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()->_reexecute; 628 } 629 PreserveReexecuteState::~PreserveReexecuteState() { 630 _kit->jvms()->_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 state for them here 808 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 809 should_reexecute_implied_by_bytecode(out_jvms)) { 810 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 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 1406 void GraphKit::pre_barrier(Node* ctl, 1407 Node* obj, 1408 Node* adr, 1409 uint adr_idx, 1410 Node* val, 1411 const TypeOopPtr* val_type, 1412 BasicType bt) { 1413 BarrierSet* bs = Universe::heap()->barrier_set(); 1414 set_control(ctl); 1415 switch (bs->kind()) { 1416 case BarrierSet::G1SATBCT: 1417 case BarrierSet::G1SATBCTLogging: 1418 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt); 1419 break; 1420 1421 case BarrierSet::CardTableModRef: 1422 case BarrierSet::CardTableExtension: 1423 case BarrierSet::ModRef: 1424 break; 1425 1426 case BarrierSet::Other: 1427 default : 1428 ShouldNotReachHere(); 1429 1430 } 1431 } 1432 1433 void GraphKit::post_barrier(Node* ctl, 1434 Node* store, 1435 Node* obj, 1436 Node* adr, 1437 uint adr_idx, 1438 Node* val, 1439 BasicType bt, 1440 bool use_precise) { 1441 BarrierSet* bs = Universe::heap()->barrier_set(); 1442 set_control(ctl); 1443 switch (bs->kind()) { 1444 case BarrierSet::G1SATBCT: 1445 case BarrierSet::G1SATBCTLogging: 1446 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1447 break; 1448 1449 case BarrierSet::CardTableModRef: 1450 case BarrierSet::CardTableExtension: 1451 write_barrier_post(store, obj, adr, val, use_precise); 1452 break; 1453 1454 case BarrierSet::ModRef: 1455 break; 1456 1457 case BarrierSet::Other: 1458 default : 1459 ShouldNotReachHere(); 1460 1461 } 1462 } 1463 1464 Node* GraphKit::store_oop(Node* ctl, 1465 Node* obj, 1466 Node* adr, 1467 const TypePtr* adr_type, 1468 Node* val, 1469 const TypeOopPtr* val_type, 1470 BasicType bt, 1471 bool use_precise) { 1472 1473 set_control(ctl); 1474 if (stopped()) return top(); // Dead path ? 1475 1476 assert(bt == T_OBJECT, "sanity"); 1477 assert(val != NULL, "not dead path"); 1478 uint adr_idx = C->get_alias_index(adr_type); 1479 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1480 1481 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt); 1482 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1483 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1484 return store; 1485 } 1486 1487 // Could be an array or object we don't know at compile time (unsafe ref.) 1488 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1489 Node* obj, // containing obj 1490 Node* adr, // actual adress to store val at 1491 const TypePtr* adr_type, 1492 Node* val, 1493 BasicType bt) { 1494 Compile::AliasType* at = C->alias_type(adr_type); 1495 const TypeOopPtr* val_type = NULL; 1496 if (adr_type->isa_instptr()) { 1497 if (at->field() != NULL) { 1498 // known field. This code is a copy of the do_put_xxx logic. 1499 ciField* field = at->field(); 1500 if (!field->type()->is_loaded()) { 1501 val_type = TypeInstPtr::BOTTOM; 1502 } else { 1503 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1504 } 1505 } 1506 } else if (adr_type->isa_aryptr()) { 1507 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1508 } 1509 if (val_type == NULL) { 1510 val_type = TypeInstPtr::BOTTOM; 1511 } 1512 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); 1513 } 1514 1515 1516 //-------------------------array_element_address------------------------- 1517 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1518 const TypeInt* sizetype) { 1519 uint shift = exact_log2(type2aelembytes(elembt)); 1520 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1521 1522 // short-circuit a common case (saves lots of confusing waste motion) 1523 jint idx_con = find_int_con(idx, -1); 1524 if (idx_con >= 0) { 1525 intptr_t offset = header + ((intptr_t)idx_con << shift); 1526 return basic_plus_adr(ary, offset); 1527 } 1528 1529 // must be correct type for alignment purposes 1530 Node* base = basic_plus_adr(ary, header); 1531 #ifdef _LP64 1532 // The scaled index operand to AddP must be a clean 64-bit value. 1533 // Java allows a 32-bit int to be incremented to a negative 1534 // value, which appears in a 64-bit register as a large 1535 // positive number. Using that large positive number as an 1536 // operand in pointer arithmetic has bad consequences. 1537 // On the other hand, 32-bit overflow is rare, and the possibility 1538 // can often be excluded, if we annotate the ConvI2L node with 1539 // a type assertion that its value is known to be a small positive 1540 // number. (The prior range check has ensured this.) 1541 // This assertion is used by ConvI2LNode::Ideal. 1542 int index_max = max_jint - 1; // array size is max_jint, index is one less 1543 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1544 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1545 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); 1546 #endif 1547 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); 1548 return basic_plus_adr(ary, base, scale); 1549 } 1550 1551 //-------------------------load_array_element------------------------- 1552 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1553 const Type* elemtype = arytype->elem(); 1554 BasicType elembt = elemtype->array_element_basic_type(); 1555 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1556 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1557 return ld; 1558 } 1559 1560 //-------------------------set_arguments_for_java_call------------------------- 1561 // Arguments (pre-popped from the stack) are taken from the JVMS. 1562 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1563 // Add the call arguments: 1564 uint nargs = call->method()->arg_size(); 1565 for (uint i = 0; i < nargs; i++) { 1566 Node* arg = argument(i); 1567 call->init_req(i + TypeFunc::Parms, arg); 1568 } 1569 } 1570 1571 //---------------------------set_edges_for_java_call--------------------------- 1572 // Connect a newly created call into the current JVMS. 1573 // A return value node (if any) is returned from set_edges_for_java_call. 1574 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) { 1575 1576 // Add the predefined inputs: 1577 call->init_req( TypeFunc::Control, control() ); 1578 call->init_req( TypeFunc::I_O , i_o() ); 1579 call->init_req( TypeFunc::Memory , reset_memory() ); 1580 call->init_req( TypeFunc::FramePtr, frameptr() ); 1581 call->init_req( TypeFunc::ReturnAdr, top() ); 1582 1583 add_safepoint_edges(call, must_throw); 1584 1585 Node* xcall = _gvn.transform(call); 1586 1587 if (xcall == top()) { 1588 set_control(top()); 1589 return; 1590 } 1591 assert(xcall == call, "call identity is stable"); 1592 1593 // Re-use the current map to produce the result. 1594 1595 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); 1596 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O ))); 1597 set_all_memory_call(xcall); 1598 1599 //return xcall; // no need, caller already has it 1600 } 1601 1602 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) { 1603 if (stopped()) return top(); // maybe the call folded up? 1604 1605 // Capture the return value, if any. 1606 Node* ret; 1607 if (call->method() == NULL || 1608 call->method()->return_type()->basic_type() == T_VOID) 1609 ret = top(); 1610 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 1611 1612 // Note: Since any out-of-line call can produce an exception, 1613 // we always insert an I_O projection from the call into the result. 1614 1615 make_slow_call_ex(call, env()->Throwable_klass(), false); 1616 1617 return ret; 1618 } 1619 1620 //--------------------set_predefined_input_for_runtime_call-------------------- 1621 // Reading and setting the memory state is way conservative here. 1622 // The real problem is that I am not doing real Type analysis on memory, 1623 // so I cannot distinguish card mark stores from other stores. Across a GC 1624 // point the Store Barrier and the card mark memory has to agree. I cannot 1625 // have a card mark store and its barrier split across the GC point from 1626 // either above or below. Here I get that to happen by reading ALL of memory. 1627 // A better answer would be to separate out card marks from other memory. 1628 // For now, return the input memory state, so that it can be reused 1629 // after the call, if this call has restricted memory effects. 1630 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1631 // Set fixed predefined input arguments 1632 Node* memory = reset_memory(); 1633 call->init_req( TypeFunc::Control, control() ); 1634 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1635 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1636 call->init_req( TypeFunc::FramePtr, frameptr() ); 1637 call->init_req( TypeFunc::ReturnAdr, top() ); 1638 return memory; 1639 } 1640 1641 //-------------------set_predefined_output_for_runtime_call-------------------- 1642 // Set control and memory (not i_o) from the call. 1643 // If keep_mem is not NULL, use it for the output state, 1644 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1645 // If hook_mem is NULL, this call produces no memory effects at all. 1646 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1647 // then only that memory slice is taken from the call. 1648 // In the last case, we must put an appropriate memory barrier before 1649 // the call, so as to create the correct anti-dependencies on loads 1650 // preceding the call. 1651 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1652 Node* keep_mem, 1653 const TypePtr* hook_mem) { 1654 // no i/o 1655 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); 1656 if (keep_mem) { 1657 // First clone the existing memory state 1658 set_all_memory(keep_mem); 1659 if (hook_mem != NULL) { 1660 // Make memory for the call 1661 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1662 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1663 // We also use hook_mem to extract specific effects from arraycopy stubs. 1664 set_memory(mem, hook_mem); 1665 } 1666 // ...else the call has NO memory effects. 1667 1668 // Make sure the call advertises its memory effects precisely. 1669 // This lets us build accurate anti-dependences in gcm.cpp. 1670 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1671 "call node must be constructed correctly"); 1672 } else { 1673 assert(hook_mem == NULL, ""); 1674 // This is not a "slow path" call; all memory comes from the call. 1675 set_all_memory_call(call); 1676 } 1677 } 1678 1679 //------------------------------increment_counter------------------------------ 1680 // for statistics: increment a VM counter by 1 1681 1682 void GraphKit::increment_counter(address counter_addr) { 1683 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1684 increment_counter(adr1); 1685 } 1686 1687 void GraphKit::increment_counter(Node* counter_addr) { 1688 int adr_type = Compile::AliasIdxRaw; 1689 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 1690 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 1691 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); 1692 } 1693 1694 1695 //------------------------------uncommon_trap---------------------------------- 1696 // Bail out to the interpreter in mid-method. Implemented by calling the 1697 // uncommon_trap blob. This helper function inserts a runtime call with the 1698 // right debug info. 1699 void GraphKit::uncommon_trap(int trap_request, 1700 ciKlass* klass, const char* comment, 1701 bool must_throw, 1702 bool keep_exact_action) { 1703 if (failing()) stop(); 1704 if (stopped()) return; // trap reachable? 1705 1706 // Note: If ProfileTraps is true, and if a deopt. actually 1707 // occurs here, the runtime will make sure an MDO exists. There is 1708 // no need to call method()->build_method_data() at this point. 1709 1710 #ifdef ASSERT 1711 if (!must_throw) { 1712 // Make sure the stack has at least enough depth to execute 1713 // the current bytecode. 1714 int inputs, ignore; 1715 if (compute_stack_effects(inputs, ignore)) { 1716 assert(sp() >= inputs, "must have enough JVMS stack to execute"); 1717 // It is a frequent error in library_call.cpp to issue an 1718 // uncommon trap with the _sp value already popped. 1719 } 1720 } 1721 #endif 1722 1723 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1724 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1725 1726 switch (action) { 1727 case Deoptimization::Action_maybe_recompile: 1728 case Deoptimization::Action_reinterpret: 1729 // Temporary fix for 6529811 to allow virtual calls to be sure they 1730 // get the chance to go from mono->bi->mega 1731 if (!keep_exact_action && 1732 Deoptimization::trap_request_index(trap_request) < 0 && 1733 too_many_recompiles(reason)) { 1734 // This BCI is causing too many recompilations. 1735 action = Deoptimization::Action_none; 1736 trap_request = Deoptimization::make_trap_request(reason, action); 1737 } else { 1738 C->set_trap_can_recompile(true); 1739 } 1740 break; 1741 case Deoptimization::Action_make_not_entrant: 1742 C->set_trap_can_recompile(true); 1743 break; 1744 #ifdef ASSERT 1745 case Deoptimization::Action_none: 1746 case Deoptimization::Action_make_not_compilable: 1747 break; 1748 default: 1749 assert(false, "bad action"); 1750 #endif 1751 } 1752 1753 if (TraceOptoParse) { 1754 char buf[100]; 1755 tty->print_cr("Uncommon trap %s at bci:%d", 1756 Deoptimization::format_trap_request(buf, sizeof(buf), 1757 trap_request), bci()); 1758 } 1759 1760 CompileLog* log = C->log(); 1761 if (log != NULL) { 1762 int kid = (klass == NULL)? -1: log->identify(klass); 1763 log->begin_elem("uncommon_trap bci='%d'", bci()); 1764 char buf[100]; 1765 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1766 trap_request)); 1767 if (kid >= 0) log->print(" klass='%d'", kid); 1768 if (comment != NULL) log->print(" comment='%s'", comment); 1769 log->end_elem(); 1770 } 1771 1772 // Make sure any guarding test views this path as very unlikely 1773 Node *i0 = control()->in(0); 1774 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1775 IfNode *iff = i0->as_If(); 1776 float f = iff->_prob; // Get prob 1777 if (control()->Opcode() == Op_IfTrue) { 1778 if (f > PROB_UNLIKELY_MAG(4)) 1779 iff->_prob = PROB_MIN; 1780 } else { 1781 if (f < PROB_LIKELY_MAG(4)) 1782 iff->_prob = PROB_MAX; 1783 } 1784 } 1785 1786 // Clear out dead values from the debug info. 1787 kill_dead_locals(); 1788 1789 // Now insert the uncommon trap subroutine call 1790 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); 1791 const TypePtr* no_memory_effects = NULL; 1792 // Pass the index of the class to be loaded 1793 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1794 (must_throw ? RC_MUST_THROW : 0), 1795 OptoRuntime::uncommon_trap_Type(), 1796 call_addr, "uncommon_trap", no_memory_effects, 1797 intcon(trap_request)); 1798 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 1799 "must extract request correctly from the graph"); 1800 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 1801 1802 call->set_req(TypeFunc::ReturnAdr, returnadr()); 1803 // The debug info is the only real input to this call. 1804 1805 // Halt-and-catch fire here. The above call should never return! 1806 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); 1807 _gvn.set_type_bottom(halt); 1808 root()->add_req(halt); 1809 1810 stop_and_kill_map(); 1811 } 1812 1813 1814 //--------------------------just_allocated_object------------------------------ 1815 // Report the object that was just allocated. 1816 // It must be the case that there are no intervening safepoints. 1817 // We use this to determine if an object is so "fresh" that 1818 // it does not require card marks. 1819 Node* GraphKit::just_allocated_object(Node* current_control) { 1820 if (C->recent_alloc_ctl() == current_control) 1821 return C->recent_alloc_obj(); 1822 return NULL; 1823 } 1824 1825 1826 void GraphKit::round_double_arguments(ciMethod* dest_method) { 1827 // (Note: TypeFunc::make has a cache that makes this fast.) 1828 const TypeFunc* tf = TypeFunc::make(dest_method); 1829 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 1830 for (int j = 0; j < nargs; j++) { 1831 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 1832 if( targ->basic_type() == T_DOUBLE ) { 1833 // If any parameters are doubles, they must be rounded before 1834 // the call, dstore_rounding does gvn.transform 1835 Node *arg = argument(j); 1836 arg = dstore_rounding(arg); 1837 set_argument(j, arg); 1838 } 1839 } 1840 } 1841 1842 void GraphKit::round_double_result(ciMethod* dest_method) { 1843 // A non-strict method may return a double value which has an extended 1844 // exponent, but this must not be visible in a caller which is 'strict' 1845 // If a strict caller invokes a non-strict callee, round a double result 1846 1847 BasicType result_type = dest_method->return_type()->basic_type(); 1848 assert( method() != NULL, "must have caller context"); 1849 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 1850 // Destination method's return value is on top of stack 1851 // dstore_rounding() does gvn.transform 1852 Node *result = pop_pair(); 1853 result = dstore_rounding(result); 1854 push_pair(result); 1855 } 1856 } 1857 1858 // rounding for strict float precision conformance 1859 Node* GraphKit::precision_rounding(Node* n) { 1860 return UseStrictFP && _method->flags().is_strict() 1861 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 1862 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) 1863 : n; 1864 } 1865 1866 // rounding for strict double precision conformance 1867 Node* GraphKit::dprecision_rounding(Node *n) { 1868 return UseStrictFP && _method->flags().is_strict() 1869 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 1870 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1871 : n; 1872 } 1873 1874 // rounding for non-strict double stores 1875 Node* GraphKit::dstore_rounding(Node* n) { 1876 return Matcher::strict_fp_requires_explicit_rounding 1877 && UseSSE <= 1 1878 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1879 : n; 1880 } 1881 1882 //============================================================================= 1883 // Generate a fast path/slow path idiom. Graph looks like: 1884 // [foo] indicates that 'foo' is a parameter 1885 // 1886 // [in] NULL 1887 // \ / 1888 // CmpP 1889 // Bool ne 1890 // If 1891 // / \ 1892 // True False-<2> 1893 // / | 1894 // / cast_not_null 1895 // Load | | ^ 1896 // [fast_test] | | 1897 // gvn to opt_test | | 1898 // / \ | <1> 1899 // True False | 1900 // | \\ | 1901 // [slow_call] \[fast_result] 1902 // Ctl Val \ \ 1903 // | \ \ 1904 // Catch <1> \ \ 1905 // / \ ^ \ \ 1906 // Ex No_Ex | \ \ 1907 // | \ \ | \ <2> \ 1908 // ... \ [slow_res] | | \ [null_result] 1909 // \ \--+--+--- | | 1910 // \ | / \ | / 1911 // --------Region Phi 1912 // 1913 //============================================================================= 1914 // Code is structured as a series of driver functions all called 'do_XXX' that 1915 // call a set of helper functions. Helper functions first, then drivers. 1916 1917 //------------------------------null_check_oop--------------------------------- 1918 // Null check oop. Set null-path control into Region in slot 3. 1919 // Make a cast-not-nullness use the other not-null control. Return cast. 1920 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 1921 bool never_see_null) { 1922 // Initial NULL check taken path 1923 (*null_control) = top(); 1924 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 1925 1926 // Generate uncommon_trap: 1927 if (never_see_null && (*null_control) != top()) { 1928 // If we see an unexpected null at a check-cast we record it and force a 1929 // recompile; the offending check-cast will be compiled to handle NULLs. 1930 // If we see more than one offending BCI, then all checkcasts in the 1931 // method will be compiled to handle NULLs. 1932 PreserveJVMState pjvms(this); 1933 set_control(*null_control); 1934 replace_in_map(value, null()); 1935 uncommon_trap(Deoptimization::Reason_null_check, 1936 Deoptimization::Action_make_not_entrant); 1937 (*null_control) = top(); // NULL path is dead 1938 } 1939 1940 // Cast away null-ness on the result 1941 return cast; 1942 } 1943 1944 //------------------------------opt_iff---------------------------------------- 1945 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 1946 // Return slow-path control. 1947 Node* GraphKit::opt_iff(Node* region, Node* iff) { 1948 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 1949 1950 // Fast path taken; set region slot 2 1951 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); 1952 region->init_req(2,fast_taken); // Capture fast-control 1953 1954 // Fast path not-taken, i.e. slow path 1955 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); 1956 return slow_taken; 1957 } 1958 1959 //-----------------------------make_runtime_call------------------------------- 1960 Node* GraphKit::make_runtime_call(int flags, 1961 const TypeFunc* call_type, address call_addr, 1962 const char* call_name, 1963 const TypePtr* adr_type, 1964 // The following parms are all optional. 1965 // The first NULL ends the list. 1966 Node* parm0, Node* parm1, 1967 Node* parm2, Node* parm3, 1968 Node* parm4, Node* parm5, 1969 Node* parm6, Node* parm7) { 1970 // Slow-path call 1971 int size = call_type->domain()->cnt(); 1972 bool is_leaf = !(flags & RC_NO_LEAF); 1973 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 1974 if (call_name == NULL) { 1975 assert(!is_leaf, "must supply name for leaf"); 1976 call_name = OptoRuntime::stub_name(call_addr); 1977 } 1978 CallNode* call; 1979 if (!is_leaf) { 1980 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, 1981 bci(), adr_type); 1982 } else if (flags & RC_NO_FP) { 1983 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 1984 } else { 1985 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); 1986 } 1987 1988 // The following is similar to set_edges_for_java_call, 1989 // except that the memory effects of the call are restricted to AliasIdxRaw. 1990 1991 // Slow path call has no side-effects, uses few values 1992 bool wide_in = !(flags & RC_NARROW_MEM); 1993 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 1994 1995 Node* prev_mem = NULL; 1996 if (wide_in) { 1997 prev_mem = set_predefined_input_for_runtime_call(call); 1998 } else { 1999 assert(!wide_out, "narrow in => narrow out"); 2000 Node* narrow_mem = memory(adr_type); 2001 prev_mem = reset_memory(); 2002 map()->set_memory(narrow_mem); 2003 set_predefined_input_for_runtime_call(call); 2004 } 2005 2006 // Hook each parm in order. Stop looking at the first NULL. 2007 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2008 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2009 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2010 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2011 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2012 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2013 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2014 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2015 /* close each nested if ===> */ } } } } } } } } 2016 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2017 2018 if (!is_leaf) { 2019 // Non-leaves can block and take safepoints: 2020 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2021 } 2022 // Non-leaves can throw exceptions: 2023 if (has_io) { 2024 call->set_req(TypeFunc::I_O, i_o()); 2025 } 2026 2027 if (flags & RC_UNCOMMON) { 2028 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2029 // (An "if" probability corresponds roughly to an unconditional count. 2030 // Sort of.) 2031 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2032 } 2033 2034 Node* c = _gvn.transform(call); 2035 assert(c == call, "cannot disappear"); 2036 2037 if (wide_out) { 2038 // Slow path call has full side-effects. 2039 set_predefined_output_for_runtime_call(call); 2040 } else { 2041 // Slow path call has few side-effects, and/or sets few values. 2042 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2043 } 2044 2045 if (has_io) { 2046 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); 2047 } 2048 return call; 2049 2050 } 2051 2052 //------------------------------merge_memory----------------------------------- 2053 // Merge memory from one path into the current memory state. 2054 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2055 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2056 Node* old_slice = mms.force_memory(); 2057 Node* new_slice = mms.memory2(); 2058 if (old_slice != new_slice) { 2059 PhiNode* phi; 2060 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2061 phi = new_slice->as_Phi(); 2062 #ifdef ASSERT 2063 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2064 old_slice = old_slice->in(new_path); 2065 // Caller is responsible for ensuring that any pre-existing 2066 // phis are already aware of old memory. 2067 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2068 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2069 #endif 2070 mms.set_memory(phi); 2071 } else { 2072 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2073 _gvn.set_type(phi, Type::MEMORY); 2074 phi->set_req(new_path, new_slice); 2075 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2076 } 2077 } 2078 } 2079 } 2080 2081 //------------------------------make_slow_call_ex------------------------------ 2082 // Make the exception handler hookups for the slow call 2083 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2084 if (stopped()) return; 2085 2086 // Make a catch node with just two handlers: fall-through and catch-all 2087 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2088 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); 2089 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2090 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2091 2092 { PreserveJVMState pjvms(this); 2093 set_control(excp); 2094 set_i_o(i_o); 2095 2096 if (excp != top()) { 2097 // Create an exception state also. 2098 // Use an exact type if the caller has specified a specific exception. 2099 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2100 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); 2101 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2102 } 2103 } 2104 2105 // Get the no-exception control from the CatchNode. 2106 set_control(norm); 2107 } 2108 2109 2110 //-------------------------------gen_subtype_check----------------------------- 2111 // Generate a subtyping check. Takes as input the subtype and supertype. 2112 // Returns 2 values: sets the default control() to the true path and returns 2113 // the false path. Only reads invariant memory; sets no (visible) memory. 2114 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2115 // but that's not exposed to the optimizer. This call also doesn't take in an 2116 // Object; if you wish to check an Object you need to load the Object's class 2117 // prior to coming here. 2118 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2119 // Fast check for identical types, perhaps identical constants. 2120 // The types can even be identical non-constants, in cases 2121 // involving Array.newInstance, Object.clone, etc. 2122 if (subklass == superklass) 2123 return top(); // false path is dead; no test needed. 2124 2125 if (_gvn.type(superklass)->singleton()) { 2126 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2127 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2128 2129 // In the common case of an exact superklass, try to fold up the 2130 // test before generating code. You may ask, why not just generate 2131 // the code and then let it fold up? The answer is that the generated 2132 // code will necessarily include null checks, which do not always 2133 // completely fold away. If they are also needless, then they turn 2134 // into a performance loss. Example: 2135 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2136 // Here, the type of 'fa' is often exact, so the store check 2137 // of fa[1]=x will fold up, without testing the nullness of x. 2138 switch (static_subtype_check(superk, subk)) { 2139 case SSC_always_false: 2140 { 2141 Node* always_fail = control(); 2142 set_control(top()); 2143 return always_fail; 2144 } 2145 case SSC_always_true: 2146 return top(); 2147 case SSC_easy_test: 2148 { 2149 // Just do a direct pointer compare and be done. 2150 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); 2151 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2152 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2153 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); 2154 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2155 } 2156 case SSC_full_test: 2157 break; 2158 default: 2159 ShouldNotReachHere(); 2160 } 2161 } 2162 2163 // %%% Possible further optimization: Even if the superklass is not exact, 2164 // if the subklass is the unique subtype of the superklass, the check 2165 // will always succeed. We could leave a dependency behind to ensure this. 2166 2167 // First load the super-klass's check-offset 2168 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); 2169 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2170 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); 2171 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2172 2173 // Load from the sub-klass's super-class display list, or a 1-word cache of 2174 // the secondary superclass list, or a failing value with a sentinel offset 2175 // if the super-klass is an interface or exceptionally deep in the Java 2176 // hierarchy and we have to scan the secondary superclass list the hard way. 2177 // Worst-case type is a little odd: NULL is allowed as a result (usually 2178 // klass loads can never produce a NULL). 2179 Node *chk_off_X = ConvI2X(chk_off); 2180 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); 2181 // For some types like interfaces the following loadKlass is from a 1-word 2182 // cache which is mutable so can't use immutable memory. Other 2183 // types load from the super-class display table which is immutable. 2184 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2185 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2186 2187 // Compile speed common case: ARE a subtype and we canNOT fail 2188 if( superklass == nkls ) 2189 return top(); // false path is dead; no test needed. 2190 2191 // See if we get an immediate positive hit. Happens roughly 83% of the 2192 // time. Test to see if the value loaded just previously from the subklass 2193 // is exactly the superklass. 2194 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); 2195 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); 2196 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2197 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); 2198 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); 2199 2200 // Compile speed common case: Check for being deterministic right now. If 2201 // chk_off is a constant and not equal to cacheoff then we are NOT a 2202 // subklass. In this case we need exactly the 1 test above and we can 2203 // return those results immediately. 2204 if (!might_be_cache) { 2205 Node* not_subtype_ctrl = control(); 2206 set_control(iftrue1); // We need exactly the 1 test above 2207 return not_subtype_ctrl; 2208 } 2209 2210 // Gather the various success & failures here 2211 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); 2212 record_for_igvn(r_ok_subtype); 2213 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); 2214 record_for_igvn(r_not_subtype); 2215 2216 r_ok_subtype->init_req(1, iftrue1); 2217 2218 // Check for immediate negative hit. Happens roughly 11% of the time (which 2219 // is roughly 63% of the remaining cases). Test to see if the loaded 2220 // check-offset points into the subklass display list or the 1-element 2221 // cache. If it points to the display (and NOT the cache) and the display 2222 // missed then it's not a subtype. 2223 Node *cacheoff = _gvn.intcon(cacheoff_con); 2224 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); 2225 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); 2226 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2227 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); 2228 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); 2229 2230 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2231 // No performance impact (too rare) but allows sharing of secondary arrays 2232 // which has some footprint reduction. 2233 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); 2234 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); 2235 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2236 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); 2237 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); 2238 2239 // -- Roads not taken here: -- 2240 // We could also have chosen to perform the self-check at the beginning 2241 // of this code sequence, as the assembler does. This would not pay off 2242 // the same way, since the optimizer, unlike the assembler, can perform 2243 // static type analysis to fold away many successful self-checks. 2244 // Non-foldable self checks work better here in second position, because 2245 // the initial primary superclass check subsumes a self-check for most 2246 // types. An exception would be a secondary type like array-of-interface, 2247 // which does not appear in its own primary supertype display. 2248 // Finally, we could have chosen to move the self-check into the 2249 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2250 // dependent manner. But it is worthwhile to have the check here, 2251 // where it can be perhaps be optimized. The cost in code space is 2252 // small (register compare, branch). 2253 2254 // Now do a linear scan of the secondary super-klass array. Again, no real 2255 // performance impact (too rare) but it's gotta be done. 2256 // Since the code is rarely used, there is no penalty for moving it 2257 // out of line, and it can only improve I-cache density. 2258 // The decision to inline or out-of-line this final check is platform 2259 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2260 Node* psc = _gvn.transform( 2261 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2262 2263 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); 2264 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); 2265 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2266 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); 2267 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); 2268 2269 // Return false path; set default control to true path. 2270 set_control( _gvn.transform(r_ok_subtype) ); 2271 return _gvn.transform(r_not_subtype); 2272 } 2273 2274 //----------------------------static_subtype_check----------------------------- 2275 // Shortcut important common cases when superklass is exact: 2276 // (0) superklass is java.lang.Object (can occur in reflective code) 2277 // (1) subklass is already limited to a subtype of superklass => always ok 2278 // (2) subklass does not overlap with superklass => always fail 2279 // (3) superklass has NO subtypes and we can check with a simple compare. 2280 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2281 if (StressReflectiveCode) { 2282 return SSC_full_test; // Let caller generate the general case. 2283 } 2284 2285 if (superk == env()->Object_klass()) { 2286 return SSC_always_true; // (0) this test cannot fail 2287 } 2288 2289 ciType* superelem = superk; 2290 if (superelem->is_array_klass()) 2291 superelem = superelem->as_array_klass()->base_element_type(); 2292 2293 if (!subk->is_interface()) { // cannot trust static interface types yet 2294 if (subk->is_subtype_of(superk)) { 2295 return SSC_always_true; // (1) false path dead; no dynamic test needed 2296 } 2297 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2298 !superk->is_subtype_of(subk)) { 2299 return SSC_always_false; 2300 } 2301 } 2302 2303 // If casting to an instance klass, it must have no subtypes 2304 if (superk->is_interface()) { 2305 // Cannot trust interfaces yet. 2306 // %%% S.B. superk->nof_implementors() == 1 2307 } else if (superelem->is_instance_klass()) { 2308 ciInstanceKlass* ik = superelem->as_instance_klass(); 2309 if (!ik->has_subklass() && !ik->is_interface()) { 2310 if (!ik->is_final()) { 2311 // Add a dependency if there is a chance of a later subclass. 2312 C->dependencies()->assert_leaf_type(ik); 2313 } 2314 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2315 } 2316 } else { 2317 // A primitive array type has no subtypes. 2318 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2319 } 2320 2321 return SSC_full_test; 2322 } 2323 2324 // Profile-driven exact type check: 2325 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2326 float prob, 2327 Node* *casted_receiver) { 2328 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2329 Node* recv_klass = load_object_klass(receiver); 2330 Node* want_klass = makecon(tklass); 2331 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); 2332 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2333 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2334 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); 2335 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2336 2337 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2338 assert(recv_xtype->klass_is_exact(), ""); 2339 2340 // Subsume downstream occurrences of receiver with a cast to 2341 // recv_xtype, since now we know what the type will be. 2342 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); 2343 (*casted_receiver) = _gvn.transform(cast); 2344 // (User must make the replace_in_map call.) 2345 2346 return fail; 2347 } 2348 2349 2350 //-------------------------------gen_instanceof-------------------------------- 2351 // Generate an instance-of idiom. Used by both the instance-of bytecode 2352 // and the reflective instance-of call. 2353 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { 2354 C->set_has_split_ifs(true); // Has chance for split-if optimization 2355 assert( !stopped(), "dead parse path should be checked in callers" ); 2356 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2357 "must check for not-null not-dead klass in callers"); 2358 2359 // Make the merge point 2360 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2361 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2362 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2363 C->set_has_split_ifs(true); // Has chance for split-if optimization 2364 2365 // Null check; get casted pointer; set region slot 3 2366 Node* null_ctl = top(); 2367 Node* not_null_obj = null_check_oop(subobj, &null_ctl); 2368 2369 // If not_null_obj is dead, only null-path is taken 2370 if (stopped()) { // Doing instance-of on a NULL? 2371 set_control(null_ctl); 2372 return intcon(0); 2373 } 2374 region->init_req(_null_path, null_ctl); 2375 phi ->init_req(_null_path, intcon(0)); // Set null path value 2376 2377 // Load the object's klass 2378 Node* obj_klass = load_object_klass(not_null_obj); 2379 2380 // Generate the subtype check 2381 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2382 2383 // Plug in the success path to the general merge in slot 1. 2384 region->init_req(_obj_path, control()); 2385 phi ->init_req(_obj_path, intcon(1)); 2386 2387 // Plug in the failing path to the general merge in slot 2. 2388 region->init_req(_fail_path, not_subtype_ctrl); 2389 phi ->init_req(_fail_path, intcon(0)); 2390 2391 // Return final merged results 2392 set_control( _gvn.transform(region) ); 2393 record_for_igvn(region); 2394 return _gvn.transform(phi); 2395 } 2396 2397 //-------------------------------gen_checkcast--------------------------------- 2398 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2399 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2400 // uncommon-trap paths work. Adjust stack after this call. 2401 // If failure_control is supplied and not null, it is filled in with 2402 // the control edge for the cast failure. Otherwise, an appropriate 2403 // uncommon trap or exception is thrown. 2404 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2405 Node* *failure_control) { 2406 kill_dead_locals(); // Benefit all the uncommon traps 2407 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2408 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2409 2410 // Fast cutout: Check the case that the cast is vacuously true. 2411 // This detects the common cases where the test will short-circuit 2412 // away completely. We do this before we perform the null check, 2413 // because if the test is going to turn into zero code, we don't 2414 // want a residual null check left around. (Causes a slowdown, 2415 // for example, in some objArray manipulations, such as a[i]=a[j].) 2416 if (tk->singleton()) { 2417 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2418 if (objtp != NULL && objtp->klass() != NULL) { 2419 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2420 case SSC_always_true: 2421 return obj; 2422 case SSC_always_false: 2423 // It needs a null check because a null will *pass* the cast check. 2424 // A non-null value will always produce an exception. 2425 return do_null_assert(obj, T_OBJECT); 2426 } 2427 } 2428 } 2429 2430 ciProfileData* data = NULL; 2431 if (failure_control == NULL) { // use MDO in regular case only 2432 assert(java_bc() == Bytecodes::_aastore || 2433 java_bc() == Bytecodes::_checkcast, 2434 "interpreter profiles type checks only for these BCs"); 2435 data = method()->method_data()->bci_to_data(bci()); 2436 } 2437 2438 // Make the merge point 2439 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2440 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2441 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); 2442 C->set_has_split_ifs(true); // Has chance for split-if optimization 2443 2444 // Use null-cast information if it is available 2445 bool never_see_null = false; 2446 // If we see an unexpected null at a check-cast we record it and force a 2447 // recompile; the offending check-cast will be compiled to handle NULLs. 2448 // If we see several offending BCIs, then all checkcasts in the 2449 // method will be compiled to handle NULLs. 2450 if (UncommonNullCast // Cutout for this technique 2451 && failure_control == NULL // regular case 2452 && obj != null() // And not the -Xcomp stupid case? 2453 && !too_many_traps(Deoptimization::Reason_null_check)) { 2454 // Finally, check the "null_seen" bit from the interpreter. 2455 if (data == NULL || !data->as_BitData()->null_seen()) { 2456 never_see_null = true; 2457 } 2458 } 2459 2460 // Null check; get casted pointer; set region slot 3 2461 Node* null_ctl = top(); 2462 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2463 2464 // If not_null_obj is dead, only null-path is taken 2465 if (stopped()) { // Doing instance-of on a NULL? 2466 set_control(null_ctl); 2467 return null(); 2468 } 2469 region->init_req(_null_path, null_ctl); 2470 phi ->init_req(_null_path, null()); // Set null path value 2471 2472 Node* cast_obj = NULL; // the casted version of the object 2473 2474 // If the profile has seen exactly one type, narrow to that type. 2475 // (The subsequent subtype check will always fold up.) 2476 if (UseTypeProfile && TypeProfileCasts && data != NULL && 2477 // Counter has never been decremented (due to cast failure). 2478 // ...This is a reasonable thing to expect. It is true of 2479 // all casts inserted by javac to implement generic types. 2480 data->as_CounterData()->count() >= 0 && 2481 !too_many_traps(Deoptimization::Reason_class_check)) { 2482 // (No, this isn't a call, but it's enough like a virtual call 2483 // to use the same ciMethod accessor to get the profile info...) 2484 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2485 if (profile.count() >= 0 && // no cast failures here 2486 profile.has_receiver(0) && 2487 profile.morphism() == 1) { 2488 ciKlass* exact_kls = profile.receiver(0); 2489 int ssc = static_subtype_check(tk->klass(), exact_kls); 2490 if (ssc == SSC_always_true) { 2491 // If we narrow the type to match what the type profile sees, 2492 // we can then remove the rest of the cast. 2493 // This is a win, even if the exact_kls is very specific, 2494 // because downstream operations, such as method calls, 2495 // will often benefit from the sharper type. 2496 Node* exact_obj = not_null_obj; // will get updated in place... 2497 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2498 &exact_obj); 2499 { PreserveJVMState pjvms(this); 2500 set_control(slow_ctl); 2501 uncommon_trap(Deoptimization::Reason_class_check, 2502 Deoptimization::Action_maybe_recompile); 2503 } 2504 if (failure_control != NULL) // failure is now impossible 2505 (*failure_control) = top(); 2506 replace_in_map(not_null_obj, exact_obj); 2507 // adjust the type of the phi to the exact klass: 2508 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); 2509 cast_obj = exact_obj; 2510 } 2511 // assert(cast_obj != NULL)... except maybe the profile lied to us. 2512 } 2513 } 2514 2515 if (cast_obj == NULL) { 2516 // Load the object's klass 2517 Node* obj_klass = load_object_klass(not_null_obj); 2518 2519 // Generate the subtype check 2520 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2521 2522 // Plug in success path into the merge 2523 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), 2524 not_null_obj, toop)); 2525 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2526 if (failure_control == NULL) { 2527 if (not_subtype_ctrl != top()) { // If failure is possible 2528 PreserveJVMState pjvms(this); 2529 set_control(not_subtype_ctrl); 2530 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2531 } 2532 } else { 2533 (*failure_control) = not_subtype_ctrl; 2534 } 2535 } 2536 2537 region->init_req(_obj_path, control()); 2538 phi ->init_req(_obj_path, cast_obj); 2539 2540 // A merge of NULL or Casted-NotNull obj 2541 Node* res = _gvn.transform(phi); 2542 2543 // Note I do NOT always 'replace_in_map(obj,result)' here. 2544 // if( tk->klass()->can_be_primary_super() ) 2545 // This means that if I successfully store an Object into an array-of-String 2546 // I 'forget' that the Object is really now known to be a String. I have to 2547 // do this because we don't have true union types for interfaces - if I store 2548 // a Baz into an array-of-Interface and then tell the optimizer it's an 2549 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2550 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2551 // replace_in_map( obj, res ); 2552 2553 // Return final merged results 2554 set_control( _gvn.transform(region) ); 2555 record_for_igvn(region); 2556 return res; 2557 } 2558 2559 //------------------------------next_monitor----------------------------------- 2560 // What number should be given to the next monitor? 2561 int GraphKit::next_monitor() { 2562 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2563 int next = current + C->sync_stack_slots(); 2564 // Keep the toplevel high water mark current: 2565 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2566 return current; 2567 } 2568 2569 //------------------------------insert_mem_bar--------------------------------- 2570 // Memory barrier to avoid floating things around 2571 // The membar serves as a pinch point between both control and all memory slices. 2572 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2573 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2574 mb->init_req(TypeFunc::Control, control()); 2575 mb->init_req(TypeFunc::Memory, reset_memory()); 2576 Node* membar = _gvn.transform(mb); 2577 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); 2578 set_all_memory_call(membar); 2579 return membar; 2580 } 2581 2582 //-------------------------insert_mem_bar_volatile---------------------------- 2583 // Memory barrier to avoid floating things around 2584 // The membar serves as a pinch point between both control and memory(alias_idx). 2585 // If you want to make a pinch point on all memory slices, do not use this 2586 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2587 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2588 // When Parse::do_put_xxx updates a volatile field, it appends a series 2589 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2590 // The first membar is on the same memory slice as the field store opcode. 2591 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2592 // All the other membars (for other volatile slices, including AliasIdxBot, 2593 // which stands for all unknown volatile slices) are control-dependent 2594 // on the first membar. This prevents later volatile loads or stores 2595 // from sliding up past the just-emitted store. 2596 2597 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2598 mb->set_req(TypeFunc::Control,control()); 2599 if (alias_idx == Compile::AliasIdxBot) { 2600 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2601 } else { 2602 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2603 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2604 } 2605 Node* membar = _gvn.transform(mb); 2606 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); 2607 if (alias_idx == Compile::AliasIdxBot) { 2608 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); 2609 } else { 2610 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2611 } 2612 return membar; 2613 } 2614 2615 //------------------------------shared_lock------------------------------------ 2616 // Emit locking code. 2617 FastLockNode* GraphKit::shared_lock(Node* obj) { 2618 // bci is either a monitorenter bc or InvocationEntryBci 2619 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2620 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2621 2622 if( !GenerateSynchronizationCode ) 2623 return NULL; // Not locking things? 2624 if (stopped()) // Dead monitor? 2625 return NULL; 2626 2627 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2628 2629 // Box the stack location 2630 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); 2631 Node* mem = reset_memory(); 2632 2633 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); 2634 if (PrintPreciseBiasedLockingStatistics) { 2635 // Create the counters for this fast lock. 2636 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2637 } 2638 // Add monitor to debug info for the slow path. If we block inside the 2639 // slow path and de-opt, we need the monitor hanging around 2640 map()->push_monitor( flock ); 2641 2642 const TypeFunc *tf = LockNode::lock_type(); 2643 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); 2644 2645 lock->init_req( TypeFunc::Control, control() ); 2646 lock->init_req( TypeFunc::Memory , mem ); 2647 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2648 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2649 lock->init_req( TypeFunc::ReturnAdr, top() ); 2650 2651 lock->init_req(TypeFunc::Parms + 0, obj); 2652 lock->init_req(TypeFunc::Parms + 1, box); 2653 lock->init_req(TypeFunc::Parms + 2, flock); 2654 add_safepoint_edges(lock); 2655 2656 lock = _gvn.transform( lock )->as_Lock(); 2657 2658 // lock has no side-effects, sets few values 2659 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2660 2661 insert_mem_bar(Op_MemBarAcquire); 2662 2663 // Add this to the worklist so that the lock can be eliminated 2664 record_for_igvn(lock); 2665 2666 #ifndef PRODUCT 2667 if (PrintLockStatistics) { 2668 // Update the counter for this lock. Don't bother using an atomic 2669 // operation since we don't require absolute accuracy. 2670 lock->create_lock_counter(map()->jvms()); 2671 int adr_type = Compile::AliasIdxRaw; 2672 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); 2673 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 2674 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 2675 store_to_memory(control(), counter_addr, incr, T_INT, adr_type); 2676 } 2677 #endif 2678 2679 return flock; 2680 } 2681 2682 2683 //------------------------------shared_unlock---------------------------------- 2684 // Emit unlocking code. 2685 void GraphKit::shared_unlock(Node* box, Node* obj) { 2686 // bci is either a monitorenter bc or InvocationEntryBci 2687 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2688 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2689 2690 if( !GenerateSynchronizationCode ) 2691 return; 2692 if (stopped()) { // Dead monitor? 2693 map()->pop_monitor(); // Kill monitor from debug info 2694 return; 2695 } 2696 2697 // Memory barrier to avoid floating things down past the locked region 2698 insert_mem_bar(Op_MemBarRelease); 2699 2700 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2701 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); 2702 uint raw_idx = Compile::AliasIdxRaw; 2703 unlock->init_req( TypeFunc::Control, control() ); 2704 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2705 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2706 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2707 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2708 2709 unlock->init_req(TypeFunc::Parms + 0, obj); 2710 unlock->init_req(TypeFunc::Parms + 1, box); 2711 unlock = _gvn.transform(unlock)->as_Unlock(); 2712 2713 Node* mem = reset_memory(); 2714 2715 // unlock has no side-effects, sets few values 2716 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 2717 2718 // Kill monitor from debug info 2719 map()->pop_monitor( ); 2720 } 2721 2722 //-------------------------------get_layout_helper----------------------------- 2723 // If the given klass is a constant or known to be an array, 2724 // fetch the constant layout helper value into constant_value 2725 // and return (Node*)NULL. Otherwise, load the non-constant 2726 // layout helper value, and return the node which represents it. 2727 // This two-faced routine is useful because allocation sites 2728 // almost always feature constant types. 2729 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 2730 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 2731 if (!StressReflectiveCode && inst_klass != NULL) { 2732 ciKlass* klass = inst_klass->klass(); 2733 bool xklass = inst_klass->klass_is_exact(); 2734 if (xklass || klass->is_array_klass()) { 2735 jint lhelper = klass->layout_helper(); 2736 if (lhelper != Klass::_lh_neutral_value) { 2737 constant_value = lhelper; 2738 return (Node*) NULL; 2739 } 2740 } 2741 } 2742 constant_value = Klass::_lh_neutral_value; // put in a known value 2743 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); 2744 return make_load(NULL, lhp, TypeInt::INT, T_INT); 2745 } 2746 2747 // We just put in an allocate/initialize with a big raw-memory effect. 2748 // Hook selected additional alias categories on the initialization. 2749 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 2750 MergeMemNode* init_in_merge, 2751 Node* init_out_raw) { 2752 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 2753 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 2754 2755 Node* prevmem = kit.memory(alias_idx); 2756 init_in_merge->set_memory_at(alias_idx, prevmem); 2757 kit.set_memory(init_out_raw, alias_idx); 2758 } 2759 2760 //---------------------------set_output_for_allocation------------------------- 2761 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 2762 const TypeOopPtr* oop_type, 2763 bool raw_mem_only) { 2764 int rawidx = Compile::AliasIdxRaw; 2765 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 2766 add_safepoint_edges(alloc); 2767 Node* allocx = _gvn.transform(alloc); 2768 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); 2769 // create memory projection for i_o 2770 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 2771 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); 2772 2773 // create a memory projection as for the normal control path 2774 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); 2775 set_memory(malloc, rawidx); 2776 2777 // a normal slow-call doesn't change i_o, but an allocation does 2778 // we create a separate i_o projection for the normal control path 2779 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); 2780 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); 2781 2782 // put in an initialization barrier 2783 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 2784 rawoop)->as_Initialize(); 2785 assert(alloc->initialization() == init, "2-way macro link must work"); 2786 assert(init ->allocation() == alloc, "2-way macro link must work"); 2787 if (ReduceFieldZeroing && !raw_mem_only) { 2788 // Extract memory strands which may participate in the new object's 2789 // initialization, and source them from the new InitializeNode. 2790 // This will allow us to observe initializations when they occur, 2791 // and link them properly (as a group) to the InitializeNode. 2792 assert(init->in(InitializeNode::Memory) == malloc, ""); 2793 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 2794 init->set_req(InitializeNode::Memory, minit_in); 2795 record_for_igvn(minit_in); // fold it up later, if possible 2796 Node* minit_out = memory(rawidx); 2797 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 2798 if (oop_type->isa_aryptr()) { 2799 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 2800 int elemidx = C->get_alias_index(telemref); 2801 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 2802 } else if (oop_type->isa_instptr()) { 2803 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 2804 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 2805 ciField* field = ik->nonstatic_field_at(i); 2806 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 2807 continue; // do not bother to track really large numbers of fields 2808 // Find (or create) the alias category for this field: 2809 int fieldidx = C->alias_type(field)->index(); 2810 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 2811 } 2812 } 2813 } 2814 2815 // Cast raw oop to the real thing... 2816 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); 2817 javaoop = _gvn.transform(javaoop); 2818 C->set_recent_alloc(control(), javaoop); 2819 assert(just_allocated_object(control()) == javaoop, "just allocated"); 2820 2821 #ifdef ASSERT 2822 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 2823 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 2824 "Ideal_allocation works"); 2825 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 2826 "Ideal_allocation works"); 2827 if (alloc->is_AllocateArray()) { 2828 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 2829 "Ideal_allocation works"); 2830 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 2831 "Ideal_allocation works"); 2832 } else { 2833 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 2834 } 2835 } 2836 #endif //ASSERT 2837 2838 return javaoop; 2839 } 2840 2841 //---------------------------new_instance-------------------------------------- 2842 // This routine takes a klass_node which may be constant (for a static type) 2843 // or may be non-constant (for reflective code). It will work equally well 2844 // for either, and the graph will fold nicely if the optimizer later reduces 2845 // the type to a constant. 2846 // The optional arguments are for specialized use by intrinsics: 2847 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 2848 // - If 'raw_mem_only', do not cast the result to an oop. 2849 // - If 'return_size_val', report the the total object size to the caller. 2850 Node* GraphKit::new_instance(Node* klass_node, 2851 Node* extra_slow_test, 2852 bool raw_mem_only, // affect only raw memory 2853 Node* *return_size_val) { 2854 // Compute size in doublewords 2855 // The size is always an integral number of doublewords, represented 2856 // as a positive bytewise size stored in the klass's layout_helper. 2857 // The layout_helper also encodes (in a low bit) the need for a slow path. 2858 jint layout_con = Klass::_lh_neutral_value; 2859 Node* layout_val = get_layout_helper(klass_node, layout_con); 2860 int layout_is_con = (layout_val == NULL); 2861 2862 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 2863 // Generate the initial go-slow test. It's either ALWAYS (return a 2864 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 2865 // case) a computed value derived from the layout_helper. 2866 Node* initial_slow_test = NULL; 2867 if (layout_is_con) { 2868 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2869 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 2870 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 2871 2872 } else { // reflective case 2873 // This reflective path is used by Unsafe.allocateInstance. 2874 // (It may be stress-tested by specifying StressReflectiveCode.) 2875 // Basically, we want to get into the VM is there's an illegal argument. 2876 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 2877 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); 2878 if (extra_slow_test != intcon(0)) { 2879 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); 2880 } 2881 // (Macro-expander will further convert this to a Bool, if necessary.) 2882 } 2883 2884 // Find the size in bytes. This is easy; it's the layout_helper. 2885 // The size value must be valid even if the slow path is taken. 2886 Node* size = NULL; 2887 if (layout_is_con) { 2888 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 2889 } else { // reflective case 2890 // This reflective path is used by clone and Unsafe.allocateInstance. 2891 size = ConvI2X(layout_val); 2892 2893 // Clear the low bits to extract layout_helper_size_in_bytes: 2894 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 2895 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 2896 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); 2897 } 2898 if (return_size_val != NULL) { 2899 (*return_size_val) = size; 2900 } 2901 2902 // This is a precise notnull oop of the klass. 2903 // (Actually, it need not be precise if this is a reflective allocation.) 2904 // It's what we cast the result to. 2905 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 2906 if (!tklass) tklass = TypeKlassPtr::OBJECT; 2907 const TypeOopPtr* oop_type = tklass->as_instance_type(); 2908 2909 // Now generate allocation code 2910 2911 // The entire memory state is needed for slow path of the allocation 2912 // since GC and deoptimization can happened. 2913 Node *mem = reset_memory(); 2914 set_all_memory(mem); // Create new memory state 2915 2916 AllocateNode* alloc 2917 = new (C, AllocateNode::ParmLimit) 2918 AllocateNode(C, AllocateNode::alloc_type(), 2919 control(), mem, i_o(), 2920 size, klass_node, 2921 initial_slow_test); 2922 2923 return set_output_for_allocation(alloc, oop_type, raw_mem_only); 2924 } 2925 2926 //-------------------------------new_array------------------------------------- 2927 // helper for both newarray and anewarray 2928 // The 'length' parameter is (obviously) the length of the array. 2929 // See comments on new_instance for the meaning of the other arguments. 2930 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 2931 Node* length, // number of array elements 2932 int nargs, // number of arguments to push back for uncommon trap 2933 bool raw_mem_only, // affect only raw memory 2934 Node* *return_size_val) { 2935 jint layout_con = Klass::_lh_neutral_value; 2936 Node* layout_val = get_layout_helper(klass_node, layout_con); 2937 int layout_is_con = (layout_val == NULL); 2938 2939 if (!layout_is_con && !StressReflectiveCode && 2940 !too_many_traps(Deoptimization::Reason_class_check)) { 2941 // This is a reflective array creation site. 2942 // Optimistically assume that it is a subtype of Object[], 2943 // so that we can fold up all the address arithmetic. 2944 layout_con = Klass::array_layout_helper(T_OBJECT); 2945 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); 2946 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); 2947 { BuildCutout unless(this, bol_lh, PROB_MAX); 2948 _sp += nargs; 2949 uncommon_trap(Deoptimization::Reason_class_check, 2950 Deoptimization::Action_maybe_recompile); 2951 } 2952 layout_val = NULL; 2953 layout_is_con = true; 2954 } 2955 2956 // Generate the initial go-slow test. Make sure we do not overflow 2957 // if length is huge (near 2Gig) or negative! We do not need 2958 // exact double-words here, just a close approximation of needed 2959 // double-words. We can't add any offset or rounding bits, lest we 2960 // take a size -1 of bytes and make it positive. Use an unsigned 2961 // compare, so negative sizes look hugely positive. 2962 int fast_size_limit = FastAllocateSizeLimit; 2963 if (layout_is_con) { 2964 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2965 // Increase the size limit if we have exact knowledge of array type. 2966 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 2967 fast_size_limit <<= (LogBytesPerLong - log2_esize); 2968 } 2969 2970 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); 2971 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 2972 if (initial_slow_test->is_Bool()) { 2973 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 2974 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 2975 } 2976 2977 // --- Size Computation --- 2978 // array_size = round_to_heap(array_header + (length << elem_shift)); 2979 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 2980 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 2981 // The rounding mask is strength-reduced, if possible. 2982 int round_mask = MinObjAlignmentInBytes - 1; 2983 Node* header_size = NULL; 2984 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 2985 // (T_BYTE has the weakest alignment and size restrictions...) 2986 if (layout_is_con) { 2987 int hsize = Klass::layout_helper_header_size(layout_con); 2988 int eshift = Klass::layout_helper_log2_element_size(layout_con); 2989 BasicType etype = Klass::layout_helper_element_type(layout_con); 2990 if ((round_mask & ~right_n_bits(eshift)) == 0) 2991 round_mask = 0; // strength-reduce it if it goes away completely 2992 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 2993 assert(header_size_min <= hsize, "generic minimum is smallest"); 2994 header_size_min = hsize; 2995 header_size = intcon(hsize + round_mask); 2996 } else { 2997 Node* hss = intcon(Klass::_lh_header_size_shift); 2998 Node* hsm = intcon(Klass::_lh_header_size_mask); 2999 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); 3000 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); 3001 Node* mask = intcon(round_mask); 3002 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); 3003 } 3004 3005 Node* elem_shift = NULL; 3006 if (layout_is_con) { 3007 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3008 if (eshift != 0) 3009 elem_shift = intcon(eshift); 3010 } else { 3011 // There is no need to mask or shift this value. 3012 // The semantics of LShiftINode include an implicit mask to 0x1F. 3013 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3014 elem_shift = layout_val; 3015 } 3016 3017 // Transition to native address size for all offset calculations: 3018 Node* lengthx = ConvI2X(length); 3019 Node* headerx = ConvI2X(header_size); 3020 #ifdef _LP64 3021 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3022 if (tllen != NULL && tllen->_lo < 0) { 3023 // Add a manual constraint to a positive range. Cf. array_element_address. 3024 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3025 if (size_max > tllen->_hi) size_max = tllen->_hi; 3026 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3027 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); 3028 } 3029 } 3030 #endif 3031 3032 // Combine header size (plus rounding) and body size. Then round down. 3033 // This computation cannot overflow, because it is used only in two 3034 // places, one where the length is sharply limited, and the other 3035 // after a successful allocation. 3036 Node* abody = lengthx; 3037 if (elem_shift != NULL) 3038 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); 3039 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); 3040 if (round_mask != 0) { 3041 Node* mask = MakeConX(~round_mask); 3042 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); 3043 } 3044 // else if round_mask == 0, the size computation is self-rounding 3045 3046 if (return_size_val != NULL) { 3047 // This is the size 3048 (*return_size_val) = size; 3049 } 3050 3051 // Now generate allocation code 3052 3053 // The entire memory state is needed for slow path of the allocation 3054 // since GC and deoptimization can happened. 3055 Node *mem = reset_memory(); 3056 set_all_memory(mem); // Create new memory state 3057 3058 // Create the AllocateArrayNode and its result projections 3059 AllocateArrayNode* alloc 3060 = new (C, AllocateArrayNode::ParmLimit) 3061 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3062 control(), mem, i_o(), 3063 size, klass_node, 3064 initial_slow_test, 3065 length); 3066 3067 // Cast to correct type. Note that the klass_node may be constant or not, 3068 // and in the latter case the actual array type will be inexact also. 3069 // (This happens via a non-constant argument to inline_native_newArray.) 3070 // In any case, the value of klass_node provides the desired array type. 3071 const TypeInt* length_type = _gvn.find_int_type(length); 3072 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3073 if (ary_type->isa_aryptr() && length_type != NULL) { 3074 // Try to get a better type than POS for the size 3075 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3076 } 3077 3078 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); 3079 3080 // Cast length on remaining path to be as narrow as possible 3081 if (map()->find_edge(length) >= 0) { 3082 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3083 if (ccast != length) { 3084 _gvn.set_type_bottom(ccast); 3085 record_for_igvn(ccast); 3086 replace_in_map(length, ccast); 3087 } 3088 } 3089 3090 return javaoop; 3091 } 3092 3093 // The following "Ideal_foo" functions are placed here because they recognize 3094 // the graph shapes created by the functions immediately above. 3095 3096 //---------------------------Ideal_allocation---------------------------------- 3097 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3098 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3099 if (ptr == NULL) { // reduce dumb test in callers 3100 return NULL; 3101 } 3102 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3103 ptr = ptr->in(1); 3104 if (ptr == NULL) return NULL; 3105 } 3106 if (ptr->is_Proj()) { 3107 Node* allo = ptr->in(0); 3108 if (allo != NULL && allo->is_Allocate()) { 3109 return allo->as_Allocate(); 3110 } 3111 } 3112 // Report failure to match. 3113 return NULL; 3114 } 3115 3116 // Fancy version which also strips off an offset (and reports it to caller). 3117 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3118 intptr_t& offset) { 3119 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3120 if (base == NULL) return NULL; 3121 return Ideal_allocation(base, phase); 3122 } 3123 3124 // Trace Initialize <- Proj[Parm] <- Allocate 3125 AllocateNode* InitializeNode::allocation() { 3126 Node* rawoop = in(InitializeNode::RawAddress); 3127 if (rawoop->is_Proj()) { 3128 Node* alloc = rawoop->in(0); 3129 if (alloc->is_Allocate()) { 3130 return alloc->as_Allocate(); 3131 } 3132 } 3133 return NULL; 3134 } 3135 3136 // Trace Allocate -> Proj[Parm] -> Initialize 3137 InitializeNode* AllocateNode::initialization() { 3138 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3139 if (rawoop == NULL) return NULL; 3140 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3141 Node* init = rawoop->fast_out(i); 3142 if (init->is_Initialize()) { 3143 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3144 return init->as_Initialize(); 3145 } 3146 } 3147 return NULL; 3148 } 3149 3150 //----------------------------- store barriers ---------------------------- 3151 #define __ ideal. 3152 3153 void GraphKit::sync_kit(IdealKit& ideal) { 3154 // Final sync IdealKit and graphKit. 3155 __ drain_delay_transform(); 3156 set_all_memory(__ merged_memory()); 3157 set_control(__ ctrl()); 3158 } 3159 3160 // vanilla/CMS post barrier 3161 // Insert a write-barrier store. This is to let generational GC work; we have 3162 // to flag all oop-stores before the next GC point. 3163 void GraphKit::write_barrier_post(Node* oop_store, 3164 Node* obj, 3165 Node* adr, 3166 Node* val, 3167 bool use_precise) { 3168 // No store check needed if we're storing a NULL or an old object 3169 // (latter case is probably a string constant). The concurrent 3170 // mark sweep garbage collector, however, needs to have all nonNull 3171 // oop updates flagged via card-marks. 3172 if (val != NULL && val->is_Con()) { 3173 // must be either an oop or NULL 3174 const Type* t = val->bottom_type(); 3175 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3176 // stores of null never (?) need barriers 3177 return; 3178 ciObject* con = t->is_oopptr()->const_oop(); 3179 if (con != NULL 3180 && con->is_perm() 3181 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 3182 // no store barrier needed, because no old-to-new ref created 3183 return; 3184 } 3185 3186 if (!use_precise) { 3187 // All card marks for a (non-array) instance are in one place: 3188 adr = obj; 3189 } 3190 // (Else it's an array (or unknown), and we want more precise card marks.) 3191 assert(adr != NULL, ""); 3192 3193 IdealKit ideal(gvn(), control(), merged_memory(), true); 3194 3195 // Convert the pointer to an int prior to doing math on it 3196 Node* cast = __ CastPX(__ ctrl(), adr); 3197 3198 // Divide by card size 3199 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3200 "Only one we handle so far."); 3201 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3202 3203 // Combine card table base and card offset 3204 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3205 3206 // Get the alias_index for raw card-mark memory 3207 int adr_type = Compile::AliasIdxRaw; 3208 // Smash zero into card 3209 Node* zero = __ ConI(0); 3210 BasicType bt = T_BYTE; 3211 if( !UseConcMarkSweepGC ) { 3212 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3213 } else { 3214 // Specialized path for CM store barrier 3215 __ storeCM(__ ctrl(), card_adr, zero, oop_store, bt, adr_type); 3216 } 3217 3218 // Final sync IdealKit and GraphKit. 3219 sync_kit(ideal); 3220 } 3221 3222 // G1 pre/post barriers 3223 void GraphKit::g1_write_barrier_pre(Node* obj, 3224 Node* adr, 3225 uint alias_idx, 3226 Node* val, 3227 const TypeOopPtr* val_type, 3228 BasicType bt) { 3229 IdealKit ideal(gvn(), control(), merged_memory(), true); 3230 3231 Node* tls = __ thread(); // ThreadLocalStorage 3232 3233 Node* no_ctrl = NULL; 3234 Node* no_base = __ top(); 3235 Node* zero = __ ConI(0); 3236 3237 float likely = PROB_LIKELY(0.999); 3238 float unlikely = PROB_UNLIKELY(0.999); 3239 3240 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3241 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3242 3243 // Offsets into the thread 3244 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3245 PtrQueue::byte_offset_of_active()); 3246 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3247 PtrQueue::byte_offset_of_index()); 3248 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3249 PtrQueue::byte_offset_of_buf()); 3250 // Now the actual pointers into the thread 3251 3252 // set_control( ctl); 3253 3254 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3255 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3256 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3257 3258 // Now some of the values 3259 3260 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3261 3262 // if (!marking) 3263 __ if_then(marking, BoolTest::ne, zero); { 3264 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3265 3266 const Type* t1 = adr->bottom_type(); 3267 const Type* t2 = val->bottom_type(); 3268 3269 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx); 3270 // if (orig != NULL) 3271 __ if_then(orig, BoolTest::ne, null()); { 3272 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3273 3274 // load original value 3275 // alias_idx correct?? 3276 3277 // is the queue for this thread full? 3278 __ if_then(index, BoolTest::ne, zero, likely); { 3279 3280 // decrement the index 3281 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3282 Node* next_indexX = next_index; 3283 #ifdef _LP64 3284 // We could refine the type for what it's worth 3285 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3286 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3287 #endif 3288 3289 // Now get the buffer location we will log the original value into and store it 3290 Node *log_addr = __ AddP(no_base, buffer, next_indexX); 3291 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw); 3292 3293 // update the index 3294 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3295 3296 } __ else_(); { 3297 3298 // logging buffer is full, call the runtime 3299 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3300 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls); 3301 } __ end_if(); // (!index) 3302 } __ end_if(); // (orig != NULL) 3303 } __ end_if(); // (!marking) 3304 3305 // Final sync IdealKit and GraphKit. 3306 sync_kit(ideal); 3307 } 3308 3309 // 3310 // Update the card table and add card address to the queue 3311 // 3312 void GraphKit::g1_mark_card(IdealKit& ideal, 3313 Node* card_adr, 3314 Node* oop_store, 3315 Node* index, 3316 Node* index_adr, 3317 Node* buffer, 3318 const TypeFunc* tf) { 3319 3320 Node* zero = __ ConI(0); 3321 Node* no_base = __ top(); 3322 BasicType card_bt = T_BYTE; 3323 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3324 __ storeCM(__ ctrl(), card_adr, zero, oop_store, card_bt, Compile::AliasIdxRaw); 3325 3326 // Now do the queue work 3327 __ if_then(index, BoolTest::ne, zero); { 3328 3329 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3330 Node* next_indexX = next_index; 3331 #ifdef _LP64 3332 // We could refine the type for what it's worth 3333 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3334 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3335 #endif // _LP64 3336 Node* log_addr = __ AddP(no_base, buffer, next_indexX); 3337 3338 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3339 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3340 3341 } __ else_(); { 3342 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3343 } __ end_if(); 3344 3345 } 3346 3347 void GraphKit::g1_write_barrier_post(Node* oop_store, 3348 Node* obj, 3349 Node* adr, 3350 uint alias_idx, 3351 Node* val, 3352 BasicType bt, 3353 bool use_precise) { 3354 // If we are writing a NULL then we need no post barrier 3355 3356 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3357 // Must be NULL 3358 const Type* t = val->bottom_type(); 3359 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3360 // No post barrier if writing NULLx 3361 return; 3362 } 3363 3364 if (!use_precise) { 3365 // All card marks for a (non-array) instance are in one place: 3366 adr = obj; 3367 } 3368 // (Else it's an array (or unknown), and we want more precise card marks.) 3369 assert(adr != NULL, ""); 3370 3371 IdealKit ideal(gvn(), control(), merged_memory(), true); 3372 3373 Node* tls = __ thread(); // ThreadLocalStorage 3374 3375 Node* no_ctrl = NULL; 3376 Node* no_base = __ top(); 3377 float likely = PROB_LIKELY(0.999); 3378 float unlikely = PROB_UNLIKELY(0.999); 3379 Node* zero = __ ConI(0); 3380 Node* zeroX = __ ConX(0); 3381 3382 // Get the alias_index for raw card-mark memory 3383 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3384 3385 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3386 3387 // Offsets into the thread 3388 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3389 PtrQueue::byte_offset_of_index()); 3390 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3391 PtrQueue::byte_offset_of_buf()); 3392 3393 // Pointers into the thread 3394 3395 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3396 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3397 3398 // Now some values 3399 3400 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3401 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3402 3403 3404 // Convert the store obj pointer to an int prior to doing math on it 3405 // Must use ctrl to prevent "integerized oop" existing across safepoint 3406 Node* cast = __ CastPX(__ ctrl(), adr); 3407 3408 // Divide pointer by card size 3409 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3410 3411 // Combine card table base and card offset 3412 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3413 3414 // If we know the value being stored does it cross regions? 3415 3416 if (val != NULL) { 3417 // Does the store cause us to cross regions? 3418 3419 // Should be able to do an unsigned compare of region_size instead of 3420 // and extra shift. Do we have an unsigned compare?? 3421 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3422 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3423 3424 // if (xor_res == 0) same region so skip 3425 __ if_then(xor_res, BoolTest::ne, zeroX); { 3426 3427 // No barrier if we are storing a NULL 3428 __ if_then(val, BoolTest::ne, null(), unlikely); { 3429 3430 // Ok must mark the card if not already dirty 3431 3432 // load the original value of the card 3433 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3434 3435 __ if_then(card_val, BoolTest::ne, zero); { 3436 g1_mark_card(ideal, card_adr, oop_store, index, index_adr, buffer, tf); 3437 } __ end_if(); 3438 } __ end_if(); 3439 } __ end_if(); 3440 } else { 3441 // Object.clone() instrinsic uses this path. 3442 g1_mark_card(ideal, card_adr, oop_store, index, index_adr, buffer, tf); 3443 } 3444 3445 // Final sync IdealKit and GraphKit. 3446 sync_kit(ideal); 3447 } 3448 #undef __ 3449