1 /* 2 * Copyright (c) 1997, 2012, 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, 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", 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 popframe_extra_args); 429 // This pc doesn't have to be perfect just good enough to identify the frame 430 // as interpreted so the skeleton frame will be walkable 431 // The correct pc will be set when the skeleton frame is completely filled out 432 // The final pc we store in the loop is wrong and will be overwritten below 433 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 434 435 callee_parameters = array->element(index)->method()->size_of_parameters(); 436 callee_locals = array->element(index)->method()->max_locals(); 437 popframe_extra_args = 0; 438 } 439 440 // Compute whether the root vframe returns a float or double value. 441 BasicType return_type; 442 { 443 HandleMark hm; 444 methodHandle method(thread, array->element(0)->method()); 445 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 446 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 447 } 448 449 // Compute information for handling adapters and adjusting the frame size of the caller. 450 int caller_adjustment = 0; 451 452 // Compute the amount the oldest interpreter frame will have to adjust 453 // its caller's stack by. If the caller is a compiled frame then 454 // we pretend that the callee has no parameters so that the 455 // extension counts for the full amount of locals and not just 456 // locals-parms. This is because without a c2i adapter the parm 457 // area as created by the compiled frame will not be usable by 458 // the interpreter. (Depending on the calling convention there 459 // may not even be enough space). 460 461 // QQQ I'd rather see this pushed down into last_frame_adjust 462 // and have it take the sender (aka caller). 463 464 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 465 caller_adjustment = last_frame_adjust(0, callee_locals); 466 } else if (callee_locals > callee_parameters) { 467 // The caller frame may need extending to accommodate 468 // non-parameter locals of the first unpacked interpreted frame. 469 // Compute that adjustment. 470 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 471 } 472 473 // If the sender is deoptimized the we must retrieve the address of the handler 474 // since the frame will "magically" show the original pc before the deopt 475 // and we'd undo the deopt. 476 477 frame_pcs[0] = deopt_sender.raw_pc(); 478 479 #ifndef SHARK 480 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 481 #endif // SHARK 482 483 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 484 caller_adjustment * BytesPerWord, 485 caller_was_method_handle ? 0 : callee_parameters, 486 number_of_frames, 487 frame_sizes, 488 frame_pcs, 489 return_type); 490 // On some platforms, we need a way to pass some platform dependent 491 // information to the unpacking code so the skeletal frames come out 492 // correct (initial fp value, unextended sp, ...) 493 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 494 495 if (array->frames() > 1) { 496 if (VerifyStack && TraceDeoptimization) { 497 ttyLocker ttyl; 498 tty->print_cr("Deoptimizing method containing inlining"); 499 } 500 } 501 502 array->set_unroll_block(info); 503 return info; 504 } 505 506 // Called to cleanup deoptimization data structures in normal case 507 // after unpacking to stack and when stack overflow error occurs 508 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 509 vframeArray *array) { 510 511 // Get array if coming from exception 512 if (array == NULL) { 513 array = thread->vframe_array_head(); 514 } 515 thread->set_vframe_array_head(NULL); 516 517 // Free the previous UnrollBlock 518 vframeArray* old_array = thread->vframe_array_last(); 519 thread->set_vframe_array_last(array); 520 521 if (old_array != NULL) { 522 UnrollBlock* old_info = old_array->unroll_block(); 523 old_array->set_unroll_block(NULL); 524 delete old_info; 525 delete old_array; 526 } 527 528 // Deallocate any resource creating in this routine and any ResourceObjs allocated 529 // inside the vframeArray (StackValueCollections) 530 531 delete thread->deopt_mark(); 532 thread->set_deopt_mark(NULL); 533 thread->set_deopt_nmethod(NULL); 534 535 536 if (JvmtiExport::can_pop_frame()) { 537 #ifndef CC_INTERP 538 // Regardless of whether we entered this routine with the pending 539 // popframe condition bit set, we should always clear it now 540 thread->clear_popframe_condition(); 541 #else 542 // C++ interpeter will clear has_pending_popframe when it enters 543 // with method_resume. For deopt_resume2 we clear it now. 544 if (thread->popframe_forcing_deopt_reexecution()) 545 thread->clear_popframe_condition(); 546 #endif /* CC_INTERP */ 547 } 548 549 // unpack_frames() is called at the end of the deoptimization handler 550 // and (in C2) at the end of the uncommon trap handler. Note this fact 551 // so that an asynchronous stack walker can work again. This counter is 552 // incremented at the beginning of fetch_unroll_info() and (in C2) at 553 // the beginning of uncommon_trap(). 554 thread->dec_in_deopt_handler(); 555 } 556 557 558 // Return BasicType of value being returned 559 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 560 561 // We are already active int he special DeoptResourceMark any ResourceObj's we 562 // allocate will be freed at the end of the routine. 563 564 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 565 // but makes the entry a little slower. There is however a little dance we have to 566 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 567 ResetNoHandleMark rnhm; // No-op in release/product versions 568 HandleMark hm; 569 570 frame stub_frame = thread->last_frame(); 571 572 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 573 // must point to the vframeArray for the unpack frame. 574 vframeArray* array = thread->vframe_array_head(); 575 576 #ifndef PRODUCT 577 if (TraceDeoptimization) { 578 ttyLocker ttyl; 579 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 580 } 581 #endif 582 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 583 stub_frame.pc(), stub_frame.sp(), exec_mode); 584 585 UnrollBlock* info = array->unroll_block(); 586 587 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 588 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 589 590 BasicType bt = info->return_type(); 591 592 // If we have an exception pending, claim that the return type is an oop 593 // so the deopt_blob does not overwrite the exception_oop. 594 595 if (exec_mode == Unpack_exception) 596 bt = T_OBJECT; 597 598 // Cleanup thread deopt data 599 cleanup_deopt_info(thread, array); 600 601 #ifndef PRODUCT 602 if (VerifyStack) { 603 ResourceMark res_mark; 604 605 thread->validate_frame_layout(); 606 607 // Verify that the just-unpacked frames match the interpreter's 608 // notions of expression stack and locals 609 vframeArray* cur_array = thread->vframe_array_last(); 610 RegisterMap rm(thread, false); 611 rm.set_include_argument_oops(false); 612 bool is_top_frame = true; 613 int callee_size_of_parameters = 0; 614 int callee_max_locals = 0; 615 for (int i = 0; i < cur_array->frames(); i++) { 616 vframeArrayElement* el = cur_array->element(i); 617 frame* iframe = el->iframe(); 618 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 619 620 // Get the oop map for this bci 621 InterpreterOopMap mask; 622 int cur_invoke_parameter_size = 0; 623 bool try_next_mask = false; 624 int next_mask_expression_stack_size = -1; 625 int top_frame_expression_stack_adjustment = 0; 626 methodHandle mh(thread, iframe->interpreter_frame_method()); 627 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 628 BytecodeStream str(mh); 629 str.set_start(iframe->interpreter_frame_bci()); 630 int max_bci = mh->code_size(); 631 // Get to the next bytecode if possible 632 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 633 // Check to see if we can grab the number of outgoing arguments 634 // at an uncommon trap for an invoke (where the compiler 635 // generates debug info before the invoke has executed) 636 Bytecodes::Code cur_code = str.next(); 637 if (cur_code == Bytecodes::_invokevirtual || 638 cur_code == Bytecodes::_invokespecial || 639 cur_code == Bytecodes::_invokestatic || 640 cur_code == Bytecodes::_invokeinterface) { 641 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 642 Symbol* signature = invoke.signature(); 643 ArgumentSizeComputer asc(signature); 644 cur_invoke_parameter_size = asc.size(); 645 if (cur_code != Bytecodes::_invokestatic) { 646 // Add in receiver 647 ++cur_invoke_parameter_size; 648 } 649 } 650 if (str.bci() < max_bci) { 651 Bytecodes::Code bc = str.next(); 652 if (bc >= 0) { 653 // The interpreter oop map generator reports results before 654 // the current bytecode has executed except in the case of 655 // calls. It seems to be hard to tell whether the compiler 656 // has emitted debug information matching the "state before" 657 // a given bytecode or the state after, so we try both 658 switch (cur_code) { 659 case Bytecodes::_invokevirtual: 660 case Bytecodes::_invokespecial: 661 case Bytecodes::_invokestatic: 662 case Bytecodes::_invokeinterface: 663 case Bytecodes::_athrow: 664 break; 665 default: { 666 InterpreterOopMap next_mask; 667 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 668 next_mask_expression_stack_size = next_mask.expression_stack_size(); 669 // Need to subtract off the size of the result type of 670 // the bytecode because this is not described in the 671 // debug info but returned to the interpreter in the TOS 672 // caching register 673 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 674 if (bytecode_result_type != T_ILLEGAL) { 675 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 676 } 677 assert(top_frame_expression_stack_adjustment >= 0, ""); 678 try_next_mask = true; 679 break; 680 } 681 } 682 } 683 } 684 685 // Verify stack depth and oops in frame 686 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 687 if (!( 688 /* SPARC */ 689 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 690 /* x86 */ 691 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 692 (try_next_mask && 693 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 694 top_frame_expression_stack_adjustment))) || 695 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 696 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 697 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 698 )) { 699 ttyLocker ttyl; 700 701 // Print out some information that will help us debug the problem 702 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 703 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 704 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 705 iframe->interpreter_frame_expression_stack_size()); 706 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 707 tty->print_cr(" try_next_mask = %d", try_next_mask); 708 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 709 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 710 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 711 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 712 tty->print_cr(" exec_mode = %d", exec_mode); 713 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 714 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 715 tty->print_cr(" Interpreted frames:"); 716 for (int k = 0; k < cur_array->frames(); k++) { 717 vframeArrayElement* el = cur_array->element(k); 718 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 719 } 720 cur_array->print_on_2(tty); 721 guarantee(false, "wrong number of expression stack elements during deopt"); 722 } 723 VerifyOopClosure verify; 724 iframe->oops_interpreted_do(&verify, &rm, false); 725 callee_size_of_parameters = mh->size_of_parameters(); 726 callee_max_locals = mh->max_locals(); 727 is_top_frame = false; 728 } 729 } 730 #endif /* !PRODUCT */ 731 732 733 return bt; 734 JRT_END 735 736 737 int Deoptimization::deoptimize_dependents() { 738 Threads::deoptimized_wrt_marked_nmethods(); 739 return 0; 740 } 741 742 743 #ifdef COMPILER2 744 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 745 Handle pending_exception(thread->pending_exception()); 746 const char* exception_file = thread->exception_file(); 747 int exception_line = thread->exception_line(); 748 thread->clear_pending_exception(); 749 750 for (int i = 0; i < objects->length(); i++) { 751 assert(objects->at(i)->is_object(), "invalid debug information"); 752 ObjectValue* sv = (ObjectValue*) objects->at(i); 753 754 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 755 oop obj = NULL; 756 757 if (k->oop_is_instance()) { 758 InstanceKlass* ik = InstanceKlass::cast(k()); 759 obj = ik->allocate_instance(CHECK_(false)); 760 } else if (k->oop_is_typeArray()) { 761 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 762 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 763 int len = sv->field_size() / type2size[ak->element_type()]; 764 obj = ak->allocate(len, CHECK_(false)); 765 } else if (k->oop_is_objArray()) { 766 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 767 obj = ak->allocate(sv->field_size(), CHECK_(false)); 768 } 769 770 assert(obj != NULL, "allocation failed"); 771 assert(sv->value().is_null(), "redundant reallocation"); 772 sv->set_value(obj); 773 } 774 775 if (pending_exception.not_null()) { 776 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 777 } 778 779 return true; 780 } 781 782 // This assumes that the fields are stored in ObjectValue in the same order 783 // they are yielded by do_nonstatic_fields. 784 class FieldReassigner: public FieldClosure { 785 frame* _fr; 786 RegisterMap* _reg_map; 787 ObjectValue* _sv; 788 InstanceKlass* _ik; 789 oop _obj; 790 791 int _i; 792 public: 793 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 794 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 795 796 int i() const { return _i; } 797 798 799 void do_field(fieldDescriptor* fd) { 800 intptr_t val; 801 StackValue* value = 802 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 803 int offset = fd->offset(); 804 switch (fd->field_type()) { 805 case T_OBJECT: case T_ARRAY: 806 assert(value->type() == T_OBJECT, "Agreement."); 807 _obj->obj_field_put(offset, value->get_obj()()); 808 break; 809 810 case T_LONG: case T_DOUBLE: { 811 assert(value->type() == T_INT, "Agreement."); 812 StackValue* low = 813 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 814 #ifdef _LP64 815 jlong res = (jlong)low->get_int(); 816 #else 817 #ifdef SPARC 818 // For SPARC we have to swap high and low words. 819 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 820 #else 821 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 822 #endif //SPARC 823 #endif 824 _obj->long_field_put(offset, res); 825 break; 826 } 827 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 828 case T_INT: case T_FLOAT: // 4 bytes. 829 assert(value->type() == T_INT, "Agreement."); 830 val = value->get_int(); 831 _obj->int_field_put(offset, (jint)*((jint*)&val)); 832 break; 833 834 case T_SHORT: case T_CHAR: // 2 bytes 835 assert(value->type() == T_INT, "Agreement."); 836 val = value->get_int(); 837 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 838 break; 839 840 case T_BOOLEAN: case T_BYTE: // 1 byte 841 assert(value->type() == T_INT, "Agreement."); 842 val = value->get_int(); 843 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 844 break; 845 846 default: 847 ShouldNotReachHere(); 848 } 849 _i++; 850 } 851 }; 852 853 // restore elements of an eliminated type array 854 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 855 int index = 0; 856 intptr_t val; 857 858 for (int i = 0; i < sv->field_size(); i++) { 859 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 860 switch(type) { 861 case T_LONG: case T_DOUBLE: { 862 assert(value->type() == T_INT, "Agreement."); 863 StackValue* low = 864 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 865 #ifdef _LP64 866 jlong res = (jlong)low->get_int(); 867 #else 868 #ifdef SPARC 869 // For SPARC we have to swap high and low words. 870 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 871 #else 872 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 873 #endif //SPARC 874 #endif 875 obj->long_at_put(index, res); 876 break; 877 } 878 879 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 880 case T_INT: case T_FLOAT: // 4 bytes. 881 assert(value->type() == T_INT, "Agreement."); 882 val = value->get_int(); 883 obj->int_at_put(index, (jint)*((jint*)&val)); 884 break; 885 886 case T_SHORT: case T_CHAR: // 2 bytes 887 assert(value->type() == T_INT, "Agreement."); 888 val = value->get_int(); 889 obj->short_at_put(index, (jshort)*((jint*)&val)); 890 break; 891 892 case T_BOOLEAN: case T_BYTE: // 1 byte 893 assert(value->type() == T_INT, "Agreement."); 894 val = value->get_int(); 895 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 896 break; 897 898 default: 899 ShouldNotReachHere(); 900 } 901 index++; 902 } 903 } 904 905 906 // restore fields of an eliminated object array 907 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 908 for (int i = 0; i < sv->field_size(); i++) { 909 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 910 assert(value->type() == T_OBJECT, "object element expected"); 911 obj->obj_at_put(i, value->get_obj()()); 912 } 913 } 914 915 916 // restore fields of all eliminated objects and arrays 917 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 918 for (int i = 0; i < objects->length(); i++) { 919 ObjectValue* sv = (ObjectValue*) objects->at(i); 920 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 921 Handle obj = sv->value(); 922 assert(obj.not_null(), "reallocation was missed"); 923 924 if (k->oop_is_instance()) { 925 InstanceKlass* ik = InstanceKlass::cast(k()); 926 FieldReassigner reassign(fr, reg_map, sv, obj()); 927 ik->do_nonstatic_fields(&reassign); 928 } else if (k->oop_is_typeArray()) { 929 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 930 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 931 } else if (k->oop_is_objArray()) { 932 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 933 } 934 } 935 } 936 937 938 // relock objects for which synchronization was eliminated 939 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 940 for (int i = 0; i < monitors->length(); i++) { 941 MonitorInfo* mon_info = monitors->at(i); 942 if (mon_info->eliminated()) { 943 assert(mon_info->owner() != NULL, "reallocation was missed"); 944 Handle obj = Handle(mon_info->owner()); 945 markOop mark = obj->mark(); 946 if (UseBiasedLocking && mark->has_bias_pattern()) { 947 // New allocated objects may have the mark set to anonymously biased. 948 // Also the deoptimized method may called methods with synchronization 949 // where the thread-local object is bias locked to the current thread. 950 assert(mark->is_biased_anonymously() || 951 mark->biased_locker() == thread, "should be locked to current thread"); 952 // Reset mark word to unbiased prototype. 953 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 954 obj->set_mark(unbiased_prototype); 955 } 956 BasicLock* lock = mon_info->lock(); 957 ObjectSynchronizer::slow_enter(obj, lock, thread); 958 } 959 assert(mon_info->owner()->is_locked(), "object must be locked now"); 960 } 961 } 962 963 964 #ifndef PRODUCT 965 // print information about reallocated objects 966 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 967 fieldDescriptor fd; 968 969 for (int i = 0; i < objects->length(); i++) { 970 ObjectValue* sv = (ObjectValue*) objects->at(i); 971 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 972 Handle obj = sv->value(); 973 974 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 975 k->print_value(); 976 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 977 tty->cr(); 978 979 if (Verbose) { 980 k->oop_print_on(obj(), tty); 981 } 982 } 983 } 984 #endif 985 #endif // COMPILER2 986 987 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 988 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 989 990 #ifndef PRODUCT 991 if (TraceDeoptimization) { 992 ttyLocker ttyl; 993 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 994 fr.print_on(tty); 995 tty->print_cr(" Virtual frames (innermost first):"); 996 for (int index = 0; index < chunk->length(); index++) { 997 compiledVFrame* vf = chunk->at(index); 998 tty->print(" %2d - ", index); 999 vf->print_value(); 1000 int bci = chunk->at(index)->raw_bci(); 1001 const char* code_name; 1002 if (bci == SynchronizationEntryBCI) { 1003 code_name = "sync entry"; 1004 } else { 1005 Bytecodes::Code code = vf->method()->code_at(bci); 1006 code_name = Bytecodes::name(code); 1007 } 1008 tty->print(" - %s", code_name); 1009 tty->print_cr(" @ bci %d ", bci); 1010 if (Verbose) { 1011 vf->print(); 1012 tty->cr(); 1013 } 1014 } 1015 } 1016 #endif 1017 1018 // Register map for next frame (used for stack crawl). We capture 1019 // the state of the deopt'ing frame's caller. Thus if we need to 1020 // stuff a C2I adapter we can properly fill in the callee-save 1021 // register locations. 1022 frame caller = fr.sender(reg_map); 1023 int frame_size = caller.sp() - fr.sp(); 1024 1025 frame sender = caller; 1026 1027 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1028 // the vframeArray containing the unpacking information is allocated in the C heap. 1029 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1030 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1031 1032 // Compare the vframeArray to the collected vframes 1033 assert(array->structural_compare(thread, chunk), "just checking"); 1034 1035 #ifndef PRODUCT 1036 if (TraceDeoptimization) { 1037 ttyLocker ttyl; 1038 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1039 } 1040 #endif // PRODUCT 1041 1042 return array; 1043 } 1044 1045 1046 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1047 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1048 for (int i = 0; i < monitors->length(); i++) { 1049 MonitorInfo* mon_info = monitors->at(i); 1050 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1051 objects_to_revoke->append(Handle(mon_info->owner())); 1052 } 1053 } 1054 } 1055 1056 1057 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1058 if (!UseBiasedLocking) { 1059 return; 1060 } 1061 1062 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1063 1064 // Unfortunately we don't have a RegisterMap available in most of 1065 // the places we want to call this routine so we need to walk the 1066 // stack again to update the register map. 1067 if (map == NULL || !map->update_map()) { 1068 StackFrameStream sfs(thread, true); 1069 bool found = false; 1070 while (!found && !sfs.is_done()) { 1071 frame* cur = sfs.current(); 1072 sfs.next(); 1073 found = cur->id() == fr.id(); 1074 } 1075 assert(found, "frame to be deoptimized not found on target thread's stack"); 1076 map = sfs.register_map(); 1077 } 1078 1079 vframe* vf = vframe::new_vframe(&fr, map, thread); 1080 compiledVFrame* cvf = compiledVFrame::cast(vf); 1081 // Revoke monitors' biases in all scopes 1082 while (!cvf->is_top()) { 1083 collect_monitors(cvf, objects_to_revoke); 1084 cvf = compiledVFrame::cast(cvf->sender()); 1085 } 1086 collect_monitors(cvf, objects_to_revoke); 1087 1088 if (SafepointSynchronize::is_at_safepoint()) { 1089 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1090 } else { 1091 BiasedLocking::revoke(objects_to_revoke); 1092 } 1093 } 1094 1095 1096 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1097 if (!UseBiasedLocking) { 1098 return; 1099 } 1100 1101 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1102 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1103 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1104 if (jt->has_last_Java_frame()) { 1105 StackFrameStream sfs(jt, true); 1106 while (!sfs.is_done()) { 1107 frame* cur = sfs.current(); 1108 if (cb->contains(cur->pc())) { 1109 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1110 compiledVFrame* cvf = compiledVFrame::cast(vf); 1111 // Revoke monitors' biases in all scopes 1112 while (!cvf->is_top()) { 1113 collect_monitors(cvf, objects_to_revoke); 1114 cvf = compiledVFrame::cast(cvf->sender()); 1115 } 1116 collect_monitors(cvf, objects_to_revoke); 1117 } 1118 sfs.next(); 1119 } 1120 } 1121 } 1122 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1123 } 1124 1125 1126 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1127 assert(fr.can_be_deoptimized(), "checking frame type"); 1128 1129 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1130 1131 // Patch the nmethod so that when execution returns to it we will 1132 // deopt the execution state and return to the interpreter. 1133 fr.deoptimize(thread); 1134 } 1135 1136 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1137 // Deoptimize only if the frame comes from compile code. 1138 // Do not deoptimize the frame which is already patched 1139 // during the execution of the loops below. 1140 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1141 return; 1142 } 1143 ResourceMark rm; 1144 DeoptimizationMarker dm; 1145 if (UseBiasedLocking) { 1146 revoke_biases_of_monitors(thread, fr, map); 1147 } 1148 deoptimize_single_frame(thread, fr); 1149 1150 } 1151 1152 1153 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1154 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1155 "can only deoptimize other thread at a safepoint"); 1156 // Compute frame and register map based on thread and sp. 1157 RegisterMap reg_map(thread, UseBiasedLocking); 1158 frame fr = thread->last_frame(); 1159 while (fr.id() != id) { 1160 fr = fr.sender(®_map); 1161 } 1162 deoptimize(thread, fr, ®_map); 1163 } 1164 1165 1166 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1167 if (thread == Thread::current()) { 1168 Deoptimization::deoptimize_frame_internal(thread, id); 1169 } else { 1170 VM_DeoptimizeFrame deopt(thread, id); 1171 VMThread::execute(&deopt); 1172 } 1173 } 1174 1175 1176 // JVMTI PopFrame support 1177 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1178 { 1179 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1180 } 1181 JRT_END 1182 1183 1184 #if defined(COMPILER2) || defined(SHARK) 1185 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1186 // in case of an unresolved klass entry, load the class. 1187 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1188 Klass* tk = constant_pool->klass_at(index, CHECK); 1189 return; 1190 } 1191 1192 if (!constant_pool->tag_at(index).is_symbol()) return; 1193 1194 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1195 Symbol* symbol = constant_pool->symbol_at(index); 1196 1197 // class name? 1198 if (symbol->byte_at(0) != '(') { 1199 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1200 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1201 return; 1202 } 1203 1204 // then it must be a signature! 1205 ResourceMark rm(THREAD); 1206 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1207 if (ss.is_object()) { 1208 Symbol* class_name = ss.as_symbol(CHECK); 1209 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1210 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1211 } 1212 } 1213 } 1214 1215 1216 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1217 EXCEPTION_MARK; 1218 load_class_by_index(constant_pool, index, THREAD); 1219 if (HAS_PENDING_EXCEPTION) { 1220 // Exception happened during classloading. We ignore the exception here, since it 1221 // is going to be rethrown since the current activation is going to be deoptimzied and 1222 // the interpreter will re-execute the bytecode. 1223 CLEAR_PENDING_EXCEPTION; 1224 } 1225 } 1226 1227 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1228 HandleMark hm; 1229 1230 // uncommon_trap() is called at the beginning of the uncommon trap 1231 // handler. Note this fact before we start generating temporary frames 1232 // that can confuse an asynchronous stack walker. This counter is 1233 // decremented at the end of unpack_frames(). 1234 thread->inc_in_deopt_handler(); 1235 1236 // We need to update the map if we have biased locking. 1237 RegisterMap reg_map(thread, UseBiasedLocking); 1238 frame stub_frame = thread->last_frame(); 1239 frame fr = stub_frame.sender(®_map); 1240 // Make sure the calling nmethod is not getting deoptimized and removed 1241 // before we are done with it. 1242 nmethodLocker nl(fr.pc()); 1243 1244 // Log a message 1245 Events::log_deopt_message(thread, "Uncommon trap %d fr.pc " INTPTR_FORMAT, 1246 trap_request, fr.pc()); 1247 1248 { 1249 ResourceMark rm; 1250 1251 // Revoke biases of any monitors in the frame to ensure we can migrate them 1252 revoke_biases_of_monitors(thread, fr, ®_map); 1253 1254 DeoptReason reason = trap_request_reason(trap_request); 1255 DeoptAction action = trap_request_action(trap_request); 1256 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1257 1258 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1259 compiledVFrame* cvf = compiledVFrame::cast(vf); 1260 1261 nmethod* nm = cvf->code(); 1262 1263 ScopeDesc* trap_scope = cvf->scope(); 1264 methodHandle trap_method = trap_scope->method(); 1265 int trap_bci = trap_scope->bci(); 1266 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1267 1268 // Record this event in the histogram. 1269 gather_statistics(reason, action, trap_bc); 1270 1271 // Ensure that we can record deopt. history: 1272 bool create_if_missing = ProfileTraps; 1273 1274 MethodData* trap_mdo = 1275 get_method_data(thread, trap_method, create_if_missing); 1276 1277 // Print a bunch of diagnostics, if requested. 1278 if (TraceDeoptimization || LogCompilation) { 1279 ResourceMark rm; 1280 ttyLocker ttyl; 1281 char buf[100]; 1282 if (xtty != NULL) { 1283 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1284 os::current_thread_id(), 1285 format_trap_request(buf, sizeof(buf), trap_request)); 1286 nm->log_identity(xtty); 1287 } 1288 Symbol* class_name = NULL; 1289 bool unresolved = false; 1290 if (unloaded_class_index >= 0) { 1291 constantPoolHandle constants (THREAD, trap_method->constants()); 1292 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1293 class_name = constants->klass_name_at(unloaded_class_index); 1294 unresolved = true; 1295 if (xtty != NULL) 1296 xtty->print(" unresolved='1'"); 1297 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1298 class_name = constants->symbol_at(unloaded_class_index); 1299 } 1300 if (xtty != NULL) 1301 xtty->name(class_name); 1302 } 1303 if (xtty != NULL && trap_mdo != NULL) { 1304 // Dump the relevant MDO state. 1305 // This is the deopt count for the current reason, any previous 1306 // reasons or recompiles seen at this point. 1307 int dcnt = trap_mdo->trap_count(reason); 1308 if (dcnt != 0) 1309 xtty->print(" count='%d'", dcnt); 1310 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1311 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1312 if (dos != 0) { 1313 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1314 if (trap_state_is_recompiled(dos)) { 1315 int recnt2 = trap_mdo->overflow_recompile_count(); 1316 if (recnt2 != 0) 1317 xtty->print(" recompiles2='%d'", recnt2); 1318 } 1319 } 1320 } 1321 if (xtty != NULL) { 1322 xtty->stamp(); 1323 xtty->end_head(); 1324 } 1325 if (TraceDeoptimization) { // make noise on the tty 1326 tty->print("Uncommon trap occurred in"); 1327 nm->method()->print_short_name(tty); 1328 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d", 1329 fr.pc(), 1330 os::current_thread_id(), 1331 trap_reason_name(reason), 1332 trap_action_name(action), 1333 unloaded_class_index); 1334 if (class_name != NULL) { 1335 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1336 class_name->print_symbol_on(tty); 1337 } 1338 tty->cr(); 1339 } 1340 if (xtty != NULL) { 1341 // Log the precise location of the trap. 1342 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1343 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1344 xtty->method(sd->method()); 1345 xtty->end_elem(); 1346 if (sd->is_top()) break; 1347 } 1348 xtty->tail("uncommon_trap"); 1349 } 1350 } 1351 // (End diagnostic printout.) 1352 1353 // Load class if necessary 1354 if (unloaded_class_index >= 0) { 1355 constantPoolHandle constants(THREAD, trap_method->constants()); 1356 load_class_by_index(constants, unloaded_class_index); 1357 } 1358 1359 // Flush the nmethod if necessary and desirable. 1360 // 1361 // We need to avoid situations where we are re-flushing the nmethod 1362 // because of a hot deoptimization site. Repeated flushes at the same 1363 // point need to be detected by the compiler and avoided. If the compiler 1364 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1365 // module must take measures to avoid an infinite cycle of recompilation 1366 // and deoptimization. There are several such measures: 1367 // 1368 // 1. If a recompilation is ordered a second time at some site X 1369 // and for the same reason R, the action is adjusted to 'reinterpret', 1370 // to give the interpreter time to exercise the method more thoroughly. 1371 // If this happens, the method's overflow_recompile_count is incremented. 1372 // 1373 // 2. If the compiler fails to reduce the deoptimization rate, then 1374 // the method's overflow_recompile_count will begin to exceed the set 1375 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1376 // is adjusted to 'make_not_compilable', and the method is abandoned 1377 // to the interpreter. This is a performance hit for hot methods, 1378 // but is better than a disastrous infinite cycle of recompilations. 1379 // (Actually, only the method containing the site X is abandoned.) 1380 // 1381 // 3. In parallel with the previous measures, if the total number of 1382 // recompilations of a method exceeds the much larger set limit 1383 // PerMethodRecompilationCutoff, the method is abandoned. 1384 // This should only happen if the method is very large and has 1385 // many "lukewarm" deoptimizations. The code which enforces this 1386 // limit is elsewhere (class nmethod, class Method). 1387 // 1388 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1389 // to recompile at each bytecode independently of the per-BCI cutoff. 1390 // 1391 // The decision to update code is up to the compiler, and is encoded 1392 // in the Action_xxx code. If the compiler requests Action_none 1393 // no trap state is changed, no compiled code is changed, and the 1394 // computation suffers along in the interpreter. 1395 // 1396 // The other action codes specify various tactics for decompilation 1397 // and recompilation. Action_maybe_recompile is the loosest, and 1398 // allows the compiled code to stay around until enough traps are seen, 1399 // and until the compiler gets around to recompiling the trapping method. 1400 // 1401 // The other actions cause immediate removal of the present code. 1402 1403 bool update_trap_state = true; 1404 bool make_not_entrant = false; 1405 bool make_not_compilable = false; 1406 bool reprofile = false; 1407 switch (action) { 1408 case Action_none: 1409 // Keep the old code. 1410 update_trap_state = false; 1411 break; 1412 case Action_maybe_recompile: 1413 // Do not need to invalidate the present code, but we can 1414 // initiate another 1415 // Start compiler without (necessarily) invalidating the nmethod. 1416 // The system will tolerate the old code, but new code should be 1417 // generated when possible. 1418 break; 1419 case Action_reinterpret: 1420 // Go back into the interpreter for a while, and then consider 1421 // recompiling form scratch. 1422 make_not_entrant = true; 1423 // Reset invocation counter for outer most method. 1424 // This will allow the interpreter to exercise the bytecodes 1425 // for a while before recompiling. 1426 // By contrast, Action_make_not_entrant is immediate. 1427 // 1428 // Note that the compiler will track null_check, null_assert, 1429 // range_check, and class_check events and log them as if they 1430 // had been traps taken from compiled code. This will update 1431 // the MDO trap history so that the next compilation will 1432 // properly detect hot trap sites. 1433 reprofile = true; 1434 break; 1435 case Action_make_not_entrant: 1436 // Request immediate recompilation, and get rid of the old code. 1437 // Make them not entrant, so next time they are called they get 1438 // recompiled. Unloaded classes are loaded now so recompile before next 1439 // time they are called. Same for uninitialized. The interpreter will 1440 // link the missing class, if any. 1441 make_not_entrant = true; 1442 break; 1443 case Action_make_not_compilable: 1444 // Give up on compiling this method at all. 1445 make_not_entrant = true; 1446 make_not_compilable = true; 1447 break; 1448 default: 1449 ShouldNotReachHere(); 1450 } 1451 1452 // Setting +ProfileTraps fixes the following, on all platforms: 1453 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1454 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1455 // recompile relies on a MethodData* to record heroic opt failures. 1456 1457 // Whether the interpreter is producing MDO data or not, we also need 1458 // to use the MDO to detect hot deoptimization points and control 1459 // aggressive optimization. 1460 bool inc_recompile_count = false; 1461 ProfileData* pdata = NULL; 1462 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1463 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity"); 1464 uint this_trap_count = 0; 1465 bool maybe_prior_trap = false; 1466 bool maybe_prior_recompile = false; 1467 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1468 //outputs: 1469 this_trap_count, 1470 maybe_prior_trap, 1471 maybe_prior_recompile); 1472 // Because the interpreter also counts null, div0, range, and class 1473 // checks, these traps from compiled code are double-counted. 1474 // This is harmless; it just means that the PerXTrapLimit values 1475 // are in effect a little smaller than they look. 1476 1477 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1478 if (per_bc_reason != Reason_none) { 1479 // Now take action based on the partially known per-BCI history. 1480 if (maybe_prior_trap 1481 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1482 // If there are too many traps at this BCI, force a recompile. 1483 // This will allow the compiler to see the limit overflow, and 1484 // take corrective action, if possible. The compiler generally 1485 // does not use the exact PerBytecodeTrapLimit value, but instead 1486 // changes its tactics if it sees any traps at all. This provides 1487 // a little hysteresis, delaying a recompile until a trap happens 1488 // several times. 1489 // 1490 // Actually, since there is only one bit of counter per BCI, 1491 // the possible per-BCI counts are {0,1,(per-method count)}. 1492 // This produces accurate results if in fact there is only 1493 // one hot trap site, but begins to get fuzzy if there are 1494 // many sites. For example, if there are ten sites each 1495 // trapping two or more times, they each get the blame for 1496 // all of their traps. 1497 make_not_entrant = true; 1498 } 1499 1500 // Detect repeated recompilation at the same BCI, and enforce a limit. 1501 if (make_not_entrant && maybe_prior_recompile) { 1502 // More than one recompile at this point. 1503 inc_recompile_count = maybe_prior_trap; 1504 } 1505 } else { 1506 // For reasons which are not recorded per-bytecode, we simply 1507 // force recompiles unconditionally. 1508 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1509 make_not_entrant = true; 1510 } 1511 1512 // Go back to the compiler if there are too many traps in this method. 1513 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1514 // If there are too many traps in this method, force a recompile. 1515 // This will allow the compiler to see the limit overflow, and 1516 // take corrective action, if possible. 1517 // (This condition is an unlikely backstop only, because the 1518 // PerBytecodeTrapLimit is more likely to take effect first, 1519 // if it is applicable.) 1520 make_not_entrant = true; 1521 } 1522 1523 // Here's more hysteresis: If there has been a recompile at 1524 // this trap point already, run the method in the interpreter 1525 // for a while to exercise it more thoroughly. 1526 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1527 reprofile = true; 1528 } 1529 1530 } 1531 1532 // Take requested actions on the method: 1533 1534 // Recompile 1535 if (make_not_entrant) { 1536 if (!nm->make_not_entrant()) { 1537 return; // the call did not change nmethod's state 1538 } 1539 1540 if (pdata != NULL) { 1541 // Record the recompilation event, if any. 1542 int tstate0 = pdata->trap_state(); 1543 int tstate1 = trap_state_set_recompiled(tstate0, true); 1544 if (tstate1 != tstate0) 1545 pdata->set_trap_state(tstate1); 1546 } 1547 } 1548 1549 if (inc_recompile_count) { 1550 trap_mdo->inc_overflow_recompile_count(); 1551 if ((uint)trap_mdo->overflow_recompile_count() > 1552 (uint)PerBytecodeRecompilationCutoff) { 1553 // Give up on the method containing the bad BCI. 1554 if (trap_method() == nm->method()) { 1555 make_not_compilable = true; 1556 } else { 1557 trap_method->set_not_compilable(CompLevel_full_optimization); 1558 // But give grace to the enclosing nm->method(). 1559 } 1560 } 1561 } 1562 1563 // Reprofile 1564 if (reprofile) { 1565 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1566 } 1567 1568 // Give up compiling 1569 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1570 assert(make_not_entrant, "consistent"); 1571 nm->method()->set_not_compilable(CompLevel_full_optimization); 1572 } 1573 1574 } // Free marked resources 1575 1576 } 1577 JRT_END 1578 1579 MethodData* 1580 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1581 bool create_if_missing) { 1582 Thread* THREAD = thread; 1583 MethodData* mdo = m()->method_data(); 1584 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1585 // Build an MDO. Ignore errors like OutOfMemory; 1586 // that simply means we won't have an MDO to update. 1587 Method::build_interpreter_method_data(m, THREAD); 1588 if (HAS_PENDING_EXCEPTION) { 1589 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1590 CLEAR_PENDING_EXCEPTION; 1591 } 1592 mdo = m()->method_data(); 1593 } 1594 return mdo; 1595 } 1596 1597 ProfileData* 1598 Deoptimization::query_update_method_data(MethodData* trap_mdo, 1599 int trap_bci, 1600 Deoptimization::DeoptReason reason, 1601 //outputs: 1602 uint& ret_this_trap_count, 1603 bool& ret_maybe_prior_trap, 1604 bool& ret_maybe_prior_recompile) { 1605 uint prior_trap_count = trap_mdo->trap_count(reason); 1606 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1607 1608 // If the runtime cannot find a place to store trap history, 1609 // it is estimated based on the general condition of the method. 1610 // If the method has ever been recompiled, or has ever incurred 1611 // a trap with the present reason , then this BCI is assumed 1612 // (pessimistically) to be the culprit. 1613 bool maybe_prior_trap = (prior_trap_count != 0); 1614 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1615 ProfileData* pdata = NULL; 1616 1617 1618 // For reasons which are recorded per bytecode, we check per-BCI data. 1619 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1620 if (per_bc_reason != Reason_none) { 1621 // Find the profile data for this BCI. If there isn't one, 1622 // try to allocate one from the MDO's set of spares. 1623 // This will let us detect a repeated trap at this point. 1624 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1625 1626 if (pdata != NULL) { 1627 // Query the trap state of this profile datum. 1628 int tstate0 = pdata->trap_state(); 1629 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1630 maybe_prior_trap = false; 1631 if (!trap_state_is_recompiled(tstate0)) 1632 maybe_prior_recompile = false; 1633 1634 // Update the trap state of this profile datum. 1635 int tstate1 = tstate0; 1636 // Record the reason. 1637 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1638 // Store the updated state on the MDO, for next time. 1639 if (tstate1 != tstate0) 1640 pdata->set_trap_state(tstate1); 1641 } else { 1642 if (LogCompilation && xtty != NULL) { 1643 ttyLocker ttyl; 1644 // Missing MDP? Leave a small complaint in the log. 1645 xtty->elem("missing_mdp bci='%d'", trap_bci); 1646 } 1647 } 1648 } 1649 1650 // Return results: 1651 ret_this_trap_count = this_trap_count; 1652 ret_maybe_prior_trap = maybe_prior_trap; 1653 ret_maybe_prior_recompile = maybe_prior_recompile; 1654 return pdata; 1655 } 1656 1657 void 1658 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1659 ResourceMark rm; 1660 // Ignored outputs: 1661 uint ignore_this_trap_count; 1662 bool ignore_maybe_prior_trap; 1663 bool ignore_maybe_prior_recompile; 1664 query_update_method_data(trap_mdo, trap_bci, 1665 (DeoptReason)reason, 1666 ignore_this_trap_count, 1667 ignore_maybe_prior_trap, 1668 ignore_maybe_prior_recompile); 1669 } 1670 1671 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1672 1673 // Still in Java no safepoints 1674 { 1675 // This enters VM and may safepoint 1676 uncommon_trap_inner(thread, trap_request); 1677 } 1678 return fetch_unroll_info_helper(thread); 1679 } 1680 1681 // Local derived constants. 1682 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1683 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1684 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1685 1686 //---------------------------trap_state_reason--------------------------------- 1687 Deoptimization::DeoptReason 1688 Deoptimization::trap_state_reason(int trap_state) { 1689 // This assert provides the link between the width of DataLayout::trap_bits 1690 // and the encoding of "recorded" reasons. It ensures there are enough 1691 // bits to store all needed reasons in the per-BCI MDO profile. 1692 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1693 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1694 trap_state -= recompile_bit; 1695 if (trap_state == DS_REASON_MASK) { 1696 return Reason_many; 1697 } else { 1698 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1699 return (DeoptReason)trap_state; 1700 } 1701 } 1702 //-------------------------trap_state_has_reason------------------------------- 1703 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1704 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1705 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1706 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1707 trap_state -= recompile_bit; 1708 if (trap_state == DS_REASON_MASK) { 1709 return -1; // true, unspecifically (bottom of state lattice) 1710 } else if (trap_state == reason) { 1711 return 1; // true, definitely 1712 } else if (trap_state == 0) { 1713 return 0; // false, definitely (top of state lattice) 1714 } else { 1715 return 0; // false, definitely 1716 } 1717 } 1718 //-------------------------trap_state_add_reason------------------------------- 1719 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1720 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1721 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1722 trap_state -= recompile_bit; 1723 if (trap_state == DS_REASON_MASK) { 1724 return trap_state + recompile_bit; // already at state lattice bottom 1725 } else if (trap_state == reason) { 1726 return trap_state + recompile_bit; // the condition is already true 1727 } else if (trap_state == 0) { 1728 return reason + recompile_bit; // no condition has yet been true 1729 } else { 1730 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1731 } 1732 } 1733 //-----------------------trap_state_is_recompiled------------------------------ 1734 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1735 return (trap_state & DS_RECOMPILE_BIT) != 0; 1736 } 1737 //-----------------------trap_state_set_recompiled----------------------------- 1738 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1739 if (z) return trap_state | DS_RECOMPILE_BIT; 1740 else return trap_state & ~DS_RECOMPILE_BIT; 1741 } 1742 //---------------------------format_trap_state--------------------------------- 1743 // This is used for debugging and diagnostics, including hotspot.log output. 1744 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1745 int trap_state) { 1746 DeoptReason reason = trap_state_reason(trap_state); 1747 bool recomp_flag = trap_state_is_recompiled(trap_state); 1748 // Re-encode the state from its decoded components. 1749 int decoded_state = 0; 1750 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1751 decoded_state = trap_state_add_reason(decoded_state, reason); 1752 if (recomp_flag) 1753 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1754 // If the state re-encodes properly, format it symbolically. 1755 // Because this routine is used for debugging and diagnostics, 1756 // be robust even if the state is a strange value. 1757 size_t len; 1758 if (decoded_state != trap_state) { 1759 // Random buggy state that doesn't decode?? 1760 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1761 } else { 1762 len = jio_snprintf(buf, buflen, "%s%s", 1763 trap_reason_name(reason), 1764 recomp_flag ? " recompiled" : ""); 1765 } 1766 if (len >= buflen) 1767 buf[buflen-1] = '\0'; 1768 return buf; 1769 } 1770 1771 1772 //--------------------------------statics-------------------------------------- 1773 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1774 = Deoptimization::Action_reinterpret; 1775 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1776 // Note: Keep this in sync. with enum DeoptReason. 1777 "none", 1778 "null_check", 1779 "null_assert", 1780 "range_check", 1781 "class_check", 1782 "array_check", 1783 "intrinsic", 1784 "bimorphic", 1785 "unloaded", 1786 "uninitialized", 1787 "unreached", 1788 "unhandled", 1789 "constraint", 1790 "div0_check", 1791 "age", 1792 "predicate", 1793 "loop_limit_check" 1794 }; 1795 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1796 // Note: Keep this in sync. with enum DeoptAction. 1797 "none", 1798 "maybe_recompile", 1799 "reinterpret", 1800 "make_not_entrant", 1801 "make_not_compilable" 1802 }; 1803 1804 const char* Deoptimization::trap_reason_name(int reason) { 1805 if (reason == Reason_many) return "many"; 1806 if ((uint)reason < Reason_LIMIT) 1807 return _trap_reason_name[reason]; 1808 static char buf[20]; 1809 sprintf(buf, "reason%d", reason); 1810 return buf; 1811 } 1812 const char* Deoptimization::trap_action_name(int action) { 1813 if ((uint)action < Action_LIMIT) 1814 return _trap_action_name[action]; 1815 static char buf[20]; 1816 sprintf(buf, "action%d", action); 1817 return buf; 1818 } 1819 1820 // This is used for debugging and diagnostics, including hotspot.log output. 1821 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1822 int trap_request) { 1823 jint unloaded_class_index = trap_request_index(trap_request); 1824 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1825 const char* action = trap_action_name(trap_request_action(trap_request)); 1826 size_t len; 1827 if (unloaded_class_index < 0) { 1828 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1829 reason, action); 1830 } else { 1831 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1832 reason, action, unloaded_class_index); 1833 } 1834 if (len >= buflen) 1835 buf[buflen-1] = '\0'; 1836 return buf; 1837 } 1838 1839 juint Deoptimization::_deoptimization_hist 1840 [Deoptimization::Reason_LIMIT] 1841 [1 + Deoptimization::Action_LIMIT] 1842 [Deoptimization::BC_CASE_LIMIT] 1843 = {0}; 1844 1845 enum { 1846 LSB_BITS = 8, 1847 LSB_MASK = right_n_bits(LSB_BITS) 1848 }; 1849 1850 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1851 Bytecodes::Code bc) { 1852 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1853 assert(action >= 0 && action < Action_LIMIT, "oob"); 1854 _deoptimization_hist[Reason_none][0][0] += 1; // total 1855 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1856 juint* cases = _deoptimization_hist[reason][1+action]; 1857 juint* bc_counter_addr = NULL; 1858 juint bc_counter = 0; 1859 // Look for an unused counter, or an exact match to this BC. 1860 if (bc != Bytecodes::_illegal) { 1861 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1862 juint* counter_addr = &cases[bc_case]; 1863 juint counter = *counter_addr; 1864 if ((counter == 0 && bc_counter_addr == NULL) 1865 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1866 // this counter is either free or is already devoted to this BC 1867 bc_counter_addr = counter_addr; 1868 bc_counter = counter | bc; 1869 } 1870 } 1871 } 1872 if (bc_counter_addr == NULL) { 1873 // Overflow, or no given bytecode. 1874 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1875 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1876 } 1877 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1878 } 1879 1880 jint Deoptimization::total_deoptimization_count() { 1881 return _deoptimization_hist[Reason_none][0][0]; 1882 } 1883 1884 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1885 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1886 return _deoptimization_hist[reason][0][0]; 1887 } 1888 1889 void Deoptimization::print_statistics() { 1890 juint total = total_deoptimization_count(); 1891 juint account = total; 1892 if (total != 0) { 1893 ttyLocker ttyl; 1894 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1895 tty->print_cr("Deoptimization traps recorded:"); 1896 #define PRINT_STAT_LINE(name, r) \ 1897 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1898 PRINT_STAT_LINE("total", total); 1899 // For each non-zero entry in the histogram, print the reason, 1900 // the action, and (if specifically known) the type of bytecode. 1901 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1902 for (int action = 0; action < Action_LIMIT; action++) { 1903 juint* cases = _deoptimization_hist[reason][1+action]; 1904 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1905 juint counter = cases[bc_case]; 1906 if (counter != 0) { 1907 char name[1*K]; 1908 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1909 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1910 bc = Bytecodes::_illegal; 1911 sprintf(name, "%s/%s/%s", 1912 trap_reason_name(reason), 1913 trap_action_name(action), 1914 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1915 juint r = counter >> LSB_BITS; 1916 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1917 account -= r; 1918 } 1919 } 1920 } 1921 } 1922 if (account != 0) { 1923 PRINT_STAT_LINE("unaccounted", account); 1924 } 1925 #undef PRINT_STAT_LINE 1926 if (xtty != NULL) xtty->tail("statistics"); 1927 } 1928 } 1929 #else // COMPILER2 || SHARK 1930 1931 1932 // Stubs for C1 only system. 1933 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1934 return false; 1935 } 1936 1937 const char* Deoptimization::trap_reason_name(int reason) { 1938 return "unknown"; 1939 } 1940 1941 void Deoptimization::print_statistics() { 1942 // no output 1943 } 1944 1945 void 1946 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1947 // no udpate 1948 } 1949 1950 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1951 return 0; 1952 } 1953 1954 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1955 Bytecodes::Code bc) { 1956 // no update 1957 } 1958 1959 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1960 int trap_state) { 1961 jio_snprintf(buf, buflen, "#%d", trap_state); 1962 return buf; 1963 } 1964 1965 #endif // COMPILER2 || SHARK