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