1 /* 2 * Copyright (c) 1997, 2014, 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/method.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_32 85 # include "adfiles/ad_ppc_32.hpp" 86 #endif 87 #ifdef TARGET_ARCH_MODEL_ppc_64 88 # include "adfiles/ad_ppc_64.hpp" 89 #endif 90 #endif // COMPILER2 91 92 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 93 94 bool DeoptimizationMarker::_is_active = false; 95 96 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 97 int caller_adjustment, 98 int caller_actual_parameters, 99 int number_of_frames, 100 intptr_t* frame_sizes, 101 address* frame_pcs, 102 BasicType return_type) { 103 _size_of_deoptimized_frame = size_of_deoptimized_frame; 104 _caller_adjustment = caller_adjustment; 105 _caller_actual_parameters = caller_actual_parameters; 106 _number_of_frames = number_of_frames; 107 _frame_sizes = frame_sizes; 108 _frame_pcs = frame_pcs; 109 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 110 _return_type = return_type; 111 _initial_info = 0; 112 // PD (x86 only) 113 _counter_temp = 0; 114 _unpack_kind = 0; 115 _sender_sp_temp = 0; 116 117 _total_frame_sizes = size_of_frames(); 118 } 119 120 121 Deoptimization::UnrollBlock::~UnrollBlock() { 122 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler); 123 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler); 124 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler); 125 } 126 127 128 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 129 assert(register_number < RegisterMap::reg_count, "checking register number"); 130 return &_register_block[register_number * 2]; 131 } 132 133 134 135 int Deoptimization::UnrollBlock::size_of_frames() const { 136 // Acount first for the adjustment of the initial frame 137 int result = _caller_adjustment; 138 for (int index = 0; index < number_of_frames(); index++) { 139 result += frame_sizes()[index]; 140 } 141 return result; 142 } 143 144 145 void Deoptimization::UnrollBlock::print() { 146 ttyLocker ttyl; 147 tty->print_cr("UnrollBlock"); 148 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 149 tty->print( " frame_sizes: "); 150 for (int index = 0; index < number_of_frames(); index++) { 151 tty->print("%d ", frame_sizes()[index]); 152 } 153 tty->cr(); 154 } 155 156 157 // In order to make fetch_unroll_info work properly with escape 158 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 159 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 160 // of previously eliminated objects occurs in realloc_objects, which is 161 // called from the method fetch_unroll_info_helper below. 162 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 163 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 164 // but makes the entry a little slower. There is however a little dance we have to 165 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 166 167 // fetch_unroll_info() is called at the beginning of the deoptimization 168 // handler. Note this fact before we start generating temporary frames 169 // that can confuse an asynchronous stack walker. This counter is 170 // decremented at the end of unpack_frames(). 171 thread->inc_in_deopt_handler(); 172 173 return fetch_unroll_info_helper(thread); 174 JRT_END 175 176 177 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 178 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 179 180 // Note: there is a safepoint safety issue here. No matter whether we enter 181 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 182 // the vframeArray is created. 183 // 184 185 // Allocate our special deoptimization ResourceMark 186 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 187 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 188 thread->set_deopt_mark(dmark); 189 190 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 191 RegisterMap map(thread, true); 192 RegisterMap dummy_map(thread, false); 193 // Now get the deoptee with a valid map 194 frame deoptee = stub_frame.sender(&map); 195 // Set the deoptee nmethod 196 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 197 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 198 199 if (VerifyStack) { 200 thread->validate_frame_layout(); 201 } 202 203 // Create a growable array of VFrames where each VFrame represents an inlined 204 // Java frame. This storage is allocated with the usual system arena. 205 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 206 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 207 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 208 while (!vf->is_top()) { 209 assert(vf->is_compiled_frame(), "Wrong frame type"); 210 chunk->push(compiledVFrame::cast(vf)); 211 vf = vf->sender(); 212 } 213 assert(vf->is_compiled_frame(), "Wrong frame type"); 214 chunk->push(compiledVFrame::cast(vf)); 215 216 bool realloc_failures = false; 217 218 #ifdef COMPILER2 219 // Reallocate the non-escaping objects and restore their fields. Then 220 // relock objects if synchronization on them was eliminated. 221 if (DoEscapeAnalysis || EliminateNestedLocks) { 222 if (EliminateAllocations) { 223 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 224 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 225 226 // The flag return_oop() indicates call sites which return oop 227 // in compiled code. Such sites include java method calls, 228 // runtime calls (for example, used to allocate new objects/arrays 229 // on slow code path) and any other calls generated in compiled code. 230 // It is not guaranteed that we can get such information here only 231 // by analyzing bytecode in deoptimized frames. This is why this flag 232 // is set during method compilation (see Compile::Process_OopMap_Node()). 233 // If the previous frame was popped, we don't have a result. 234 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution(); 235 Handle return_value; 236 if (save_oop_result) { 237 // Reallocation may trigger GC. If deoptimization happened on return from 238 // call which returns oop we need to save it since it is not in oopmap. 239 oop result = deoptee.saved_oop_result(&map); 240 assert(result == NULL || result->is_oop(), "must be oop"); 241 return_value = Handle(thread, result); 242 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 243 if (TraceDeoptimization) { 244 ttyLocker ttyl; 245 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread); 246 } 247 } 248 if (objects != NULL) { 249 JRT_BLOCK 250 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD); 251 JRT_END 252 reassign_fields(&deoptee, &map, objects, realloc_failures); 253 #ifndef PRODUCT 254 if (TraceDeoptimization) { 255 ttyLocker ttyl; 256 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 257 print_objects(objects, realloc_failures); 258 } 259 #endif 260 } 261 if (save_oop_result) { 262 // Restore result. 263 deoptee.set_saved_oop_result(&map, return_value()); 264 } 265 } 266 if (EliminateLocks) { 267 #ifndef PRODUCT 268 bool first = true; 269 #endif 270 for (int i = 0; i < chunk->length(); i++) { 271 compiledVFrame* cvf = chunk->at(i); 272 assert (cvf->scope() != NULL,"expect only compiled java frames"); 273 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 274 if (monitors->is_nonempty()) { 275 relock_objects(monitors, thread, realloc_failures); 276 #ifndef PRODUCT 277 if (TraceDeoptimization) { 278 ttyLocker ttyl; 279 for (int j = 0; j < monitors->length(); j++) { 280 MonitorInfo* mi = monitors->at(j); 281 if (mi->eliminated()) { 282 if (first) { 283 first = false; 284 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 285 } 286 if (mi->owner_is_scalar_replaced()) { 287 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 288 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 289 } else { 290 tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner()); 291 } 292 } 293 } 294 } 295 #endif 296 } 297 } 298 } 299 } 300 #endif // COMPILER2 301 // Ensure that no safepoint is taken after pointers have been stored 302 // in fields of rematerialized objects. If a safepoint occurs from here on 303 // out the java state residing in the vframeArray will be missed. 304 No_Safepoint_Verifier no_safepoint; 305 306 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 307 #ifdef COMPILER2 308 if (realloc_failures) { 309 pop_frames_failed_reallocs(thread, array); 310 } 311 #endif 312 313 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 314 thread->set_vframe_array_head(array); 315 316 // Now that the vframeArray has been created if we have any deferred local writes 317 // added by jvmti then we can free up that structure as the data is now in the 318 // vframeArray 319 320 if (thread->deferred_locals() != NULL) { 321 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 322 int i = 0; 323 do { 324 // Because of inlining we could have multiple vframes for a single frame 325 // and several of the vframes could have deferred writes. Find them all. 326 if (list->at(i)->id() == array->original().id()) { 327 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 328 list->remove_at(i); 329 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 330 delete dlv; 331 } else { 332 i++; 333 } 334 } while ( i < list->length() ); 335 if (list->length() == 0) { 336 thread->set_deferred_locals(NULL); 337 // free the list and elements back to C heap. 338 delete list; 339 } 340 341 } 342 343 #ifndef SHARK 344 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 345 CodeBlob* cb = stub_frame.cb(); 346 // Verify we have the right vframeArray 347 assert(cb->frame_size() >= 0, "Unexpected frame size"); 348 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 349 350 // If the deopt call site is a MethodHandle invoke call site we have 351 // to adjust the unpack_sp. 352 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 353 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 354 unpack_sp = deoptee.unextended_sp(); 355 356 #ifdef ASSERT 357 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 358 #endif 359 #else 360 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 361 #endif // !SHARK 362 363 // This is a guarantee instead of an assert because if vframe doesn't match 364 // we will unpack the wrong deoptimized frame and wind up in strange places 365 // where it will be very difficult to figure out what went wrong. Better 366 // to die an early death here than some very obscure death later when the 367 // trail is cold. 368 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 369 // in that it will fail to detect a problem when there is one. This needs 370 // more work in tiger timeframe. 371 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 372 373 int number_of_frames = array->frames(); 374 375 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 376 // virtual activation, which is the reverse of the elements in the vframes array. 377 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 378 // +1 because we always have an interpreter return address for the final slot. 379 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 380 int popframe_extra_args = 0; 381 // Create an interpreter return address for the stub to use as its return 382 // address so the skeletal frames are perfectly walkable 383 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 384 385 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 386 // activation be put back on the expression stack of the caller for reexecution 387 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 388 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 389 } 390 391 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 392 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 393 // than simply use array->sender.pc(). This requires us to walk the current set of frames 394 // 395 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 396 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 397 398 // It's possible that the number of paramters at the call site is 399 // different than number of arguments in the callee when method 400 // handles are used. If the caller is interpreted get the real 401 // value so that the proper amount of space can be added to it's 402 // frame. 403 bool caller_was_method_handle = false; 404 if (deopt_sender.is_interpreted_frame()) { 405 methodHandle method = deopt_sender.interpreter_frame_method(); 406 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 407 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 408 // Method handle invokes may involve fairly arbitrary chains of 409 // calls so it's impossible to know how much actual space the 410 // caller has for locals. 411 caller_was_method_handle = true; 412 } 413 } 414 415 // 416 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 417 // frame_sizes/frame_pcs[1] next oldest frame (int) 418 // frame_sizes/frame_pcs[n] youngest frame (int) 419 // 420 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 421 // owns the space for the return address to it's caller). Confusing ain't it. 422 // 423 // The vframe array can address vframes with indices running from 424 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 425 // When we create the skeletal frames we need the oldest frame to be in the zero slot 426 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 427 // so things look a little strange in this loop. 428 // 429 int callee_parameters = 0; 430 int callee_locals = 0; 431 for (int index = 0; index < array->frames(); index++ ) { 432 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 433 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 434 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 435 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 436 callee_locals, 437 index == 0, 438 popframe_extra_args); 439 // This pc doesn't have to be perfect just good enough to identify the frame 440 // as interpreted so the skeleton frame will be walkable 441 // The correct pc will be set when the skeleton frame is completely filled out 442 // The final pc we store in the loop is wrong and will be overwritten below 443 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 444 445 callee_parameters = array->element(index)->method()->size_of_parameters(); 446 callee_locals = array->element(index)->method()->max_locals(); 447 popframe_extra_args = 0; 448 } 449 450 // Compute whether the root vframe returns a float or double value. 451 BasicType return_type; 452 { 453 HandleMark hm; 454 methodHandle method(thread, array->element(0)->method()); 455 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 456 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 457 } 458 459 // Compute information for handling adapters and adjusting the frame size of the caller. 460 int caller_adjustment = 0; 461 462 // Compute the amount the oldest interpreter frame will have to adjust 463 // its caller's stack by. If the caller is a compiled frame then 464 // we pretend that the callee has no parameters so that the 465 // extension counts for the full amount of locals and not just 466 // locals-parms. This is because without a c2i adapter the parm 467 // area as created by the compiled frame will not be usable by 468 // the interpreter. (Depending on the calling convention there 469 // may not even be enough space). 470 471 // QQQ I'd rather see this pushed down into last_frame_adjust 472 // and have it take the sender (aka caller). 473 474 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 475 caller_adjustment = last_frame_adjust(0, callee_locals); 476 } else if (callee_locals > callee_parameters) { 477 // The caller frame may need extending to accommodate 478 // non-parameter locals of the first unpacked interpreted frame. 479 // Compute that adjustment. 480 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 481 } 482 483 // If the sender is deoptimized the we must retrieve the address of the handler 484 // since the frame will "magically" show the original pc before the deopt 485 // and we'd undo the deopt. 486 487 frame_pcs[0] = deopt_sender.raw_pc(); 488 489 #ifndef SHARK 490 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 491 #endif // SHARK 492 493 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 494 caller_adjustment * BytesPerWord, 495 caller_was_method_handle ? 0 : callee_parameters, 496 number_of_frames, 497 frame_sizes, 498 frame_pcs, 499 return_type); 500 // On some platforms, we need a way to pass some platform dependent 501 // information to the unpacking code so the skeletal frames come out 502 // correct (initial fp value, unextended sp, ...) 503 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 504 505 if (array->frames() > 1) { 506 if (VerifyStack && TraceDeoptimization) { 507 ttyLocker ttyl; 508 tty->print_cr("Deoptimizing method containing inlining"); 509 } 510 } 511 512 array->set_unroll_block(info); 513 return info; 514 } 515 516 // Called to cleanup deoptimization data structures in normal case 517 // after unpacking to stack and when stack overflow error occurs 518 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 519 vframeArray *array) { 520 521 // Get array if coming from exception 522 if (array == NULL) { 523 array = thread->vframe_array_head(); 524 } 525 thread->set_vframe_array_head(NULL); 526 527 // Free the previous UnrollBlock 528 vframeArray* old_array = thread->vframe_array_last(); 529 thread->set_vframe_array_last(array); 530 531 if (old_array != NULL) { 532 UnrollBlock* old_info = old_array->unroll_block(); 533 old_array->set_unroll_block(NULL); 534 delete old_info; 535 delete old_array; 536 } 537 538 // Deallocate any resource creating in this routine and any ResourceObjs allocated 539 // inside the vframeArray (StackValueCollections) 540 541 delete thread->deopt_mark(); 542 thread->set_deopt_mark(NULL); 543 thread->set_deopt_nmethod(NULL); 544 545 546 if (JvmtiExport::can_pop_frame()) { 547 #ifndef CC_INTERP 548 // Regardless of whether we entered this routine with the pending 549 // popframe condition bit set, we should always clear it now 550 thread->clear_popframe_condition(); 551 #else 552 // C++ interpeter will clear has_pending_popframe when it enters 553 // with method_resume. For deopt_resume2 we clear it now. 554 if (thread->popframe_forcing_deopt_reexecution()) 555 thread->clear_popframe_condition(); 556 #endif /* CC_INTERP */ 557 } 558 559 // unpack_frames() is called at the end of the deoptimization handler 560 // and (in C2) at the end of the uncommon trap handler. Note this fact 561 // so that an asynchronous stack walker can work again. This counter is 562 // incremented at the beginning of fetch_unroll_info() and (in C2) at 563 // the beginning of uncommon_trap(). 564 thread->dec_in_deopt_handler(); 565 } 566 567 568 // Return BasicType of value being returned 569 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 570 571 // We are already active int he special DeoptResourceMark any ResourceObj's we 572 // allocate will be freed at the end of the routine. 573 574 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 575 // but makes the entry a little slower. There is however a little dance we have to 576 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 577 ResetNoHandleMark rnhm; // No-op in release/product versions 578 HandleMark hm; 579 580 frame stub_frame = thread->last_frame(); 581 582 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 583 // must point to the vframeArray for the unpack frame. 584 vframeArray* array = thread->vframe_array_head(); 585 586 #ifndef PRODUCT 587 if (TraceDeoptimization) { 588 ttyLocker ttyl; 589 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 590 } 591 #endif 592 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 593 stub_frame.pc(), stub_frame.sp(), exec_mode); 594 595 UnrollBlock* info = array->unroll_block(); 596 597 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 598 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 599 600 BasicType bt = info->return_type(); 601 602 // If we have an exception pending, claim that the return type is an oop 603 // so the deopt_blob does not overwrite the exception_oop. 604 605 if (exec_mode == Unpack_exception) 606 bt = T_OBJECT; 607 608 // Cleanup thread deopt data 609 cleanup_deopt_info(thread, array); 610 611 #ifndef PRODUCT 612 if (VerifyStack) { 613 ResourceMark res_mark; 614 615 thread->validate_frame_layout(); 616 617 // Verify that the just-unpacked frames match the interpreter's 618 // notions of expression stack and locals 619 vframeArray* cur_array = thread->vframe_array_last(); 620 RegisterMap rm(thread, false); 621 rm.set_include_argument_oops(false); 622 bool is_top_frame = true; 623 int callee_size_of_parameters = 0; 624 int callee_max_locals = 0; 625 for (int i = 0; i < cur_array->frames(); i++) { 626 vframeArrayElement* el = cur_array->element(i); 627 frame* iframe = el->iframe(); 628 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 629 630 // Get the oop map for this bci 631 InterpreterOopMap mask; 632 int cur_invoke_parameter_size = 0; 633 bool try_next_mask = false; 634 int next_mask_expression_stack_size = -1; 635 int top_frame_expression_stack_adjustment = 0; 636 methodHandle mh(thread, iframe->interpreter_frame_method()); 637 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 638 BytecodeStream str(mh); 639 str.set_start(iframe->interpreter_frame_bci()); 640 int max_bci = mh->code_size(); 641 // Get to the next bytecode if possible 642 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 643 // Check to see if we can grab the number of outgoing arguments 644 // at an uncommon trap for an invoke (where the compiler 645 // generates debug info before the invoke has executed) 646 Bytecodes::Code cur_code = str.next(); 647 if (cur_code == Bytecodes::_invokevirtual || 648 cur_code == Bytecodes::_invokespecial || 649 cur_code == Bytecodes::_invokestatic || 650 cur_code == Bytecodes::_invokeinterface || 651 cur_code == Bytecodes::_invokedynamic) { 652 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 653 Symbol* signature = invoke.signature(); 654 ArgumentSizeComputer asc(signature); 655 cur_invoke_parameter_size = asc.size(); 656 if (invoke.has_receiver()) { 657 // Add in receiver 658 ++cur_invoke_parameter_size; 659 } 660 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 661 callee_size_of_parameters++; 662 } 663 } 664 if (str.bci() < max_bci) { 665 Bytecodes::Code bc = str.next(); 666 if (bc >= 0) { 667 // The interpreter oop map generator reports results before 668 // the current bytecode has executed except in the case of 669 // calls. It seems to be hard to tell whether the compiler 670 // has emitted debug information matching the "state before" 671 // a given bytecode or the state after, so we try both 672 switch (cur_code) { 673 case Bytecodes::_invokevirtual: 674 case Bytecodes::_invokespecial: 675 case Bytecodes::_invokestatic: 676 case Bytecodes::_invokeinterface: 677 case Bytecodes::_invokedynamic: 678 case Bytecodes::_athrow: 679 break; 680 default: { 681 InterpreterOopMap next_mask; 682 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 683 next_mask_expression_stack_size = next_mask.expression_stack_size(); 684 // Need to subtract off the size of the result type of 685 // the bytecode because this is not described in the 686 // debug info but returned to the interpreter in the TOS 687 // caching register 688 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 689 if (bytecode_result_type != T_ILLEGAL) { 690 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 691 } 692 assert(top_frame_expression_stack_adjustment >= 0, ""); 693 try_next_mask = true; 694 break; 695 } 696 } 697 } 698 } 699 700 // Verify stack depth and oops in frame 701 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 702 if (!( 703 /* SPARC */ 704 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 705 /* x86 */ 706 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 707 (try_next_mask && 708 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 709 top_frame_expression_stack_adjustment))) || 710 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 711 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 712 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 713 )) { 714 ttyLocker ttyl; 715 716 // Print out some information that will help us debug the problem 717 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 718 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 719 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 720 iframe->interpreter_frame_expression_stack_size()); 721 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 722 tty->print_cr(" try_next_mask = %d", try_next_mask); 723 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 724 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 725 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 726 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 727 tty->print_cr(" exec_mode = %d", exec_mode); 728 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 729 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 730 tty->print_cr(" Interpreted frames:"); 731 for (int k = 0; k < cur_array->frames(); k++) { 732 vframeArrayElement* el = cur_array->element(k); 733 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 734 } 735 cur_array->print_on_2(tty); 736 guarantee(false, "wrong number of expression stack elements during deopt"); 737 } 738 VerifyOopClosure verify; 739 iframe->oops_interpreted_do(&verify, NULL, &rm, false); 740 callee_size_of_parameters = mh->size_of_parameters(); 741 callee_max_locals = mh->max_locals(); 742 is_top_frame = false; 743 } 744 } 745 #endif /* !PRODUCT */ 746 747 748 return bt; 749 JRT_END 750 751 752 int Deoptimization::deoptimize_dependents() { 753 Threads::deoptimized_wrt_marked_nmethods(); 754 return 0; 755 } 756 757 758 #ifdef COMPILER2 759 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 760 Handle pending_exception(thread->pending_exception()); 761 const char* exception_file = thread->exception_file(); 762 int exception_line = thread->exception_line(); 763 thread->clear_pending_exception(); 764 765 bool failures = false; 766 767 for (int i = 0; i < objects->length(); i++) { 768 assert(objects->at(i)->is_object(), "invalid debug information"); 769 ObjectValue* sv = (ObjectValue*) objects->at(i); 770 771 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 772 oop obj = NULL; 773 774 if (k->oop_is_instance()) { 775 InstanceKlass* ik = InstanceKlass::cast(k()); 776 obj = ik->allocate_instance(THREAD); 777 } else if (k->oop_is_typeArray()) { 778 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 779 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 780 int len = sv->field_size() / type2size[ak->element_type()]; 781 obj = ak->allocate(len, THREAD); 782 } else if (k->oop_is_objArray()) { 783 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 784 obj = ak->allocate(sv->field_size(), THREAD); 785 } 786 787 if (obj == NULL) { 788 failures = true; 789 } 790 791 assert(sv->value().is_null(), "redundant reallocation"); 792 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 793 CLEAR_PENDING_EXCEPTION; 794 sv->set_value(obj); 795 } 796 797 if (failures) { 798 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 799 } else if (pending_exception.not_null()) { 800 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 801 } 802 803 return failures; 804 } 805 806 // This assumes that the fields are stored in ObjectValue in the same order 807 // they are yielded by do_nonstatic_fields. 808 class FieldReassigner: public FieldClosure { 809 frame* _fr; 810 RegisterMap* _reg_map; 811 ObjectValue* _sv; 812 InstanceKlass* _ik; 813 oop _obj; 814 815 int _i; 816 public: 817 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 818 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 819 820 int i() const { return _i; } 821 822 823 void do_field(fieldDescriptor* fd) { 824 intptr_t val; 825 StackValue* value = 826 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 827 int offset = fd->offset(); 828 switch (fd->field_type()) { 829 case T_OBJECT: case T_ARRAY: 830 assert(value->type() == T_OBJECT, "Agreement."); 831 _obj->obj_field_put(offset, value->get_obj()()); 832 break; 833 834 case T_LONG: case T_DOUBLE: { 835 assert(value->type() == T_INT, "Agreement."); 836 StackValue* low = 837 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 838 #ifdef _LP64 839 jlong res = (jlong)low->get_int(); 840 #else 841 #ifdef SPARC 842 // For SPARC we have to swap high and low words. 843 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 844 #else 845 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 846 #endif //SPARC 847 #endif 848 _obj->long_field_put(offset, res); 849 break; 850 } 851 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 852 case T_INT: case T_FLOAT: // 4 bytes. 853 assert(value->type() == T_INT, "Agreement."); 854 val = value->get_int(); 855 _obj->int_field_put(offset, (jint)*((jint*)&val)); 856 break; 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_field_put(offset, (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_field_put(offset, (jboolean)*((jint*)&val)); 868 break; 869 870 default: 871 ShouldNotReachHere(); 872 } 873 _i++; 874 } 875 }; 876 877 // restore elements of an eliminated type array 878 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 879 int index = 0; 880 intptr_t val; 881 882 for (int i = 0; i < sv->field_size(); i++) { 883 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 884 switch(type) { 885 case T_LONG: case T_DOUBLE: { 886 assert(value->type() == T_INT, "Agreement."); 887 StackValue* low = 888 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 889 #ifdef _LP64 890 jlong res = (jlong)low->get_int(); 891 #else 892 #ifdef SPARC 893 // For SPARC we have to swap high and low words. 894 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 895 #else 896 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 897 #endif //SPARC 898 #endif 899 obj->long_at_put(index, res); 900 break; 901 } 902 903 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 904 case T_INT: case T_FLOAT: // 4 bytes. 905 assert(value->type() == T_INT, "Agreement."); 906 val = value->get_int(); 907 obj->int_at_put(index, (jint)*((jint*)&val)); 908 break; 909 910 case T_SHORT: case T_CHAR: // 2 bytes 911 assert(value->type() == T_INT, "Agreement."); 912 val = value->get_int(); 913 obj->short_at_put(index, (jshort)*((jint*)&val)); 914 break; 915 916 case T_BOOLEAN: case T_BYTE: // 1 byte 917 assert(value->type() == T_INT, "Agreement."); 918 val = value->get_int(); 919 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 920 break; 921 922 default: 923 ShouldNotReachHere(); 924 } 925 index++; 926 } 927 } 928 929 930 // restore fields of an eliminated object array 931 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 932 for (int i = 0; i < sv->field_size(); i++) { 933 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 934 assert(value->type() == T_OBJECT, "object element expected"); 935 obj->obj_at_put(i, value->get_obj()()); 936 } 937 } 938 939 940 // restore fields of all eliminated objects and arrays 941 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 942 for (int i = 0; i < objects->length(); i++) { 943 ObjectValue* sv = (ObjectValue*) objects->at(i); 944 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 945 Handle obj = sv->value(); 946 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 947 if (obj.is_null()) { 948 continue; 949 } 950 951 if (k->oop_is_instance()) { 952 InstanceKlass* ik = InstanceKlass::cast(k()); 953 FieldReassigner reassign(fr, reg_map, sv, obj()); 954 ik->do_nonstatic_fields(&reassign); 955 } else if (k->oop_is_typeArray()) { 956 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 957 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 958 } else if (k->oop_is_objArray()) { 959 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 960 } 961 } 962 } 963 964 965 // relock objects for which synchronization was eliminated 966 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 967 for (int i = 0; i < monitors->length(); i++) { 968 MonitorInfo* mon_info = monitors->at(i); 969 if (mon_info->eliminated()) { 970 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 971 if (!mon_info->owner_is_scalar_replaced()) { 972 Handle obj = Handle(mon_info->owner()); 973 markOop mark = obj->mark(); 974 if (UseBiasedLocking && mark->has_bias_pattern()) { 975 // New allocated objects may have the mark set to anonymously biased. 976 // Also the deoptimized method may called methods with synchronization 977 // where the thread-local object is bias locked to the current thread. 978 assert(mark->is_biased_anonymously() || 979 mark->biased_locker() == thread, "should be locked to current thread"); 980 // Reset mark word to unbiased prototype. 981 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 982 obj->set_mark(unbiased_prototype); 983 } 984 BasicLock* lock = mon_info->lock(); 985 ObjectSynchronizer::slow_enter(obj, lock, thread); 986 assert(mon_info->owner()->is_locked(), "object must be locked now"); 987 } 988 } 989 } 990 } 991 992 993 #ifndef PRODUCT 994 // print information about reallocated objects 995 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 996 fieldDescriptor fd; 997 998 for (int i = 0; i < objects->length(); i++) { 999 ObjectValue* sv = (ObjectValue*) objects->at(i); 1000 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 1001 Handle obj = sv->value(); 1002 1003 tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()()); 1004 k->print_value(); 1005 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 1006 if (obj.is_null()) { 1007 tty->print(" allocation failed"); 1008 } else { 1009 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1010 } 1011 tty->cr(); 1012 1013 if (Verbose && !obj.is_null()) { 1014 k->oop_print_on(obj(), tty); 1015 } 1016 } 1017 } 1018 #endif 1019 #endif // COMPILER2 1020 1021 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1022 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 1023 1024 #ifndef PRODUCT 1025 if (TraceDeoptimization) { 1026 ttyLocker ttyl; 1027 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 1028 fr.print_on(tty); 1029 tty->print_cr(" Virtual frames (innermost first):"); 1030 for (int index = 0; index < chunk->length(); index++) { 1031 compiledVFrame* vf = chunk->at(index); 1032 tty->print(" %2d - ", index); 1033 vf->print_value(); 1034 int bci = chunk->at(index)->raw_bci(); 1035 const char* code_name; 1036 if (bci == SynchronizationEntryBCI) { 1037 code_name = "sync entry"; 1038 } else { 1039 Bytecodes::Code code = vf->method()->code_at(bci); 1040 code_name = Bytecodes::name(code); 1041 } 1042 tty->print(" - %s", code_name); 1043 tty->print_cr(" @ bci %d ", bci); 1044 if (Verbose) { 1045 vf->print(); 1046 tty->cr(); 1047 } 1048 } 1049 } 1050 #endif 1051 1052 // Register map for next frame (used for stack crawl). We capture 1053 // the state of the deopt'ing frame's caller. Thus if we need to 1054 // stuff a C2I adapter we can properly fill in the callee-save 1055 // register locations. 1056 frame caller = fr.sender(reg_map); 1057 int frame_size = caller.sp() - fr.sp(); 1058 1059 frame sender = caller; 1060 1061 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1062 // the vframeArray containing the unpacking information is allocated in the C heap. 1063 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1064 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1065 1066 // Compare the vframeArray to the collected vframes 1067 assert(array->structural_compare(thread, chunk), "just checking"); 1068 1069 #ifndef PRODUCT 1070 if (TraceDeoptimization) { 1071 ttyLocker ttyl; 1072 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1073 } 1074 #endif // PRODUCT 1075 1076 return array; 1077 } 1078 1079 #ifdef COMPILER2 1080 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1081 // Reallocation of some scalar replaced objects failed. Record 1082 // that we need to pop all the interpreter frames for the 1083 // deoptimized compiled frame. 1084 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1085 thread->set_frames_to_pop_failed_realloc(array->frames()); 1086 // Unlock all monitors here otherwise the interpreter will see a 1087 // mix of locked and unlocked monitors (because of failed 1088 // reallocations of synchronized objects) and be confused. 1089 for (int i = 0; i < array->frames(); i++) { 1090 MonitorChunk* monitors = array->element(i)->monitors(); 1091 if (monitors != NULL) { 1092 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1093 BasicObjectLock* src = monitors->at(j); 1094 if (src->obj() != NULL) { 1095 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1096 } 1097 } 1098 array->element(i)->free_monitors(thread); 1099 #ifdef ASSERT 1100 array->element(i)->set_removed_monitors(); 1101 #endif 1102 } 1103 } 1104 } 1105 #endif 1106 1107 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1108 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1109 for (int i = 0; i < monitors->length(); i++) { 1110 MonitorInfo* mon_info = monitors->at(i); 1111 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1112 objects_to_revoke->append(Handle(mon_info->owner())); 1113 } 1114 } 1115 } 1116 1117 1118 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1119 if (!UseBiasedLocking) { 1120 return; 1121 } 1122 1123 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1124 1125 // Unfortunately we don't have a RegisterMap available in most of 1126 // the places we want to call this routine so we need to walk the 1127 // stack again to update the register map. 1128 if (map == NULL || !map->update_map()) { 1129 StackFrameStream sfs(thread, true); 1130 bool found = false; 1131 while (!found && !sfs.is_done()) { 1132 frame* cur = sfs.current(); 1133 sfs.next(); 1134 found = cur->id() == fr.id(); 1135 } 1136 assert(found, "frame to be deoptimized not found on target thread's stack"); 1137 map = sfs.register_map(); 1138 } 1139 1140 vframe* vf = vframe::new_vframe(&fr, map, thread); 1141 compiledVFrame* cvf = compiledVFrame::cast(vf); 1142 // Revoke monitors' biases in all scopes 1143 while (!cvf->is_top()) { 1144 collect_monitors(cvf, objects_to_revoke); 1145 cvf = compiledVFrame::cast(cvf->sender()); 1146 } 1147 collect_monitors(cvf, objects_to_revoke); 1148 1149 if (SafepointSynchronize::is_at_safepoint()) { 1150 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1151 } else { 1152 BiasedLocking::revoke(objects_to_revoke); 1153 } 1154 } 1155 1156 1157 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1158 if (!UseBiasedLocking) { 1159 return; 1160 } 1161 1162 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1163 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1164 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1165 if (jt->has_last_Java_frame()) { 1166 StackFrameStream sfs(jt, true); 1167 while (!sfs.is_done()) { 1168 frame* cur = sfs.current(); 1169 if (cb->contains(cur->pc())) { 1170 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1171 compiledVFrame* cvf = compiledVFrame::cast(vf); 1172 // Revoke monitors' biases in all scopes 1173 while (!cvf->is_top()) { 1174 collect_monitors(cvf, objects_to_revoke); 1175 cvf = compiledVFrame::cast(cvf->sender()); 1176 } 1177 collect_monitors(cvf, objects_to_revoke); 1178 } 1179 sfs.next(); 1180 } 1181 } 1182 } 1183 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1184 } 1185 1186 1187 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1188 assert(fr.can_be_deoptimized(), "checking frame type"); 1189 1190 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1191 1192 // Patch the nmethod so that when execution returns to it we will 1193 // deopt the execution state and return to the interpreter. 1194 fr.deoptimize(thread); 1195 } 1196 1197 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1198 // Deoptimize only if the frame comes from compile code. 1199 // Do not deoptimize the frame which is already patched 1200 // during the execution of the loops below. 1201 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1202 return; 1203 } 1204 ResourceMark rm; 1205 DeoptimizationMarker dm; 1206 if (UseBiasedLocking) { 1207 revoke_biases_of_monitors(thread, fr, map); 1208 } 1209 deoptimize_single_frame(thread, fr); 1210 1211 } 1212 1213 1214 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1215 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1216 "can only deoptimize other thread at a safepoint"); 1217 // Compute frame and register map based on thread and sp. 1218 RegisterMap reg_map(thread, UseBiasedLocking); 1219 frame fr = thread->last_frame(); 1220 while (fr.id() != id) { 1221 fr = fr.sender(®_map); 1222 } 1223 deoptimize(thread, fr, ®_map); 1224 } 1225 1226 1227 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1228 if (thread == Thread::current()) { 1229 Deoptimization::deoptimize_frame_internal(thread, id); 1230 } else { 1231 VM_DeoptimizeFrame deopt(thread, id); 1232 VMThread::execute(&deopt); 1233 } 1234 } 1235 1236 1237 // JVMTI PopFrame support 1238 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1239 { 1240 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1241 } 1242 JRT_END 1243 1244 1245 #if defined(COMPILER2) || defined(SHARK) 1246 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1247 // in case of an unresolved klass entry, load the class. 1248 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1249 Klass* tk = constant_pool->klass_at(index, CHECK); 1250 return; 1251 } 1252 1253 if (!constant_pool->tag_at(index).is_symbol()) return; 1254 1255 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1256 Symbol* symbol = constant_pool->symbol_at(index); 1257 1258 // class name? 1259 if (symbol->byte_at(0) != '(') { 1260 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1261 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1262 return; 1263 } 1264 1265 // then it must be a signature! 1266 ResourceMark rm(THREAD); 1267 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1268 if (ss.is_object()) { 1269 Symbol* class_name = ss.as_symbol(CHECK); 1270 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1271 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1272 } 1273 } 1274 } 1275 1276 1277 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1278 EXCEPTION_MARK; 1279 load_class_by_index(constant_pool, index, THREAD); 1280 if (HAS_PENDING_EXCEPTION) { 1281 // Exception happened during classloading. We ignore the exception here, since it 1282 // is going to be rethrown since the current activation is going to be deoptimized and 1283 // the interpreter will re-execute the bytecode. 1284 CLEAR_PENDING_EXCEPTION; 1285 // Class loading called java code which may have caused a stack 1286 // overflow. If the exception was thrown right before the return 1287 // to the runtime the stack is no longer guarded. Reguard the 1288 // stack otherwise if we return to the uncommon trap blob and the 1289 // stack bang causes a stack overflow we crash. 1290 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1291 JavaThread* thread = (JavaThread*)THREAD; 1292 bool guard_pages_enabled = thread->stack_yellow_zone_enabled(); 1293 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1294 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1295 } 1296 } 1297 1298 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1299 HandleMark hm; 1300 1301 // uncommon_trap() is called at the beginning of the uncommon trap 1302 // handler. Note this fact before we start generating temporary frames 1303 // that can confuse an asynchronous stack walker. This counter is 1304 // decremented at the end of unpack_frames(). 1305 thread->inc_in_deopt_handler(); 1306 1307 // We need to update the map if we have biased locking. 1308 RegisterMap reg_map(thread, UseBiasedLocking); 1309 frame stub_frame = thread->last_frame(); 1310 frame fr = stub_frame.sender(®_map); 1311 // Make sure the calling nmethod is not getting deoptimized and removed 1312 // before we are done with it. 1313 nmethodLocker nl(fr.pc()); 1314 1315 // Log a message 1316 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT, 1317 trap_request, fr.pc()); 1318 1319 { 1320 ResourceMark rm; 1321 1322 // Revoke biases of any monitors in the frame to ensure we can migrate them 1323 revoke_biases_of_monitors(thread, fr, ®_map); 1324 1325 DeoptReason reason = trap_request_reason(trap_request); 1326 DeoptAction action = trap_request_action(trap_request); 1327 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1328 1329 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1330 compiledVFrame* cvf = compiledVFrame::cast(vf); 1331 1332 nmethod* nm = cvf->code(); 1333 1334 ScopeDesc* trap_scope = cvf->scope(); 1335 methodHandle trap_method = trap_scope->method(); 1336 int trap_bci = trap_scope->bci(); 1337 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1338 1339 // Record this event in the histogram. 1340 gather_statistics(reason, action, trap_bc); 1341 1342 // Ensure that we can record deopt. history: 1343 // Need MDO to record RTM code generation state. 1344 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 1345 1346 MethodData* trap_mdo = 1347 get_method_data(thread, trap_method, create_if_missing); 1348 1349 // Log a message 1350 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1351 trap_reason_name(reason), trap_action_name(action), fr.pc(), 1352 trap_method->name_and_sig_as_C_string(), trap_bci); 1353 1354 // Print a bunch of diagnostics, if requested. 1355 if (TraceDeoptimization || LogCompilation) { 1356 ResourceMark rm; 1357 ttyLocker ttyl; 1358 char buf[100]; 1359 if (xtty != NULL) { 1360 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1361 os::current_thread_id(), 1362 format_trap_request(buf, sizeof(buf), trap_request)); 1363 nm->log_identity(xtty); 1364 } 1365 Symbol* class_name = NULL; 1366 bool unresolved = false; 1367 if (unloaded_class_index >= 0) { 1368 constantPoolHandle constants (THREAD, trap_method->constants()); 1369 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1370 class_name = constants->klass_name_at(unloaded_class_index); 1371 unresolved = true; 1372 if (xtty != NULL) 1373 xtty->print(" unresolved='1'"); 1374 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1375 class_name = constants->symbol_at(unloaded_class_index); 1376 } 1377 if (xtty != NULL) 1378 xtty->name(class_name); 1379 } 1380 if (xtty != NULL && trap_mdo != NULL) { 1381 // Dump the relevant MDO state. 1382 // This is the deopt count for the current reason, any previous 1383 // reasons or recompiles seen at this point. 1384 int dcnt = trap_mdo->trap_count(reason); 1385 if (dcnt != 0) 1386 xtty->print(" count='%d'", dcnt); 1387 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1388 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1389 if (dos != 0) { 1390 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1391 if (trap_state_is_recompiled(dos)) { 1392 int recnt2 = trap_mdo->overflow_recompile_count(); 1393 if (recnt2 != 0) 1394 xtty->print(" recompiles2='%d'", recnt2); 1395 } 1396 } 1397 } 1398 if (xtty != NULL) { 1399 xtty->stamp(); 1400 xtty->end_head(); 1401 } 1402 if (TraceDeoptimization) { // make noise on the tty 1403 tty->print("Uncommon trap occurred in"); 1404 nm->method()->print_short_name(tty); 1405 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d", 1406 fr.pc(), 1407 os::current_thread_id(), 1408 trap_reason_name(reason), 1409 trap_action_name(action), 1410 unloaded_class_index); 1411 if (class_name != NULL) { 1412 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1413 class_name->print_symbol_on(tty); 1414 } 1415 tty->cr(); 1416 } 1417 if (xtty != NULL) { 1418 // Log the precise location of the trap. 1419 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1420 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1421 xtty->method(sd->method()); 1422 xtty->end_elem(); 1423 if (sd->is_top()) break; 1424 } 1425 xtty->tail("uncommon_trap"); 1426 } 1427 } 1428 // (End diagnostic printout.) 1429 1430 // Load class if necessary 1431 if (unloaded_class_index >= 0) { 1432 constantPoolHandle constants(THREAD, trap_method->constants()); 1433 load_class_by_index(constants, unloaded_class_index); 1434 } 1435 1436 // Flush the nmethod if necessary and desirable. 1437 // 1438 // We need to avoid situations where we are re-flushing the nmethod 1439 // because of a hot deoptimization site. Repeated flushes at the same 1440 // point need to be detected by the compiler and avoided. If the compiler 1441 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1442 // module must take measures to avoid an infinite cycle of recompilation 1443 // and deoptimization. There are several such measures: 1444 // 1445 // 1. If a recompilation is ordered a second time at some site X 1446 // and for the same reason R, the action is adjusted to 'reinterpret', 1447 // to give the interpreter time to exercise the method more thoroughly. 1448 // If this happens, the method's overflow_recompile_count is incremented. 1449 // 1450 // 2. If the compiler fails to reduce the deoptimization rate, then 1451 // the method's overflow_recompile_count will begin to exceed the set 1452 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1453 // is adjusted to 'make_not_compilable', and the method is abandoned 1454 // to the interpreter. This is a performance hit for hot methods, 1455 // but is better than a disastrous infinite cycle of recompilations. 1456 // (Actually, only the method containing the site X is abandoned.) 1457 // 1458 // 3. In parallel with the previous measures, if the total number of 1459 // recompilations of a method exceeds the much larger set limit 1460 // PerMethodRecompilationCutoff, the method is abandoned. 1461 // This should only happen if the method is very large and has 1462 // many "lukewarm" deoptimizations. The code which enforces this 1463 // limit is elsewhere (class nmethod, class Method). 1464 // 1465 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1466 // to recompile at each bytecode independently of the per-BCI cutoff. 1467 // 1468 // The decision to update code is up to the compiler, and is encoded 1469 // in the Action_xxx code. If the compiler requests Action_none 1470 // no trap state is changed, no compiled code is changed, and the 1471 // computation suffers along in the interpreter. 1472 // 1473 // The other action codes specify various tactics for decompilation 1474 // and recompilation. Action_maybe_recompile is the loosest, and 1475 // allows the compiled code to stay around until enough traps are seen, 1476 // and until the compiler gets around to recompiling the trapping method. 1477 // 1478 // The other actions cause immediate removal of the present code. 1479 1480 bool update_trap_state = true; 1481 bool make_not_entrant = false; 1482 bool make_not_compilable = false; 1483 bool reprofile = false; 1484 switch (action) { 1485 case Action_none: 1486 // Keep the old code. 1487 update_trap_state = false; 1488 break; 1489 case Action_maybe_recompile: 1490 // Do not need to invalidate the present code, but we can 1491 // initiate another 1492 // Start compiler without (necessarily) invalidating the nmethod. 1493 // The system will tolerate the old code, but new code should be 1494 // generated when possible. 1495 break; 1496 case Action_reinterpret: 1497 // Go back into the interpreter for a while, and then consider 1498 // recompiling form scratch. 1499 make_not_entrant = true; 1500 // Reset invocation counter for outer most method. 1501 // This will allow the interpreter to exercise the bytecodes 1502 // for a while before recompiling. 1503 // By contrast, Action_make_not_entrant is immediate. 1504 // 1505 // Note that the compiler will track null_check, null_assert, 1506 // range_check, and class_check events and log them as if they 1507 // had been traps taken from compiled code. This will update 1508 // the MDO trap history so that the next compilation will 1509 // properly detect hot trap sites. 1510 reprofile = true; 1511 break; 1512 case Action_make_not_entrant: 1513 // Request immediate recompilation, and get rid of the old code. 1514 // Make them not entrant, so next time they are called they get 1515 // recompiled. Unloaded classes are loaded now so recompile before next 1516 // time they are called. Same for uninitialized. The interpreter will 1517 // link the missing class, if any. 1518 make_not_entrant = true; 1519 break; 1520 case Action_make_not_compilable: 1521 // Give up on compiling this method at all. 1522 make_not_entrant = true; 1523 make_not_compilable = true; 1524 break; 1525 default: 1526 ShouldNotReachHere(); 1527 } 1528 1529 // Setting +ProfileTraps fixes the following, on all platforms: 1530 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1531 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1532 // recompile relies on a MethodData* to record heroic opt failures. 1533 1534 // Whether the interpreter is producing MDO data or not, we also need 1535 // to use the MDO to detect hot deoptimization points and control 1536 // aggressive optimization. 1537 bool inc_recompile_count = false; 1538 ProfileData* pdata = NULL; 1539 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1540 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity"); 1541 uint this_trap_count = 0; 1542 bool maybe_prior_trap = false; 1543 bool maybe_prior_recompile = false; 1544 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1545 nm->method(), 1546 //outputs: 1547 this_trap_count, 1548 maybe_prior_trap, 1549 maybe_prior_recompile); 1550 // Because the interpreter also counts null, div0, range, and class 1551 // checks, these traps from compiled code are double-counted. 1552 // This is harmless; it just means that the PerXTrapLimit values 1553 // are in effect a little smaller than they look. 1554 1555 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1556 if (per_bc_reason != Reason_none) { 1557 // Now take action based on the partially known per-BCI history. 1558 if (maybe_prior_trap 1559 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1560 // If there are too many traps at this BCI, force a recompile. 1561 // This will allow the compiler to see the limit overflow, and 1562 // take corrective action, if possible. The compiler generally 1563 // does not use the exact PerBytecodeTrapLimit value, but instead 1564 // changes its tactics if it sees any traps at all. This provides 1565 // a little hysteresis, delaying a recompile until a trap happens 1566 // several times. 1567 // 1568 // Actually, since there is only one bit of counter per BCI, 1569 // the possible per-BCI counts are {0,1,(per-method count)}. 1570 // This produces accurate results if in fact there is only 1571 // one hot trap site, but begins to get fuzzy if there are 1572 // many sites. For example, if there are ten sites each 1573 // trapping two or more times, they each get the blame for 1574 // all of their traps. 1575 make_not_entrant = true; 1576 } 1577 1578 // Detect repeated recompilation at the same BCI, and enforce a limit. 1579 if (make_not_entrant && maybe_prior_recompile) { 1580 // More than one recompile at this point. 1581 inc_recompile_count = maybe_prior_trap; 1582 } 1583 } else { 1584 // For reasons which are not recorded per-bytecode, we simply 1585 // force recompiles unconditionally. 1586 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1587 make_not_entrant = true; 1588 } 1589 1590 // Go back to the compiler if there are too many traps in this method. 1591 if (this_trap_count >= per_method_trap_limit(reason)) { 1592 // If there are too many traps in this method, force a recompile. 1593 // This will allow the compiler to see the limit overflow, and 1594 // take corrective action, if possible. 1595 // (This condition is an unlikely backstop only, because the 1596 // PerBytecodeTrapLimit is more likely to take effect first, 1597 // if it is applicable.) 1598 make_not_entrant = true; 1599 } 1600 1601 // Here's more hysteresis: If there has been a recompile at 1602 // this trap point already, run the method in the interpreter 1603 // for a while to exercise it more thoroughly. 1604 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1605 reprofile = true; 1606 } 1607 1608 } 1609 1610 // Take requested actions on the method: 1611 1612 // Recompile 1613 if (make_not_entrant) { 1614 if (!nm->make_not_entrant()) { 1615 return; // the call did not change nmethod's state 1616 } 1617 1618 if (pdata != NULL) { 1619 // Record the recompilation event, if any. 1620 int tstate0 = pdata->trap_state(); 1621 int tstate1 = trap_state_set_recompiled(tstate0, true); 1622 if (tstate1 != tstate0) 1623 pdata->set_trap_state(tstate1); 1624 } 1625 1626 #if INCLUDE_RTM_OPT 1627 // Restart collecting RTM locking abort statistic if the method 1628 // is recompiled for a reason other than RTM state change. 1629 // Assume that in new recompiled code the statistic could be different, 1630 // for example, due to different inlining. 1631 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1632 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 1633 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1634 } 1635 #endif 1636 } 1637 1638 if (inc_recompile_count) { 1639 trap_mdo->inc_overflow_recompile_count(); 1640 if ((uint)trap_mdo->overflow_recompile_count() > 1641 (uint)PerBytecodeRecompilationCutoff) { 1642 // Give up on the method containing the bad BCI. 1643 if (trap_method() == nm->method()) { 1644 make_not_compilable = true; 1645 } else { 1646 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1647 // But give grace to the enclosing nm->method(). 1648 } 1649 } 1650 } 1651 1652 // Reprofile 1653 if (reprofile) { 1654 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1655 } 1656 1657 // Give up compiling 1658 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1659 assert(make_not_entrant, "consistent"); 1660 nm->method()->set_not_compilable(CompLevel_full_optimization); 1661 } 1662 1663 } // Free marked resources 1664 1665 } 1666 JRT_END 1667 1668 MethodData* 1669 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1670 bool create_if_missing) { 1671 Thread* THREAD = thread; 1672 MethodData* mdo = m()->method_data(); 1673 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1674 // Build an MDO. Ignore errors like OutOfMemory; 1675 // that simply means we won't have an MDO to update. 1676 Method::build_interpreter_method_data(m, THREAD); 1677 if (HAS_PENDING_EXCEPTION) { 1678 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1679 CLEAR_PENDING_EXCEPTION; 1680 } 1681 mdo = m()->method_data(); 1682 } 1683 return mdo; 1684 } 1685 1686 ProfileData* 1687 Deoptimization::query_update_method_data(MethodData* trap_mdo, 1688 int trap_bci, 1689 Deoptimization::DeoptReason reason, 1690 Method* compiled_method, 1691 //outputs: 1692 uint& ret_this_trap_count, 1693 bool& ret_maybe_prior_trap, 1694 bool& ret_maybe_prior_recompile) { 1695 uint prior_trap_count = trap_mdo->trap_count(reason); 1696 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1697 1698 // If the runtime cannot find a place to store trap history, 1699 // it is estimated based on the general condition of the method. 1700 // If the method has ever been recompiled, or has ever incurred 1701 // a trap with the present reason , then this BCI is assumed 1702 // (pessimistically) to be the culprit. 1703 bool maybe_prior_trap = (prior_trap_count != 0); 1704 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1705 ProfileData* pdata = NULL; 1706 1707 1708 // For reasons which are recorded per bytecode, we check per-BCI data. 1709 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1710 if (per_bc_reason != Reason_none) { 1711 // Find the profile data for this BCI. If there isn't one, 1712 // try to allocate one from the MDO's set of spares. 1713 // This will let us detect a repeated trap at this point. 1714 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 1715 1716 if (pdata != NULL) { 1717 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 1718 if (LogCompilation && xtty != NULL) { 1719 ttyLocker ttyl; 1720 // no more room for speculative traps in this MDO 1721 xtty->elem("speculative_traps_oom"); 1722 } 1723 } 1724 // Query the trap state of this profile datum. 1725 int tstate0 = pdata->trap_state(); 1726 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1727 maybe_prior_trap = false; 1728 if (!trap_state_is_recompiled(tstate0)) 1729 maybe_prior_recompile = false; 1730 1731 // Update the trap state of this profile datum. 1732 int tstate1 = tstate0; 1733 // Record the reason. 1734 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1735 // Store the updated state on the MDO, for next time. 1736 if (tstate1 != tstate0) 1737 pdata->set_trap_state(tstate1); 1738 } else { 1739 if (LogCompilation && xtty != NULL) { 1740 ttyLocker ttyl; 1741 // Missing MDP? Leave a small complaint in the log. 1742 xtty->elem("missing_mdp bci='%d'", trap_bci); 1743 } 1744 } 1745 } 1746 1747 // Return results: 1748 ret_this_trap_count = this_trap_count; 1749 ret_maybe_prior_trap = maybe_prior_trap; 1750 ret_maybe_prior_recompile = maybe_prior_recompile; 1751 return pdata; 1752 } 1753 1754 void 1755 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1756 ResourceMark rm; 1757 // Ignored outputs: 1758 uint ignore_this_trap_count; 1759 bool ignore_maybe_prior_trap; 1760 bool ignore_maybe_prior_recompile; 1761 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 1762 query_update_method_data(trap_mdo, trap_bci, 1763 (DeoptReason)reason, 1764 NULL, 1765 ignore_this_trap_count, 1766 ignore_maybe_prior_trap, 1767 ignore_maybe_prior_recompile); 1768 } 1769 1770 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1771 1772 // Still in Java no safepoints 1773 { 1774 // This enters VM and may safepoint 1775 uncommon_trap_inner(thread, trap_request); 1776 } 1777 return fetch_unroll_info_helper(thread); 1778 } 1779 1780 // Local derived constants. 1781 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1782 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1783 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1784 1785 //---------------------------trap_state_reason--------------------------------- 1786 Deoptimization::DeoptReason 1787 Deoptimization::trap_state_reason(int trap_state) { 1788 // This assert provides the link between the width of DataLayout::trap_bits 1789 // and the encoding of "recorded" reasons. It ensures there are enough 1790 // bits to store all needed reasons in the per-BCI MDO profile. 1791 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1792 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1793 trap_state -= recompile_bit; 1794 if (trap_state == DS_REASON_MASK) { 1795 return Reason_many; 1796 } else { 1797 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1798 return (DeoptReason)trap_state; 1799 } 1800 } 1801 //-------------------------trap_state_has_reason------------------------------- 1802 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1803 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1804 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1805 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1806 trap_state -= recompile_bit; 1807 if (trap_state == DS_REASON_MASK) { 1808 return -1; // true, unspecifically (bottom of state lattice) 1809 } else if (trap_state == reason) { 1810 return 1; // true, definitely 1811 } else if (trap_state == 0) { 1812 return 0; // false, definitely (top of state lattice) 1813 } else { 1814 return 0; // false, definitely 1815 } 1816 } 1817 //-------------------------trap_state_add_reason------------------------------- 1818 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1819 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1820 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1821 trap_state -= recompile_bit; 1822 if (trap_state == DS_REASON_MASK) { 1823 return trap_state + recompile_bit; // already at state lattice bottom 1824 } else if (trap_state == reason) { 1825 return trap_state + recompile_bit; // the condition is already true 1826 } else if (trap_state == 0) { 1827 return reason + recompile_bit; // no condition has yet been true 1828 } else { 1829 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1830 } 1831 } 1832 //-----------------------trap_state_is_recompiled------------------------------ 1833 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1834 return (trap_state & DS_RECOMPILE_BIT) != 0; 1835 } 1836 //-----------------------trap_state_set_recompiled----------------------------- 1837 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1838 if (z) return trap_state | DS_RECOMPILE_BIT; 1839 else return trap_state & ~DS_RECOMPILE_BIT; 1840 } 1841 //---------------------------format_trap_state--------------------------------- 1842 // This is used for debugging and diagnostics, including LogFile output. 1843 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1844 int trap_state) { 1845 DeoptReason reason = trap_state_reason(trap_state); 1846 bool recomp_flag = trap_state_is_recompiled(trap_state); 1847 // Re-encode the state from its decoded components. 1848 int decoded_state = 0; 1849 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1850 decoded_state = trap_state_add_reason(decoded_state, reason); 1851 if (recomp_flag) 1852 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1853 // If the state re-encodes properly, format it symbolically. 1854 // Because this routine is used for debugging and diagnostics, 1855 // be robust even if the state is a strange value. 1856 size_t len; 1857 if (decoded_state != trap_state) { 1858 // Random buggy state that doesn't decode?? 1859 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1860 } else { 1861 len = jio_snprintf(buf, buflen, "%s%s", 1862 trap_reason_name(reason), 1863 recomp_flag ? " recompiled" : ""); 1864 } 1865 if (len >= buflen) 1866 buf[buflen-1] = '\0'; 1867 return buf; 1868 } 1869 1870 1871 //--------------------------------statics-------------------------------------- 1872 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1873 = Deoptimization::Action_reinterpret; 1874 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1875 // Note: Keep this in sync. with enum DeoptReason. 1876 "none", 1877 "null_check", 1878 "null_assert", 1879 "range_check", 1880 "class_check", 1881 "array_check", 1882 "intrinsic", 1883 "bimorphic", 1884 "unloaded", 1885 "uninitialized", 1886 "unreached", 1887 "unhandled", 1888 "constraint", 1889 "div0_check", 1890 "age", 1891 "predicate", 1892 "loop_limit_check", 1893 "speculate_class_check", 1894 "rtm_state_change", 1895 "unstable_if" 1896 }; 1897 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1898 // Note: Keep this in sync. with enum DeoptAction. 1899 "none", 1900 "maybe_recompile", 1901 "reinterpret", 1902 "make_not_entrant", 1903 "make_not_compilable" 1904 }; 1905 1906 const char* Deoptimization::trap_reason_name(int reason) { 1907 if (reason == Reason_many) return "many"; 1908 if ((uint)reason < Reason_LIMIT) 1909 return _trap_reason_name[reason]; 1910 static char buf[20]; 1911 sprintf(buf, "reason%d", reason); 1912 return buf; 1913 } 1914 const char* Deoptimization::trap_action_name(int action) { 1915 if ((uint)action < Action_LIMIT) 1916 return _trap_action_name[action]; 1917 static char buf[20]; 1918 sprintf(buf, "action%d", action); 1919 return buf; 1920 } 1921 1922 // This is used for debugging and diagnostics, including LogFile output. 1923 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1924 int trap_request) { 1925 jint unloaded_class_index = trap_request_index(trap_request); 1926 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1927 const char* action = trap_action_name(trap_request_action(trap_request)); 1928 size_t len; 1929 if (unloaded_class_index < 0) { 1930 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1931 reason, action); 1932 } else { 1933 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1934 reason, action, unloaded_class_index); 1935 } 1936 if (len >= buflen) 1937 buf[buflen-1] = '\0'; 1938 return buf; 1939 } 1940 1941 juint Deoptimization::_deoptimization_hist 1942 [Deoptimization::Reason_LIMIT] 1943 [1 + Deoptimization::Action_LIMIT] 1944 [Deoptimization::BC_CASE_LIMIT] 1945 = {0}; 1946 1947 enum { 1948 LSB_BITS = 8, 1949 LSB_MASK = right_n_bits(LSB_BITS) 1950 }; 1951 1952 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1953 Bytecodes::Code bc) { 1954 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1955 assert(action >= 0 && action < Action_LIMIT, "oob"); 1956 _deoptimization_hist[Reason_none][0][0] += 1; // total 1957 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1958 juint* cases = _deoptimization_hist[reason][1+action]; 1959 juint* bc_counter_addr = NULL; 1960 juint bc_counter = 0; 1961 // Look for an unused counter, or an exact match to this BC. 1962 if (bc != Bytecodes::_illegal) { 1963 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1964 juint* counter_addr = &cases[bc_case]; 1965 juint counter = *counter_addr; 1966 if ((counter == 0 && bc_counter_addr == NULL) 1967 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1968 // this counter is either free or is already devoted to this BC 1969 bc_counter_addr = counter_addr; 1970 bc_counter = counter | bc; 1971 } 1972 } 1973 } 1974 if (bc_counter_addr == NULL) { 1975 // Overflow, or no given bytecode. 1976 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1977 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1978 } 1979 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1980 } 1981 1982 jint Deoptimization::total_deoptimization_count() { 1983 return _deoptimization_hist[Reason_none][0][0]; 1984 } 1985 1986 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1987 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1988 return _deoptimization_hist[reason][0][0]; 1989 } 1990 1991 void Deoptimization::print_statistics() { 1992 juint total = total_deoptimization_count(); 1993 juint account = total; 1994 if (total != 0) { 1995 ttyLocker ttyl; 1996 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1997 tty->print_cr("Deoptimization traps recorded:"); 1998 #define PRINT_STAT_LINE(name, r) \ 1999 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 2000 PRINT_STAT_LINE("total", total); 2001 // For each non-zero entry in the histogram, print the reason, 2002 // the action, and (if specifically known) the type of bytecode. 2003 for (int reason = 0; reason < Reason_LIMIT; reason++) { 2004 for (int action = 0; action < Action_LIMIT; action++) { 2005 juint* cases = _deoptimization_hist[reason][1+action]; 2006 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 2007 juint counter = cases[bc_case]; 2008 if (counter != 0) { 2009 char name[1*K]; 2010 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2011 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2012 bc = Bytecodes::_illegal; 2013 sprintf(name, "%s/%s/%s", 2014 trap_reason_name(reason), 2015 trap_action_name(action), 2016 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2017 juint r = counter >> LSB_BITS; 2018 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2019 account -= r; 2020 } 2021 } 2022 } 2023 } 2024 if (account != 0) { 2025 PRINT_STAT_LINE("unaccounted", account); 2026 } 2027 #undef PRINT_STAT_LINE 2028 if (xtty != NULL) xtty->tail("statistics"); 2029 } 2030 } 2031 #else // COMPILER2 || SHARK 2032 2033 2034 // Stubs for C1 only system. 2035 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2036 return false; 2037 } 2038 2039 const char* Deoptimization::trap_reason_name(int reason) { 2040 return "unknown"; 2041 } 2042 2043 void Deoptimization::print_statistics() { 2044 // no output 2045 } 2046 2047 void 2048 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2049 // no udpate 2050 } 2051 2052 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2053 return 0; 2054 } 2055 2056 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2057 Bytecodes::Code bc) { 2058 // no update 2059 } 2060 2061 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2062 int trap_state) { 2063 jio_snprintf(buf, buflen, "#%d", trap_state); 2064 return buf; 2065 } 2066 2067 #endif // COMPILER2 || SHARK