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