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