1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "code/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 #if defined AD_MD_HPP 70 # include AD_MD_HPP 71 #elif defined TARGET_ARCH_MODEL_x86_32 72 # include "adfiles/ad_x86_32.hpp" 73 #elif defined TARGET_ARCH_MODEL_x86_64 74 # include "adfiles/ad_x86_64.hpp" 75 #elif defined TARGET_ARCH_MODEL_sparc 76 # include "adfiles/ad_sparc.hpp" 77 #elif defined TARGET_ARCH_MODEL_zero 78 # include "adfiles/ad_zero.hpp" 79 #elif defined TARGET_ARCH_MODEL_ppc_64 80 # include "adfiles/ad_ppc_64.hpp" 81 #endif 82 #endif // COMPILER2 83 84 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC 85 86 bool DeoptimizationMarker::_is_active = false; 87 88 Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 89 int caller_adjustment, 90 int caller_actual_parameters, 91 int number_of_frames, 92 intptr_t* frame_sizes, 93 address* frame_pcs, 94 BasicType return_type) { 95 _size_of_deoptimized_frame = size_of_deoptimized_frame; 96 _caller_adjustment = caller_adjustment; 97 _caller_actual_parameters = caller_actual_parameters; 98 _number_of_frames = number_of_frames; 99 _frame_sizes = frame_sizes; 100 _frame_pcs = frame_pcs; 101 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 102 _return_type = return_type; 103 _initial_info = 0; 104 // PD (x86 only) 105 _counter_temp = 0; 106 _unpack_kind = 0; 107 _sender_sp_temp = 0; 108 109 _total_frame_sizes = size_of_frames(); 110 } 111 112 113 Deoptimization::UnrollBlock::~UnrollBlock() { 114 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler); 115 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler); 116 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler); 117 } 118 119 120 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 121 assert(register_number < RegisterMap::reg_count, "checking register number"); 122 return &_register_block[register_number * 2]; 123 } 124 125 126 127 int Deoptimization::UnrollBlock::size_of_frames() const { 128 // Acount first for the adjustment of the initial frame 129 int result = _caller_adjustment; 130 for (int index = 0; index < number_of_frames(); index++) { 131 result += frame_sizes()[index]; 132 } 133 return result; 134 } 135 136 137 void Deoptimization::UnrollBlock::print() { 138 ttyLocker ttyl; 139 tty->print_cr("UnrollBlock"); 140 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 141 tty->print( " frame_sizes: "); 142 for (int index = 0; index < number_of_frames(); index++) { 143 tty->print("%d ", frame_sizes()[index]); 144 } 145 tty->cr(); 146 } 147 148 149 // In order to make fetch_unroll_info work properly with escape 150 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 151 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 152 // of previously eliminated objects occurs in realloc_objects, which is 153 // called from the method fetch_unroll_info_helper below. 154 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 155 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 156 // but makes the entry a little slower. There is however a little dance we have to 157 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 158 159 // fetch_unroll_info() is called at the beginning of the deoptimization 160 // handler. Note this fact before we start generating temporary frames 161 // that can confuse an asynchronous stack walker. This counter is 162 // decremented at the end of unpack_frames(). 163 thread->inc_in_deopt_handler(); 164 165 return fetch_unroll_info_helper(thread); 166 JRT_END 167 168 169 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 170 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 171 172 // Note: there is a safepoint safety issue here. No matter whether we enter 173 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 174 // the vframeArray is created. 175 // 176 177 // Allocate our special deoptimization ResourceMark 178 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 179 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 180 thread->set_deopt_mark(dmark); 181 182 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 183 RegisterMap map(thread, true); 184 RegisterMap dummy_map(thread, false); 185 // Now get the deoptee with a valid map 186 frame deoptee = stub_frame.sender(&map); 187 // Set the deoptee nmethod 188 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 189 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 190 191 if (VerifyStack) { 192 thread->validate_frame_layout(); 193 } 194 195 // Create a growable array of VFrames where each VFrame represents an inlined 196 // Java frame. This storage is allocated with the usual system arena. 197 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 198 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 199 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 200 while (!vf->is_top()) { 201 assert(vf->is_compiled_frame(), "Wrong frame type"); 202 chunk->push(compiledVFrame::cast(vf)); 203 vf = vf->sender(); 204 } 205 assert(vf->is_compiled_frame(), "Wrong frame type"); 206 chunk->push(compiledVFrame::cast(vf)); 207 208 bool realloc_failures = false; 209 210 #ifdef COMPILER2 211 // Reallocate the non-escaping objects and restore their fields. Then 212 // relock objects if synchronization on them was eliminated. 213 if (DoEscapeAnalysis || EliminateNestedLocks) { 214 if (EliminateAllocations) { 215 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 216 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 217 218 // The flag return_oop() indicates call sites which return oop 219 // in compiled code. Such sites include java method calls, 220 // runtime calls (for example, used to allocate new objects/arrays 221 // on slow code path) and any other calls generated in compiled code. 222 // It is not guaranteed that we can get such information here only 223 // by analyzing bytecode in deoptimized frames. This is why this flag 224 // is set during method compilation (see Compile::Process_OopMap_Node()). 225 // If the previous frame was popped, we don't have a result. 226 bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution(); 227 Handle return_value; 228 if (save_oop_result) { 229 // Reallocation may trigger GC. If deoptimization happened on return from 230 // call which returns oop we need to save it since it is not in oopmap. 231 oop result = deoptee.saved_oop_result(&map); 232 assert(result == NULL || result->is_oop(), "must be oop"); 233 return_value = Handle(thread, result); 234 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 235 if (TraceDeoptimization) { 236 ttyLocker ttyl; 237 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, (void *)result, thread); 238 } 239 } 240 if (objects != NULL) { 241 JRT_BLOCK 242 realloc_failures = realloc_objects(thread, &deoptee, objects, THREAD); 243 JRT_END 244 reassign_fields(&deoptee, &map, objects, realloc_failures); 245 #ifndef PRODUCT 246 if (TraceDeoptimization) { 247 ttyLocker ttyl; 248 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 249 print_objects(objects, realloc_failures); 250 } 251 #endif 252 } 253 if (save_oop_result) { 254 // Restore result. 255 deoptee.set_saved_oop_result(&map, return_value()); 256 } 257 } 258 if (EliminateLocks) { 259 #ifndef PRODUCT 260 bool first = true; 261 #endif 262 for (int i = 0; i < chunk->length(); i++) { 263 compiledVFrame* cvf = chunk->at(i); 264 assert (cvf->scope() != NULL,"expect only compiled java frames"); 265 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 266 if (monitors->is_nonempty()) { 267 relock_objects(monitors, thread, realloc_failures); 268 #ifndef PRODUCT 269 if (TraceDeoptimization) { 270 ttyLocker ttyl; 271 for (int j = 0; j < monitors->length(); j++) { 272 MonitorInfo* mi = monitors->at(j); 273 if (mi->eliminated()) { 274 if (first) { 275 first = false; 276 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 277 } 278 if (mi->owner_is_scalar_replaced()) { 279 Klass* k = java_lang_Class::as_Klass(mi->owner_klass()); 280 tty->print_cr(" failed reallocation for klass %s", k->external_name()); 281 } else { 282 tty->print_cr(" object <" INTPTR_FORMAT "> locked", (void *)mi->owner()); 283 } 284 } 285 } 286 } 287 #endif 288 } 289 } 290 } 291 } 292 #endif // COMPILER2 293 // Ensure that no safepoint is taken after pointers have been stored 294 // in fields of rematerialized objects. If a safepoint occurs from here on 295 // out the java state residing in the vframeArray will be missed. 296 No_Safepoint_Verifier no_safepoint; 297 298 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures); 299 #ifdef COMPILER2 300 if (realloc_failures) { 301 pop_frames_failed_reallocs(thread, array); 302 } 303 #endif 304 305 assert(thread->vframe_array_head() == NULL, "Pending deopt!"); 306 thread->set_vframe_array_head(array); 307 308 // Now that the vframeArray has been created if we have any deferred local writes 309 // added by jvmti then we can free up that structure as the data is now in the 310 // vframeArray 311 312 if (thread->deferred_locals() != NULL) { 313 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 314 int i = 0; 315 do { 316 // Because of inlining we could have multiple vframes for a single frame 317 // and several of the vframes could have deferred writes. Find them all. 318 if (list->at(i)->id() == array->original().id()) { 319 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 320 list->remove_at(i); 321 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 322 delete dlv; 323 } else { 324 i++; 325 } 326 } while ( i < list->length() ); 327 if (list->length() == 0) { 328 thread->set_deferred_locals(NULL); 329 // free the list and elements back to C heap. 330 delete list; 331 } 332 333 } 334 335 #ifndef SHARK 336 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 337 CodeBlob* cb = stub_frame.cb(); 338 // Verify we have the right vframeArray 339 assert(cb->frame_size() >= 0, "Unexpected frame size"); 340 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 341 342 // If the deopt call site is a MethodHandle invoke call site we have 343 // to adjust the unpack_sp. 344 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 345 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 346 unpack_sp = deoptee.unextended_sp(); 347 348 #ifdef ASSERT 349 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 350 #endif 351 #else 352 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 353 #endif // !SHARK 354 355 // This is a guarantee instead of an assert because if vframe doesn't match 356 // we will unpack the wrong deoptimized frame and wind up in strange places 357 // where it will be very difficult to figure out what went wrong. Better 358 // to die an early death here than some very obscure death later when the 359 // trail is cold. 360 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 361 // in that it will fail to detect a problem when there is one. This needs 362 // more work in tiger timeframe. 363 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 364 365 int number_of_frames = array->frames(); 366 367 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 368 // virtual activation, which is the reverse of the elements in the vframes array. 369 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 370 // +1 because we always have an interpreter return address for the final slot. 371 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 372 int popframe_extra_args = 0; 373 // Create an interpreter return address for the stub to use as its return 374 // address so the skeletal frames are perfectly walkable 375 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 376 377 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 378 // activation be put back on the expression stack of the caller for reexecution 379 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 380 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 381 } 382 383 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 384 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 385 // than simply use array->sender.pc(). This requires us to walk the current set of frames 386 // 387 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 388 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 389 390 // It's possible that the number of paramters at the call site is 391 // different than number of arguments in the callee when method 392 // handles are used. If the caller is interpreted get the real 393 // value so that the proper amount of space can be added to it's 394 // frame. 395 bool caller_was_method_handle = false; 396 if (deopt_sender.is_interpreted_frame()) { 397 methodHandle method = deopt_sender.interpreter_frame_method(); 398 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 399 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 400 // Method handle invokes may involve fairly arbitrary chains of 401 // calls so it's impossible to know how much actual space the 402 // caller has for locals. 403 caller_was_method_handle = true; 404 } 405 } 406 407 // 408 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 409 // frame_sizes/frame_pcs[1] next oldest frame (int) 410 // frame_sizes/frame_pcs[n] youngest frame (int) 411 // 412 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 413 // owns the space for the return address to it's caller). Confusing ain't it. 414 // 415 // The vframe array can address vframes with indices running from 416 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 417 // When we create the skeletal frames we need the oldest frame to be in the zero slot 418 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 419 // so things look a little strange in this loop. 420 // 421 int callee_parameters = 0; 422 int callee_locals = 0; 423 for (int index = 0; index < array->frames(); index++ ) { 424 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 425 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 426 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 427 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 428 callee_locals, 429 index == 0, 430 popframe_extra_args); 431 // This pc doesn't have to be perfect just good enough to identify the frame 432 // as interpreted so the skeleton frame will be walkable 433 // The correct pc will be set when the skeleton frame is completely filled out 434 // The final pc we store in the loop is wrong and will be overwritten below 435 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 436 437 callee_parameters = array->element(index)->method()->size_of_parameters(); 438 callee_locals = array->element(index)->method()->max_locals(); 439 popframe_extra_args = 0; 440 } 441 442 // Compute whether the root vframe returns a float or double value. 443 BasicType return_type; 444 { 445 HandleMark hm; 446 methodHandle method(thread, array->element(0)->method()); 447 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 448 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 449 } 450 451 // Compute information for handling adapters and adjusting the frame size of the caller. 452 int caller_adjustment = 0; 453 454 // Compute the amount the oldest interpreter frame will have to adjust 455 // its caller's stack by. If the caller is a compiled frame then 456 // we pretend that the callee has no parameters so that the 457 // extension counts for the full amount of locals and not just 458 // locals-parms. This is because without a c2i adapter the parm 459 // area as created by the compiled frame will not be usable by 460 // the interpreter. (Depending on the calling convention there 461 // may not even be enough space). 462 463 // QQQ I'd rather see this pushed down into last_frame_adjust 464 // and have it take the sender (aka caller). 465 466 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 467 caller_adjustment = last_frame_adjust(0, callee_locals); 468 } else if (callee_locals > callee_parameters) { 469 // The caller frame may need extending to accommodate 470 // non-parameter locals of the first unpacked interpreted frame. 471 // Compute that adjustment. 472 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 473 } 474 475 // If the sender is deoptimized the we must retrieve the address of the handler 476 // since the frame will "magically" show the original pc before the deopt 477 // and we'd undo the deopt. 478 479 frame_pcs[0] = deopt_sender.raw_pc(); 480 481 #ifndef SHARK 482 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 483 #endif // SHARK 484 485 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 486 caller_adjustment * BytesPerWord, 487 caller_was_method_handle ? 0 : callee_parameters, 488 number_of_frames, 489 frame_sizes, 490 frame_pcs, 491 return_type); 492 // On some platforms, we need a way to pass some platform dependent 493 // information to the unpacking code so the skeletal frames come out 494 // correct (initial fp value, unextended sp, ...) 495 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 496 497 if (array->frames() > 1) { 498 if (VerifyStack && TraceDeoptimization) { 499 ttyLocker ttyl; 500 tty->print_cr("Deoptimizing method containing inlining"); 501 } 502 } 503 504 array->set_unroll_block(info); 505 return info; 506 } 507 508 // Called to cleanup deoptimization data structures in normal case 509 // after unpacking to stack and when stack overflow error occurs 510 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 511 vframeArray *array) { 512 513 // Get array if coming from exception 514 if (array == NULL) { 515 array = thread->vframe_array_head(); 516 } 517 thread->set_vframe_array_head(NULL); 518 519 // Free the previous UnrollBlock 520 vframeArray* old_array = thread->vframe_array_last(); 521 thread->set_vframe_array_last(array); 522 523 if (old_array != NULL) { 524 UnrollBlock* old_info = old_array->unroll_block(); 525 old_array->set_unroll_block(NULL); 526 delete old_info; 527 delete old_array; 528 } 529 530 // Deallocate any resource creating in this routine and any ResourceObjs allocated 531 // inside the vframeArray (StackValueCollections) 532 533 delete thread->deopt_mark(); 534 thread->set_deopt_mark(NULL); 535 thread->set_deopt_nmethod(NULL); 536 537 538 if (JvmtiExport::can_pop_frame()) { 539 #ifndef CC_INTERP 540 // Regardless of whether we entered this routine with the pending 541 // popframe condition bit set, we should always clear it now 542 thread->clear_popframe_condition(); 543 #else 544 // C++ interpeter will clear has_pending_popframe when it enters 545 // with method_resume. For deopt_resume2 we clear it now. 546 if (thread->popframe_forcing_deopt_reexecution()) 547 thread->clear_popframe_condition(); 548 #endif /* CC_INTERP */ 549 } 550 551 // unpack_frames() is called at the end of the deoptimization handler 552 // and (in C2) at the end of the uncommon trap handler. Note this fact 553 // so that an asynchronous stack walker can work again. This counter is 554 // incremented at the beginning of fetch_unroll_info() and (in C2) at 555 // the beginning of uncommon_trap(). 556 thread->dec_in_deopt_handler(); 557 } 558 559 560 // Return BasicType of value being returned 561 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 562 563 // We are already active int he special DeoptResourceMark any ResourceObj's we 564 // allocate will be freed at the end of the routine. 565 566 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 567 // but makes the entry a little slower. There is however a little dance we have to 568 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 569 ResetNoHandleMark rnhm; // No-op in release/product versions 570 HandleMark hm; 571 572 frame stub_frame = thread->last_frame(); 573 574 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 575 // must point to the vframeArray for the unpack frame. 576 vframeArray* array = thread->vframe_array_head(); 577 578 #ifndef PRODUCT 579 if (TraceDeoptimization) { 580 ttyLocker ttyl; 581 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 582 } 583 #endif 584 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 585 stub_frame.pc(), stub_frame.sp(), exec_mode); 586 587 UnrollBlock* info = array->unroll_block(); 588 589 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 590 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 591 592 BasicType bt = info->return_type(); 593 594 // If we have an exception pending, claim that the return type is an oop 595 // so the deopt_blob does not overwrite the exception_oop. 596 597 if (exec_mode == Unpack_exception) 598 bt = T_OBJECT; 599 600 // Cleanup thread deopt data 601 cleanup_deopt_info(thread, array); 602 603 #ifndef PRODUCT 604 if (VerifyStack) { 605 ResourceMark res_mark; 606 607 thread->validate_frame_layout(); 608 609 // Verify that the just-unpacked frames match the interpreter's 610 // notions of expression stack and locals 611 vframeArray* cur_array = thread->vframe_array_last(); 612 RegisterMap rm(thread, false); 613 rm.set_include_argument_oops(false); 614 bool is_top_frame = true; 615 int callee_size_of_parameters = 0; 616 int callee_max_locals = 0; 617 for (int i = 0; i < cur_array->frames(); i++) { 618 vframeArrayElement* el = cur_array->element(i); 619 frame* iframe = el->iframe(); 620 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 621 622 // Get the oop map for this bci 623 InterpreterOopMap mask; 624 int cur_invoke_parameter_size = 0; 625 bool try_next_mask = false; 626 int next_mask_expression_stack_size = -1; 627 int top_frame_expression_stack_adjustment = 0; 628 methodHandle mh(thread, iframe->interpreter_frame_method()); 629 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 630 BytecodeStream str(mh); 631 str.set_start(iframe->interpreter_frame_bci()); 632 int max_bci = mh->code_size(); 633 // Get to the next bytecode if possible 634 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 635 // Check to see if we can grab the number of outgoing arguments 636 // at an uncommon trap for an invoke (where the compiler 637 // generates debug info before the invoke has executed) 638 Bytecodes::Code cur_code = str.next(); 639 if (cur_code == Bytecodes::_invokevirtual || 640 cur_code == Bytecodes::_invokespecial || 641 cur_code == Bytecodes::_invokestatic || 642 cur_code == Bytecodes::_invokeinterface || 643 cur_code == Bytecodes::_invokedynamic) { 644 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 645 Symbol* signature = invoke.signature(); 646 ArgumentSizeComputer asc(signature); 647 cur_invoke_parameter_size = asc.size(); 648 if (invoke.has_receiver()) { 649 // Add in receiver 650 ++cur_invoke_parameter_size; 651 } 652 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 653 callee_size_of_parameters++; 654 } 655 } 656 if (str.bci() < max_bci) { 657 Bytecodes::Code bc = str.next(); 658 if (bc >= 0) { 659 // The interpreter oop map generator reports results before 660 // the current bytecode has executed except in the case of 661 // calls. It seems to be hard to tell whether the compiler 662 // has emitted debug information matching the "state before" 663 // a given bytecode or the state after, so we try both 664 switch (cur_code) { 665 case Bytecodes::_invokevirtual: 666 case Bytecodes::_invokespecial: 667 case Bytecodes::_invokestatic: 668 case Bytecodes::_invokeinterface: 669 case Bytecodes::_invokedynamic: 670 case Bytecodes::_athrow: 671 break; 672 default: { 673 InterpreterOopMap next_mask; 674 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 675 next_mask_expression_stack_size = next_mask.expression_stack_size(); 676 // Need to subtract off the size of the result type of 677 // the bytecode because this is not described in the 678 // debug info but returned to the interpreter in the TOS 679 // caching register 680 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 681 if (bytecode_result_type != T_ILLEGAL) { 682 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 683 } 684 assert(top_frame_expression_stack_adjustment >= 0, ""); 685 try_next_mask = true; 686 break; 687 } 688 } 689 } 690 } 691 692 // Verify stack depth and oops in frame 693 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 694 if (!( 695 /* SPARC */ 696 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 697 /* x86 */ 698 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 699 (try_next_mask && 700 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 701 top_frame_expression_stack_adjustment))) || 702 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 703 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) && 704 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 705 )) { 706 ttyLocker ttyl; 707 708 // Print out some information that will help us debug the problem 709 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 710 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 711 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 712 iframe->interpreter_frame_expression_stack_size()); 713 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 714 tty->print_cr(" try_next_mask = %d", try_next_mask); 715 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 716 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 717 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 718 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 719 tty->print_cr(" exec_mode = %d", exec_mode); 720 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 721 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 722 tty->print_cr(" Interpreted frames:"); 723 for (int k = 0; k < cur_array->frames(); k++) { 724 vframeArrayElement* el = cur_array->element(k); 725 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 726 } 727 cur_array->print_on_2(tty); 728 guarantee(false, "wrong number of expression stack elements during deopt"); 729 } 730 VerifyOopClosure verify; 731 iframe->oops_interpreted_do(&verify, NULL, &rm, false); 732 callee_size_of_parameters = mh->size_of_parameters(); 733 callee_max_locals = mh->max_locals(); 734 is_top_frame = false; 735 } 736 } 737 #endif /* !PRODUCT */ 738 739 740 return bt; 741 JRT_END 742 743 744 int Deoptimization::deoptimize_dependents() { 745 Threads::deoptimized_wrt_marked_nmethods(); 746 return 0; 747 } 748 749 750 #ifdef COMPILER2 751 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 752 Handle pending_exception(thread->pending_exception()); 753 const char* exception_file = thread->exception_file(); 754 int exception_line = thread->exception_line(); 755 thread->clear_pending_exception(); 756 757 bool failures = false; 758 759 for (int i = 0; i < objects->length(); i++) { 760 assert(objects->at(i)->is_object(), "invalid debug information"); 761 ObjectValue* sv = (ObjectValue*) objects->at(i); 762 763 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 764 oop obj = NULL; 765 766 if (k->oop_is_instance()) { 767 InstanceKlass* ik = InstanceKlass::cast(k()); 768 obj = ik->allocate_instance(THREAD); 769 } else if (k->oop_is_typeArray()) { 770 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 771 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 772 int len = sv->field_size() / type2size[ak->element_type()]; 773 obj = ak->allocate(len, THREAD); 774 } else if (k->oop_is_objArray()) { 775 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 776 obj = ak->allocate(sv->field_size(), THREAD); 777 } 778 779 if (obj == NULL) { 780 failures = true; 781 } 782 783 assert(sv->value().is_null(), "redundant reallocation"); 784 assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception"); 785 CLEAR_PENDING_EXCEPTION; 786 sv->set_value(obj); 787 } 788 789 if (failures) { 790 THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures); 791 } else if (pending_exception.not_null()) { 792 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 793 } 794 795 return failures; 796 } 797 798 // This assumes that the fields are stored in ObjectValue in the same order 799 // they are yielded by do_nonstatic_fields. 800 class FieldReassigner: public FieldClosure { 801 frame* _fr; 802 RegisterMap* _reg_map; 803 ObjectValue* _sv; 804 InstanceKlass* _ik; 805 oop _obj; 806 807 int _i; 808 public: 809 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 810 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 811 812 int i() const { return _i; } 813 814 815 void do_field(fieldDescriptor* fd) { 816 intptr_t val; 817 StackValue* value = 818 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 819 int offset = fd->offset(); 820 switch (fd->field_type()) { 821 case T_OBJECT: case T_ARRAY: 822 assert(value->type() == T_OBJECT, "Agreement."); 823 _obj->obj_field_put(offset, value->get_obj()()); 824 break; 825 826 case T_LONG: case T_DOUBLE: { 827 assert(value->type() == T_INT, "Agreement."); 828 StackValue* low = 829 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 830 #ifdef _LP64 831 jlong res = (jlong)low->get_int(); 832 #else 833 #ifdef SPARC 834 // For SPARC we have to swap high and low words. 835 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 836 #else 837 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 838 #endif //SPARC 839 #endif 840 _obj->long_field_put(offset, res); 841 break; 842 } 843 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 844 case T_INT: case T_FLOAT: // 4 bytes. 845 assert(value->type() == T_INT, "Agreement."); 846 val = value->get_int(); 847 _obj->int_field_put(offset, (jint)*((jint*)&val)); 848 break; 849 850 case T_SHORT: case T_CHAR: // 2 bytes 851 assert(value->type() == T_INT, "Agreement."); 852 val = value->get_int(); 853 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 854 break; 855 856 case T_BOOLEAN: case T_BYTE: // 1 byte 857 assert(value->type() == T_INT, "Agreement."); 858 val = value->get_int(); 859 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 860 break; 861 862 default: 863 ShouldNotReachHere(); 864 } 865 _i++; 866 } 867 }; 868 869 // restore elements of an eliminated type array 870 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 871 int index = 0; 872 intptr_t val; 873 874 for (int i = 0; i < sv->field_size(); i++) { 875 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 876 switch(type) { 877 case T_LONG: case T_DOUBLE: { 878 assert(value->type() == T_INT, "Agreement."); 879 StackValue* low = 880 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 881 #ifdef _LP64 882 jlong res = (jlong)low->get_int(); 883 #else 884 #ifdef SPARC 885 // For SPARC we have to swap high and low words. 886 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 887 #else 888 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 889 #endif //SPARC 890 #endif 891 obj->long_at_put(index, res); 892 break; 893 } 894 895 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 896 case T_INT: case T_FLOAT: // 4 bytes. 897 assert(value->type() == T_INT, "Agreement."); 898 val = value->get_int(); 899 obj->int_at_put(index, (jint)*((jint*)&val)); 900 break; 901 902 case T_SHORT: case T_CHAR: // 2 bytes 903 assert(value->type() == T_INT, "Agreement."); 904 val = value->get_int(); 905 obj->short_at_put(index, (jshort)*((jint*)&val)); 906 break; 907 908 case T_BOOLEAN: case T_BYTE: // 1 byte 909 assert(value->type() == T_INT, "Agreement."); 910 val = value->get_int(); 911 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 912 break; 913 914 default: 915 ShouldNotReachHere(); 916 } 917 index++; 918 } 919 } 920 921 922 // restore fields of an eliminated object array 923 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 924 for (int i = 0; i < sv->field_size(); i++) { 925 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 926 assert(value->type() == T_OBJECT, "object element expected"); 927 obj->obj_at_put(i, value->get_obj()()); 928 } 929 } 930 931 932 // restore fields of all eliminated objects and arrays 933 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 934 for (int i = 0; i < objects->length(); i++) { 935 ObjectValue* sv = (ObjectValue*) objects->at(i); 936 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 937 Handle obj = sv->value(); 938 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 939 if (obj.is_null()) { 940 continue; 941 } 942 943 if (k->oop_is_instance()) { 944 InstanceKlass* ik = InstanceKlass::cast(k()); 945 FieldReassigner reassign(fr, reg_map, sv, obj()); 946 ik->do_nonstatic_fields(&reassign); 947 } else if (k->oop_is_typeArray()) { 948 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 949 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 950 } else if (k->oop_is_objArray()) { 951 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 952 } 953 } 954 } 955 956 957 // relock objects for which synchronization was eliminated 958 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread, bool realloc_failures) { 959 for (int i = 0; i < monitors->length(); i++) { 960 MonitorInfo* mon_info = monitors->at(i); 961 if (mon_info->eliminated()) { 962 assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed"); 963 if (!mon_info->owner_is_scalar_replaced()) { 964 Handle obj = Handle(mon_info->owner()); 965 markOop mark = obj->mark(); 966 if (UseBiasedLocking && mark->has_bias_pattern()) { 967 // New allocated objects may have the mark set to anonymously biased. 968 // Also the deoptimized method may called methods with synchronization 969 // where the thread-local object is bias locked to the current thread. 970 assert(mark->is_biased_anonymously() || 971 mark->biased_locker() == thread, "should be locked to current thread"); 972 // Reset mark word to unbiased prototype. 973 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 974 obj->set_mark(unbiased_prototype); 975 } 976 BasicLock* lock = mon_info->lock(); 977 ObjectSynchronizer::slow_enter(obj, lock, thread); 978 assert(mon_info->owner()->is_locked(), "object must be locked now"); 979 } 980 } 981 } 982 } 983 984 985 #ifndef PRODUCT 986 // print information about reallocated objects 987 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) { 988 fieldDescriptor fd; 989 990 for (int i = 0; i < objects->length(); i++) { 991 ObjectValue* sv = (ObjectValue*) objects->at(i); 992 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 993 Handle obj = sv->value(); 994 995 tty->print(" object <" INTPTR_FORMAT "> of type ", (void *)sv->value()()); 996 k->print_value(); 997 assert(obj.not_null() || realloc_failures, "reallocation was missed"); 998 if (obj.is_null()) { 999 tty->print(" allocation failed"); 1000 } else { 1001 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 1002 } 1003 tty->cr(); 1004 1005 if (Verbose && !obj.is_null()) { 1006 k->oop_print_on(obj(), tty); 1007 } 1008 } 1009 } 1010 #endif 1011 #endif // COMPILER2 1012 1013 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) { 1014 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 1015 1016 #ifndef PRODUCT 1017 if (TraceDeoptimization) { 1018 ttyLocker ttyl; 1019 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 1020 fr.print_on(tty); 1021 tty->print_cr(" Virtual frames (innermost first):"); 1022 for (int index = 0; index < chunk->length(); index++) { 1023 compiledVFrame* vf = chunk->at(index); 1024 tty->print(" %2d - ", index); 1025 vf->print_value(); 1026 int bci = chunk->at(index)->raw_bci(); 1027 const char* code_name; 1028 if (bci == SynchronizationEntryBCI) { 1029 code_name = "sync entry"; 1030 } else { 1031 Bytecodes::Code code = vf->method()->code_at(bci); 1032 code_name = Bytecodes::name(code); 1033 } 1034 tty->print(" - %s", code_name); 1035 tty->print_cr(" @ bci %d ", bci); 1036 if (Verbose) { 1037 vf->print(); 1038 tty->cr(); 1039 } 1040 } 1041 } 1042 #endif 1043 1044 // Register map for next frame (used for stack crawl). We capture 1045 // the state of the deopt'ing frame's caller. Thus if we need to 1046 // stuff a C2I adapter we can properly fill in the callee-save 1047 // register locations. 1048 frame caller = fr.sender(reg_map); 1049 int frame_size = caller.sp() - fr.sp(); 1050 1051 frame sender = caller; 1052 1053 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1054 // the vframeArray containing the unpacking information is allocated in the C heap. 1055 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1056 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures); 1057 1058 // Compare the vframeArray to the collected vframes 1059 assert(array->structural_compare(thread, chunk), "just checking"); 1060 1061 #ifndef PRODUCT 1062 if (TraceDeoptimization) { 1063 ttyLocker ttyl; 1064 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1065 } 1066 #endif // PRODUCT 1067 1068 return array; 1069 } 1070 1071 #ifdef COMPILER2 1072 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) { 1073 // Reallocation of some scalar replaced objects failed. Record 1074 // that we need to pop all the interpreter frames for the 1075 // deoptimized compiled frame. 1076 assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?"); 1077 thread->set_frames_to_pop_failed_realloc(array->frames()); 1078 // Unlock all monitors here otherwise the interpreter will see a 1079 // mix of locked and unlocked monitors (because of failed 1080 // reallocations of synchronized objects) and be confused. 1081 for (int i = 0; i < array->frames(); i++) { 1082 MonitorChunk* monitors = array->element(i)->monitors(); 1083 if (monitors != NULL) { 1084 for (int j = 0; j < monitors->number_of_monitors(); j++) { 1085 BasicObjectLock* src = monitors->at(j); 1086 if (src->obj() != NULL) { 1087 ObjectSynchronizer::fast_exit(src->obj(), src->lock(), thread); 1088 } 1089 } 1090 array->element(i)->free_monitors(thread); 1091 #ifdef ASSERT 1092 array->element(i)->set_removed_monitors(); 1093 #endif 1094 } 1095 } 1096 } 1097 #endif 1098 1099 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1100 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1101 for (int i = 0; i < monitors->length(); i++) { 1102 MonitorInfo* mon_info = monitors->at(i); 1103 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1104 objects_to_revoke->append(Handle(mon_info->owner())); 1105 } 1106 } 1107 } 1108 1109 1110 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1111 if (!UseBiasedLocking) { 1112 return; 1113 } 1114 1115 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1116 1117 // Unfortunately we don't have a RegisterMap available in most of 1118 // the places we want to call this routine so we need to walk the 1119 // stack again to update the register map. 1120 if (map == NULL || !map->update_map()) { 1121 StackFrameStream sfs(thread, true); 1122 bool found = false; 1123 while (!found && !sfs.is_done()) { 1124 frame* cur = sfs.current(); 1125 sfs.next(); 1126 found = cur->id() == fr.id(); 1127 } 1128 assert(found, "frame to be deoptimized not found on target thread's stack"); 1129 map = sfs.register_map(); 1130 } 1131 1132 vframe* vf = vframe::new_vframe(&fr, map, thread); 1133 compiledVFrame* cvf = compiledVFrame::cast(vf); 1134 // Revoke monitors' biases in all scopes 1135 while (!cvf->is_top()) { 1136 collect_monitors(cvf, objects_to_revoke); 1137 cvf = compiledVFrame::cast(cvf->sender()); 1138 } 1139 collect_monitors(cvf, objects_to_revoke); 1140 1141 if (SafepointSynchronize::is_at_safepoint()) { 1142 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1143 } else { 1144 BiasedLocking::revoke(objects_to_revoke); 1145 } 1146 } 1147 1148 1149 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1150 if (!UseBiasedLocking) { 1151 return; 1152 } 1153 1154 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1155 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1156 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1157 if (jt->has_last_Java_frame()) { 1158 StackFrameStream sfs(jt, true); 1159 while (!sfs.is_done()) { 1160 frame* cur = sfs.current(); 1161 if (cb->contains(cur->pc())) { 1162 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1163 compiledVFrame* cvf = compiledVFrame::cast(vf); 1164 // Revoke monitors' biases in all scopes 1165 while (!cvf->is_top()) { 1166 collect_monitors(cvf, objects_to_revoke); 1167 cvf = compiledVFrame::cast(cvf->sender()); 1168 } 1169 collect_monitors(cvf, objects_to_revoke); 1170 } 1171 sfs.next(); 1172 } 1173 } 1174 } 1175 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1176 } 1177 1178 1179 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1180 assert(fr.can_be_deoptimized(), "checking frame type"); 1181 1182 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1183 1184 // Patch the nmethod so that when execution returns to it we will 1185 // deopt the execution state and return to the interpreter. 1186 fr.deoptimize(thread); 1187 } 1188 1189 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1190 // Deoptimize only if the frame comes from compile code. 1191 // Do not deoptimize the frame which is already patched 1192 // during the execution of the loops below. 1193 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1194 return; 1195 } 1196 ResourceMark rm; 1197 DeoptimizationMarker dm; 1198 if (UseBiasedLocking) { 1199 revoke_biases_of_monitors(thread, fr, map); 1200 } 1201 deoptimize_single_frame(thread, fr); 1202 1203 } 1204 1205 1206 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1207 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1208 "can only deoptimize other thread at a safepoint"); 1209 // Compute frame and register map based on thread and sp. 1210 RegisterMap reg_map(thread, UseBiasedLocking); 1211 frame fr = thread->last_frame(); 1212 while (fr.id() != id) { 1213 fr = fr.sender(®_map); 1214 } 1215 deoptimize(thread, fr, ®_map); 1216 } 1217 1218 1219 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1220 if (thread == Thread::current()) { 1221 Deoptimization::deoptimize_frame_internal(thread, id); 1222 } else { 1223 VM_DeoptimizeFrame deopt(thread, id); 1224 VMThread::execute(&deopt); 1225 } 1226 } 1227 1228 1229 // JVMTI PopFrame support 1230 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1231 { 1232 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1233 } 1234 JRT_END 1235 1236 1237 #if defined(COMPILER2) || defined(SHARK) 1238 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1239 // in case of an unresolved klass entry, load the class. 1240 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1241 Klass* tk = constant_pool->klass_at(index, CHECK); 1242 return; 1243 } 1244 1245 if (!constant_pool->tag_at(index).is_symbol()) return; 1246 1247 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1248 Symbol* symbol = constant_pool->symbol_at(index); 1249 1250 // class name? 1251 if (symbol->byte_at(0) != '(') { 1252 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1253 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1254 return; 1255 } 1256 1257 // then it must be a signature! 1258 ResourceMark rm(THREAD); 1259 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1260 if (ss.is_object()) { 1261 Symbol* class_name = ss.as_symbol(CHECK); 1262 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1263 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1264 } 1265 } 1266 } 1267 1268 1269 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1270 EXCEPTION_MARK; 1271 load_class_by_index(constant_pool, index, THREAD); 1272 if (HAS_PENDING_EXCEPTION) { 1273 // Exception happened during classloading. We ignore the exception here, since it 1274 // is going to be rethrown since the current activation is going to be deoptimized and 1275 // the interpreter will re-execute the bytecode. 1276 CLEAR_PENDING_EXCEPTION; 1277 // Class loading called java code which may have caused a stack 1278 // overflow. If the exception was thrown right before the return 1279 // to the runtime the stack is no longer guarded. Reguard the 1280 // stack otherwise if we return to the uncommon trap blob and the 1281 // stack bang causes a stack overflow we crash. 1282 assert(THREAD->is_Java_thread(), "only a java thread can be here"); 1283 JavaThread* thread = (JavaThread*)THREAD; 1284 bool guard_pages_enabled = thread->stack_yellow_zone_enabled(); 1285 if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack(); 1286 assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash"); 1287 } 1288 } 1289 1290 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1291 HandleMark hm; 1292 1293 // uncommon_trap() is called at the beginning of the uncommon trap 1294 // handler. Note this fact before we start generating temporary frames 1295 // that can confuse an asynchronous stack walker. This counter is 1296 // decremented at the end of unpack_frames(). 1297 thread->inc_in_deopt_handler(); 1298 1299 // We need to update the map if we have biased locking. 1300 RegisterMap reg_map(thread, UseBiasedLocking); 1301 frame stub_frame = thread->last_frame(); 1302 frame fr = stub_frame.sender(®_map); 1303 // Make sure the calling nmethod is not getting deoptimized and removed 1304 // before we are done with it. 1305 nmethodLocker nl(fr.pc()); 1306 1307 // Log a message 1308 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT, 1309 trap_request, fr.pc()); 1310 1311 { 1312 ResourceMark rm; 1313 1314 // Revoke biases of any monitors in the frame to ensure we can migrate them 1315 revoke_biases_of_monitors(thread, fr, ®_map); 1316 1317 DeoptReason reason = trap_request_reason(trap_request); 1318 DeoptAction action = trap_request_action(trap_request); 1319 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1320 1321 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1322 compiledVFrame* cvf = compiledVFrame::cast(vf); 1323 1324 nmethod* nm = cvf->code(); 1325 1326 ScopeDesc* trap_scope = cvf->scope(); 1327 methodHandle trap_method = trap_scope->method(); 1328 int trap_bci = trap_scope->bci(); 1329 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1330 1331 // Record this event in the histogram. 1332 gather_statistics(reason, action, trap_bc); 1333 1334 // Ensure that we can record deopt. history: 1335 // Need MDO to record RTM code generation state. 1336 bool create_if_missing = ProfileTraps RTM_OPT_ONLY( || UseRTMLocking ); 1337 1338 MethodData* trap_mdo = 1339 get_method_data(thread, trap_method, create_if_missing); 1340 1341 // Log a message 1342 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1343 trap_reason_name(reason), trap_action_name(action), fr.pc(), 1344 trap_method->name_and_sig_as_C_string(), trap_bci); 1345 1346 // Print a bunch of diagnostics, if requested. 1347 if (TraceDeoptimization || LogCompilation) { 1348 ResourceMark rm; 1349 ttyLocker ttyl; 1350 char buf[100]; 1351 if (xtty != NULL) { 1352 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1353 os::current_thread_id(), 1354 format_trap_request(buf, sizeof(buf), trap_request)); 1355 nm->log_identity(xtty); 1356 } 1357 Symbol* class_name = NULL; 1358 bool unresolved = false; 1359 if (unloaded_class_index >= 0) { 1360 constantPoolHandle constants (THREAD, trap_method->constants()); 1361 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1362 class_name = constants->klass_name_at(unloaded_class_index); 1363 unresolved = true; 1364 if (xtty != NULL) 1365 xtty->print(" unresolved='1'"); 1366 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1367 class_name = constants->symbol_at(unloaded_class_index); 1368 } 1369 if (xtty != NULL) 1370 xtty->name(class_name); 1371 } 1372 if (xtty != NULL && trap_mdo != NULL) { 1373 // Dump the relevant MDO state. 1374 // This is the deopt count for the current reason, any previous 1375 // reasons or recompiles seen at this point. 1376 int dcnt = trap_mdo->trap_count(reason); 1377 if (dcnt != 0) 1378 xtty->print(" count='%d'", dcnt); 1379 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1380 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1381 if (dos != 0) { 1382 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1383 if (trap_state_is_recompiled(dos)) { 1384 int recnt2 = trap_mdo->overflow_recompile_count(); 1385 if (recnt2 != 0) 1386 xtty->print(" recompiles2='%d'", recnt2); 1387 } 1388 } 1389 } 1390 if (xtty != NULL) { 1391 xtty->stamp(); 1392 xtty->end_head(); 1393 } 1394 if (TraceDeoptimization) { // make noise on the tty 1395 tty->print("Uncommon trap occurred in"); 1396 nm->method()->print_short_name(tty); 1397 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d", 1398 fr.pc(), 1399 os::current_thread_id(), 1400 trap_reason_name(reason), 1401 trap_action_name(action), 1402 unloaded_class_index); 1403 if (class_name != NULL) { 1404 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1405 class_name->print_symbol_on(tty); 1406 } 1407 tty->cr(); 1408 } 1409 if (xtty != NULL) { 1410 // Log the precise location of the trap. 1411 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1412 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1413 xtty->method(sd->method()); 1414 xtty->end_elem(); 1415 if (sd->is_top()) break; 1416 } 1417 xtty->tail("uncommon_trap"); 1418 } 1419 } 1420 // (End diagnostic printout.) 1421 1422 // Load class if necessary 1423 if (unloaded_class_index >= 0) { 1424 constantPoolHandle constants(THREAD, trap_method->constants()); 1425 load_class_by_index(constants, unloaded_class_index); 1426 } 1427 1428 // Flush the nmethod if necessary and desirable. 1429 // 1430 // We need to avoid situations where we are re-flushing the nmethod 1431 // because of a hot deoptimization site. Repeated flushes at the same 1432 // point need to be detected by the compiler and avoided. If the compiler 1433 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1434 // module must take measures to avoid an infinite cycle of recompilation 1435 // and deoptimization. There are several such measures: 1436 // 1437 // 1. If a recompilation is ordered a second time at some site X 1438 // and for the same reason R, the action is adjusted to 'reinterpret', 1439 // to give the interpreter time to exercise the method more thoroughly. 1440 // If this happens, the method's overflow_recompile_count is incremented. 1441 // 1442 // 2. If the compiler fails to reduce the deoptimization rate, then 1443 // the method's overflow_recompile_count will begin to exceed the set 1444 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1445 // is adjusted to 'make_not_compilable', and the method is abandoned 1446 // to the interpreter. This is a performance hit for hot methods, 1447 // but is better than a disastrous infinite cycle of recompilations. 1448 // (Actually, only the method containing the site X is abandoned.) 1449 // 1450 // 3. In parallel with the previous measures, if the total number of 1451 // recompilations of a method exceeds the much larger set limit 1452 // PerMethodRecompilationCutoff, the method is abandoned. 1453 // This should only happen if the method is very large and has 1454 // many "lukewarm" deoptimizations. The code which enforces this 1455 // limit is elsewhere (class nmethod, class Method). 1456 // 1457 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1458 // to recompile at each bytecode independently of the per-BCI cutoff. 1459 // 1460 // The decision to update code is up to the compiler, and is encoded 1461 // in the Action_xxx code. If the compiler requests Action_none 1462 // no trap state is changed, no compiled code is changed, and the 1463 // computation suffers along in the interpreter. 1464 // 1465 // The other action codes specify various tactics for decompilation 1466 // and recompilation. Action_maybe_recompile is the loosest, and 1467 // allows the compiled code to stay around until enough traps are seen, 1468 // and until the compiler gets around to recompiling the trapping method. 1469 // 1470 // The other actions cause immediate removal of the present code. 1471 1472 bool update_trap_state = true; 1473 bool make_not_entrant = false; 1474 bool make_not_compilable = false; 1475 bool reprofile = false; 1476 switch (action) { 1477 case Action_none: 1478 // Keep the old code. 1479 update_trap_state = false; 1480 break; 1481 case Action_maybe_recompile: 1482 // Do not need to invalidate the present code, but we can 1483 // initiate another 1484 // Start compiler without (necessarily) invalidating the nmethod. 1485 // The system will tolerate the old code, but new code should be 1486 // generated when possible. 1487 break; 1488 case Action_reinterpret: 1489 // Go back into the interpreter for a while, and then consider 1490 // recompiling form scratch. 1491 make_not_entrant = true; 1492 // Reset invocation counter for outer most method. 1493 // This will allow the interpreter to exercise the bytecodes 1494 // for a while before recompiling. 1495 // By contrast, Action_make_not_entrant is immediate. 1496 // 1497 // Note that the compiler will track null_check, null_assert, 1498 // range_check, and class_check events and log them as if they 1499 // had been traps taken from compiled code. This will update 1500 // the MDO trap history so that the next compilation will 1501 // properly detect hot trap sites. 1502 reprofile = true; 1503 break; 1504 case Action_make_not_entrant: 1505 // Request immediate recompilation, and get rid of the old code. 1506 // Make them not entrant, so next time they are called they get 1507 // recompiled. Unloaded classes are loaded now so recompile before next 1508 // time they are called. Same for uninitialized. The interpreter will 1509 // link the missing class, if any. 1510 make_not_entrant = true; 1511 break; 1512 case Action_make_not_compilable: 1513 // Give up on compiling this method at all. 1514 make_not_entrant = true; 1515 make_not_compilable = true; 1516 break; 1517 default: 1518 ShouldNotReachHere(); 1519 } 1520 1521 // Setting +ProfileTraps fixes the following, on all platforms: 1522 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1523 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1524 // recompile relies on a MethodData* to record heroic opt failures. 1525 1526 // Whether the interpreter is producing MDO data or not, we also need 1527 // to use the MDO to detect hot deoptimization points and control 1528 // aggressive optimization. 1529 bool inc_recompile_count = false; 1530 ProfileData* pdata = NULL; 1531 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1532 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity"); 1533 uint this_trap_count = 0; 1534 bool maybe_prior_trap = false; 1535 bool maybe_prior_recompile = false; 1536 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1537 nm->method(), 1538 //outputs: 1539 this_trap_count, 1540 maybe_prior_trap, 1541 maybe_prior_recompile); 1542 // Because the interpreter also counts null, div0, range, and class 1543 // checks, these traps from compiled code are double-counted. 1544 // This is harmless; it just means that the PerXTrapLimit values 1545 // are in effect a little smaller than they look. 1546 1547 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1548 if (per_bc_reason != Reason_none) { 1549 // Now take action based on the partially known per-BCI history. 1550 if (maybe_prior_trap 1551 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1552 // If there are too many traps at this BCI, force a recompile. 1553 // This will allow the compiler to see the limit overflow, and 1554 // take corrective action, if possible. The compiler generally 1555 // does not use the exact PerBytecodeTrapLimit value, but instead 1556 // changes its tactics if it sees any traps at all. This provides 1557 // a little hysteresis, delaying a recompile until a trap happens 1558 // several times. 1559 // 1560 // Actually, since there is only one bit of counter per BCI, 1561 // the possible per-BCI counts are {0,1,(per-method count)}. 1562 // This produces accurate results if in fact there is only 1563 // one hot trap site, but begins to get fuzzy if there are 1564 // many sites. For example, if there are ten sites each 1565 // trapping two or more times, they each get the blame for 1566 // all of their traps. 1567 make_not_entrant = true; 1568 } 1569 1570 // Detect repeated recompilation at the same BCI, and enforce a limit. 1571 if (make_not_entrant && maybe_prior_recompile) { 1572 // More than one recompile at this point. 1573 inc_recompile_count = maybe_prior_trap; 1574 } 1575 } else { 1576 // For reasons which are not recorded per-bytecode, we simply 1577 // force recompiles unconditionally. 1578 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1579 make_not_entrant = true; 1580 } 1581 1582 // Go back to the compiler if there are too many traps in this method. 1583 if (this_trap_count >= per_method_trap_limit(reason)) { 1584 // If there are too many traps in this method, force a recompile. 1585 // This will allow the compiler to see the limit overflow, and 1586 // take corrective action, if possible. 1587 // (This condition is an unlikely backstop only, because the 1588 // PerBytecodeTrapLimit is more likely to take effect first, 1589 // if it is applicable.) 1590 make_not_entrant = true; 1591 } 1592 1593 // Here's more hysteresis: If there has been a recompile at 1594 // this trap point already, run the method in the interpreter 1595 // for a while to exercise it more thoroughly. 1596 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1597 reprofile = true; 1598 } 1599 1600 } 1601 1602 // Take requested actions on the method: 1603 1604 // Recompile 1605 if (make_not_entrant) { 1606 if (!nm->make_not_entrant()) { 1607 return; // the call did not change nmethod's state 1608 } 1609 1610 if (pdata != NULL) { 1611 // Record the recompilation event, if any. 1612 int tstate0 = pdata->trap_state(); 1613 int tstate1 = trap_state_set_recompiled(tstate0, true); 1614 if (tstate1 != tstate0) 1615 pdata->set_trap_state(tstate1); 1616 } 1617 1618 #if INCLUDE_RTM_OPT 1619 // Restart collecting RTM locking abort statistic if the method 1620 // is recompiled for a reason other than RTM state change. 1621 // Assume that in new recompiled code the statistic could be different, 1622 // for example, due to different inlining. 1623 if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) && 1624 UseRTMDeopt && (nm->rtm_state() != ProfileRTM)) { 1625 trap_mdo->atomic_set_rtm_state(ProfileRTM); 1626 } 1627 #endif 1628 } 1629 1630 if (inc_recompile_count) { 1631 trap_mdo->inc_overflow_recompile_count(); 1632 if ((uint)trap_mdo->overflow_recompile_count() > 1633 (uint)PerBytecodeRecompilationCutoff) { 1634 // Give up on the method containing the bad BCI. 1635 if (trap_method() == nm->method()) { 1636 make_not_compilable = true; 1637 } else { 1638 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1639 // But give grace to the enclosing nm->method(). 1640 } 1641 } 1642 } 1643 1644 // Reprofile 1645 if (reprofile) { 1646 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1647 } 1648 1649 // Give up compiling 1650 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1651 assert(make_not_entrant, "consistent"); 1652 nm->method()->set_not_compilable(CompLevel_full_optimization); 1653 } 1654 1655 } // Free marked resources 1656 1657 } 1658 JRT_END 1659 1660 MethodData* 1661 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1662 bool create_if_missing) { 1663 Thread* THREAD = thread; 1664 MethodData* mdo = m()->method_data(); 1665 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1666 // Build an MDO. Ignore errors like OutOfMemory; 1667 // that simply means we won't have an MDO to update. 1668 Method::build_interpreter_method_data(m, THREAD); 1669 if (HAS_PENDING_EXCEPTION) { 1670 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1671 CLEAR_PENDING_EXCEPTION; 1672 } 1673 mdo = m()->method_data(); 1674 } 1675 return mdo; 1676 } 1677 1678 ProfileData* 1679 Deoptimization::query_update_method_data(MethodData* trap_mdo, 1680 int trap_bci, 1681 Deoptimization::DeoptReason reason, 1682 Method* compiled_method, 1683 //outputs: 1684 uint& ret_this_trap_count, 1685 bool& ret_maybe_prior_trap, 1686 bool& ret_maybe_prior_recompile) { 1687 uint prior_trap_count = trap_mdo->trap_count(reason); 1688 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1689 1690 // If the runtime cannot find a place to store trap history, 1691 // it is estimated based on the general condition of the method. 1692 // If the method has ever been recompiled, or has ever incurred 1693 // a trap with the present reason , then this BCI is assumed 1694 // (pessimistically) to be the culprit. 1695 bool maybe_prior_trap = (prior_trap_count != 0); 1696 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1697 ProfileData* pdata = NULL; 1698 1699 1700 // For reasons which are recorded per bytecode, we check per-BCI data. 1701 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1702 if (per_bc_reason != Reason_none) { 1703 // Find the profile data for this BCI. If there isn't one, 1704 // try to allocate one from the MDO's set of spares. 1705 // This will let us detect a repeated trap at this point. 1706 pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL); 1707 1708 if (pdata != NULL) { 1709 if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) { 1710 if (LogCompilation && xtty != NULL) { 1711 ttyLocker ttyl; 1712 // no more room for speculative traps in this MDO 1713 xtty->elem("speculative_traps_oom"); 1714 } 1715 } 1716 // Query the trap state of this profile datum. 1717 int tstate0 = pdata->trap_state(); 1718 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1719 maybe_prior_trap = false; 1720 if (!trap_state_is_recompiled(tstate0)) 1721 maybe_prior_recompile = false; 1722 1723 // Update the trap state of this profile datum. 1724 int tstate1 = tstate0; 1725 // Record the reason. 1726 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1727 // Store the updated state on the MDO, for next time. 1728 if (tstate1 != tstate0) 1729 pdata->set_trap_state(tstate1); 1730 } else { 1731 if (LogCompilation && xtty != NULL) { 1732 ttyLocker ttyl; 1733 // Missing MDP? Leave a small complaint in the log. 1734 xtty->elem("missing_mdp bci='%d'", trap_bci); 1735 } 1736 } 1737 } 1738 1739 // Return results: 1740 ret_this_trap_count = this_trap_count; 1741 ret_maybe_prior_trap = maybe_prior_trap; 1742 ret_maybe_prior_recompile = maybe_prior_recompile; 1743 return pdata; 1744 } 1745 1746 void 1747 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1748 ResourceMark rm; 1749 // Ignored outputs: 1750 uint ignore_this_trap_count; 1751 bool ignore_maybe_prior_trap; 1752 bool ignore_maybe_prior_recompile; 1753 assert(!reason_is_speculate(reason), "reason speculate only used by compiler"); 1754 query_update_method_data(trap_mdo, trap_bci, 1755 (DeoptReason)reason, 1756 NULL, 1757 ignore_this_trap_count, 1758 ignore_maybe_prior_trap, 1759 ignore_maybe_prior_recompile); 1760 } 1761 1762 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1763 1764 // Still in Java no safepoints 1765 { 1766 // This enters VM and may safepoint 1767 uncommon_trap_inner(thread, trap_request); 1768 } 1769 return fetch_unroll_info_helper(thread); 1770 } 1771 1772 // Local derived constants. 1773 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1774 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1775 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1776 1777 //---------------------------trap_state_reason--------------------------------- 1778 Deoptimization::DeoptReason 1779 Deoptimization::trap_state_reason(int trap_state) { 1780 // This assert provides the link between the width of DataLayout::trap_bits 1781 // and the encoding of "recorded" reasons. It ensures there are enough 1782 // bits to store all needed reasons in the per-BCI MDO profile. 1783 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1784 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1785 trap_state -= recompile_bit; 1786 if (trap_state == DS_REASON_MASK) { 1787 return Reason_many; 1788 } else { 1789 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1790 return (DeoptReason)trap_state; 1791 } 1792 } 1793 //-------------------------trap_state_has_reason------------------------------- 1794 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1795 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1796 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1797 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1798 trap_state -= recompile_bit; 1799 if (trap_state == DS_REASON_MASK) { 1800 return -1; // true, unspecifically (bottom of state lattice) 1801 } else if (trap_state == reason) { 1802 return 1; // true, definitely 1803 } else if (trap_state == 0) { 1804 return 0; // false, definitely (top of state lattice) 1805 } else { 1806 return 0; // false, definitely 1807 } 1808 } 1809 //-------------------------trap_state_add_reason------------------------------- 1810 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1811 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1812 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1813 trap_state -= recompile_bit; 1814 if (trap_state == DS_REASON_MASK) { 1815 return trap_state + recompile_bit; // already at state lattice bottom 1816 } else if (trap_state == reason) { 1817 return trap_state + recompile_bit; // the condition is already true 1818 } else if (trap_state == 0) { 1819 return reason + recompile_bit; // no condition has yet been true 1820 } else { 1821 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1822 } 1823 } 1824 //-----------------------trap_state_is_recompiled------------------------------ 1825 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1826 return (trap_state & DS_RECOMPILE_BIT) != 0; 1827 } 1828 //-----------------------trap_state_set_recompiled----------------------------- 1829 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1830 if (z) return trap_state | DS_RECOMPILE_BIT; 1831 else return trap_state & ~DS_RECOMPILE_BIT; 1832 } 1833 //---------------------------format_trap_state--------------------------------- 1834 // This is used for debugging and diagnostics, including LogFile output. 1835 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1836 int trap_state) { 1837 DeoptReason reason = trap_state_reason(trap_state); 1838 bool recomp_flag = trap_state_is_recompiled(trap_state); 1839 // Re-encode the state from its decoded components. 1840 int decoded_state = 0; 1841 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1842 decoded_state = trap_state_add_reason(decoded_state, reason); 1843 if (recomp_flag) 1844 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1845 // If the state re-encodes properly, format it symbolically. 1846 // Because this routine is used for debugging and diagnostics, 1847 // be robust even if the state is a strange value. 1848 size_t len; 1849 if (decoded_state != trap_state) { 1850 // Random buggy state that doesn't decode?? 1851 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1852 } else { 1853 len = jio_snprintf(buf, buflen, "%s%s", 1854 trap_reason_name(reason), 1855 recomp_flag ? " recompiled" : ""); 1856 } 1857 if (len >= buflen) 1858 buf[buflen-1] = '\0'; 1859 return buf; 1860 } 1861 1862 1863 //--------------------------------statics-------------------------------------- 1864 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1865 = Deoptimization::Action_reinterpret; 1866 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1867 // Note: Keep this in sync. with enum DeoptReason. 1868 "none", 1869 "null_check", 1870 "null_assert", 1871 "range_check", 1872 "class_check", 1873 "array_check", 1874 "intrinsic", 1875 "bimorphic", 1876 "unloaded", 1877 "uninitialized", 1878 "unreached", 1879 "unhandled", 1880 "constraint", 1881 "div0_check", 1882 "age", 1883 "predicate", 1884 "loop_limit_check", 1885 "speculate_class_check", 1886 "rtm_state_change", 1887 "unstable_if" 1888 }; 1889 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1890 // Note: Keep this in sync. with enum DeoptAction. 1891 "none", 1892 "maybe_recompile", 1893 "reinterpret", 1894 "make_not_entrant", 1895 "make_not_compilable" 1896 }; 1897 1898 const char* Deoptimization::trap_reason_name(int reason) { 1899 if (reason == Reason_many) return "many"; 1900 if ((uint)reason < Reason_LIMIT) 1901 return _trap_reason_name[reason]; 1902 static char buf[20]; 1903 sprintf(buf, "reason%d", reason); 1904 return buf; 1905 } 1906 const char* Deoptimization::trap_action_name(int action) { 1907 if ((uint)action < Action_LIMIT) 1908 return _trap_action_name[action]; 1909 static char buf[20]; 1910 sprintf(buf, "action%d", action); 1911 return buf; 1912 } 1913 1914 // This is used for debugging and diagnostics, including LogFile output. 1915 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1916 int trap_request) { 1917 jint unloaded_class_index = trap_request_index(trap_request); 1918 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1919 const char* action = trap_action_name(trap_request_action(trap_request)); 1920 size_t len; 1921 if (unloaded_class_index < 0) { 1922 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1923 reason, action); 1924 } else { 1925 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1926 reason, action, unloaded_class_index); 1927 } 1928 if (len >= buflen) 1929 buf[buflen-1] = '\0'; 1930 return buf; 1931 } 1932 1933 juint Deoptimization::_deoptimization_hist 1934 [Deoptimization::Reason_LIMIT] 1935 [1 + Deoptimization::Action_LIMIT] 1936 [Deoptimization::BC_CASE_LIMIT] 1937 = {0}; 1938 1939 enum { 1940 LSB_BITS = 8, 1941 LSB_MASK = right_n_bits(LSB_BITS) 1942 }; 1943 1944 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1945 Bytecodes::Code bc) { 1946 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1947 assert(action >= 0 && action < Action_LIMIT, "oob"); 1948 _deoptimization_hist[Reason_none][0][0] += 1; // total 1949 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1950 juint* cases = _deoptimization_hist[reason][1+action]; 1951 juint* bc_counter_addr = NULL; 1952 juint bc_counter = 0; 1953 // Look for an unused counter, or an exact match to this BC. 1954 if (bc != Bytecodes::_illegal) { 1955 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1956 juint* counter_addr = &cases[bc_case]; 1957 juint counter = *counter_addr; 1958 if ((counter == 0 && bc_counter_addr == NULL) 1959 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1960 // this counter is either free or is already devoted to this BC 1961 bc_counter_addr = counter_addr; 1962 bc_counter = counter | bc; 1963 } 1964 } 1965 } 1966 if (bc_counter_addr == NULL) { 1967 // Overflow, or no given bytecode. 1968 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1969 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1970 } 1971 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1972 } 1973 1974 jint Deoptimization::total_deoptimization_count() { 1975 return _deoptimization_hist[Reason_none][0][0]; 1976 } 1977 1978 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1979 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1980 return _deoptimization_hist[reason][0][0]; 1981 } 1982 1983 void Deoptimization::print_statistics() { 1984 juint total = total_deoptimization_count(); 1985 juint account = total; 1986 if (total != 0) { 1987 ttyLocker ttyl; 1988 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1989 tty->print_cr("Deoptimization traps recorded:"); 1990 #define PRINT_STAT_LINE(name, r) \ 1991 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1992 PRINT_STAT_LINE("total", total); 1993 // For each non-zero entry in the histogram, print the reason, 1994 // the action, and (if specifically known) the type of bytecode. 1995 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1996 for (int action = 0; action < Action_LIMIT; action++) { 1997 juint* cases = _deoptimization_hist[reason][1+action]; 1998 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1999 juint counter = cases[bc_case]; 2000 if (counter != 0) { 2001 char name[1*K]; 2002 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 2003 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 2004 bc = Bytecodes::_illegal; 2005 sprintf(name, "%s/%s/%s", 2006 trap_reason_name(reason), 2007 trap_action_name(action), 2008 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 2009 juint r = counter >> LSB_BITS; 2010 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 2011 account -= r; 2012 } 2013 } 2014 } 2015 } 2016 if (account != 0) { 2017 PRINT_STAT_LINE("unaccounted", account); 2018 } 2019 #undef PRINT_STAT_LINE 2020 if (xtty != NULL) xtty->tail("statistics"); 2021 } 2022 } 2023 #else // COMPILER2 || SHARK 2024 2025 2026 // Stubs for C1 only system. 2027 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 2028 return false; 2029 } 2030 2031 const char* Deoptimization::trap_reason_name(int reason) { 2032 return "unknown"; 2033 } 2034 2035 void Deoptimization::print_statistics() { 2036 // no output 2037 } 2038 2039 void 2040 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 2041 // no udpate 2042 } 2043 2044 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 2045 return 0; 2046 } 2047 2048 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 2049 Bytecodes::Code bc) { 2050 // no update 2051 } 2052 2053 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 2054 int trap_state) { 2055 jio_snprintf(buf, buflen, "#%d", trap_state); 2056 return buf; 2057 } 2058 2059 #endif // COMPILER2 || SHARK