class_name->print_symbol_on(tty); 1832 } 1833 tty->cr(); 1834 } 1835 if (xtty != NULL) { 1836 // Log the precise location of the trap. 1837 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1838 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1839 xtty->method(sd->method()); 1840 xtty->end_elem(); 1841 if (sd->is_top()) break; 1842 } 1843 xtty->tail("uncommon_trap"); 1844 } 1845 } 1846 // (End diagnostic printout.) 1847 1848 // Load class if necessary 1849 if (unloaded_class_index >= 0) { 1850 constantPoolHandle constants(THREAD, trap_method->constants()); 1851 load_class_by_index(constants, unloaded_class_index); 1852 } 1853 1854 // Flush the nmethod if necessary and desirable. 1855 // 1856 // We need to avoid situations where we are re-flushing the nmethod 1857 // because of a hot deoptimization site. Repeated flushes at the same 1858 // point need to be detected by the compiler and avoided. If the compiler 1859 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1860 // module must take measures to avoid an infinite cycle of recompilation 1861 // and deoptimization. There are several such measures: 1862 // 1863 // 1. If a recompilation is ordered a second time at some site X 1864 // and for the same reason R, the action is adjusted to 'reinterpret', 1865 // to give the interpreter time to exercise the method more thoroughly. 1866 // If this happens, the method's overflow_recompile_count is incremented. 1867 // 1868 // 2. If the compiler fails to reduce the deoptimization rate, then 1869 // the method's overflow_recompile_count will begin to exceed the set 1870 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1871 // is adjusted to 'make_not_compilable', and the method is abandoned 1872 // to the interpreter. This is a performance hit for hot methods, 1873 // but is better than a disastrous infinite cycle of recompilations. 1874 // (Actually, only the method containing the site X is abandoned.) 1875 // 1876 // 3. In parallel with the previous measures, if the total number of 1877 // recompilations of a method exceeds the much larger set limit 1878 // PerMethodRecompilationCutoff, the method is abandoned. 1879 // This should only happen if the method is very large and has 1880 // many "lukewarm" deoptimizations. The code which enforces this 1881 // limit is elsewhere (class nmethod, class Method). 1882 // 1883 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1884 // to recompile at each bytecode independently of the per-BCI cutoff. 1885 // 1886 // The decision to update code is up to the compiler, and is encoded 1887 // in the Action_xxx code. If the compiler requests Action_none 1888 // no trap state is changed, no compiled code is changed, and the 1889 // computation suffers along in the interpreter. 1890 // 1891 // The other action codes specify various tactics for decompilation 1892 // and recompilation. Action_maybe_recompile is the loosest, and 1893 // allows the compiled code to stay around until enough traps are seen, 1894 // and until the compiler gets around to recompiling the trapping method. 1895 // 1896 // The other actions cause immediate removal of the present code. 1897 1898 // Traps caused by injected profile shouldn't pollute trap counts. 1899 bool injected_profile_trap = trap_method->has_injected_profile() && 1900 (reason == Reason_intrinsic || reason == Reason_unreached); 1901 1902 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1903 bool make_not_entrant = false; 1904 bool make_not_compilable = false; 1905 bool reprofile = false; 1906 switch (action) { 1907 case Action_none: 1908 // Keep the old code. 1909 update_trap_state = false; 1910 break; 1911 case Action_maybe_recompile: 1912 // Do not need to invalidate the present code, but we can 1913 // initiate another 1914 // Start compiler without (necessarily) invalidating the nmethod. 1915 // The system will tolerate the old code, but new code should be 1916 // generated when possible. 1917 break; 1918 case Action_reinterpret: 1919 // Go back into the interpreter for a while, and then consider 1920 // recompiling form scratch. 1921 make_not_entrant = true; 1922 // Reset invocation counter for outer most method. 1923 // This will allow the interpreter to exercise the bytecodes 1924 // for a while before recompiling. 1925 // By contrast, Action_make_not_entrant is immediate. 1926 // 1927 // Note that the compiler will track null_check, null_assert, 1928 // range_check, and class_check events and log them as if they 1929 // had been traps taken from compiled code. This will update 1930 // the MDO trap history so that the next compilation will 1931 // properly detect hot trap sites. 1932 reprofile = true; 1933 break; 1934 case Action_make_not_entrant: 1935 // Request immediate recompilation, and get rid of the old code. 1936 // Make them not entrant, so next time they are called they get 1937 // recompiled. Unloaded classes are loaded now so recompile before next 1938 // time they are called. Same for uninitialized. The interpreter will 1939 // link the missing class, if any. 1940 make_not_entrant = true; 1941 break; 1942 case Action_make_not_compilable: 1943 // Give up on compiling this method at all. 1944 make_not_entrant = true; 1945 make_not_compilable = true; 1946 break; 1947 default: 1948 ShouldNotReachHere(); 1949 } 1950 1951 // Setting +ProfileTraps fixes the following, on all platforms: 1952 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1953 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1954 // recompile relies on a MethodData* to record heroic opt failures. 1955 1956 // Whether the interpreter is producing MDO data or not, we also need 1957 // to use the MDO to detect hot deoptimization points and control 1958 // aggressive optimization. 1959 bool inc_recompile_count = false; 1960 ProfileData* pdata = NULL; 1961 if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) { 1962 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1963 uint this_trap_count = 0; 1964 bool maybe_prior_trap = false; 1965 bool maybe_prior_recompile = false; 1966 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1967 #if INCLUDE_JVMCI 1968 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1969 #endif 1970 nm->method(), 1971 //outputs: 1972 this_trap_count, 1973 maybe_prior_trap, 1974 maybe_prior_recompile); 1975 // Because the interpreter also counts null, div0, range, and class 1976 // checks, these traps from compiled code are double-counted. 1977 // This is harmless; it just means that the PerXTrapLimit values 1978 // are in effect a little smaller than they look. 1979 1980 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1981 if (per_bc_reason != Reason_none) { 1982 // Now take action based on the partially known per-BCI history. 1983 if (maybe_prior_trap 1984 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1985 // If there are too many traps at this BCI, force a recompile. 1986 // This will allow the compiler to see the limit overflow, and 1987 // take corrective action, if possible. The compiler generally 1988 // does not use the exact PerBytecodeTrapLimit value, but instead 1989 // changes its tactics if it sees any traps at all. This provides 1990 // a little hysteresis, delaying a recompile until a trap happens 1991 // several times. 1992 // 1993 // Actually, since there is only one bit of counter per BCI, 1994 // the possible per-BCI counts are {0,1,(per-method count)}. 1995 // This produces accurate results if in fact there is only 1996 // one hot trap site, but begins to get fuzzy if there are 1997 // many sites. For example, if there are ten sites each 1998 // trapping two or more times, they each get the blame for 1999 // all of their traps. 2000 make_not_entrant = true; 2001 } 2002 2003 // Detect repeated recompilation at the same BCI, and enforce a limit. 2004 if (make_not_entrant && maybe_prior_recompile) { 2005 // More than one recompile at this point. 2006 inc_recompile_count = maybe_prior_trap; 2007 } 2008 } else { 2009 // For reasons which are not recorded per-bytecode, we simply 2010 // force recompiles unconditionally. 2011 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 2012 make_not_entrant = true; 2013 } 2014 2015 // Go back to the compiler if there are too many traps in this method. 2016 if (this_trap_count >= per_method_trap_limit(reason)) { 2017 // If there are too many traps in this method, force a recompile. 2018 // This will allow the compiler to see the limit overflow, and 2019 // take corrective action, if possible. 2020 // (This condition is an unlikely backstop only, because the 2021 // PerBytecodeTrapLimit is more likely to take effect first, 2022 // if it is applicable.) 2023 make_not_entrant = true; 2024 } 2025 2026 // Here's more hysteresis: If there has been a recompile at 2027 // this trap point already, run the method in the interpreter 2028 // for a while to exercise it more thoroughly. 2029 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 2030 reprofile = true; 2031 } 2032 } 2033 2034 // Take requested actions on the method: 2035 2036 // Recompile 2037 if (make_not_entrant) { 2038 if (!nm->make_not_entrant()) { 2039 return; // the call did not change nmethod's state 2040 } 2041 2042 if (pdata != NULL) { 2043 // Record the recompilation event, if any. 2044 int tstate0 = pdata->trap_state(); 2045 int tstate1 = trap_state_set_recompiled(tstate0, true); 2046 if (tstate1 != tstate0) 2047 pdata->set_trap_state(tstate1); 2048 } 2049 2050 #if INCLUDE_RTM_OPT 2051 // Restart collecting RTM locking abort statistic if the method 2052 // is recompiled for a reason other than RTM state change. 2053 // Assume that in new recompiled code the statistic could be different, 2054 // for example, due to different inlining. 2055 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 2056 UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) { 2057 trap_mdo->atomic_set_rtm_state(ProfileRTM); 2058 } 2059 #endif 2060 // For code aging we count traps separately here, using make_not_entrant() 2061 // as a guard against simultaneous deopts in multiple threads. 2062 if (reason == Reason_tenured && trap_mdo != NULL) { 2063 trap_mdo->inc_tenure_traps(); 2064 } 2065 } 2066 2067 if (inc_recompile_count) { 2068 trap_mdo->inc_overflow_recompile_count(); 2069 if ((uint)trap_mdo->overflow_recompile_count() > 2070 (uint)PerBytecodeRecompilationCutoff) { 2071 // Give up on the method containing the bad BCI. 2072 if (trap_method() == nm->method()) { 2073 make_not_compilable = true; 2074 } else { 2075 trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization); 2076 // But give grace to the enclosing nm->method(). 2077 } 2078 } 2079 } 2080 2081 // Reprofile 2082 if (reprofile) { 2083 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 2084 } 2085 2086 // Give up compiling 2087 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 2088 assert(make_not_entrant, "consistent"); 2089 nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization); 2090 } 2091 2092 } // Free marked resources 2093 2094 } 2095 JRT_END 2096 2097 ProfileData* 2098 Deoptimization::query_update_method_data(MethodData* trap_mdo, 2099 int trap_bci, 2100 Deoptimization::DeoptReason reason, 2101 bool update_total_trap_count, 2102 #if INCLUDE_JVMCI 2103 bool is_osr, 2104 #endif 2105 Method* compiled_method, 2106 //outputs: 2107 uint& ret_this_trap_count, 2108 bool& ret_maybe_prior_trap, 2109 bool& ret_maybe_prior_recompile) { 2110 bool maybe_prior_trap = false; 2111 bool maybe_prior_recompile = false; 2112 uint this_trap_count = 0; 2113 if (update_total_trap_count) { 2114 uint idx = reason; 2115 #if INCLUDE_JVMCI 2116 if (is_osr) { 2117 idx += Reason_LIMIT; 2118 } 2119 #endif 2120 uint prior_trap_count = trap_mdo->trap_count(idx); 2121 this_trap_count = trap_mdo->inc_trap_count(idx); 2122 2123 // If the runtime cannot find a place to store trap history, 2124 // it is estimated based on the general condition of the method. 2125 // If the method has ever been recompiled, or has ever incurred 2126 // a trap with the present reason , then this BCI is assumed 2127 // (pessimistically) to be the culprit. 2128 maybe_prior_trap = (prior_trap_count != 0); 2129 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 2130 } 2131 ProfileData* pdata = NULL; 2132 2133 2134 // For reasons which are recorded per bytecode, we check per-BCI data. 2135 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 2136 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 2137 if (per_bc_reason != Reason_none) { 2138 // Find the profile data for this BCI. If there isn't one, 2139 // try to allocate one from the MDO's set of spares. 2140 // This will let us detect a repeated trap at this point. 2141 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 2142 2143 if (pdata != NULL) { 2144 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 2145 if (LogCompilation && xtty != NULL) { 2146 ttyLocker ttyl; 2147 // no more room for speculative traps in this MDO 2148 xtty->elem("speculative_traps_oom"); 2149 } 2150 } 2151 // Query the trap state of this profile datum. 2152 int tstate0 = pdata->trap_state(); 2153 if (!trap_state_has_reason(tstate0, per_bc_reason)) 2154 maybe_prior_trap = false; 2155 if (!trap_state_is_recompiled(tstate0)) 2156 maybe_prior_recompile = false; 2157 2158 // Update the trap state of this profile datum. 2159 int tstate1 = tstate0; 2160 // Record the reason. 2161 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 2162 // Store the updated state on the MDO, for next time. 2163 if (tstate1 != tstate0) 2164 pdata->set_trap_state(tstate1); 2165 } else { 2166 if (LogCompilation && xtty != NULL) { 2167 ttyLocker ttyl; 2168 // Missing MDP? Leave a small complaint in the log. 2169 xtty->elem("missing_mdp bci='%d'", trap_bci); 2170 } 2171 } 2172 } 2173 2174 // Return results: 2175 ret_this_trap_count = this_trap_count; 2176 ret_maybe_prior_trap = maybe_prior_trap; 2177 ret_maybe_prior_recompile = maybe_prior_recompile; 2178 return pdata; 2179 } 2180 2181 void 2182 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2183 ResourceMark rm; 2184 // Ignored outputs: 2185 uint ignore_this_trap_count; 2186 bool ignore_maybe_prior_trap; 2187 bool ignore_maybe_prior_recompile; 2188 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2189 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2190 bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler); 2191 query_update_method_data(trap_mdo, trap_bci, 2192 (DeoptReason)reason, 2193 update_total_counts, 2194 #if INCLUDE_JVMCI 2195 false, 2196 #endif 2197 NULL, 2198 ignore_this_trap_count, 2199 ignore_maybe_prior_trap, 2200 ignore_maybe_prior_recompile); 2201 } 2202 2203 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2204 if (TraceDeoptimization) { 2205 tty->print("Uncommon trap "); 2206 } 2207 // Still in Java no safepoints 2208 { 2209 // This enters VM and may safepoint 2210 uncommon_trap_inner(thread, trap_request); 2211 } 2212 return fetch_unroll_info_helper(thread, exec_mode); 2213 } 2214 2215 // Local derived constants. 2216 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2217 const int DS_REASON_MASK = ((uint)DataLayout::trap_mask) >> 1; 2218 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2219 2220 //---------------------------trap_state_reason--------------------------------- 2221 Deoptimization::DeoptReason 2222 Deoptimization::trap_state_reason(int trap_state) { 2223 // This assert provides the link between the width of DataLayout::trap_bits 2224 // and the encoding of "recorded" reasons. It ensures there are enough 2225 // bits to store all needed reasons in the per-BCI MDO profile. 2226 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2227 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2228 trap_state -= recompile_bit; 2229 if (trap_state == DS_REASON_MASK) { 2230 return Reason_many; 2231 } else { 2232 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2233 return (DeoptReason)trap_state; 2234 } 2235 } 2236 //-------------------------trap_state_has_reason------------------------------- 2237 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2238 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2239 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2240 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2241 trap_state -= recompile_bit; 2242 if (trap_state == DS_REASON_MASK) { 2243 return -1; // true, unspecifically (bottom of state lattice) 2244 } else if (trap_state == reason) { 2245 return 1; // true, definitely 2246 } else if (trap_state == 0) { 2247 return 0; // false, definitely (top of state lattice) 2248 } else { 2249 return 0; // false, definitely 2250 } 2251 } 2252 //-------------------------trap_state_add_reason------------------------------- 2253 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2254 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2255 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2256 trap_state -= recompile_bit; 2257 if (trap_state == DS_REASON_MASK) { 2258 return trap_state + recompile_bit; // already at state lattice bottom 2259 } else if (trap_state == reason) { 2260 return trap_state + recompile_bit; // the condition is already true 2261 } else if (trap_state == 0) { 2262 return reason + recompile_bit; // no condition has yet been true 2263 } else { 2264 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2265 } 2266 } 2267 //-----------------------trap_state_is_recompiled------------------------------ 2268 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2269 return (trap_state & DS_RECOMPILE_BIT) != 0; 2270 } 2271 //-----------------------trap_state_set_recompiled----------------------------- 2272 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2273 if (z) return trap_state | DS_RECOMPILE_BIT; 2274 else return trap_state & ~DS_RECOMPILE_BIT; 2275 } 2276 //---------------------------format_trap_state--------------------------------- 2277 // This is used for debugging and diagnostics, including LogFile output. 2278 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2279 int trap_state) { 2280 assert(buflen > 0, "sanity"); 2281 DeoptReason reason = trap_state_reason(trap_state); 2282 bool recomp_flag = trap_state_is_recompiled(trap_state); 2283 // Re-encode the state from its decoded components. 2284 int decoded_state = 0; 2285 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2286 decoded_state = trap_state_add_reason(decoded_state, reason); 2287 if (recomp_flag) 2288 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2289 // If the state re-encodes properly, format it symbolically. 2290 // Because this routine is used for debugging and diagnostics, 2291 // be robust even if the state is a strange value. 2292 size_t len; 2293 if (decoded_state != trap_state) { 2294 // Random buggy state that doesn't decode?? 2295 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2296 } else { 2297 len = jio_snprintf(buf, buflen, "%s%s", 2298 trap_reason_name(reason), 2299 recomp_flag ? " recompiled" : ""); 2300 } 2301 return buf; 2302 } 2303 2304 2305 //--------------------------------statics-------------------------------------- 2306 const char* Deoptimization::_trap_reason_name[] = { 2307 // Note: Keep this in sync. with enum DeoptReason. 2308 "none", 2309 "null_check", 2310 "null_assert" JVMCI_ONLY("_or_unreached0"), 2311 "range_check", 2312 "class_check", 2313 "array_check", 2314 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2315 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2316 "profile_predicate", 2317 "unloaded", 2318 "uninitialized", 2319 "initialized", 2320 "unreached", 2321 "unhandled", 2322 "constraint", 2323 "div0_check", 2324 "age", 2325 "predicate", 2326 "loop_limit_check", 2327 "speculate_class_check", 2328 "speculate_null_check", 2329 "speculate_null_assert", 2330 "rtm_state_change", 2331 "unstable_if", 2332 "unstable_fused_if", 2333 #if INCLUDE_JVMCI 2334 "aliasing", 2335 "transfer_to_interpreter", 2336 "not_compiled_exception_handler", 2337 "unresolved", 2338 "jsr_mismatch", 2339 #endif 2340 "tenured" 2341 }; 2342 const char* Deoptimization::_trap_action_name[] = { 2343 // Note: Keep this in sync. with enum DeoptAction. 2344 "none", 2345 "maybe_recompile", 2346 "reinterpret", 2347 "make_not_entrant", 2348 "make_not_compilable" 2349 }; 2350 2351 const char* Deoptimization::trap_reason_name(int reason) { 2352 // Check that every reason has a name 2353 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2354 2355 if (reason == Reason_many) return "many"; 2356 if ((uint)reason < Reason_LIMIT) 2357 return _trap_reason_name[reason]; 2358 static char buf[20]; 2359 sprintf(buf, "reason%d", reason); 2360 return buf; 2361 } 2362 const char* Deoptimization::trap_action_name(int action) { 2363 // Check that every action has a name 2364 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2365 2366 if ((uint)action < Action_LIMIT) 2367 return _trap_action_name[action]; 2368 static char buf[20]; 2369 sprintf(buf, "action%d", action); 2370 return buf; 2371 } 2372 2373 // This is used for debugging and diagnostics, including LogFile output. 2374 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2375 int trap_request) { 2376 jint unloaded_class_index = trap_request_index(trap_request); 2377 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2378 const char* action = trap_action_name(trap_request_action(trap_request)); 2379 #if INCLUDE_JVMCI 2380 int debug_id = trap_request_debug_id(trap_request); 2381 #endif 2382 size_t len; 2383 if (unloaded_class_index < 0) { 2384 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2385 reason, action 2386 #if INCLUDE_JVMCI 2387 ,debug_id 2388 #endif 2389 ); 2390 } else { 2391 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2392 reason, action, unloaded_class_index 2393 #if INCLUDE_JVMCI 2394 ,debug_id 2395 #endif 2396 ); 2397 } 2398 return buf; 2399 } 2400 2401 juint Deoptimization::_deoptimization_hist 2402 [Deoptimization::Reason_LIMIT] 2403 [1 + Deoptimization::Action_LIMIT] 2404 [Deoptimization::BC_CASE_LIMIT] 2405 = {0}; 2406 2407 enum { 2408 LSB_BITS = 8, 2409 LSB_MASK = right_n_bits(LSB_BITS) 2410 }; 2411 2412 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2413 Bytecodes::Code bc) { 2414 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2415 assert(action >= 0 && action < Action_LIMIT, "oob"); 2416 _deoptimization_hist[Reason_none][0][0] += 1; // total 2417 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2418 juint* cases = _deoptimization_hist[reason][1+action]; 2419 juint* bc_counter_addr = NULL; 2420 juint bc_counter = 0; 2421 // Look for an unused counter, or an exact match to this BC. 2422 if (bc != Bytecodes::_illegal) { 2423 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2424 juint* counter_addr = &cases[bc_case]; 2425 juint counter = *counter_addr; 2426 if ((counter == 0 && bc_counter_addr == NULL) 2427 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2428 // this counter is either free or is already devoted to this BC 2429 bc_counter_addr = counter_addr; 2430 bc_counter = counter | bc; 2431 } 2432 } 2433 } 2434 if (bc_counter_addr == NULL) { 2435 // Overflow, or no given bytecode. 2436 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2437 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2438 } 2439 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2440 } 2441 2442 jint Deoptimization::total_deoptimization_count() { 2443 return _deoptimization_hist[Reason_none][0][0]; 2444 } 2445 2446 void Deoptimization::print_statistics() { 2447 juint total = total_deoptimization_count(); 2448 juint account = total; 2449 if (total != 0) { 2450 ttyLocker ttyl; 2451 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2452 tty->print_cr("Deoptimization traps recorded:"); 2453 #define PRINT_STAT_LINE(name, r) \ 2454 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2455 PRINT_STAT_LINE("total", total); 2456 // For each non-zero entry in the histogram, print the reason, 2457 // the action, and (if specifically known) the type of bytecode. 2458 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2459 for (int action = 0; action < Action_LIMIT; action++) { 2460 juint* cases = _deoptimization_hist[reason][1+action]; 2461 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2462 juint counter = cases[bc_case]; 2463 if (counter != 0) { 2464 char name[1*K]; 2465 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2466 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2467 bc = Bytecodes::_illegal; 2468 sprintf(name, "%s/%s/%s", 2469 trap_reason_name(reason), 2470 trap_action_name(action), 2471 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2472 juint r = counter >> LSB_BITS; 2473 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2474 account -= r; 2475 } 2476 } 2477 } 2478 } 2479 if (account != 0) { 2480 PRINT_STAT_LINE("unaccounted", account); 2481 } 2482 #undef PRINT_STAT_LINE 2483 if (xtty != NULL) xtty->tail("statistics"); 2484 } 2485 } 2486 #else // COMPILER2_OR_JVMCI 2487 2488 2489 // Stubs for C1 only system. 2490 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2491 return false; 2492 } 2493 2494 const char* Deoptimization::trap_reason_name(int reason) { 2495 return "unknown"; 2496 } 2497 2498 void Deoptimization::print_statistics() { 2499 // no output 2500 } 2501 2502 void 2503 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2504 // no udpate 2505 } 2506 2507 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2508 return 0; 2509 } 2510 2511 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2512 Bytecodes::Code bc) { 2513 // no update 2514 } 2515 2516 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2517 int trap_state) { 2518 jio_snprintf(buf, buflen, "#%d", trap_state); 2519 return buf; 2520 } 2521 2522 #endif // COMPILER2_OR_JVMCI