deoptimization.cpp revision 5946:b5c8a61d7fa0
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_32 85# include "adfiles/ad_ppc_32.hpp" 86#endif 87#ifdef TARGET_ARCH_MODEL_ppc_64 88# include "adfiles/ad_ppc_64.hpp" 89#endif 90#endif // COMPILER2 91 92bool DeoptimizationMarker::_is_active = false; 93 94Deoptimization::UnrollBlock::UnrollBlock(int size_of_deoptimized_frame, 95 int caller_adjustment, 96 int caller_actual_parameters, 97 int number_of_frames, 98 intptr_t* frame_sizes, 99 address* frame_pcs, 100 BasicType return_type) { 101 _size_of_deoptimized_frame = size_of_deoptimized_frame; 102 _caller_adjustment = caller_adjustment; 103 _caller_actual_parameters = caller_actual_parameters; 104 _number_of_frames = number_of_frames; 105 _frame_sizes = frame_sizes; 106 _frame_pcs = frame_pcs; 107 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler); 108 _return_type = return_type; 109 _initial_info = 0; 110 // PD (x86 only) 111 _counter_temp = 0; 112 _unpack_kind = 0; 113 _sender_sp_temp = 0; 114 115 _total_frame_sizes = size_of_frames(); 116} 117 118 119Deoptimization::UnrollBlock::~UnrollBlock() { 120 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes, mtCompiler); 121 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs, mtCompiler); 122 FREE_C_HEAP_ARRAY(intptr_t, _register_block, mtCompiler); 123} 124 125 126intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 127 assert(register_number < RegisterMap::reg_count, "checking register number"); 128 return &_register_block[register_number * 2]; 129} 130 131 132 133int Deoptimization::UnrollBlock::size_of_frames() const { 134 // Acount first for the adjustment of the initial frame 135 int result = _caller_adjustment; 136 for (int index = 0; index < number_of_frames(); index++) { 137 result += frame_sizes()[index]; 138 } 139 return result; 140} 141 142 143void Deoptimization::UnrollBlock::print() { 144 ttyLocker ttyl; 145 tty->print_cr("UnrollBlock"); 146 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 147 tty->print( " frame_sizes: "); 148 for (int index = 0; index < number_of_frames(); index++) { 149 tty->print("%d ", frame_sizes()[index]); 150 } 151 tty->cr(); 152} 153 154 155// In order to make fetch_unroll_info work properly with escape 156// analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 157// ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 158// of previously eliminated objects occurs in realloc_objects, which is 159// called from the method fetch_unroll_info_helper below. 160JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 161 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 162 // but makes the entry a little slower. There is however a little dance we have to 163 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 164 165 // fetch_unroll_info() is called at the beginning of the deoptimization 166 // handler. Note this fact before we start generating temporary frames 167 // that can confuse an asynchronous stack walker. This counter is 168 // decremented at the end of unpack_frames(). 169 thread->inc_in_deopt_handler(); 170 171 return fetch_unroll_info_helper(thread); 172JRT_END 173 174 175// This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 176Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 177 178 // Note: there is a safepoint safety issue here. No matter whether we enter 179 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 180 // the vframeArray is created. 181 // 182 183 // Allocate our special deoptimization ResourceMark 184 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 185 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 186 thread->set_deopt_mark(dmark); 187 188 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 189 RegisterMap map(thread, true); 190 RegisterMap dummy_map(thread, false); 191 // Now get the deoptee with a valid map 192 frame deoptee = stub_frame.sender(&map); 193 // Set the deoptee nmethod 194 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 195 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 196 197 if (VerifyStack) { 198 thread->validate_frame_layout(); 199 } 200 201 // Create a growable array of VFrames where each VFrame represents an inlined 202 // Java frame. This storage is allocated with the usual system arena. 203 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 204 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 205 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 206 while (!vf->is_top()) { 207 assert(vf->is_compiled_frame(), "Wrong frame type"); 208 chunk->push(compiledVFrame::cast(vf)); 209 vf = vf->sender(); 210 } 211 assert(vf->is_compiled_frame(), "Wrong frame type"); 212 chunk->push(compiledVFrame::cast(vf)); 213 214#ifdef COMPILER2 215 // Reallocate the non-escaping objects and restore their fields. Then 216 // relock objects if synchronization on them was eliminated. 217 if (DoEscapeAnalysis || EliminateNestedLocks) { 218 if (EliminateAllocations) { 219 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 220 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 221 222 // The flag return_oop() indicates call sites which return oop 223 // in compiled code. Such sites include java method calls, 224 // runtime calls (for example, used to allocate new objects/arrays 225 // on slow code path) and any other calls generated in compiled code. 226 // It is not guaranteed that we can get such information here only 227 // by analyzing bytecode in deoptimized frames. This is why this flag 228 // is set during method compilation (see Compile::Process_OopMap_Node()). 229 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 230 Handle return_value; 231 if (save_oop_result) { 232 // Reallocation may trigger GC. If deoptimization happened on return from 233 // call which returns oop we need to save it since it is not in oopmap. 234 oop result = deoptee.saved_oop_result(&map); 235 assert(result == NULL || result->is_oop(), "must be oop"); 236 return_value = Handle(thread, result); 237 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 238 if (TraceDeoptimization) { 239 ttyLocker ttyl; 240 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); 241 } 242 } 243 bool reallocated = false; 244 if (objects != NULL) { 245 JRT_BLOCK 246 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 247 JRT_END 248 } 249 if (reallocated) { 250 reassign_fields(&deoptee, &map, objects); 251#ifndef PRODUCT 252 if (TraceDeoptimization) { 253 ttyLocker ttyl; 254 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 255 print_objects(objects); 256 } 257#endif 258 } 259 if (save_oop_result) { 260 // Restore result. 261 deoptee.set_saved_oop_result(&map, return_value()); 262 } 263 } 264 if (EliminateLocks) { 265#ifndef PRODUCT 266 bool first = true; 267#endif 268 for (int i = 0; i < chunk->length(); i++) { 269 compiledVFrame* cvf = chunk->at(i); 270 assert (cvf->scope() != NULL,"expect only compiled java frames"); 271 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 272 if (monitors->is_nonempty()) { 273 relock_objects(monitors, thread); 274#ifndef PRODUCT 275 if (TraceDeoptimization) { 276 ttyLocker ttyl; 277 for (int j = 0; j < monitors->length(); j++) { 278 MonitorInfo* mi = monitors->at(j); 279 if (mi->eliminated()) { 280 if (first) { 281 first = false; 282 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 283 } 284 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 285 } 286 } 287 } 288#endif 289 } 290 } 291 } 292 } 293#endif // COMPILER2 294 // Ensure that no safepoint is taken after pointers have been stored 295 // in fields of rematerialized objects. If a safepoint occurs from here on 296 // out the java state residing in the vframeArray will be missed. 297 No_Safepoint_Verifier no_safepoint; 298 299 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 300 301 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 302 thread->set_vframe_array_head(array); 303 304 // Now that the vframeArray has been created if we have any deferred local writes 305 // added by jvmti then we can free up that structure as the data is now in the 306 // vframeArray 307 308 if (thread->deferred_locals() != NULL) { 309 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 310 int i = 0; 311 do { 312 // Because of inlining we could have multiple vframes for a single frame 313 // and several of the vframes could have deferred writes. Find them all. 314 if (list->at(i)->id() == array->original().id()) { 315 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 316 list->remove_at(i); 317 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 318 delete dlv; 319 } else { 320 i++; 321 } 322 } while ( i < list->length() ); 323 if (list->length() == 0) { 324 thread->set_deferred_locals(NULL); 325 // free the list and elements back to C heap. 326 delete list; 327 } 328 329 } 330 331#ifndef SHARK 332 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 333 CodeBlob* cb = stub_frame.cb(); 334 // Verify we have the right vframeArray 335 assert(cb->frame_size() >= 0, "Unexpected frame size"); 336 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 337 338 // If the deopt call site is a MethodHandle invoke call site we have 339 // to adjust the unpack_sp. 340 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 341 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 342 unpack_sp = deoptee.unextended_sp(); 343 344#ifdef ASSERT 345 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 346#endif 347#else 348 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 349#endif // !SHARK 350 351 // This is a guarantee instead of an assert because if vframe doesn't match 352 // we will unpack the wrong deoptimized frame and wind up in strange places 353 // where it will be very difficult to figure out what went wrong. Better 354 // to die an early death here than some very obscure death later when the 355 // trail is cold. 356 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 357 // in that it will fail to detect a problem when there is one. This needs 358 // more work in tiger timeframe. 359 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 360 361 int number_of_frames = array->frames(); 362 363 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 364 // virtual activation, which is the reverse of the elements in the vframes array. 365 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler); 366 // +1 because we always have an interpreter return address for the final slot. 367 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler); 368 int popframe_extra_args = 0; 369 // Create an interpreter return address for the stub to use as its return 370 // address so the skeletal frames are perfectly walkable 371 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 372 373 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 374 // activation be put back on the expression stack of the caller for reexecution 375 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 376 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 377 } 378 379 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 380 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 381 // than simply use array->sender.pc(). This requires us to walk the current set of frames 382 // 383 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 384 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 385 386 // It's possible that the number of paramters at the call site is 387 // different than number of arguments in the callee when method 388 // handles are used. If the caller is interpreted get the real 389 // value so that the proper amount of space can be added to it's 390 // frame. 391 bool caller_was_method_handle = false; 392 if (deopt_sender.is_interpreted_frame()) { 393 methodHandle method = deopt_sender.interpreter_frame_method(); 394 Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci()); 395 if (cur.is_invokedynamic() || cur.is_invokehandle()) { 396 // Method handle invokes may involve fairly arbitrary chains of 397 // calls so it's impossible to know how much actual space the 398 // caller has for locals. 399 caller_was_method_handle = true; 400 } 401 } 402 403 // 404 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 405 // frame_sizes/frame_pcs[1] next oldest frame (int) 406 // frame_sizes/frame_pcs[n] youngest frame (int) 407 // 408 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 409 // owns the space for the return address to it's caller). Confusing ain't it. 410 // 411 // The vframe array can address vframes with indices running from 412 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 413 // When we create the skeletal frames we need the oldest frame to be in the zero slot 414 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 415 // so things look a little strange in this loop. 416 // 417 int callee_parameters = 0; 418 int callee_locals = 0; 419 for (int index = 0; index < array->frames(); index++ ) { 420 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 421 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 422 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 423 int caller_parms = callee_parameters; 424 if ((index == array->frames() - 1) && caller_was_method_handle) { 425 caller_parms = 0; 426 } 427 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(caller_parms, 428 callee_parameters, 429 callee_locals, 430 index == 0, 431 index == array->frames() - 1, 432 popframe_extra_args); 433 // This pc doesn't have to be perfect just good enough to identify the frame 434 // as interpreted so the skeleton frame will be walkable 435 // The correct pc will be set when the skeleton frame is completely filled out 436 // The final pc we store in the loop is wrong and will be overwritten below 437 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 438 439 callee_parameters = array->element(index)->method()->size_of_parameters(); 440 callee_locals = array->element(index)->method()->max_locals(); 441 popframe_extra_args = 0; 442 } 443 444 // Compute whether the root vframe returns a float or double value. 445 BasicType return_type; 446 { 447 HandleMark hm; 448 methodHandle method(thread, array->element(0)->method()); 449 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 450 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 451 } 452 453 // Compute information for handling adapters and adjusting the frame size of the caller. 454 int caller_adjustment = 0; 455 456 // Compute the amount the oldest interpreter frame will have to adjust 457 // its caller's stack by. If the caller is a compiled frame then 458 // we pretend that the callee has no parameters so that the 459 // extension counts for the full amount of locals and not just 460 // locals-parms. This is because without a c2i adapter the parm 461 // area as created by the compiled frame will not be usable by 462 // the interpreter. (Depending on the calling convention there 463 // may not even be enough space). 464 465 // QQQ I'd rather see this pushed down into last_frame_adjust 466 // and have it take the sender (aka caller). 467 468 if (deopt_sender.is_compiled_frame() || caller_was_method_handle) { 469 caller_adjustment = last_frame_adjust(0, callee_locals); 470 } else if (callee_locals > callee_parameters) { 471 // The caller frame may need extending to accommodate 472 // non-parameter locals of the first unpacked interpreted frame. 473 // Compute that adjustment. 474 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 475 } 476 477 // If the sender is deoptimized the we must retrieve the address of the handler 478 // since the frame will "magically" show the original pc before the deopt 479 // and we'd undo the deopt. 480 481 frame_pcs[0] = deopt_sender.raw_pc(); 482 483#ifndef SHARK 484 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 485#endif // SHARK 486 487 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 488 caller_adjustment * BytesPerWord, 489 caller_was_method_handle ? 0 : callee_parameters, 490 number_of_frames, 491 frame_sizes, 492 frame_pcs, 493 return_type); 494 // On some platforms, we need a way to pass some platform dependent 495 // information to the unpacking code so the skeletal frames come out 496 // correct (initial fp value, unextended sp, ...) 497 info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info()); 498 499 if (array->frames() > 1) { 500 if (VerifyStack && TraceDeoptimization) { 501 ttyLocker ttyl; 502 tty->print_cr("Deoptimizing method containing inlining"); 503 } 504 } 505 506 array->set_unroll_block(info); 507 return info; 508} 509 510// Called to cleanup deoptimization data structures in normal case 511// after unpacking to stack and when stack overflow error occurs 512void Deoptimization::cleanup_deopt_info(JavaThread *thread, 513 vframeArray *array) { 514 515 // Get array if coming from exception 516 if (array == NULL) { 517 array = thread->vframe_array_head(); 518 } 519 thread->set_vframe_array_head(NULL); 520 521 // Free the previous UnrollBlock 522 vframeArray* old_array = thread->vframe_array_last(); 523 thread->set_vframe_array_last(array); 524 525 if (old_array != NULL) { 526 UnrollBlock* old_info = old_array->unroll_block(); 527 old_array->set_unroll_block(NULL); 528 delete old_info; 529 delete old_array; 530 } 531 532 // Deallocate any resource creating in this routine and any ResourceObjs allocated 533 // inside the vframeArray (StackValueCollections) 534 535 delete thread->deopt_mark(); 536 thread->set_deopt_mark(NULL); 537 thread->set_deopt_nmethod(NULL); 538 539 540 if (JvmtiExport::can_pop_frame()) { 541#ifndef CC_INTERP 542 // Regardless of whether we entered this routine with the pending 543 // popframe condition bit set, we should always clear it now 544 thread->clear_popframe_condition(); 545#else 546 // C++ interpeter will clear has_pending_popframe when it enters 547 // with method_resume. For deopt_resume2 we clear it now. 548 if (thread->popframe_forcing_deopt_reexecution()) 549 thread->clear_popframe_condition(); 550#endif /* CC_INTERP */ 551 } 552 553 // unpack_frames() is called at the end of the deoptimization handler 554 // and (in C2) at the end of the uncommon trap handler. Note this fact 555 // so that an asynchronous stack walker can work again. This counter is 556 // incremented at the beginning of fetch_unroll_info() and (in C2) at 557 // the beginning of uncommon_trap(). 558 thread->dec_in_deopt_handler(); 559} 560 561 562// Return BasicType of value being returned 563JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 564 565 // We are already active int he special DeoptResourceMark any ResourceObj's we 566 // allocate will be freed at the end of the routine. 567 568 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 569 // but makes the entry a little slower. There is however a little dance we have to 570 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 571 ResetNoHandleMark rnhm; // No-op in release/product versions 572 HandleMark hm; 573 574 frame stub_frame = thread->last_frame(); 575 576 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 577 // must point to the vframeArray for the unpack frame. 578 vframeArray* array = thread->vframe_array_head(); 579 580#ifndef PRODUCT 581 if (TraceDeoptimization) { 582 ttyLocker ttyl; 583 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 584 } 585#endif 586 Events::log(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d", 587 stub_frame.pc(), stub_frame.sp(), exec_mode); 588 589 UnrollBlock* info = array->unroll_block(); 590 591 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 592 array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters()); 593 594 BasicType bt = info->return_type(); 595 596 // If we have an exception pending, claim that the return type is an oop 597 // so the deopt_blob does not overwrite the exception_oop. 598 599 if (exec_mode == Unpack_exception) 600 bt = T_OBJECT; 601 602 // Cleanup thread deopt data 603 cleanup_deopt_info(thread, array); 604 605#ifndef PRODUCT 606 if (VerifyStack) { 607 ResourceMark res_mark; 608 609 thread->validate_frame_layout(); 610 611 // Verify that the just-unpacked frames match the interpreter's 612 // notions of expression stack and locals 613 vframeArray* cur_array = thread->vframe_array_last(); 614 RegisterMap rm(thread, false); 615 rm.set_include_argument_oops(false); 616 bool is_top_frame = true; 617 int callee_size_of_parameters = 0; 618 int callee_max_locals = 0; 619 for (int i = 0; i < cur_array->frames(); i++) { 620 vframeArrayElement* el = cur_array->element(i); 621 frame* iframe = el->iframe(); 622 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 623 624 // Get the oop map for this bci 625 InterpreterOopMap mask; 626 int cur_invoke_parameter_size = 0; 627 bool try_next_mask = false; 628 int next_mask_expression_stack_size = -1; 629 int top_frame_expression_stack_adjustment = 0; 630 methodHandle mh(thread, iframe->interpreter_frame_method()); 631 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 632 BytecodeStream str(mh); 633 str.set_start(iframe->interpreter_frame_bci()); 634 int max_bci = mh->code_size(); 635 // Get to the next bytecode if possible 636 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 637 // Check to see if we can grab the number of outgoing arguments 638 // at an uncommon trap for an invoke (where the compiler 639 // generates debug info before the invoke has executed) 640 Bytecodes::Code cur_code = str.next(); 641 if (cur_code == Bytecodes::_invokevirtual || 642 cur_code == Bytecodes::_invokespecial || 643 cur_code == Bytecodes::_invokestatic || 644 cur_code == Bytecodes::_invokeinterface || 645 cur_code == Bytecodes::_invokedynamic) { 646 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 647 Symbol* signature = invoke.signature(); 648 ArgumentSizeComputer asc(signature); 649 cur_invoke_parameter_size = asc.size(); 650 if (invoke.has_receiver()) { 651 // Add in receiver 652 ++cur_invoke_parameter_size; 653 } 654 if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) { 655 callee_size_of_parameters++; 656 } 657 } 658 if (str.bci() < max_bci) { 659 Bytecodes::Code bc = str.next(); 660 if (bc >= 0) { 661 // The interpreter oop map generator reports results before 662 // the current bytecode has executed except in the case of 663 // calls. It seems to be hard to tell whether the compiler 664 // has emitted debug information matching the "state before" 665 // a given bytecode or the state after, so we try both 666 switch (cur_code) { 667 case Bytecodes::_invokevirtual: 668 case Bytecodes::_invokespecial: 669 case Bytecodes::_invokestatic: 670 case Bytecodes::_invokeinterface: 671 case Bytecodes::_invokedynamic: 672 case Bytecodes::_athrow: 673 break; 674 default: { 675 InterpreterOopMap next_mask; 676 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 677 next_mask_expression_stack_size = next_mask.expression_stack_size(); 678 // Need to subtract off the size of the result type of 679 // the bytecode because this is not described in the 680 // debug info but returned to the interpreter in the TOS 681 // caching register 682 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 683 if (bytecode_result_type != T_ILLEGAL) { 684 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 685 } 686 assert(top_frame_expression_stack_adjustment >= 0, ""); 687 try_next_mask = true; 688 break; 689 } 690 } 691 } 692 } 693 694 // Verify stack depth and oops in frame 695 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 696 if (!( 697 /* SPARC */ 698 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 699 /* x86 */ 700 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 701 (try_next_mask && 702 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 703 top_frame_expression_stack_adjustment))) || 704 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 705 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 706 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 707 )) { 708 ttyLocker ttyl; 709 710 // Print out some information that will help us debug the problem 711 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 712 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 713 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 714 iframe->interpreter_frame_expression_stack_size()); 715 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 716 tty->print_cr(" try_next_mask = %d", try_next_mask); 717 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 718 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 719 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 720 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 721 tty->print_cr(" exec_mode = %d", exec_mode); 722 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 723 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 724 tty->print_cr(" Interpreted frames:"); 725 for (int k = 0; k < cur_array->frames(); k++) { 726 vframeArrayElement* el = cur_array->element(k); 727 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 728 } 729 cur_array->print_on_2(tty); 730 guarantee(false, "wrong number of expression stack elements during deopt"); 731 } 732 VerifyOopClosure verify; 733 iframe->oops_interpreted_do(&verify, NULL, &rm, false); 734 callee_size_of_parameters = mh->size_of_parameters(); 735 callee_max_locals = mh->max_locals(); 736 is_top_frame = false; 737 } 738 } 739#endif /* !PRODUCT */ 740 741 742 return bt; 743JRT_END 744 745 746int Deoptimization::deoptimize_dependents() { 747 Threads::deoptimized_wrt_marked_nmethods(); 748 return 0; 749} 750 751 752#ifdef COMPILER2 753bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 754 Handle pending_exception(thread->pending_exception()); 755 const char* exception_file = thread->exception_file(); 756 int exception_line = thread->exception_line(); 757 thread->clear_pending_exception(); 758 759 for (int i = 0; i < objects->length(); i++) { 760 assert(objects->at(i)->is_object(), "invalid debug information"); 761 ObjectValue* sv = (ObjectValue*) objects->at(i); 762 763 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 764 oop obj = NULL; 765 766 if (k->oop_is_instance()) { 767 InstanceKlass* ik = InstanceKlass::cast(k()); 768 obj = ik->allocate_instance(CHECK_(false)); 769 } else if (k->oop_is_typeArray()) { 770 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 771 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 772 int len = sv->field_size() / type2size[ak->element_type()]; 773 obj = ak->allocate(len, CHECK_(false)); 774 } else if (k->oop_is_objArray()) { 775 ObjArrayKlass* ak = ObjArrayKlass::cast(k()); 776 obj = ak->allocate(sv->field_size(), CHECK_(false)); 777 } 778 779 assert(obj != NULL, "allocation failed"); 780 assert(sv->value().is_null(), "redundant reallocation"); 781 sv->set_value(obj); 782 } 783 784 if (pending_exception.not_null()) { 785 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 786 } 787 788 return true; 789} 790 791// This assumes that the fields are stored in ObjectValue in the same order 792// they are yielded by do_nonstatic_fields. 793class FieldReassigner: public FieldClosure { 794 frame* _fr; 795 RegisterMap* _reg_map; 796 ObjectValue* _sv; 797 InstanceKlass* _ik; 798 oop _obj; 799 800 int _i; 801public: 802 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 803 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 804 805 int i() const { return _i; } 806 807 808 void do_field(fieldDescriptor* fd) { 809 intptr_t val; 810 StackValue* value = 811 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 812 int offset = fd->offset(); 813 switch (fd->field_type()) { 814 case T_OBJECT: case T_ARRAY: 815 assert(value->type() == T_OBJECT, "Agreement."); 816 _obj->obj_field_put(offset, value->get_obj()()); 817 break; 818 819 case T_LONG: case T_DOUBLE: { 820 assert(value->type() == T_INT, "Agreement."); 821 StackValue* low = 822 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 823#ifdef _LP64 824 jlong res = (jlong)low->get_int(); 825#else 826#ifdef SPARC 827 // For SPARC we have to swap high and low words. 828 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 829#else 830 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 831#endif //SPARC 832#endif 833 _obj->long_field_put(offset, res); 834 break; 835 } 836 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 837 case T_INT: case T_FLOAT: // 4 bytes. 838 assert(value->type() == T_INT, "Agreement."); 839 val = value->get_int(); 840 _obj->int_field_put(offset, (jint)*((jint*)&val)); 841 break; 842 843 case T_SHORT: case T_CHAR: // 2 bytes 844 assert(value->type() == T_INT, "Agreement."); 845 val = value->get_int(); 846 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 847 break; 848 849 case T_BOOLEAN: case T_BYTE: // 1 byte 850 assert(value->type() == T_INT, "Agreement."); 851 val = value->get_int(); 852 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 853 break; 854 855 default: 856 ShouldNotReachHere(); 857 } 858 _i++; 859 } 860}; 861 862// restore elements of an eliminated type array 863void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 864 int index = 0; 865 intptr_t val; 866 867 for (int i = 0; i < sv->field_size(); i++) { 868 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 869 switch(type) { 870 case T_LONG: case T_DOUBLE: { 871 assert(value->type() == T_INT, "Agreement."); 872 StackValue* low = 873 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 874#ifdef _LP64 875 jlong res = (jlong)low->get_int(); 876#else 877#ifdef SPARC 878 // For SPARC we have to swap high and low words. 879 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 880#else 881 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 882#endif //SPARC 883#endif 884 obj->long_at_put(index, res); 885 break; 886 } 887 888 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 889 case T_INT: case T_FLOAT: // 4 bytes. 890 assert(value->type() == T_INT, "Agreement."); 891 val = value->get_int(); 892 obj->int_at_put(index, (jint)*((jint*)&val)); 893 break; 894 895 case T_SHORT: case T_CHAR: // 2 bytes 896 assert(value->type() == T_INT, "Agreement."); 897 val = value->get_int(); 898 obj->short_at_put(index, (jshort)*((jint*)&val)); 899 break; 900 901 case T_BOOLEAN: case T_BYTE: // 1 byte 902 assert(value->type() == T_INT, "Agreement."); 903 val = value->get_int(); 904 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 905 break; 906 907 default: 908 ShouldNotReachHere(); 909 } 910 index++; 911 } 912} 913 914 915// restore fields of an eliminated object array 916void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 917 for (int i = 0; i < sv->field_size(); i++) { 918 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 919 assert(value->type() == T_OBJECT, "object element expected"); 920 obj->obj_at_put(i, value->get_obj()()); 921 } 922} 923 924 925// restore fields of all eliminated objects and arrays 926void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 927 for (int i = 0; i < objects->length(); i++) { 928 ObjectValue* sv = (ObjectValue*) objects->at(i); 929 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 930 Handle obj = sv->value(); 931 assert(obj.not_null(), "reallocation was missed"); 932 933 if (k->oop_is_instance()) { 934 InstanceKlass* ik = InstanceKlass::cast(k()); 935 FieldReassigner reassign(fr, reg_map, sv, obj()); 936 ik->do_nonstatic_fields(&reassign); 937 } else if (k->oop_is_typeArray()) { 938 TypeArrayKlass* ak = TypeArrayKlass::cast(k()); 939 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 940 } else if (k->oop_is_objArray()) { 941 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 942 } 943 } 944} 945 946 947// relock objects for which synchronization was eliminated 948void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 949 for (int i = 0; i < monitors->length(); i++) { 950 MonitorInfo* mon_info = monitors->at(i); 951 if (mon_info->eliminated()) { 952 assert(mon_info->owner() != NULL, "reallocation was missed"); 953 Handle obj = Handle(mon_info->owner()); 954 markOop mark = obj->mark(); 955 if (UseBiasedLocking && mark->has_bias_pattern()) { 956 // New allocated objects may have the mark set to anonymously biased. 957 // Also the deoptimized method may called methods with synchronization 958 // where the thread-local object is bias locked to the current thread. 959 assert(mark->is_biased_anonymously() || 960 mark->biased_locker() == thread, "should be locked to current thread"); 961 // Reset mark word to unbiased prototype. 962 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 963 obj->set_mark(unbiased_prototype); 964 } 965 BasicLock* lock = mon_info->lock(); 966 ObjectSynchronizer::slow_enter(obj, lock, thread); 967 } 968 assert(mon_info->owner()->is_locked(), "object must be locked now"); 969 } 970} 971 972 973#ifndef PRODUCT 974// print information about reallocated objects 975void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 976 fieldDescriptor fd; 977 978 for (int i = 0; i < objects->length(); i++) { 979 ObjectValue* sv = (ObjectValue*) objects->at(i); 980 KlassHandle k(java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()())); 981 Handle obj = sv->value(); 982 983 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 984 k->print_value(); 985 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 986 tty->cr(); 987 988 if (Verbose) { 989 k->oop_print_on(obj(), tty); 990 } 991 } 992} 993#endif 994#endif // COMPILER2 995 996vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 997 Events::log(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, fr.pc(), fr.sp()); 998 999#ifndef PRODUCT 1000 if (TraceDeoptimization) { 1001 ttyLocker ttyl; 1002 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 1003 fr.print_on(tty); 1004 tty->print_cr(" Virtual frames (innermost first):"); 1005 for (int index = 0; index < chunk->length(); index++) { 1006 compiledVFrame* vf = chunk->at(index); 1007 tty->print(" %2d - ", index); 1008 vf->print_value(); 1009 int bci = chunk->at(index)->raw_bci(); 1010 const char* code_name; 1011 if (bci == SynchronizationEntryBCI) { 1012 code_name = "sync entry"; 1013 } else { 1014 Bytecodes::Code code = vf->method()->code_at(bci); 1015 code_name = Bytecodes::name(code); 1016 } 1017 tty->print(" - %s", code_name); 1018 tty->print_cr(" @ bci %d ", bci); 1019 if (Verbose) { 1020 vf->print(); 1021 tty->cr(); 1022 } 1023 } 1024 } 1025#endif 1026 1027 // Register map for next frame (used for stack crawl). We capture 1028 // the state of the deopt'ing frame's caller. Thus if we need to 1029 // stuff a C2I adapter we can properly fill in the callee-save 1030 // register locations. 1031 frame caller = fr.sender(reg_map); 1032 int frame_size = caller.sp() - fr.sp(); 1033 1034 frame sender = caller; 1035 1036 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1037 // the vframeArray containing the unpacking information is allocated in the C heap. 1038 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1039 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1040 1041 // Compare the vframeArray to the collected vframes 1042 assert(array->structural_compare(thread, chunk), "just checking"); 1043 1044#ifndef PRODUCT 1045 if (TraceDeoptimization) { 1046 ttyLocker ttyl; 1047 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1048 } 1049#endif // PRODUCT 1050 1051 return array; 1052} 1053 1054 1055static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1056 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1057 for (int i = 0; i < monitors->length(); i++) { 1058 MonitorInfo* mon_info = monitors->at(i); 1059 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1060 objects_to_revoke->append(Handle(mon_info->owner())); 1061 } 1062 } 1063} 1064 1065 1066void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1067 if (!UseBiasedLocking) { 1068 return; 1069 } 1070 1071 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1072 1073 // Unfortunately we don't have a RegisterMap available in most of 1074 // the places we want to call this routine so we need to walk the 1075 // stack again to update the register map. 1076 if (map == NULL || !map->update_map()) { 1077 StackFrameStream sfs(thread, true); 1078 bool found = false; 1079 while (!found && !sfs.is_done()) { 1080 frame* cur = sfs.current(); 1081 sfs.next(); 1082 found = cur->id() == fr.id(); 1083 } 1084 assert(found, "frame to be deoptimized not found on target thread's stack"); 1085 map = sfs.register_map(); 1086 } 1087 1088 vframe* vf = vframe::new_vframe(&fr, map, thread); 1089 compiledVFrame* cvf = compiledVFrame::cast(vf); 1090 // Revoke monitors' biases in all scopes 1091 while (!cvf->is_top()) { 1092 collect_monitors(cvf, objects_to_revoke); 1093 cvf = compiledVFrame::cast(cvf->sender()); 1094 } 1095 collect_monitors(cvf, objects_to_revoke); 1096 1097 if (SafepointSynchronize::is_at_safepoint()) { 1098 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1099 } else { 1100 BiasedLocking::revoke(objects_to_revoke); 1101 } 1102} 1103 1104 1105void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1106 if (!UseBiasedLocking) { 1107 return; 1108 } 1109 1110 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1111 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1112 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1113 if (jt->has_last_Java_frame()) { 1114 StackFrameStream sfs(jt, true); 1115 while (!sfs.is_done()) { 1116 frame* cur = sfs.current(); 1117 if (cb->contains(cur->pc())) { 1118 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1119 compiledVFrame* cvf = compiledVFrame::cast(vf); 1120 // Revoke monitors' biases in all scopes 1121 while (!cvf->is_top()) { 1122 collect_monitors(cvf, objects_to_revoke); 1123 cvf = compiledVFrame::cast(cvf->sender()); 1124 } 1125 collect_monitors(cvf, objects_to_revoke); 1126 } 1127 sfs.next(); 1128 } 1129 } 1130 } 1131 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1132} 1133 1134 1135void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1136 assert(fr.can_be_deoptimized(), "checking frame type"); 1137 1138 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1139 1140 // Patch the nmethod so that when execution returns to it we will 1141 // deopt the execution state and return to the interpreter. 1142 fr.deoptimize(thread); 1143} 1144 1145void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1146 // Deoptimize only if the frame comes from compile code. 1147 // Do not deoptimize the frame which is already patched 1148 // during the execution of the loops below. 1149 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1150 return; 1151 } 1152 ResourceMark rm; 1153 DeoptimizationMarker dm; 1154 if (UseBiasedLocking) { 1155 revoke_biases_of_monitors(thread, fr, map); 1156 } 1157 deoptimize_single_frame(thread, fr); 1158 1159} 1160 1161 1162void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1163 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1164 "can only deoptimize other thread at a safepoint"); 1165 // Compute frame and register map based on thread and sp. 1166 RegisterMap reg_map(thread, UseBiasedLocking); 1167 frame fr = thread->last_frame(); 1168 while (fr.id() != id) { 1169 fr = fr.sender(®_map); 1170 } 1171 deoptimize(thread, fr, ®_map); 1172} 1173 1174 1175void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1176 if (thread == Thread::current()) { 1177 Deoptimization::deoptimize_frame_internal(thread, id); 1178 } else { 1179 VM_DeoptimizeFrame deopt(thread, id); 1180 VMThread::execute(&deopt); 1181 } 1182} 1183 1184 1185// JVMTI PopFrame support 1186JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1187{ 1188 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1189} 1190JRT_END 1191 1192 1193#if defined(COMPILER2) || defined(SHARK) 1194void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1195 // in case of an unresolved klass entry, load the class. 1196 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1197 Klass* tk = constant_pool->klass_at(index, CHECK); 1198 return; 1199 } 1200 1201 if (!constant_pool->tag_at(index).is_symbol()) return; 1202 1203 Handle class_loader (THREAD, constant_pool->pool_holder()->class_loader()); 1204 Symbol* symbol = constant_pool->symbol_at(index); 1205 1206 // class name? 1207 if (symbol->byte_at(0) != '(') { 1208 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1209 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1210 return; 1211 } 1212 1213 // then it must be a signature! 1214 ResourceMark rm(THREAD); 1215 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1216 if (ss.is_object()) { 1217 Symbol* class_name = ss.as_symbol(CHECK); 1218 Handle protection_domain (THREAD, constant_pool->pool_holder()->protection_domain()); 1219 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1220 } 1221 } 1222} 1223 1224 1225void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1226 EXCEPTION_MARK; 1227 load_class_by_index(constant_pool, index, THREAD); 1228 if (HAS_PENDING_EXCEPTION) { 1229 // Exception happened during classloading. We ignore the exception here, since it 1230 // is going to be rethrown since the current activation is going to be deoptimzied and 1231 // the interpreter will re-execute the bytecode. 1232 CLEAR_PENDING_EXCEPTION; 1233 } 1234} 1235 1236JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1237 HandleMark hm; 1238 1239 // uncommon_trap() is called at the beginning of the uncommon trap 1240 // handler. Note this fact before we start generating temporary frames 1241 // that can confuse an asynchronous stack walker. This counter is 1242 // decremented at the end of unpack_frames(). 1243 thread->inc_in_deopt_handler(); 1244 1245 // We need to update the map if we have biased locking. 1246 RegisterMap reg_map(thread, UseBiasedLocking); 1247 frame stub_frame = thread->last_frame(); 1248 frame fr = stub_frame.sender(®_map); 1249 // Make sure the calling nmethod is not getting deoptimized and removed 1250 // before we are done with it. 1251 nmethodLocker nl(fr.pc()); 1252 1253 // Log a message 1254 Events::log(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT, 1255 trap_request, fr.pc()); 1256 1257 { 1258 ResourceMark rm; 1259 1260 // Revoke biases of any monitors in the frame to ensure we can migrate them 1261 revoke_biases_of_monitors(thread, fr, ®_map); 1262 1263 DeoptReason reason = trap_request_reason(trap_request); 1264 DeoptAction action = trap_request_action(trap_request); 1265 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1266 1267 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1268 compiledVFrame* cvf = compiledVFrame::cast(vf); 1269 1270 nmethod* nm = cvf->code(); 1271 1272 ScopeDesc* trap_scope = cvf->scope(); 1273 methodHandle trap_method = trap_scope->method(); 1274 int trap_bci = trap_scope->bci(); 1275 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1276 1277 // Record this event in the histogram. 1278 gather_statistics(reason, action, trap_bc); 1279 1280 // Ensure that we can record deopt. history: 1281 bool create_if_missing = ProfileTraps; 1282 1283 MethodData* trap_mdo = 1284 get_method_data(thread, trap_method, create_if_missing); 1285 1286 // Log a message 1287 Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d", 1288 trap_reason_name(reason), trap_action_name(action), fr.pc(), 1289 trap_method->name_and_sig_as_C_string(), trap_bci); 1290 1291 // Print a bunch of diagnostics, if requested. 1292 if (TraceDeoptimization || LogCompilation) { 1293 ResourceMark rm; 1294 ttyLocker ttyl; 1295 char buf[100]; 1296 if (xtty != NULL) { 1297 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1298 os::current_thread_id(), 1299 format_trap_request(buf, sizeof(buf), trap_request)); 1300 nm->log_identity(xtty); 1301 } 1302 Symbol* class_name = NULL; 1303 bool unresolved = false; 1304 if (unloaded_class_index >= 0) { 1305 constantPoolHandle constants (THREAD, trap_method->constants()); 1306 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1307 class_name = constants->klass_name_at(unloaded_class_index); 1308 unresolved = true; 1309 if (xtty != NULL) 1310 xtty->print(" unresolved='1'"); 1311 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1312 class_name = constants->symbol_at(unloaded_class_index); 1313 } 1314 if (xtty != NULL) 1315 xtty->name(class_name); 1316 } 1317 if (xtty != NULL && trap_mdo != NULL) { 1318 // Dump the relevant MDO state. 1319 // This is the deopt count for the current reason, any previous 1320 // reasons or recompiles seen at this point. 1321 int dcnt = trap_mdo->trap_count(reason); 1322 if (dcnt != 0) 1323 xtty->print(" count='%d'", dcnt); 1324 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1325 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1326 if (dos != 0) { 1327 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1328 if (trap_state_is_recompiled(dos)) { 1329 int recnt2 = trap_mdo->overflow_recompile_count(); 1330 if (recnt2 != 0) 1331 xtty->print(" recompiles2='%d'", recnt2); 1332 } 1333 } 1334 } 1335 if (xtty != NULL) { 1336 xtty->stamp(); 1337 xtty->end_head(); 1338 } 1339 if (TraceDeoptimization) { // make noise on the tty 1340 tty->print("Uncommon trap occurred in"); 1341 nm->method()->print_short_name(tty); 1342 tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d", 1343 fr.pc(), 1344 os::current_thread_id(), 1345 trap_reason_name(reason), 1346 trap_action_name(action), 1347 unloaded_class_index); 1348 if (class_name != NULL) { 1349 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1350 class_name->print_symbol_on(tty); 1351 } 1352 tty->cr(); 1353 } 1354 if (xtty != NULL) { 1355 // Log the precise location of the trap. 1356 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1357 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1358 xtty->method(sd->method()); 1359 xtty->end_elem(); 1360 if (sd->is_top()) break; 1361 } 1362 xtty->tail("uncommon_trap"); 1363 } 1364 } 1365 // (End diagnostic printout.) 1366 1367 // Load class if necessary 1368 if (unloaded_class_index >= 0) { 1369 constantPoolHandle constants(THREAD, trap_method->constants()); 1370 load_class_by_index(constants, unloaded_class_index); 1371 } 1372 1373 // Flush the nmethod if necessary and desirable. 1374 // 1375 // We need to avoid situations where we are re-flushing the nmethod 1376 // because of a hot deoptimization site. Repeated flushes at the same 1377 // point need to be detected by the compiler and avoided. If the compiler 1378 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1379 // module must take measures to avoid an infinite cycle of recompilation 1380 // and deoptimization. There are several such measures: 1381 // 1382 // 1. If a recompilation is ordered a second time at some site X 1383 // and for the same reason R, the action is adjusted to 'reinterpret', 1384 // to give the interpreter time to exercise the method more thoroughly. 1385 // If this happens, the method's overflow_recompile_count is incremented. 1386 // 1387 // 2. If the compiler fails to reduce the deoptimization rate, then 1388 // the method's overflow_recompile_count will begin to exceed the set 1389 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1390 // is adjusted to 'make_not_compilable', and the method is abandoned 1391 // to the interpreter. This is a performance hit for hot methods, 1392 // but is better than a disastrous infinite cycle of recompilations. 1393 // (Actually, only the method containing the site X is abandoned.) 1394 // 1395 // 3. In parallel with the previous measures, if the total number of 1396 // recompilations of a method exceeds the much larger set limit 1397 // PerMethodRecompilationCutoff, the method is abandoned. 1398 // This should only happen if the method is very large and has 1399 // many "lukewarm" deoptimizations. The code which enforces this 1400 // limit is elsewhere (class nmethod, class Method). 1401 // 1402 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1403 // to recompile at each bytecode independently of the per-BCI cutoff. 1404 // 1405 // The decision to update code is up to the compiler, and is encoded 1406 // in the Action_xxx code. If the compiler requests Action_none 1407 // no trap state is changed, no compiled code is changed, and the 1408 // computation suffers along in the interpreter. 1409 // 1410 // The other action codes specify various tactics for decompilation 1411 // and recompilation. Action_maybe_recompile is the loosest, and 1412 // allows the compiled code to stay around until enough traps are seen, 1413 // and until the compiler gets around to recompiling the trapping method. 1414 // 1415 // The other actions cause immediate removal of the present code. 1416 1417 bool update_trap_state = true; 1418 bool make_not_entrant = false; 1419 bool make_not_compilable = false; 1420 bool reprofile = false; 1421 switch (action) { 1422 case Action_none: 1423 // Keep the old code. 1424 update_trap_state = false; 1425 break; 1426 case Action_maybe_recompile: 1427 // Do not need to invalidate the present code, but we can 1428 // initiate another 1429 // Start compiler without (necessarily) invalidating the nmethod. 1430 // The system will tolerate the old code, but new code should be 1431 // generated when possible. 1432 break; 1433 case Action_reinterpret: 1434 // Go back into the interpreter for a while, and then consider 1435 // recompiling form scratch. 1436 make_not_entrant = true; 1437 // Reset invocation counter for outer most method. 1438 // This will allow the interpreter to exercise the bytecodes 1439 // for a while before recompiling. 1440 // By contrast, Action_make_not_entrant is immediate. 1441 // 1442 // Note that the compiler will track null_check, null_assert, 1443 // range_check, and class_check events and log them as if they 1444 // had been traps taken from compiled code. This will update 1445 // the MDO trap history so that the next compilation will 1446 // properly detect hot trap sites. 1447 reprofile = true; 1448 break; 1449 case Action_make_not_entrant: 1450 // Request immediate recompilation, and get rid of the old code. 1451 // Make them not entrant, so next time they are called they get 1452 // recompiled. Unloaded classes are loaded now so recompile before next 1453 // time they are called. Same for uninitialized. The interpreter will 1454 // link the missing class, if any. 1455 make_not_entrant = true; 1456 break; 1457 case Action_make_not_compilable: 1458 // Give up on compiling this method at all. 1459 make_not_entrant = true; 1460 make_not_compilable = true; 1461 break; 1462 default: 1463 ShouldNotReachHere(); 1464 } 1465 1466 // Setting +ProfileTraps fixes the following, on all platforms: 1467 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1468 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1469 // recompile relies on a MethodData* to record heroic opt failures. 1470 1471 // Whether the interpreter is producing MDO data or not, we also need 1472 // to use the MDO to detect hot deoptimization points and control 1473 // aggressive optimization. 1474 bool inc_recompile_count = false; 1475 ProfileData* pdata = NULL; 1476 if (ProfileTraps && update_trap_state && trap_mdo != NULL) { 1477 assert(trap_mdo == get_method_data(thread, trap_method, false), "sanity"); 1478 uint this_trap_count = 0; 1479 bool maybe_prior_trap = false; 1480 bool maybe_prior_recompile = false; 1481 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1482 //outputs: 1483 this_trap_count, 1484 maybe_prior_trap, 1485 maybe_prior_recompile); 1486 // Because the interpreter also counts null, div0, range, and class 1487 // checks, these traps from compiled code are double-counted. 1488 // This is harmless; it just means that the PerXTrapLimit values 1489 // are in effect a little smaller than they look. 1490 1491 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1492 if (per_bc_reason != Reason_none) { 1493 // Now take action based on the partially known per-BCI history. 1494 if (maybe_prior_trap 1495 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1496 // If there are too many traps at this BCI, force a recompile. 1497 // This will allow the compiler to see the limit overflow, and 1498 // take corrective action, if possible. The compiler generally 1499 // does not use the exact PerBytecodeTrapLimit value, but instead 1500 // changes its tactics if it sees any traps at all. This provides 1501 // a little hysteresis, delaying a recompile until a trap happens 1502 // several times. 1503 // 1504 // Actually, since there is only one bit of counter per BCI, 1505 // the possible per-BCI counts are {0,1,(per-method count)}. 1506 // This produces accurate results if in fact there is only 1507 // one hot trap site, but begins to get fuzzy if there are 1508 // many sites. For example, if there are ten sites each 1509 // trapping two or more times, they each get the blame for 1510 // all of their traps. 1511 make_not_entrant = true; 1512 } 1513 1514 // Detect repeated recompilation at the same BCI, and enforce a limit. 1515 if (make_not_entrant && maybe_prior_recompile) { 1516 // More than one recompile at this point. 1517 inc_recompile_count = maybe_prior_trap; 1518 } 1519 } else { 1520 // For reasons which are not recorded per-bytecode, we simply 1521 // force recompiles unconditionally. 1522 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1523 make_not_entrant = true; 1524 } 1525 1526 // Go back to the compiler if there are too many traps in this method. 1527 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1528 // If there are too many traps in this method, force a recompile. 1529 // This will allow the compiler to see the limit overflow, and 1530 // take corrective action, if possible. 1531 // (This condition is an unlikely backstop only, because the 1532 // PerBytecodeTrapLimit is more likely to take effect first, 1533 // if it is applicable.) 1534 make_not_entrant = true; 1535 } 1536 1537 // Here's more hysteresis: If there has been a recompile at 1538 // this trap point already, run the method in the interpreter 1539 // for a while to exercise it more thoroughly. 1540 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1541 reprofile = true; 1542 } 1543 1544 } 1545 1546 // Take requested actions on the method: 1547 1548 // Recompile 1549 if (make_not_entrant) { 1550 if (!nm->make_not_entrant()) { 1551 return; // the call did not change nmethod's state 1552 } 1553 1554 if (pdata != NULL) { 1555 // Record the recompilation event, if any. 1556 int tstate0 = pdata->trap_state(); 1557 int tstate1 = trap_state_set_recompiled(tstate0, true); 1558 if (tstate1 != tstate0) 1559 pdata->set_trap_state(tstate1); 1560 } 1561 } 1562 1563 if (inc_recompile_count) { 1564 trap_mdo->inc_overflow_recompile_count(); 1565 if ((uint)trap_mdo->overflow_recompile_count() > 1566 (uint)PerBytecodeRecompilationCutoff) { 1567 // Give up on the method containing the bad BCI. 1568 if (trap_method() == nm->method()) { 1569 make_not_compilable = true; 1570 } else { 1571 trap_method->set_not_compilable(CompLevel_full_optimization, true, "overflow_recompile_count > PerBytecodeRecompilationCutoff"); 1572 // But give grace to the enclosing nm->method(). 1573 } 1574 } 1575 } 1576 1577 // Reprofile 1578 if (reprofile) { 1579 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1580 } 1581 1582 // Give up compiling 1583 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1584 assert(make_not_entrant, "consistent"); 1585 nm->method()->set_not_compilable(CompLevel_full_optimization); 1586 } 1587 1588 } // Free marked resources 1589 1590} 1591JRT_END 1592 1593MethodData* 1594Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1595 bool create_if_missing) { 1596 Thread* THREAD = thread; 1597 MethodData* mdo = m()->method_data(); 1598 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1599 // Build an MDO. Ignore errors like OutOfMemory; 1600 // that simply means we won't have an MDO to update. 1601 Method::build_interpreter_method_data(m, THREAD); 1602 if (HAS_PENDING_EXCEPTION) { 1603 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1604 CLEAR_PENDING_EXCEPTION; 1605 } 1606 mdo = m()->method_data(); 1607 } 1608 return mdo; 1609} 1610 1611ProfileData* 1612Deoptimization::query_update_method_data(MethodData* trap_mdo, 1613 int trap_bci, 1614 Deoptimization::DeoptReason reason, 1615 //outputs: 1616 uint& ret_this_trap_count, 1617 bool& ret_maybe_prior_trap, 1618 bool& ret_maybe_prior_recompile) { 1619 uint prior_trap_count = trap_mdo->trap_count(reason); 1620 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1621 1622 // If the runtime cannot find a place to store trap history, 1623 // it is estimated based on the general condition of the method. 1624 // If the method has ever been recompiled, or has ever incurred 1625 // a trap with the present reason , then this BCI is assumed 1626 // (pessimistically) to be the culprit. 1627 bool maybe_prior_trap = (prior_trap_count != 0); 1628 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1629 ProfileData* pdata = NULL; 1630 1631 1632 // For reasons which are recorded per bytecode, we check per-BCI data. 1633 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1634 if (per_bc_reason != Reason_none) { 1635 // Find the profile data for this BCI. If there isn't one, 1636 // try to allocate one from the MDO's set of spares. 1637 // This will let us detect a repeated trap at this point. 1638 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1639 1640 if (pdata != NULL) { 1641 // Query the trap state of this profile datum. 1642 int tstate0 = pdata->trap_state(); 1643 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1644 maybe_prior_trap = false; 1645 if (!trap_state_is_recompiled(tstate0)) 1646 maybe_prior_recompile = false; 1647 1648 // Update the trap state of this profile datum. 1649 int tstate1 = tstate0; 1650 // Record the reason. 1651 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1652 // Store the updated state on the MDO, for next time. 1653 if (tstate1 != tstate0) 1654 pdata->set_trap_state(tstate1); 1655 } else { 1656 if (LogCompilation && xtty != NULL) { 1657 ttyLocker ttyl; 1658 // Missing MDP? Leave a small complaint in the log. 1659 xtty->elem("missing_mdp bci='%d'", trap_bci); 1660 } 1661 } 1662 } 1663 1664 // Return results: 1665 ret_this_trap_count = this_trap_count; 1666 ret_maybe_prior_trap = maybe_prior_trap; 1667 ret_maybe_prior_recompile = maybe_prior_recompile; 1668 return pdata; 1669} 1670 1671void 1672Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1673 ResourceMark rm; 1674 // Ignored outputs: 1675 uint ignore_this_trap_count; 1676 bool ignore_maybe_prior_trap; 1677 bool ignore_maybe_prior_recompile; 1678 query_update_method_data(trap_mdo, trap_bci, 1679 (DeoptReason)reason, 1680 ignore_this_trap_count, 1681 ignore_maybe_prior_trap, 1682 ignore_maybe_prior_recompile); 1683} 1684 1685Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1686 1687 // Still in Java no safepoints 1688 { 1689 // This enters VM and may safepoint 1690 uncommon_trap_inner(thread, trap_request); 1691 } 1692 return fetch_unroll_info_helper(thread); 1693} 1694 1695// Local derived constants. 1696// Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1697const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1698const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1699 1700//---------------------------trap_state_reason--------------------------------- 1701Deoptimization::DeoptReason 1702Deoptimization::trap_state_reason(int trap_state) { 1703 // This assert provides the link between the width of DataLayout::trap_bits 1704 // and the encoding of "recorded" reasons. It ensures there are enough 1705 // bits to store all needed reasons in the per-BCI MDO profile. 1706 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1707 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1708 trap_state -= recompile_bit; 1709 if (trap_state == DS_REASON_MASK) { 1710 return Reason_many; 1711 } else { 1712 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1713 return (DeoptReason)trap_state; 1714 } 1715} 1716//-------------------------trap_state_has_reason------------------------------- 1717int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1718 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1719 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1720 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1721 trap_state -= recompile_bit; 1722 if (trap_state == DS_REASON_MASK) { 1723 return -1; // true, unspecifically (bottom of state lattice) 1724 } else if (trap_state == reason) { 1725 return 1; // true, definitely 1726 } else if (trap_state == 0) { 1727 return 0; // false, definitely (top of state lattice) 1728 } else { 1729 return 0; // false, definitely 1730 } 1731} 1732//-------------------------trap_state_add_reason------------------------------- 1733int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1734 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1735 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1736 trap_state -= recompile_bit; 1737 if (trap_state == DS_REASON_MASK) { 1738 return trap_state + recompile_bit; // already at state lattice bottom 1739 } else if (trap_state == reason) { 1740 return trap_state + recompile_bit; // the condition is already true 1741 } else if (trap_state == 0) { 1742 return reason + recompile_bit; // no condition has yet been true 1743 } else { 1744 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1745 } 1746} 1747//-----------------------trap_state_is_recompiled------------------------------ 1748bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1749 return (trap_state & DS_RECOMPILE_BIT) != 0; 1750} 1751//-----------------------trap_state_set_recompiled----------------------------- 1752int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1753 if (z) return trap_state | DS_RECOMPILE_BIT; 1754 else return trap_state & ~DS_RECOMPILE_BIT; 1755} 1756//---------------------------format_trap_state--------------------------------- 1757// This is used for debugging and diagnostics, including hotspot.log output. 1758const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1759 int trap_state) { 1760 DeoptReason reason = trap_state_reason(trap_state); 1761 bool recomp_flag = trap_state_is_recompiled(trap_state); 1762 // Re-encode the state from its decoded components. 1763 int decoded_state = 0; 1764 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1765 decoded_state = trap_state_add_reason(decoded_state, reason); 1766 if (recomp_flag) 1767 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1768 // If the state re-encodes properly, format it symbolically. 1769 // Because this routine is used for debugging and diagnostics, 1770 // be robust even if the state is a strange value. 1771 size_t len; 1772 if (decoded_state != trap_state) { 1773 // Random buggy state that doesn't decode?? 1774 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1775 } else { 1776 len = jio_snprintf(buf, buflen, "%s%s", 1777 trap_reason_name(reason), 1778 recomp_flag ? " recompiled" : ""); 1779 } 1780 if (len >= buflen) 1781 buf[buflen-1] = '\0'; 1782 return buf; 1783} 1784 1785 1786//--------------------------------statics-------------------------------------- 1787Deoptimization::DeoptAction Deoptimization::_unloaded_action 1788 = Deoptimization::Action_reinterpret; 1789const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1790 // Note: Keep this in sync. with enum DeoptReason. 1791 "none", 1792 "null_check", 1793 "null_assert", 1794 "range_check", 1795 "class_check", 1796 "array_check", 1797 "intrinsic", 1798 "bimorphic", 1799 "unloaded", 1800 "uninitialized", 1801 "unreached", 1802 "unhandled", 1803 "constraint", 1804 "div0_check", 1805 "age", 1806 "predicate", 1807 "loop_limit_check" 1808}; 1809const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1810 // Note: Keep this in sync. with enum DeoptAction. 1811 "none", 1812 "maybe_recompile", 1813 "reinterpret", 1814 "make_not_entrant", 1815 "make_not_compilable" 1816}; 1817 1818const char* Deoptimization::trap_reason_name(int reason) { 1819 if (reason == Reason_many) return "many"; 1820 if ((uint)reason < Reason_LIMIT) 1821 return _trap_reason_name[reason]; 1822 static char buf[20]; 1823 sprintf(buf, "reason%d", reason); 1824 return buf; 1825} 1826const char* Deoptimization::trap_action_name(int action) { 1827 if ((uint)action < Action_LIMIT) 1828 return _trap_action_name[action]; 1829 static char buf[20]; 1830 sprintf(buf, "action%d", action); 1831 return buf; 1832} 1833 1834// This is used for debugging and diagnostics, including hotspot.log output. 1835const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1836 int trap_request) { 1837 jint unloaded_class_index = trap_request_index(trap_request); 1838 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1839 const char* action = trap_action_name(trap_request_action(trap_request)); 1840 size_t len; 1841 if (unloaded_class_index < 0) { 1842 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1843 reason, action); 1844 } else { 1845 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1846 reason, action, unloaded_class_index); 1847 } 1848 if (len >= buflen) 1849 buf[buflen-1] = '\0'; 1850 return buf; 1851} 1852 1853juint Deoptimization::_deoptimization_hist 1854 [Deoptimization::Reason_LIMIT] 1855 [1 + Deoptimization::Action_LIMIT] 1856 [Deoptimization::BC_CASE_LIMIT] 1857 = {0}; 1858 1859enum { 1860 LSB_BITS = 8, 1861 LSB_MASK = right_n_bits(LSB_BITS) 1862}; 1863 1864void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1865 Bytecodes::Code bc) { 1866 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1867 assert(action >= 0 && action < Action_LIMIT, "oob"); 1868 _deoptimization_hist[Reason_none][0][0] += 1; // total 1869 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1870 juint* cases = _deoptimization_hist[reason][1+action]; 1871 juint* bc_counter_addr = NULL; 1872 juint bc_counter = 0; 1873 // Look for an unused counter, or an exact match to this BC. 1874 if (bc != Bytecodes::_illegal) { 1875 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1876 juint* counter_addr = &cases[bc_case]; 1877 juint counter = *counter_addr; 1878 if ((counter == 0 && bc_counter_addr == NULL) 1879 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1880 // this counter is either free or is already devoted to this BC 1881 bc_counter_addr = counter_addr; 1882 bc_counter = counter | bc; 1883 } 1884 } 1885 } 1886 if (bc_counter_addr == NULL) { 1887 // Overflow, or no given bytecode. 1888 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1889 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1890 } 1891 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1892} 1893 1894jint Deoptimization::total_deoptimization_count() { 1895 return _deoptimization_hist[Reason_none][0][0]; 1896} 1897 1898jint Deoptimization::deoptimization_count(DeoptReason reason) { 1899 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1900 return _deoptimization_hist[reason][0][0]; 1901} 1902 1903void Deoptimization::print_statistics() { 1904 juint total = total_deoptimization_count(); 1905 juint account = total; 1906 if (total != 0) { 1907 ttyLocker ttyl; 1908 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1909 tty->print_cr("Deoptimization traps recorded:"); 1910 #define PRINT_STAT_LINE(name, r) \ 1911 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1912 PRINT_STAT_LINE("total", total); 1913 // For each non-zero entry in the histogram, print the reason, 1914 // the action, and (if specifically known) the type of bytecode. 1915 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1916 for (int action = 0; action < Action_LIMIT; action++) { 1917 juint* cases = _deoptimization_hist[reason][1+action]; 1918 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1919 juint counter = cases[bc_case]; 1920 if (counter != 0) { 1921 char name[1*K]; 1922 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1923 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1924 bc = Bytecodes::_illegal; 1925 sprintf(name, "%s/%s/%s", 1926 trap_reason_name(reason), 1927 trap_action_name(action), 1928 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1929 juint r = counter >> LSB_BITS; 1930 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1931 account -= r; 1932 } 1933 } 1934 } 1935 } 1936 if (account != 0) { 1937 PRINT_STAT_LINE("unaccounted", account); 1938 } 1939 #undef PRINT_STAT_LINE 1940 if (xtty != NULL) xtty->tail("statistics"); 1941 } 1942} 1943#else // COMPILER2 || SHARK 1944 1945 1946// Stubs for C1 only system. 1947bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1948 return false; 1949} 1950 1951const char* Deoptimization::trap_reason_name(int reason) { 1952 return "unknown"; 1953} 1954 1955void Deoptimization::print_statistics() { 1956 // no output 1957} 1958 1959void 1960Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) { 1961 // no udpate 1962} 1963 1964int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1965 return 0; 1966} 1967 1968void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1969 Bytecodes::Code bc) { 1970 // no update 1971} 1972 1973const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1974 int trap_state) { 1975 jio_snprintf(buf, buflen, "#%d", trap_state); 1976 return buf; 1977} 1978 1979#endif // COMPILER2 || SHARK 1980