1 /* 2 * Copyright 2001-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_graphKit.cpp.incl" 27 28 //----------------------------GraphKit----------------------------------------- 29 // Main utility constructor. 30 GraphKit::GraphKit(JVMState* jvms) 31 : Phase(Phase::Parser), 32 _env(C->env()), 33 _gvn(*C->initial_gvn()) 34 { 35 _exceptions = jvms->map()->next_exception(); 36 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 37 set_jvms(jvms); 38 } 39 40 // Private constructor for parser. 41 GraphKit::GraphKit() 42 : Phase(Phase::Parser), 43 _env(C->env()), 44 _gvn(*C->initial_gvn()) 45 { 46 _exceptions = NULL; 47 set_map(NULL); 48 debug_only(_sp = -99); 49 debug_only(set_bci(-99)); 50 } 51 52 53 54 //---------------------------clean_stack--------------------------------------- 55 // Clear away rubbish from the stack area of the JVM state. 56 // This destroys any arguments that may be waiting on the stack. 57 void GraphKit::clean_stack(int from_sp) { 58 SafePointNode* map = this->map(); 59 JVMState* jvms = this->jvms(); 60 int stk_size = jvms->stk_size(); 61 int stkoff = jvms->stkoff(); 62 Node* top = this->top(); 63 for (int i = from_sp; i < stk_size; i++) { 64 if (map->in(stkoff + i) != top) { 65 map->set_req(stkoff + i, top); 66 } 67 } 68 } 69 70 71 //--------------------------------sync_jvms----------------------------------- 72 // Make sure our current jvms agrees with our parse state. 73 JVMState* GraphKit::sync_jvms() const { 74 JVMState* jvms = this->jvms(); 75 jvms->set_bci(bci()); // Record the new bci in the JVMState 76 jvms->set_sp(sp()); // Record the new sp in the JVMState 77 assert(jvms_in_sync(), "jvms is now in sync"); 78 return jvms; 79 } 80 81 #ifdef ASSERT 82 bool GraphKit::jvms_in_sync() const { 83 Parse* parse = is_Parse(); 84 if (parse == NULL) { 85 if (bci() != jvms()->bci()) return false; 86 if (sp() != (int)jvms()->sp()) return false; 87 return true; 88 } 89 if (jvms()->method() != parse->method()) return false; 90 if (jvms()->bci() != parse->bci()) return false; 91 int jvms_sp = jvms()->sp(); 92 if (jvms_sp != parse->sp()) return false; 93 int jvms_depth = jvms()->depth(); 94 if (jvms_depth != parse->depth()) return false; 95 return true; 96 } 97 98 // Local helper checks for special internal merge points 99 // used to accumulate and merge exception states. 100 // They are marked by the region's in(0) edge being the map itself. 101 // Such merge points must never "escape" into the parser at large, 102 // until they have been handed to gvn.transform. 103 static bool is_hidden_merge(Node* reg) { 104 if (reg == NULL) return false; 105 if (reg->is_Phi()) { 106 reg = reg->in(0); 107 if (reg == NULL) return false; 108 } 109 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 110 } 111 112 void GraphKit::verify_map() const { 113 if (map() == NULL) return; // null map is OK 114 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 115 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 116 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 117 } 118 119 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 120 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 121 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 122 } 123 #endif 124 125 //---------------------------stop_and_kill_map--------------------------------- 126 // Set _map to NULL, signalling a stop to further bytecode execution. 127 // First smash the current map's control to a constant, to mark it dead. 128 void GraphKit::stop_and_kill_map() { 129 SafePointNode* dead_map = stop(); 130 if (dead_map != NULL) { 131 dead_map->disconnect_inputs(NULL); // Mark the map as killed. 132 assert(dead_map->is_killed(), "must be so marked"); 133 } 134 } 135 136 137 //--------------------------------stopped-------------------------------------- 138 // Tell if _map is NULL, or control is top. 139 bool GraphKit::stopped() { 140 if (map() == NULL) return true; 141 else if (control() == top()) return true; 142 else return false; 143 } 144 145 146 //-----------------------------has_ex_handler---------------------------------- 147 // Tell if this method or any caller method has exception handlers. 148 bool GraphKit::has_ex_handler() { 149 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 150 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 151 return true; 152 } 153 } 154 return false; 155 } 156 157 //------------------------------save_ex_oop------------------------------------ 158 // Save an exception without blowing stack contents or other JVM state. 159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 160 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 161 ex_map->add_req(ex_oop); 162 debug_only(verify_exception_state(ex_map)); 163 } 164 165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 166 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 167 Node* ex_oop = ex_map->in(ex_map->req()-1); 168 if (clear_it) ex_map->del_req(ex_map->req()-1); 169 return ex_oop; 170 } 171 172 //-----------------------------saved_ex_oop------------------------------------ 173 // Recover a saved exception from its map. 174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 175 return common_saved_ex_oop(ex_map, false); 176 } 177 178 //--------------------------clear_saved_ex_oop--------------------------------- 179 // Erase a previously saved exception from its map. 180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 181 return common_saved_ex_oop(ex_map, true); 182 } 183 184 #ifdef ASSERT 185 //---------------------------has_saved_ex_oop---------------------------------- 186 // Erase a previously saved exception from its map. 187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 188 return ex_map->req() == ex_map->jvms()->endoff()+1; 189 } 190 #endif 191 192 //-------------------------make_exception_state-------------------------------- 193 // Turn the current JVM state into an exception state, appending the ex_oop. 194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 195 sync_jvms(); 196 SafePointNode* ex_map = stop(); // do not manipulate this map any more 197 set_saved_ex_oop(ex_map, ex_oop); 198 return ex_map; 199 } 200 201 202 //--------------------------add_exception_state-------------------------------- 203 // Add an exception to my list of exceptions. 204 void GraphKit::add_exception_state(SafePointNode* ex_map) { 205 if (ex_map == NULL || ex_map->control() == top()) { 206 return; 207 } 208 #ifdef ASSERT 209 verify_exception_state(ex_map); 210 if (has_exceptions()) { 211 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 212 } 213 #endif 214 215 // If there is already an exception of exactly this type, merge with it. 216 // In particular, null-checks and other low-level exceptions common up here. 217 Node* ex_oop = saved_ex_oop(ex_map); 218 const Type* ex_type = _gvn.type(ex_oop); 219 if (ex_oop == top()) { 220 // No action needed. 221 return; 222 } 223 assert(ex_type->isa_instptr(), "exception must be an instance"); 224 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 225 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 226 // We check sp also because call bytecodes can generate exceptions 227 // both before and after arguments are popped! 228 if (ex_type2 == ex_type 229 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 230 combine_exception_states(ex_map, e2); 231 return; 232 } 233 } 234 235 // No pre-existing exception of the same type. Chain it on the list. 236 push_exception_state(ex_map); 237 } 238 239 //-----------------------add_exception_states_from----------------------------- 240 void GraphKit::add_exception_states_from(JVMState* jvms) { 241 SafePointNode* ex_map = jvms->map()->next_exception(); 242 if (ex_map != NULL) { 243 jvms->map()->set_next_exception(NULL); 244 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 245 next_map = ex_map->next_exception(); 246 ex_map->set_next_exception(NULL); 247 add_exception_state(ex_map); 248 } 249 } 250 } 251 252 //-----------------------transfer_exceptions_into_jvms------------------------- 253 JVMState* GraphKit::transfer_exceptions_into_jvms() { 254 if (map() == NULL) { 255 // We need a JVMS to carry the exceptions, but the map has gone away. 256 // Create a scratch JVMS, cloned from any of the exception states... 257 if (has_exceptions()) { 258 _map = _exceptions; 259 _map = clone_map(); 260 _map->set_next_exception(NULL); 261 clear_saved_ex_oop(_map); 262 debug_only(verify_map()); 263 } else { 264 // ...or created from scratch 265 JVMState* jvms = new (C) JVMState(_method, NULL); 266 jvms->set_bci(_bci); 267 jvms->set_sp(_sp); 268 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms)); 269 set_jvms(jvms); 270 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 271 set_all_memory(top()); 272 while (map()->req() < jvms->endoff()) map()->add_req(top()); 273 } 274 // (This is a kludge, in case you didn't notice.) 275 set_control(top()); 276 } 277 JVMState* jvms = sync_jvms(); 278 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 279 jvms->map()->set_next_exception(_exceptions); 280 _exceptions = NULL; // done with this set of exceptions 281 return jvms; 282 } 283 284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 285 assert(is_hidden_merge(dstphi), "must be a special merge node"); 286 assert(is_hidden_merge(srcphi), "must be a special merge node"); 287 uint limit = srcphi->req(); 288 for (uint i = PhiNode::Input; i < limit; i++) { 289 dstphi->add_req(srcphi->in(i)); 290 } 291 } 292 static inline void add_one_req(Node* dstphi, Node* src) { 293 assert(is_hidden_merge(dstphi), "must be a special merge node"); 294 assert(!is_hidden_merge(src), "must not be a special merge node"); 295 dstphi->add_req(src); 296 } 297 298 //-----------------------combine_exception_states------------------------------ 299 // This helper function combines exception states by building phis on a 300 // specially marked state-merging region. These regions and phis are 301 // untransformed, and can build up gradually. The region is marked by 302 // having a control input of its exception map, rather than NULL. Such 303 // regions do not appear except in this function, and in use_exception_state. 304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 305 if (failing()) return; // dying anyway... 306 JVMState* ex_jvms = ex_map->_jvms; 307 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 308 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 309 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 310 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 311 assert(ex_map->req() == phi_map->req(), "matching maps"); 312 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 313 Node* hidden_merge_mark = root(); 314 Node* region = phi_map->control(); 315 MergeMemNode* phi_mem = phi_map->merged_memory(); 316 MergeMemNode* ex_mem = ex_map->merged_memory(); 317 if (region->in(0) != hidden_merge_mark) { 318 // The control input is not (yet) a specially-marked region in phi_map. 319 // Make it so, and build some phis. 320 region = new (C, 2) RegionNode(2); 321 _gvn.set_type(region, Type::CONTROL); 322 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 323 region->init_req(1, phi_map->control()); 324 phi_map->set_control(region); 325 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 326 record_for_igvn(io_phi); 327 _gvn.set_type(io_phi, Type::ABIO); 328 phi_map->set_i_o(io_phi); 329 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 330 Node* m = mms.memory(); 331 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 332 record_for_igvn(m_phi); 333 _gvn.set_type(m_phi, Type::MEMORY); 334 mms.set_memory(m_phi); 335 } 336 } 337 338 // Either or both of phi_map and ex_map might already be converted into phis. 339 Node* ex_control = ex_map->control(); 340 // if there is special marking on ex_map also, we add multiple edges from src 341 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 342 // how wide was the destination phi_map, originally? 343 uint orig_width = region->req(); 344 345 if (add_multiple) { 346 add_n_reqs(region, ex_control); 347 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 348 } else { 349 // ex_map has no merges, so we just add single edges everywhere 350 add_one_req(region, ex_control); 351 add_one_req(phi_map->i_o(), ex_map->i_o()); 352 } 353 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 354 if (mms.is_empty()) { 355 // get a copy of the base memory, and patch some inputs into it 356 const TypePtr* adr_type = mms.adr_type(C); 357 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 358 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 359 mms.set_memory(phi); 360 // Prepare to append interesting stuff onto the newly sliced phi: 361 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 362 } 363 // Append stuff from ex_map: 364 if (add_multiple) { 365 add_n_reqs(mms.memory(), mms.memory2()); 366 } else { 367 add_one_req(mms.memory(), mms.memory2()); 368 } 369 } 370 uint limit = ex_map->req(); 371 for (uint i = TypeFunc::Parms; i < limit; i++) { 372 // Skip everything in the JVMS after tos. (The ex_oop follows.) 373 if (i == tos) i = ex_jvms->monoff(); 374 Node* src = ex_map->in(i); 375 Node* dst = phi_map->in(i); 376 if (src != dst) { 377 PhiNode* phi; 378 if (dst->in(0) != region) { 379 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 380 record_for_igvn(phi); 381 _gvn.set_type(phi, phi->type()); 382 phi_map->set_req(i, dst); 383 // Prepare to append interesting stuff onto the new phi: 384 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 385 } else { 386 assert(dst->is_Phi(), "nobody else uses a hidden region"); 387 phi = (PhiNode*)dst; 388 } 389 if (add_multiple && src->in(0) == ex_control) { 390 // Both are phis. 391 add_n_reqs(dst, src); 392 } else { 393 while (dst->req() < region->req()) add_one_req(dst, src); 394 } 395 const Type* srctype = _gvn.type(src); 396 if (phi->type() != srctype) { 397 const Type* dsttype = phi->type()->meet(srctype); 398 if (phi->type() != dsttype) { 399 phi->set_type(dsttype); 400 _gvn.set_type(phi, dsttype); 401 } 402 } 403 } 404 } 405 } 406 407 //--------------------------use_exception_state-------------------------------- 408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 409 if (failing()) { stop(); return top(); } 410 Node* region = phi_map->control(); 411 Node* hidden_merge_mark = root(); 412 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 413 Node* ex_oop = clear_saved_ex_oop(phi_map); 414 if (region->in(0) == hidden_merge_mark) { 415 // Special marking for internal ex-states. Process the phis now. 416 region->set_req(0, region); // now it's an ordinary region 417 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 418 // Note: Setting the jvms also sets the bci and sp. 419 set_control(_gvn.transform(region)); 420 uint tos = jvms()->stkoff() + sp(); 421 for (uint i = 1; i < tos; i++) { 422 Node* x = phi_map->in(i); 423 if (x->in(0) == region) { 424 assert(x->is_Phi(), "expected a special phi"); 425 phi_map->set_req(i, _gvn.transform(x)); 426 } 427 } 428 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 429 Node* x = mms.memory(); 430 if (x->in(0) == region) { 431 assert(x->is_Phi(), "nobody else uses a hidden region"); 432 mms.set_memory(_gvn.transform(x)); 433 } 434 } 435 if (ex_oop->in(0) == region) { 436 assert(ex_oop->is_Phi(), "expected a special phi"); 437 ex_oop = _gvn.transform(ex_oop); 438 } 439 } else { 440 set_jvms(phi_map->jvms()); 441 } 442 443 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 444 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 445 return ex_oop; 446 } 447 448 //---------------------------------java_bc------------------------------------- 449 Bytecodes::Code GraphKit::java_bc() const { 450 ciMethod* method = this->method(); 451 int bci = this->bci(); 452 if (method != NULL && bci != InvocationEntryBci) 453 return method->java_code_at_bci(bci); 454 else 455 return Bytecodes::_illegal; 456 } 457 458 //------------------------------builtin_throw---------------------------------- 459 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 460 bool must_throw = true; 461 462 if (env()->jvmti_can_post_exceptions()) { 463 // Do not try anything fancy if we're notifying the VM on every throw. 464 // Cf. case Bytecodes::_athrow in parse2.cpp. 465 uncommon_trap(reason, Deoptimization::Action_none, 466 (ciKlass*)NULL, (char*)NULL, must_throw); 467 return; 468 } 469 470 // If this particular condition has not yet happened at this 471 // bytecode, then use the uncommon trap mechanism, and allow for 472 // a future recompilation if several traps occur here. 473 // If the throw is hot, try to use a more complicated inline mechanism 474 // which keeps execution inside the compiled code. 475 bool treat_throw_as_hot = false; 476 ciMethodData* md = method()->method_data(); 477 478 if (ProfileTraps) { 479 if (too_many_traps(reason)) { 480 treat_throw_as_hot = true; 481 } 482 // (If there is no MDO at all, assume it is early in 483 // execution, and that any deopts are part of the 484 // startup transient, and don't need to be remembered.) 485 486 // Also, if there is a local exception handler, treat all throws 487 // as hot if there has been at least one in this method. 488 if (C->trap_count(reason) != 0 489 && method()->method_data()->trap_count(reason) != 0 490 && has_ex_handler()) { 491 treat_throw_as_hot = true; 492 } 493 } 494 495 // If this throw happens frequently, an uncommon trap might cause 496 // a performance pothole. If there is a local exception handler, 497 // and if this particular bytecode appears to be deoptimizing often, 498 // let us handle the throw inline, with a preconstructed instance. 499 // Note: If the deopt count has blown up, the uncommon trap 500 // runtime is going to flush this nmethod, not matter what. 501 if (treat_throw_as_hot 502 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 503 // If the throw is local, we use a pre-existing instance and 504 // punt on the backtrace. This would lead to a missing backtrace 505 // (a repeat of 4292742) if the backtrace object is ever asked 506 // for its backtrace. 507 // Fixing this remaining case of 4292742 requires some flavor of 508 // escape analysis. Leave that for the future. 509 ciInstance* ex_obj = NULL; 510 switch (reason) { 511 case Deoptimization::Reason_null_check: 512 ex_obj = env()->NullPointerException_instance(); 513 break; 514 case Deoptimization::Reason_div0_check: 515 ex_obj = env()->ArithmeticException_instance(); 516 break; 517 case Deoptimization::Reason_range_check: 518 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 519 break; 520 case Deoptimization::Reason_class_check: 521 if (java_bc() == Bytecodes::_aastore) { 522 ex_obj = env()->ArrayStoreException_instance(); 523 } else { 524 ex_obj = env()->ClassCastException_instance(); 525 } 526 break; 527 } 528 if (failing()) { stop(); return; } // exception allocation might fail 529 if (ex_obj != NULL) { 530 // Cheat with a preallocated exception object. 531 if (C->log() != NULL) 532 C->log()->elem("hot_throw preallocated='1' reason='%s'", 533 Deoptimization::trap_reason_name(reason)); 534 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 535 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 536 537 // Clear the detail message of the preallocated exception object. 538 // Weblogic sometimes mutates the detail message of exceptions 539 // using reflection. 540 int offset = java_lang_Throwable::get_detailMessage_offset(); 541 const TypePtr* adr_typ = ex_con->add_offset(offset); 542 543 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 544 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT); 545 546 add_exception_state(make_exception_state(ex_node)); 547 return; 548 } 549 } 550 551 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 552 // It won't be much cheaper than bailing to the interp., since we'll 553 // have to pass up all the debug-info, and the runtime will have to 554 // create the stack trace. 555 556 // Usual case: Bail to interpreter. 557 // Reserve the right to recompile if we haven't seen anything yet. 558 559 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 560 if (treat_throw_as_hot 561 && (method()->method_data()->trap_recompiled_at(bci()) 562 || C->too_many_traps(reason))) { 563 // We cannot afford to take more traps here. Suffer in the interpreter. 564 if (C->log() != NULL) 565 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 566 Deoptimization::trap_reason_name(reason), 567 C->trap_count(reason)); 568 action = Deoptimization::Action_none; 569 } 570 571 // "must_throw" prunes the JVM state to include only the stack, if there 572 // are no local exception handlers. This should cut down on register 573 // allocation time and code size, by drastically reducing the number 574 // of in-edges on the call to the uncommon trap. 575 576 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 577 } 578 579 580 //----------------------------PreserveJVMState--------------------------------- 581 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 582 debug_only(kit->verify_map()); 583 _kit = kit; 584 _map = kit->map(); // preserve the map 585 _sp = kit->sp(); 586 kit->set_map(clone_map ? kit->clone_map() : NULL); 587 #ifdef ASSERT 588 _bci = kit->bci(); 589 Parse* parser = kit->is_Parse(); 590 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 591 _block = block; 592 #endif 593 } 594 PreserveJVMState::~PreserveJVMState() { 595 GraphKit* kit = _kit; 596 #ifdef ASSERT 597 assert(kit->bci() == _bci, "bci must not shift"); 598 Parse* parser = kit->is_Parse(); 599 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 600 assert(block == _block, "block must not shift"); 601 #endif 602 kit->set_map(_map); 603 kit->set_sp(_sp); 604 } 605 606 607 //-----------------------------BuildCutout------------------------------------- 608 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 609 : PreserveJVMState(kit) 610 { 611 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 612 SafePointNode* outer_map = _map; // preserved map is caller's 613 SafePointNode* inner_map = kit->map(); 614 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 615 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) )); 616 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) )); 617 } 618 BuildCutout::~BuildCutout() { 619 GraphKit* kit = _kit; 620 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 621 } 622 623 //---------------------------PreserveReexecuteState---------------------------- 624 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 625 _kit = kit; 626 _sp = kit->sp(); 627 _reexecute = kit->jvms()->should_reexecute(); 628 } 629 PreserveReexecuteState::~PreserveReexecuteState() { 630 _kit->jvms()->set_reexecute(_reexecute); 631 _kit->set_sp(_sp); 632 } 633 634 //------------------------------clone_map-------------------------------------- 635 // Implementation of PreserveJVMState 636 // 637 // Only clone_map(...) here. If this function is only used in the 638 // PreserveJVMState class we may want to get rid of this extra 639 // function eventually and do it all there. 640 641 SafePointNode* GraphKit::clone_map() { 642 if (map() == NULL) return NULL; 643 644 // Clone the memory edge first 645 Node* mem = MergeMemNode::make(C, map()->memory()); 646 gvn().set_type_bottom(mem); 647 648 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 649 JVMState* jvms = this->jvms(); 650 JVMState* clonejvms = jvms->clone_shallow(C); 651 clonemap->set_memory(mem); 652 clonemap->set_jvms(clonejvms); 653 clonejvms->set_map(clonemap); 654 record_for_igvn(clonemap); 655 gvn().set_type_bottom(clonemap); 656 return clonemap; 657 } 658 659 660 //-----------------------------set_map_clone----------------------------------- 661 void GraphKit::set_map_clone(SafePointNode* m) { 662 _map = m; 663 _map = clone_map(); 664 _map->set_next_exception(NULL); 665 debug_only(verify_map()); 666 } 667 668 669 //----------------------------kill_dead_locals--------------------------------- 670 // Detect any locals which are known to be dead, and force them to top. 671 void GraphKit::kill_dead_locals() { 672 // Consult the liveness information for the locals. If any 673 // of them are unused, then they can be replaced by top(). This 674 // should help register allocation time and cut down on the size 675 // of the deoptimization information. 676 677 // This call is made from many of the bytecode handling 678 // subroutines called from the Big Switch in do_one_bytecode. 679 // Every bytecode which might include a slow path is responsible 680 // for killing its dead locals. The more consistent we 681 // are about killing deads, the fewer useless phis will be 682 // constructed for them at various merge points. 683 684 // bci can be -1 (InvocationEntryBci). We return the entry 685 // liveness for the method. 686 687 if (method() == NULL || method()->code_size() == 0) { 688 // We are building a graph for a call to a native method. 689 // All locals are live. 690 return; 691 } 692 693 ResourceMark rm; 694 695 // Consult the liveness information for the locals. If any 696 // of them are unused, then they can be replaced by top(). This 697 // should help register allocation time and cut down on the size 698 // of the deoptimization information. 699 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 700 701 int len = (int)live_locals.size(); 702 assert(len <= jvms()->loc_size(), "too many live locals"); 703 for (int local = 0; local < len; local++) { 704 if (!live_locals.at(local)) { 705 set_local(local, top()); 706 } 707 } 708 } 709 710 #ifdef ASSERT 711 //-------------------------dead_locals_are_killed------------------------------ 712 // Return true if all dead locals are set to top in the map. 713 // Used to assert "clean" debug info at various points. 714 bool GraphKit::dead_locals_are_killed() { 715 if (method() == NULL || method()->code_size() == 0) { 716 // No locals need to be dead, so all is as it should be. 717 return true; 718 } 719 720 // Make sure somebody called kill_dead_locals upstream. 721 ResourceMark rm; 722 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 723 if (jvms->loc_size() == 0) continue; // no locals to consult 724 SafePointNode* map = jvms->map(); 725 ciMethod* method = jvms->method(); 726 int bci = jvms->bci(); 727 if (jvms == this->jvms()) { 728 bci = this->bci(); // it might not yet be synched 729 } 730 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 731 int len = (int)live_locals.size(); 732 if (!live_locals.is_valid() || len == 0) 733 // This method is trivial, or is poisoned by a breakpoint. 734 return true; 735 assert(len == jvms->loc_size(), "live map consistent with locals map"); 736 for (int local = 0; local < len; local++) { 737 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 738 if (PrintMiscellaneous && (Verbose || WizardMode)) { 739 tty->print_cr("Zombie local %d: ", local); 740 jvms->dump(); 741 } 742 return false; 743 } 744 } 745 } 746 return true; 747 } 748 749 #endif //ASSERT 750 751 // Helper function for enforcing certain bytecodes to reexecute if 752 // deoptimization happens 753 static bool should_reexecute_implied_by_bytecode(JVMState *jvms) { 754 ciMethod* cur_method = jvms->method(); 755 int cur_bci = jvms->bci(); 756 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 757 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 758 return Interpreter::bytecode_should_reexecute(code); 759 } else 760 return false; 761 } 762 763 // Helper function for adding JVMState and debug information to node 764 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 765 // Add the safepoint edges to the call (or other safepoint). 766 767 // Make sure dead locals are set to top. This 768 // should help register allocation time and cut down on the size 769 // of the deoptimization information. 770 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 771 772 // Walk the inline list to fill in the correct set of JVMState's 773 // Also fill in the associated edges for each JVMState. 774 775 JVMState* youngest_jvms = sync_jvms(); 776 777 // Do we need debug info here? If it is a SafePoint and this method 778 // cannot de-opt, then we do NOT need any debug info. 779 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint); 780 781 // If we are guaranteed to throw, we can prune everything but the 782 // input to the current bytecode. 783 bool can_prune_locals = false; 784 uint stack_slots_not_pruned = 0; 785 int inputs = 0, depth = 0; 786 if (must_throw) { 787 assert(method() == youngest_jvms->method(), "sanity"); 788 if (compute_stack_effects(inputs, depth)) { 789 can_prune_locals = true; 790 stack_slots_not_pruned = inputs; 791 } 792 } 793 794 if (env()->jvmti_can_examine_or_deopt_anywhere()) { 795 // At any safepoint, this method can get breakpointed, which would 796 // then require an immediate deoptimization. 797 full_info = true; 798 can_prune_locals = false; // do not prune locals 799 stack_slots_not_pruned = 0; 800 } 801 802 // do not scribble on the input jvms 803 JVMState* out_jvms = youngest_jvms->clone_deep(C); 804 call->set_jvms(out_jvms); // Start jvms list for call node 805 806 // For a known set of bytecodes, the interpreter should reexecute them if 807 // deoptimization happens. We set the reexecute bit for them here 808 if (out_jvms->should_reexecute() == JVMState::RE_Undefined && //don't touch defined values 809 should_reexecute_implied_by_bytecode(out_jvms)) { 810 out_jvms->set_reexecute(JVMState::RE_True); 811 } 812 813 // Presize the call: 814 debug_only(uint non_debug_edges = call->req()); 815 call->add_req_batch(top(), youngest_jvms->debug_depth()); 816 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 817 818 // Set up edges so that the call looks like this: 819 // Call [state:] ctl io mem fptr retadr 820 // [parms:] parm0 ... parmN 821 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 822 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 823 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 824 // Note that caller debug info precedes callee debug info. 825 826 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 827 uint debug_ptr = call->req(); 828 829 // Loop over the map input edges associated with jvms, add them 830 // to the call node, & reset all offsets to match call node array. 831 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 832 uint debug_end = debug_ptr; 833 uint debug_start = debug_ptr - in_jvms->debug_size(); 834 debug_ptr = debug_start; // back up the ptr 835 836 uint p = debug_start; // walks forward in [debug_start, debug_end) 837 uint j, k, l; 838 SafePointNode* in_map = in_jvms->map(); 839 out_jvms->set_map(call); 840 841 if (can_prune_locals) { 842 assert(in_jvms->method() == out_jvms->method(), "sanity"); 843 // If the current throw can reach an exception handler in this JVMS, 844 // then we must keep everything live that can reach that handler. 845 // As a quick and dirty approximation, we look for any handlers at all. 846 if (in_jvms->method()->has_exception_handlers()) { 847 can_prune_locals = false; 848 } 849 } 850 851 // Add the Locals 852 k = in_jvms->locoff(); 853 l = in_jvms->loc_size(); 854 out_jvms->set_locoff(p); 855 if (full_info && !can_prune_locals) { 856 for (j = 0; j < l; j++) 857 call->set_req(p++, in_map->in(k+j)); 858 } else { 859 p += l; // already set to top above by add_req_batch 860 } 861 862 // Add the Expression Stack 863 k = in_jvms->stkoff(); 864 l = in_jvms->sp(); 865 out_jvms->set_stkoff(p); 866 if (full_info && !can_prune_locals) { 867 for (j = 0; j < l; j++) 868 call->set_req(p++, in_map->in(k+j)); 869 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 870 // Divide stack into {S0,...,S1}, where S0 is set to top. 871 uint s1 = stack_slots_not_pruned; 872 stack_slots_not_pruned = 0; // for next iteration 873 if (s1 > l) s1 = l; 874 uint s0 = l - s1; 875 p += s0; // skip the tops preinstalled by add_req_batch 876 for (j = s0; j < l; j++) 877 call->set_req(p++, in_map->in(k+j)); 878 } else { 879 p += l; // already set to top above by add_req_batch 880 } 881 882 // Add the Monitors 883 k = in_jvms->monoff(); 884 l = in_jvms->mon_size(); 885 out_jvms->set_monoff(p); 886 for (j = 0; j < l; j++) 887 call->set_req(p++, in_map->in(k+j)); 888 889 // Copy any scalar object fields. 890 k = in_jvms->scloff(); 891 l = in_jvms->scl_size(); 892 out_jvms->set_scloff(p); 893 for (j = 0; j < l; j++) 894 call->set_req(p++, in_map->in(k+j)); 895 896 // Finish the new jvms. 897 out_jvms->set_endoff(p); 898 899 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 900 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 901 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 902 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 903 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 904 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 905 906 // Update the two tail pointers in parallel. 907 out_jvms = out_jvms->caller(); 908 in_jvms = in_jvms->caller(); 909 } 910 911 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 912 913 // Test the correctness of JVMState::debug_xxx accessors: 914 assert(call->jvms()->debug_start() == non_debug_edges, ""); 915 assert(call->jvms()->debug_end() == call->req(), ""); 916 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 917 } 918 919 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 920 Bytecodes::Code code = java_bc(); 921 if (code == Bytecodes::_wide) { 922 code = method()->java_code_at_bci(bci() + 1); 923 } 924 925 BasicType rtype = T_ILLEGAL; 926 int rsize = 0; 927 928 if (code != Bytecodes::_illegal) { 929 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 930 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 931 if (rtype < T_CONFLICT) 932 rsize = type2size[rtype]; 933 } 934 935 switch (code) { 936 case Bytecodes::_illegal: 937 return false; 938 939 case Bytecodes::_ldc: 940 case Bytecodes::_ldc_w: 941 case Bytecodes::_ldc2_w: 942 inputs = 0; 943 break; 944 945 case Bytecodes::_dup: inputs = 1; break; 946 case Bytecodes::_dup_x1: inputs = 2; break; 947 case Bytecodes::_dup_x2: inputs = 3; break; 948 case Bytecodes::_dup2: inputs = 2; break; 949 case Bytecodes::_dup2_x1: inputs = 3; break; 950 case Bytecodes::_dup2_x2: inputs = 4; break; 951 case Bytecodes::_swap: inputs = 2; break; 952 case Bytecodes::_arraylength: inputs = 1; break; 953 954 case Bytecodes::_getstatic: 955 case Bytecodes::_putstatic: 956 case Bytecodes::_getfield: 957 case Bytecodes::_putfield: 958 { 959 bool is_get = (depth >= 0), is_static = (depth & 1); 960 bool ignore; 961 ciBytecodeStream iter(method()); 962 iter.reset_to_bci(bci()); 963 iter.next(); 964 ciField* field = iter.get_field(ignore); 965 int size = field->type()->size(); 966 inputs = (is_static ? 0 : 1); 967 if (is_get) { 968 depth = size - inputs; 969 } else { 970 inputs += size; // putxxx pops the value from the stack 971 depth = - inputs; 972 } 973 } 974 break; 975 976 case Bytecodes::_invokevirtual: 977 case Bytecodes::_invokespecial: 978 case Bytecodes::_invokestatic: 979 case Bytecodes::_invokedynamic: 980 case Bytecodes::_invokeinterface: 981 { 982 bool is_static = (depth == 0); 983 bool ignore; 984 ciBytecodeStream iter(method()); 985 iter.reset_to_bci(bci()); 986 iter.next(); 987 ciMethod* method = iter.get_method(ignore); 988 inputs = method->arg_size_no_receiver(); 989 if (!is_static) inputs += 1; 990 int size = method->return_type()->size(); 991 depth = size - inputs; 992 } 993 break; 994 995 case Bytecodes::_multianewarray: 996 { 997 ciBytecodeStream iter(method()); 998 iter.reset_to_bci(bci()); 999 iter.next(); 1000 inputs = iter.get_dimensions(); 1001 assert(rsize == 1, ""); 1002 depth = rsize - inputs; 1003 } 1004 break; 1005 1006 case Bytecodes::_ireturn: 1007 case Bytecodes::_lreturn: 1008 case Bytecodes::_freturn: 1009 case Bytecodes::_dreturn: 1010 case Bytecodes::_areturn: 1011 assert(rsize = -depth, ""); 1012 inputs = rsize; 1013 break; 1014 1015 case Bytecodes::_jsr: 1016 case Bytecodes::_jsr_w: 1017 inputs = 0; 1018 depth = 1; // S.B. depth=1, not zero 1019 break; 1020 1021 default: 1022 // bytecode produces a typed result 1023 inputs = rsize - depth; 1024 assert(inputs >= 0, ""); 1025 break; 1026 } 1027 1028 #ifdef ASSERT 1029 // spot check 1030 int outputs = depth + inputs; 1031 assert(outputs >= 0, "sanity"); 1032 switch (code) { 1033 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1034 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1035 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1036 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1037 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1038 } 1039 #endif //ASSERT 1040 1041 return true; 1042 } 1043 1044 1045 1046 //------------------------------basic_plus_adr--------------------------------- 1047 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1048 // short-circuit a common case 1049 if (offset == intcon(0)) return ptr; 1050 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); 1051 } 1052 1053 Node* GraphKit::ConvI2L(Node* offset) { 1054 // short-circuit a common case 1055 jint offset_con = find_int_con(offset, Type::OffsetBot); 1056 if (offset_con != Type::OffsetBot) { 1057 return longcon((long) offset_con); 1058 } 1059 return _gvn.transform( new (C, 2) ConvI2LNode(offset)); 1060 } 1061 Node* GraphKit::ConvL2I(Node* offset) { 1062 // short-circuit a common case 1063 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1064 if (offset_con != (jlong)Type::OffsetBot) { 1065 return intcon((int) offset_con); 1066 } 1067 return _gvn.transform( new (C, 2) ConvL2INode(offset)); 1068 } 1069 1070 //-------------------------load_object_klass----------------------------------- 1071 Node* GraphKit::load_object_klass(Node* obj) { 1072 // Special-case a fresh allocation to avoid building nodes: 1073 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1074 if (akls != NULL) return akls; 1075 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1076 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1077 } 1078 1079 //-------------------------load_array_length----------------------------------- 1080 Node* GraphKit::load_array_length(Node* array) { 1081 // Special-case a fresh allocation to avoid building nodes: 1082 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1083 Node *alen; 1084 if (alloc == NULL) { 1085 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1086 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1087 } else { 1088 alen = alloc->Ideal_length(); 1089 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_aryptr(), &_gvn); 1090 if (ccast != alen) { 1091 alen = _gvn.transform(ccast); 1092 } 1093 } 1094 return alen; 1095 } 1096 1097 //------------------------------do_null_check---------------------------------- 1098 // Helper function to do a NULL pointer check. Returned value is 1099 // the incoming address with NULL casted away. You are allowed to use the 1100 // not-null value only if you are control dependent on the test. 1101 extern int explicit_null_checks_inserted, 1102 explicit_null_checks_elided; 1103 Node* GraphKit::null_check_common(Node* value, BasicType type, 1104 // optional arguments for variations: 1105 bool assert_null, 1106 Node* *null_control) { 1107 assert(!assert_null || null_control == NULL, "not both at once"); 1108 if (stopped()) return top(); 1109 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1110 // For some performance testing, we may wish to suppress null checking. 1111 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1112 return value; 1113 } 1114 explicit_null_checks_inserted++; 1115 1116 // Construct NULL check 1117 Node *chk = NULL; 1118 switch(type) { 1119 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1120 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; 1121 case T_ARRAY : // fall through 1122 type = T_OBJECT; // simplify further tests 1123 case T_OBJECT : { 1124 const Type *t = _gvn.type( value ); 1125 1126 const TypeInstPtr* tp = t->isa_instptr(); 1127 if (tp != NULL && !tp->klass()->is_loaded() 1128 // Only for do_null_check, not any of its siblings: 1129 && !assert_null && null_control == NULL) { 1130 // Usually, any field access or invocation on an unloaded oop type 1131 // will simply fail to link, since the statically linked class is 1132 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1133 // the static class is loaded but the sharper oop type is not. 1134 // Rather than checking for this obscure case in lots of places, 1135 // we simply observe that a null check on an unloaded class 1136 // will always be followed by a nonsense operation, so we 1137 // can just issue the uncommon trap here. 1138 // Our access to the unloaded class will only be correct 1139 // after it has been loaded and initialized, which requires 1140 // a trip through the interpreter. 1141 #ifndef PRODUCT 1142 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1143 #endif 1144 uncommon_trap(Deoptimization::Reason_unloaded, 1145 Deoptimization::Action_reinterpret, 1146 tp->klass(), "!loaded"); 1147 return top(); 1148 } 1149 1150 if (assert_null) { 1151 // See if the type is contained in NULL_PTR. 1152 // If so, then the value is already null. 1153 if (t->higher_equal(TypePtr::NULL_PTR)) { 1154 explicit_null_checks_elided++; 1155 return value; // Elided null assert quickly! 1156 } 1157 } else { 1158 // See if mixing in the NULL pointer changes type. 1159 // If so, then the NULL pointer was not allowed in the original 1160 // type. In other words, "value" was not-null. 1161 if (t->meet(TypePtr::NULL_PTR) != t) { 1162 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1163 explicit_null_checks_elided++; 1164 return value; // Elided null check quickly! 1165 } 1166 } 1167 chk = new (C, 3) CmpPNode( value, null() ); 1168 break; 1169 } 1170 1171 default : ShouldNotReachHere(); 1172 } 1173 assert(chk != NULL, "sanity check"); 1174 chk = _gvn.transform(chk); 1175 1176 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1177 BoolNode *btst = new (C, 2) BoolNode( chk, btest); 1178 Node *tst = _gvn.transform( btst ); 1179 1180 //----------- 1181 // if peephole optimizations occurred, a prior test existed. 1182 // If a prior test existed, maybe it dominates as we can avoid this test. 1183 if (tst != btst && type == T_OBJECT) { 1184 // At this point we want to scan up the CFG to see if we can 1185 // find an identical test (and so avoid this test altogether). 1186 Node *cfg = control(); 1187 int depth = 0; 1188 while( depth < 16 ) { // Limit search depth for speed 1189 if( cfg->Opcode() == Op_IfTrue && 1190 cfg->in(0)->in(1) == tst ) { 1191 // Found prior test. Use "cast_not_null" to construct an identical 1192 // CastPP (and hence hash to) as already exists for the prior test. 1193 // Return that casted value. 1194 if (assert_null) { 1195 replace_in_map(value, null()); 1196 return null(); // do not issue the redundant test 1197 } 1198 Node *oldcontrol = control(); 1199 set_control(cfg); 1200 Node *res = cast_not_null(value); 1201 set_control(oldcontrol); 1202 explicit_null_checks_elided++; 1203 return res; 1204 } 1205 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1206 if (cfg == NULL) break; // Quit at region nodes 1207 depth++; 1208 } 1209 } 1210 1211 //----------- 1212 // Branch to failure if null 1213 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1214 Deoptimization::DeoptReason reason; 1215 if (assert_null) 1216 reason = Deoptimization::Reason_null_assert; 1217 else if (type == T_OBJECT) 1218 reason = Deoptimization::Reason_null_check; 1219 else 1220 reason = Deoptimization::Reason_div0_check; 1221 1222 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1223 // ciMethodData::has_trap_at will return a conservative -1 if any 1224 // must-be-null assertion has failed. This could cause performance 1225 // problems for a method after its first do_null_assert failure. 1226 // Consider using 'Reason_class_check' instead? 1227 1228 // To cause an implicit null check, we set the not-null probability 1229 // to the maximum (PROB_MAX). For an explicit check the probability 1230 // is set to a smaller value. 1231 if (null_control != NULL || too_many_traps(reason)) { 1232 // probability is less likely 1233 ok_prob = PROB_LIKELY_MAG(3); 1234 } else if (!assert_null && 1235 (ImplicitNullCheckThreshold > 0) && 1236 method() != NULL && 1237 (method()->method_data()->trap_count(reason) 1238 >= (uint)ImplicitNullCheckThreshold)) { 1239 ok_prob = PROB_LIKELY_MAG(3); 1240 } 1241 1242 if (null_control != NULL) { 1243 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1244 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); 1245 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); 1246 if (null_true == top()) 1247 explicit_null_checks_elided++; 1248 (*null_control) = null_true; 1249 } else { 1250 BuildCutout unless(this, tst, ok_prob); 1251 // Check for optimizer eliding test at parse time 1252 if (stopped()) { 1253 // Failure not possible; do not bother making uncommon trap. 1254 explicit_null_checks_elided++; 1255 } else if (assert_null) { 1256 uncommon_trap(reason, 1257 Deoptimization::Action_make_not_entrant, 1258 NULL, "assert_null"); 1259 } else { 1260 replace_in_map(value, zerocon(type)); 1261 builtin_throw(reason); 1262 } 1263 } 1264 1265 // Must throw exception, fall-thru not possible? 1266 if (stopped()) { 1267 return top(); // No result 1268 } 1269 1270 if (assert_null) { 1271 // Cast obj to null on this path. 1272 replace_in_map(value, zerocon(type)); 1273 return zerocon(type); 1274 } 1275 1276 // Cast obj to not-null on this path, if there is no null_control. 1277 // (If there is a null_control, a non-null value may come back to haunt us.) 1278 if (type == T_OBJECT) { 1279 Node* cast = cast_not_null(value, false); 1280 if (null_control == NULL || (*null_control) == top()) 1281 replace_in_map(value, cast); 1282 value = cast; 1283 } 1284 1285 return value; 1286 } 1287 1288 1289 //------------------------------cast_not_null---------------------------------- 1290 // Cast obj to not-null on this path 1291 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1292 const Type *t = _gvn.type(obj); 1293 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1294 // Object is already not-null? 1295 if( t == t_not_null ) return obj; 1296 1297 Node *cast = new (C, 2) CastPPNode(obj,t_not_null); 1298 cast->init_req(0, control()); 1299 cast = _gvn.transform( cast ); 1300 1301 // Scan for instances of 'obj' in the current JVM mapping. 1302 // These instances are known to be not-null after the test. 1303 if (do_replace_in_map) 1304 replace_in_map(obj, cast); 1305 1306 return cast; // Return casted value 1307 } 1308 1309 1310 //--------------------------replace_in_map------------------------------------- 1311 void GraphKit::replace_in_map(Node* old, Node* neww) { 1312 this->map()->replace_edge(old, neww); 1313 1314 // Note: This operation potentially replaces any edge 1315 // on the map. This includes locals, stack, and monitors 1316 // of the current (innermost) JVM state. 1317 1318 // We can consider replacing in caller maps. 1319 // The idea would be that an inlined function's null checks 1320 // can be shared with the entire inlining tree. 1321 // The expense of doing this is that the PreserveJVMState class 1322 // would have to preserve caller states too, with a deep copy. 1323 } 1324 1325 1326 1327 //============================================================================= 1328 //--------------------------------memory--------------------------------------- 1329 Node* GraphKit::memory(uint alias_idx) { 1330 MergeMemNode* mem = merged_memory(); 1331 Node* p = mem->memory_at(alias_idx); 1332 _gvn.set_type(p, Type::MEMORY); // must be mapped 1333 return p; 1334 } 1335 1336 //-----------------------------reset_memory------------------------------------ 1337 Node* GraphKit::reset_memory() { 1338 Node* mem = map()->memory(); 1339 // do not use this node for any more parsing! 1340 debug_only( map()->set_memory((Node*)NULL) ); 1341 return _gvn.transform( mem ); 1342 } 1343 1344 //------------------------------set_all_memory--------------------------------- 1345 void GraphKit::set_all_memory(Node* newmem) { 1346 Node* mergemem = MergeMemNode::make(C, newmem); 1347 gvn().set_type_bottom(mergemem); 1348 map()->set_memory(mergemem); 1349 } 1350 1351 //------------------------------set_all_memory_call---------------------------- 1352 void GraphKit::set_all_memory_call(Node* call) { 1353 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1354 set_all_memory(newmem); 1355 } 1356 1357 //============================================================================= 1358 // 1359 // parser factory methods for MemNodes 1360 // 1361 // These are layered on top of the factory methods in LoadNode and StoreNode, 1362 // and integrate with the parser's memory state and _gvn engine. 1363 // 1364 1365 // factory methods in "int adr_idx" 1366 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1367 int adr_idx, 1368 bool require_atomic_access) { 1369 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1370 const TypePtr* adr_type = NULL; // debug-mode-only argument 1371 debug_only(adr_type = C->get_adr_type(adr_idx)); 1372 Node* mem = memory(adr_idx); 1373 Node* ld; 1374 if (require_atomic_access && bt == T_LONG) { 1375 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); 1376 } else { 1377 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); 1378 } 1379 return _gvn.transform(ld); 1380 } 1381 1382 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1383 int adr_idx, 1384 bool require_atomic_access) { 1385 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1386 const TypePtr* adr_type = NULL; 1387 debug_only(adr_type = C->get_adr_type(adr_idx)); 1388 Node *mem = memory(adr_idx); 1389 Node* st; 1390 if (require_atomic_access && bt == T_LONG) { 1391 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); 1392 } else { 1393 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); 1394 } 1395 st = _gvn.transform(st); 1396 set_memory(st, adr_idx); 1397 // Back-to-back stores can only remove intermediate store with DU info 1398 // so push on worklist for optimizer. 1399 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1400 record_for_igvn(st); 1401 1402 return st; 1403 } 1404 1405 void GraphKit::pre_barrier(Node* ctl, 1406 Node* obj, 1407 Node* adr, 1408 uint adr_idx, 1409 Node *val, 1410 const TypeOopPtr* val_type, 1411 BasicType bt) { 1412 BarrierSet* bs = Universe::heap()->barrier_set(); 1413 set_control(ctl); 1414 switch (bs->kind()) { 1415 case BarrierSet::G1SATBCT: 1416 case BarrierSet::G1SATBCTLogging: 1417 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt); 1418 break; 1419 1420 case BarrierSet::CardTableModRef: 1421 case BarrierSet::CardTableExtension: 1422 case BarrierSet::ModRef: 1423 break; 1424 1425 case BarrierSet::Other: 1426 default : 1427 ShouldNotReachHere(); 1428 1429 } 1430 } 1431 1432 void GraphKit::post_barrier(Node* ctl, 1433 Node* store, 1434 Node* obj, 1435 Node* adr, 1436 uint adr_idx, 1437 Node *val, 1438 BasicType bt, 1439 bool use_precise) { 1440 BarrierSet* bs = Universe::heap()->barrier_set(); 1441 set_control(ctl); 1442 switch (bs->kind()) { 1443 case BarrierSet::G1SATBCT: 1444 case BarrierSet::G1SATBCTLogging: 1445 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1446 break; 1447 1448 case BarrierSet::CardTableModRef: 1449 case BarrierSet::CardTableExtension: 1450 write_barrier_post(store, obj, adr, val, use_precise); 1451 break; 1452 1453 case BarrierSet::ModRef: 1454 break; 1455 1456 case BarrierSet::Other: 1457 default : 1458 ShouldNotReachHere(); 1459 1460 } 1461 } 1462 1463 Node* GraphKit::store_oop_to_object(Node* ctl, 1464 Node* obj, 1465 Node* adr, 1466 const TypePtr* adr_type, 1467 Node *val, 1468 const TypeOopPtr* val_type, 1469 BasicType bt) { 1470 uint adr_idx = C->get_alias_index(adr_type); 1471 Node* store; 1472 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt); 1473 store = store_to_memory(control(), adr, val, bt, adr_idx); 1474 post_barrier(control(), store, obj, adr, adr_idx, val, bt, false); 1475 return store; 1476 } 1477 1478 Node* GraphKit::store_oop_to_array(Node* ctl, 1479 Node* obj, 1480 Node* adr, 1481 const TypePtr* adr_type, 1482 Node *val, 1483 const TypeOopPtr* val_type, 1484 BasicType bt) { 1485 uint adr_idx = C->get_alias_index(adr_type); 1486 Node* store; 1487 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt); 1488 store = store_to_memory(control(), adr, val, bt, adr_idx); 1489 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true); 1490 return store; 1491 } 1492 1493 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1494 Node* obj, 1495 Node* adr, 1496 const TypePtr* adr_type, 1497 Node *val, 1498 BasicType bt) { 1499 Compile::AliasType* at = C->alias_type(adr_type); 1500 const TypeOopPtr* val_type = NULL; 1501 if (adr_type->isa_instptr()) { 1502 if (at->field() != NULL) { 1503 // known field. This code is a copy of the do_put_xxx logic. 1504 ciField* field = at->field(); 1505 if (!field->type()->is_loaded()) { 1506 val_type = TypeInstPtr::BOTTOM; 1507 } else { 1508 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1509 } 1510 } 1511 } else if (adr_type->isa_aryptr()) { 1512 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1513 } 1514 if (val_type == NULL) { 1515 val_type = TypeInstPtr::BOTTOM; 1516 } 1517 1518 uint adr_idx = at->index(); 1519 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt); 1520 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1521 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true); 1522 return store; 1523 } 1524 1525 1526 //-------------------------array_element_address------------------------- 1527 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1528 const TypeInt* sizetype) { 1529 uint shift = exact_log2(type2aelembytes(elembt)); 1530 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1531 1532 // short-circuit a common case (saves lots of confusing waste motion) 1533 jint idx_con = find_int_con(idx, -1); 1534 if (idx_con >= 0) { 1535 intptr_t offset = header + ((intptr_t)idx_con << shift); 1536 return basic_plus_adr(ary, offset); 1537 } 1538 1539 // must be correct type for alignment purposes 1540 Node* base = basic_plus_adr(ary, header); 1541 #ifdef _LP64 1542 // The scaled index operand to AddP must be a clean 64-bit value. 1543 // Java allows a 32-bit int to be incremented to a negative 1544 // value, which appears in a 64-bit register as a large 1545 // positive number. Using that large positive number as an 1546 // operand in pointer arithmetic has bad consequences. 1547 // On the other hand, 32-bit overflow is rare, and the possibility 1548 // can often be excluded, if we annotate the ConvI2L node with 1549 // a type assertion that its value is known to be a small positive 1550 // number. (The prior range check has ensured this.) 1551 // This assertion is used by ConvI2LNode::Ideal. 1552 int index_max = max_jint - 1; // array size is max_jint, index is one less 1553 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1554 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1555 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); 1556 #endif 1557 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); 1558 return basic_plus_adr(ary, base, scale); 1559 } 1560 1561 //-------------------------load_array_element------------------------- 1562 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1563 const Type* elemtype = arytype->elem(); 1564 BasicType elembt = elemtype->array_element_basic_type(); 1565 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1566 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1567 return ld; 1568 } 1569 1570 //-------------------------set_arguments_for_java_call------------------------- 1571 // Arguments (pre-popped from the stack) are taken from the JVMS. 1572 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1573 // Add the call arguments: 1574 uint nargs = call->method()->arg_size(); 1575 for (uint i = 0; i < nargs; i++) { 1576 Node* arg = argument(i); 1577 call->init_req(i + TypeFunc::Parms, arg); 1578 } 1579 } 1580 1581 //---------------------------set_edges_for_java_call--------------------------- 1582 // Connect a newly created call into the current JVMS. 1583 // A return value node (if any) is returned from set_edges_for_java_call. 1584 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) { 1585 1586 // Add the predefined inputs: 1587 call->init_req( TypeFunc::Control, control() ); 1588 call->init_req( TypeFunc::I_O , i_o() ); 1589 call->init_req( TypeFunc::Memory , reset_memory() ); 1590 call->init_req( TypeFunc::FramePtr, frameptr() ); 1591 call->init_req( TypeFunc::ReturnAdr, top() ); 1592 1593 add_safepoint_edges(call, must_throw); 1594 1595 Node* xcall = _gvn.transform(call); 1596 1597 if (xcall == top()) { 1598 set_control(top()); 1599 return; 1600 } 1601 assert(xcall == call, "call identity is stable"); 1602 1603 // Re-use the current map to produce the result. 1604 1605 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); 1606 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O ))); 1607 set_all_memory_call(xcall); 1608 1609 //return xcall; // no need, caller already has it 1610 } 1611 1612 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) { 1613 if (stopped()) return top(); // maybe the call folded up? 1614 1615 // Capture the return value, if any. 1616 Node* ret; 1617 if (call->method() == NULL || 1618 call->method()->return_type()->basic_type() == T_VOID) 1619 ret = top(); 1620 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 1621 1622 // Note: Since any out-of-line call can produce an exception, 1623 // we always insert an I_O projection from the call into the result. 1624 1625 make_slow_call_ex(call, env()->Throwable_klass(), false); 1626 1627 return ret; 1628 } 1629 1630 //--------------------set_predefined_input_for_runtime_call-------------------- 1631 // Reading and setting the memory state is way conservative here. 1632 // The real problem is that I am not doing real Type analysis on memory, 1633 // so I cannot distinguish card mark stores from other stores. Across a GC 1634 // point the Store Barrier and the card mark memory has to agree. I cannot 1635 // have a card mark store and its barrier split across the GC point from 1636 // either above or below. Here I get that to happen by reading ALL of memory. 1637 // A better answer would be to separate out card marks from other memory. 1638 // For now, return the input memory state, so that it can be reused 1639 // after the call, if this call has restricted memory effects. 1640 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1641 // Set fixed predefined input arguments 1642 Node* memory = reset_memory(); 1643 call->init_req( TypeFunc::Control, control() ); 1644 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1645 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1646 call->init_req( TypeFunc::FramePtr, frameptr() ); 1647 call->init_req( TypeFunc::ReturnAdr, top() ); 1648 return memory; 1649 } 1650 1651 //-------------------set_predefined_output_for_runtime_call-------------------- 1652 // Set control and memory (not i_o) from the call. 1653 // If keep_mem is not NULL, use it for the output state, 1654 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1655 // If hook_mem is NULL, this call produces no memory effects at all. 1656 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1657 // then only that memory slice is taken from the call. 1658 // In the last case, we must put an appropriate memory barrier before 1659 // the call, so as to create the correct anti-dependencies on loads 1660 // preceding the call. 1661 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1662 Node* keep_mem, 1663 const TypePtr* hook_mem) { 1664 // no i/o 1665 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); 1666 if (keep_mem) { 1667 // First clone the existing memory state 1668 set_all_memory(keep_mem); 1669 if (hook_mem != NULL) { 1670 // Make memory for the call 1671 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1672 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1673 // We also use hook_mem to extract specific effects from arraycopy stubs. 1674 set_memory(mem, hook_mem); 1675 } 1676 // ...else the call has NO memory effects. 1677 1678 // Make sure the call advertises its memory effects precisely. 1679 // This lets us build accurate anti-dependences in gcm.cpp. 1680 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1681 "call node must be constructed correctly"); 1682 } else { 1683 assert(hook_mem == NULL, ""); 1684 // This is not a "slow path" call; all memory comes from the call. 1685 set_all_memory_call(call); 1686 } 1687 } 1688 1689 //------------------------------increment_counter------------------------------ 1690 // for statistics: increment a VM counter by 1 1691 1692 void GraphKit::increment_counter(address counter_addr) { 1693 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1694 increment_counter(adr1); 1695 } 1696 1697 void GraphKit::increment_counter(Node* counter_addr) { 1698 int adr_type = Compile::AliasIdxRaw; 1699 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 1700 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 1701 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); 1702 } 1703 1704 1705 //------------------------------uncommon_trap---------------------------------- 1706 // Bail out to the interpreter in mid-method. Implemented by calling the 1707 // uncommon_trap blob. This helper function inserts a runtime call with the 1708 // right debug info. 1709 void GraphKit::uncommon_trap(int trap_request, 1710 ciKlass* klass, const char* comment, 1711 bool must_throw, 1712 bool keep_exact_action) { 1713 if (failing()) stop(); 1714 if (stopped()) return; // trap reachable? 1715 1716 // Note: If ProfileTraps is true, and if a deopt. actually 1717 // occurs here, the runtime will make sure an MDO exists. There is 1718 // no need to call method()->build_method_data() at this point. 1719 1720 #ifdef ASSERT 1721 if (!must_throw) { 1722 // Make sure the stack has at least enough depth to execute 1723 // the current bytecode. 1724 int inputs, ignore; 1725 if (compute_stack_effects(inputs, ignore)) { 1726 assert(sp() >= inputs, "must have enough JVMS stack to execute"); 1727 // It is a frequent error in library_call.cpp to issue an 1728 // uncommon trap with the _sp value already popped. 1729 } 1730 } 1731 #endif 1732 1733 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1734 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1735 1736 switch (action) { 1737 case Deoptimization::Action_maybe_recompile: 1738 case Deoptimization::Action_reinterpret: 1739 // Temporary fix for 6529811 to allow virtual calls to be sure they 1740 // get the chance to go from mono->bi->mega 1741 if (!keep_exact_action && 1742 Deoptimization::trap_request_index(trap_request) < 0 && 1743 too_many_recompiles(reason)) { 1744 // This BCI is causing too many recompilations. 1745 action = Deoptimization::Action_none; 1746 trap_request = Deoptimization::make_trap_request(reason, action); 1747 } else { 1748 C->set_trap_can_recompile(true); 1749 } 1750 break; 1751 case Deoptimization::Action_make_not_entrant: 1752 C->set_trap_can_recompile(true); 1753 break; 1754 #ifdef ASSERT 1755 case Deoptimization::Action_none: 1756 case Deoptimization::Action_make_not_compilable: 1757 break; 1758 default: 1759 assert(false, "bad action"); 1760 #endif 1761 } 1762 1763 if (TraceOptoParse) { 1764 char buf[100]; 1765 tty->print_cr("Uncommon trap %s at bci:%d", 1766 Deoptimization::format_trap_request(buf, sizeof(buf), 1767 trap_request), bci()); 1768 } 1769 1770 CompileLog* log = C->log(); 1771 if (log != NULL) { 1772 int kid = (klass == NULL)? -1: log->identify(klass); 1773 log->begin_elem("uncommon_trap bci='%d'", bci()); 1774 char buf[100]; 1775 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1776 trap_request)); 1777 if (kid >= 0) log->print(" klass='%d'", kid); 1778 if (comment != NULL) log->print(" comment='%s'", comment); 1779 log->end_elem(); 1780 } 1781 1782 // Make sure any guarding test views this path as very unlikely 1783 Node *i0 = control()->in(0); 1784 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1785 IfNode *iff = i0->as_If(); 1786 float f = iff->_prob; // Get prob 1787 if (control()->Opcode() == Op_IfTrue) { 1788 if (f > PROB_UNLIKELY_MAG(4)) 1789 iff->_prob = PROB_MIN; 1790 } else { 1791 if (f < PROB_LIKELY_MAG(4)) 1792 iff->_prob = PROB_MAX; 1793 } 1794 } 1795 1796 // Clear out dead values from the debug info. 1797 kill_dead_locals(); 1798 1799 // Now insert the uncommon trap subroutine call 1800 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); 1801 const TypePtr* no_memory_effects = NULL; 1802 // Pass the index of the class to be loaded 1803 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1804 (must_throw ? RC_MUST_THROW : 0), 1805 OptoRuntime::uncommon_trap_Type(), 1806 call_addr, "uncommon_trap", no_memory_effects, 1807 intcon(trap_request)); 1808 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 1809 "must extract request correctly from the graph"); 1810 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 1811 1812 call->set_req(TypeFunc::ReturnAdr, returnadr()); 1813 // The debug info is the only real input to this call. 1814 1815 // Halt-and-catch fire here. The above call should never return! 1816 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); 1817 _gvn.set_type_bottom(halt); 1818 root()->add_req(halt); 1819 1820 stop_and_kill_map(); 1821 } 1822 1823 1824 //--------------------------just_allocated_object------------------------------ 1825 // Report the object that was just allocated. 1826 // It must be the case that there are no intervening safepoints. 1827 // We use this to determine if an object is so "fresh" that 1828 // it does not require card marks. 1829 Node* GraphKit::just_allocated_object(Node* current_control) { 1830 if (C->recent_alloc_ctl() == current_control) 1831 return C->recent_alloc_obj(); 1832 return NULL; 1833 } 1834 1835 1836 //------------------------------store_barrier---------------------------------- 1837 // Insert a write-barrier store. This is to let generational GC work; we have 1838 // to flag all oop-stores before the next GC point. 1839 void GraphKit::write_barrier_post(Node* oop_store, Node* obj, Node* adr, 1840 Node* val, bool use_precise) { 1841 // No store check needed if we're storing a NULL or an old object 1842 // (latter case is probably a string constant). The concurrent 1843 // mark sweep garbage collector, however, needs to have all nonNull 1844 // oop updates flagged via card-marks. 1845 if (val != NULL && val->is_Con()) { 1846 // must be either an oop or NULL 1847 const Type* t = val->bottom_type(); 1848 if (t == TypePtr::NULL_PTR || t == Type::TOP) 1849 // stores of null never (?) need barriers 1850 return; 1851 ciObject* con = t->is_oopptr()->const_oop(); 1852 if (con != NULL 1853 && con->is_perm() 1854 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 1855 // no store barrier needed, because no old-to-new ref created 1856 return; 1857 } 1858 1859 if (use_ReduceInitialCardMarks() 1860 && obj == just_allocated_object(control())) { 1861 // We can skip marks on a freshly-allocated object. 1862 // Keep this code in sync with do_eager_card_mark in runtime.cpp. 1863 // That routine eagerly marks the occasional object which is produced 1864 // by the slow path, so that we don't have to do it here. 1865 return; 1866 } 1867 1868 if (!use_precise) { 1869 // All card marks for a (non-array) instance are in one place: 1870 adr = obj; 1871 } 1872 // (Else it's an array (or unknown), and we want more precise card marks.) 1873 assert(adr != NULL, ""); 1874 1875 // Get the alias_index for raw card-mark memory 1876 int adr_type = Compile::AliasIdxRaw; 1877 // Convert the pointer to an int prior to doing math on it 1878 Node* cast = _gvn.transform(new (C, 2) CastP2XNode(control(), adr)); 1879 // Divide by card size 1880 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 1881 "Only one we handle so far."); 1882 CardTableModRefBS* ct = 1883 (CardTableModRefBS*)(Universe::heap()->barrier_set()); 1884 Node *b = _gvn.transform(new (C, 3) URShiftXNode( cast, _gvn.intcon(CardTableModRefBS::card_shift) )); 1885 // We store into a byte array, so do not bother to left-shift by zero 1886 Node *c = byte_map_base_node(); 1887 // Combine 1888 Node *sb_ctl = control(); 1889 Node *sb_adr = _gvn.transform(new (C, 4) AddPNode( top()/*no base ptr*/, c, b )); 1890 Node *sb_val = _gvn.intcon(0); 1891 // Smash zero into card 1892 if( !UseConcMarkSweepGC ) { 1893 BasicType bt = T_BYTE; 1894 store_to_memory(sb_ctl, sb_adr, sb_val, bt, adr_type); 1895 } else { 1896 // Specialized path for CM store barrier 1897 cms_card_mark( sb_ctl, sb_adr, sb_val, oop_store); 1898 } 1899 } 1900 1901 // Specialized path for CMS store barrier 1902 void GraphKit::cms_card_mark(Node* ctl, Node* adr, Node* val, Node *oop_store) { 1903 BasicType bt = T_BYTE; 1904 int adr_idx = Compile::AliasIdxRaw; 1905 Node* mem = memory(adr_idx); 1906 1907 // The type input is NULL in PRODUCT builds 1908 const TypePtr* type = NULL; 1909 debug_only(type = C->get_adr_type(adr_idx)); 1910 1911 // Add required edge to oop_store, optimizer does not support precedence edges. 1912 // Convert required edge to precedence edge before allocation. 1913 Node *store = _gvn.transform( new (C, 5) StoreCMNode(ctl, mem, adr, type, val, oop_store) ); 1914 set_memory(store, adr_idx); 1915 1916 // For CMS, back-to-back card-marks can only remove the first one 1917 // and this requires DU info. Push on worklist for optimizer. 1918 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1919 record_for_igvn(store); 1920 } 1921 1922 1923 void GraphKit::round_double_arguments(ciMethod* dest_method) { 1924 // (Note: TypeFunc::make has a cache that makes this fast.) 1925 const TypeFunc* tf = TypeFunc::make(dest_method); 1926 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 1927 for (int j = 0; j < nargs; j++) { 1928 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 1929 if( targ->basic_type() == T_DOUBLE ) { 1930 // If any parameters are doubles, they must be rounded before 1931 // the call, dstore_rounding does gvn.transform 1932 Node *arg = argument(j); 1933 arg = dstore_rounding(arg); 1934 set_argument(j, arg); 1935 } 1936 } 1937 } 1938 1939 void GraphKit::round_double_result(ciMethod* dest_method) { 1940 // A non-strict method may return a double value which has an extended 1941 // exponent, but this must not be visible in a caller which is 'strict' 1942 // If a strict caller invokes a non-strict callee, round a double result 1943 1944 BasicType result_type = dest_method->return_type()->basic_type(); 1945 assert( method() != NULL, "must have caller context"); 1946 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 1947 // Destination method's return value is on top of stack 1948 // dstore_rounding() does gvn.transform 1949 Node *result = pop_pair(); 1950 result = dstore_rounding(result); 1951 push_pair(result); 1952 } 1953 } 1954 1955 // rounding for strict float precision conformance 1956 Node* GraphKit::precision_rounding(Node* n) { 1957 return UseStrictFP && _method->flags().is_strict() 1958 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 1959 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) 1960 : n; 1961 } 1962 1963 // rounding for strict double precision conformance 1964 Node* GraphKit::dprecision_rounding(Node *n) { 1965 return UseStrictFP && _method->flags().is_strict() 1966 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 1967 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1968 : n; 1969 } 1970 1971 // rounding for non-strict double stores 1972 Node* GraphKit::dstore_rounding(Node* n) { 1973 return Matcher::strict_fp_requires_explicit_rounding 1974 && UseSSE <= 1 1975 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1976 : n; 1977 } 1978 1979 //============================================================================= 1980 // Generate a fast path/slow path idiom. Graph looks like: 1981 // [foo] indicates that 'foo' is a parameter 1982 // 1983 // [in] NULL 1984 // \ / 1985 // CmpP 1986 // Bool ne 1987 // If 1988 // / \ 1989 // True False-<2> 1990 // / | 1991 // / cast_not_null 1992 // Load | | ^ 1993 // [fast_test] | | 1994 // gvn to opt_test | | 1995 // / \ | <1> 1996 // True False | 1997 // | \\ | 1998 // [slow_call] \[fast_result] 1999 // Ctl Val \ \ 2000 // | \ \ 2001 // Catch <1> \ \ 2002 // / \ ^ \ \ 2003 // Ex No_Ex | \ \ 2004 // | \ \ | \ <2> \ 2005 // ... \ [slow_res] | | \ [null_result] 2006 // \ \--+--+--- | | 2007 // \ | / \ | / 2008 // --------Region Phi 2009 // 2010 //============================================================================= 2011 // Code is structured as a series of driver functions all called 'do_XXX' that 2012 // call a set of helper functions. Helper functions first, then drivers. 2013 2014 //------------------------------null_check_oop--------------------------------- 2015 // Null check oop. Set null-path control into Region in slot 3. 2016 // Make a cast-not-nullness use the other not-null control. Return cast. 2017 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2018 bool never_see_null) { 2019 // Initial NULL check taken path 2020 (*null_control) = top(); 2021 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2022 2023 // Generate uncommon_trap: 2024 if (never_see_null && (*null_control) != top()) { 2025 // If we see an unexpected null at a check-cast we record it and force a 2026 // recompile; the offending check-cast will be compiled to handle NULLs. 2027 // If we see more than one offending BCI, then all checkcasts in the 2028 // method will be compiled to handle NULLs. 2029 PreserveJVMState pjvms(this); 2030 set_control(*null_control); 2031 replace_in_map(value, null()); 2032 uncommon_trap(Deoptimization::Reason_null_check, 2033 Deoptimization::Action_make_not_entrant); 2034 (*null_control) = top(); // NULL path is dead 2035 } 2036 2037 // Cast away null-ness on the result 2038 return cast; 2039 } 2040 2041 //------------------------------opt_iff---------------------------------------- 2042 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2043 // Return slow-path control. 2044 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2045 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2046 2047 // Fast path taken; set region slot 2 2048 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); 2049 region->init_req(2,fast_taken); // Capture fast-control 2050 2051 // Fast path not-taken, i.e. slow path 2052 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); 2053 return slow_taken; 2054 } 2055 2056 //-----------------------------make_runtime_call------------------------------- 2057 Node* GraphKit::make_runtime_call(int flags, 2058 const TypeFunc* call_type, address call_addr, 2059 const char* call_name, 2060 const TypePtr* adr_type, 2061 // The following parms are all optional. 2062 // The first NULL ends the list. 2063 Node* parm0, Node* parm1, 2064 Node* parm2, Node* parm3, 2065 Node* parm4, Node* parm5, 2066 Node* parm6, Node* parm7) { 2067 // Slow-path call 2068 int size = call_type->domain()->cnt(); 2069 bool is_leaf = !(flags & RC_NO_LEAF); 2070 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2071 if (call_name == NULL) { 2072 assert(!is_leaf, "must supply name for leaf"); 2073 call_name = OptoRuntime::stub_name(call_addr); 2074 } 2075 CallNode* call; 2076 if (!is_leaf) { 2077 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, 2078 bci(), adr_type); 2079 } else if (flags & RC_NO_FP) { 2080 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2081 } else { 2082 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); 2083 } 2084 2085 // The following is similar to set_edges_for_java_call, 2086 // except that the memory effects of the call are restricted to AliasIdxRaw. 2087 2088 // Slow path call has no side-effects, uses few values 2089 bool wide_in = !(flags & RC_NARROW_MEM); 2090 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2091 2092 Node* prev_mem = NULL; 2093 if (wide_in) { 2094 prev_mem = set_predefined_input_for_runtime_call(call); 2095 } else { 2096 assert(!wide_out, "narrow in => narrow out"); 2097 Node* narrow_mem = memory(adr_type); 2098 prev_mem = reset_memory(); 2099 map()->set_memory(narrow_mem); 2100 set_predefined_input_for_runtime_call(call); 2101 } 2102 2103 // Hook each parm in order. Stop looking at the first NULL. 2104 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2105 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2106 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2107 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2108 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2109 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2110 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2111 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2112 /* close each nested if ===> */ } } } } } } } } 2113 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2114 2115 if (!is_leaf) { 2116 // Non-leaves can block and take safepoints: 2117 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2118 } 2119 // Non-leaves can throw exceptions: 2120 if (has_io) { 2121 call->set_req(TypeFunc::I_O, i_o()); 2122 } 2123 2124 if (flags & RC_UNCOMMON) { 2125 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2126 // (An "if" probability corresponds roughly to an unconditional count. 2127 // Sort of.) 2128 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2129 } 2130 2131 Node* c = _gvn.transform(call); 2132 assert(c == call, "cannot disappear"); 2133 2134 if (wide_out) { 2135 // Slow path call has full side-effects. 2136 set_predefined_output_for_runtime_call(call); 2137 } else { 2138 // Slow path call has few side-effects, and/or sets few values. 2139 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2140 } 2141 2142 if (has_io) { 2143 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); 2144 } 2145 return call; 2146 2147 } 2148 2149 //------------------------------merge_memory----------------------------------- 2150 // Merge memory from one path into the current memory state. 2151 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2152 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2153 Node* old_slice = mms.force_memory(); 2154 Node* new_slice = mms.memory2(); 2155 if (old_slice != new_slice) { 2156 PhiNode* phi; 2157 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2158 phi = new_slice->as_Phi(); 2159 #ifdef ASSERT 2160 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2161 old_slice = old_slice->in(new_path); 2162 // Caller is responsible for ensuring that any pre-existing 2163 // phis are already aware of old memory. 2164 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2165 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2166 #endif 2167 mms.set_memory(phi); 2168 } else { 2169 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2170 _gvn.set_type(phi, Type::MEMORY); 2171 phi->set_req(new_path, new_slice); 2172 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2173 } 2174 } 2175 } 2176 } 2177 2178 //------------------------------make_slow_call_ex------------------------------ 2179 // Make the exception handler hookups for the slow call 2180 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2181 if (stopped()) return; 2182 2183 // Make a catch node with just two handlers: fall-through and catch-all 2184 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2185 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); 2186 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2187 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2188 2189 { PreserveJVMState pjvms(this); 2190 set_control(excp); 2191 set_i_o(i_o); 2192 2193 if (excp != top()) { 2194 // Create an exception state also. 2195 // Use an exact type if the caller has specified a specific exception. 2196 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2197 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); 2198 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2199 } 2200 } 2201 2202 // Get the no-exception control from the CatchNode. 2203 set_control(norm); 2204 } 2205 2206 2207 //-------------------------------gen_subtype_check----------------------------- 2208 // Generate a subtyping check. Takes as input the subtype and supertype. 2209 // Returns 2 values: sets the default control() to the true path and returns 2210 // the false path. Only reads invariant memory; sets no (visible) memory. 2211 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2212 // but that's not exposed to the optimizer. This call also doesn't take in an 2213 // Object; if you wish to check an Object you need to load the Object's class 2214 // prior to coming here. 2215 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2216 // Fast check for identical types, perhaps identical constants. 2217 // The types can even be identical non-constants, in cases 2218 // involving Array.newInstance, Object.clone, etc. 2219 if (subklass == superklass) 2220 return top(); // false path is dead; no test needed. 2221 2222 if (_gvn.type(superklass)->singleton()) { 2223 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2224 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2225 2226 // In the common case of an exact superklass, try to fold up the 2227 // test before generating code. You may ask, why not just generate 2228 // the code and then let it fold up? The answer is that the generated 2229 // code will necessarily include null checks, which do not always 2230 // completely fold away. If they are also needless, then they turn 2231 // into a performance loss. Example: 2232 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2233 // Here, the type of 'fa' is often exact, so the store check 2234 // of fa[1]=x will fold up, without testing the nullness of x. 2235 switch (static_subtype_check(superk, subk)) { 2236 case SSC_always_false: 2237 { 2238 Node* always_fail = control(); 2239 set_control(top()); 2240 return always_fail; 2241 } 2242 case SSC_always_true: 2243 return top(); 2244 case SSC_easy_test: 2245 { 2246 // Just do a direct pointer compare and be done. 2247 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); 2248 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2249 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2250 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); 2251 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2252 } 2253 case SSC_full_test: 2254 break; 2255 default: 2256 ShouldNotReachHere(); 2257 } 2258 } 2259 2260 // %%% Possible further optimization: Even if the superklass is not exact, 2261 // if the subklass is the unique subtype of the superklass, the check 2262 // will always succeed. We could leave a dependency behind to ensure this. 2263 2264 // First load the super-klass's check-offset 2265 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); 2266 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2267 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); 2268 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2269 2270 // Load from the sub-klass's super-class display list, or a 1-word cache of 2271 // the secondary superclass list, or a failing value with a sentinel offset 2272 // if the super-klass is an interface or exceptionally deep in the Java 2273 // hierarchy and we have to scan the secondary superclass list the hard way. 2274 // Worst-case type is a little odd: NULL is allowed as a result (usually 2275 // klass loads can never produce a NULL). 2276 Node *chk_off_X = ConvI2X(chk_off); 2277 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); 2278 // For some types like interfaces the following loadKlass is from a 1-word 2279 // cache which is mutable so can't use immutable memory. Other 2280 // types load from the super-class display table which is immutable. 2281 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2282 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2283 2284 // Compile speed common case: ARE a subtype and we canNOT fail 2285 if( superklass == nkls ) 2286 return top(); // false path is dead; no test needed. 2287 2288 // See if we get an immediate positive hit. Happens roughly 83% of the 2289 // time. Test to see if the value loaded just previously from the subklass 2290 // is exactly the superklass. 2291 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); 2292 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); 2293 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2294 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); 2295 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); 2296 2297 // Compile speed common case: Check for being deterministic right now. If 2298 // chk_off is a constant and not equal to cacheoff then we are NOT a 2299 // subklass. In this case we need exactly the 1 test above and we can 2300 // return those results immediately. 2301 if (!might_be_cache) { 2302 Node* not_subtype_ctrl = control(); 2303 set_control(iftrue1); // We need exactly the 1 test above 2304 return not_subtype_ctrl; 2305 } 2306 2307 // Gather the various success & failures here 2308 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); 2309 record_for_igvn(r_ok_subtype); 2310 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); 2311 record_for_igvn(r_not_subtype); 2312 2313 r_ok_subtype->init_req(1, iftrue1); 2314 2315 // Check for immediate negative hit. Happens roughly 11% of the time (which 2316 // is roughly 63% of the remaining cases). Test to see if the loaded 2317 // check-offset points into the subklass display list or the 1-element 2318 // cache. If it points to the display (and NOT the cache) and the display 2319 // missed then it's not a subtype. 2320 Node *cacheoff = _gvn.intcon(cacheoff_con); 2321 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); 2322 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); 2323 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2324 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); 2325 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); 2326 2327 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2328 // No performance impact (too rare) but allows sharing of secondary arrays 2329 // which has some footprint reduction. 2330 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); 2331 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); 2332 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2333 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); 2334 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); 2335 2336 // -- Roads not taken here: -- 2337 // We could also have chosen to perform the self-check at the beginning 2338 // of this code sequence, as the assembler does. This would not pay off 2339 // the same way, since the optimizer, unlike the assembler, can perform 2340 // static type analysis to fold away many successful self-checks. 2341 // Non-foldable self checks work better here in second position, because 2342 // the initial primary superclass check subsumes a self-check for most 2343 // types. An exception would be a secondary type like array-of-interface, 2344 // which does not appear in its own primary supertype display. 2345 // Finally, we could have chosen to move the self-check into the 2346 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2347 // dependent manner. But it is worthwhile to have the check here, 2348 // where it can be perhaps be optimized. The cost in code space is 2349 // small (register compare, branch). 2350 2351 // Now do a linear scan of the secondary super-klass array. Again, no real 2352 // performance impact (too rare) but it's gotta be done. 2353 // Since the code is rarely used, there is no penalty for moving it 2354 // out of line, and it can only improve I-cache density. 2355 // The decision to inline or out-of-line this final check is platform 2356 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2357 Node* psc = _gvn.transform( 2358 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2359 2360 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); 2361 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); 2362 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2363 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); 2364 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); 2365 2366 // Return false path; set default control to true path. 2367 set_control( _gvn.transform(r_ok_subtype) ); 2368 return _gvn.transform(r_not_subtype); 2369 } 2370 2371 //----------------------------static_subtype_check----------------------------- 2372 // Shortcut important common cases when superklass is exact: 2373 // (0) superklass is java.lang.Object (can occur in reflective code) 2374 // (1) subklass is already limited to a subtype of superklass => always ok 2375 // (2) subklass does not overlap with superklass => always fail 2376 // (3) superklass has NO subtypes and we can check with a simple compare. 2377 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2378 if (StressReflectiveCode) { 2379 return SSC_full_test; // Let caller generate the general case. 2380 } 2381 2382 if (superk == env()->Object_klass()) { 2383 return SSC_always_true; // (0) this test cannot fail 2384 } 2385 2386 ciType* superelem = superk; 2387 if (superelem->is_array_klass()) 2388 superelem = superelem->as_array_klass()->base_element_type(); 2389 2390 if (!subk->is_interface()) { // cannot trust static interface types yet 2391 if (subk->is_subtype_of(superk)) { 2392 return SSC_always_true; // (1) false path dead; no dynamic test needed 2393 } 2394 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2395 !superk->is_subtype_of(subk)) { 2396 return SSC_always_false; 2397 } 2398 } 2399 2400 // If casting to an instance klass, it must have no subtypes 2401 if (superk->is_interface()) { 2402 // Cannot trust interfaces yet. 2403 // %%% S.B. superk->nof_implementors() == 1 2404 } else if (superelem->is_instance_klass()) { 2405 ciInstanceKlass* ik = superelem->as_instance_klass(); 2406 if (!ik->has_subklass() && !ik->is_interface()) { 2407 if (!ik->is_final()) { 2408 // Add a dependency if there is a chance of a later subclass. 2409 C->dependencies()->assert_leaf_type(ik); 2410 } 2411 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2412 } 2413 } else { 2414 // A primitive array type has no subtypes. 2415 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2416 } 2417 2418 return SSC_full_test; 2419 } 2420 2421 // Profile-driven exact type check: 2422 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2423 float prob, 2424 Node* *casted_receiver) { 2425 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2426 Node* recv_klass = load_object_klass(receiver); 2427 Node* want_klass = makecon(tklass); 2428 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); 2429 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2430 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2431 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); 2432 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2433 2434 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2435 assert(recv_xtype->klass_is_exact(), ""); 2436 2437 // Subsume downstream occurrences of receiver with a cast to 2438 // recv_xtype, since now we know what the type will be. 2439 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); 2440 (*casted_receiver) = _gvn.transform(cast); 2441 // (User must make the replace_in_map call.) 2442 2443 return fail; 2444 } 2445 2446 2447 //-------------------------------gen_instanceof-------------------------------- 2448 // Generate an instance-of idiom. Used by both the instance-of bytecode 2449 // and the reflective instance-of call. 2450 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { 2451 C->set_has_split_ifs(true); // Has chance for split-if optimization 2452 assert( !stopped(), "dead parse path should be checked in callers" ); 2453 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2454 "must check for not-null not-dead klass in callers"); 2455 2456 // Make the merge point 2457 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2458 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2459 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2460 C->set_has_split_ifs(true); // Has chance for split-if optimization 2461 2462 // Null check; get casted pointer; set region slot 3 2463 Node* null_ctl = top(); 2464 Node* not_null_obj = null_check_oop(subobj, &null_ctl); 2465 2466 // If not_null_obj is dead, only null-path is taken 2467 if (stopped()) { // Doing instance-of on a NULL? 2468 set_control(null_ctl); 2469 return intcon(0); 2470 } 2471 region->init_req(_null_path, null_ctl); 2472 phi ->init_req(_null_path, intcon(0)); // Set null path value 2473 2474 // Load the object's klass 2475 Node* obj_klass = load_object_klass(not_null_obj); 2476 2477 // Generate the subtype check 2478 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2479 2480 // Plug in the success path to the general merge in slot 1. 2481 region->init_req(_obj_path, control()); 2482 phi ->init_req(_obj_path, intcon(1)); 2483 2484 // Plug in the failing path to the general merge in slot 2. 2485 region->init_req(_fail_path, not_subtype_ctrl); 2486 phi ->init_req(_fail_path, intcon(0)); 2487 2488 // Return final merged results 2489 set_control( _gvn.transform(region) ); 2490 record_for_igvn(region); 2491 return _gvn.transform(phi); 2492 } 2493 2494 //-------------------------------gen_checkcast--------------------------------- 2495 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2496 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2497 // uncommon-trap paths work. Adjust stack after this call. 2498 // If failure_control is supplied and not null, it is filled in with 2499 // the control edge for the cast failure. Otherwise, an appropriate 2500 // uncommon trap or exception is thrown. 2501 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2502 Node* *failure_control) { 2503 kill_dead_locals(); // Benefit all the uncommon traps 2504 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2505 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2506 2507 // Fast cutout: Check the case that the cast is vacuously true. 2508 // This detects the common cases where the test will short-circuit 2509 // away completely. We do this before we perform the null check, 2510 // because if the test is going to turn into zero code, we don't 2511 // want a residual null check left around. (Causes a slowdown, 2512 // for example, in some objArray manipulations, such as a[i]=a[j].) 2513 if (tk->singleton()) { 2514 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2515 if (objtp != NULL && objtp->klass() != NULL) { 2516 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2517 case SSC_always_true: 2518 return obj; 2519 case SSC_always_false: 2520 // It needs a null check because a null will *pass* the cast check. 2521 // A non-null value will always produce an exception. 2522 return do_null_assert(obj, T_OBJECT); 2523 } 2524 } 2525 } 2526 2527 ciProfileData* data = NULL; 2528 if (failure_control == NULL) { // use MDO in regular case only 2529 assert(java_bc() == Bytecodes::_aastore || 2530 java_bc() == Bytecodes::_checkcast, 2531 "interpreter profiles type checks only for these BCs"); 2532 data = method()->method_data()->bci_to_data(bci()); 2533 } 2534 2535 // Make the merge point 2536 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2537 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2538 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); 2539 C->set_has_split_ifs(true); // Has chance for split-if optimization 2540 2541 // Use null-cast information if it is available 2542 bool never_see_null = false; 2543 // If we see an unexpected null at a check-cast we record it and force a 2544 // recompile; the offending check-cast will be compiled to handle NULLs. 2545 // If we see several offending BCIs, then all checkcasts in the 2546 // method will be compiled to handle NULLs. 2547 if (UncommonNullCast // Cutout for this technique 2548 && failure_control == NULL // regular case 2549 && obj != null() // And not the -Xcomp stupid case? 2550 && !too_many_traps(Deoptimization::Reason_null_check)) { 2551 // Finally, check the "null_seen" bit from the interpreter. 2552 if (data == NULL || !data->as_BitData()->null_seen()) { 2553 never_see_null = true; 2554 } 2555 } 2556 2557 // Null check; get casted pointer; set region slot 3 2558 Node* null_ctl = top(); 2559 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2560 2561 // If not_null_obj is dead, only null-path is taken 2562 if (stopped()) { // Doing instance-of on a NULL? 2563 set_control(null_ctl); 2564 return null(); 2565 } 2566 region->init_req(_null_path, null_ctl); 2567 phi ->init_req(_null_path, null()); // Set null path value 2568 2569 Node* cast_obj = NULL; // the casted version of the object 2570 2571 // If the profile has seen exactly one type, narrow to that type. 2572 // (The subsequent subtype check will always fold up.) 2573 if (UseTypeProfile && TypeProfileCasts && data != NULL && 2574 // Counter has never been decremented (due to cast failure). 2575 // ...This is a reasonable thing to expect. It is true of 2576 // all casts inserted by javac to implement generic types. 2577 data->as_CounterData()->count() >= 0 && 2578 !too_many_traps(Deoptimization::Reason_class_check)) { 2579 // (No, this isn't a call, but it's enough like a virtual call 2580 // to use the same ciMethod accessor to get the profile info...) 2581 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2582 if (profile.count() >= 0 && // no cast failures here 2583 profile.has_receiver(0) && 2584 profile.morphism() == 1) { 2585 ciKlass* exact_kls = profile.receiver(0); 2586 int ssc = static_subtype_check(tk->klass(), exact_kls); 2587 if (ssc == SSC_always_true) { 2588 // If we narrow the type to match what the type profile sees, 2589 // we can then remove the rest of the cast. 2590 // This is a win, even if the exact_kls is very specific, 2591 // because downstream operations, such as method calls, 2592 // will often benefit from the sharper type. 2593 Node* exact_obj = not_null_obj; // will get updated in place... 2594 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2595 &exact_obj); 2596 { PreserveJVMState pjvms(this); 2597 set_control(slow_ctl); 2598 uncommon_trap(Deoptimization::Reason_class_check, 2599 Deoptimization::Action_maybe_recompile); 2600 } 2601 if (failure_control != NULL) // failure is now impossible 2602 (*failure_control) = top(); 2603 replace_in_map(not_null_obj, exact_obj); 2604 // adjust the type of the phi to the exact klass: 2605 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); 2606 cast_obj = exact_obj; 2607 } 2608 // assert(cast_obj != NULL)... except maybe the profile lied to us. 2609 } 2610 } 2611 2612 if (cast_obj == NULL) { 2613 // Load the object's klass 2614 Node* obj_klass = load_object_klass(not_null_obj); 2615 2616 // Generate the subtype check 2617 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2618 2619 // Plug in success path into the merge 2620 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), 2621 not_null_obj, toop)); 2622 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2623 if (failure_control == NULL) { 2624 if (not_subtype_ctrl != top()) { // If failure is possible 2625 PreserveJVMState pjvms(this); 2626 set_control(not_subtype_ctrl); 2627 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2628 } 2629 } else { 2630 (*failure_control) = not_subtype_ctrl; 2631 } 2632 } 2633 2634 region->init_req(_obj_path, control()); 2635 phi ->init_req(_obj_path, cast_obj); 2636 2637 // A merge of NULL or Casted-NotNull obj 2638 Node* res = _gvn.transform(phi); 2639 2640 // Note I do NOT always 'replace_in_map(obj,result)' here. 2641 // if( tk->klass()->can_be_primary_super() ) 2642 // This means that if I successfully store an Object into an array-of-String 2643 // I 'forget' that the Object is really now known to be a String. I have to 2644 // do this because we don't have true union types for interfaces - if I store 2645 // a Baz into an array-of-Interface and then tell the optimizer it's an 2646 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2647 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2648 // replace_in_map( obj, res ); 2649 2650 // Return final merged results 2651 set_control( _gvn.transform(region) ); 2652 record_for_igvn(region); 2653 return res; 2654 } 2655 2656 //------------------------------next_monitor----------------------------------- 2657 // What number should be given to the next monitor? 2658 int GraphKit::next_monitor() { 2659 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2660 int next = current + C->sync_stack_slots(); 2661 // Keep the toplevel high water mark current: 2662 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2663 return current; 2664 } 2665 2666 //------------------------------insert_mem_bar--------------------------------- 2667 // Memory barrier to avoid floating things around 2668 // The membar serves as a pinch point between both control and all memory slices. 2669 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2670 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2671 mb->init_req(TypeFunc::Control, control()); 2672 mb->init_req(TypeFunc::Memory, reset_memory()); 2673 Node* membar = _gvn.transform(mb); 2674 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); 2675 set_all_memory_call(membar); 2676 return membar; 2677 } 2678 2679 //-------------------------insert_mem_bar_volatile---------------------------- 2680 // Memory barrier to avoid floating things around 2681 // The membar serves as a pinch point between both control and memory(alias_idx). 2682 // If you want to make a pinch point on all memory slices, do not use this 2683 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2684 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2685 // When Parse::do_put_xxx updates a volatile field, it appends a series 2686 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2687 // The first membar is on the same memory slice as the field store opcode. 2688 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2689 // All the other membars (for other volatile slices, including AliasIdxBot, 2690 // which stands for all unknown volatile slices) are control-dependent 2691 // on the first membar. This prevents later volatile loads or stores 2692 // from sliding up past the just-emitted store. 2693 2694 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2695 mb->set_req(TypeFunc::Control,control()); 2696 if (alias_idx == Compile::AliasIdxBot) { 2697 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2698 } else { 2699 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2700 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2701 } 2702 Node* membar = _gvn.transform(mb); 2703 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); 2704 if (alias_idx == Compile::AliasIdxBot) { 2705 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); 2706 } else { 2707 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2708 } 2709 return membar; 2710 } 2711 2712 //------------------------------shared_lock------------------------------------ 2713 // Emit locking code. 2714 FastLockNode* GraphKit::shared_lock(Node* obj) { 2715 // bci is either a monitorenter bc or InvocationEntryBci 2716 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2717 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2718 2719 if( !GenerateSynchronizationCode ) 2720 return NULL; // Not locking things? 2721 if (stopped()) // Dead monitor? 2722 return NULL; 2723 2724 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2725 2726 // Box the stack location 2727 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); 2728 Node* mem = reset_memory(); 2729 2730 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); 2731 if (PrintPreciseBiasedLockingStatistics) { 2732 // Create the counters for this fast lock. 2733 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2734 } 2735 // Add monitor to debug info for the slow path. If we block inside the 2736 // slow path and de-opt, we need the monitor hanging around 2737 map()->push_monitor( flock ); 2738 2739 const TypeFunc *tf = LockNode::lock_type(); 2740 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); 2741 2742 lock->init_req( TypeFunc::Control, control() ); 2743 lock->init_req( TypeFunc::Memory , mem ); 2744 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2745 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2746 lock->init_req( TypeFunc::ReturnAdr, top() ); 2747 2748 lock->init_req(TypeFunc::Parms + 0, obj); 2749 lock->init_req(TypeFunc::Parms + 1, box); 2750 lock->init_req(TypeFunc::Parms + 2, flock); 2751 add_safepoint_edges(lock); 2752 2753 lock = _gvn.transform( lock )->as_Lock(); 2754 2755 // lock has no side-effects, sets few values 2756 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2757 2758 insert_mem_bar(Op_MemBarAcquire); 2759 2760 // Add this to the worklist so that the lock can be eliminated 2761 record_for_igvn(lock); 2762 2763 #ifndef PRODUCT 2764 if (PrintLockStatistics) { 2765 // Update the counter for this lock. Don't bother using an atomic 2766 // operation since we don't require absolute accuracy. 2767 lock->create_lock_counter(map()->jvms()); 2768 int adr_type = Compile::AliasIdxRaw; 2769 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); 2770 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 2771 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 2772 store_to_memory(control(), counter_addr, incr, T_INT, adr_type); 2773 } 2774 #endif 2775 2776 return flock; 2777 } 2778 2779 2780 //------------------------------shared_unlock---------------------------------- 2781 // Emit unlocking code. 2782 void GraphKit::shared_unlock(Node* box, Node* obj) { 2783 // bci is either a monitorenter bc or InvocationEntryBci 2784 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2785 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2786 2787 if( !GenerateSynchronizationCode ) 2788 return; 2789 if (stopped()) { // Dead monitor? 2790 map()->pop_monitor(); // Kill monitor from debug info 2791 return; 2792 } 2793 2794 // Memory barrier to avoid floating things down past the locked region 2795 insert_mem_bar(Op_MemBarRelease); 2796 2797 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2798 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); 2799 uint raw_idx = Compile::AliasIdxRaw; 2800 unlock->init_req( TypeFunc::Control, control() ); 2801 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2802 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2803 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2804 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2805 2806 unlock->init_req(TypeFunc::Parms + 0, obj); 2807 unlock->init_req(TypeFunc::Parms + 1, box); 2808 unlock = _gvn.transform(unlock)->as_Unlock(); 2809 2810 Node* mem = reset_memory(); 2811 2812 // unlock has no side-effects, sets few values 2813 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 2814 2815 // Kill monitor from debug info 2816 map()->pop_monitor( ); 2817 } 2818 2819 //-------------------------------get_layout_helper----------------------------- 2820 // If the given klass is a constant or known to be an array, 2821 // fetch the constant layout helper value into constant_value 2822 // and return (Node*)NULL. Otherwise, load the non-constant 2823 // layout helper value, and return the node which represents it. 2824 // This two-faced routine is useful because allocation sites 2825 // almost always feature constant types. 2826 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 2827 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 2828 if (!StressReflectiveCode && inst_klass != NULL) { 2829 ciKlass* klass = inst_klass->klass(); 2830 bool xklass = inst_klass->klass_is_exact(); 2831 if (xklass || klass->is_array_klass()) { 2832 jint lhelper = klass->layout_helper(); 2833 if (lhelper != Klass::_lh_neutral_value) { 2834 constant_value = lhelper; 2835 return (Node*) NULL; 2836 } 2837 } 2838 } 2839 constant_value = Klass::_lh_neutral_value; // put in a known value 2840 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); 2841 return make_load(NULL, lhp, TypeInt::INT, T_INT); 2842 } 2843 2844 // We just put in an allocate/initialize with a big raw-memory effect. 2845 // Hook selected additional alias categories on the initialization. 2846 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 2847 MergeMemNode* init_in_merge, 2848 Node* init_out_raw) { 2849 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 2850 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 2851 2852 Node* prevmem = kit.memory(alias_idx); 2853 init_in_merge->set_memory_at(alias_idx, prevmem); 2854 kit.set_memory(init_out_raw, alias_idx); 2855 } 2856 2857 //---------------------------set_output_for_allocation------------------------- 2858 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 2859 const TypeOopPtr* oop_type, 2860 bool raw_mem_only) { 2861 int rawidx = Compile::AliasIdxRaw; 2862 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 2863 add_safepoint_edges(alloc); 2864 Node* allocx = _gvn.transform(alloc); 2865 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); 2866 // create memory projection for i_o 2867 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 2868 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); 2869 2870 // create a memory projection as for the normal control path 2871 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); 2872 set_memory(malloc, rawidx); 2873 2874 // a normal slow-call doesn't change i_o, but an allocation does 2875 // we create a separate i_o projection for the normal control path 2876 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); 2877 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); 2878 2879 // put in an initialization barrier 2880 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 2881 rawoop)->as_Initialize(); 2882 assert(alloc->initialization() == init, "2-way macro link must work"); 2883 assert(init ->allocation() == alloc, "2-way macro link must work"); 2884 if (ReduceFieldZeroing && !raw_mem_only) { 2885 // Extract memory strands which may participate in the new object's 2886 // initialization, and source them from the new InitializeNode. 2887 // This will allow us to observe initializations when they occur, 2888 // and link them properly (as a group) to the InitializeNode. 2889 assert(init->in(InitializeNode::Memory) == malloc, ""); 2890 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 2891 init->set_req(InitializeNode::Memory, minit_in); 2892 record_for_igvn(minit_in); // fold it up later, if possible 2893 Node* minit_out = memory(rawidx); 2894 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 2895 if (oop_type->isa_aryptr()) { 2896 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 2897 int elemidx = C->get_alias_index(telemref); 2898 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 2899 } else if (oop_type->isa_instptr()) { 2900 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 2901 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 2902 ciField* field = ik->nonstatic_field_at(i); 2903 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 2904 continue; // do not bother to track really large numbers of fields 2905 // Find (or create) the alias category for this field: 2906 int fieldidx = C->alias_type(field)->index(); 2907 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 2908 } 2909 } 2910 } 2911 2912 // Cast raw oop to the real thing... 2913 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); 2914 javaoop = _gvn.transform(javaoop); 2915 C->set_recent_alloc(control(), javaoop); 2916 assert(just_allocated_object(control()) == javaoop, "just allocated"); 2917 2918 #ifdef ASSERT 2919 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 2920 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 2921 "Ideal_allocation works"); 2922 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 2923 "Ideal_allocation works"); 2924 if (alloc->is_AllocateArray()) { 2925 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 2926 "Ideal_allocation works"); 2927 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 2928 "Ideal_allocation works"); 2929 } else { 2930 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 2931 } 2932 } 2933 #endif //ASSERT 2934 2935 return javaoop; 2936 } 2937 2938 //---------------------------new_instance-------------------------------------- 2939 // This routine takes a klass_node which may be constant (for a static type) 2940 // or may be non-constant (for reflective code). It will work equally well 2941 // for either, and the graph will fold nicely if the optimizer later reduces 2942 // the type to a constant. 2943 // The optional arguments are for specialized use by intrinsics: 2944 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 2945 // - If 'raw_mem_only', do not cast the result to an oop. 2946 // - If 'return_size_val', report the the total object size to the caller. 2947 Node* GraphKit::new_instance(Node* klass_node, 2948 Node* extra_slow_test, 2949 bool raw_mem_only, // affect only raw memory 2950 Node* *return_size_val) { 2951 // Compute size in doublewords 2952 // The size is always an integral number of doublewords, represented 2953 // as a positive bytewise size stored in the klass's layout_helper. 2954 // The layout_helper also encodes (in a low bit) the need for a slow path. 2955 jint layout_con = Klass::_lh_neutral_value; 2956 Node* layout_val = get_layout_helper(klass_node, layout_con); 2957 int layout_is_con = (layout_val == NULL); 2958 2959 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 2960 // Generate the initial go-slow test. It's either ALWAYS (return a 2961 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 2962 // case) a computed value derived from the layout_helper. 2963 Node* initial_slow_test = NULL; 2964 if (layout_is_con) { 2965 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2966 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 2967 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 2968 2969 } else { // reflective case 2970 // This reflective path is used by Unsafe.allocateInstance. 2971 // (It may be stress-tested by specifying StressReflectiveCode.) 2972 // Basically, we want to get into the VM is there's an illegal argument. 2973 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 2974 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); 2975 if (extra_slow_test != intcon(0)) { 2976 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); 2977 } 2978 // (Macro-expander will further convert this to a Bool, if necessary.) 2979 } 2980 2981 // Find the size in bytes. This is easy; it's the layout_helper. 2982 // The size value must be valid even if the slow path is taken. 2983 Node* size = NULL; 2984 if (layout_is_con) { 2985 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 2986 } else { // reflective case 2987 // This reflective path is used by clone and Unsafe.allocateInstance. 2988 size = ConvI2X(layout_val); 2989 2990 // Clear the low bits to extract layout_helper_size_in_bytes: 2991 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 2992 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 2993 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); 2994 } 2995 if (return_size_val != NULL) { 2996 (*return_size_val) = size; 2997 } 2998 2999 // This is a precise notnull oop of the klass. 3000 // (Actually, it need not be precise if this is a reflective allocation.) 3001 // It's what we cast the result to. 3002 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3003 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3004 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3005 3006 // Now generate allocation code 3007 3008 // The entire memory state is needed for slow path of the allocation 3009 // since GC and deoptimization can happened. 3010 Node *mem = reset_memory(); 3011 set_all_memory(mem); // Create new memory state 3012 3013 AllocateNode* alloc 3014 = new (C, AllocateNode::ParmLimit) 3015 AllocateNode(C, AllocateNode::alloc_type(), 3016 control(), mem, i_o(), 3017 size, klass_node, 3018 initial_slow_test); 3019 3020 return set_output_for_allocation(alloc, oop_type, raw_mem_only); 3021 } 3022 3023 //-------------------------------new_array------------------------------------- 3024 // helper for both newarray and anewarray 3025 // The 'length' parameter is (obviously) the length of the array. 3026 // See comments on new_instance for the meaning of the other arguments. 3027 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3028 Node* length, // number of array elements 3029 int nargs, // number of arguments to push back for uncommon trap 3030 bool raw_mem_only, // affect only raw memory 3031 Node* *return_size_val) { 3032 jint layout_con = Klass::_lh_neutral_value; 3033 Node* layout_val = get_layout_helper(klass_node, layout_con); 3034 int layout_is_con = (layout_val == NULL); 3035 3036 if (!layout_is_con && !StressReflectiveCode && 3037 !too_many_traps(Deoptimization::Reason_class_check)) { 3038 // This is a reflective array creation site. 3039 // Optimistically assume that it is a subtype of Object[], 3040 // so that we can fold up all the address arithmetic. 3041 layout_con = Klass::array_layout_helper(T_OBJECT); 3042 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); 3043 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); 3044 { BuildCutout unless(this, bol_lh, PROB_MAX); 3045 _sp += nargs; 3046 uncommon_trap(Deoptimization::Reason_class_check, 3047 Deoptimization::Action_maybe_recompile); 3048 } 3049 layout_val = NULL; 3050 layout_is_con = true; 3051 } 3052 3053 // Generate the initial go-slow test. Make sure we do not overflow 3054 // if length is huge (near 2Gig) or negative! We do not need 3055 // exact double-words here, just a close approximation of needed 3056 // double-words. We can't add any offset or rounding bits, lest we 3057 // take a size -1 of bytes and make it positive. Use an unsigned 3058 // compare, so negative sizes look hugely positive. 3059 int fast_size_limit = FastAllocateSizeLimit; 3060 if (layout_is_con) { 3061 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3062 // Increase the size limit if we have exact knowledge of array type. 3063 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3064 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3065 } 3066 3067 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); 3068 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3069 if (initial_slow_test->is_Bool()) { 3070 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3071 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3072 } 3073 3074 // --- Size Computation --- 3075 // array_size = round_to_heap(array_header + (length << elem_shift)); 3076 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3077 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3078 // The rounding mask is strength-reduced, if possible. 3079 int round_mask = MinObjAlignmentInBytes - 1; 3080 Node* header_size = NULL; 3081 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3082 // (T_BYTE has the weakest alignment and size restrictions...) 3083 if (layout_is_con) { 3084 int hsize = Klass::layout_helper_header_size(layout_con); 3085 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3086 BasicType etype = Klass::layout_helper_element_type(layout_con); 3087 if ((round_mask & ~right_n_bits(eshift)) == 0) 3088 round_mask = 0; // strength-reduce it if it goes away completely 3089 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3090 assert(header_size_min <= hsize, "generic minimum is smallest"); 3091 header_size_min = hsize; 3092 header_size = intcon(hsize + round_mask); 3093 } else { 3094 Node* hss = intcon(Klass::_lh_header_size_shift); 3095 Node* hsm = intcon(Klass::_lh_header_size_mask); 3096 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); 3097 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); 3098 Node* mask = intcon(round_mask); 3099 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); 3100 } 3101 3102 Node* elem_shift = NULL; 3103 if (layout_is_con) { 3104 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3105 if (eshift != 0) 3106 elem_shift = intcon(eshift); 3107 } else { 3108 // There is no need to mask or shift this value. 3109 // The semantics of LShiftINode include an implicit mask to 0x1F. 3110 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3111 elem_shift = layout_val; 3112 } 3113 3114 // Transition to native address size for all offset calculations: 3115 Node* lengthx = ConvI2X(length); 3116 Node* headerx = ConvI2X(header_size); 3117 #ifdef _LP64 3118 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3119 if (tllen != NULL && tllen->_lo < 0) { 3120 // Add a manual constraint to a positive range. Cf. array_element_address. 3121 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3122 if (size_max > tllen->_hi) size_max = tllen->_hi; 3123 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3124 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); 3125 } 3126 } 3127 #endif 3128 3129 // Combine header size (plus rounding) and body size. Then round down. 3130 // This computation cannot overflow, because it is used only in two 3131 // places, one where the length is sharply limited, and the other 3132 // after a successful allocation. 3133 Node* abody = lengthx; 3134 if (elem_shift != NULL) 3135 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); 3136 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); 3137 if (round_mask != 0) { 3138 Node* mask = MakeConX(~round_mask); 3139 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); 3140 } 3141 // else if round_mask == 0, the size computation is self-rounding 3142 3143 if (return_size_val != NULL) { 3144 // This is the size 3145 (*return_size_val) = size; 3146 } 3147 3148 // Now generate allocation code 3149 3150 // The entire memory state is needed for slow path of the allocation 3151 // since GC and deoptimization can happened. 3152 Node *mem = reset_memory(); 3153 set_all_memory(mem); // Create new memory state 3154 3155 // Create the AllocateArrayNode and its result projections 3156 AllocateArrayNode* alloc 3157 = new (C, AllocateArrayNode::ParmLimit) 3158 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3159 control(), mem, i_o(), 3160 size, klass_node, 3161 initial_slow_test, 3162 length); 3163 3164 // Cast to correct type. Note that the klass_node may be constant or not, 3165 // and in the latter case the actual array type will be inexact also. 3166 // (This happens via a non-constant argument to inline_native_newArray.) 3167 // In any case, the value of klass_node provides the desired array type. 3168 const TypeInt* length_type = _gvn.find_int_type(length); 3169 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3170 if (ary_type->isa_aryptr() && length_type != NULL) { 3171 // Try to get a better type than POS for the size 3172 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3173 } 3174 3175 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); 3176 3177 // Cast length on remaining path to be as narrow as possible 3178 if (map()->find_edge(length) >= 0) { 3179 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3180 if (ccast != length) { 3181 _gvn.set_type_bottom(ccast); 3182 record_for_igvn(ccast); 3183 replace_in_map(length, ccast); 3184 } 3185 } 3186 3187 return javaoop; 3188 } 3189 3190 // The following "Ideal_foo" functions are placed here because they recognize 3191 // the graph shapes created by the functions immediately above. 3192 3193 //---------------------------Ideal_allocation---------------------------------- 3194 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3195 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3196 if (ptr == NULL) { // reduce dumb test in callers 3197 return NULL; 3198 } 3199 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3200 ptr = ptr->in(1); 3201 if (ptr == NULL) return NULL; 3202 } 3203 if (ptr->is_Proj()) { 3204 Node* allo = ptr->in(0); 3205 if (allo != NULL && allo->is_Allocate()) { 3206 return allo->as_Allocate(); 3207 } 3208 } 3209 // Report failure to match. 3210 return NULL; 3211 } 3212 3213 // Fancy version which also strips off an offset (and reports it to caller). 3214 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3215 intptr_t& offset) { 3216 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3217 if (base == NULL) return NULL; 3218 return Ideal_allocation(base, phase); 3219 } 3220 3221 // Trace Initialize <- Proj[Parm] <- Allocate 3222 AllocateNode* InitializeNode::allocation() { 3223 Node* rawoop = in(InitializeNode::RawAddress); 3224 if (rawoop->is_Proj()) { 3225 Node* alloc = rawoop->in(0); 3226 if (alloc->is_Allocate()) { 3227 return alloc->as_Allocate(); 3228 } 3229 } 3230 return NULL; 3231 } 3232 3233 // Trace Allocate -> Proj[Parm] -> Initialize 3234 InitializeNode* AllocateNode::initialization() { 3235 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3236 if (rawoop == NULL) return NULL; 3237 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3238 Node* init = rawoop->fast_out(i); 3239 if (init->is_Initialize()) { 3240 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3241 return init->as_Initialize(); 3242 } 3243 } 3244 return NULL; 3245 } 3246 3247 void GraphKit::g1_write_barrier_pre(Node* obj, 3248 Node* adr, 3249 uint alias_idx, 3250 Node* val, 3251 const TypeOopPtr* val_type, 3252 BasicType bt) { 3253 IdealKit ideal(gvn(), control(), merged_memory(), true); 3254 #define __ ideal. 3255 __ declares_done(); 3256 3257 Node* thread = __ thread(); 3258 3259 Node* no_ctrl = NULL; 3260 Node* no_base = __ top(); 3261 Node* zero = __ ConI(0); 3262 3263 float likely = PROB_LIKELY(0.999); 3264 float unlikely = PROB_UNLIKELY(0.999); 3265 3266 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3267 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3268 3269 // Offsets into the thread 3270 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3271 PtrQueue::byte_offset_of_active()); 3272 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3273 PtrQueue::byte_offset_of_index()); 3274 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3275 PtrQueue::byte_offset_of_buf()); 3276 // Now the actual pointers into the thread 3277 3278 // set_control( ctl); 3279 3280 Node* marking_adr = __ AddP(no_base, thread, __ ConX(marking_offset)); 3281 Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset)); 3282 Node* index_adr = __ AddP(no_base, thread, __ ConX(index_offset)); 3283 3284 // Now some of the values 3285 3286 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3287 3288 // if (!marking) 3289 __ if_then(marking, BoolTest::ne, zero); { 3290 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3291 3292 const Type* t1 = adr->bottom_type(); 3293 const Type* t2 = val->bottom_type(); 3294 3295 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx); 3296 // if (orig != NULL) 3297 __ if_then(orig, BoolTest::ne, null()); { 3298 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3299 3300 // load original value 3301 // alias_idx correct?? 3302 3303 // is the queue for this thread full? 3304 __ if_then(index, BoolTest::ne, zero, likely); { 3305 3306 // decrement the index 3307 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3308 Node* next_indexX = next_index; 3309 #ifdef _LP64 3310 // We could refine the type for what it's worth 3311 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3312 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3313 #endif // _LP64 3314 3315 // Now get the buffer location we will log the original value into and store it 3316 3317 Node *log_addr = __ AddP(no_base, buffer, next_indexX); 3318 // __ store(__ ctrl(), log_addr, orig, T_OBJECT, C->get_alias_index(TypeOopPtr::BOTTOM)); 3319 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw); 3320 3321 3322 // update the index 3323 // __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3324 // This is a hack to force this store to occur before the oop store that is coming up 3325 __ store(__ ctrl(), index_adr, next_index, T_INT, C->get_alias_index(TypeOopPtr::BOTTOM)); 3326 3327 } __ else_(); { 3328 3329 // logging buffer is full, call the runtime 3330 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3331 // __ make_leaf_call(tf, OptoRuntime::g1_wb_pre_Java(), "g1_wb_pre", orig, thread); 3332 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, thread); 3333 } __ end_if(); 3334 } __ end_if(); 3335 } __ end_if(); 3336 3337 __ drain_delay_transform(); 3338 set_control( __ ctrl()); 3339 set_all_memory( __ merged_memory()); 3340 3341 #undef __ 3342 } 3343 3344 // 3345 // Update the card table and add card address to the queue 3346 // 3347 void GraphKit::g1_mark_card(IdealKit* ideal, Node* card_adr, Node* store, Node* index, Node* index_adr, Node* buffer, const TypeFunc* tf) { 3348 #define __ ideal-> 3349 Node* zero = __ ConI(0); 3350 Node* no_base = __ top(); 3351 BasicType card_bt = T_BYTE; 3352 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3353 __ storeCM(__ ctrl(), card_adr, zero, store, card_bt, Compile::AliasIdxRaw); 3354 3355 // Now do the queue work 3356 __ if_then(index, BoolTest::ne, zero); { 3357 3358 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3359 Node* next_indexX = next_index; 3360 #ifdef _LP64 3361 // We could refine the type for what it's worth 3362 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3363 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3364 #endif // _LP64 3365 Node* log_addr = __ AddP(no_base, buffer, next_indexX); 3366 3367 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3368 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3369 3370 } __ else_(); { 3371 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3372 } __ end_if(); 3373 #undef __ 3374 } 3375 3376 void GraphKit::g1_write_barrier_post(Node* store, 3377 Node* obj, 3378 Node* adr, 3379 uint alias_idx, 3380 Node* val, 3381 BasicType bt, 3382 bool use_precise) { 3383 // If we are writing a NULL then we need no post barrier 3384 3385 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3386 // Must be NULL 3387 const Type* t = val->bottom_type(); 3388 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3389 // No post barrier if writing NULLx 3390 return; 3391 } 3392 3393 if (!use_precise) { 3394 // All card marks for a (non-array) instance are in one place: 3395 adr = obj; 3396 } 3397 // (Else it's an array (or unknown), and we want more precise card marks.) 3398 assert(adr != NULL, ""); 3399 3400 IdealKit ideal(gvn(), control(), merged_memory(), true); 3401 #define __ ideal. 3402 __ declares_done(); 3403 3404 Node* thread = __ thread(); 3405 3406 Node* no_ctrl = NULL; 3407 Node* no_base = __ top(); 3408 float likely = PROB_LIKELY(0.999); 3409 float unlikely = PROB_UNLIKELY(0.999); 3410 Node* zero = __ ConI(0); 3411 Node* zeroX = __ ConX(0); 3412 3413 // Get the alias_index for raw card-mark memory 3414 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3415 3416 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3417 3418 // Offsets into the thread 3419 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3420 PtrQueue::byte_offset_of_index()); 3421 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3422 PtrQueue::byte_offset_of_buf()); 3423 3424 // Pointers into the thread 3425 3426 Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset)); 3427 Node* index_adr = __ AddP(no_base, thread, __ ConX(index_offset)); 3428 3429 // Now some values 3430 3431 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3432 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3433 3434 3435 // Convert the store obj pointer to an int prior to doing math on it 3436 // Use addr not obj gets accurate card marks 3437 3438 // Node* cast = __ CastPX(no_ctrl, adr /* obj */); 3439 3440 // Must use ctrl to prevent "integerized oop" existing across safepoint 3441 Node* cast = __ CastPX(__ ctrl(), ( use_precise ? adr : obj )); 3442 3443 // Divide pointer by card size 3444 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3445 3446 // Combine card table base and card offset 3447 Node *card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3448 3449 // If we know the value being stored does it cross regions? 3450 3451 if (val != NULL) { 3452 // Does the store cause us to cross regions? 3453 3454 // Should be able to do an unsigned compare of region_size instead of 3455 // and extra shift. Do we have an unsigned compare?? 3456 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3457 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3458 3459 // if (xor_res == 0) same region so skip 3460 __ if_then(xor_res, BoolTest::ne, zeroX); { 3461 3462 // No barrier if we are storing a NULL 3463 __ if_then(val, BoolTest::ne, null(), unlikely); { 3464 3465 // Ok must mark the card if not already dirty 3466 3467 // load the original value of the card 3468 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3469 3470 __ if_then(card_val, BoolTest::ne, zero); { 3471 g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf); 3472 } __ end_if(); 3473 } __ end_if(); 3474 } __ end_if(); 3475 } else { 3476 g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf); 3477 } 3478 3479 3480 __ drain_delay_transform(); 3481 set_control( __ ctrl()); 3482 set_all_memory( __ merged_memory()); 3483 #undef __ 3484 3485 }