1 /* 2 * Copyright (c) 1997, 2011, 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/methodOop.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 number_of_frames, 94 intptr_t* frame_sizes, 95 address* frame_pcs, 96 BasicType return_type) { 97 _size_of_deoptimized_frame = size_of_deoptimized_frame; 98 _caller_adjustment = caller_adjustment; 99 _number_of_frames = number_of_frames; 100 _frame_sizes = frame_sizes; 101 _frame_pcs = frame_pcs; 102 _register_block = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2); 103 _return_type = return_type; 104 // PD (x86 only) 105 _counter_temp = 0; 106 _initial_fp = 0; 107 _unpack_kind = 0; 108 _sender_sp_temp = 0; 109 110 _total_frame_sizes = size_of_frames(); 111 } 112 113 114 Deoptimization::UnrollBlock::~UnrollBlock() { 115 FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes); 116 FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs); 117 FREE_C_HEAP_ARRAY(intptr_t, _register_block); 118 } 119 120 121 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const { 122 assert(register_number < RegisterMap::reg_count, "checking register number"); 123 return &_register_block[register_number * 2]; 124 } 125 126 127 128 int Deoptimization::UnrollBlock::size_of_frames() const { 129 // Acount first for the adjustment of the initial frame 130 int result = _caller_adjustment; 131 for (int index = 0; index < number_of_frames(); index++) { 132 result += frame_sizes()[index]; 133 } 134 return result; 135 } 136 137 138 void Deoptimization::UnrollBlock::print() { 139 ttyLocker ttyl; 140 tty->print_cr("UnrollBlock"); 141 tty->print_cr(" size_of_deoptimized_frame = %d", _size_of_deoptimized_frame); 142 tty->print( " frame_sizes: "); 143 for (int index = 0; index < number_of_frames(); index++) { 144 tty->print("%d ", frame_sizes()[index]); 145 } 146 tty->cr(); 147 } 148 149 150 // In order to make fetch_unroll_info work properly with escape 151 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and 152 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation 153 // of previously eliminated objects occurs in realloc_objects, which is 154 // called from the method fetch_unroll_info_helper below. 155 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread)) 156 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 157 // but makes the entry a little slower. There is however a little dance we have to 158 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 159 160 // fetch_unroll_info() is called at the beginning of the deoptimization 161 // handler. Note this fact before we start generating temporary frames 162 // that can confuse an asynchronous stack walker. This counter is 163 // decremented at the end of unpack_frames(). 164 thread->inc_in_deopt_handler(); 165 166 return fetch_unroll_info_helper(thread); 167 JRT_END 168 169 170 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap) 171 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread) { 172 173 // Note: there is a safepoint safety issue here. No matter whether we enter 174 // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once 175 // the vframeArray is created. 176 // 177 178 // Allocate our special deoptimization ResourceMark 179 DeoptResourceMark* dmark = new DeoptResourceMark(thread); 180 assert(thread->deopt_mark() == NULL, "Pending deopt!"); 181 thread->set_deopt_mark(dmark); 182 183 frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect 184 RegisterMap map(thread, true); 185 RegisterMap dummy_map(thread, false); 186 // Now get the deoptee with a valid map 187 frame deoptee = stub_frame.sender(&map); 188 // Set the deoptee nmethod 189 assert(thread->deopt_nmethod() == NULL, "Pending deopt!"); 190 thread->set_deopt_nmethod(deoptee.cb()->as_nmethod_or_null()); 191 192 // Create a growable array of VFrames where each VFrame represents an inlined 193 // Java frame. This storage is allocated with the usual system arena. 194 assert(deoptee.is_compiled_frame(), "Wrong frame type"); 195 GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10); 196 vframe* vf = vframe::new_vframe(&deoptee, &map, thread); 197 while (!vf->is_top()) { 198 assert(vf->is_compiled_frame(), "Wrong frame type"); 199 chunk->push(compiledVFrame::cast(vf)); 200 vf = vf->sender(); 201 } 202 assert(vf->is_compiled_frame(), "Wrong frame type"); 203 chunk->push(compiledVFrame::cast(vf)); 204 205 #ifdef COMPILER2 206 // Reallocate the non-escaping objects and restore their fields. Then 207 // relock objects if synchronization on them was eliminated. 208 if (DoEscapeAnalysis) { 209 if (EliminateAllocations) { 210 assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames"); 211 GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects(); 212 213 // The flag return_oop() indicates call sites which return oop 214 // in compiled code. Such sites include java method calls, 215 // runtime calls (for example, used to allocate new objects/arrays 216 // on slow code path) and any other calls generated in compiled code. 217 // It is not guaranteed that we can get such information here only 218 // by analyzing bytecode in deoptimized frames. This is why this flag 219 // is set during method compilation (see Compile::Process_OopMap_Node()). 220 bool save_oop_result = chunk->at(0)->scope()->return_oop(); 221 Handle return_value; 222 if (save_oop_result) { 223 // Reallocation may trigger GC. If deoptimization happened on return from 224 // call which returns oop we need to save it since it is not in oopmap. 225 oop result = deoptee.saved_oop_result(&map); 226 assert(result == NULL || result->is_oop(), "must be oop"); 227 return_value = Handle(thread, result); 228 assert(Universe::heap()->is_in_or_null(result), "must be heap pointer"); 229 if (TraceDeoptimization) { 230 tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, result, thread); 231 } 232 } 233 bool reallocated = false; 234 if (objects != NULL) { 235 JRT_BLOCK 236 reallocated = realloc_objects(thread, &deoptee, objects, THREAD); 237 JRT_END 238 } 239 if (reallocated) { 240 reassign_fields(&deoptee, &map, objects); 241 #ifndef PRODUCT 242 if (TraceDeoptimization) { 243 ttyLocker ttyl; 244 tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, thread); 245 print_objects(objects); 246 } 247 #endif 248 } 249 if (save_oop_result) { 250 // Restore result. 251 deoptee.set_saved_oop_result(&map, return_value()); 252 } 253 } 254 if (EliminateLocks) { 255 #ifndef PRODUCT 256 bool first = true; 257 #endif 258 for (int i = 0; i < chunk->length(); i++) { 259 compiledVFrame* cvf = chunk->at(i); 260 assert (cvf->scope() != NULL,"expect only compiled java frames"); 261 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 262 if (monitors->is_nonempty()) { 263 relock_objects(monitors, thread); 264 #ifndef PRODUCT 265 if (TraceDeoptimization) { 266 ttyLocker ttyl; 267 for (int j = 0; j < monitors->length(); j++) { 268 MonitorInfo* mi = monitors->at(j); 269 if (mi->eliminated()) { 270 if (first) { 271 first = false; 272 tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, thread); 273 } 274 tty->print_cr(" object <" INTPTR_FORMAT "> locked", mi->owner()); 275 } 276 } 277 } 278 #endif 279 } 280 } 281 } 282 } 283 #endif // COMPILER2 284 // Ensure that no safepoint is taken after pointers have been stored 285 // in fields of rematerialized objects. If a safepoint occurs from here on 286 // out the java state residing in the vframeArray will be missed. 287 No_Safepoint_Verifier no_safepoint; 288 289 vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk); 290 291 assert(thread->vframe_array_head() == NULL, "Pending deopt!");; 292 thread->set_vframe_array_head(array); 293 294 // Now that the vframeArray has been created if we have any deferred local writes 295 // added by jvmti then we can free up that structure as the data is now in the 296 // vframeArray 297 298 if (thread->deferred_locals() != NULL) { 299 GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals(); 300 int i = 0; 301 do { 302 // Because of inlining we could have multiple vframes for a single frame 303 // and several of the vframes could have deferred writes. Find them all. 304 if (list->at(i)->id() == array->original().id()) { 305 jvmtiDeferredLocalVariableSet* dlv = list->at(i); 306 list->remove_at(i); 307 // individual jvmtiDeferredLocalVariableSet are CHeapObj's 308 delete dlv; 309 } else { 310 i++; 311 } 312 } while ( i < list->length() ); 313 if (list->length() == 0) { 314 thread->set_deferred_locals(NULL); 315 // free the list and elements back to C heap. 316 delete list; 317 } 318 319 } 320 321 #ifndef SHARK 322 // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info. 323 CodeBlob* cb = stub_frame.cb(); 324 // Verify we have the right vframeArray 325 assert(cb->frame_size() >= 0, "Unexpected frame size"); 326 intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size(); 327 328 // If the deopt call site is a MethodHandle invoke call site we have 329 // to adjust the unpack_sp. 330 nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null(); 331 if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc())) 332 unpack_sp = deoptee.unextended_sp(); 333 334 #ifdef ASSERT 335 assert(cb->is_deoptimization_stub() || cb->is_uncommon_trap_stub(), "just checking"); 336 Events::log("fetch unroll sp " INTPTR_FORMAT, unpack_sp); 337 #endif 338 #else 339 intptr_t* unpack_sp = stub_frame.sender(&dummy_map).unextended_sp(); 340 #endif // !SHARK 341 342 // This is a guarantee instead of an assert because if vframe doesn't match 343 // we will unpack the wrong deoptimized frame and wind up in strange places 344 // where it will be very difficult to figure out what went wrong. Better 345 // to die an early death here than some very obscure death later when the 346 // trail is cold. 347 // Note: on ia64 this guarantee can be fooled by frames with no memory stack 348 // in that it will fail to detect a problem when there is one. This needs 349 // more work in tiger timeframe. 350 guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack"); 351 352 int number_of_frames = array->frames(); 353 354 // Compute the vframes' sizes. Note that frame_sizes[] entries are ordered from outermost to innermost 355 // virtual activation, which is the reverse of the elements in the vframes array. 356 intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames); 357 // +1 because we always have an interpreter return address for the final slot. 358 address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1); 359 int callee_parameters = 0; 360 int callee_locals = 0; 361 int popframe_extra_args = 0; 362 // Create an interpreter return address for the stub to use as its return 363 // address so the skeletal frames are perfectly walkable 364 frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0); 365 366 // PopFrame requires that the preserved incoming arguments from the recently-popped topmost 367 // activation be put back on the expression stack of the caller for reexecution 368 if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) { 369 popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words()); 370 } 371 372 // 373 // frame_sizes/frame_pcs[0] oldest frame (int or c2i) 374 // frame_sizes/frame_pcs[1] next oldest frame (int) 375 // frame_sizes/frame_pcs[n] youngest frame (int) 376 // 377 // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame 378 // owns the space for the return address to it's caller). Confusing ain't it. 379 // 380 // The vframe array can address vframes with indices running from 381 // 0.._frames-1. Index 0 is the youngest frame and _frame - 1 is the oldest (root) frame. 382 // When we create the skeletal frames we need the oldest frame to be in the zero slot 383 // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk. 384 // so things look a little strange in this loop. 385 // 386 for (int index = 0; index < array->frames(); index++ ) { 387 // frame[number_of_frames - 1 ] = on_stack_size(youngest) 388 // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest)) 389 // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest))) 390 frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters, 391 callee_locals, 392 index == 0, 393 popframe_extra_args); 394 // This pc doesn't have to be perfect just good enough to identify the frame 395 // as interpreted so the skeleton frame will be walkable 396 // The correct pc will be set when the skeleton frame is completely filled out 397 // The final pc we store in the loop is wrong and will be overwritten below 398 frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset; 399 400 callee_parameters = array->element(index)->method()->size_of_parameters(); 401 callee_locals = array->element(index)->method()->max_locals(); 402 popframe_extra_args = 0; 403 } 404 405 // Compute whether the root vframe returns a float or double value. 406 BasicType return_type; 407 { 408 HandleMark hm; 409 methodHandle method(thread, array->element(0)->method()); 410 Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci()); 411 return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL; 412 } 413 414 // Compute information for handling adapters and adjusting the frame size of the caller. 415 int caller_adjustment = 0; 416 417 // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized 418 // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather 419 // than simply use array->sender.pc(). This requires us to walk the current set of frames 420 // 421 frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame 422 deopt_sender = deopt_sender.sender(&dummy_map); // Now deoptee caller 423 424 // Compute the amount the oldest interpreter frame will have to adjust 425 // its caller's stack by. If the caller is a compiled frame then 426 // we pretend that the callee has no parameters so that the 427 // extension counts for the full amount of locals and not just 428 // locals-parms. This is because without a c2i adapter the parm 429 // area as created by the compiled frame will not be usable by 430 // the interpreter. (Depending on the calling convention there 431 // may not even be enough space). 432 433 // QQQ I'd rather see this pushed down into last_frame_adjust 434 // and have it take the sender (aka caller). 435 436 if (deopt_sender.is_compiled_frame()) { 437 caller_adjustment = last_frame_adjust(0, callee_locals); 438 } else if (callee_locals > callee_parameters) { 439 // The caller frame may need extending to accommodate 440 // non-parameter locals of the first unpacked interpreted frame. 441 // Compute that adjustment. 442 caller_adjustment = last_frame_adjust(callee_parameters, callee_locals); 443 } 444 445 446 // If the sender is deoptimized the we must retrieve the address of the handler 447 // since the frame will "magically" show the original pc before the deopt 448 // and we'd undo the deopt. 449 450 frame_pcs[0] = deopt_sender.raw_pc(); 451 452 #ifndef SHARK 453 assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc"); 454 #endif // SHARK 455 456 UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord, 457 caller_adjustment * BytesPerWord, 458 number_of_frames, 459 frame_sizes, 460 frame_pcs, 461 return_type); 462 #if defined(IA32) || defined(AMD64) 463 // We need a way to pass fp to the unpacking code so the skeletal frames 464 // come out correct. This is only needed for x86 because of c2 using ebp 465 // as an allocatable register. So this update is useless (and harmless) 466 // on the other platforms. It would be nice to do this in a different 467 // way but even the old style deoptimization had a problem with deriving 468 // this value. NEEDS_CLEANUP 469 // Note: now that c1 is using c2's deopt blob we must do this on all 470 // x86 based platforms 471 intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes()); 472 *fp_addr = array->sender().fp(); // was adapter_caller 473 #endif /* IA32 || AMD64 */ 474 475 if (array->frames() > 1) { 476 if (VerifyStack && TraceDeoptimization) { 477 tty->print_cr("Deoptimizing method containing inlining"); 478 } 479 } 480 481 array->set_unroll_block(info); 482 return info; 483 } 484 485 // Called to cleanup deoptimization data structures in normal case 486 // after unpacking to stack and when stack overflow error occurs 487 void Deoptimization::cleanup_deopt_info(JavaThread *thread, 488 vframeArray *array) { 489 490 // Get array if coming from exception 491 if (array == NULL) { 492 array = thread->vframe_array_head(); 493 } 494 thread->set_vframe_array_head(NULL); 495 496 // Free the previous UnrollBlock 497 vframeArray* old_array = thread->vframe_array_last(); 498 thread->set_vframe_array_last(array); 499 500 if (old_array != NULL) { 501 UnrollBlock* old_info = old_array->unroll_block(); 502 old_array->set_unroll_block(NULL); 503 delete old_info; 504 delete old_array; 505 } 506 507 // Deallocate any resource creating in this routine and any ResourceObjs allocated 508 // inside the vframeArray (StackValueCollections) 509 510 delete thread->deopt_mark(); 511 thread->set_deopt_mark(NULL); 512 thread->set_deopt_nmethod(NULL); 513 514 515 if (JvmtiExport::can_pop_frame()) { 516 #ifndef CC_INTERP 517 // Regardless of whether we entered this routine with the pending 518 // popframe condition bit set, we should always clear it now 519 thread->clear_popframe_condition(); 520 #else 521 // C++ interpeter will clear has_pending_popframe when it enters 522 // with method_resume. For deopt_resume2 we clear it now. 523 if (thread->popframe_forcing_deopt_reexecution()) 524 thread->clear_popframe_condition(); 525 #endif /* CC_INTERP */ 526 } 527 528 // unpack_frames() is called at the end of the deoptimization handler 529 // and (in C2) at the end of the uncommon trap handler. Note this fact 530 // so that an asynchronous stack walker can work again. This counter is 531 // incremented at the beginning of fetch_unroll_info() and (in C2) at 532 // the beginning of uncommon_trap(). 533 thread->dec_in_deopt_handler(); 534 } 535 536 537 // Return BasicType of value being returned 538 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)) 539 540 // We are already active int he special DeoptResourceMark any ResourceObj's we 541 // allocate will be freed at the end of the routine. 542 543 // It is actually ok to allocate handles in a leaf method. It causes no safepoints, 544 // but makes the entry a little slower. There is however a little dance we have to 545 // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro 546 ResetNoHandleMark rnhm; // No-op in release/product versions 547 HandleMark hm; 548 549 frame stub_frame = thread->last_frame(); 550 551 // Since the frame to unpack is the top frame of this thread, the vframe_array_head 552 // must point to the vframeArray for the unpack frame. 553 vframeArray* array = thread->vframe_array_head(); 554 555 #ifndef PRODUCT 556 if (TraceDeoptimization) { 557 tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d", thread, array, exec_mode); 558 } 559 #endif 560 561 UnrollBlock* info = array->unroll_block(); 562 563 // Unpack the interpreter frames and any adapter frame (c2 only) we might create. 564 array->unpack_to_stack(stub_frame, exec_mode); 565 566 BasicType bt = info->return_type(); 567 568 // If we have an exception pending, claim that the return type is an oop 569 // so the deopt_blob does not overwrite the exception_oop. 570 571 if (exec_mode == Unpack_exception) 572 bt = T_OBJECT; 573 574 // Cleanup thread deopt data 575 cleanup_deopt_info(thread, array); 576 577 #ifndef PRODUCT 578 if (VerifyStack) { 579 ResourceMark res_mark; 580 581 // Verify that the just-unpacked frames match the interpreter's 582 // notions of expression stack and locals 583 vframeArray* cur_array = thread->vframe_array_last(); 584 RegisterMap rm(thread, false); 585 rm.set_include_argument_oops(false); 586 bool is_top_frame = true; 587 int callee_size_of_parameters = 0; 588 int callee_max_locals = 0; 589 for (int i = 0; i < cur_array->frames(); i++) { 590 vframeArrayElement* el = cur_array->element(i); 591 frame* iframe = el->iframe(); 592 guarantee(iframe->is_interpreted_frame(), "Wrong frame type"); 593 594 // Get the oop map for this bci 595 InterpreterOopMap mask; 596 int cur_invoke_parameter_size = 0; 597 bool try_next_mask = false; 598 int next_mask_expression_stack_size = -1; 599 int top_frame_expression_stack_adjustment = 0; 600 methodHandle mh(thread, iframe->interpreter_frame_method()); 601 OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask); 602 BytecodeStream str(mh); 603 str.set_start(iframe->interpreter_frame_bci()); 604 int max_bci = mh->code_size(); 605 // Get to the next bytecode if possible 606 assert(str.bci() < max_bci, "bci in interpreter frame out of bounds"); 607 // Check to see if we can grab the number of outgoing arguments 608 // at an uncommon trap for an invoke (where the compiler 609 // generates debug info before the invoke has executed) 610 Bytecodes::Code cur_code = str.next(); 611 if (cur_code == Bytecodes::_invokevirtual || 612 cur_code == Bytecodes::_invokespecial || 613 cur_code == Bytecodes::_invokestatic || 614 cur_code == Bytecodes::_invokeinterface) { 615 Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci()); 616 Symbol* signature = invoke.signature(); 617 ArgumentSizeComputer asc(signature); 618 cur_invoke_parameter_size = asc.size(); 619 if (cur_code != Bytecodes::_invokestatic) { 620 // Add in receiver 621 ++cur_invoke_parameter_size; 622 } 623 } 624 if (str.bci() < max_bci) { 625 Bytecodes::Code bc = str.next(); 626 if (bc >= 0) { 627 // The interpreter oop map generator reports results before 628 // the current bytecode has executed except in the case of 629 // calls. It seems to be hard to tell whether the compiler 630 // has emitted debug information matching the "state before" 631 // a given bytecode or the state after, so we try both 632 switch (cur_code) { 633 case Bytecodes::_invokevirtual: 634 case Bytecodes::_invokespecial: 635 case Bytecodes::_invokestatic: 636 case Bytecodes::_invokeinterface: 637 case Bytecodes::_athrow: 638 break; 639 default: { 640 InterpreterOopMap next_mask; 641 OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask); 642 next_mask_expression_stack_size = next_mask.expression_stack_size(); 643 // Need to subtract off the size of the result type of 644 // the bytecode because this is not described in the 645 // debug info but returned to the interpreter in the TOS 646 // caching register 647 BasicType bytecode_result_type = Bytecodes::result_type(cur_code); 648 if (bytecode_result_type != T_ILLEGAL) { 649 top_frame_expression_stack_adjustment = type2size[bytecode_result_type]; 650 } 651 assert(top_frame_expression_stack_adjustment >= 0, ""); 652 try_next_mask = true; 653 break; 654 } 655 } 656 } 657 } 658 659 // Verify stack depth and oops in frame 660 // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc) 661 if (!( 662 /* SPARC */ 663 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) || 664 /* x86 */ 665 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) || 666 (try_next_mask && 667 (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size - 668 top_frame_expression_stack_adjustment))) || 669 (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) || 670 (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute) && 671 (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size)) 672 )) { 673 ttyLocker ttyl; 674 675 // Print out some information that will help us debug the problem 676 tty->print_cr("Wrong number of expression stack elements during deoptimization"); 677 tty->print_cr(" Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1); 678 tty->print_cr(" Fabricated interpreter frame had %d expression stack elements", 679 iframe->interpreter_frame_expression_stack_size()); 680 tty->print_cr(" Interpreter oop map had %d expression stack elements", mask.expression_stack_size()); 681 tty->print_cr(" try_next_mask = %d", try_next_mask); 682 tty->print_cr(" next_mask_expression_stack_size = %d", next_mask_expression_stack_size); 683 tty->print_cr(" callee_size_of_parameters = %d", callee_size_of_parameters); 684 tty->print_cr(" callee_max_locals = %d", callee_max_locals); 685 tty->print_cr(" top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment); 686 tty->print_cr(" exec_mode = %d", exec_mode); 687 tty->print_cr(" cur_invoke_parameter_size = %d", cur_invoke_parameter_size); 688 tty->print_cr(" Thread = " INTPTR_FORMAT ", thread ID = " UINTX_FORMAT, thread, thread->osthread()->thread_id()); 689 tty->print_cr(" Interpreted frames:"); 690 for (int k = 0; k < cur_array->frames(); k++) { 691 vframeArrayElement* el = cur_array->element(k); 692 tty->print_cr(" %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci()); 693 } 694 cur_array->print_on_2(tty); 695 guarantee(false, "wrong number of expression stack elements during deopt"); 696 } 697 VerifyOopClosure verify; 698 iframe->oops_interpreted_do(&verify, &rm, false); 699 callee_size_of_parameters = mh->size_of_parameters(); 700 callee_max_locals = mh->max_locals(); 701 is_top_frame = false; 702 } 703 } 704 #endif /* !PRODUCT */ 705 706 707 return bt; 708 JRT_END 709 710 711 int Deoptimization::deoptimize_dependents() { 712 Threads::deoptimized_wrt_marked_nmethods(); 713 return 0; 714 } 715 716 717 #ifdef COMPILER2 718 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, GrowableArray<ScopeValue*>* objects, TRAPS) { 719 Handle pending_exception(thread->pending_exception()); 720 const char* exception_file = thread->exception_file(); 721 int exception_line = thread->exception_line(); 722 thread->clear_pending_exception(); 723 724 for (int i = 0; i < objects->length(); i++) { 725 assert(objects->at(i)->is_object(), "invalid debug information"); 726 ObjectValue* sv = (ObjectValue*) objects->at(i); 727 728 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 729 oop obj = NULL; 730 731 if (k->oop_is_instance()) { 732 instanceKlass* ik = instanceKlass::cast(k()); 733 obj = ik->allocate_instance(CHECK_(false)); 734 } else if (k->oop_is_typeArray()) { 735 typeArrayKlass* ak = typeArrayKlass::cast(k()); 736 assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length"); 737 int len = sv->field_size() / type2size[ak->element_type()]; 738 obj = ak->allocate(len, CHECK_(false)); 739 } else if (k->oop_is_objArray()) { 740 objArrayKlass* ak = objArrayKlass::cast(k()); 741 obj = ak->allocate(sv->field_size(), CHECK_(false)); 742 } 743 744 assert(obj != NULL, "allocation failed"); 745 assert(sv->value().is_null(), "redundant reallocation"); 746 sv->set_value(obj); 747 } 748 749 if (pending_exception.not_null()) { 750 thread->set_pending_exception(pending_exception(), exception_file, exception_line); 751 } 752 753 return true; 754 } 755 756 // This assumes that the fields are stored in ObjectValue in the same order 757 // they are yielded by do_nonstatic_fields. 758 class FieldReassigner: public FieldClosure { 759 frame* _fr; 760 RegisterMap* _reg_map; 761 ObjectValue* _sv; 762 instanceKlass* _ik; 763 oop _obj; 764 765 int _i; 766 public: 767 FieldReassigner(frame* fr, RegisterMap* reg_map, ObjectValue* sv, oop obj) : 768 _fr(fr), _reg_map(reg_map), _sv(sv), _obj(obj), _i(0) {} 769 770 int i() const { return _i; } 771 772 773 void do_field(fieldDescriptor* fd) { 774 intptr_t val; 775 StackValue* value = 776 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(i())); 777 int offset = fd->offset(); 778 switch (fd->field_type()) { 779 case T_OBJECT: case T_ARRAY: 780 assert(value->type() == T_OBJECT, "Agreement."); 781 _obj->obj_field_put(offset, value->get_obj()()); 782 break; 783 784 case T_LONG: case T_DOUBLE: { 785 assert(value->type() == T_INT, "Agreement."); 786 StackValue* low = 787 StackValue::create_stack_value(_fr, _reg_map, _sv->field_at(++_i)); 788 #ifdef _LP64 789 jlong res = (jlong)low->get_int(); 790 #else 791 #ifdef SPARC 792 // For SPARC we have to swap high and low words. 793 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 794 #else 795 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 796 #endif //SPARC 797 #endif 798 _obj->long_field_put(offset, res); 799 break; 800 } 801 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 802 case T_INT: case T_FLOAT: // 4 bytes. 803 assert(value->type() == T_INT, "Agreement."); 804 val = value->get_int(); 805 _obj->int_field_put(offset, (jint)*((jint*)&val)); 806 break; 807 808 case T_SHORT: case T_CHAR: // 2 bytes 809 assert(value->type() == T_INT, "Agreement."); 810 val = value->get_int(); 811 _obj->short_field_put(offset, (jshort)*((jint*)&val)); 812 break; 813 814 case T_BOOLEAN: case T_BYTE: // 1 byte 815 assert(value->type() == T_INT, "Agreement."); 816 val = value->get_int(); 817 _obj->bool_field_put(offset, (jboolean)*((jint*)&val)); 818 break; 819 820 default: 821 ShouldNotReachHere(); 822 } 823 _i++; 824 } 825 }; 826 827 // restore elements of an eliminated type array 828 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) { 829 int index = 0; 830 intptr_t val; 831 832 for (int i = 0; i < sv->field_size(); i++) { 833 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 834 switch(type) { 835 case T_LONG: case T_DOUBLE: { 836 assert(value->type() == T_INT, "Agreement."); 837 StackValue* low = 838 StackValue::create_stack_value(fr, reg_map, sv->field_at(++i)); 839 #ifdef _LP64 840 jlong res = (jlong)low->get_int(); 841 #else 842 #ifdef SPARC 843 // For SPARC we have to swap high and low words. 844 jlong res = jlong_from((jint)low->get_int(), (jint)value->get_int()); 845 #else 846 jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int()); 847 #endif //SPARC 848 #endif 849 obj->long_at_put(index, res); 850 break; 851 } 852 853 // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem. 854 case T_INT: case T_FLOAT: // 4 bytes. 855 assert(value->type() == T_INT, "Agreement."); 856 val = value->get_int(); 857 obj->int_at_put(index, (jint)*((jint*)&val)); 858 break; 859 860 case T_SHORT: case T_CHAR: // 2 bytes 861 assert(value->type() == T_INT, "Agreement."); 862 val = value->get_int(); 863 obj->short_at_put(index, (jshort)*((jint*)&val)); 864 break; 865 866 case T_BOOLEAN: case T_BYTE: // 1 byte 867 assert(value->type() == T_INT, "Agreement."); 868 val = value->get_int(); 869 obj->bool_at_put(index, (jboolean)*((jint*)&val)); 870 break; 871 872 default: 873 ShouldNotReachHere(); 874 } 875 index++; 876 } 877 } 878 879 880 // restore fields of an eliminated object array 881 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) { 882 for (int i = 0; i < sv->field_size(); i++) { 883 StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i)); 884 assert(value->type() == T_OBJECT, "object element expected"); 885 obj->obj_at_put(i, value->get_obj()()); 886 } 887 } 888 889 890 // restore fields of all eliminated objects and arrays 891 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects) { 892 for (int i = 0; i < objects->length(); i++) { 893 ObjectValue* sv = (ObjectValue*) objects->at(i); 894 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 895 Handle obj = sv->value(); 896 assert(obj.not_null(), "reallocation was missed"); 897 898 if (k->oop_is_instance()) { 899 instanceKlass* ik = instanceKlass::cast(k()); 900 FieldReassigner reassign(fr, reg_map, sv, obj()); 901 ik->do_nonstatic_fields(&reassign); 902 } else if (k->oop_is_typeArray()) { 903 typeArrayKlass* ak = typeArrayKlass::cast(k()); 904 reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type()); 905 } else if (k->oop_is_objArray()) { 906 reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj()); 907 } 908 } 909 } 910 911 912 // relock objects for which synchronization was eliminated 913 void Deoptimization::relock_objects(GrowableArray<MonitorInfo*>* monitors, JavaThread* thread) { 914 for (int i = 0; i < monitors->length(); i++) { 915 MonitorInfo* mon_info = monitors->at(i); 916 if (mon_info->eliminated()) { 917 assert(mon_info->owner() != NULL, "reallocation was missed"); 918 Handle obj = Handle(mon_info->owner()); 919 markOop mark = obj->mark(); 920 if (UseBiasedLocking && mark->has_bias_pattern()) { 921 // New allocated objects may have the mark set to anonymously biased. 922 // Also the deoptimized method may called methods with synchronization 923 // where the thread-local object is bias locked to the current thread. 924 assert(mark->is_biased_anonymously() || 925 mark->biased_locker() == thread, "should be locked to current thread"); 926 // Reset mark word to unbiased prototype. 927 markOop unbiased_prototype = markOopDesc::prototype()->set_age(mark->age()); 928 obj->set_mark(unbiased_prototype); 929 } 930 BasicLock* lock = mon_info->lock(); 931 ObjectSynchronizer::slow_enter(obj, lock, thread); 932 } 933 assert(mon_info->owner()->is_locked(), "object must be locked now"); 934 } 935 } 936 937 938 #ifndef PRODUCT 939 // print information about reallocated objects 940 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects) { 941 fieldDescriptor fd; 942 943 for (int i = 0; i < objects->length(); i++) { 944 ObjectValue* sv = (ObjectValue*) objects->at(i); 945 KlassHandle k(((ConstantOopReadValue*) sv->klass())->value()()); 946 Handle obj = sv->value(); 947 948 tty->print(" object <" INTPTR_FORMAT "> of type ", sv->value()()); 949 k->as_klassOop()->print_value(); 950 tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize); 951 tty->cr(); 952 953 if (Verbose) { 954 k->oop_print_on(obj(), tty); 955 } 956 } 957 } 958 #endif 959 #endif // COMPILER2 960 961 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk) { 962 963 #ifndef PRODUCT 964 if (TraceDeoptimization) { 965 ttyLocker ttyl; 966 tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", thread); 967 fr.print_on(tty); 968 tty->print_cr(" Virtual frames (innermost first):"); 969 for (int index = 0; index < chunk->length(); index++) { 970 compiledVFrame* vf = chunk->at(index); 971 tty->print(" %2d - ", index); 972 vf->print_value(); 973 int bci = chunk->at(index)->raw_bci(); 974 const char* code_name; 975 if (bci == SynchronizationEntryBCI) { 976 code_name = "sync entry"; 977 } else { 978 Bytecodes::Code code = vf->method()->code_at(bci); 979 code_name = Bytecodes::name(code); 980 } 981 tty->print(" - %s", code_name); 982 tty->print_cr(" @ bci %d ", bci); 983 if (Verbose) { 984 vf->print(); 985 tty->cr(); 986 } 987 } 988 } 989 #endif 990 991 // Register map for next frame (used for stack crawl). We capture 992 // the state of the deopt'ing frame's caller. Thus if we need to 993 // stuff a C2I adapter we can properly fill in the callee-save 994 // register locations. 995 frame caller = fr.sender(reg_map); 996 int frame_size = caller.sp() - fr.sp(); 997 998 frame sender = caller; 999 1000 // Since the Java thread being deoptimized will eventually adjust it's own stack, 1001 // the vframeArray containing the unpacking information is allocated in the C heap. 1002 // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames(). 1003 vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr); 1004 1005 // Compare the vframeArray to the collected vframes 1006 assert(array->structural_compare(thread, chunk), "just checking"); 1007 Events::log("# vframes = %d", (intptr_t)chunk->length()); 1008 1009 #ifndef PRODUCT 1010 if (TraceDeoptimization) { 1011 ttyLocker ttyl; 1012 tty->print_cr(" Created vframeArray " INTPTR_FORMAT, array); 1013 } 1014 #endif // PRODUCT 1015 1016 return array; 1017 } 1018 1019 1020 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke) { 1021 GrowableArray<MonitorInfo*>* monitors = cvf->monitors(); 1022 for (int i = 0; i < monitors->length(); i++) { 1023 MonitorInfo* mon_info = monitors->at(i); 1024 if (!mon_info->eliminated() && mon_info->owner() != NULL) { 1025 objects_to_revoke->append(Handle(mon_info->owner())); 1026 } 1027 } 1028 } 1029 1030 1031 void Deoptimization::revoke_biases_of_monitors(JavaThread* thread, frame fr, RegisterMap* map) { 1032 if (!UseBiasedLocking) { 1033 return; 1034 } 1035 1036 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1037 1038 // Unfortunately we don't have a RegisterMap available in most of 1039 // the places we want to call this routine so we need to walk the 1040 // stack again to update the register map. 1041 if (map == NULL || !map->update_map()) { 1042 StackFrameStream sfs(thread, true); 1043 bool found = false; 1044 while (!found && !sfs.is_done()) { 1045 frame* cur = sfs.current(); 1046 sfs.next(); 1047 found = cur->id() == fr.id(); 1048 } 1049 assert(found, "frame to be deoptimized not found on target thread's stack"); 1050 map = sfs.register_map(); 1051 } 1052 1053 vframe* vf = vframe::new_vframe(&fr, map, thread); 1054 compiledVFrame* cvf = compiledVFrame::cast(vf); 1055 // Revoke monitors' biases in all scopes 1056 while (!cvf->is_top()) { 1057 collect_monitors(cvf, objects_to_revoke); 1058 cvf = compiledVFrame::cast(cvf->sender()); 1059 } 1060 collect_monitors(cvf, objects_to_revoke); 1061 1062 if (SafepointSynchronize::is_at_safepoint()) { 1063 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1064 } else { 1065 BiasedLocking::revoke(objects_to_revoke); 1066 } 1067 } 1068 1069 1070 void Deoptimization::revoke_biases_of_monitors(CodeBlob* cb) { 1071 if (!UseBiasedLocking) { 1072 return; 1073 } 1074 1075 assert(SafepointSynchronize::is_at_safepoint(), "must only be called from safepoint"); 1076 GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>(); 1077 for (JavaThread* jt = Threads::first(); jt != NULL ; jt = jt->next()) { 1078 if (jt->has_last_Java_frame()) { 1079 StackFrameStream sfs(jt, true); 1080 while (!sfs.is_done()) { 1081 frame* cur = sfs.current(); 1082 if (cb->contains(cur->pc())) { 1083 vframe* vf = vframe::new_vframe(cur, sfs.register_map(), jt); 1084 compiledVFrame* cvf = compiledVFrame::cast(vf); 1085 // Revoke monitors' biases in all scopes 1086 while (!cvf->is_top()) { 1087 collect_monitors(cvf, objects_to_revoke); 1088 cvf = compiledVFrame::cast(cvf->sender()); 1089 } 1090 collect_monitors(cvf, objects_to_revoke); 1091 } 1092 sfs.next(); 1093 } 1094 } 1095 } 1096 BiasedLocking::revoke_at_safepoint(objects_to_revoke); 1097 } 1098 1099 1100 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr) { 1101 assert(fr.can_be_deoptimized(), "checking frame type"); 1102 1103 gather_statistics(Reason_constraint, Action_none, Bytecodes::_illegal); 1104 1105 EventMark m("Deoptimization (pc=" INTPTR_FORMAT ", sp=" INTPTR_FORMAT ")", fr.pc(), fr.id()); 1106 1107 // Patch the nmethod so that when execution returns to it we will 1108 // deopt the execution state and return to the interpreter. 1109 fr.deoptimize(thread); 1110 } 1111 1112 void Deoptimization::deoptimize(JavaThread* thread, frame fr, RegisterMap *map) { 1113 // Deoptimize only if the frame comes from compile code. 1114 // Do not deoptimize the frame which is already patched 1115 // during the execution of the loops below. 1116 if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) { 1117 return; 1118 } 1119 ResourceMark rm; 1120 DeoptimizationMarker dm; 1121 if (UseBiasedLocking) { 1122 revoke_biases_of_monitors(thread, fr, map); 1123 } 1124 deoptimize_single_frame(thread, fr); 1125 1126 } 1127 1128 1129 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id) { 1130 assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(), 1131 "can only deoptimize other thread at a safepoint"); 1132 // Compute frame and register map based on thread and sp. 1133 RegisterMap reg_map(thread, UseBiasedLocking); 1134 frame fr = thread->last_frame(); 1135 while (fr.id() != id) { 1136 fr = fr.sender(®_map); 1137 } 1138 deoptimize(thread, fr, ®_map); 1139 } 1140 1141 1142 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) { 1143 if (thread == Thread::current()) { 1144 Deoptimization::deoptimize_frame_internal(thread, id); 1145 } else { 1146 VM_DeoptimizeFrame deopt(thread, id); 1147 VMThread::execute(&deopt); 1148 } 1149 } 1150 1151 1152 // JVMTI PopFrame support 1153 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address)) 1154 { 1155 thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address); 1156 } 1157 JRT_END 1158 1159 1160 #if defined(COMPILER2) || defined(SHARK) 1161 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index, TRAPS) { 1162 // in case of an unresolved klass entry, load the class. 1163 if (constant_pool->tag_at(index).is_unresolved_klass()) { 1164 klassOop tk = constant_pool->klass_at(index, CHECK); 1165 return; 1166 } 1167 1168 if (!constant_pool->tag_at(index).is_symbol()) return; 1169 1170 Handle class_loader (THREAD, instanceKlass::cast(constant_pool->pool_holder())->class_loader()); 1171 Symbol* symbol = constant_pool->symbol_at(index); 1172 1173 // class name? 1174 if (symbol->byte_at(0) != '(') { 1175 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1176 SystemDictionary::resolve_or_null(symbol, class_loader, protection_domain, CHECK); 1177 return; 1178 } 1179 1180 // then it must be a signature! 1181 ResourceMark rm(THREAD); 1182 for (SignatureStream ss(symbol); !ss.is_done(); ss.next()) { 1183 if (ss.is_object()) { 1184 Symbol* class_name = ss.as_symbol(CHECK); 1185 Handle protection_domain (THREAD, Klass::cast(constant_pool->pool_holder())->protection_domain()); 1186 SystemDictionary::resolve_or_null(class_name, class_loader, protection_domain, CHECK); 1187 } 1188 } 1189 } 1190 1191 1192 void Deoptimization::load_class_by_index(constantPoolHandle constant_pool, int index) { 1193 EXCEPTION_MARK; 1194 load_class_by_index(constant_pool, index, THREAD); 1195 if (HAS_PENDING_EXCEPTION) { 1196 // Exception happened during classloading. We ignore the exception here, since it 1197 // is going to be rethrown since the current activation is going to be deoptimzied and 1198 // the interpreter will re-execute the bytecode. 1199 CLEAR_PENDING_EXCEPTION; 1200 } 1201 } 1202 1203 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) { 1204 HandleMark hm; 1205 1206 // uncommon_trap() is called at the beginning of the uncommon trap 1207 // handler. Note this fact before we start generating temporary frames 1208 // that can confuse an asynchronous stack walker. This counter is 1209 // decremented at the end of unpack_frames(). 1210 thread->inc_in_deopt_handler(); 1211 1212 // We need to update the map if we have biased locking. 1213 RegisterMap reg_map(thread, UseBiasedLocking); 1214 frame stub_frame = thread->last_frame(); 1215 frame fr = stub_frame.sender(®_map); 1216 // Make sure the calling nmethod is not getting deoptimized and removed 1217 // before we are done with it. 1218 nmethodLocker nl(fr.pc()); 1219 1220 { 1221 ResourceMark rm; 1222 1223 // Revoke biases of any monitors in the frame to ensure we can migrate them 1224 revoke_biases_of_monitors(thread, fr, ®_map); 1225 1226 DeoptReason reason = trap_request_reason(trap_request); 1227 DeoptAction action = trap_request_action(trap_request); 1228 jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1 1229 1230 Events::log("Uncommon trap occurred @" INTPTR_FORMAT " unloaded_class_index = %d", fr.pc(), (int) trap_request); 1231 vframe* vf = vframe::new_vframe(&fr, ®_map, thread); 1232 compiledVFrame* cvf = compiledVFrame::cast(vf); 1233 1234 nmethod* nm = cvf->code(); 1235 1236 ScopeDesc* trap_scope = cvf->scope(); 1237 methodHandle trap_method = trap_scope->method(); 1238 int trap_bci = trap_scope->bci(); 1239 Bytecodes::Code trap_bc = trap_method->java_code_at(trap_bci); 1240 1241 // Record this event in the histogram. 1242 gather_statistics(reason, action, trap_bc); 1243 1244 // Ensure that we can record deopt. history: 1245 bool create_if_missing = ProfileTraps; 1246 1247 methodDataHandle trap_mdo 1248 (THREAD, get_method_data(thread, trap_method, create_if_missing)); 1249 1250 // Print a bunch of diagnostics, if requested. 1251 if (TraceDeoptimization || LogCompilation) { 1252 ResourceMark rm; 1253 ttyLocker ttyl; 1254 char buf[100]; 1255 if (xtty != NULL) { 1256 xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT"' %s", 1257 os::current_thread_id(), 1258 format_trap_request(buf, sizeof(buf), trap_request)); 1259 nm->log_identity(xtty); 1260 } 1261 Symbol* class_name = NULL; 1262 bool unresolved = false; 1263 if (unloaded_class_index >= 0) { 1264 constantPoolHandle constants (THREAD, trap_method->constants()); 1265 if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) { 1266 class_name = constants->klass_name_at(unloaded_class_index); 1267 unresolved = true; 1268 if (xtty != NULL) 1269 xtty->print(" unresolved='1'"); 1270 } else if (constants->tag_at(unloaded_class_index).is_symbol()) { 1271 class_name = constants->symbol_at(unloaded_class_index); 1272 } 1273 if (xtty != NULL) 1274 xtty->name(class_name); 1275 } 1276 if (xtty != NULL && trap_mdo.not_null()) { 1277 // Dump the relevant MDO state. 1278 // This is the deopt count for the current reason, any previous 1279 // reasons or recompiles seen at this point. 1280 int dcnt = trap_mdo->trap_count(reason); 1281 if (dcnt != 0) 1282 xtty->print(" count='%d'", dcnt); 1283 ProfileData* pdata = trap_mdo->bci_to_data(trap_bci); 1284 int dos = (pdata == NULL)? 0: pdata->trap_state(); 1285 if (dos != 0) { 1286 xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos)); 1287 if (trap_state_is_recompiled(dos)) { 1288 int recnt2 = trap_mdo->overflow_recompile_count(); 1289 if (recnt2 != 0) 1290 xtty->print(" recompiles2='%d'", recnt2); 1291 } 1292 } 1293 } 1294 if (xtty != NULL) { 1295 xtty->stamp(); 1296 xtty->end_head(); 1297 } 1298 if (TraceDeoptimization) { // make noise on the tty 1299 tty->print("Uncommon trap occurred in"); 1300 nm->method()->print_short_name(tty); 1301 tty->print(" (@" INTPTR_FORMAT ") thread=%d reason=%s action=%s unloaded_class_index=%d", 1302 fr.pc(), 1303 (int) os::current_thread_id(), 1304 trap_reason_name(reason), 1305 trap_action_name(action), 1306 unloaded_class_index); 1307 if (class_name != NULL) { 1308 tty->print(unresolved ? " unresolved class: " : " symbol: "); 1309 class_name->print_symbol_on(tty); 1310 } 1311 tty->cr(); 1312 } 1313 if (xtty != NULL) { 1314 // Log the precise location of the trap. 1315 for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) { 1316 xtty->begin_elem("jvms bci='%d'", sd->bci()); 1317 xtty->method(sd->method()); 1318 xtty->end_elem(); 1319 if (sd->is_top()) break; 1320 } 1321 xtty->tail("uncommon_trap"); 1322 } 1323 } 1324 // (End diagnostic printout.) 1325 1326 // Load class if necessary 1327 if (unloaded_class_index >= 0) { 1328 constantPoolHandle constants(THREAD, trap_method->constants()); 1329 load_class_by_index(constants, unloaded_class_index); 1330 } 1331 1332 // Flush the nmethod if necessary and desirable. 1333 // 1334 // We need to avoid situations where we are re-flushing the nmethod 1335 // because of a hot deoptimization site. Repeated flushes at the same 1336 // point need to be detected by the compiler and avoided. If the compiler 1337 // cannot avoid them (or has a bug and "refuses" to avoid them), this 1338 // module must take measures to avoid an infinite cycle of recompilation 1339 // and deoptimization. There are several such measures: 1340 // 1341 // 1. If a recompilation is ordered a second time at some site X 1342 // and for the same reason R, the action is adjusted to 'reinterpret', 1343 // to give the interpreter time to exercise the method more thoroughly. 1344 // If this happens, the method's overflow_recompile_count is incremented. 1345 // 1346 // 2. If the compiler fails to reduce the deoptimization rate, then 1347 // the method's overflow_recompile_count will begin to exceed the set 1348 // limit PerBytecodeRecompilationCutoff. If this happens, the action 1349 // is adjusted to 'make_not_compilable', and the method is abandoned 1350 // to the interpreter. This is a performance hit for hot methods, 1351 // but is better than a disastrous infinite cycle of recompilations. 1352 // (Actually, only the method containing the site X is abandoned.) 1353 // 1354 // 3. In parallel with the previous measures, if the total number of 1355 // recompilations of a method exceeds the much larger set limit 1356 // PerMethodRecompilationCutoff, the method is abandoned. 1357 // This should only happen if the method is very large and has 1358 // many "lukewarm" deoptimizations. The code which enforces this 1359 // limit is elsewhere (class nmethod, class methodOopDesc). 1360 // 1361 // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance 1362 // to recompile at each bytecode independently of the per-BCI cutoff. 1363 // 1364 // The decision to update code is up to the compiler, and is encoded 1365 // in the Action_xxx code. If the compiler requests Action_none 1366 // no trap state is changed, no compiled code is changed, and the 1367 // computation suffers along in the interpreter. 1368 // 1369 // The other action codes specify various tactics for decompilation 1370 // and recompilation. Action_maybe_recompile is the loosest, and 1371 // allows the compiled code to stay around until enough traps are seen, 1372 // and until the compiler gets around to recompiling the trapping method. 1373 // 1374 // The other actions cause immediate removal of the present code. 1375 1376 bool update_trap_state = true; 1377 bool make_not_entrant = false; 1378 bool make_not_compilable = false; 1379 bool reprofile = false; 1380 switch (action) { 1381 case Action_none: 1382 // Keep the old code. 1383 update_trap_state = false; 1384 break; 1385 case Action_maybe_recompile: 1386 // Do not need to invalidate the present code, but we can 1387 // initiate another 1388 // Start compiler without (necessarily) invalidating the nmethod. 1389 // The system will tolerate the old code, but new code should be 1390 // generated when possible. 1391 break; 1392 case Action_reinterpret: 1393 // Go back into the interpreter for a while, and then consider 1394 // recompiling form scratch. 1395 make_not_entrant = true; 1396 // Reset invocation counter for outer most method. 1397 // This will allow the interpreter to exercise the bytecodes 1398 // for a while before recompiling. 1399 // By contrast, Action_make_not_entrant is immediate. 1400 // 1401 // Note that the compiler will track null_check, null_assert, 1402 // range_check, and class_check events and log them as if they 1403 // had been traps taken from compiled code. This will update 1404 // the MDO trap history so that the next compilation will 1405 // properly detect hot trap sites. 1406 reprofile = true; 1407 break; 1408 case Action_make_not_entrant: 1409 // Request immediate recompilation, and get rid of the old code. 1410 // Make them not entrant, so next time they are called they get 1411 // recompiled. Unloaded classes are loaded now so recompile before next 1412 // time they are called. Same for uninitialized. The interpreter will 1413 // link the missing class, if any. 1414 make_not_entrant = true; 1415 break; 1416 case Action_make_not_compilable: 1417 // Give up on compiling this method at all. 1418 make_not_entrant = true; 1419 make_not_compilable = true; 1420 break; 1421 default: 1422 ShouldNotReachHere(); 1423 } 1424 1425 // Setting +ProfileTraps fixes the following, on all platforms: 1426 // 4852688: ProfileInterpreter is off by default for ia64. The result is 1427 // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the 1428 // recompile relies on a methodDataOop to record heroic opt failures. 1429 1430 // Whether the interpreter is producing MDO data or not, we also need 1431 // to use the MDO to detect hot deoptimization points and control 1432 // aggressive optimization. 1433 bool inc_recompile_count = false; 1434 ProfileData* pdata = NULL; 1435 if (ProfileTraps && update_trap_state && trap_mdo.not_null()) { 1436 assert(trap_mdo() == get_method_data(thread, trap_method, false), "sanity"); 1437 uint this_trap_count = 0; 1438 bool maybe_prior_trap = false; 1439 bool maybe_prior_recompile = false; 1440 pdata = query_update_method_data(trap_mdo, trap_bci, reason, 1441 //outputs: 1442 this_trap_count, 1443 maybe_prior_trap, 1444 maybe_prior_recompile); 1445 // Because the interpreter also counts null, div0, range, and class 1446 // checks, these traps from compiled code are double-counted. 1447 // This is harmless; it just means that the PerXTrapLimit values 1448 // are in effect a little smaller than they look. 1449 1450 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1451 if (per_bc_reason != Reason_none) { 1452 // Now take action based on the partially known per-BCI history. 1453 if (maybe_prior_trap 1454 && this_trap_count >= (uint)PerBytecodeTrapLimit) { 1455 // If there are too many traps at this BCI, force a recompile. 1456 // This will allow the compiler to see the limit overflow, and 1457 // take corrective action, if possible. The compiler generally 1458 // does not use the exact PerBytecodeTrapLimit value, but instead 1459 // changes its tactics if it sees any traps at all. This provides 1460 // a little hysteresis, delaying a recompile until a trap happens 1461 // several times. 1462 // 1463 // Actually, since there is only one bit of counter per BCI, 1464 // the possible per-BCI counts are {0,1,(per-method count)}. 1465 // This produces accurate results if in fact there is only 1466 // one hot trap site, but begins to get fuzzy if there are 1467 // many sites. For example, if there are ten sites each 1468 // trapping two or more times, they each get the blame for 1469 // all of their traps. 1470 make_not_entrant = true; 1471 } 1472 1473 // Detect repeated recompilation at the same BCI, and enforce a limit. 1474 if (make_not_entrant && maybe_prior_recompile) { 1475 // More than one recompile at this point. 1476 inc_recompile_count = maybe_prior_trap; 1477 } 1478 } else { 1479 // For reasons which are not recorded per-bytecode, we simply 1480 // force recompiles unconditionally. 1481 // (Note that PerMethodRecompilationCutoff is enforced elsewhere.) 1482 make_not_entrant = true; 1483 } 1484 1485 // Go back to the compiler if there are too many traps in this method. 1486 if (this_trap_count >= (uint)PerMethodTrapLimit) { 1487 // If there are too many traps in this method, force a recompile. 1488 // This will allow the compiler to see the limit overflow, and 1489 // take corrective action, if possible. 1490 // (This condition is an unlikely backstop only, because the 1491 // PerBytecodeTrapLimit is more likely to take effect first, 1492 // if it is applicable.) 1493 make_not_entrant = true; 1494 } 1495 1496 // Here's more hysteresis: If there has been a recompile at 1497 // this trap point already, run the method in the interpreter 1498 // for a while to exercise it more thoroughly. 1499 if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) { 1500 reprofile = true; 1501 } 1502 1503 } 1504 1505 // Take requested actions on the method: 1506 1507 // Recompile 1508 if (make_not_entrant) { 1509 if (!nm->make_not_entrant()) { 1510 return; // the call did not change nmethod's state 1511 } 1512 1513 if (pdata != NULL) { 1514 // Record the recompilation event, if any. 1515 int tstate0 = pdata->trap_state(); 1516 int tstate1 = trap_state_set_recompiled(tstate0, true); 1517 if (tstate1 != tstate0) 1518 pdata->set_trap_state(tstate1); 1519 } 1520 } 1521 1522 if (inc_recompile_count) { 1523 trap_mdo->inc_overflow_recompile_count(); 1524 if ((uint)trap_mdo->overflow_recompile_count() > 1525 (uint)PerBytecodeRecompilationCutoff) { 1526 // Give up on the method containing the bad BCI. 1527 if (trap_method() == nm->method()) { 1528 make_not_compilable = true; 1529 } else { 1530 trap_method->set_not_compilable(CompLevel_full_optimization); 1531 // But give grace to the enclosing nm->method(). 1532 } 1533 } 1534 } 1535 1536 // Reprofile 1537 if (reprofile) { 1538 CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method()); 1539 } 1540 1541 // Give up compiling 1542 if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) { 1543 assert(make_not_entrant, "consistent"); 1544 nm->method()->set_not_compilable(CompLevel_full_optimization); 1545 } 1546 1547 } // Free marked resources 1548 1549 } 1550 JRT_END 1551 1552 methodDataOop 1553 Deoptimization::get_method_data(JavaThread* thread, methodHandle m, 1554 bool create_if_missing) { 1555 Thread* THREAD = thread; 1556 methodDataOop mdo = m()->method_data(); 1557 if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) { 1558 // Build an MDO. Ignore errors like OutOfMemory; 1559 // that simply means we won't have an MDO to update. 1560 methodOopDesc::build_interpreter_method_data(m, THREAD); 1561 if (HAS_PENDING_EXCEPTION) { 1562 assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here"); 1563 CLEAR_PENDING_EXCEPTION; 1564 } 1565 mdo = m()->method_data(); 1566 } 1567 return mdo; 1568 } 1569 1570 ProfileData* 1571 Deoptimization::query_update_method_data(methodDataHandle trap_mdo, 1572 int trap_bci, 1573 Deoptimization::DeoptReason reason, 1574 //outputs: 1575 uint& ret_this_trap_count, 1576 bool& ret_maybe_prior_trap, 1577 bool& ret_maybe_prior_recompile) { 1578 uint prior_trap_count = trap_mdo->trap_count(reason); 1579 uint this_trap_count = trap_mdo->inc_trap_count(reason); 1580 1581 // If the runtime cannot find a place to store trap history, 1582 // it is estimated based on the general condition of the method. 1583 // If the method has ever been recompiled, or has ever incurred 1584 // a trap with the present reason , then this BCI is assumed 1585 // (pessimistically) to be the culprit. 1586 bool maybe_prior_trap = (prior_trap_count != 0); 1587 bool maybe_prior_recompile = (trap_mdo->decompile_count() != 0); 1588 ProfileData* pdata = NULL; 1589 1590 1591 // For reasons which are recorded per bytecode, we check per-BCI data. 1592 DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason); 1593 if (per_bc_reason != Reason_none) { 1594 // Find the profile data for this BCI. If there isn't one, 1595 // try to allocate one from the MDO's set of spares. 1596 // This will let us detect a repeated trap at this point. 1597 pdata = trap_mdo->allocate_bci_to_data(trap_bci); 1598 1599 if (pdata != NULL) { 1600 // Query the trap state of this profile datum. 1601 int tstate0 = pdata->trap_state(); 1602 if (!trap_state_has_reason(tstate0, per_bc_reason)) 1603 maybe_prior_trap = false; 1604 if (!trap_state_is_recompiled(tstate0)) 1605 maybe_prior_recompile = false; 1606 1607 // Update the trap state of this profile datum. 1608 int tstate1 = tstate0; 1609 // Record the reason. 1610 tstate1 = trap_state_add_reason(tstate1, per_bc_reason); 1611 // Store the updated state on the MDO, for next time. 1612 if (tstate1 != tstate0) 1613 pdata->set_trap_state(tstate1); 1614 } else { 1615 if (LogCompilation && xtty != NULL) { 1616 ttyLocker ttyl; 1617 // Missing MDP? Leave a small complaint in the log. 1618 xtty->elem("missing_mdp bci='%d'", trap_bci); 1619 } 1620 } 1621 } 1622 1623 // Return results: 1624 ret_this_trap_count = this_trap_count; 1625 ret_maybe_prior_trap = maybe_prior_trap; 1626 ret_maybe_prior_recompile = maybe_prior_recompile; 1627 return pdata; 1628 } 1629 1630 void 1631 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1632 ResourceMark rm; 1633 // Ignored outputs: 1634 uint ignore_this_trap_count; 1635 bool ignore_maybe_prior_trap; 1636 bool ignore_maybe_prior_recompile; 1637 query_update_method_data(trap_mdo, trap_bci, 1638 (DeoptReason)reason, 1639 ignore_this_trap_count, 1640 ignore_maybe_prior_trap, 1641 ignore_maybe_prior_recompile); 1642 } 1643 1644 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request) { 1645 1646 // Still in Java no safepoints 1647 { 1648 // This enters VM and may safepoint 1649 uncommon_trap_inner(thread, trap_request); 1650 } 1651 return fetch_unroll_info_helper(thread); 1652 } 1653 1654 // Local derived constants. 1655 // Further breakdown of DataLayout::trap_state, as promised by DataLayout. 1656 const int DS_REASON_MASK = DataLayout::trap_mask >> 1; 1657 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK; 1658 1659 //---------------------------trap_state_reason--------------------------------- 1660 Deoptimization::DeoptReason 1661 Deoptimization::trap_state_reason(int trap_state) { 1662 // This assert provides the link between the width of DataLayout::trap_bits 1663 // and the encoding of "recorded" reasons. It ensures there are enough 1664 // bits to store all needed reasons in the per-BCI MDO profile. 1665 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1666 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1667 trap_state -= recompile_bit; 1668 if (trap_state == DS_REASON_MASK) { 1669 return Reason_many; 1670 } else { 1671 assert((int)Reason_none == 0, "state=0 => Reason_none"); 1672 return (DeoptReason)trap_state; 1673 } 1674 } 1675 //-------------------------trap_state_has_reason------------------------------- 1676 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1677 assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason"); 1678 assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits"); 1679 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1680 trap_state -= recompile_bit; 1681 if (trap_state == DS_REASON_MASK) { 1682 return -1; // true, unspecifically (bottom of state lattice) 1683 } else if (trap_state == reason) { 1684 return 1; // true, definitely 1685 } else if (trap_state == 0) { 1686 return 0; // false, definitely (top of state lattice) 1687 } else { 1688 return 0; // false, definitely 1689 } 1690 } 1691 //-------------------------trap_state_add_reason------------------------------- 1692 int Deoptimization::trap_state_add_reason(int trap_state, int reason) { 1693 assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason"); 1694 int recompile_bit = (trap_state & DS_RECOMPILE_BIT); 1695 trap_state -= recompile_bit; 1696 if (trap_state == DS_REASON_MASK) { 1697 return trap_state + recompile_bit; // already at state lattice bottom 1698 } else if (trap_state == reason) { 1699 return trap_state + recompile_bit; // the condition is already true 1700 } else if (trap_state == 0) { 1701 return reason + recompile_bit; // no condition has yet been true 1702 } else { 1703 return DS_REASON_MASK + recompile_bit; // fall to state lattice bottom 1704 } 1705 } 1706 //-----------------------trap_state_is_recompiled------------------------------ 1707 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1708 return (trap_state & DS_RECOMPILE_BIT) != 0; 1709 } 1710 //-----------------------trap_state_set_recompiled----------------------------- 1711 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) { 1712 if (z) return trap_state | DS_RECOMPILE_BIT; 1713 else return trap_state & ~DS_RECOMPILE_BIT; 1714 } 1715 //---------------------------format_trap_state--------------------------------- 1716 // This is used for debugging and diagnostics, including hotspot.log output. 1717 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1718 int trap_state) { 1719 DeoptReason reason = trap_state_reason(trap_state); 1720 bool recomp_flag = trap_state_is_recompiled(trap_state); 1721 // Re-encode the state from its decoded components. 1722 int decoded_state = 0; 1723 if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many) 1724 decoded_state = trap_state_add_reason(decoded_state, reason); 1725 if (recomp_flag) 1726 decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag); 1727 // If the state re-encodes properly, format it symbolically. 1728 // Because this routine is used for debugging and diagnostics, 1729 // be robust even if the state is a strange value. 1730 size_t len; 1731 if (decoded_state != trap_state) { 1732 // Random buggy state that doesn't decode?? 1733 len = jio_snprintf(buf, buflen, "#%d", trap_state); 1734 } else { 1735 len = jio_snprintf(buf, buflen, "%s%s", 1736 trap_reason_name(reason), 1737 recomp_flag ? " recompiled" : ""); 1738 } 1739 if (len >= buflen) 1740 buf[buflen-1] = '\0'; 1741 return buf; 1742 } 1743 1744 1745 //--------------------------------statics-------------------------------------- 1746 Deoptimization::DeoptAction Deoptimization::_unloaded_action 1747 = Deoptimization::Action_reinterpret; 1748 const char* Deoptimization::_trap_reason_name[Reason_LIMIT] = { 1749 // Note: Keep this in sync. with enum DeoptReason. 1750 "none", 1751 "null_check", 1752 "null_assert", 1753 "range_check", 1754 "class_check", 1755 "array_check", 1756 "intrinsic", 1757 "bimorphic", 1758 "unloaded", 1759 "uninitialized", 1760 "unreached", 1761 "unhandled", 1762 "constraint", 1763 "div0_check", 1764 "age", 1765 "predicate" 1766 }; 1767 const char* Deoptimization::_trap_action_name[Action_LIMIT] = { 1768 // Note: Keep this in sync. with enum DeoptAction. 1769 "none", 1770 "maybe_recompile", 1771 "reinterpret", 1772 "make_not_entrant", 1773 "make_not_compilable" 1774 }; 1775 1776 const char* Deoptimization::trap_reason_name(int reason) { 1777 if (reason == Reason_many) return "many"; 1778 if ((uint)reason < Reason_LIMIT) 1779 return _trap_reason_name[reason]; 1780 static char buf[20]; 1781 sprintf(buf, "reason%d", reason); 1782 return buf; 1783 } 1784 const char* Deoptimization::trap_action_name(int action) { 1785 if ((uint)action < Action_LIMIT) 1786 return _trap_action_name[action]; 1787 static char buf[20]; 1788 sprintf(buf, "action%d", action); 1789 return buf; 1790 } 1791 1792 // This is used for debugging and diagnostics, including hotspot.log output. 1793 const char* Deoptimization::format_trap_request(char* buf, size_t buflen, 1794 int trap_request) { 1795 jint unloaded_class_index = trap_request_index(trap_request); 1796 const char* reason = trap_reason_name(trap_request_reason(trap_request)); 1797 const char* action = trap_action_name(trap_request_action(trap_request)); 1798 size_t len; 1799 if (unloaded_class_index < 0) { 1800 len = jio_snprintf(buf, buflen, "reason='%s' action='%s'", 1801 reason, action); 1802 } else { 1803 len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'", 1804 reason, action, unloaded_class_index); 1805 } 1806 if (len >= buflen) 1807 buf[buflen-1] = '\0'; 1808 return buf; 1809 } 1810 1811 juint Deoptimization::_deoptimization_hist 1812 [Deoptimization::Reason_LIMIT] 1813 [1 + Deoptimization::Action_LIMIT] 1814 [Deoptimization::BC_CASE_LIMIT] 1815 = {0}; 1816 1817 enum { 1818 LSB_BITS = 8, 1819 LSB_MASK = right_n_bits(LSB_BITS) 1820 }; 1821 1822 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1823 Bytecodes::Code bc) { 1824 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1825 assert(action >= 0 && action < Action_LIMIT, "oob"); 1826 _deoptimization_hist[Reason_none][0][0] += 1; // total 1827 _deoptimization_hist[reason][0][0] += 1; // per-reason total 1828 juint* cases = _deoptimization_hist[reason][1+action]; 1829 juint* bc_counter_addr = NULL; 1830 juint bc_counter = 0; 1831 // Look for an unused counter, or an exact match to this BC. 1832 if (bc != Bytecodes::_illegal) { 1833 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1834 juint* counter_addr = &cases[bc_case]; 1835 juint counter = *counter_addr; 1836 if ((counter == 0 && bc_counter_addr == NULL) 1837 || (Bytecodes::Code)(counter & LSB_MASK) == bc) { 1838 // this counter is either free or is already devoted to this BC 1839 bc_counter_addr = counter_addr; 1840 bc_counter = counter | bc; 1841 } 1842 } 1843 } 1844 if (bc_counter_addr == NULL) { 1845 // Overflow, or no given bytecode. 1846 bc_counter_addr = &cases[BC_CASE_LIMIT-1]; 1847 bc_counter = (*bc_counter_addr & ~LSB_MASK); // clear LSB 1848 } 1849 *bc_counter_addr = bc_counter + (1 << LSB_BITS); 1850 } 1851 1852 jint Deoptimization::total_deoptimization_count() { 1853 return _deoptimization_hist[Reason_none][0][0]; 1854 } 1855 1856 jint Deoptimization::deoptimization_count(DeoptReason reason) { 1857 assert(reason >= 0 && reason < Reason_LIMIT, "oob"); 1858 return _deoptimization_hist[reason][0][0]; 1859 } 1860 1861 void Deoptimization::print_statistics() { 1862 juint total = total_deoptimization_count(); 1863 juint account = total; 1864 if (total != 0) { 1865 ttyLocker ttyl; 1866 if (xtty != NULL) xtty->head("statistics type='deoptimization'"); 1867 tty->print_cr("Deoptimization traps recorded:"); 1868 #define PRINT_STAT_LINE(name, r) \ 1869 tty->print_cr(" %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name); 1870 PRINT_STAT_LINE("total", total); 1871 // For each non-zero entry in the histogram, print the reason, 1872 // the action, and (if specifically known) the type of bytecode. 1873 for (int reason = 0; reason < Reason_LIMIT; reason++) { 1874 for (int action = 0; action < Action_LIMIT; action++) { 1875 juint* cases = _deoptimization_hist[reason][1+action]; 1876 for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) { 1877 juint counter = cases[bc_case]; 1878 if (counter != 0) { 1879 char name[1*K]; 1880 Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK); 1881 if (bc_case == BC_CASE_LIMIT && (int)bc == 0) 1882 bc = Bytecodes::_illegal; 1883 sprintf(name, "%s/%s/%s", 1884 trap_reason_name(reason), 1885 trap_action_name(action), 1886 Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other"); 1887 juint r = counter >> LSB_BITS; 1888 tty->print_cr(" %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total); 1889 account -= r; 1890 } 1891 } 1892 } 1893 } 1894 if (account != 0) { 1895 PRINT_STAT_LINE("unaccounted", account); 1896 } 1897 #undef PRINT_STAT_LINE 1898 if (xtty != NULL) xtty->tail("statistics"); 1899 } 1900 } 1901 #else // COMPILER2 || SHARK 1902 1903 1904 // Stubs for C1 only system. 1905 bool Deoptimization::trap_state_is_recompiled(int trap_state) { 1906 return false; 1907 } 1908 1909 const char* Deoptimization::trap_reason_name(int reason) { 1910 return "unknown"; 1911 } 1912 1913 void Deoptimization::print_statistics() { 1914 // no output 1915 } 1916 1917 void 1918 Deoptimization::update_method_data_from_interpreter(methodDataHandle trap_mdo, int trap_bci, int reason) { 1919 // no udpate 1920 } 1921 1922 int Deoptimization::trap_state_has_reason(int trap_state, int reason) { 1923 return 0; 1924 } 1925 1926 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action, 1927 Bytecodes::Code bc) { 1928 // no update 1929 } 1930 1931 const char* Deoptimization::format_trap_state(char* buf, size_t buflen, 1932 int trap_state) { 1933 jio_snprintf(buf, buflen, "#%d", trap_state); 1934 return buf; 1935 } 1936 1937 #endif // COMPILER2 || SHARK