1 /* 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "code/debugInfoRec.hpp" 28 #include "code/nmethod.hpp" 29 #include "code/pcDesc.hpp" 30 #include "code/scopeDesc.hpp" 31 #include "interpreter/bytecode.hpp" 32 #include "interpreter/interpreter.hpp" 33 #include "interpreter/oopMapCache.hpp" 34 #include "memory/allocation.inline.hpp" 35 #include "memory/oopFactory.hpp" 36 #include "memory/resourceArea.hpp" 37 #include "oops/methodOop.hpp" 38 #include "oops/oop.inline.hpp" 39 #include "prims/jvmtiThreadState.hpp" 40 #include "runtime/biasedLocking.hpp" 41 #include "runtime/compilationPolicy.hpp" 42 #include "runtime/deoptimization.hpp" 43 #include "runtime/interfaceSupport.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 #include "runtime/signature.hpp" 46 #include "runtime/stubRoutines.hpp" 47 #include "runtime/thread.hpp" 48 #include "runtime/vframe.hpp" 49 #include "runtime/vframeArray.hpp" 50 #include "runtime/vframe_hp.hpp" 51 #include "utilities/events.hpp" 52 #include "utilities/xmlstream.hpp" 53 #ifdef TARGET_ARCH_x86 54 # include "vmreg_x86.inline.hpp" 55 #endif 56 #ifdef TARGET_ARCH_sparc 57 # include "vmreg_sparc.inline.hpp" 58 #endif 59 #ifdef TARGET_ARCH_zero 60 # include "vmreg_zero.inline.hpp" 61 #endif 62 #ifdef TARGET_ARCH_arm 63 # include "vmreg_arm.inline.hpp" 64 #endif 65 #ifdef TARGET_ARCH_ppc 66 # include "vmreg_ppc.inline.hpp" 67 #endif 68 #ifdef COMPILER2 69 #ifdef TARGET_ARCH_MODEL_x86_32 70 # include "adfiles/ad_x86_32.hpp" 71 #endif 72 #ifdef TARGET_ARCH_MODEL_x86_64 73 # include "adfiles/ad_x86_64.hpp" 74 #endif 75 #ifdef TARGET_ARCH_MODEL_sparc 76 # include "adfiles/ad_sparc.hpp" 77 #endif 78 #ifdef TARGET_ARCH_MODEL_zero 79 # include "adfiles/ad_zero.hpp" 80 #endif 81 #ifdef TARGET_ARCH_MODEL_arm 82 # include "adfiles/ad_arm.hpp" 83 #endif 84 #ifdef TARGET_ARCH_MODEL_ppc 85 # include "adfiles/ad_ppc.hpp" 86 #endif 87 #endif 88 89 bool DeoptimizationMarker::_is_active = false; 90 91 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 92 int caller_adjustment, 93 int caller_actual_parameters, 94 int number_of_frames, 95 intptr_t* frame_sizes, 96 address* frame_pcs, 97 BasicType return_type) { 98 _size_of_deoptimized_frame = size_of_deoptimized_frame; 99 _caller_adjustment = caller_adjustment; 100 _caller_actual_parameters = caller_actual_parameters; 101 _number_of_frames = number_of_frames; 102 _frame_sizes = frame_sizes; 103 _frame_pcs = frame_pcs; 104 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); 105 _return_type = return_type; 106 _initial_fp = 0; 107 // PD (x86 only) 108 _counter_temp = 0; 109 _unpack_kind = 0; 110 _sender_sp_temp = 0; 111 112 _total_frame_sizes = size_of_frames(); 113 } 114 115 116 Deoptimization::UnrollBlock::~UnrollBlock() { 117 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 118 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 119 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 120 } 121 122 123 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 124 assert(register_number < RegisterMap::reg_count, "checking register number"); 125 return &_register_block[register_number * 2]; 126 } 127 128 129 130 int Deoptimization::UnrollBlock::size_of_frames() const { 131 // Acount first for the adjustment of the initial frame 132 int result = _caller_adjustment; 133 for (int index = 0; index < number_of_frames(); index++) { 134 result += frame_sizes()[index]; 135 } 136 return result; 137 } 138 139 140 void Deoptimization::UnrollBlock::print() { 141 ttyLocker ttyl; 142 tty->print_cr("UnrollBlock"); 143 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 144 tty->print( " frame_sizes: "); 145 for (int index = 0; index < number_of_frames(); index++) { 146 tty->print("%d ", frame_sizes()[index]); 147 } 148 tty->cr(); 149 } 150 151 152 // In order to make fetch_unroll_info work properly with escape 153 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 154 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 155 // of previously eliminated objects occurs in realloc_objects, which is 156 // called from the method fetch_unroll_info_helper below. 157 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 158 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 159 // but makes the entry a little slower. There is however a little dance we have to 160 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 161 162 // fetch_unroll_info() is called at the beginning of the deoptimization 163 // handler. Note this fact before we start generating temporary frames 164 // that can confuse an asynchronous stack walker. This counter is 165 // decremented at the end of unpack_frames(). 166 thread->inc_in_deopt_handler(); 167 168 return fetch_unroll_info_helper(thread); 169 JRT_END 170 171 172 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 173 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 174 175 // Note: there is a safepoint safety issue here. No matter whether we enter 176 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 177 // the vframeArray is created. 178 // 179 180 // Allocate our special deoptimization ResourceMark 181 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 182 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 183 thread->set_deopt_mark(dmark); 184 185 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 186 RegisterMap map(thread, true); 187 RegisterMap dummy_map(thread, false); 188 // Now get the deoptee with a valid map 189 frame deoptee = stub_frame.sender(&map); 190 // Set the deoptee nmethod 191 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 192 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 193 194 if (VerifyStack) { 195 thread->validate_frame_layout(); 196 } 197 198 // Create a growable array of VFrames where each VFrame represents an inlined 199 // Java frame. This storage is allocated with the usual system arena. 200 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 201 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 202 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 203 while (!vf->is_top()) { 204 assert(vf->is_compiled_frame(), "Wrong frame type"); 205 chunk->push(compiledVFrame::cast(vf)); 206 vf = vf->sender(); 207 } 208 assert(vf->is_compiled_frame(), "Wrong frame type"); 209 chunk->push(compiledVFrame::cast(vf)); 210 211 #ifdef COMPILER2 212 // Reallocate the non-escaping objects and restore their fields. Then 213 // relock objects if synchronization on them was eliminated. 214 if (DoEscapeAnalysis) { 215 if (EliminateAllocations) { 216 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 217 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 218 219 // The flag return_oop() indicates call sites which return oop 220 // in compiled code. Such sites include java method calls, 221 // runtime calls (for example, used to allocate new objects/arrays 222 // on slow code path) and any other calls generated in compiled code. 223 // It is not guaranteed that we can get such information here only 224 // by analyzing bytecode in deoptimized frames. This is why this flag 225 // is set during method compilation (see Compile::Process_OopMap_Node()). 226 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 227 Handle return_value; 228 if (save_oop_result) { 229 // Reallocation may trigger GC. If deoptimization happened on return from 230 // call which returns oop we need to save it since it is not in oopmap. 231 oop result = deoptee.saved_oop_result(&map); 232 assert(result == NULL || result->is_oop(), "must be oop"); 233 return_value = Handle(thread, result); 234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 235 if (TraceDeoptimization) { 236 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); 237 } 238 } 239 bool reallocated = false; 240 if (objects != NULL) { 241 JRT_BLOCK 242 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 243 JRT_END 244 } 245 if (reallocated) { 246 reassign_fields(&deoptee, &map, objects); 247 #ifndef PRODUCT 248 if (TraceDeoptimization) { 249 ttyLocker ttyl; 250 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 251 print_objects(objects); 252 } 253 #endif 254 } 255 if (save_oop_result) { 256 // Restore result. 257 deoptee.set_saved_oop_result(&map, return_value()); 258 } 259 } 260 if (EliminateLocks) { 261 #ifndef PRODUCT 262 bool first = true; 263 #endif 264 for (int i = 0; i < chunk->length(); i++) { 265 compiledVFrame* cvf = chunk->at(i); 266 assert (cvf->scope() != NULL,"expect only compiled java frames"); 267 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 268 if (monitors->is_nonempty()) { 269 relock_objects(monitors, thread); 270 #ifndef PRODUCT 271 if (TraceDeoptimization) { 272 ttyLocker ttyl; 273 for (int j = 0; j < monitors->length(); j++) { 274 MonitorInfo* mi = monitors->at(j); 275 if (mi->eliminated()) { 276 if (first) { 277 first = false; 278 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 279 } 280 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 281 } 282 } 283 } 284 #endif 285 } 286 } 287 } 288 } 289 #endif // COMPILER2 290 // Ensure that no safepoint is taken after pointers have been stored 291 // in fields of rematerialized objects. If a safepoint occurs from here on 292 // out the java state residing in the vframeArray will be missed. 293 No_Safepoint_Verifier no_safepoint; 294 295 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 296 297 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 298 thread->set_vframe_array_head(array); 299 300 // Now that the vframeArray has been created if we have any deferred local writes 301 // added by jvmti then we can free up that structure as the data is now in the 302 // vframeArray 303 304 if (thread->deferred_locals() != NULL) { 305 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 306 int i = 0; 307 do { 308 // Because of inlining we could have multiple vframes for a single frame 309 // and several of the vframes could have deferred writes. Find them all. 310 if (list->at(i)->id() == array->original().id()) { 311 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 312 list->remove_at(i); 313 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 314 delete dlv; 315 } else { 316 i++; 317 } 318 } while ( i < list->length() ); 319 if (list->length() == 0) { 320 thread->set_deferred_locals(NULL); 321 // free the list and elements back to C heap. 322 delete list; 323 } 324 325 } 326 327 #ifndef SHARK 328 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 329 CodeBlob* cb = stub_frame.cb(); 330 // Verify we have the right vframeArray 331 assert(cb->frame_size() >= 0, "Unexpected frame size"); 332 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 333 334 // If the deopt call site is a MethodHandle invoke call site we have 335 // to adjust the unpack_sp. 336 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 337 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 338 unpack_sp = deoptee.unextended_sp(); 339 340 #ifdef ASSERT 341 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 342 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); 343 #endif 344 #else 345 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 346 #endif // !SHARK 347 348 // This is a guarantee instead of an assert because if vframe doesn't match 349 // we will unpack the wrong deoptimized frame and wind up in strange places 350 // where it will be very difficult to figure out what went wrong. Better 351 // to die an early death here than some very obscure death later when the 352 // trail is cold. 353 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 354 // in that it will fail to detect a problem when there is one. This needs 355 // more work in tiger timeframe. 356 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 357 358 int number_of_frames = array->frames(); 359 360 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 361 // virtual activation, which is the reverse of the elements in the vframes array. 362 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 363 // +1 because we always have an interpreter return address for the final slot. 364 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 365 int callee_parameters = 0; 366 int callee_locals = 0; 367 int popframe_extra_args = 0; 368 // Create an interpreter return address for the stub to use as its return 369 // address so the skeletal frames are perfectly walkable 370 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 371 372 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 373 // activation be put back on the expression stack of the caller for reexecution 374 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 375 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 376 } 377 378 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 379 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 380 // than simply use array->sender.pc(). This requires us to walk the current set of frames 381 // 382 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 383 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 384 385 // It's possible that the number of paramters at the call site is 386 // different than number of arguments in the callee when method 387 // handles are used. If the caller is interpreted get the real 388 // value so that the proper amount of space can be added to it's 389 // frame. 390 int caller_actual_parameters = callee_parameters; 391 if (deopt_sender.is_interpreted_frame()) { 392 methodHandle method = deopt_sender.interpreter_frame_method(); 393 Bytecode_invoke cur = Bytecode_invoke_check(method, 394 deopt_sender.interpreter_frame_bci()); 395 Symbol* signature = method->constants()->signature_ref_at(cur.index()); 396 ArgumentSizeComputer asc(signature); 397 caller_actual_parameters = asc.size() + (cur.has_receiver() ? 1 : 0); 398 } 399 400 // 401 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 402 // frame_sizes/frame_pcs[1] next oldest frame (int) 403 // frame_sizes/frame_pcs[n] youngest frame (int) 404 // 405 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 406 // owns the space for the return address to it's caller). Confusing ain't it. 407 // 408 // The vframe array can address vframes with indices running from 409 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 410 // When we create the skeletal frames we need the oldest frame to be in the zero slot 411 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 412 // so things look a little strange in this loop. 413 // 414 for (int index = 0; index < array->frames(); index++ ) { 415 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 416 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 417 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 418 int caller_parms = callee_parameters; 419 if (index == array->frames() - 1) { 420 // Use the value from the interpreted caller 421 caller_parms = caller_actual_parameters; 422 } 423 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms, 424 callee_parameters, 425 callee_locals, 426 index == 0, 427 popframe_extra_args); 428 // This pc doesn't have to be perfect just good enough to identify the frame 429 // as interpreted so the skeleton frame will be walkable 430 // The correct pc will be set when the skeleton frame is completely filled out 431 // The final pc we store in the loop is wrong and will be overwritten below 432 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 433 434 callee_parameters = array->element(index)->method()->size_of_parameters(); 435 callee_locals = array->element(index)->method()->max_locals(); 436 popframe_extra_args = 0; 437 } 438 439 // Compute whether the root vframe returns a float or double value. 440 BasicType return_type; 441 { 442 HandleMark hm; 443 methodHandle method(thread, array->element(0)->method()); 444 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 445 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 446 } 447 448 // Compute information for handling adapters and adjusting the frame size of the caller. 449 int caller_adjustment = 0; 450 451 // Compute the amount the oldest interpreter frame will have to adjust 452 // its caller's stack by. If the caller is a compiled frame then 453 // we pretend that the callee has no parameters so that the 454 // extension counts for the full amount of locals and not just 455 // locals-parms. This is because without a c2i adapter the parm 456 // area as created by the compiled frame will not be usable by 457 // the interpreter. (Depending on the calling convention there 458 // may not even be enough space). 459 460 // QQQ I'd rather see this pushed down into last_frame_adjust 461 // and have it take the sender (aka caller). 462 463 if (deopt_sender.is_compiled_frame()) { 464 caller_adjustment = last_frame_adjust(0, callee_locals); 465 } else if (callee_locals > caller_actual_parameters) { 466 // The caller frame may need extending to accommodate 467 // non-parameter locals of the first unpacked interpreted frame. 468 // Compute that adjustment. 469 caller_adjustment = last_frame_adjust(caller_actual_parameters, callee_locals); 470 } 471 472 // If the sender is deoptimized the we must retrieve the address of the handler 473 // since the frame will "magically" show the original pc before the deopt 474 // and we'd undo the deopt. 475 476 frame_pcs[0] = deopt_sender.raw_pc(); 477 478 #ifndef SHARK 479 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 480 #endif // SHARK 481 482 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 483 caller_adjustment * BytesPerWord, 484 caller_actual_parameters, 485 number_of_frames, 486 frame_sizes, 487 frame_pcs, 488 return_type); 489 // On some platforms, we need a way to pass some platform dependent 490 // information to the unpacking code so the skeletal frames come out 491 // correct (initial fp value, unextended sp, ...) 492 info->set_initial_fp((intptr_t) array->sender().initial_deoptimization_info()); 493 494 if (array->frames() > 1) { 495 if (VerifyStack && TraceDeoptimization) { 496 tty->print_cr("Deoptimizing method containing inlining"); 497 } 498 } 499 500 array->set_unroll_block(info); 501 return info; 502 } 503 504 // Called to cleanup deoptimization data structures in normal case 505 // after unpacking to stack and when stack overflow error occurs 506 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 507 vframeArray *array) { 508 509 // Get array if coming from exception 510 if (array == NULL) { 511 array = thread->vframe_array_head(); 512 } 513 thread->set_vframe_array_head(NULL); 514 515 // Free the previous UnrollBlock 516 vframeArray* old_array = thread->vframe_array_last(); 517 thread->set_vframe_array_last(array); 518 519 if (old_array != NULL) { 520 UnrollBlock* old_info = old_array->unroll_block(); 521 old_array->set_unroll_block(NULL); 522 delete old_info; 523 delete old_array; 524 } 525 526 // Deallocate any resource creating in this routine and any ResourceObjs allocated 527 // inside the vframeArray (StackValueCollections) 528 529 delete thread->deopt_mark(); 530 thread->set_deopt_mark(NULL); 531 thread->set_deopt_nmethod(NULL); 532 533 534 if (JvmtiExport::can_pop_frame()) { 535 #ifndef CC_INTERP 536 // Regardless of whether we entered this routine with the pending 537 // popframe condition bit set, we should always clear it now 538 thread->clear_popframe_condition(); 539 #else 540 // C++ interpeter will clear has_pending_popframe when it enters 541 // with method_resume. For deopt_resume2 we clear it now. 542 if (thread->popframe_forcing_deopt_reexecution()) 543 thread->clear_popframe_condition(); 544 #endif /* CC_INTERP */ 545 } 546 547 // unpack_frames() is called at the end of the deoptimization handler 548 // and (in C2) at the end of the uncommon trap handler. Note this fact 549 // so that an asynchronous stack walker can work again. This counter is 550 // incremented at the beginning of fetch_unroll_info() and (in C2) at 551 // the beginning of uncommon_trap(). 552 thread->dec_in_deopt_handler(); 553 } 554 555 556 // Return BasicType of value being returned 557 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 558 559 // We are already active int he special DeoptResourceMark any ResourceObj's we 560 // allocate will be freed at the end of the routine. 561 562 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 563 // but makes the entry a little slower. There is however a little dance we have to 564 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 565 ResetNoHandleMark rnhm; // No-op in release/product versions 566 HandleMark hm; 567 568 frame stub_frame = thread->last_frame(); 569 570 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 571 // must point to the vframeArray for the unpack frame. 572 vframeArray* array = thread->vframe_array_head(); 573 574 #ifndef PRODUCT 575 if (TraceDeoptimization) { 576 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 577 } 578 #endif 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 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 637 Symbol* signature = invoke.signature(); 638 ArgumentSizeComputer asc(signature); 639 cur_invoke_parameter_size = asc.size(); 640 if (cur_code != Bytecodes::_invokestatic) { 641 // Add in receiver 642 ++cur_invoke_parameter_size; 643 } 644 } 645 if (str.bci() < max_bci) { 646 Bytecodes::Code bc = str.next(); 647 if (bc >= 0) { 648 // The interpreter oop map generator reports results before 649 // the current bytecode has executed except in the case of 650 // calls. It seems to be hard to tell whether the compiler 651 // has emitted debug information matching the "state before" 652 // a given bytecode or the state after, so we try both 653 switch (cur_code) { 654 case Bytecodes::_invokevirtual: 655 case Bytecodes::_invokespecial: 656 case Bytecodes::_invokestatic: 657 case Bytecodes::_invokeinterface: 658 case Bytecodes::_athrow: 659 break; 660 default: { 661 InterpreterOopMap next_mask; 662 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 663 next_mask_expression_stack_size = next_mask.expression_stack_size(); 664 // Need to subtract off the size of the result type of 665 // the bytecode because this is not described in the 666 // debug info but returned to the interpreter in the TOS 667 // caching register 668 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 669 if (bytecode_result_type != T_ILLEGAL) { 670 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 671 } 672 assert(top_frame_expression_stack_adjustment >= 0, ""); 673 try_next_mask = true; 674 break; 675 } 676 } 677 } 678 } 679 680 // Verify stack depth and oops in frame 681 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 682 if (!( 683 /* SPARC */ 684 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 685 /* x86 */ 686 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 687 (try_next_mask && 688 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 689 top_frame_expression_stack_adjustment))) || 690 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 691 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 692 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 693 )) { 694 ttyLocker ttyl; 695 696 // Print out some information that will help us debug the problem 697 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 698 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 699 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 700 iframe->interpreter_frame_expression_stack_size()); 701 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 702 tty->print_cr(" try_next_mask = %d", try_next_mask); 703 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 704 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 705 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 706 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 707 tty->print_cr(" exec_mode = %d", exec_mode); 708 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 709 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 710 tty->print_cr(" Interpreted frames:"); 711 for (int k = 0; k < cur_array->frames(); k++) { 712 vframeArrayElement* el = cur_array->element(k); 713 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 714 } 715 cur_array->print_on_2(tty); 716 guarantee(false, "wrong number of expression stack elements during deopt"); 717 } 718 VerifyOopClosure verify; 719 iframe->oops_interpreted_do(&verify, &rm, false); 720 callee_size_of_parameters = mh->size_of_parameters(); 721 callee_max_locals = mh->max_locals(); 722 is_top_frame = false; 723 } 724 } 725 #endif /* !PRODUCT */ 726 727 728 return bt; 729 JRT_END 730 731 732 int Deoptimization::deoptimize_dependents() { 733 Threads::deoptimized_wrt_marked_nmethods(); 734 return 0; 735 } 736 737 738 #ifdef COMPILER2 739 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 740 Handle pending_exception(thread->pending_exception()); 741 const char* exception_file = thread->exception_file(); 742 int exception_line = thread->exception_line(); 743 thread->clear_pending_exception(); 744 745 for (int i = 0; i < objects->length(); i++) { 746 assert(objects->at(i)->is_object(), "invalid debug information"); 747 ObjectValue* sv = (ObjectValue*) objects->at(i); 748 749 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 750 oop obj = NULL; 751 752 if (k->oop_is_instance()) { 753 instanceKlass* ik = instanceKlass::cast(k()); 754 obj = ik->allocate_instance(CHECK_(false)); 755 } else if (k->oop_is_typeArray()) { 756 typeArrayKlass* ak = typeArrayKlass::cast(k()); 757 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 758 int len = sv->field_size() / type2size[ak->element_type()]; 759 obj = ak->allocate(len, CHECK_(false)); 760 } else if (k->oop_is_objArray()) { 761 objArrayKlass* ak = objArrayKlass::cast(k()); 762 obj = ak->allocate(sv->field_size(), CHECK_(false)); 763 } 764 765 assert(obj != NULL, "allocation failed"); 766 assert(sv->value().is_null(), "redundant reallocation"); 767 sv->set_value(obj); 768 } 769 770 if (pending_exception.not_null()) { 771 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 772 } 773 774 return true; 775 } 776 777 // This assumes that the fields are stored in ObjectValue in the same order 778 // they are yielded by do_nonstatic_fields. 779 class FieldReassigner: public FieldClosure { 780 frame* _fr; 781 RegisterMap* _reg_map; 782 ObjectValue* _sv; 783 instanceKlass* _ik; 784 oop _obj; 785 786 int _i; 787 public: 788 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 789 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 790 791 int i() const { return _i; } 792 793 794 void do_field(fieldDescriptor* fd) { 795 intptr_t val; 796 StackValue* value = 797 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 798 int offset = fd->offset(); 799 switch (fd->field_type()) { 800 case T_OBJECT: case T_ARRAY: 801 assert(value->type() == T_OBJECT, "Agreement."); 802 _obj->obj_field_put(offset, value->get_obj()()); 803 break; 804 805 case T_LONG: case T_DOUBLE: { 806 assert(value->type() == T_INT, "Agreement."); 807 StackValue* low = 808 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 809 #ifdef _LP64 810 jlong res = (jlong)low->get_int(); 811 #else 812 #ifdef SPARC 813 // For SPARC we have to swap high and low words. 814 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 815 #else 816 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 817 #endif //SPARC 818 #endif 819 _obj->long_field_put(offset, res); 820 break; 821 } 822 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 823 case T_INT: case T_FLOAT: // 4 bytes. 824 assert(value->type() == T_INT, "Agreement."); 825 val = value->get_int(); 826 _obj->int_field_put(offset, (jint)*((jint*)&val)); 827 break; 828 829 case T_SHORT: case T_CHAR: // 2 bytes 830 assert(value->type() == T_INT, "Agreement."); 831 val = value->get_int(); 832 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 833 break; 834 835 case T_BOOLEAN: case T_BYTE: // 1 byte 836 assert(value->type() == T_INT, "Agreement."); 837 val = value->get_int(); 838 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 839 break; 840 841 default: 842 ShouldNotReachHere(); 843 } 844 _i++; 845 } 846 }; 847 848 // restore elements of an eliminated type array 849 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 850 int index = 0; 851 intptr_t val; 852 853 for (int i = 0; i < sv->field_size(); i++) { 854 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 855 switch(type) { 856 case T_LONG: case T_DOUBLE: { 857 assert(value->type() == T_INT, "Agreement."); 858 StackValue* low = 859 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 860 #ifdef _LP64 861 jlong res = (jlong)low->get_int(); 862 #else 863 #ifdef SPARC 864 // For SPARC we have to swap high and low words. 865 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 866 #else 867 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 868 #endif //SPARC 869 #endif 870 obj->long_at_put(index, res); 871 break; 872 } 873 874 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 875 case T_INT: case T_FLOAT: // 4 bytes. 876 assert(value->type() == T_INT, "Agreement."); 877 val = value->get_int(); 878 obj->int_at_put(index, (jint)*((jint*)&val)); 879 break; 880 881 case T_SHORT: case T_CHAR: // 2 bytes 882 assert(value->type() == T_INT, "Agreement."); 883 val = value->get_int(); 884 obj->short_at_put(index, (jshort)*((jint*)&val)); 885 break; 886 887 case T_BOOLEAN: case T_BYTE: // 1 byte 888 assert(value->type() == T_INT, "Agreement."); 889 val = value->get_int(); 890 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 891 break; 892 893 default: 894 ShouldNotReachHere(); 895 } 896 index++; 897 } 898 } 899 900 901 // restore fields of an eliminated object array 902 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 903 for (int i = 0; i < sv->field_size(); i++) { 904 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 905 assert(value->type() == T_OBJECT, "object element expected"); 906 obj->obj_at_put(i, value->get_obj()()); 907 } 908 } 909 910 911 // restore fields of all eliminated objects and arrays 912 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 913 for (int i = 0; i < objects->length(); i++) { 914 ObjectValue* sv = (ObjectValue*) objects->at(i); 915 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 916 Handle obj = sv->value(); 917 assert(obj.not_null(), "reallocation was missed"); 918 919 if (k->oop_is_instance()) { 920 instanceKlass* ik = instanceKlass::cast(k()); 921 FieldReassigner reassign(fr, reg_map, sv, obj()); 922 ik->do_nonstatic_fields(&reassign); 923 } else if (k->oop_is_typeArray()) { 924 typeArrayKlass* ak = typeArrayKlass::cast(k()); 925 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 926 } else if (k->oop_is_objArray()) { 927 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 928 } 929 } 930 } 931 932 933 // relock objects for which synchronization was eliminated 934 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 935 for (int i = 0; i < monitors->length(); i++) { 936 MonitorInfo* mon_info = monitors->at(i); 937 if (mon_info->eliminated()) { 938 assert(mon_info->owner() != NULL, "reallocation was missed"); 939 Handle obj = Handle(mon_info->owner()); 940 markOop mark = obj->mark(); 941 if (UseBiasedLocking && mark->has_bias_pattern()) { 942 // New allocated objects may have the mark set to anonymously biased. 943 // Also the deoptimized method may called methods with synchronization 944 // where the thread-local object is bias locked to the current thread. 945 assert(mark->is_biased_anonymously() || 946 mark->biased_locker() == thread, "should be locked to current thread"); 947 // Reset mark word to unbiased prototype. 948 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 949 obj->set_mark(unbiased_prototype); 950 } 951 BasicLock* lock = mon_info->lock(); 952 ObjectSynchronizer::slow_enter(obj, lock, thread); 953 } 954 assert(mon_info->owner()->is_locked(), "object must be locked now"); 955 } 956 } 957 958 959 #ifndef PRODUCT 960 // print information about reallocated objects 961 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 962 fieldDescriptor fd; 963 964 for (int i = 0; i < objects->length(); i++) { 965 ObjectValue* sv = (ObjectValue*) objects->at(i); 966 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 967 Handle obj = sv->value(); 968 969 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 970 k->as_klassOop()->print_value(); 971 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 972 tty->cr(); 973 974 if (Verbose) { 975 k->oop_print_on(obj(), tty); 976 } 977 } 978 } 979 #endif 980 #endif // COMPILER2 981 982 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 983 984 #ifndef PRODUCT 985 if (TraceDeoptimization) { 986 ttyLocker ttyl; 987 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 988 fr.print_on(tty); 989 tty->print_cr(" Virtual frames (innermost first):"); 990 for (int index = 0; index < chunk->length(); index++) { 991 compiledVFrame* vf = chunk->at(index); 992 tty->print(" %2d - ", index); 993 vf->print_value(); 994 int bci = chunk->at(index)->raw_bci(); 995 const char* code_name; 996 if (bci == SynchronizationEntryBCI) { 997 code_name = "sync entry"; 998 } else { 999 Bytecodes::Code code = vf->method()->code_at(bci); 1000 code_name = Bytecodes::name(code); 1001 } 1002 tty->print(" - %s", code_name); 1003 tty->print_cr(" @ bci %d ", bci); 1004 if (Verbose) { 1005 vf->print(); 1006 tty->cr(); 1007 } 1008 } 1009 } 1010 #endif 1011 1012 // Register map for next frame (used for stack crawl). We capture 1013 // the state of the deopt'ing frame's caller. Thus if we need to 1014 // stuff a C2I adapter we can properly fill in the callee-save 1015 // register locations. 1016 frame caller = fr.sender(reg_map); 1017 int frame_size = caller.sp() - fr.sp(); 1018 1019 frame sender = caller; 1020 1021 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1022 // the vframeArray containing the unpacking information is allocated in the C heap. 1023 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1024 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1025 1026 // Compare the vframeArray to the collected vframes 1027 assert(array->structural_compare(thread, chunk), "just checking"); 1028 Events::log("# vframes = %d", (intptr_t)chunk->length()); 1029 1030 #ifndef PRODUCT 1031 if (TraceDeoptimization) { 1032 ttyLocker ttyl; 1033 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1034 } 1035 #endif // PRODUCT 1036 1037 return array; 1038 } 1039 1040 1041 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1042 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1043 for (int i = 0; i < monitors->length(); i++) { 1044 MonitorInfo* mon_info = monitors->at(i); 1045 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1046 objects_to_revoke->append(Handle(mon_info->owner())); 1047 } 1048 } 1049 } 1050 1051 1052 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1053 if (!UseBiasedLocking) { 1054 return; 1055 } 1056 1057 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1058 1059 // Unfortunately we don't have a RegisterMap available in most of 1060 // the places we want to call this routine so we need to walk the 1061 // stack again to update the register map. 1062 if (map == NULL || !map->update_map()) { 1063 StackFrameStream sfs(thread, true); 1064 bool found = false; 1065 while (!found && !sfs.is_done()) { 1066 frame* cur = sfs.current(); 1067 sfs.next(); 1068 found = cur->id() == fr.id(); 1069 } 1070 assert(found, "frame to be deoptimized not found on target thread's stack"); 1071 map = sfs.register_map(); 1072 } 1073 1074 vframe* vf = vframe::new_vframe(&fr, map, thread); 1075 compiledVFrame* cvf = compiledVFrame::cast(vf); 1076 // Revoke monitors' biases in all scopes 1077 while (!cvf->is_top()) { 1078 collect_monitors(cvf, objects_to_revoke); 1079 cvf = compiledVFrame::cast(cvf->sender()); 1080 } 1081 collect_monitors(cvf, objects_to_revoke); 1082 1083 if (SafepointSynchronize::is_at_safepoint()) { 1084 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1085 } else { 1086 BiasedLocking::revoke(objects_to_revoke); 1087 } 1088 } 1089 1090 1091 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1092 if (!UseBiasedLocking) { 1093 return; 1094 } 1095 1096 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1097 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1098 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1099 if (jt->has_last_Java_frame()) { 1100 StackFrameStream sfs(jt, true); 1101 while (!sfs.is_done()) { 1102 frame* cur = sfs.current(); 1103 if (cb->contains(cur->pc())) { 1104 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1105 compiledVFrame* cvf = compiledVFrame::cast(vf); 1106 // Revoke monitors' biases in all scopes 1107 while (!cvf->is_top()) { 1108 collect_monitors(cvf, objects_to_revoke); 1109 cvf = compiledVFrame::cast(cvf->sender()); 1110 } 1111 collect_monitors(cvf, objects_to_revoke); 1112 } 1113 sfs.next(); 1114 } 1115 } 1116 } 1117 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1118 } 1119 1120 1121 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1122 assert(fr.can_be_deoptimized(), "checking frame type"); 1123 1124 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1125 1126 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); 1127 1128 // Patch the nmethod so that when execution returns to it we will 1129 // deopt the execution state and return to the interpreter. 1130 fr.deoptimize(thread); 1131 } 1132 1133 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1134 // Deoptimize only if the frame comes from compile code. 1135 // Do not deoptimize the frame which is already patched 1136 // during the execution of the loops below. 1137 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1138 return; 1139 } 1140 ResourceMark rm; 1141 DeoptimizationMarker dm; 1142 if (UseBiasedLocking) { 1143 revoke_biases_of_monitors(thread, fr, map); 1144 } 1145 deoptimize_single_frame(thread, fr); 1146 1147 } 1148 1149 1150 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1151 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1152 "can only deoptimize other thread at a safepoint"); 1153 // Compute frame and register map based on thread and sp. 1154 RegisterMap reg_map(thread, UseBiasedLocking); 1155 frame fr = thread->last_frame(); 1156 while (fr.id() != id) { 1157 fr = fr.sender(®_map); 1158 } 1159 deoptimize(thread, fr, ®_map); 1160 } 1161 1162 1163 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1164 if (thread == Thread::current()) { 1165 Deoptimization::deoptimize_frame_internal(thread, id); 1166 } else { 1167 VM_DeoptimizeFrame deopt(thread, id); 1168 VMThread::execute(&deopt); 1169 } 1170 } 1171 1172 1173 // JVMTI PopFrame support 1174 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1175 { 1176 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1177 } 1178 JRT_END 1179 1180 1181 #if defined(COMPILER2) || defined(SHARK) 1182 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1183 // in case of an unresolved klass entry, load the class. 1184 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1185 klassOop tk = constant_pool->klass_at(index, CHECK); 1186 return; 1187 } 1188 1189 if (!constant_pool->tag_at(index).is_symbol()) return; 1190 1191 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1192 Symbol* symbol = constant_pool->symbol_at(index); 1193 1194 // class name? 1195 if (symbol->byte_at(0) != '(') { 1196 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1197 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1198 return; 1199 } 1200 1201 // then it must be a signature! 1202 ResourceMark rm(THREAD); 1203 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1204 if (ss.is_object()) { 1205 Symbol* class_name = ss.as_symbol(CHECK); 1206 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1207 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1208 } 1209 } 1210 } 1211 1212 1213 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1214 EXCEPTION_MARK; 1215 load_class_by_index(constant_pool, index, THREAD); 1216 if (HAS_PENDING_EXCEPTION) { 1217 // Exception happened during classloading. We ignore the exception here, since it 1218 // is going to be rethrown since the current activation is going to be deoptimzied and 1219 // the interpreter will re-execute the bytecode. 1220 CLEAR_PENDING_EXCEPTION; 1221 } 1222 } 1223 1224 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1225 HandleMark hm; 1226 1227 // uncommon_trap() is called at the beginning of the uncommon trap 1228 // handler. Note this fact before we start generating temporary frames 1229 // that can confuse an asynchronous stack walker. This counter is 1230 // decremented at the end of unpack_frames(). 1231 thread->inc_in_deopt_handler(); 1232 1233 // We need to update the map if we have biased locking. 1234 RegisterMap reg_map(thread, UseBiasedLocking); 1235 frame stub_frame = thread->last_frame(); 1236 frame fr = stub_frame.sender(®_map); 1237 // Make sure the calling nmethod is not getting deoptimized and removed 1238 // before we are done with it. 1239 nmethodLocker nl(fr.pc()); 1240 1241 { 1242 ResourceMark rm; 1243 1244 // Revoke biases of any monitors in the frame to ensure we can migrate them 1245 revoke_biases_of_monitors(thread, fr, ®_map); 1246 1247 DeoptReason reason = trap_request_reason(trap_request); 1248 DeoptAction action = trap_request_action(trap_request); 1249 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1250 1251 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); 1252 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1253 compiledVFrame* cvf = compiledVFrame::cast(vf); 1254 1255 nmethod* nm = cvf->code(); 1256 1257 ScopeDesc* trap_scope = cvf->scope(); 1258 methodHandle trap_method = trap_scope->method(); 1259 int trap_bci = trap_scope->bci(); 1260 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1261 1262 // Record this event in the histogram. 1263 gather_statistics(reason, action, trap_bc); 1264 1265 // Ensure that we can record deopt. history: 1266 bool create_if_missing = ProfileTraps; 1267 1268 methodDataHandle trap_mdo 1269 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1270 1271 // Print a bunch of diagnostics, if requested. 1272 if (TraceDeoptimization || LogCompilation) { 1273 ResourceMark rm; 1274 ttyLocker ttyl; 1275 char buf[100]; 1276 if (xtty != NULL) { 1277 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1278 os::current_thread_id(), 1279 format_trap_request(buf, sizeof(buf), trap_request)); 1280 nm->log_identity(xtty); 1281 } 1282 Symbol* class_name = NULL; 1283 bool unresolved = false; 1284 if (unloaded_class_index >= 0) { 1285 constantPoolHandle constants (THREAD, trap_method->constants()); 1286 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1287 class_name = constants->klass_name_at(unloaded_class_index); 1288 unresolved = true; 1289 if (xtty != NULL) 1290 xtty->print(" unresolved='1'"); 1291 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1292 class_name = constants->symbol_at(unloaded_class_index); 1293 } 1294 if (xtty != NULL) 1295 xtty->name(class_name); 1296 } 1297 if (xtty != NULL && trap_mdo.not_null()) { 1298 // Dump the relevant MDO state. 1299 // This is the deopt count for the current reason, any previous 1300 // reasons or recompiles seen at this point. 1301 int dcnt = trap_mdo->trap_count(reason); 1302 if (dcnt != 0) 1303 xtty->print(" count='%d'", dcnt); 1304 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1305 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1306 if (dos != 0) { 1307 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1308 if (trap_state_is_recompiled(dos)) { 1309 int recnt2 = trap_mdo->overflow_recompile_count(); 1310 if (recnt2 != 0) 1311 xtty->print(" recompiles2='%d'", recnt2); 1312 } 1313 } 1314 } 1315 if (xtty != NULL) { 1316 xtty->stamp(); 1317 xtty->end_head(); 1318 } 1319 if (TraceDeoptimization) { // make noise on the tty 1320 tty->print("Uncommon trap occurred in"); 1321 nm->method()->print_short_name(tty); 1322 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1323 fr.pc(), 1324 (int) os::current_thread_id(), 1325 trap_reason_name(reason), 1326 trap_action_name(action), 1327 unloaded_class_index); 1328 if (class_name != NULL) { 1329 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1330 class_name->print_symbol_on(tty); 1331 } 1332 tty->cr(); 1333 } 1334 if (xtty != NULL) { 1335 // Log the precise location of the trap. 1336 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1337 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1338 xtty->method(sd->method()); 1339 xtty->end_elem(); 1340 if (sd->is_top()) break; 1341 } 1342 xtty->tail("uncommon_trap"); 1343 } 1344 } 1345 // (End diagnostic printout.) 1346 1347 // Load class if necessary 1348 if (unloaded_class_index >= 0) { 1349 constantPoolHandle constants(THREAD, trap_method->constants()); 1350 load_class_by_index(constants, unloaded_class_index); 1351 } 1352 1353 // Flush the nmethod if necessary and desirable. 1354 // 1355 // We need to avoid situations where we are re-flushing the nmethod 1356 // because of a hot deoptimization site. Repeated flushes at the same 1357 // point need to be detected by the compiler and avoided. If the compiler 1358 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1359 // module must take measures to avoid an infinite cycle of recompilation 1360 // and deoptimization. There are several such measures: 1361 // 1362 // 1. If a recompilation is ordered a second time at some site X 1363 // and for the same reason R, the action is adjusted to 'reinterpret', 1364 // to give the interpreter time to exercise the method more thoroughly. 1365 // If this happens, the method's overflow_recompile_count is incremented. 1366 // 1367 // 2. If the compiler fails to reduce the deoptimization rate, then 1368 // the method's overflow_recompile_count will begin to exceed the set 1369 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1370 // is adjusted to 'make_not_compilable', and the method is abandoned 1371 // to the interpreter. This is a performance hit for hot methods, 1372 // but is better than a disastrous infinite cycle of recompilations. 1373 // (Actually, only the method containing the site X is abandoned.) 1374 // 1375 // 3. In parallel with the previous measures, if the total number of 1376 // recompilations of a method exceeds the much larger set limit 1377 // PerMethodRecompilationCutoff, the method is abandoned. 1378 // This should only happen if the method is very large and has 1379 // many "lukewarm" deoptimizations. The code which enforces this 1380 // limit is elsewhere (class nmethod, class methodOopDesc). 1381 // 1382 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1383 // to recompile at each bytecode independently of the per-BCI cutoff. 1384 // 1385 // The decision to update code is up to the compiler, and is encoded 1386 // in the Action_xxx code. If the compiler requests Action_none 1387 // no trap state is changed, no compiled code is changed, and the 1388 // computation suffers along in the interpreter. 1389 // 1390 // The other action codes specify various tactics for decompilation 1391 // and recompilation. Action_maybe_recompile is the loosest, and 1392 // allows the compiled code to stay around until enough traps are seen, 1393 // and until the compiler gets around to recompiling the trapping method. 1394 // 1395 // The other actions cause immediate removal of the present code. 1396 1397 bool update_trap_state = true; 1398 bool make_not_entrant = false; 1399 bool make_not_compilable = false; 1400 bool reprofile = false; 1401 switch (action) { 1402 case Action_none: 1403 // Keep the old code. 1404 update_trap_state = false; 1405 break; 1406 case Action_maybe_recompile: 1407 // Do not need to invalidate the present code, but we can 1408 // initiate another 1409 // Start compiler without (necessarily) invalidating the nmethod. 1410 // The system will tolerate the old code, but new code should be 1411 // generated when possible. 1412 break; 1413 case Action_reinterpret: 1414 // Go back into the interpreter for a while, and then consider 1415 // recompiling form scratch. 1416 make_not_entrant = true; 1417 // Reset invocation counter for outer most method. 1418 // This will allow the interpreter to exercise the bytecodes 1419 // for a while before recompiling. 1420 // By contrast, Action_make_not_entrant is immediate. 1421 // 1422 // Note that the compiler will track null_check, null_assert, 1423 // range_check, and class_check events and log them as if they 1424 // had been traps taken from compiled code. This will update 1425 // the MDO trap history so that the next compilation will 1426 // properly detect hot trap sites. 1427 reprofile = true; 1428 break; 1429 case Action_make_not_entrant: 1430 // Request immediate recompilation, and get rid of the old code. 1431 // Make them not entrant, so next time they are called they get 1432 // recompiled. Unloaded classes are loaded now so recompile before next 1433 // time they are called. Same for uninitialized. The interpreter will 1434 // link the missing class, if any. 1435 make_not_entrant = true; 1436 break; 1437 case Action_make_not_compilable: 1438 // Give up on compiling this method at all. 1439 make_not_entrant = true; 1440 make_not_compilable = true; 1441 break; 1442 default: 1443 ShouldNotReachHere(); 1444 } 1445 1446 // Setting +ProfileTraps fixes the following, on all platforms: 1447 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1448 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1449 // recompile relies on a methodDataOop to record heroic opt failures. 1450 1451 // Whether the interpreter is producing MDO data or not, we also need 1452 // to use the MDO to detect hot deoptimization points and control 1453 // aggressive optimization. 1454 bool inc_recompile_count = false; 1455 ProfileData* pdata = NULL; 1456 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1457 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1458 uint this_trap_count = 0; 1459 bool maybe_prior_trap = false; 1460 bool maybe_prior_recompile = false; 1461 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1462 //outputs: 1463 this_trap_count, 1464 maybe_prior_trap, 1465 maybe_prior_recompile); 1466 // Because the interpreter also counts null, div0, range, and class 1467 // checks, these traps from compiled code are double-counted. 1468 // This is harmless; it just means that the PerXTrapLimit values 1469 // are in effect a little smaller than they look. 1470 1471 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1472 if (per_bc_reason != Reason_none) { 1473 // Now take action based on the partially known per-BCI history. 1474 if (maybe_prior_trap 1475 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1476 // If there are too many traps at this BCI, force a recompile. 1477 // This will allow the compiler to see the limit overflow, and 1478 // take corrective action, if possible. The compiler generally 1479 // does not use the exact PerBytecodeTrapLimit value, but instead 1480 // changes its tactics if it sees any traps at all. This provides 1481 // a little hysteresis, delaying a recompile until a trap happens 1482 // several times. 1483 // 1484 // Actually, since there is only one bit of counter per BCI, 1485 // the possible per-BCI counts are {0,1,(per-method count)}. 1486 // This produces accurate results if in fact there is only 1487 // one hot trap site, but begins to get fuzzy if there are 1488 // many sites. For example, if there are ten sites each 1489 // trapping two or more times, they each get the blame for 1490 // all of their traps. 1491 make_not_entrant = true; 1492 } 1493 1494 // Detect repeated recompilation at the same BCI, and enforce a limit. 1495 if (make_not_entrant && maybe_prior_recompile) { 1496 // More than one recompile at this point. 1497 inc_recompile_count = maybe_prior_trap; 1498 } 1499 } else { 1500 // For reasons which are not recorded per-bytecode, we simply 1501 // force recompiles unconditionally. 1502 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1503 make_not_entrant = true; 1504 } 1505 1506 // Go back to the compiler if there are too many traps in this method. 1507 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1508 // If there are too many traps in this method, force a recompile. 1509 // This will allow the compiler to see the limit overflow, and 1510 // take corrective action, if possible. 1511 // (This condition is an unlikely backstop only, because the 1512 // PerBytecodeTrapLimit is more likely to take effect first, 1513 // if it is applicable.) 1514 make_not_entrant = true; 1515 } 1516 1517 // Here's more hysteresis: If there has been a recompile at 1518 // this trap point already, run the method in the interpreter 1519 // for a while to exercise it more thoroughly. 1520 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1521 reprofile = true; 1522 } 1523 1524 } 1525 1526 // Take requested actions on the method: 1527 1528 // Recompile 1529 if (make_not_entrant) { 1530 if (!nm->make_not_entrant()) { 1531 return; // the call did not change nmethod's state 1532 } 1533 1534 if (pdata != NULL) { 1535 // Record the recompilation event, if any. 1536 int tstate0 = pdata->trap_state(); 1537 int tstate1 = trap_state_set_recompiled(tstate0, true); 1538 if (tstate1 != tstate0) 1539 pdata->set_trap_state(tstate1); 1540 } 1541 } 1542 1543 if (inc_recompile_count) { 1544 trap_mdo->inc_overflow_recompile_count(); 1545 if ((uint)trap_mdo->overflow_recompile_count() > 1546 (uint)PerBytecodeRecompilationCutoff) { 1547 // Give up on the method containing the bad BCI. 1548 if (trap_method() == nm->method()) { 1549 make_not_compilable = true; 1550 } else { 1551 trap_method->set_not_compilable(CompLevel_full_optimization); 1552 // But give grace to the enclosing nm->method(). 1553 } 1554 } 1555 } 1556 1557 // Reprofile 1558 if (reprofile) { 1559 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1560 } 1561 1562 // Give up compiling 1563 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1564 assert(make_not_entrant, "consistent"); 1565 nm->method()->set_not_compilable(CompLevel_full_optimization); 1566 } 1567 1568 } // Free marked resources 1569 1570 } 1571 JRT_END 1572 1573 methodDataOop 1574 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1575 bool create_if_missing) { 1576 Thread* THREAD = thread; 1577 methodDataOop mdo = m()->method_data(); 1578 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1579 // Build an MDO. Ignore errors like OutOfMemory; 1580 // that simply means we won't have an MDO to update. 1581 methodOopDesc::build_interpreter_method_data(m, THREAD); 1582 if (HAS_PENDING_EXCEPTION) { 1583 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1584 CLEAR_PENDING_EXCEPTION; 1585 } 1586 mdo = m()->method_data(); 1587 } 1588 return mdo; 1589 } 1590 1591 ProfileData* 1592 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1593 int trap_bci, 1594 Deoptimization::DeoptReason reason, 1595 //outputs: 1596 uint& ret_this_trap_count, 1597 bool& ret_maybe_prior_trap, 1598 bool& ret_maybe_prior_recompile) { 1599 uint prior_trap_count = trap_mdo->trap_count(reason); 1600 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1601 1602 // If the runtime cannot find a place to store trap history, 1603 // it is estimated based on the general condition of the method. 1604 // If the method has ever been recompiled, or has ever incurred 1605 // a trap with the present reason , then this BCI is assumed 1606 // (pessimistically) to be the culprit. 1607 bool maybe_prior_trap = (prior_trap_count != 0); 1608 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1609 ProfileData* pdata = NULL; 1610 1611 1612 // For reasons which are recorded per bytecode, we check per-BCI data. 1613 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1614 if (per_bc_reason != Reason_none) { 1615 // Find the profile data for this BCI. If there isn't one, 1616 // try to allocate one from the MDO's set of spares. 1617 // This will let us detect a repeated trap at this point. 1618 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1619 1620 if (pdata != NULL) { 1621 // Query the trap state of this profile datum. 1622 int tstate0 = pdata->trap_state(); 1623 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1624 maybe_prior_trap = false; 1625 if (!trap_state_is_recompiled(tstate0)) 1626 maybe_prior_recompile = false; 1627 1628 // Update the trap state of this profile datum. 1629 int tstate1 = tstate0; 1630 // Record the reason. 1631 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1632 // Store the updated state on the MDO, for next time. 1633 if (tstate1 != tstate0) 1634 pdata->set_trap_state(tstate1); 1635 } else { 1636 if (LogCompilation && xtty != NULL) { 1637 ttyLocker ttyl; 1638 // Missing MDP? Leave a small complaint in the log. 1639 xtty->elem("missing_mdp bci='%d'", trap_bci); 1640 } 1641 } 1642 } 1643 1644 // Return results: 1645 ret_this_trap_count = this_trap_count; 1646 ret_maybe_prior_trap = maybe_prior_trap; 1647 ret_maybe_prior_recompile = maybe_prior_recompile; 1648 return pdata; 1649 } 1650 1651 void 1652 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1653 ResourceMark rm; 1654 // Ignored outputs: 1655 uint ignore_this_trap_count; 1656 bool ignore_maybe_prior_trap; 1657 bool ignore_maybe_prior_recompile; 1658 query_update_method_data(trap_mdo, trap_bci, 1659 (DeoptReason)reason, 1660 ignore_this_trap_count, 1661 ignore_maybe_prior_trap, 1662 ignore_maybe_prior_recompile); 1663 } 1664 1665 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1666 1667 // Still in Java no safepoints 1668 { 1669 // This enters VM and may safepoint 1670 uncommon_trap_inner(thread, trap_request); 1671 } 1672 return fetch_unroll_info_helper(thread); 1673 } 1674 1675 // Local derived constants. 1676 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1677 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1678 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1679 1680 //---------------------------trap_state_reason--------------------------------- 1681 Deoptimization::DeoptReason 1682 Deoptimization::trap_state_reason(int trap_state) { 1683 // This assert provides the link between the width of DataLayout::trap_bits 1684 // and the encoding of "recorded" reasons. It ensures there are enough 1685 // bits to store all needed reasons in the per-BCI MDO profile. 1686 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1687 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1688 trap_state -= recompile_bit; 1689 if (trap_state == DS_REASON_MASK) { 1690 return Reason_many; 1691 } else { 1692 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1693 return (DeoptReason)trap_state; 1694 } 1695 } 1696 //-------------------------trap_state_has_reason------------------------------- 1697 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1698 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1699 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1700 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1701 trap_state -= recompile_bit; 1702 if (trap_state == DS_REASON_MASK) { 1703 return -1; // true, unspecifically (bottom of state lattice) 1704 } else if (trap_state == reason) { 1705 return 1; // true, definitely 1706 } else if (trap_state == 0) { 1707 return 0; // false, definitely (top of state lattice) 1708 } else { 1709 return 0; // false, definitely 1710 } 1711 } 1712 //-------------------------trap_state_add_reason------------------------------- 1713 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1714 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1715 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1716 trap_state -= recompile_bit; 1717 if (trap_state == DS_REASON_MASK) { 1718 return trap_state + recompile_bit; // already at state lattice bottom 1719 } else if (trap_state == reason) { 1720 return trap_state + recompile_bit; // the condition is already true 1721 } else if (trap_state == 0) { 1722 return reason + recompile_bit; // no condition has yet been true 1723 } else { 1724 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1725 } 1726 } 1727 //-----------------------trap_state_is_recompiled------------------------------ 1728 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1729 return (trap_state & DS_RECOMPILE_BIT) != 0; 1730 } 1731 //-----------------------trap_state_set_recompiled----------------------------- 1732 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1733 if (z) return trap_state | DS_RECOMPILE_BIT; 1734 else return trap_state & ~DS_RECOMPILE_BIT; 1735 } 1736 //---------------------------format_trap_state--------------------------------- 1737 // This is used for debugging and diagnostics, including hotspot.log output. 1738 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1739 int trap_state) { 1740 DeoptReason reason = trap_state_reason(trap_state); 1741 bool recomp_flag = trap_state_is_recompiled(trap_state); 1742 // Re-encode the state from its decoded components. 1743 int decoded_state = 0; 1744 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1745 decoded_state = trap_state_add_reason(decoded_state, reason); 1746 if (recomp_flag) 1747 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1748 // If the state re-encodes properly, format it symbolically. 1749 // Because this routine is used for debugging and diagnostics, 1750 // be robust even if the state is a strange value. 1751 size_t len; 1752 if (decoded_state != trap_state) { 1753 // Random buggy state that doesn't decode?? 1754 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1755 } else { 1756 len = jio_snprintf(buf, buflen, "%s%s", 1757 trap_reason_name(reason), 1758 recomp_flag ? " recompiled" : ""); 1759 } 1760 if (len >= buflen) 1761 buf[buflen-1] = '\0'; 1762 return buf; 1763 } 1764 1765 1766 //--------------------------------statics-------------------------------------- 1767 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1768 = Deoptimization::Action_reinterpret; 1769 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1770 // Note: Keep this in sync. with enum DeoptReason. 1771 "none", 1772 "null_check", 1773 "null_assert", 1774 "range_check", 1775 "class_check", 1776 "array_check", 1777 "intrinsic", 1778 "bimorphic", 1779 "unloaded", 1780 "uninitialized", 1781 "unreached", 1782 "unhandled", 1783 "constraint", 1784 "div0_check", 1785 "age", 1786 "predicate", 1787 "loop_limit_check" 1788 }; 1789 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1790 // Note: Keep this in sync. with enum DeoptAction. 1791 "none", 1792 "maybe_recompile", 1793 "reinterpret", 1794 "make_not_entrant", 1795 "make_not_compilable" 1796 }; 1797 1798 const char* Deoptimization::trap_reason_name(int reason) { 1799 if (reason == Reason_many) return "many"; 1800 if ((uint)reason < Reason_LIMIT) 1801 return _trap_reason_name[reason]; 1802 static char buf[20]; 1803 sprintf(buf, "reason%d", reason); 1804 return buf; 1805 } 1806 const char* Deoptimization::trap_action_name(int action) { 1807 if ((uint)action < Action_LIMIT) 1808 return _trap_action_name[action]; 1809 static char buf[20]; 1810 sprintf(buf, "action%d", action); 1811 return buf; 1812 } 1813 1814 // This is used for debugging and diagnostics, including hotspot.log output. 1815 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1816 int trap_request) { 1817 jint unloaded_class_index = trap_request_index(trap_request); 1818 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1819 const char* action = trap_action_name(trap_request_action(trap_request)); 1820 size_t len; 1821 if (unloaded_class_index < 0) { 1822 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1823 reason, action); 1824 } else { 1825 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1826 reason, action, unloaded_class_index); 1827 } 1828 if (len >= buflen) 1829 buf[buflen-1] = '\0'; 1830 return buf; 1831 } 1832 1833 juint Deoptimization::_deoptimization_hist 1834 [Deoptimization::Reason_LIMIT] 1835 [1 + Deoptimization::Action_LIMIT] 1836 [Deoptimization::BC_CASE_LIMIT] 1837 = {0}; 1838 1839 enum { 1840 LSB_BITS = 8, 1841 LSB_MASK = right_n_bits(LSB_BITS) 1842 }; 1843 1844 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1845 Bytecodes::Code bc) { 1846 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1847 assert(action >= 0 && action < Action_LIMIT, "oob"); 1848 _deoptimization_hist[Reason_none][0][0] += 1; // total 1849 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1850 juint* cases = _deoptimization_hist[reason][1+action]; 1851 juint* bc_counter_addr = NULL; 1852 juint bc_counter = 0; 1853 // Look for an unused counter, or an exact match to this BC. 1854 if (bc != Bytecodes::_illegal) { 1855 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1856 juint* counter_addr = &cases[bc_case]; 1857 juint counter = *counter_addr; 1858 if ((counter == 0 && bc_counter_addr == NULL) 1859 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1860 // this counter is either free or is already devoted to this BC 1861 bc_counter_addr = counter_addr; 1862 bc_counter = counter | bc; 1863 } 1864 } 1865 } 1866 if (bc_counter_addr == NULL) { 1867 // Overflow, or no given bytecode. 1868 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1869 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1870 } 1871 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1872 } 1873 1874 jint Deoptimization::total_deoptimization_count() { 1875 return _deoptimization_hist[Reason_none][0][0]; 1876 } 1877 1878 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1879 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1880 return _deoptimization_hist[reason][0][0]; 1881 } 1882 1883 void Deoptimization::print_statistics() { 1884 juint total = total_deoptimization_count(); 1885 juint account = total; 1886 if (total != 0) { 1887 ttyLocker ttyl; 1888 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1889 tty->print_cr("Deoptimization traps recorded:"); 1890 #define PRINT_STAT_LINE(name, r) \ 1891 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1892 PRINT_STAT_LINE("total", total); 1893 // For each non-zero entry in the histogram, print the reason, 1894 // the action, and (if specifically known) the type of bytecode. 1895 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1896 for (int action = 0; action < Action_LIMIT; action++) { 1897 juint* cases = _deoptimization_hist[reason][1+action]; 1898 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1899 juint counter = cases[bc_case]; 1900 if (counter != 0) { 1901 char name[1*K]; 1902 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1903 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1904 bc = Bytecodes::_illegal; 1905 sprintf(name, "%s/%s/%s", 1906 trap_reason_name(reason), 1907 trap_action_name(action), 1908 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1909 juint r = counter >> LSB_BITS; 1910 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1911 account -= r; 1912 } 1913 } 1914 } 1915 } 1916 if (account != 0) { 1917 PRINT_STAT_LINE("unaccounted", account); 1918 } 1919 #undef PRINT_STAT_LINE 1920 if (xtty != NULL) xtty->tail("statistics"); 1921 } 1922 } 1923 #else // COMPILER2 || SHARK 1924 1925 1926 // Stubs for C1 only system. 1927 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1928 return false; 1929 } 1930 1931 const char* Deoptimization::trap_reason_name(int reason) { 1932 return "unknown"; 1933 } 1934 1935 void Deoptimization::print_statistics() { 1936 // no output 1937 } 1938 1939 void 1940 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1941 // no udpate 1942 } 1943 1944 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1945 return 0; 1946 } 1947 1948 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1949 Bytecodes::Code bc) { 1950 // no update 1951 } 1952 1953 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1954 int trap_state) { 1955 jio_snprintf(buf, buflen, "#%d", trap_state); 1956 return buf; 1957 } 1958 1959 #endif // COMPILER2 || SHARK