1 /* 2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "code/codeCache.hpp" 28 #include "code/debugInfoRec.hpp" 29 #include "code/nmethod.hpp" 30 #include "code/pcDesc.hpp" 31 #include "code/scopeDesc.hpp" 32 #include "interpreter/bytecode.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "interpreter/oopMapCache.hpp" 35 #include "memory/allocation.inline.hpp" 36 #include "memory/oopFactory.hpp" 37 #include "memory/resourceArea.hpp" 38 #include "oops/method.hpp" 39 #include "oops/objArrayOop.inline.hpp" 40 #include "oops/oop.inline.hpp" 41 #include "oops/fieldStreams.hpp" 42 #include "oops/valueKlass.hpp" 43 #include "oops/verifyOopClosure.hpp" 44 #include "prims/jvmtiThreadState.hpp" 45 #include "runtime/biasedLocking.hpp" 46 #include "runtime/compilationPolicy.hpp" 47 #include "runtime/deoptimization.hpp" 48 #include "runtime/interfaceSupport.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 #include "runtime/signature.hpp" 51 #include "runtime/stubRoutines.hpp" 52 #include "runtime/thread.hpp" 53 #include "runtime/vframe.hpp" 54 #include "runtime/vframeArray.hpp" 55 #include "runtime/vframe_hp.hpp" 56 #include "utilities/events.hpp" 57 #include "utilities/xmlstream.hpp" 58 59 #if INCLUDE_JVMCI 60 #include "jvmci/jvmciRuntime.hpp" 61 #include "jvmci/jvmciJavaClasses.hpp" 62 #endif 63 64 65 bool DeoptimizationMarker::_is_active = false; 66 67 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 68 int caller_adjustment, 69 int caller_actual_parameters, 70 int number_of_frames, 71 intptr_t* frame_sizes, 72 address* frame_pcs, 73 BasicType return_type, 74 int exec_mode) { 75 _size_of_deoptimized_frame = size_of_deoptimized_frame; 76 _caller_adjustment = caller_adjustment; 77 _caller_actual_parameters = caller_actual_parameters; 78 _number_of_frames = number_of_frames; 79 _frame_sizes = frame_sizes; 80 _frame_pcs = frame_pcs; 81 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 82 _return_type = return_type; 83 _initial_info = 0; 84 // PD (x86 only) 85 _counter_temp = 0; 86 _unpack_kind = exec_mode; 87 _sender_sp_temp = 0; 88 89 _total_frame_sizes = size_of_frames(); 90 assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode"); 91 } 92 93 94 Deoptimization::UnrollBlock::~UnrollBlock() { 95 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 96 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 97 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 98 } 99 100 101 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 102 assert(register_number < RegisterMap::reg_count, "checking register number"); 103 return &_register_block[register_number * 2]; 104 } 105 106 107 108 int Deoptimization::UnrollBlock::size_of_frames() const { 109 // Acount first for the adjustment of the initial frame 110 int result = _caller_adjustment; 111 for (int index = 0; index < number_of_frames(); index++) { 112 result += frame_sizes()[index]; 113 } 114 return result; 115 } 116 117 118 void Deoptimization::UnrollBlock::print() { 119 ttyLocker ttyl; 120 tty->print_cr("UnrollBlock"); 121 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 122 tty->print( " frame_sizes: "); 123 for (int index = 0; index < number_of_frames(); index++) { 124 tty->print(INTX_FORMAT " ", frame_sizes()[index]); 125 } 126 tty->cr(); 127 } 128 129 130 // In order to make fetch_unroll_info work properly with escape 131 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 132 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 133 // of previously eliminated objects occurs in realloc_objects, which is 134 // called from the method fetch_unroll_info_helper below. 135 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode)) 136 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 137 // but makes the entry a little slower. There is however a little dance we have to 138 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 139 140 // fetch_unroll_info() is called at the beginning of the deoptimization 141 // handler. Note this fact before we start generating temporary frames 142 // that can confuse an asynchronous stack walker. This counter is 143 // decremented at the end of unpack_frames(). 144 if (TraceDeoptimization) { 145 tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread)); 146 } 147 thread->inc_in_deopt_handler(); 148 149 return fetch_unroll_info_helper(thread, exec_mode); 150 JRT_END 151 152 153 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 154 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) { 155 156 // Note: there is a safepoint safety issue here. No matter whether we enter 157 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 158 // the vframeArray is created. 159 // 160 161 // Allocate our special deoptimization ResourceMark 162 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 163 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 164 thread->set_deopt_mark(dmark); 165 166 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 167 RegisterMap map(thread, true); 168 RegisterMap dummy_map(thread, false); 169 // Now get the deoptee with a valid map 170 frame deoptee = stub_frame.sender(&map); 171 // Set the deoptee nmethod 172 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 173 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 174 bool skip_internal = thread->deopt_nmethod() != NULL && !thread->deopt_nmethod()->compiler()->is_jvmci(); 175 176 if (VerifyStack) { 177 thread->validate_frame_layout(); 178 } 179 180 // Create a growable array of VFrames where each VFrame represents an inlined 181 // Java frame. This storage is allocated with the usual system arena. 182 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 183 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 184 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 185 while (!vf->is_top()) { 186 assert(vf->is_compiled_frame(), "Wrong frame type"); 187 chunk->push(compiledVFrame::cast(vf)); 188 vf = vf->sender(); 189 } 190 assert(vf->is_compiled_frame(), "Wrong frame type"); 191 chunk->push(compiledVFrame::cast(vf)); 192 193 ScopeDesc* trap_scope = chunk->at(0)->scope(); 194 Handle exceptionObject; 195 if (trap_scope->rethrow_exception()) { 196 if (PrintDeoptimizationDetails) { 197 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()); 198 } 199 GrowableArray<ScopeValue*>* expressions = trap_scope->expressions(); 200 guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw"); 201 ScopeValue* topOfStack = expressions->top(); 202 exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj(); 203 assert(exceptionObject() != NULL, "exception oop can not be null"); 204 } 205 206 bool realloc_failures = false; 207 208 #if defined(COMPILER2) || INCLUDE_JVMCI 209 // Reallocate the non-escaping objects and restore their fields. Then 210 // relock objects if synchronization on them was eliminated. 211 #ifndef INCLUDE_JVMCI 212 if (DoEscapeAnalysis || EliminateNestedLocks) { 213 if (EliminateAllocations) { 214 #endif // INCLUDE_JVMCI 215 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 216 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 217 218 // The flag return_oop() indicates call sites which return oop 219 // in compiled code. Such sites include java method calls, 220 // runtime calls (for example, used to allocate new objects/arrays 221 // on slow code path) and any other calls generated in compiled code. 222 // It is not guaranteed that we can get such information here only 223 // by analyzing bytecode in deoptimized frames. This is why this flag 224 // is set during method compilation (see Compile::Process_OopMap_Node()). 225 // If the previous frame was popped, we don't have a result. 226 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution(); 227 Handle return_value; 228 if (save_oop_result) { 229 // Reallocation may trigger GC. If deoptimization happened on return from 230 // call which returns oop we need to save it since it is not in oopmap. 231 oop result = deoptee.saved_oop_result(&map); 232 assert(result == NULL || result->is_oop(), "must be oop"); 233 return_value = Handle(thread, result); 234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 235 if (TraceDeoptimization) { 236 ttyLocker ttyl; 237 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread)); 238 } 239 } 240 if (objects != NULL) { 241 JRT_BLOCK 242 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD); 243 reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal, THREAD); 244 JRT_END 245 #ifndef PRODUCT 246 if (TraceDeoptimization) { 247 ttyLocker ttyl; 248 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 249 print_objects(objects, realloc_failures); 250 } 251 #endif 252 } 253 if (save_oop_result) { 254 // Restore result. 255 deoptee.set_saved_oop_result(&map, return_value()); 256 } 257 #ifndef INCLUDE_JVMCI 258 } 259 if (EliminateLocks) { 260 #endif // INCLUDE_JVMCI 261 #ifndef PRODUCT 262 bool first = true; 263 #endif 264 for (int i = 0; i < chunk->length(); i++) { 265 compiledVFrame* cvf = chunk->at(i); 266 assert (cvf->scope() != NULL,"expect only compiled java frames"); 267 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 268 if (monitors->is_nonempty()) { 269 relock_objects(monitors, thread, realloc_failures); 270 #ifndef PRODUCT 271 if (PrintDeoptimizationDetails) { 272 ttyLocker ttyl; 273 for (int j = 0; j < monitors->length(); j++) { 274 MonitorInfo* mi = monitors->at(j); 275 if (mi->eliminated()) { 276 if (first) { 277 first = false; 278 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread)); 279 } 280 if (mi->owner_is_scalar_replaced()) { 281 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 282 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 283 } else { 284 tty->print_cr(" object <" INTPTR_FORMAT "> locked", p2i(mi->owner())); 285 } 286 } 287 } 288 } 289 #endif // !PRODUCT 290 } 291 } 292 #ifndef INCLUDE_JVMCI 293 } 294 } 295 #endif // INCLUDE_JVMCI 296 #endif // COMPILER2 || INCLUDE_JVMCI 297 298 // Ensure that no safepoint is taken after pointers have been stored 299 // in fields of rematerialized objects. If a safepoint occurs from here on 300 // out the java state residing in the vframeArray will be missed. 301 NoSafepointVerifier no_safepoint; 302 303 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 304 #if defined(COMPILER2) || INCLUDE_JVMCI 305 if (realloc_failures) { 306 pop_frames_failed_reallocs(thread, array); 307 } 308 #endif 309 310 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 311 thread->set_vframe_array_head(array); 312 313 // Now that the vframeArray has been created if we have any deferred local writes 314 // added by jvmti then we can free up that structure as the data is now in the 315 // vframeArray 316 317 if (thread->deferred_locals() != NULL) { 318 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 319 int i = 0; 320 do { 321 // Because of inlining we could have multiple vframes for a single frame 322 // and several of the vframes could have deferred writes. Find them all. 323 if (list->at(i)->id() == array->original().id()) { 324 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 325 list->remove_at(i); 326 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 327 delete dlv; 328 } else { 329 i++; 330 } 331 } while ( i < list->length() ); 332 if (list->length() == 0) { 333 thread->set_deferred_locals(NULL); 334 // free the list and elements back to C heap. 335 delete list; 336 } 337 338 } 339 340 #ifndef SHARK 341 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 342 CodeBlob* cb = stub_frame.cb(); 343 // Verify we have the right vframeArray 344 assert(cb->frame_size() >= 0, "Unexpected frame size"); 345 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 346 347 // If the deopt call site is a MethodHandle invoke call site we have 348 // to adjust the unpack_sp. 349 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 350 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 351 unpack_sp = deoptee.unextended_sp(); 352 353 #ifdef ASSERT 354 assert(cb->is_deoptimization_stub() || 355 cb->is_uncommon_trap_stub() || 356 strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 || 357 strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0, 358 "unexpected code blob: %s", cb->name()); 359 #endif 360 #else 361 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 362 #endif // !SHARK 363 364 // This is a guarantee instead of an assert because if vframe doesn't match 365 // we will unpack the wrong deoptimized frame and wind up in strange places 366 // where it will be very difficult to figure out what went wrong. Better 367 // to die an early death here than some very obscure death later when the 368 // trail is cold. 369 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 370 // in that it will fail to detect a problem when there is one. This needs 371 // more work in tiger timeframe. 372 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 373 374 int number_of_frames = array->frames(); 375 376 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 377 // virtual activation, which is the reverse of the elements in the vframes array. 378 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 379 // +1 because we always have an interpreter return address for the final slot. 380 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 381 int popframe_extra_args = 0; 382 // Create an interpreter return address for the stub to use as its return 383 // address so the skeletal frames are perfectly walkable 384 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 385 386 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 387 // activation be put back on the expression stack of the caller for reexecution 388 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 389 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 390 } 391 392 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 393 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 394 // than simply use array->sender.pc(). This requires us to walk the current set of frames 395 // 396 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 397 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 398 399 // It's possible that the number of parameters at the call site is 400 // different than number of arguments in the callee when method 401 // handles are used. If the caller is interpreted get the real 402 // value so that the proper amount of space can be added to it's 403 // frame. 404 bool caller_was_method_handle = false; 405 if (deopt_sender.is_interpreted_frame()) { 406 methodHandle method = deopt_sender.interpreter_frame_method(); 407 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 408 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 409 // Method handle invokes may involve fairly arbitrary chains of 410 // calls so it's impossible to know how much actual space the 411 // caller has for locals. 412 caller_was_method_handle = true; 413 } 414 } 415 416 // 417 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 418 // frame_sizes/frame_pcs[1] next oldest frame (int) 419 // frame_sizes/frame_pcs[n] youngest frame (int) 420 // 421 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 422 // owns the space for the return address to it's caller). Confusing ain't it. 423 // 424 // The vframe array can address vframes with indices running from 425 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 426 // When we create the skeletal frames we need the oldest frame to be in the zero slot 427 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 428 // so things look a little strange in this loop. 429 // 430 int callee_parameters = 0; 431 int callee_locals = 0; 432 for (int index = 0; index < array->frames(); index++ ) { 433 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 434 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 435 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 436 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 437 callee_locals, 438 index == 0, 439 popframe_extra_args); 440 // This pc doesn't have to be perfect just good enough to identify the frame 441 // as interpreted so the skeleton frame will be walkable 442 // The correct pc will be set when the skeleton frame is completely filled out 443 // The final pc we store in the loop is wrong and will be overwritten below 444 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 445 446 callee_parameters = array->element(index)->method()->size_of_parameters(); 447 callee_locals = array->element(index)->method()->max_locals(); 448 popframe_extra_args = 0; 449 } 450 451 // Compute whether the root vframe returns a float or double value. 452 BasicType return_type; 453 { 454 HandleMark hm; 455 methodHandle method(thread, array->element(0)->method()); 456 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 457 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 458 } 459 460 // Compute information for handling adapters and adjusting the frame size of the caller. 461 int caller_adjustment = 0; 462 463 // Compute the amount the oldest interpreter frame will have to adjust 464 // its caller's stack by. If the caller is a compiled frame then 465 // we pretend that the callee has no parameters so that the 466 // extension counts for the full amount of locals and not just 467 // locals-parms. This is because without a c2i adapter the parm 468 // area as created by the compiled frame will not be usable by 469 // the interpreter. (Depending on the calling convention there 470 // may not even be enough space). 471 472 // QQQ I'd rather see this pushed down into last_frame_adjust 473 // and have it take the sender (aka caller). 474 475 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 476 caller_adjustment = last_frame_adjust(0, callee_locals); 477 } else if (callee_locals > callee_parameters) { 478 // The caller frame may need extending to accommodate 479 // non-parameter locals of the first unpacked interpreted frame. 480 // Compute that adjustment. 481 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 482 } 483 484 // If the sender is deoptimized the we must retrieve the address of the handler 485 // since the frame will "magically" show the original pc before the deopt 486 // and we'd undo the deopt. 487 488 frame_pcs[0] = deopt_sender.raw_pc(); 489 490 #ifndef SHARK 491 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 492 #endif // SHARK 493 494 #ifdef INCLUDE_JVMCI 495 if (exceptionObject() != NULL) { 496 thread->set_exception_oop(exceptionObject()); 497 exec_mode = Unpack_exception; 498 } 499 #endif 500 501 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 502 caller_adjustment * BytesPerWord, 503 caller_was_method_handle ? 0 : callee_parameters, 504 number_of_frames, 505 frame_sizes, 506 frame_pcs, 507 return_type, 508 exec_mode); 509 // On some platforms, we need a way to pass some platform dependent 510 // information to the unpacking code so the skeletal frames come out 511 // correct (initial fp value, unextended sp, ...) 512 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 513 514 if (array->frames() > 1) { 515 if (VerifyStack && TraceDeoptimization) { 516 ttyLocker ttyl; 517 tty->print_cr("Deoptimizing method containing inlining"); 518 } 519 } 520 521 array->set_unroll_block(info); 522 return info; 523 } 524 525 // Called to cleanup deoptimization data structures in normal case 526 // after unpacking to stack and when stack overflow error occurs 527 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 528 vframeArray *array) { 529 530 // Get array if coming from exception 531 if (array == NULL) { 532 array = thread->vframe_array_head(); 533 } 534 thread->set_vframe_array_head(NULL); 535 536 // Free the previous UnrollBlock 537 vframeArray* old_array = thread->vframe_array_last(); 538 thread->set_vframe_array_last(array); 539 540 if (old_array != NULL) { 541 UnrollBlock* old_info = old_array->unroll_block(); 542 old_array->set_unroll_block(NULL); 543 delete old_info; 544 delete old_array; 545 } 546 547 // Deallocate any resource creating in this routine and any ResourceObjs allocated 548 // inside the vframeArray (StackValueCollections) 549 550 delete thread->deopt_mark(); 551 thread->set_deopt_mark(NULL); 552 thread->set_deopt_nmethod(NULL); 553 554 555 if (JvmtiExport::can_pop_frame()) { 556 #ifndef CC_INTERP 557 // Regardless of whether we entered this routine with the pending 558 // popframe condition bit set, we should always clear it now 559 thread->clear_popframe_condition(); 560 #else 561 // C++ interpreter will clear has_pending_popframe when it enters 562 // with method_resume. For deopt_resume2 we clear it now. 563 if (thread->popframe_forcing_deopt_reexecution()) 564 thread->clear_popframe_condition(); 565 #endif /* CC_INTERP */ 566 } 567 568 // unpack_frames() is called at the end of the deoptimization handler 569 // and (in C2) at the end of the uncommon trap handler. Note this fact 570 // so that an asynchronous stack walker can work again. This counter is 571 // incremented at the beginning of fetch_unroll_info() and (in C2) at 572 // the beginning of uncommon_trap(). 573 thread->dec_in_deopt_handler(); 574 } 575 576 // Moved from cpu directories because none of the cpus has callee save values. 577 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp. 578 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) { 579 580 // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in 581 // the days we had adapter frames. When we deoptimize a situation where a 582 // compiled caller calls a compiled caller will have registers it expects 583 // to survive the call to the callee. If we deoptimize the callee the only 584 // way we can restore these registers is to have the oldest interpreter 585 // frame that we create restore these values. That is what this routine 586 // will accomplish. 587 588 // At the moment we have modified c2 to not have any callee save registers 589 // so this problem does not exist and this routine is just a place holder. 590 591 assert(f->is_interpreted_frame(), "must be interpreted"); 592 } 593 594 // Return BasicType of value being returned 595 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 596 597 // We are already active int he special DeoptResourceMark any ResourceObj's we 598 // allocate will be freed at the end of the routine. 599 600 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 601 // but makes the entry a little slower. There is however a little dance we have to 602 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 603 ResetNoHandleMark rnhm; // No-op in release/product versions 604 HandleMark hm; 605 606 frame stub_frame = thread->last_frame(); 607 608 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 609 // must point to the vframeArray for the unpack frame. 610 vframeArray* array = thread->vframe_array_head(); 611 612 #ifndef PRODUCT 613 if (TraceDeoptimization) { 614 ttyLocker ttyl; 615 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", 616 p2i(thread), p2i(array), exec_mode); 617 } 618 #endif 619 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 620 p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode); 621 622 UnrollBlock* info = array->unroll_block(); 623 624 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 625 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 626 627 BasicType bt = info->return_type(); 628 629 // If we have an exception pending, claim that the return type is an oop 630 // so the deopt_blob does not overwrite the exception_oop. 631 632 if (exec_mode == Unpack_exception) 633 bt = T_OBJECT; 634 635 // Cleanup thread deopt data 636 cleanup_deopt_info(thread, array); 637 638 #ifndef PRODUCT 639 if (VerifyStack) { 640 ResourceMark res_mark; 641 642 thread->validate_frame_layout(); 643 644 // Verify that the just-unpacked frames match the interpreter's 645 // notions of expression stack and locals 646 vframeArray* cur_array = thread->vframe_array_last(); 647 RegisterMap rm(thread, false); 648 rm.set_include_argument_oops(false); 649 bool is_top_frame = true; 650 int callee_size_of_parameters = 0; 651 int callee_max_locals = 0; 652 for (int i = 0; i < cur_array->frames(); i++) { 653 vframeArrayElement* el = cur_array->element(i); 654 frame* iframe = el->iframe(); 655 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 656 657 // Get the oop map for this bci 658 InterpreterOopMap mask; 659 int cur_invoke_parameter_size = 0; 660 bool try_next_mask = false; 661 int next_mask_expression_stack_size = -1; 662 int top_frame_expression_stack_adjustment = 0; 663 methodHandle mh(thread, iframe->interpreter_frame_method()); 664 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 665 BytecodeStream str(mh); 666 str.set_start(iframe->interpreter_frame_bci()); 667 int max_bci = mh->code_size(); 668 // Get to the next bytecode if possible 669 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 670 // Check to see if we can grab the number of outgoing arguments 671 // at an uncommon trap for an invoke (where the compiler 672 // generates debug info before the invoke has executed) 673 Bytecodes::Code cur_code = str.next(); 674 if (cur_code == Bytecodes::_invokevirtual || 675 cur_code == Bytecodes::_invokedirect || 676 cur_code == Bytecodes::_invokespecial || 677 cur_code == Bytecodes::_invokestatic || 678 cur_code == Bytecodes::_invokeinterface || 679 cur_code == Bytecodes::_invokedynamic) { 680 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 681 Symbol* signature = invoke.signature(); 682 ArgumentSizeComputer asc(signature); 683 cur_invoke_parameter_size = asc.size(); 684 if (invoke.has_receiver()) { 685 // Add in receiver 686 ++cur_invoke_parameter_size; 687 } 688 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 689 callee_size_of_parameters++; 690 } 691 } 692 if (str.bci() < max_bci) { 693 Bytecodes::Code bc = str.next(); 694 if (bc >= 0) { 695 // The interpreter oop map generator reports results before 696 // the current bytecode has executed except in the case of 697 // calls. It seems to be hard to tell whether the compiler 698 // has emitted debug information matching the "state before" 699 // a given bytecode or the state after, so we try both 700 switch (cur_code) { 701 case Bytecodes::_invokevirtual: 702 case Bytecodes::_invokedirect: 703 case Bytecodes::_invokespecial: 704 case Bytecodes::_invokestatic: 705 case Bytecodes::_invokeinterface: 706 case Bytecodes::_invokedynamic: 707 case Bytecodes::_athrow: 708 break; 709 default: { 710 InterpreterOopMap next_mask; 711 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 712 next_mask_expression_stack_size = next_mask.expression_stack_size(); 713 // Need to subtract off the size of the result type of 714 // the bytecode because this is not described in the 715 // debug info but returned to the interpreter in the TOS 716 // caching register 717 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 718 if (bytecode_result_type != T_ILLEGAL) { 719 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 720 } 721 assert(top_frame_expression_stack_adjustment >= 0, ""); 722 try_next_mask = true; 723 break; 724 } 725 } 726 } 727 } 728 729 // Verify stack depth and oops in frame 730 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 731 if (!( 732 /* SPARC */ 733 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 734 /* x86 */ 735 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 736 (try_next_mask && 737 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 738 top_frame_expression_stack_adjustment))) || 739 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 740 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 741 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 742 )) { 743 ttyLocker ttyl; 744 745 // Print out some information that will help us debug the problem 746 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 747 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 748 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 749 iframe->interpreter_frame_expression_stack_size()); 750 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 751 tty->print_cr(" try_next_mask = %d", try_next_mask); 752 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 753 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 754 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 755 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 756 tty->print_cr(" exec_mode = %d", exec_mode); 757 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 758 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id()); 759 tty->print_cr(" Interpreted frames:"); 760 for (int k = 0; k < cur_array->frames(); k++) { 761 vframeArrayElement* el = cur_array->element(k); 762 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 763 } 764 cur_array->print_on_2(tty); 765 guarantee(false, "wrong number of expression stack elements during deopt"); 766 } 767 VerifyOopClosure verify; 768 iframe->oops_interpreted_do(&verify, NULL, &rm, false); 769 callee_size_of_parameters = mh->size_of_parameters(); 770 callee_max_locals = mh->max_locals(); 771 is_top_frame = false; 772 } 773 } 774 #endif /* !PRODUCT */ 775 776 777 return bt; 778 JRT_END 779 780 781 int Deoptimization::deoptimize_dependents() { 782 Threads::deoptimized_wrt_marked_nmethods(); 783 return 0; 784 } 785 786 Deoptimization::DeoptAction Deoptimization::_unloaded_action 787 = Deoptimization::Action_reinterpret; 788 789 #if defined(COMPILER2) || INCLUDE_JVMCI 790 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 791 Handle pending_exception(thread->pending_exception()); 792 const char* exception_file = thread->exception_file(); 793 int exception_line = thread->exception_line(); 794 thread->clear_pending_exception(); 795 796 bool failures = false; 797 798 for (int i = 0; i < objects->length(); i++) { 799 assert(objects->at(i)->is_object(), "invalid debug information"); 800 ObjectValue* sv = (ObjectValue*) objects->at(i); 801 802 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 803 oop obj = NULL; 804 805 if (k->is_instance_klass()) { 806 InstanceKlass* ik = InstanceKlass::cast(k()); 807 obj = ik->allocate_instance(THREAD); 808 } else if (k->is_typeArray_klass()) { 809 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 810 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 811 int len = sv->field_size() / type2size[ak->element_type()]; 812 obj = ak->allocate(len, THREAD); 813 } else if (k->is_objArray_klass()) { 814 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 815 obj = ak->allocate(sv->field_size(), THREAD); 816 } 817 818 if (obj == NULL) { 819 failures = true; 820 } 821 822 assert(sv->value().is_null(), "redundant reallocation"); 823 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 824 CLEAR_PENDING_EXCEPTION; 825 sv->set_value(obj); 826 } 827 828 if (failures) { 829 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 830 } else if (pending_exception.not_null()) { 831 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 832 } 833 834 return failures; 835 } 836 837 // restore elements of an eliminated type array 838 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 839 int index = 0; 840 intptr_t val; 841 842 for (int i = 0; i < sv->field_size(); i++) { 843 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 844 switch(type) { 845 case T_LONG: case T_DOUBLE: { 846 assert(value->type() == T_INT, "Agreement."); 847 StackValue* low = 848 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 849 #ifdef _LP64 850 jlong res = (jlong)low->get_int(); 851 #else 852 #ifdef SPARC 853 // For SPARC we have to swap high and low words. 854 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 855 #else 856 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 857 #endif //SPARC 858 #endif 859 obj->long_at_put(index, res); 860 break; 861 } 862 863 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 864 case T_INT: case T_FLOAT: { // 4 bytes. 865 assert(value->type() == T_INT, "Agreement."); 866 bool big_value = false; 867 if (i + 1 < sv->field_size() && type == T_INT) { 868 if (sv->field_at(i)->is_location()) { 869 Location::Type type = ((LocationValue*) sv->field_at(i))->location().type(); 870 if (type == Location::dbl || type == Location::lng) { 871 big_value = true; 872 } 873 } else if (sv->field_at(i)->is_constant_int()) { 874 ScopeValue* next_scope_field = sv->field_at(i + 1); 875 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 876 big_value = true; 877 } 878 } 879 } 880 881 if (big_value) { 882 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 883 #ifdef _LP64 884 jlong res = (jlong)low->get_int(); 885 #else 886 #ifdef SPARC 887 // For SPARC we have to swap high and low words. 888 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 889 #else 890 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 891 #endif //SPARC 892 #endif 893 obj->int_at_put(index, (jint)*((jint*)&res)); 894 obj->int_at_put(++index, (jint)*(((jint*)&res) + 1)); 895 } else { 896 val = value->get_int(); 897 obj->int_at_put(index, (jint)*((jint*)&val)); 898 } 899 break; 900 } 901 902 case T_SHORT: case T_CHAR: // 2 bytes 903 assert(value->type() == T_INT, "Agreement."); 904 val = value->get_int(); 905 obj->short_at_put(index, (jshort)*((jint*)&val)); 906 break; 907 908 case T_BOOLEAN: case T_BYTE: // 1 byte 909 assert(value->type() == T_INT, "Agreement."); 910 val = value->get_int(); 911 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 912 break; 913 914 default: 915 ShouldNotReachHere(); 916 } 917 index++; 918 } 919 } 920 921 922 // restore fields of an eliminated object array 923 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 924 for (int i = 0; i < sv->field_size(); i++) { 925 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 926 assert(value->type() == T_OBJECT, "object element expected"); 927 obj->obj_at_put(i, value->get_obj()()); 928 } 929 } 930 931 class ReassignedField { 932 public: 933 int _offset; 934 BasicType _type; 935 InstanceKlass* _klass; 936 public: 937 ReassignedField() { 938 _offset = 0; 939 _type = T_ILLEGAL; 940 _klass = NULL; 941 } 942 }; 943 944 int compare(ReassignedField* left, ReassignedField* right) { 945 return left->_offset - right->_offset; 946 } 947 948 // Restore fields of an eliminated instance object using the same field order 949 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true) 950 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal, int base_offset, TRAPS) { 951 if (klass->superklass() != NULL) { 952 svIndex = reassign_fields_by_klass(klass->superklass(), fr, reg_map, sv, svIndex, obj, skip_internal, 0, CHECK_0); 953 } 954 955 GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>(); 956 for (AllFieldStream fs(klass); !fs.done(); fs.next()) { 957 if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) { 958 ReassignedField field; 959 field._offset = fs.offset(); 960 field._type = FieldType::basic_type(fs.signature()); 961 if (field._type == T_VALUETYPE) { 962 // Resolve klass of flattened value type field 963 SignatureStream ss(fs.signature(), false); 964 Klass* vk = ss.as_klass(Handle(klass->class_loader()), Handle(klass->protection_domain()), SignatureStream::NCDFError, CHECK_0); 965 assert(vk->is_value(), "must be a ValueKlass"); 966 field._klass = InstanceKlass::cast(vk); 967 } 968 fields->append(field); 969 } 970 } 971 fields->sort(compare); 972 for (int i = 0; i < fields->length(); i++) { 973 intptr_t val; 974 ScopeValue* scope_field = sv->field_at(svIndex); 975 StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field); 976 int offset = base_offset + fields->at(i)._offset; 977 BasicType type = fields->at(i)._type; 978 switch (type) { 979 case T_OBJECT: case T_ARRAY: 980 assert(value->type() == T_OBJECT, "Agreement."); 981 obj->obj_field_put(offset, value->get_obj()()); 982 break; 983 984 case T_VALUETYPE: { 985 // Recursively re-assign flattened value type fields 986 InstanceKlass* vk = fields->at(i)._klass; 987 assert(vk != NULL, "must be resolved"); 988 offset -= ValueKlass::cast(vk)->first_field_offset(); // Adjust offset to omit oop header 989 svIndex = reassign_fields_by_klass(vk, fr, reg_map, sv, svIndex, obj, skip_internal, offset, CHECK_0); 990 continue; // Continue because we don't need to increment svIndex 991 } 992 993 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 994 case T_INT: case T_FLOAT: { // 4 bytes. 995 assert(value->type() == T_INT, "Agreement."); 996 bool big_value = false; 997 if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) { 998 if (scope_field->is_location()) { 999 Location::Type type = ((LocationValue*) scope_field)->location().type(); 1000 if (type == Location::dbl || type == Location::lng) { 1001 big_value = true; 1002 } 1003 } 1004 if (scope_field->is_constant_int()) { 1005 ScopeValue* next_scope_field = sv->field_at(svIndex + 1); 1006 if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) { 1007 big_value = true; 1008 } 1009 } 1010 } 1011 1012 if (big_value) { 1013 i++; 1014 assert(i < fields->length(), "second T_INT field needed"); 1015 assert(fields->at(i)._type == T_INT, "T_INT field needed"); 1016 } else { 1017 val = value->get_int(); 1018 obj->int_field_put(offset, (jint)*((jint*)&val)); 1019 break; 1020 } 1021 } 1022 /* no break */ 1023 1024 case T_LONG: case T_DOUBLE: { 1025 assert(value->type() == T_INT, "Agreement."); 1026 StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex)); 1027 #ifdef _LP64 1028 jlong res = (jlong)low->get_int(); 1029 #else 1030 #ifdef SPARC 1031 // For SPARC we have to swap high and low words. 1032 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 1033 #else 1034 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 1035 #endif //SPARC 1036 #endif 1037 obj->long_field_put(offset, res); 1038 break; 1039 } 1040 1041 case T_SHORT: case T_CHAR: // 2 bytes 1042 assert(value->type() == T_INT, "Agreement."); 1043 val = value->get_int(); 1044 obj->short_field_put(offset, (jshort)*((jint*)&val)); 1045 break; 1046 1047 case T_BOOLEAN: case T_BYTE: // 1 byte 1048 assert(value->type() == T_INT, "Agreement."); 1049 val = value->get_int(); 1050 obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 1051 break; 1052 1053 default: 1054 ShouldNotReachHere(); 1055 } 1056 svIndex++; 1057 } 1058 return svIndex; 1059 } 1060 1061 // restore fields of all eliminated objects and arrays 1062 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal, TRAPS) { 1063 for (int i = 0; i < objects->length(); i++) { 1064 ObjectValue* sv = (ObjectValue*) objects->at(i); 1065 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1066 Handle obj = sv->value(); 1067 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1068 if (PrintDeoptimizationDetails) { 1069 tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string()); 1070 } 1071 if (obj.is_null()) { 1072 continue; 1073 } 1074 1075 if (k->is_instance_klass()) { 1076 InstanceKlass* ik = InstanceKlass::cast(k()); 1077 reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal, 0, CHECK); 1078 } else if (k->is_typeArray_klass()) { 1079 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 1080 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 1081 } else if (k->is_objArray_klass()) { 1082 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 1083 } 1084 } 1085 } 1086 1087 1088 // relock objects for which synchronization was eliminated 1089 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 1090 for (int i = 0; i < monitors->length(); i++) { 1091 MonitorInfo* mon_info = monitors->at(i); 1092 if (mon_info->eliminated()) { 1093 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 1094 if (!mon_info->owner_is_scalar_replaced()) { 1095 Handle obj = Handle(mon_info->owner()); 1096 markOop mark = obj->mark(); 1097 if (UseBiasedLocking && mark->has_bias_pattern()) { 1098 // New allocated objects may have the mark set to anonymously biased. 1099 // Also the deoptimized method may called methods with synchronization 1100 // where the thread-local object is bias locked to the current thread. 1101 assert(mark->is_biased_anonymously() || 1102 mark->biased_locker() == thread, "should be locked to current thread"); 1103 // Reset mark word to unbiased prototype. 1104 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 1105 obj->set_mark(unbiased_prototype); 1106 } 1107 BasicLock* lock = mon_info->lock(); 1108 ObjectSynchronizer::slow_enter(obj, lock, thread); 1109 assert(mon_info->owner()->is_locked(), "object must be locked now"); 1110 } 1111 } 1112 } 1113 } 1114 1115 1116 #ifndef PRODUCT 1117 // print information about reallocated objects 1118 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 1119 fieldDescriptor fd; 1120 1121 for (int i = 0; i < objects->length(); i++) { 1122 ObjectValue* sv = (ObjectValue*) objects->at(i); 1123 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1124 Handle obj = sv->value(); 1125 1126 tty->print(" object <" INTPTR_FORMAT "> of type ", p2i(sv->value()())); 1127 k->print_value(); 1128 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1129 if (obj.is_null()) { 1130 tty->print(" allocation failed"); 1131 } else { 1132 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1133 } 1134 tty->cr(); 1135 1136 if (Verbose && !obj.is_null()) { 1137 k->oop_print_on(obj(), tty); 1138 } 1139 } 1140 } 1141 #endif 1142 #endif // COMPILER2 || INCLUDE_JVMCI 1143 1144 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1145 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp())); 1146 1147 #ifndef PRODUCT 1148 if (PrintDeoptimizationDetails) { 1149 ttyLocker ttyl; 1150 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread)); 1151 fr.print_on(tty); 1152 tty->print_cr(" Virtual frames (innermost first):"); 1153 for (int index = 0; index < chunk->length(); index++) { 1154 compiledVFrame* vf = chunk->at(index); 1155 tty->print(" %2d - ", index); 1156 vf->print_value(); 1157 int bci = chunk->at(index)->raw_bci(); 1158 const char* code_name; 1159 if (bci == SynchronizationEntryBCI) { 1160 code_name = "sync entry"; 1161 } else { 1162 Bytecodes::Code code = vf->method()->code_at(bci); 1163 code_name = Bytecodes::name(code); 1164 } 1165 tty->print(" - %s", code_name); 1166 tty->print_cr(" @ bci %d ", bci); 1167 if (Verbose) { 1168 vf->print(); 1169 tty->cr(); 1170 } 1171 } 1172 } 1173 #endif 1174 1175 // Register map for next frame (used for stack crawl). We capture 1176 // the state of the deopt'ing frame's caller. Thus if we need to 1177 // stuff a C2I adapter we can properly fill in the callee-save 1178 // register locations. 1179 frame caller = fr.sender(reg_map); 1180 int frame_size = caller.sp() - fr.sp(); 1181 1182 frame sender = caller; 1183 1184 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1185 // the vframeArray containing the unpacking information is allocated in the C heap. 1186 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1187 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1188 1189 // Compare the vframeArray to the collected vframes 1190 assert(array->structural_compare(thread, chunk), "just checking"); 1191 1192 #ifndef PRODUCT 1193 if (PrintDeoptimizationDetails) { 1194 ttyLocker ttyl; 1195 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, p2i(array)); 1196 } 1197 #endif // PRODUCT 1198 1199 return array; 1200 } 1201 1202 #if defined(COMPILER2) || INCLUDE_JVMCI 1203 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1204 // Reallocation of some scalar replaced objects failed. Record 1205 // that we need to pop all the interpreter frames for the 1206 // deoptimized compiled frame. 1207 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1208 thread->set_frames_to_pop_failed_realloc(array->frames()); 1209 // Unlock all monitors here otherwise the interpreter will see a 1210 // mix of locked and unlocked monitors (because of failed 1211 // reallocations of synchronized objects) and be confused. 1212 for (int i = 0; i < array->frames(); i++) { 1213 MonitorChunk* monitors = array->element(i)->monitors(); 1214 if (monitors != NULL) { 1215 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1216 BasicObjectLock* src = monitors->at(j); 1217 if (src->obj() != NULL) { 1218 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1219 } 1220 } 1221 array->element(i)->free_monitors(thread); 1222 #ifdef ASSERT 1223 array->element(i)->set_removed_monitors(); 1224 #endif 1225 } 1226 } 1227 } 1228 #endif 1229 1230 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1231 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1232 for (int i = 0; i < monitors->length(); i++) { 1233 MonitorInfo* mon_info = monitors->at(i); 1234 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1235 objects_to_revoke->append(Handle(mon_info->owner())); 1236 } 1237 } 1238 } 1239 1240 1241 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1242 if (!UseBiasedLocking) { 1243 return; 1244 } 1245 1246 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1247 1248 // Unfortunately we don't have a RegisterMap available in most of 1249 // the places we want to call this routine so we need to walk the 1250 // stack again to update the register map. 1251 if (map == NULL || !map->update_map()) { 1252 StackFrameStream sfs(thread, true); 1253 bool found = false; 1254 while (!found && !sfs.is_done()) { 1255 frame* cur = sfs.current(); 1256 sfs.next(); 1257 found = cur->id() == fr.id(); 1258 } 1259 assert(found, "frame to be deoptimized not found on target thread's stack"); 1260 map = sfs.register_map(); 1261 } 1262 1263 vframe* vf = vframe::new_vframe(&fr, map, thread); 1264 compiledVFrame* cvf = compiledVFrame::cast(vf); 1265 // Revoke monitors' biases in all scopes 1266 while (!cvf->is_top()) { 1267 collect_monitors(cvf, objects_to_revoke); 1268 cvf = compiledVFrame::cast(cvf->sender()); 1269 } 1270 collect_monitors(cvf, objects_to_revoke); 1271 1272 if (SafepointSynchronize::is_at_safepoint()) { 1273 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1274 } else { 1275 BiasedLocking::revoke(objects_to_revoke); 1276 } 1277 } 1278 1279 1280 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1281 if (!UseBiasedLocking) { 1282 return; 1283 } 1284 1285 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1286 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1287 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1288 if (jt->has_last_Java_frame()) { 1289 StackFrameStream sfs(jt, true); 1290 while (!sfs.is_done()) { 1291 frame* cur = sfs.current(); 1292 if (cb->contains(cur->pc())) { 1293 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1294 compiledVFrame* cvf = compiledVFrame::cast(vf); 1295 // Revoke monitors' biases in all scopes 1296 while (!cvf->is_top()) { 1297 collect_monitors(cvf, objects_to_revoke); 1298 cvf = compiledVFrame::cast(cvf->sender()); 1299 } 1300 collect_monitors(cvf, objects_to_revoke); 1301 } 1302 sfs.next(); 1303 } 1304 } 1305 } 1306 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1307 } 1308 1309 1310 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) { 1311 assert(fr.can_be_deoptimized(), "checking frame type"); 1312 1313 gather_statistics(reason, Action_none, Bytecodes::_illegal); 1314 1315 if (LogCompilation && xtty != NULL) { 1316 nmethod* nm = fr.cb()->as_nmethod_or_null(); 1317 assert(nm != NULL, "only compiled methods can deopt"); 1318 1319 ttyLocker ttyl; 1320 xtty->begin_head("deoptimized thread='" UINTX_FORMAT "'", (uintx)thread->osthread()->thread_id()); 1321 nm->log_identity(xtty); 1322 xtty->end_head(); 1323 for (ScopeDesc* sd = nm->scope_desc_at(fr.pc()); ; sd = sd->sender()) { 1324 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1325 xtty->method(sd->method()); 1326 xtty->end_elem(); 1327 if (sd->is_top()) break; 1328 } 1329 xtty->tail("deoptimized"); 1330 } 1331 1332 // Patch the compiled method so that when execution returns to it we will 1333 // deopt the execution state and return to the interpreter. 1334 fr.deoptimize(thread); 1335 } 1336 1337 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1338 deoptimize(thread, fr, map, Reason_constraint); 1339 } 1340 1341 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map, DeoptReason reason) { 1342 // Deoptimize only if the frame comes from compile code. 1343 // Do not deoptimize the frame which is already patched 1344 // during the execution of the loops below. 1345 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1346 return; 1347 } 1348 ResourceMark rm; 1349 DeoptimizationMarker dm; 1350 if (UseBiasedLocking) { 1351 revoke_biases_of_monitors(thread, fr, map); 1352 } 1353 deoptimize_single_frame(thread, fr, reason); 1354 1355 } 1356 1357 1358 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1359 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1360 "can only deoptimize other thread at a safepoint"); 1361 // Compute frame and register map based on thread and sp. 1362 RegisterMap reg_map(thread, UseBiasedLocking); 1363 frame fr = thread->last_frame(); 1364 while (fr.id() != id) { 1365 fr = fr.sender(®_map); 1366 } 1367 deoptimize(thread, fr, ®_map, reason); 1368 } 1369 1370 1371 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) { 1372 if (thread == Thread::current()) { 1373 Deoptimization::deoptimize_frame_internal(thread, id, reason); 1374 } else { 1375 VM_DeoptimizeFrame deopt(thread, id, reason); 1376 VMThread::execute(&deopt); 1377 } 1378 } 1379 1380 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1381 deoptimize_frame(thread, id, Reason_constraint); 1382 } 1383 1384 // JVMTI PopFrame support 1385 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1386 { 1387 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1388 } 1389 JRT_END 1390 1391 MethodData* 1392 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1393 bool create_if_missing) { 1394 Thread* THREAD = thread; 1395 MethodData* mdo = m()->method_data(); 1396 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1397 // Build an MDO. Ignore errors like OutOfMemory; 1398 // that simply means we won't have an MDO to update. 1399 Method::build_interpreter_method_data(m, THREAD); 1400 if (HAS_PENDING_EXCEPTION) { 1401 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1402 CLEAR_PENDING_EXCEPTION; 1403 } 1404 mdo = m()->method_data(); 1405 } 1406 return mdo; 1407 } 1408 1409 #if defined(COMPILER2) || defined(SHARK) || INCLUDE_JVMCI 1410 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) { 1411 // in case of an unresolved klass entry, load the class. 1412 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1413 Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK); 1414 return; 1415 } 1416 1417 if (!constant_pool->tag_at(index).is_symbol()) return; 1418 1419 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1420 Symbol* symbol = constant_pool->symbol_at(index); 1421 1422 // class name? 1423 if (symbol->byte_at(0) != '(') { 1424 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1425 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1426 return; 1427 } 1428 1429 // then it must be a signature! 1430 ResourceMark rm(THREAD); 1431 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1432 if (ss.is_object()) { 1433 Symbol* class_name = ss.as_symbol(CHECK); 1434 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1435 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1436 } 1437 } 1438 } 1439 1440 1441 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) { 1442 EXCEPTION_MARK; 1443 load_class_by_index(constant_pool, index, THREAD); 1444 if (HAS_PENDING_EXCEPTION) { 1445 // Exception happened during classloading. We ignore the exception here, since it 1446 // is going to be rethrown since the current activation is going to be deoptimized and 1447 // the interpreter will re-execute the bytecode. 1448 CLEAR_PENDING_EXCEPTION; 1449 // Class loading called java code which may have caused a stack 1450 // overflow. If the exception was thrown right before the return 1451 // to the runtime the stack is no longer guarded. Reguard the 1452 // stack otherwise if we return to the uncommon trap blob and the 1453 // stack bang causes a stack overflow we crash. 1454 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1455 JavaThread* thread = (JavaThread*)THREAD; 1456 bool guard_pages_enabled = thread->stack_guards_enabled(); 1457 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1458 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1459 } 1460 } 1461 1462 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1463 HandleMark hm; 1464 1465 // uncommon_trap() is called at the beginning of the uncommon trap 1466 // handler. Note this fact before we start generating temporary frames 1467 // that can confuse an asynchronous stack walker. This counter is 1468 // decremented at the end of unpack_frames(). 1469 thread->inc_in_deopt_handler(); 1470 1471 // We need to update the map if we have biased locking. 1472 #if INCLUDE_JVMCI 1473 // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid 1474 RegisterMap reg_map(thread, true); 1475 #else 1476 RegisterMap reg_map(thread, UseBiasedLocking); 1477 #endif 1478 frame stub_frame = thread->last_frame(); 1479 frame fr = stub_frame.sender(®_map); 1480 // Make sure the calling nmethod is not getting deoptimized and removed 1481 // before we are done with it. 1482 nmethodLocker nl(fr.pc()); 1483 1484 // Log a message 1485 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT, 1486 trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin()); 1487 1488 { 1489 ResourceMark rm; 1490 1491 // Revoke biases of any monitors in the frame to ensure we can migrate them 1492 revoke_biases_of_monitors(thread, fr, ®_map); 1493 1494 DeoptReason reason = trap_request_reason(trap_request); 1495 DeoptAction action = trap_request_action(trap_request); 1496 #if INCLUDE_JVMCI 1497 int debug_id = trap_request_debug_id(trap_request); 1498 #endif 1499 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1500 1501 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1502 compiledVFrame* cvf = compiledVFrame::cast(vf); 1503 1504 nmethod* nm = cvf->code(); 1505 1506 ScopeDesc* trap_scope = cvf->scope(); 1507 1508 if (TraceDeoptimization) { 1509 ttyLocker ttyl; 1510 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() 1511 #if INCLUDE_JVMCI 1512 , debug_id 1513 #endif 1514 ); 1515 } 1516 1517 methodHandle trap_method = trap_scope->method(); 1518 int trap_bci = trap_scope->bci(); 1519 #if INCLUDE_JVMCI 1520 oop speculation = thread->pending_failed_speculation(); 1521 if (nm->is_compiled_by_jvmci()) { 1522 if (speculation != NULL) { 1523 oop speculation_log = nm->speculation_log(); 1524 if (speculation_log != NULL) { 1525 if (TraceDeoptimization || TraceUncollectedSpeculations) { 1526 if (HotSpotSpeculationLog::lastFailed(speculation_log) != NULL) { 1527 tty->print_cr("A speculation that was not collected by the compiler is being overwritten"); 1528 } 1529 } 1530 if (TraceDeoptimization) { 1531 tty->print_cr("Saving speculation to speculation log"); 1532 } 1533 HotSpotSpeculationLog::set_lastFailed(speculation_log, speculation); 1534 } else { 1535 if (TraceDeoptimization) { 1536 tty->print_cr("Speculation present but no speculation log"); 1537 } 1538 } 1539 thread->set_pending_failed_speculation(NULL); 1540 } else { 1541 if (TraceDeoptimization) { 1542 tty->print_cr("No speculation"); 1543 } 1544 } 1545 } else { 1546 assert(speculation == NULL, "There should not be a speculation for method compiled by non-JVMCI compilers"); 1547 } 1548 1549 if (trap_bci == SynchronizationEntryBCI) { 1550 trap_bci = 0; 1551 thread->set_pending_monitorenter(true); 1552 } 1553 1554 if (reason == Deoptimization::Reason_transfer_to_interpreter) { 1555 thread->set_pending_transfer_to_interpreter(true); 1556 } 1557 #endif 1558 1559 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1560 // Record this event in the histogram. 1561 gather_statistics(reason, action, trap_bc); 1562 1563 // Ensure that we can record deopt. history: 1564 // Need MDO to record RTM code generation state. 1565 bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking ); 1566 1567 methodHandle profiled_method; 1568 #if INCLUDE_JVMCI 1569 if (nm->is_compiled_by_jvmci()) { 1570 profiled_method = nm->method(); 1571 } else { 1572 profiled_method = trap_method; 1573 } 1574 #else 1575 profiled_method = trap_method; 1576 #endif 1577 1578 MethodData* trap_mdo = 1579 get_method_data(thread, profiled_method, create_if_missing); 1580 1581 // Log a message 1582 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1583 trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()), 1584 trap_method->name_and_sig_as_C_string(), trap_bci); 1585 1586 // Print a bunch of diagnostics, if requested. 1587 if (TraceDeoptimization || LogCompilation) { 1588 ResourceMark rm; 1589 ttyLocker ttyl; 1590 char buf[100]; 1591 if (xtty != NULL) { 1592 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s", 1593 os::current_thread_id(), 1594 format_trap_request(buf, sizeof(buf), trap_request)); 1595 nm->log_identity(xtty); 1596 } 1597 Symbol* class_name = NULL; 1598 bool unresolved = false; 1599 if (unloaded_class_index >= 0) { 1600 constantPoolHandle constants (THREAD, trap_method->constants()); 1601 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1602 class_name = constants->klass_name_at(unloaded_class_index); 1603 unresolved = true; 1604 if (xtty != NULL) 1605 xtty->print(" unresolved='1'"); 1606 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1607 class_name = constants->symbol_at(unloaded_class_index); 1608 } 1609 if (xtty != NULL) 1610 xtty->name(class_name); 1611 } 1612 if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) { 1613 // Dump the relevant MDO state. 1614 // This is the deopt count for the current reason, any previous 1615 // reasons or recompiles seen at this point. 1616 int dcnt = trap_mdo->trap_count(reason); 1617 if (dcnt != 0) 1618 xtty->print(" count='%d'", dcnt); 1619 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1620 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1621 if (dos != 0) { 1622 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1623 if (trap_state_is_recompiled(dos)) { 1624 int recnt2 = trap_mdo->overflow_recompile_count(); 1625 if (recnt2 != 0) 1626 xtty->print(" recompiles2='%d'", recnt2); 1627 } 1628 } 1629 } 1630 if (xtty != NULL) { 1631 xtty->stamp(); 1632 xtty->end_head(); 1633 } 1634 if (TraceDeoptimization) { // make noise on the tty 1635 tty->print("Uncommon trap occurred in"); 1636 nm->method()->print_short_name(tty); 1637 tty->print(" compiler=%s compile_id=%d", nm->compiler() == NULL ? "" : nm->compiler()->name(), nm->compile_id()); 1638 #if INCLUDE_JVMCI 1639 oop installedCode = nm->jvmci_installed_code(); 1640 if (installedCode != NULL) { 1641 oop installedCodeName = NULL; 1642 if (installedCode->is_a(InstalledCode::klass())) { 1643 installedCodeName = InstalledCode::name(installedCode); 1644 } 1645 if (installedCodeName != NULL) { 1646 tty->print(" (JVMCI: installedCodeName=%s) ", java_lang_String::as_utf8_string(installedCodeName)); 1647 } else { 1648 tty->print(" (JVMCI: installed code has no name) "); 1649 } 1650 } else if (nm->is_compiled_by_jvmci()) { 1651 tty->print(" (JVMCI: no installed code) "); 1652 } 1653 #endif 1654 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"), 1655 p2i(fr.pc()), 1656 os::current_thread_id(), 1657 trap_reason_name(reason), 1658 trap_action_name(action), 1659 unloaded_class_index 1660 #if INCLUDE_JVMCI 1661 , debug_id 1662 #endif 1663 ); 1664 if (class_name != NULL) { 1665 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1666 class_name->print_symbol_on(tty); 1667 } 1668 tty->cr(); 1669 } 1670 if (xtty != NULL) { 1671 // Log the precise location of the trap. 1672 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1673 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1674 xtty->method(sd->method()); 1675 xtty->end_elem(); 1676 if (sd->is_top()) break; 1677 } 1678 xtty->tail("uncommon_trap"); 1679 } 1680 } 1681 // (End diagnostic printout.) 1682 1683 // Load class if necessary 1684 if (unloaded_class_index >= 0) { 1685 constantPoolHandle constants(THREAD, trap_method->constants()); 1686 load_class_by_index(constants, unloaded_class_index); 1687 } 1688 1689 // Flush the nmethod if necessary and desirable. 1690 // 1691 // We need to avoid situations where we are re-flushing the nmethod 1692 // because of a hot deoptimization site. Repeated flushes at the same 1693 // point need to be detected by the compiler and avoided. If the compiler 1694 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1695 // module must take measures to avoid an infinite cycle of recompilation 1696 // and deoptimization. There are several such measures: 1697 // 1698 // 1. If a recompilation is ordered a second time at some site X 1699 // and for the same reason R, the action is adjusted to 'reinterpret', 1700 // to give the interpreter time to exercise the method more thoroughly. 1701 // If this happens, the method's overflow_recompile_count is incremented. 1702 // 1703 // 2. If the compiler fails to reduce the deoptimization rate, then 1704 // the method's overflow_recompile_count will begin to exceed the set 1705 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1706 // is adjusted to 'make_not_compilable', and the method is abandoned 1707 // to the interpreter. This is a performance hit for hot methods, 1708 // but is better than a disastrous infinite cycle of recompilations. 1709 // (Actually, only the method containing the site X is abandoned.) 1710 // 1711 // 3. In parallel with the previous measures, if the total number of 1712 // recompilations of a method exceeds the much larger set limit 1713 // PerMethodRecompilationCutoff, the method is abandoned. 1714 // This should only happen if the method is very large and has 1715 // many "lukewarm" deoptimizations. The code which enforces this 1716 // limit is elsewhere (class nmethod, class Method). 1717 // 1718 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1719 // to recompile at each bytecode independently of the per-BCI cutoff. 1720 // 1721 // The decision to update code is up to the compiler, and is encoded 1722 // in the Action_xxx code. If the compiler requests Action_none 1723 // no trap state is changed, no compiled code is changed, and the 1724 // computation suffers along in the interpreter. 1725 // 1726 // The other action codes specify various tactics for decompilation 1727 // and recompilation. Action_maybe_recompile is the loosest, and 1728 // allows the compiled code to stay around until enough traps are seen, 1729 // and until the compiler gets around to recompiling the trapping method. 1730 // 1731 // The other actions cause immediate removal of the present code. 1732 1733 // Traps caused by injected profile shouldn't pollute trap counts. 1734 bool injected_profile_trap = trap_method->has_injected_profile() && 1735 (reason == Reason_intrinsic || reason == Reason_unreached); 1736 1737 bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap; 1738 bool make_not_entrant = false; 1739 bool make_not_compilable = false; 1740 bool reprofile = false; 1741 switch (action) { 1742 case Action_none: 1743 // Keep the old code. 1744 update_trap_state = false; 1745 break; 1746 case Action_maybe_recompile: 1747 // Do not need to invalidate the present code, but we can 1748 // initiate another 1749 // Start compiler without (necessarily) invalidating the nmethod. 1750 // The system will tolerate the old code, but new code should be 1751 // generated when possible. 1752 break; 1753 case Action_reinterpret: 1754 // Go back into the interpreter for a while, and then consider 1755 // recompiling form scratch. 1756 make_not_entrant = true; 1757 // Reset invocation counter for outer most method. 1758 // This will allow the interpreter to exercise the bytecodes 1759 // for a while before recompiling. 1760 // By contrast, Action_make_not_entrant is immediate. 1761 // 1762 // Note that the compiler will track null_check, null_assert, 1763 // range_check, and class_check events and log them as if they 1764 // had been traps taken from compiled code. This will update 1765 // the MDO trap history so that the next compilation will 1766 // properly detect hot trap sites. 1767 reprofile = true; 1768 break; 1769 case Action_make_not_entrant: 1770 // Request immediate recompilation, and get rid of the old code. 1771 // Make them not entrant, so next time they are called they get 1772 // recompiled. Unloaded classes are loaded now so recompile before next 1773 // time they are called. Same for uninitialized. The interpreter will 1774 // link the missing class, if any. 1775 make_not_entrant = true; 1776 break; 1777 case Action_make_not_compilable: 1778 // Give up on compiling this method at all. 1779 make_not_entrant = true; 1780 make_not_compilable = true; 1781 break; 1782 default: 1783 ShouldNotReachHere(); 1784 } 1785 1786 // Setting +ProfileTraps fixes the following, on all platforms: 1787 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1788 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1789 // recompile relies on a MethodData* to record heroic opt failures. 1790 1791 // Whether the interpreter is producing MDO data or not, we also need 1792 // to use the MDO to detect hot deoptimization points and control 1793 // aggressive optimization. 1794 bool inc_recompile_count = false; 1795 ProfileData* pdata = NULL; 1796 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1797 assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity"); 1798 uint this_trap_count = 0; 1799 bool maybe_prior_trap = false; 1800 bool maybe_prior_recompile = false; 1801 pdata = query_update_method_data(trap_mdo, trap_bci, reason, true, 1802 #if INCLUDE_JVMCI 1803 nm->is_compiled_by_jvmci() && nm->is_osr_method(), 1804 #endif 1805 nm->method(), 1806 //outputs: 1807 this_trap_count, 1808 maybe_prior_trap, 1809 maybe_prior_recompile); 1810 // Because the interpreter also counts null, div0, range, and class 1811 // checks, these traps from compiled code are double-counted. 1812 // This is harmless; it just means that the PerXTrapLimit values 1813 // are in effect a little smaller than they look. 1814 1815 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1816 if (per_bc_reason != Reason_none) { 1817 // Now take action based on the partially known per-BCI history. 1818 if (maybe_prior_trap 1819 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1820 // If there are too many traps at this BCI, force a recompile. 1821 // This will allow the compiler to see the limit overflow, and 1822 // take corrective action, if possible. The compiler generally 1823 // does not use the exact PerBytecodeTrapLimit value, but instead 1824 // changes its tactics if it sees any traps at all. This provides 1825 // a little hysteresis, delaying a recompile until a trap happens 1826 // several times. 1827 // 1828 // Actually, since there is only one bit of counter per BCI, 1829 // the possible per-BCI counts are {0,1,(per-method count)}. 1830 // This produces accurate results if in fact there is only 1831 // one hot trap site, but begins to get fuzzy if there are 1832 // many sites. For example, if there are ten sites each 1833 // trapping two or more times, they each get the blame for 1834 // all of their traps. 1835 make_not_entrant = true; 1836 } 1837 1838 // Detect repeated recompilation at the same BCI, and enforce a limit. 1839 if (make_not_entrant && maybe_prior_recompile) { 1840 // More than one recompile at this point. 1841 inc_recompile_count = maybe_prior_trap; 1842 } 1843 } else { 1844 // For reasons which are not recorded per-bytecode, we simply 1845 // force recompiles unconditionally. 1846 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1847 make_not_entrant = true; 1848 } 1849 1850 // Go back to the compiler if there are too many traps in this method. 1851 if (this_trap_count >= per_method_trap_limit(reason)) { 1852 // If there are too many traps in this method, force a recompile. 1853 // This will allow the compiler to see the limit overflow, and 1854 // take corrective action, if possible. 1855 // (This condition is an unlikely backstop only, because the 1856 // PerBytecodeTrapLimit is more likely to take effect first, 1857 // if it is applicable.) 1858 make_not_entrant = true; 1859 } 1860 1861 // Here's more hysteresis: If there has been a recompile at 1862 // this trap point already, run the method in the interpreter 1863 // for a while to exercise it more thoroughly. 1864 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1865 reprofile = true; 1866 } 1867 } 1868 1869 // Take requested actions on the method: 1870 1871 // Recompile 1872 if (make_not_entrant) { 1873 if (!nm->make_not_entrant()) { 1874 return; // the call did not change nmethod's state 1875 } 1876 1877 if (pdata != NULL) { 1878 // Record the recompilation event, if any. 1879 int tstate0 = pdata->trap_state(); 1880 int tstate1 = trap_state_set_recompiled(tstate0, true); 1881 if (tstate1 != tstate0) 1882 pdata->set_trap_state(tstate1); 1883 } 1884 1885 #if INCLUDE_RTM_OPT 1886 // Restart collecting RTM locking abort statistic if the method 1887 // is recompiled for a reason other than RTM state change. 1888 // Assume that in new recompiled code the statistic could be different, 1889 // for example, due to different inlining. 1890 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1891 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 1892 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1893 } 1894 #endif 1895 // For code aging we count traps separately here, using make_not_entrant() 1896 // as a guard against simultaneous deopts in multiple threads. 1897 if (reason == Reason_tenured && trap_mdo != NULL) { 1898 trap_mdo->inc_tenure_traps(); 1899 } 1900 } 1901 1902 if (inc_recompile_count) { 1903 trap_mdo->inc_overflow_recompile_count(); 1904 if ((uint)trap_mdo->overflow_recompile_count() > 1905 (uint)PerBytecodeRecompilationCutoff) { 1906 // Give up on the method containing the bad BCI. 1907 if (trap_method() == nm->method()) { 1908 make_not_compilable = true; 1909 } else { 1910 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1911 // But give grace to the enclosing nm->method(). 1912 } 1913 } 1914 } 1915 1916 // Reprofile 1917 if (reprofile) { 1918 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1919 } 1920 1921 // Give up compiling 1922 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1923 assert(make_not_entrant, "consistent"); 1924 nm->method()->set_not_compilable(CompLevel_full_optimization); 1925 } 1926 1927 } // Free marked resources 1928 1929 } 1930 JRT_END 1931 1932 ProfileData* 1933 Deoptimization::query_update_method_data(MethodData* trap_mdo, 1934 int trap_bci, 1935 Deoptimization::DeoptReason reason, 1936 bool update_total_trap_count, 1937 #if INCLUDE_JVMCI 1938 bool is_osr, 1939 #endif 1940 Method* compiled_method, 1941 //outputs: 1942 uint& ret_this_trap_count, 1943 bool& ret_maybe_prior_trap, 1944 bool& ret_maybe_prior_recompile) { 1945 bool maybe_prior_trap = false; 1946 bool maybe_prior_recompile = false; 1947 uint this_trap_count = 0; 1948 if (update_total_trap_count) { 1949 uint idx = reason; 1950 #if INCLUDE_JVMCI 1951 if (is_osr) { 1952 idx += Reason_LIMIT; 1953 } 1954 #endif 1955 uint prior_trap_count = trap_mdo->trap_count(idx); 1956 this_trap_count = trap_mdo->inc_trap_count(idx); 1957 1958 // If the runtime cannot find a place to store trap history, 1959 // it is estimated based on the general condition of the method. 1960 // If the method has ever been recompiled, or has ever incurred 1961 // a trap with the present reason , then this BCI is assumed 1962 // (pessimistically) to be the culprit. 1963 maybe_prior_trap = (prior_trap_count != 0); 1964 maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1965 } 1966 ProfileData* pdata = NULL; 1967 1968 1969 // For reasons which are recorded per bytecode, we check per-BCI data. 1970 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1971 assert(per_bc_reason != Reason_none || update_total_trap_count, "must be"); 1972 if (per_bc_reason != Reason_none) { 1973 // Find the profile data for this BCI. If there isn't one, 1974 // try to allocate one from the MDO's set of spares. 1975 // This will let us detect a repeated trap at this point. 1976 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 1977 1978 if (pdata != NULL) { 1979 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 1980 if (LogCompilation && xtty != NULL) { 1981 ttyLocker ttyl; 1982 // no more room for speculative traps in this MDO 1983 xtty->elem("speculative_traps_oom"); 1984 } 1985 } 1986 // Query the trap state of this profile datum. 1987 int tstate0 = pdata->trap_state(); 1988 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1989 maybe_prior_trap = false; 1990 if (!trap_state_is_recompiled(tstate0)) 1991 maybe_prior_recompile = false; 1992 1993 // Update the trap state of this profile datum. 1994 int tstate1 = tstate0; 1995 // Record the reason. 1996 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1997 // Store the updated state on the MDO, for next time. 1998 if (tstate1 != tstate0) 1999 pdata->set_trap_state(tstate1); 2000 } else { 2001 if (LogCompilation && xtty != NULL) { 2002 ttyLocker ttyl; 2003 // Missing MDP? Leave a small complaint in the log. 2004 xtty->elem("missing_mdp bci='%d'", trap_bci); 2005 } 2006 } 2007 } 2008 2009 // Return results: 2010 ret_this_trap_count = this_trap_count; 2011 ret_maybe_prior_trap = maybe_prior_trap; 2012 ret_maybe_prior_recompile = maybe_prior_recompile; 2013 return pdata; 2014 } 2015 2016 void 2017 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2018 ResourceMark rm; 2019 // Ignored outputs: 2020 uint ignore_this_trap_count; 2021 bool ignore_maybe_prior_trap; 2022 bool ignore_maybe_prior_recompile; 2023 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 2024 // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts 2025 bool update_total_counts = JVMCI_ONLY(false) NOT_JVMCI(true); 2026 query_update_method_data(trap_mdo, trap_bci, 2027 (DeoptReason)reason, 2028 update_total_counts, 2029 #if INCLUDE_JVMCI 2030 false, 2031 #endif 2032 NULL, 2033 ignore_this_trap_count, 2034 ignore_maybe_prior_trap, 2035 ignore_maybe_prior_recompile); 2036 } 2037 2038 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) { 2039 if (TraceDeoptimization) { 2040 tty->print("Uncommon trap "); 2041 } 2042 // Still in Java no safepoints 2043 { 2044 // This enters VM and may safepoint 2045 uncommon_trap_inner(thread, trap_request); 2046 } 2047 return fetch_unroll_info_helper(thread, exec_mode); 2048 } 2049 2050 // Local derived constants. 2051 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 2052 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 2053 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 2054 2055 //---------------------------trap_state_reason--------------------------------- 2056 Deoptimization::DeoptReason 2057 Deoptimization::trap_state_reason(int trap_state) { 2058 // This assert provides the link between the width of DataLayout::trap_bits 2059 // and the encoding of "recorded" reasons. It ensures there are enough 2060 // bits to store all needed reasons in the per-BCI MDO profile. 2061 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2062 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2063 trap_state -= recompile_bit; 2064 if (trap_state == DS_REASON_MASK) { 2065 return Reason_many; 2066 } else { 2067 assert((int)Reason_none == 0, "state=0 => Reason_none"); 2068 return (DeoptReason)trap_state; 2069 } 2070 } 2071 //-------------------------trap_state_has_reason------------------------------- 2072 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2073 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 2074 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 2075 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2076 trap_state -= recompile_bit; 2077 if (trap_state == DS_REASON_MASK) { 2078 return -1; // true, unspecifically (bottom of state lattice) 2079 } else if (trap_state == reason) { 2080 return 1; // true, definitely 2081 } else if (trap_state == 0) { 2082 return 0; // false, definitely (top of state lattice) 2083 } else { 2084 return 0; // false, definitely 2085 } 2086 } 2087 //-------------------------trap_state_add_reason------------------------------- 2088 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 2089 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 2090 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 2091 trap_state -= recompile_bit; 2092 if (trap_state == DS_REASON_MASK) { 2093 return trap_state + recompile_bit; // already at state lattice bottom 2094 } else if (trap_state == reason) { 2095 return trap_state + recompile_bit; // the condition is already true 2096 } else if (trap_state == 0) { 2097 return reason + recompile_bit; // no condition has yet been true 2098 } else { 2099 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 2100 } 2101 } 2102 //-----------------------trap_state_is_recompiled------------------------------ 2103 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2104 return (trap_state & DS_RECOMPILE_BIT) != 0; 2105 } 2106 //-----------------------trap_state_set_recompiled----------------------------- 2107 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 2108 if (z) return trap_state | DS_RECOMPILE_BIT; 2109 else return trap_state & ~DS_RECOMPILE_BIT; 2110 } 2111 //---------------------------format_trap_state--------------------------------- 2112 // This is used for debugging and diagnostics, including LogFile output. 2113 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2114 int trap_state) { 2115 assert(buflen > 0, "sanity"); 2116 DeoptReason reason = trap_state_reason(trap_state); 2117 bool recomp_flag = trap_state_is_recompiled(trap_state); 2118 // Re-encode the state from its decoded components. 2119 int decoded_state = 0; 2120 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 2121 decoded_state = trap_state_add_reason(decoded_state, reason); 2122 if (recomp_flag) 2123 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 2124 // If the state re-encodes properly, format it symbolically. 2125 // Because this routine is used for debugging and diagnostics, 2126 // be robust even if the state is a strange value. 2127 size_t len; 2128 if (decoded_state != trap_state) { 2129 // Random buggy state that doesn't decode?? 2130 len = jio_snprintf(buf, buflen, "#%d", trap_state); 2131 } else { 2132 len = jio_snprintf(buf, buflen, "%s%s", 2133 trap_reason_name(reason), 2134 recomp_flag ? " recompiled" : ""); 2135 } 2136 return buf; 2137 } 2138 2139 2140 //--------------------------------statics-------------------------------------- 2141 const char* Deoptimization::_trap_reason_name[] = { 2142 // Note: Keep this in sync. with enum DeoptReason. 2143 "none", 2144 "null_check", 2145 "null_assert" JVMCI_ONLY("_or_unreached0"), 2146 "range_check", 2147 "class_check", 2148 "array_check", 2149 "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"), 2150 "bimorphic" JVMCI_ONLY("_or_optimized_type_check"), 2151 "unloaded", 2152 "uninitialized", 2153 "unreached", 2154 "unhandled", 2155 "constraint", 2156 "div0_check", 2157 "age", 2158 "predicate", 2159 "loop_limit_check", 2160 "speculate_class_check", 2161 "speculate_null_check", 2162 "rtm_state_change", 2163 "unstable_if", 2164 "unstable_fused_if", 2165 #if INCLUDE_JVMCI 2166 "aliasing", 2167 "transfer_to_interpreter", 2168 "not_compiled_exception_handler", 2169 "unresolved", 2170 "jsr_mismatch", 2171 #endif 2172 "tenured" 2173 }; 2174 const char* Deoptimization::_trap_action_name[] = { 2175 // Note: Keep this in sync. with enum DeoptAction. 2176 "none", 2177 "maybe_recompile", 2178 "reinterpret", 2179 "make_not_entrant", 2180 "make_not_compilable" 2181 }; 2182 2183 const char* Deoptimization::trap_reason_name(int reason) { 2184 // Check that every reason has a name 2185 STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT); 2186 2187 if (reason == Reason_many) return "many"; 2188 if ((uint)reason < Reason_LIMIT) 2189 return _trap_reason_name[reason]; 2190 static char buf[20]; 2191 sprintf(buf, "reason%d", reason); 2192 return buf; 2193 } 2194 const char* Deoptimization::trap_action_name(int action) { 2195 // Check that every action has a name 2196 STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT); 2197 2198 if ((uint)action < Action_LIMIT) 2199 return _trap_action_name[action]; 2200 static char buf[20]; 2201 sprintf(buf, "action%d", action); 2202 return buf; 2203 } 2204 2205 // This is used for debugging and diagnostics, including LogFile output. 2206 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 2207 int trap_request) { 2208 jint unloaded_class_index = trap_request_index(trap_request); 2209 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 2210 const char* action = trap_action_name(trap_request_action(trap_request)); 2211 #if INCLUDE_JVMCI 2212 int debug_id = trap_request_debug_id(trap_request); 2213 #endif 2214 size_t len; 2215 if (unloaded_class_index < 0) { 2216 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"), 2217 reason, action 2218 #if INCLUDE_JVMCI 2219 ,debug_id 2220 #endif 2221 ); 2222 } else { 2223 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"), 2224 reason, action, unloaded_class_index 2225 #if INCLUDE_JVMCI 2226 ,debug_id 2227 #endif 2228 ); 2229 } 2230 return buf; 2231 } 2232 2233 juint Deoptimization::_deoptimization_hist 2234 [Deoptimization::Reason_LIMIT] 2235 [1 + Deoptimization::Action_LIMIT] 2236 [Deoptimization::BC_CASE_LIMIT] 2237 = {0}; 2238 2239 enum { 2240 LSB_BITS = 8, 2241 LSB_MASK = right_n_bits(LSB_BITS) 2242 }; 2243 2244 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2245 Bytecodes::Code bc) { 2246 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2247 assert(action >= 0 && action < Action_LIMIT, "oob"); 2248 _deoptimization_hist[Reason_none][0][0] += 1; // total 2249 _deoptimization_hist[reason][0][0] += 1; // per-reason total 2250 juint* cases = _deoptimization_hist[reason][1+action]; 2251 juint* bc_counter_addr = NULL; 2252 juint bc_counter = 0; 2253 // Look for an unused counter, or an exact match to this BC. 2254 if (bc != Bytecodes::_illegal) { 2255 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2256 juint* counter_addr = &cases[bc_case]; 2257 juint counter = *counter_addr; 2258 if ((counter == 0 && bc_counter_addr == NULL) 2259 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 2260 // this counter is either free or is already devoted to this BC 2261 bc_counter_addr = counter_addr; 2262 bc_counter = counter | bc; 2263 } 2264 } 2265 } 2266 if (bc_counter_addr == NULL) { 2267 // Overflow, or no given bytecode. 2268 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 2269 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 2270 } 2271 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 2272 } 2273 2274 jint Deoptimization::total_deoptimization_count() { 2275 return _deoptimization_hist[Reason_none][0][0]; 2276 } 2277 2278 jint Deoptimization::deoptimization_count(DeoptReason reason) { 2279 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 2280 return _deoptimization_hist[reason][0][0]; 2281 } 2282 2283 void Deoptimization::print_statistics() { 2284 juint total = total_deoptimization_count(); 2285 juint account = total; 2286 if (total != 0) { 2287 ttyLocker ttyl; 2288 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 2289 tty->print_cr("Deoptimization traps recorded:"); 2290 #define PRINT_STAT_LINE(name, r) \ 2291 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2292 PRINT_STAT_LINE("total", total); 2293 // For each non-zero entry in the histogram, print the reason, 2294 // the action, and (if specifically known) the type of bytecode. 2295 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2296 for (int action = 0; action < Action_LIMIT; action++) { 2297 juint* cases = _deoptimization_hist[reason][1+action]; 2298 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2299 juint counter = cases[bc_case]; 2300 if (counter != 0) { 2301 char name[1*K]; 2302 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2303 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2304 bc = Bytecodes::_illegal; 2305 sprintf(name, "%s/%s/%s", 2306 trap_reason_name(reason), 2307 trap_action_name(action), 2308 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2309 juint r = counter >> LSB_BITS; 2310 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2311 account -= r; 2312 } 2313 } 2314 } 2315 } 2316 if (account != 0) { 2317 PRINT_STAT_LINE("unaccounted", account); 2318 } 2319 #undef PRINT_STAT_LINE 2320 if (xtty != NULL) xtty->tail("statistics"); 2321 } 2322 } 2323 #else // COMPILER2 || SHARK || INCLUDE_JVMCI 2324 2325 2326 // Stubs for C1 only system. 2327 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2328 return false; 2329 } 2330 2331 const char* Deoptimization::trap_reason_name(int reason) { 2332 return "unknown"; 2333 } 2334 2335 void Deoptimization::print_statistics() { 2336 // no output 2337 } 2338 2339 void 2340 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2341 // no udpate 2342 } 2343 2344 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2345 return 0; 2346 } 2347 2348 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2349 Bytecodes::Code bc) { 2350 // no update 2351 } 2352 2353 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2354 int trap_state) { 2355 jio_snprintf(buf, buflen, "#%d", trap_state); 2356 return buf; 2357 } 2358 2359 #endif // COMPILER2 || SHARK || INCLUDE_JVMCI