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 assert(!kit->stopped(), "must call stopped() before"); 626 _kit = kit; 627 _sp = kit->sp(); 628 _reexecute = kit->jvms()->_reexecute; 629 } 630 PreserveReexecuteState::~PreserveReexecuteState() { 631 if (_kit->stopped()) return; 632 _kit->jvms()->_reexecute = _reexecute; 633 _kit->set_sp(_sp); 634 } 635 636 //------------------------------clone_map-------------------------------------- 637 // Implementation of PreserveJVMState 638 // 639 // Only clone_map(...) here. If this function is only used in the 640 // PreserveJVMState class we may want to get rid of this extra 641 // function eventually and do it all there. 642 643 SafePointNode* GraphKit::clone_map() { 644 if (map() == NULL) return NULL; 645 646 // Clone the memory edge first 647 Node* mem = MergeMemNode::make(C, map()->memory()); 648 gvn().set_type_bottom(mem); 649 650 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 651 JVMState* jvms = this->jvms(); 652 JVMState* clonejvms = jvms->clone_shallow(C); 653 clonemap->set_memory(mem); 654 clonemap->set_jvms(clonejvms); 655 clonejvms->set_map(clonemap); 656 record_for_igvn(clonemap); 657 gvn().set_type_bottom(clonemap); 658 return clonemap; 659 } 660 661 662 //-----------------------------set_map_clone----------------------------------- 663 void GraphKit::set_map_clone(SafePointNode* m) { 664 _map = m; 665 _map = clone_map(); 666 _map->set_next_exception(NULL); 667 debug_only(verify_map()); 668 } 669 670 671 //----------------------------kill_dead_locals--------------------------------- 672 // Detect any locals which are known to be dead, and force them to top. 673 void GraphKit::kill_dead_locals() { 674 // Consult the liveness information for the locals. If any 675 // of them are unused, then they can be replaced by top(). This 676 // should help register allocation time and cut down on the size 677 // of the deoptimization information. 678 679 // This call is made from many of the bytecode handling 680 // subroutines called from the Big Switch in do_one_bytecode. 681 // Every bytecode which might include a slow path is responsible 682 // for killing its dead locals. The more consistent we 683 // are about killing deads, the fewer useless phis will be 684 // constructed for them at various merge points. 685 686 // bci can be -1 (InvocationEntryBci). We return the entry 687 // liveness for the method. 688 689 if (method() == NULL || method()->code_size() == 0) { 690 // We are building a graph for a call to a native method. 691 // All locals are live. 692 return; 693 } 694 695 ResourceMark rm; 696 697 // Consult the liveness information for the locals. If any 698 // of them are unused, then they can be replaced by top(). This 699 // should help register allocation time and cut down on the size 700 // of the deoptimization information. 701 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 702 703 int len = (int)live_locals.size(); 704 assert(len <= jvms()->loc_size(), "too many live locals"); 705 for (int local = 0; local < len; local++) { 706 if (!live_locals.at(local)) { 707 set_local(local, top()); 708 } 709 } 710 } 711 712 #ifdef ASSERT 713 //-------------------------dead_locals_are_killed------------------------------ 714 // Return true if all dead locals are set to top in the map. 715 // Used to assert "clean" debug info at various points. 716 bool GraphKit::dead_locals_are_killed() { 717 if (method() == NULL || method()->code_size() == 0) { 718 // No locals need to be dead, so all is as it should be. 719 return true; 720 } 721 722 // Make sure somebody called kill_dead_locals upstream. 723 ResourceMark rm; 724 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 725 if (jvms->loc_size() == 0) continue; // no locals to consult 726 SafePointNode* map = jvms->map(); 727 ciMethod* method = jvms->method(); 728 int bci = jvms->bci(); 729 if (jvms == this->jvms()) { 730 bci = this->bci(); // it might not yet be synched 731 } 732 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 733 int len = (int)live_locals.size(); 734 if (!live_locals.is_valid() || len == 0) 735 // This method is trivial, or is poisoned by a breakpoint. 736 return true; 737 assert(len == jvms->loc_size(), "live map consistent with locals map"); 738 for (int local = 0; local < len; local++) { 739 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 740 if (PrintMiscellaneous && (Verbose || WizardMode)) { 741 tty->print_cr("Zombie local %d: ", local); 742 jvms->dump(); 743 } 744 return false; 745 } 746 } 747 } 748 return true; 749 } 750 751 #endif //ASSERT 752 753 // Helper function for enforcing certain bytecodes to reexecute if 754 // deoptimization happens 755 static bool should_reexecute_implied_by_bytecode(JVMState *jvms) { 756 ciMethod* cur_method = jvms->method(); 757 int cur_bci = jvms->bci(); 758 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 759 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 760 return Interpreter::bytecode_should_reexecute(code); 761 } else 762 return false; 763 } 764 765 // Helper function for adding JVMState and debug information to node 766 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 767 // Add the safepoint edges to the call (or other safepoint). 768 769 // Make sure dead locals are set to top. This 770 // should help register allocation time and cut down on the size 771 // of the deoptimization information. 772 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 773 774 // Walk the inline list to fill in the correct set of JVMState's 775 // Also fill in the associated edges for each JVMState. 776 777 JVMState* youngest_jvms = sync_jvms(); 778 779 // Do we need debug info here? If it is a SafePoint and this method 780 // cannot de-opt, then we do NOT need any debug info. 781 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint); 782 783 // If we are guaranteed to throw, we can prune everything but the 784 // input to the current bytecode. 785 bool can_prune_locals = false; 786 uint stack_slots_not_pruned = 0; 787 int inputs = 0, depth = 0; 788 if (must_throw) { 789 assert(method() == youngest_jvms->method(), "sanity"); 790 if (compute_stack_effects(inputs, depth)) { 791 can_prune_locals = true; 792 stack_slots_not_pruned = inputs; 793 } 794 } 795 796 if (env()->jvmti_can_examine_or_deopt_anywhere()) { 797 // At any safepoint, this method can get breakpointed, which would 798 // then require an immediate deoptimization. 799 full_info = true; 800 can_prune_locals = false; // do not prune locals 801 stack_slots_not_pruned = 0; 802 } 803 804 // do not scribble on the input jvms 805 JVMState* out_jvms = youngest_jvms->clone_deep(C); 806 call->set_jvms(out_jvms); // Start jvms list for call node 807 808 // For a known set of bytecodes, the interpreter should reexecute them if 809 // deoptimization happens. We set the reexecute state for them here 810 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 811 should_reexecute_implied_by_bytecode(out_jvms)) { 812 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 813 } 814 815 // Presize the call: 816 debug_only(uint non_debug_edges = call->req()); 817 call->add_req_batch(top(), youngest_jvms->debug_depth()); 818 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 819 820 // Set up edges so that the call looks like this: 821 // Call [state:] ctl io mem fptr retadr 822 // [parms:] parm0 ... parmN 823 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 824 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 825 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 826 // Note that caller debug info precedes callee debug info. 827 828 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 829 uint debug_ptr = call->req(); 830 831 // Loop over the map input edges associated with jvms, add them 832 // to the call node, & reset all offsets to match call node array. 833 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 834 uint debug_end = debug_ptr; 835 uint debug_start = debug_ptr - in_jvms->debug_size(); 836 debug_ptr = debug_start; // back up the ptr 837 838 uint p = debug_start; // walks forward in [debug_start, debug_end) 839 uint j, k, l; 840 SafePointNode* in_map = in_jvms->map(); 841 out_jvms->set_map(call); 842 843 if (can_prune_locals) { 844 assert(in_jvms->method() == out_jvms->method(), "sanity"); 845 // If the current throw can reach an exception handler in this JVMS, 846 // then we must keep everything live that can reach that handler. 847 // As a quick and dirty approximation, we look for any handlers at all. 848 if (in_jvms->method()->has_exception_handlers()) { 849 can_prune_locals = false; 850 } 851 } 852 853 // Add the Locals 854 k = in_jvms->locoff(); 855 l = in_jvms->loc_size(); 856 out_jvms->set_locoff(p); 857 if (full_info && !can_prune_locals) { 858 for (j = 0; j < l; j++) 859 call->set_req(p++, in_map->in(k+j)); 860 } else { 861 p += l; // already set to top above by add_req_batch 862 } 863 864 // Add the Expression Stack 865 k = in_jvms->stkoff(); 866 l = in_jvms->sp(); 867 out_jvms->set_stkoff(p); 868 if (full_info && !can_prune_locals) { 869 for (j = 0; j < l; j++) 870 call->set_req(p++, in_map->in(k+j)); 871 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 872 // Divide stack into {S0,...,S1}, where S0 is set to top. 873 uint s1 = stack_slots_not_pruned; 874 stack_slots_not_pruned = 0; // for next iteration 875 if (s1 > l) s1 = l; 876 uint s0 = l - s1; 877 p += s0; // skip the tops preinstalled by add_req_batch 878 for (j = s0; j < l; j++) 879 call->set_req(p++, in_map->in(k+j)); 880 } else { 881 p += l; // already set to top above by add_req_batch 882 } 883 884 // Add the Monitors 885 k = in_jvms->monoff(); 886 l = in_jvms->mon_size(); 887 out_jvms->set_monoff(p); 888 for (j = 0; j < l; j++) 889 call->set_req(p++, in_map->in(k+j)); 890 891 // Copy any scalar object fields. 892 k = in_jvms->scloff(); 893 l = in_jvms->scl_size(); 894 out_jvms->set_scloff(p); 895 for (j = 0; j < l; j++) 896 call->set_req(p++, in_map->in(k+j)); 897 898 // Finish the new jvms. 899 out_jvms->set_endoff(p); 900 901 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 902 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 903 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 904 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 905 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 906 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 907 908 // Update the two tail pointers in parallel. 909 out_jvms = out_jvms->caller(); 910 in_jvms = in_jvms->caller(); 911 } 912 913 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 914 915 // Test the correctness of JVMState::debug_xxx accessors: 916 assert(call->jvms()->debug_start() == non_debug_edges, ""); 917 assert(call->jvms()->debug_end() == call->req(), ""); 918 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 919 } 920 921 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 922 Bytecodes::Code code = java_bc(); 923 if (code == Bytecodes::_wide) { 924 code = method()->java_code_at_bci(bci() + 1); 925 } 926 927 BasicType rtype = T_ILLEGAL; 928 int rsize = 0; 929 930 if (code != Bytecodes::_illegal) { 931 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 932 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 933 if (rtype < T_CONFLICT) 934 rsize = type2size[rtype]; 935 } 936 937 switch (code) { 938 case Bytecodes::_illegal: 939 return false; 940 941 case Bytecodes::_ldc: 942 case Bytecodes::_ldc_w: 943 case Bytecodes::_ldc2_w: 944 inputs = 0; 945 break; 946 947 case Bytecodes::_dup: inputs = 1; break; 948 case Bytecodes::_dup_x1: inputs = 2; break; 949 case Bytecodes::_dup_x2: inputs = 3; break; 950 case Bytecodes::_dup2: inputs = 2; break; 951 case Bytecodes::_dup2_x1: inputs = 3; break; 952 case Bytecodes::_dup2_x2: inputs = 4; break; 953 case Bytecodes::_swap: inputs = 2; break; 954 case Bytecodes::_arraylength: inputs = 1; break; 955 956 case Bytecodes::_getstatic: 957 case Bytecodes::_putstatic: 958 case Bytecodes::_getfield: 959 case Bytecodes::_putfield: 960 { 961 bool is_get = (depth >= 0), is_static = (depth & 1); 962 bool ignore; 963 ciBytecodeStream iter(method()); 964 iter.reset_to_bci(bci()); 965 iter.next(); 966 ciField* field = iter.get_field(ignore); 967 int size = field->type()->size(); 968 inputs = (is_static ? 0 : 1); 969 if (is_get) { 970 depth = size - inputs; 971 } else { 972 inputs += size; // putxxx pops the value from the stack 973 depth = - inputs; 974 } 975 } 976 break; 977 978 case Bytecodes::_invokevirtual: 979 case Bytecodes::_invokespecial: 980 case Bytecodes::_invokestatic: 981 case Bytecodes::_invokedynamic: 982 case Bytecodes::_invokeinterface: 983 { 984 bool is_static = (depth == 0); 985 bool ignore; 986 ciBytecodeStream iter(method()); 987 iter.reset_to_bci(bci()); 988 iter.next(); 989 ciMethod* method = iter.get_method(ignore); 990 inputs = method->arg_size_no_receiver(); 991 if (!is_static) inputs += 1; 992 int size = method->return_type()->size(); 993 depth = size - inputs; 994 } 995 break; 996 997 case Bytecodes::_multianewarray: 998 { 999 ciBytecodeStream iter(method()); 1000 iter.reset_to_bci(bci()); 1001 iter.next(); 1002 inputs = iter.get_dimensions(); 1003 assert(rsize == 1, ""); 1004 depth = rsize - inputs; 1005 } 1006 break; 1007 1008 case Bytecodes::_ireturn: 1009 case Bytecodes::_lreturn: 1010 case Bytecodes::_freturn: 1011 case Bytecodes::_dreturn: 1012 case Bytecodes::_areturn: 1013 assert(rsize = -depth, ""); 1014 inputs = rsize; 1015 break; 1016 1017 case Bytecodes::_jsr: 1018 case Bytecodes::_jsr_w: 1019 inputs = 0; 1020 depth = 1; // S.B. depth=1, not zero 1021 break; 1022 1023 default: 1024 // bytecode produces a typed result 1025 inputs = rsize - depth; 1026 assert(inputs >= 0, ""); 1027 break; 1028 } 1029 1030 #ifdef ASSERT 1031 // spot check 1032 int outputs = depth + inputs; 1033 assert(outputs >= 0, "sanity"); 1034 switch (code) { 1035 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1036 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1037 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1038 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1039 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1040 } 1041 #endif //ASSERT 1042 1043 return true; 1044 } 1045 1046 1047 1048 //------------------------------basic_plus_adr--------------------------------- 1049 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1050 // short-circuit a common case 1051 if (offset == intcon(0)) return ptr; 1052 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); 1053 } 1054 1055 Node* GraphKit::ConvI2L(Node* offset) { 1056 // short-circuit a common case 1057 jint offset_con = find_int_con(offset, Type::OffsetBot); 1058 if (offset_con != Type::OffsetBot) { 1059 return longcon((long) offset_con); 1060 } 1061 return _gvn.transform( new (C, 2) ConvI2LNode(offset)); 1062 } 1063 Node* GraphKit::ConvL2I(Node* offset) { 1064 // short-circuit a common case 1065 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1066 if (offset_con != (jlong)Type::OffsetBot) { 1067 return intcon((int) offset_con); 1068 } 1069 return _gvn.transform( new (C, 2) ConvL2INode(offset)); 1070 } 1071 1072 //-------------------------load_object_klass----------------------------------- 1073 Node* GraphKit::load_object_klass(Node* obj) { 1074 // Special-case a fresh allocation to avoid building nodes: 1075 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1076 if (akls != NULL) return akls; 1077 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1078 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1079 } 1080 1081 //-------------------------load_array_length----------------------------------- 1082 Node* GraphKit::load_array_length(Node* array) { 1083 // Special-case a fresh allocation to avoid building nodes: 1084 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1085 Node *alen; 1086 if (alloc == NULL) { 1087 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1088 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1089 } else { 1090 alen = alloc->Ideal_length(); 1091 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1092 if (ccast != alen) { 1093 alen = _gvn.transform(ccast); 1094 } 1095 } 1096 return alen; 1097 } 1098 1099 //------------------------------do_null_check---------------------------------- 1100 // Helper function to do a NULL pointer check. Returned value is 1101 // the incoming address with NULL casted away. You are allowed to use the 1102 // not-null value only if you are control dependent on the test. 1103 extern int explicit_null_checks_inserted, 1104 explicit_null_checks_elided; 1105 Node* GraphKit::null_check_common(Node* value, BasicType type, 1106 // optional arguments for variations: 1107 bool assert_null, 1108 Node* *null_control) { 1109 assert(!assert_null || null_control == NULL, "not both at once"); 1110 if (stopped()) return top(); 1111 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1112 // For some performance testing, we may wish to suppress null checking. 1113 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1114 return value; 1115 } 1116 explicit_null_checks_inserted++; 1117 1118 // Construct NULL check 1119 Node *chk = NULL; 1120 switch(type) { 1121 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1122 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; 1123 case T_ARRAY : // fall through 1124 type = T_OBJECT; // simplify further tests 1125 case T_OBJECT : { 1126 const Type *t = _gvn.type( value ); 1127 1128 const TypeOopPtr* tp = t->isa_oopptr(); 1129 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1130 // Only for do_null_check, not any of its siblings: 1131 && !assert_null && null_control == NULL) { 1132 // Usually, any field access or invocation on an unloaded oop type 1133 // will simply fail to link, since the statically linked class is 1134 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1135 // the static class is loaded but the sharper oop type is not. 1136 // Rather than checking for this obscure case in lots of places, 1137 // we simply observe that a null check on an unloaded class 1138 // will always be followed by a nonsense operation, so we 1139 // can just issue the uncommon trap here. 1140 // Our access to the unloaded class will only be correct 1141 // after it has been loaded and initialized, which requires 1142 // a trip through the interpreter. 1143 #ifndef PRODUCT 1144 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1145 #endif 1146 uncommon_trap(Deoptimization::Reason_unloaded, 1147 Deoptimization::Action_reinterpret, 1148 tp->klass(), "!loaded"); 1149 return top(); 1150 } 1151 1152 if (assert_null) { 1153 // See if the type is contained in NULL_PTR. 1154 // If so, then the value is already null. 1155 if (t->higher_equal(TypePtr::NULL_PTR)) { 1156 explicit_null_checks_elided++; 1157 return value; // Elided null assert quickly! 1158 } 1159 } else { 1160 // See if mixing in the NULL pointer changes type. 1161 // If so, then the NULL pointer was not allowed in the original 1162 // type. In other words, "value" was not-null. 1163 if (t->meet(TypePtr::NULL_PTR) != t) { 1164 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1165 explicit_null_checks_elided++; 1166 return value; // Elided null check quickly! 1167 } 1168 } 1169 chk = new (C, 3) CmpPNode( value, null() ); 1170 break; 1171 } 1172 1173 default : ShouldNotReachHere(); 1174 } 1175 assert(chk != NULL, "sanity check"); 1176 chk = _gvn.transform(chk); 1177 1178 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1179 BoolNode *btst = new (C, 2) BoolNode( chk, btest); 1180 Node *tst = _gvn.transform( btst ); 1181 1182 //----------- 1183 // if peephole optimizations occurred, a prior test existed. 1184 // If a prior test existed, maybe it dominates as we can avoid this test. 1185 if (tst != btst && type == T_OBJECT) { 1186 // At this point we want to scan up the CFG to see if we can 1187 // find an identical test (and so avoid this test altogether). 1188 Node *cfg = control(); 1189 int depth = 0; 1190 while( depth < 16 ) { // Limit search depth for speed 1191 if( cfg->Opcode() == Op_IfTrue && 1192 cfg->in(0)->in(1) == tst ) { 1193 // Found prior test. Use "cast_not_null" to construct an identical 1194 // CastPP (and hence hash to) as already exists for the prior test. 1195 // Return that casted value. 1196 if (assert_null) { 1197 replace_in_map(value, null()); 1198 return null(); // do not issue the redundant test 1199 } 1200 Node *oldcontrol = control(); 1201 set_control(cfg); 1202 Node *res = cast_not_null(value); 1203 set_control(oldcontrol); 1204 explicit_null_checks_elided++; 1205 return res; 1206 } 1207 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1208 if (cfg == NULL) break; // Quit at region nodes 1209 depth++; 1210 } 1211 } 1212 1213 //----------- 1214 // Branch to failure if null 1215 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1216 Deoptimization::DeoptReason reason; 1217 if (assert_null) 1218 reason = Deoptimization::Reason_null_assert; 1219 else if (type == T_OBJECT) 1220 reason = Deoptimization::Reason_null_check; 1221 else 1222 reason = Deoptimization::Reason_div0_check; 1223 1224 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1225 // ciMethodData::has_trap_at will return a conservative -1 if any 1226 // must-be-null assertion has failed. This could cause performance 1227 // problems for a method after its first do_null_assert failure. 1228 // Consider using 'Reason_class_check' instead? 1229 1230 // To cause an implicit null check, we set the not-null probability 1231 // to the maximum (PROB_MAX). For an explicit check the probability 1232 // is set to a smaller value. 1233 if (null_control != NULL || too_many_traps(reason)) { 1234 // probability is less likely 1235 ok_prob = PROB_LIKELY_MAG(3); 1236 } else if (!assert_null && 1237 (ImplicitNullCheckThreshold > 0) && 1238 method() != NULL && 1239 (method()->method_data()->trap_count(reason) 1240 >= (uint)ImplicitNullCheckThreshold)) { 1241 ok_prob = PROB_LIKELY_MAG(3); 1242 } 1243 1244 if (null_control != NULL) { 1245 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1246 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); 1247 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); 1248 if (null_true == top()) 1249 explicit_null_checks_elided++; 1250 (*null_control) = null_true; 1251 } else { 1252 BuildCutout unless(this, tst, ok_prob); 1253 // Check for optimizer eliding test at parse time 1254 if (stopped()) { 1255 // Failure not possible; do not bother making uncommon trap. 1256 explicit_null_checks_elided++; 1257 } else if (assert_null) { 1258 uncommon_trap(reason, 1259 Deoptimization::Action_make_not_entrant, 1260 NULL, "assert_null"); 1261 } else { 1262 replace_in_map(value, zerocon(type)); 1263 builtin_throw(reason); 1264 } 1265 } 1266 1267 // Must throw exception, fall-thru not possible? 1268 if (stopped()) { 1269 return top(); // No result 1270 } 1271 1272 if (assert_null) { 1273 // Cast obj to null on this path. 1274 replace_in_map(value, zerocon(type)); 1275 return zerocon(type); 1276 } 1277 1278 // Cast obj to not-null on this path, if there is no null_control. 1279 // (If there is a null_control, a non-null value may come back to haunt us.) 1280 if (type == T_OBJECT) { 1281 Node* cast = cast_not_null(value, false); 1282 if (null_control == NULL || (*null_control) == top()) 1283 replace_in_map(value, cast); 1284 value = cast; 1285 } 1286 1287 return value; 1288 } 1289 1290 1291 //------------------------------cast_not_null---------------------------------- 1292 // Cast obj to not-null on this path 1293 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1294 const Type *t = _gvn.type(obj); 1295 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1296 // Object is already not-null? 1297 if( t == t_not_null ) return obj; 1298 1299 Node *cast = new (C, 2) CastPPNode(obj,t_not_null); 1300 cast->init_req(0, control()); 1301 cast = _gvn.transform( cast ); 1302 1303 // Scan for instances of 'obj' in the current JVM mapping. 1304 // These instances are known to be not-null after the test. 1305 if (do_replace_in_map) 1306 replace_in_map(obj, cast); 1307 1308 return cast; // Return casted value 1309 } 1310 1311 1312 //--------------------------replace_in_map------------------------------------- 1313 void GraphKit::replace_in_map(Node* old, Node* neww) { 1314 this->map()->replace_edge(old, neww); 1315 1316 // Note: This operation potentially replaces any edge 1317 // on the map. This includes locals, stack, and monitors 1318 // of the current (innermost) JVM state. 1319 1320 // We can consider replacing in caller maps. 1321 // The idea would be that an inlined function's null checks 1322 // can be shared with the entire inlining tree. 1323 // The expense of doing this is that the PreserveJVMState class 1324 // would have to preserve caller states too, with a deep copy. 1325 } 1326 1327 1328 1329 //============================================================================= 1330 //--------------------------------memory--------------------------------------- 1331 Node* GraphKit::memory(uint alias_idx) { 1332 MergeMemNode* mem = merged_memory(); 1333 Node* p = mem->memory_at(alias_idx); 1334 _gvn.set_type(p, Type::MEMORY); // must be mapped 1335 return p; 1336 } 1337 1338 //-----------------------------reset_memory------------------------------------ 1339 Node* GraphKit::reset_memory() { 1340 Node* mem = map()->memory(); 1341 // do not use this node for any more parsing! 1342 debug_only( map()->set_memory((Node*)NULL) ); 1343 return _gvn.transform( mem ); 1344 } 1345 1346 //------------------------------set_all_memory--------------------------------- 1347 void GraphKit::set_all_memory(Node* newmem) { 1348 Node* mergemem = MergeMemNode::make(C, newmem); 1349 gvn().set_type_bottom(mergemem); 1350 map()->set_memory(mergemem); 1351 } 1352 1353 //------------------------------set_all_memory_call---------------------------- 1354 void GraphKit::set_all_memory_call(Node* call) { 1355 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1356 set_all_memory(newmem); 1357 } 1358 1359 //============================================================================= 1360 // 1361 // parser factory methods for MemNodes 1362 // 1363 // These are layered on top of the factory methods in LoadNode and StoreNode, 1364 // and integrate with the parser's memory state and _gvn engine. 1365 // 1366 1367 // factory methods in "int adr_idx" 1368 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1369 int adr_idx, 1370 bool require_atomic_access) { 1371 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1372 const TypePtr* adr_type = NULL; // debug-mode-only argument 1373 debug_only(adr_type = C->get_adr_type(adr_idx)); 1374 Node* mem = memory(adr_idx); 1375 Node* ld; 1376 if (require_atomic_access && bt == T_LONG) { 1377 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); 1378 } else { 1379 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); 1380 } 1381 return _gvn.transform(ld); 1382 } 1383 1384 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1385 int adr_idx, 1386 bool require_atomic_access) { 1387 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1388 const TypePtr* adr_type = NULL; 1389 debug_only(adr_type = C->get_adr_type(adr_idx)); 1390 Node *mem = memory(adr_idx); 1391 Node* st; 1392 if (require_atomic_access && bt == T_LONG) { 1393 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); 1394 } else { 1395 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); 1396 } 1397 st = _gvn.transform(st); 1398 set_memory(st, adr_idx); 1399 // Back-to-back stores can only remove intermediate store with DU info 1400 // so push on worklist for optimizer. 1401 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1402 record_for_igvn(st); 1403 1404 return st; 1405 } 1406 1407 1408 void GraphKit::pre_barrier(Node* ctl, 1409 Node* obj, 1410 Node* adr, 1411 uint adr_idx, 1412 Node* val, 1413 const TypeOopPtr* val_type, 1414 BasicType bt) { 1415 BarrierSet* bs = Universe::heap()->barrier_set(); 1416 set_control(ctl); 1417 switch (bs->kind()) { 1418 case BarrierSet::G1SATBCT: 1419 case BarrierSet::G1SATBCTLogging: 1420 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt); 1421 break; 1422 1423 case BarrierSet::CardTableModRef: 1424 case BarrierSet::CardTableExtension: 1425 case BarrierSet::ModRef: 1426 break; 1427 1428 case BarrierSet::Other: 1429 default : 1430 ShouldNotReachHere(); 1431 1432 } 1433 } 1434 1435 void GraphKit::post_barrier(Node* ctl, 1436 Node* store, 1437 Node* obj, 1438 Node* adr, 1439 uint adr_idx, 1440 Node* val, 1441 BasicType bt, 1442 bool use_precise) { 1443 BarrierSet* bs = Universe::heap()->barrier_set(); 1444 set_control(ctl); 1445 switch (bs->kind()) { 1446 case BarrierSet::G1SATBCT: 1447 case BarrierSet::G1SATBCTLogging: 1448 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1449 break; 1450 1451 case BarrierSet::CardTableModRef: 1452 case BarrierSet::CardTableExtension: 1453 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1454 break; 1455 1456 case BarrierSet::ModRef: 1457 break; 1458 1459 case BarrierSet::Other: 1460 default : 1461 ShouldNotReachHere(); 1462 1463 } 1464 } 1465 1466 Node* GraphKit::store_oop(Node* ctl, 1467 Node* obj, 1468 Node* adr, 1469 const TypePtr* adr_type, 1470 Node* val, 1471 const TypeOopPtr* val_type, 1472 BasicType bt, 1473 bool use_precise) { 1474 1475 set_control(ctl); 1476 if (stopped()) return top(); // Dead path ? 1477 1478 assert(bt == T_OBJECT, "sanity"); 1479 assert(val != NULL, "not dead path"); 1480 uint adr_idx = C->get_alias_index(adr_type); 1481 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1482 1483 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt); 1484 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1485 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1486 return store; 1487 } 1488 1489 // Could be an array or object we don't know at compile time (unsafe ref.) 1490 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1491 Node* obj, // containing obj 1492 Node* adr, // actual adress to store val at 1493 const TypePtr* adr_type, 1494 Node* val, 1495 BasicType bt) { 1496 Compile::AliasType* at = C->alias_type(adr_type); 1497 const TypeOopPtr* val_type = NULL; 1498 if (adr_type->isa_instptr()) { 1499 if (at->field() != NULL) { 1500 // known field. This code is a copy of the do_put_xxx logic. 1501 ciField* field = at->field(); 1502 if (!field->type()->is_loaded()) { 1503 val_type = TypeInstPtr::BOTTOM; 1504 } else { 1505 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1506 } 1507 } 1508 } else if (adr_type->isa_aryptr()) { 1509 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1510 } 1511 if (val_type == NULL) { 1512 val_type = TypeInstPtr::BOTTOM; 1513 } 1514 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); 1515 } 1516 1517 1518 //-------------------------array_element_address------------------------- 1519 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1520 const TypeInt* sizetype) { 1521 uint shift = exact_log2(type2aelembytes(elembt)); 1522 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1523 1524 // short-circuit a common case (saves lots of confusing waste motion) 1525 jint idx_con = find_int_con(idx, -1); 1526 if (idx_con >= 0) { 1527 intptr_t offset = header + ((intptr_t)idx_con << shift); 1528 return basic_plus_adr(ary, offset); 1529 } 1530 1531 // must be correct type for alignment purposes 1532 Node* base = basic_plus_adr(ary, header); 1533 #ifdef _LP64 1534 // The scaled index operand to AddP must be a clean 64-bit value. 1535 // Java allows a 32-bit int to be incremented to a negative 1536 // value, which appears in a 64-bit register as a large 1537 // positive number. Using that large positive number as an 1538 // operand in pointer arithmetic has bad consequences. 1539 // On the other hand, 32-bit overflow is rare, and the possibility 1540 // can often be excluded, if we annotate the ConvI2L node with 1541 // a type assertion that its value is known to be a small positive 1542 // number. (The prior range check has ensured this.) 1543 // This assertion is used by ConvI2LNode::Ideal. 1544 int index_max = max_jint - 1; // array size is max_jint, index is one less 1545 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1546 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1547 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); 1548 #endif 1549 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); 1550 return basic_plus_adr(ary, base, scale); 1551 } 1552 1553 //-------------------------load_array_element------------------------- 1554 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1555 const Type* elemtype = arytype->elem(); 1556 BasicType elembt = elemtype->array_element_basic_type(); 1557 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1558 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1559 return ld; 1560 } 1561 1562 //-------------------------set_arguments_for_java_call------------------------- 1563 // Arguments (pre-popped from the stack) are taken from the JVMS. 1564 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1565 // Add the call arguments: 1566 uint nargs = call->method()->arg_size(); 1567 for (uint i = 0; i < nargs; i++) { 1568 Node* arg = argument(i); 1569 call->init_req(i + TypeFunc::Parms, arg); 1570 } 1571 } 1572 1573 //---------------------------set_edges_for_java_call--------------------------- 1574 // Connect a newly created call into the current JVMS. 1575 // A return value node (if any) is returned from set_edges_for_java_call. 1576 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) { 1577 1578 // Add the predefined inputs: 1579 call->init_req( TypeFunc::Control, control() ); 1580 call->init_req( TypeFunc::I_O , i_o() ); 1581 call->init_req( TypeFunc::Memory , reset_memory() ); 1582 call->init_req( TypeFunc::FramePtr, frameptr() ); 1583 call->init_req( TypeFunc::ReturnAdr, top() ); 1584 1585 add_safepoint_edges(call, must_throw); 1586 1587 Node* xcall = _gvn.transform(call); 1588 1589 if (xcall == top()) { 1590 set_control(top()); 1591 return; 1592 } 1593 assert(xcall == call, "call identity is stable"); 1594 1595 // Re-use the current map to produce the result. 1596 1597 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); 1598 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O ))); 1599 set_all_memory_call(xcall); 1600 1601 //return xcall; // no need, caller already has it 1602 } 1603 1604 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) { 1605 if (stopped()) return top(); // maybe the call folded up? 1606 1607 // Capture the return value, if any. 1608 Node* ret; 1609 if (call->method() == NULL || 1610 call->method()->return_type()->basic_type() == T_VOID) 1611 ret = top(); 1612 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 1613 1614 // Note: Since any out-of-line call can produce an exception, 1615 // we always insert an I_O projection from the call into the result. 1616 1617 make_slow_call_ex(call, env()->Throwable_klass(), false); 1618 1619 return ret; 1620 } 1621 1622 //--------------------set_predefined_input_for_runtime_call-------------------- 1623 // Reading and setting the memory state is way conservative here. 1624 // The real problem is that I am not doing real Type analysis on memory, 1625 // so I cannot distinguish card mark stores from other stores. Across a GC 1626 // point the Store Barrier and the card mark memory has to agree. I cannot 1627 // have a card mark store and its barrier split across the GC point from 1628 // either above or below. Here I get that to happen by reading ALL of memory. 1629 // A better answer would be to separate out card marks from other memory. 1630 // For now, return the input memory state, so that it can be reused 1631 // after the call, if this call has restricted memory effects. 1632 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1633 // Set fixed predefined input arguments 1634 Node* memory = reset_memory(); 1635 call->init_req( TypeFunc::Control, control() ); 1636 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1637 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1638 call->init_req( TypeFunc::FramePtr, frameptr() ); 1639 call->init_req( TypeFunc::ReturnAdr, top() ); 1640 return memory; 1641 } 1642 1643 //-------------------set_predefined_output_for_runtime_call-------------------- 1644 // Set control and memory (not i_o) from the call. 1645 // If keep_mem is not NULL, use it for the output state, 1646 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1647 // If hook_mem is NULL, this call produces no memory effects at all. 1648 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1649 // then only that memory slice is taken from the call. 1650 // In the last case, we must put an appropriate memory barrier before 1651 // the call, so as to create the correct anti-dependencies on loads 1652 // preceding the call. 1653 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1654 Node* keep_mem, 1655 const TypePtr* hook_mem) { 1656 // no i/o 1657 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); 1658 if (keep_mem) { 1659 // First clone the existing memory state 1660 set_all_memory(keep_mem); 1661 if (hook_mem != NULL) { 1662 // Make memory for the call 1663 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1664 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1665 // We also use hook_mem to extract specific effects from arraycopy stubs. 1666 set_memory(mem, hook_mem); 1667 } 1668 // ...else the call has NO memory effects. 1669 1670 // Make sure the call advertises its memory effects precisely. 1671 // This lets us build accurate anti-dependences in gcm.cpp. 1672 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1673 "call node must be constructed correctly"); 1674 } else { 1675 assert(hook_mem == NULL, ""); 1676 // This is not a "slow path" call; all memory comes from the call. 1677 set_all_memory_call(call); 1678 } 1679 } 1680 1681 //------------------------------increment_counter------------------------------ 1682 // for statistics: increment a VM counter by 1 1683 1684 void GraphKit::increment_counter(address counter_addr) { 1685 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1686 increment_counter(adr1); 1687 } 1688 1689 void GraphKit::increment_counter(Node* counter_addr) { 1690 int adr_type = Compile::AliasIdxRaw; 1691 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 1692 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 1693 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); 1694 } 1695 1696 1697 //------------------------------uncommon_trap---------------------------------- 1698 // Bail out to the interpreter in mid-method. Implemented by calling the 1699 // uncommon_trap blob. This helper function inserts a runtime call with the 1700 // right debug info. 1701 void GraphKit::uncommon_trap(int trap_request, 1702 ciKlass* klass, const char* comment, 1703 bool must_throw, 1704 bool keep_exact_action) { 1705 if (failing()) stop(); 1706 if (stopped()) return; // trap reachable? 1707 1708 // Note: If ProfileTraps is true, and if a deopt. actually 1709 // occurs here, the runtime will make sure an MDO exists. There is 1710 // no need to call method()->build_method_data() at this point. 1711 1712 #ifdef ASSERT 1713 if (!must_throw) { 1714 // Make sure the stack has at least enough depth to execute 1715 // the current bytecode. 1716 int inputs, ignore; 1717 if (compute_stack_effects(inputs, ignore)) { 1718 assert(sp() >= inputs, "must have enough JVMS stack to execute"); 1719 // It is a frequent error in library_call.cpp to issue an 1720 // uncommon trap with the _sp value already popped. 1721 } 1722 } 1723 #endif 1724 1725 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1726 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1727 1728 switch (action) { 1729 case Deoptimization::Action_maybe_recompile: 1730 case Deoptimization::Action_reinterpret: 1731 // Temporary fix for 6529811 to allow virtual calls to be sure they 1732 // get the chance to go from mono->bi->mega 1733 if (!keep_exact_action && 1734 Deoptimization::trap_request_index(trap_request) < 0 && 1735 too_many_recompiles(reason)) { 1736 // This BCI is causing too many recompilations. 1737 action = Deoptimization::Action_none; 1738 trap_request = Deoptimization::make_trap_request(reason, action); 1739 } else { 1740 C->set_trap_can_recompile(true); 1741 } 1742 break; 1743 case Deoptimization::Action_make_not_entrant: 1744 C->set_trap_can_recompile(true); 1745 break; 1746 #ifdef ASSERT 1747 case Deoptimization::Action_none: 1748 case Deoptimization::Action_make_not_compilable: 1749 break; 1750 default: 1751 assert(false, "bad action"); 1752 #endif 1753 } 1754 1755 if (TraceOptoParse) { 1756 char buf[100]; 1757 tty->print_cr("Uncommon trap %s at bci:%d", 1758 Deoptimization::format_trap_request(buf, sizeof(buf), 1759 trap_request), bci()); 1760 } 1761 1762 CompileLog* log = C->log(); 1763 if (log != NULL) { 1764 int kid = (klass == NULL)? -1: log->identify(klass); 1765 log->begin_elem("uncommon_trap bci='%d'", bci()); 1766 char buf[100]; 1767 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1768 trap_request)); 1769 if (kid >= 0) log->print(" klass='%d'", kid); 1770 if (comment != NULL) log->print(" comment='%s'", comment); 1771 log->end_elem(); 1772 } 1773 1774 // Make sure any guarding test views this path as very unlikely 1775 Node *i0 = control()->in(0); 1776 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1777 IfNode *iff = i0->as_If(); 1778 float f = iff->_prob; // Get prob 1779 if (control()->Opcode() == Op_IfTrue) { 1780 if (f > PROB_UNLIKELY_MAG(4)) 1781 iff->_prob = PROB_MIN; 1782 } else { 1783 if (f < PROB_LIKELY_MAG(4)) 1784 iff->_prob = PROB_MAX; 1785 } 1786 } 1787 1788 // Clear out dead values from the debug info. 1789 kill_dead_locals(); 1790 1791 // Now insert the uncommon trap subroutine call 1792 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); 1793 const TypePtr* no_memory_effects = NULL; 1794 // Pass the index of the class to be loaded 1795 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1796 (must_throw ? RC_MUST_THROW : 0), 1797 OptoRuntime::uncommon_trap_Type(), 1798 call_addr, "uncommon_trap", no_memory_effects, 1799 intcon(trap_request)); 1800 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 1801 "must extract request correctly from the graph"); 1802 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 1803 1804 call->set_req(TypeFunc::ReturnAdr, returnadr()); 1805 // The debug info is the only real input to this call. 1806 1807 // Halt-and-catch fire here. The above call should never return! 1808 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); 1809 _gvn.set_type_bottom(halt); 1810 root()->add_req(halt); 1811 1812 stop_and_kill_map(); 1813 } 1814 1815 1816 //--------------------------just_allocated_object------------------------------ 1817 // Report the object that was just allocated. 1818 // It must be the case that there are no intervening safepoints. 1819 // We use this to determine if an object is so "fresh" that 1820 // it does not require card marks. 1821 Node* GraphKit::just_allocated_object(Node* current_control) { 1822 if (C->recent_alloc_ctl() == current_control) 1823 return C->recent_alloc_obj(); 1824 return NULL; 1825 } 1826 1827 1828 void GraphKit::round_double_arguments(ciMethod* dest_method) { 1829 // (Note: TypeFunc::make has a cache that makes this fast.) 1830 const TypeFunc* tf = TypeFunc::make(dest_method); 1831 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 1832 for (int j = 0; j < nargs; j++) { 1833 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 1834 if( targ->basic_type() == T_DOUBLE ) { 1835 // If any parameters are doubles, they must be rounded before 1836 // the call, dstore_rounding does gvn.transform 1837 Node *arg = argument(j); 1838 arg = dstore_rounding(arg); 1839 set_argument(j, arg); 1840 } 1841 } 1842 } 1843 1844 void GraphKit::round_double_result(ciMethod* dest_method) { 1845 // A non-strict method may return a double value which has an extended 1846 // exponent, but this must not be visible in a caller which is 'strict' 1847 // If a strict caller invokes a non-strict callee, round a double result 1848 1849 BasicType result_type = dest_method->return_type()->basic_type(); 1850 assert( method() != NULL, "must have caller context"); 1851 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 1852 // Destination method's return value is on top of stack 1853 // dstore_rounding() does gvn.transform 1854 Node *result = pop_pair(); 1855 result = dstore_rounding(result); 1856 push_pair(result); 1857 } 1858 } 1859 1860 // rounding for strict float precision conformance 1861 Node* GraphKit::precision_rounding(Node* n) { 1862 return UseStrictFP && _method->flags().is_strict() 1863 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 1864 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) 1865 : n; 1866 } 1867 1868 // rounding for strict double precision conformance 1869 Node* GraphKit::dprecision_rounding(Node *n) { 1870 return UseStrictFP && _method->flags().is_strict() 1871 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 1872 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1873 : n; 1874 } 1875 1876 // rounding for non-strict double stores 1877 Node* GraphKit::dstore_rounding(Node* n) { 1878 return Matcher::strict_fp_requires_explicit_rounding 1879 && UseSSE <= 1 1880 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1881 : n; 1882 } 1883 1884 //============================================================================= 1885 // Generate a fast path/slow path idiom. Graph looks like: 1886 // [foo] indicates that 'foo' is a parameter 1887 // 1888 // [in] NULL 1889 // \ / 1890 // CmpP 1891 // Bool ne 1892 // If 1893 // / \ 1894 // True False-<2> 1895 // / | 1896 // / cast_not_null 1897 // Load | | ^ 1898 // [fast_test] | | 1899 // gvn to opt_test | | 1900 // / \ | <1> 1901 // True False | 1902 // | \\ | 1903 // [slow_call] \[fast_result] 1904 // Ctl Val \ \ 1905 // | \ \ 1906 // Catch <1> \ \ 1907 // / \ ^ \ \ 1908 // Ex No_Ex | \ \ 1909 // | \ \ | \ <2> \ 1910 // ... \ [slow_res] | | \ [null_result] 1911 // \ \--+--+--- | | 1912 // \ | / \ | / 1913 // --------Region Phi 1914 // 1915 //============================================================================= 1916 // Code is structured as a series of driver functions all called 'do_XXX' that 1917 // call a set of helper functions. Helper functions first, then drivers. 1918 1919 //------------------------------null_check_oop--------------------------------- 1920 // Null check oop. Set null-path control into Region in slot 3. 1921 // Make a cast-not-nullness use the other not-null control. Return cast. 1922 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 1923 bool never_see_null) { 1924 // Initial NULL check taken path 1925 (*null_control) = top(); 1926 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 1927 1928 // Generate uncommon_trap: 1929 if (never_see_null && (*null_control) != top()) { 1930 // If we see an unexpected null at a check-cast we record it and force a 1931 // recompile; the offending check-cast will be compiled to handle NULLs. 1932 // If we see more than one offending BCI, then all checkcasts in the 1933 // method will be compiled to handle NULLs. 1934 PreserveJVMState pjvms(this); 1935 set_control(*null_control); 1936 replace_in_map(value, null()); 1937 uncommon_trap(Deoptimization::Reason_null_check, 1938 Deoptimization::Action_make_not_entrant); 1939 (*null_control) = top(); // NULL path is dead 1940 } 1941 1942 // Cast away null-ness on the result 1943 return cast; 1944 } 1945 1946 //------------------------------opt_iff---------------------------------------- 1947 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 1948 // Return slow-path control. 1949 Node* GraphKit::opt_iff(Node* region, Node* iff) { 1950 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 1951 1952 // Fast path taken; set region slot 2 1953 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); 1954 region->init_req(2,fast_taken); // Capture fast-control 1955 1956 // Fast path not-taken, i.e. slow path 1957 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); 1958 return slow_taken; 1959 } 1960 1961 //-----------------------------make_runtime_call------------------------------- 1962 Node* GraphKit::make_runtime_call(int flags, 1963 const TypeFunc* call_type, address call_addr, 1964 const char* call_name, 1965 const TypePtr* adr_type, 1966 // The following parms are all optional. 1967 // The first NULL ends the list. 1968 Node* parm0, Node* parm1, 1969 Node* parm2, Node* parm3, 1970 Node* parm4, Node* parm5, 1971 Node* parm6, Node* parm7) { 1972 // Slow-path call 1973 int size = call_type->domain()->cnt(); 1974 bool is_leaf = !(flags & RC_NO_LEAF); 1975 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 1976 if (call_name == NULL) { 1977 assert(!is_leaf, "must supply name for leaf"); 1978 call_name = OptoRuntime::stub_name(call_addr); 1979 } 1980 CallNode* call; 1981 if (!is_leaf) { 1982 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, 1983 bci(), adr_type); 1984 } else if (flags & RC_NO_FP) { 1985 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 1986 } else { 1987 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); 1988 } 1989 1990 // The following is similar to set_edges_for_java_call, 1991 // except that the memory effects of the call are restricted to AliasIdxRaw. 1992 1993 // Slow path call has no side-effects, uses few values 1994 bool wide_in = !(flags & RC_NARROW_MEM); 1995 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 1996 1997 Node* prev_mem = NULL; 1998 if (wide_in) { 1999 prev_mem = set_predefined_input_for_runtime_call(call); 2000 } else { 2001 assert(!wide_out, "narrow in => narrow out"); 2002 Node* narrow_mem = memory(adr_type); 2003 prev_mem = reset_memory(); 2004 map()->set_memory(narrow_mem); 2005 set_predefined_input_for_runtime_call(call); 2006 } 2007 2008 // Hook each parm in order. Stop looking at the first NULL. 2009 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2010 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2011 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2012 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2013 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2014 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2015 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2016 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2017 /* close each nested if ===> */ } } } } } } } } 2018 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2019 2020 if (!is_leaf) { 2021 // Non-leaves can block and take safepoints: 2022 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2023 } 2024 // Non-leaves can throw exceptions: 2025 if (has_io) { 2026 call->set_req(TypeFunc::I_O, i_o()); 2027 } 2028 2029 if (flags & RC_UNCOMMON) { 2030 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2031 // (An "if" probability corresponds roughly to an unconditional count. 2032 // Sort of.) 2033 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2034 } 2035 2036 Node* c = _gvn.transform(call); 2037 assert(c == call, "cannot disappear"); 2038 2039 if (wide_out) { 2040 // Slow path call has full side-effects. 2041 set_predefined_output_for_runtime_call(call); 2042 } else { 2043 // Slow path call has few side-effects, and/or sets few values. 2044 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2045 } 2046 2047 if (has_io) { 2048 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); 2049 } 2050 return call; 2051 2052 } 2053 2054 //------------------------------merge_memory----------------------------------- 2055 // Merge memory from one path into the current memory state. 2056 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2057 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2058 Node* old_slice = mms.force_memory(); 2059 Node* new_slice = mms.memory2(); 2060 if (old_slice != new_slice) { 2061 PhiNode* phi; 2062 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2063 phi = new_slice->as_Phi(); 2064 #ifdef ASSERT 2065 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2066 old_slice = old_slice->in(new_path); 2067 // Caller is responsible for ensuring that any pre-existing 2068 // phis are already aware of old memory. 2069 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2070 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2071 #endif 2072 mms.set_memory(phi); 2073 } else { 2074 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2075 _gvn.set_type(phi, Type::MEMORY); 2076 phi->set_req(new_path, new_slice); 2077 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2078 } 2079 } 2080 } 2081 } 2082 2083 //------------------------------make_slow_call_ex------------------------------ 2084 // Make the exception handler hookups for the slow call 2085 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2086 if (stopped()) return; 2087 2088 // Make a catch node with just two handlers: fall-through and catch-all 2089 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2090 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); 2091 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2092 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2093 2094 { PreserveJVMState pjvms(this); 2095 set_control(excp); 2096 set_i_o(i_o); 2097 2098 if (excp != top()) { 2099 // Create an exception state also. 2100 // Use an exact type if the caller has specified a specific exception. 2101 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2102 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); 2103 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2104 } 2105 } 2106 2107 // Get the no-exception control from the CatchNode. 2108 set_control(norm); 2109 } 2110 2111 2112 //-------------------------------gen_subtype_check----------------------------- 2113 // Generate a subtyping check. Takes as input the subtype and supertype. 2114 // Returns 2 values: sets the default control() to the true path and returns 2115 // the false path. Only reads invariant memory; sets no (visible) memory. 2116 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2117 // but that's not exposed to the optimizer. This call also doesn't take in an 2118 // Object; if you wish to check an Object you need to load the Object's class 2119 // prior to coming here. 2120 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2121 // Fast check for identical types, perhaps identical constants. 2122 // The types can even be identical non-constants, in cases 2123 // involving Array.newInstance, Object.clone, etc. 2124 if (subklass == superklass) 2125 return top(); // false path is dead; no test needed. 2126 2127 if (_gvn.type(superklass)->singleton()) { 2128 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2129 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2130 2131 // In the common case of an exact superklass, try to fold up the 2132 // test before generating code. You may ask, why not just generate 2133 // the code and then let it fold up? The answer is that the generated 2134 // code will necessarily include null checks, which do not always 2135 // completely fold away. If they are also needless, then they turn 2136 // into a performance loss. Example: 2137 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2138 // Here, the type of 'fa' is often exact, so the store check 2139 // of fa[1]=x will fold up, without testing the nullness of x. 2140 switch (static_subtype_check(superk, subk)) { 2141 case SSC_always_false: 2142 { 2143 Node* always_fail = control(); 2144 set_control(top()); 2145 return always_fail; 2146 } 2147 case SSC_always_true: 2148 return top(); 2149 case SSC_easy_test: 2150 { 2151 // Just do a direct pointer compare and be done. 2152 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); 2153 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2154 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2155 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); 2156 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2157 } 2158 case SSC_full_test: 2159 break; 2160 default: 2161 ShouldNotReachHere(); 2162 } 2163 } 2164 2165 // %%% Possible further optimization: Even if the superklass is not exact, 2166 // if the subklass is the unique subtype of the superklass, the check 2167 // will always succeed. We could leave a dependency behind to ensure this. 2168 2169 // First load the super-klass's check-offset 2170 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); 2171 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2172 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); 2173 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2174 2175 // Load from the sub-klass's super-class display list, or a 1-word cache of 2176 // the secondary superclass list, or a failing value with a sentinel offset 2177 // if the super-klass is an interface or exceptionally deep in the Java 2178 // hierarchy and we have to scan the secondary superclass list the hard way. 2179 // Worst-case type is a little odd: NULL is allowed as a result (usually 2180 // klass loads can never produce a NULL). 2181 Node *chk_off_X = ConvI2X(chk_off); 2182 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); 2183 // For some types like interfaces the following loadKlass is from a 1-word 2184 // cache which is mutable so can't use immutable memory. Other 2185 // types load from the super-class display table which is immutable. 2186 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2187 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2188 2189 // Compile speed common case: ARE a subtype and we canNOT fail 2190 if( superklass == nkls ) 2191 return top(); // false path is dead; no test needed. 2192 2193 // See if we get an immediate positive hit. Happens roughly 83% of the 2194 // time. Test to see if the value loaded just previously from the subklass 2195 // is exactly the superklass. 2196 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); 2197 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); 2198 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2199 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); 2200 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); 2201 2202 // Compile speed common case: Check for being deterministic right now. If 2203 // chk_off is a constant and not equal to cacheoff then we are NOT a 2204 // subklass. In this case we need exactly the 1 test above and we can 2205 // return those results immediately. 2206 if (!might_be_cache) { 2207 Node* not_subtype_ctrl = control(); 2208 set_control(iftrue1); // We need exactly the 1 test above 2209 return not_subtype_ctrl; 2210 } 2211 2212 // Gather the various success & failures here 2213 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); 2214 record_for_igvn(r_ok_subtype); 2215 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); 2216 record_for_igvn(r_not_subtype); 2217 2218 r_ok_subtype->init_req(1, iftrue1); 2219 2220 // Check for immediate negative hit. Happens roughly 11% of the time (which 2221 // is roughly 63% of the remaining cases). Test to see if the loaded 2222 // check-offset points into the subklass display list or the 1-element 2223 // cache. If it points to the display (and NOT the cache) and the display 2224 // missed then it's not a subtype. 2225 Node *cacheoff = _gvn.intcon(cacheoff_con); 2226 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); 2227 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); 2228 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2229 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); 2230 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); 2231 2232 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2233 // No performance impact (too rare) but allows sharing of secondary arrays 2234 // which has some footprint reduction. 2235 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); 2236 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); 2237 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2238 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); 2239 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); 2240 2241 // -- Roads not taken here: -- 2242 // We could also have chosen to perform the self-check at the beginning 2243 // of this code sequence, as the assembler does. This would not pay off 2244 // the same way, since the optimizer, unlike the assembler, can perform 2245 // static type analysis to fold away many successful self-checks. 2246 // Non-foldable self checks work better here in second position, because 2247 // the initial primary superclass check subsumes a self-check for most 2248 // types. An exception would be a secondary type like array-of-interface, 2249 // which does not appear in its own primary supertype display. 2250 // Finally, we could have chosen to move the self-check into the 2251 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2252 // dependent manner. But it is worthwhile to have the check here, 2253 // where it can be perhaps be optimized. The cost in code space is 2254 // small (register compare, branch). 2255 2256 // Now do a linear scan of the secondary super-klass array. Again, no real 2257 // performance impact (too rare) but it's gotta be done. 2258 // Since the code is rarely used, there is no penalty for moving it 2259 // out of line, and it can only improve I-cache density. 2260 // The decision to inline or out-of-line this final check is platform 2261 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2262 Node* psc = _gvn.transform( 2263 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2264 2265 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); 2266 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); 2267 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2268 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); 2269 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); 2270 2271 // Return false path; set default control to true path. 2272 set_control( _gvn.transform(r_ok_subtype) ); 2273 return _gvn.transform(r_not_subtype); 2274 } 2275 2276 //----------------------------static_subtype_check----------------------------- 2277 // Shortcut important common cases when superklass is exact: 2278 // (0) superklass is java.lang.Object (can occur in reflective code) 2279 // (1) subklass is already limited to a subtype of superklass => always ok 2280 // (2) subklass does not overlap with superklass => always fail 2281 // (3) superklass has NO subtypes and we can check with a simple compare. 2282 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2283 if (StressReflectiveCode) { 2284 return SSC_full_test; // Let caller generate the general case. 2285 } 2286 2287 if (superk == env()->Object_klass()) { 2288 return SSC_always_true; // (0) this test cannot fail 2289 } 2290 2291 ciType* superelem = superk; 2292 if (superelem->is_array_klass()) 2293 superelem = superelem->as_array_klass()->base_element_type(); 2294 2295 if (!subk->is_interface()) { // cannot trust static interface types yet 2296 if (subk->is_subtype_of(superk)) { 2297 return SSC_always_true; // (1) false path dead; no dynamic test needed 2298 } 2299 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2300 !superk->is_subtype_of(subk)) { 2301 return SSC_always_false; 2302 } 2303 } 2304 2305 // If casting to an instance klass, it must have no subtypes 2306 if (superk->is_interface()) { 2307 // Cannot trust interfaces yet. 2308 // %%% S.B. superk->nof_implementors() == 1 2309 } else if (superelem->is_instance_klass()) { 2310 ciInstanceKlass* ik = superelem->as_instance_klass(); 2311 if (!ik->has_subklass() && !ik->is_interface()) { 2312 if (!ik->is_final()) { 2313 // Add a dependency if there is a chance of a later subclass. 2314 C->dependencies()->assert_leaf_type(ik); 2315 } 2316 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2317 } 2318 } else { 2319 // A primitive array type has no subtypes. 2320 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2321 } 2322 2323 return SSC_full_test; 2324 } 2325 2326 // Profile-driven exact type check: 2327 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2328 float prob, 2329 Node* *casted_receiver) { 2330 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2331 Node* recv_klass = load_object_klass(receiver); 2332 Node* want_klass = makecon(tklass); 2333 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); 2334 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2335 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2336 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); 2337 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2338 2339 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2340 assert(recv_xtype->klass_is_exact(), ""); 2341 2342 // Subsume downstream occurrences of receiver with a cast to 2343 // recv_xtype, since now we know what the type will be. 2344 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); 2345 (*casted_receiver) = _gvn.transform(cast); 2346 // (User must make the replace_in_map call.) 2347 2348 return fail; 2349 } 2350 2351 2352 //-------------------------------gen_instanceof-------------------------------- 2353 // Generate an instance-of idiom. Used by both the instance-of bytecode 2354 // and the reflective instance-of call. 2355 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { 2356 C->set_has_split_ifs(true); // Has chance for split-if optimization 2357 assert( !stopped(), "dead parse path should be checked in callers" ); 2358 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2359 "must check for not-null not-dead klass in callers"); 2360 2361 // Make the merge point 2362 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2363 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2364 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2365 C->set_has_split_ifs(true); // Has chance for split-if optimization 2366 2367 // Null check; get casted pointer; set region slot 3 2368 Node* null_ctl = top(); 2369 Node* not_null_obj = null_check_oop(subobj, &null_ctl); 2370 2371 // If not_null_obj is dead, only null-path is taken 2372 if (stopped()) { // Doing instance-of on a NULL? 2373 set_control(null_ctl); 2374 return intcon(0); 2375 } 2376 region->init_req(_null_path, null_ctl); 2377 phi ->init_req(_null_path, intcon(0)); // Set null path value 2378 2379 // Load the object's klass 2380 Node* obj_klass = load_object_klass(not_null_obj); 2381 2382 // Generate the subtype check 2383 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2384 2385 // Plug in the success path to the general merge in slot 1. 2386 region->init_req(_obj_path, control()); 2387 phi ->init_req(_obj_path, intcon(1)); 2388 2389 // Plug in the failing path to the general merge in slot 2. 2390 region->init_req(_fail_path, not_subtype_ctrl); 2391 phi ->init_req(_fail_path, intcon(0)); 2392 2393 // Return final merged results 2394 set_control( _gvn.transform(region) ); 2395 record_for_igvn(region); 2396 return _gvn.transform(phi); 2397 } 2398 2399 //-------------------------------gen_checkcast--------------------------------- 2400 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2401 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2402 // uncommon-trap paths work. Adjust stack after this call. 2403 // If failure_control is supplied and not null, it is filled in with 2404 // the control edge for the cast failure. Otherwise, an appropriate 2405 // uncommon trap or exception is thrown. 2406 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2407 Node* *failure_control) { 2408 kill_dead_locals(); // Benefit all the uncommon traps 2409 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2410 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2411 2412 // Fast cutout: Check the case that the cast is vacuously true. 2413 // This detects the common cases where the test will short-circuit 2414 // away completely. We do this before we perform the null check, 2415 // because if the test is going to turn into zero code, we don't 2416 // want a residual null check left around. (Causes a slowdown, 2417 // for example, in some objArray manipulations, such as a[i]=a[j].) 2418 if (tk->singleton()) { 2419 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2420 if (objtp != NULL && objtp->klass() != NULL) { 2421 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2422 case SSC_always_true: 2423 return obj; 2424 case SSC_always_false: 2425 // It needs a null check because a null will *pass* the cast check. 2426 // A non-null value will always produce an exception. 2427 return do_null_assert(obj, T_OBJECT); 2428 } 2429 } 2430 } 2431 2432 ciProfileData* data = NULL; 2433 if (failure_control == NULL) { // use MDO in regular case only 2434 assert(java_bc() == Bytecodes::_aastore || 2435 java_bc() == Bytecodes::_checkcast, 2436 "interpreter profiles type checks only for these BCs"); 2437 data = method()->method_data()->bci_to_data(bci()); 2438 } 2439 2440 // Make the merge point 2441 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2442 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2443 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); 2444 C->set_has_split_ifs(true); // Has chance for split-if optimization 2445 2446 // Use null-cast information if it is available 2447 bool never_see_null = false; 2448 // If we see an unexpected null at a check-cast we record it and force a 2449 // recompile; the offending check-cast will be compiled to handle NULLs. 2450 // If we see several offending BCIs, then all checkcasts in the 2451 // method will be compiled to handle NULLs. 2452 if (UncommonNullCast // Cutout for this technique 2453 && failure_control == NULL // regular case 2454 && obj != null() // And not the -Xcomp stupid case? 2455 && !too_many_traps(Deoptimization::Reason_null_check)) { 2456 // Finally, check the "null_seen" bit from the interpreter. 2457 if (data == NULL || !data->as_BitData()->null_seen()) { 2458 never_see_null = true; 2459 } 2460 } 2461 2462 // Null check; get casted pointer; set region slot 3 2463 Node* null_ctl = top(); 2464 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2465 2466 // If not_null_obj is dead, only null-path is taken 2467 if (stopped()) { // Doing instance-of on a NULL? 2468 set_control(null_ctl); 2469 return null(); 2470 } 2471 region->init_req(_null_path, null_ctl); 2472 phi ->init_req(_null_path, null()); // Set null path value 2473 2474 Node* cast_obj = NULL; // the casted version of the object 2475 2476 // If the profile has seen exactly one type, narrow to that type. 2477 // (The subsequent subtype check will always fold up.) 2478 if (UseTypeProfile && TypeProfileCasts && data != NULL && 2479 // Counter has never been decremented (due to cast failure). 2480 // ...This is a reasonable thing to expect. It is true of 2481 // all casts inserted by javac to implement generic types. 2482 data->as_CounterData()->count() >= 0 && 2483 !too_many_traps(Deoptimization::Reason_class_check)) { 2484 // (No, this isn't a call, but it's enough like a virtual call 2485 // to use the same ciMethod accessor to get the profile info...) 2486 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2487 if (profile.count() >= 0 && // no cast failures here 2488 profile.has_receiver(0) && 2489 profile.morphism() == 1) { 2490 ciKlass* exact_kls = profile.receiver(0); 2491 int ssc = static_subtype_check(tk->klass(), exact_kls); 2492 if (ssc == SSC_always_true) { 2493 // If we narrow the type to match what the type profile sees, 2494 // we can then remove the rest of the cast. 2495 // This is a win, even if the exact_kls is very specific, 2496 // because downstream operations, such as method calls, 2497 // will often benefit from the sharper type. 2498 Node* exact_obj = not_null_obj; // will get updated in place... 2499 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2500 &exact_obj); 2501 { PreserveJVMState pjvms(this); 2502 set_control(slow_ctl); 2503 uncommon_trap(Deoptimization::Reason_class_check, 2504 Deoptimization::Action_maybe_recompile); 2505 } 2506 if (failure_control != NULL) // failure is now impossible 2507 (*failure_control) = top(); 2508 replace_in_map(not_null_obj, exact_obj); 2509 // adjust the type of the phi to the exact klass: 2510 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); 2511 cast_obj = exact_obj; 2512 } 2513 // assert(cast_obj != NULL)... except maybe the profile lied to us. 2514 } 2515 } 2516 2517 if (cast_obj == NULL) { 2518 // Load the object's klass 2519 Node* obj_klass = load_object_klass(not_null_obj); 2520 2521 // Generate the subtype check 2522 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2523 2524 // Plug in success path into the merge 2525 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), 2526 not_null_obj, toop)); 2527 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2528 if (failure_control == NULL) { 2529 if (not_subtype_ctrl != top()) { // If failure is possible 2530 PreserveJVMState pjvms(this); 2531 set_control(not_subtype_ctrl); 2532 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2533 } 2534 } else { 2535 (*failure_control) = not_subtype_ctrl; 2536 } 2537 } 2538 2539 region->init_req(_obj_path, control()); 2540 phi ->init_req(_obj_path, cast_obj); 2541 2542 // A merge of NULL or Casted-NotNull obj 2543 Node* res = _gvn.transform(phi); 2544 2545 // Note I do NOT always 'replace_in_map(obj,result)' here. 2546 // if( tk->klass()->can_be_primary_super() ) 2547 // This means that if I successfully store an Object into an array-of-String 2548 // I 'forget' that the Object is really now known to be a String. I have to 2549 // do this because we don't have true union types for interfaces - if I store 2550 // a Baz into an array-of-Interface and then tell the optimizer it's an 2551 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2552 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2553 // replace_in_map( obj, res ); 2554 2555 // Return final merged results 2556 set_control( _gvn.transform(region) ); 2557 record_for_igvn(region); 2558 return res; 2559 } 2560 2561 //------------------------------next_monitor----------------------------------- 2562 // What number should be given to the next monitor? 2563 int GraphKit::next_monitor() { 2564 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2565 int next = current + C->sync_stack_slots(); 2566 // Keep the toplevel high water mark current: 2567 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2568 return current; 2569 } 2570 2571 //------------------------------insert_mem_bar--------------------------------- 2572 // Memory barrier to avoid floating things around 2573 // The membar serves as a pinch point between both control and all memory slices. 2574 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2575 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2576 mb->init_req(TypeFunc::Control, control()); 2577 mb->init_req(TypeFunc::Memory, reset_memory()); 2578 Node* membar = _gvn.transform(mb); 2579 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); 2580 set_all_memory_call(membar); 2581 return membar; 2582 } 2583 2584 //-------------------------insert_mem_bar_volatile---------------------------- 2585 // Memory barrier to avoid floating things around 2586 // The membar serves as a pinch point between both control and memory(alias_idx). 2587 // If you want to make a pinch point on all memory slices, do not use this 2588 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2589 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2590 // When Parse::do_put_xxx updates a volatile field, it appends a series 2591 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2592 // The first membar is on the same memory slice as the field store opcode. 2593 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2594 // All the other membars (for other volatile slices, including AliasIdxBot, 2595 // which stands for all unknown volatile slices) are control-dependent 2596 // on the first membar. This prevents later volatile loads or stores 2597 // from sliding up past the just-emitted store. 2598 2599 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2600 mb->set_req(TypeFunc::Control,control()); 2601 if (alias_idx == Compile::AliasIdxBot) { 2602 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2603 } else { 2604 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2605 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2606 } 2607 Node* membar = _gvn.transform(mb); 2608 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); 2609 if (alias_idx == Compile::AliasIdxBot) { 2610 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); 2611 } else { 2612 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2613 } 2614 return membar; 2615 } 2616 2617 //------------------------------shared_lock------------------------------------ 2618 // Emit locking code. 2619 FastLockNode* GraphKit::shared_lock(Node* obj) { 2620 // bci is either a monitorenter bc or InvocationEntryBci 2621 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2622 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2623 2624 if( !GenerateSynchronizationCode ) 2625 return NULL; // Not locking things? 2626 if (stopped()) // Dead monitor? 2627 return NULL; 2628 2629 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2630 2631 // Box the stack location 2632 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); 2633 Node* mem = reset_memory(); 2634 2635 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); 2636 if (PrintPreciseBiasedLockingStatistics) { 2637 // Create the counters for this fast lock. 2638 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2639 } 2640 // Add monitor to debug info for the slow path. If we block inside the 2641 // slow path and de-opt, we need the monitor hanging around 2642 map()->push_monitor( flock ); 2643 2644 const TypeFunc *tf = LockNode::lock_type(); 2645 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); 2646 2647 lock->init_req( TypeFunc::Control, control() ); 2648 lock->init_req( TypeFunc::Memory , mem ); 2649 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2650 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2651 lock->init_req( TypeFunc::ReturnAdr, top() ); 2652 2653 lock->init_req(TypeFunc::Parms + 0, obj); 2654 lock->init_req(TypeFunc::Parms + 1, box); 2655 lock->init_req(TypeFunc::Parms + 2, flock); 2656 add_safepoint_edges(lock); 2657 2658 lock = _gvn.transform( lock )->as_Lock(); 2659 2660 // lock has no side-effects, sets few values 2661 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2662 2663 insert_mem_bar(Op_MemBarAcquire); 2664 2665 // Add this to the worklist so that the lock can be eliminated 2666 record_for_igvn(lock); 2667 2668 #ifndef PRODUCT 2669 if (PrintLockStatistics) { 2670 // Update the counter for this lock. Don't bother using an atomic 2671 // operation since we don't require absolute accuracy. 2672 lock->create_lock_counter(map()->jvms()); 2673 int adr_type = Compile::AliasIdxRaw; 2674 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); 2675 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 2676 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 2677 store_to_memory(control(), counter_addr, incr, T_INT, adr_type); 2678 } 2679 #endif 2680 2681 return flock; 2682 } 2683 2684 2685 //------------------------------shared_unlock---------------------------------- 2686 // Emit unlocking code. 2687 void GraphKit::shared_unlock(Node* box, Node* obj) { 2688 // bci is either a monitorenter bc or InvocationEntryBci 2689 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2690 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2691 2692 if( !GenerateSynchronizationCode ) 2693 return; 2694 if (stopped()) { // Dead monitor? 2695 map()->pop_monitor(); // Kill monitor from debug info 2696 return; 2697 } 2698 2699 // Memory barrier to avoid floating things down past the locked region 2700 insert_mem_bar(Op_MemBarRelease); 2701 2702 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2703 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); 2704 uint raw_idx = Compile::AliasIdxRaw; 2705 unlock->init_req( TypeFunc::Control, control() ); 2706 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2707 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2708 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2709 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2710 2711 unlock->init_req(TypeFunc::Parms + 0, obj); 2712 unlock->init_req(TypeFunc::Parms + 1, box); 2713 unlock = _gvn.transform(unlock)->as_Unlock(); 2714 2715 Node* mem = reset_memory(); 2716 2717 // unlock has no side-effects, sets few values 2718 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 2719 2720 // Kill monitor from debug info 2721 map()->pop_monitor( ); 2722 } 2723 2724 //-------------------------------get_layout_helper----------------------------- 2725 // If the given klass is a constant or known to be an array, 2726 // fetch the constant layout helper value into constant_value 2727 // and return (Node*)NULL. Otherwise, load the non-constant 2728 // layout helper value, and return the node which represents it. 2729 // This two-faced routine is useful because allocation sites 2730 // almost always feature constant types. 2731 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 2732 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 2733 if (!StressReflectiveCode && inst_klass != NULL) { 2734 ciKlass* klass = inst_klass->klass(); 2735 bool xklass = inst_klass->klass_is_exact(); 2736 if (xklass || klass->is_array_klass()) { 2737 jint lhelper = klass->layout_helper(); 2738 if (lhelper != Klass::_lh_neutral_value) { 2739 constant_value = lhelper; 2740 return (Node*) NULL; 2741 } 2742 } 2743 } 2744 constant_value = Klass::_lh_neutral_value; // put in a known value 2745 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); 2746 return make_load(NULL, lhp, TypeInt::INT, T_INT); 2747 } 2748 2749 // We just put in an allocate/initialize with a big raw-memory effect. 2750 // Hook selected additional alias categories on the initialization. 2751 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 2752 MergeMemNode* init_in_merge, 2753 Node* init_out_raw) { 2754 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 2755 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 2756 2757 Node* prevmem = kit.memory(alias_idx); 2758 init_in_merge->set_memory_at(alias_idx, prevmem); 2759 kit.set_memory(init_out_raw, alias_idx); 2760 } 2761 2762 //---------------------------set_output_for_allocation------------------------- 2763 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 2764 const TypeOopPtr* oop_type, 2765 bool raw_mem_only) { 2766 int rawidx = Compile::AliasIdxRaw; 2767 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 2768 add_safepoint_edges(alloc); 2769 Node* allocx = _gvn.transform(alloc); 2770 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); 2771 // create memory projection for i_o 2772 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 2773 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); 2774 2775 // create a memory projection as for the normal control path 2776 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); 2777 set_memory(malloc, rawidx); 2778 2779 // a normal slow-call doesn't change i_o, but an allocation does 2780 // we create a separate i_o projection for the normal control path 2781 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); 2782 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); 2783 2784 // put in an initialization barrier 2785 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 2786 rawoop)->as_Initialize(); 2787 assert(alloc->initialization() == init, "2-way macro link must work"); 2788 assert(init ->allocation() == alloc, "2-way macro link must work"); 2789 if (ReduceFieldZeroing && !raw_mem_only) { 2790 // Extract memory strands which may participate in the new object's 2791 // initialization, and source them from the new InitializeNode. 2792 // This will allow us to observe initializations when they occur, 2793 // and link them properly (as a group) to the InitializeNode. 2794 assert(init->in(InitializeNode::Memory) == malloc, ""); 2795 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 2796 init->set_req(InitializeNode::Memory, minit_in); 2797 record_for_igvn(minit_in); // fold it up later, if possible 2798 Node* minit_out = memory(rawidx); 2799 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 2800 if (oop_type->isa_aryptr()) { 2801 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 2802 int elemidx = C->get_alias_index(telemref); 2803 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 2804 } else if (oop_type->isa_instptr()) { 2805 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 2806 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 2807 ciField* field = ik->nonstatic_field_at(i); 2808 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 2809 continue; // do not bother to track really large numbers of fields 2810 // Find (or create) the alias category for this field: 2811 int fieldidx = C->alias_type(field)->index(); 2812 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 2813 } 2814 } 2815 } 2816 2817 // Cast raw oop to the real thing... 2818 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); 2819 javaoop = _gvn.transform(javaoop); 2820 C->set_recent_alloc(control(), javaoop); 2821 assert(just_allocated_object(control()) == javaoop, "just allocated"); 2822 2823 #ifdef ASSERT 2824 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 2825 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 2826 "Ideal_allocation works"); 2827 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 2828 "Ideal_allocation works"); 2829 if (alloc->is_AllocateArray()) { 2830 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 2831 "Ideal_allocation works"); 2832 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 2833 "Ideal_allocation works"); 2834 } else { 2835 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 2836 } 2837 } 2838 #endif //ASSERT 2839 2840 return javaoop; 2841 } 2842 2843 //---------------------------new_instance-------------------------------------- 2844 // This routine takes a klass_node which may be constant (for a static type) 2845 // or may be non-constant (for reflective code). It will work equally well 2846 // for either, and the graph will fold nicely if the optimizer later reduces 2847 // the type to a constant. 2848 // The optional arguments are for specialized use by intrinsics: 2849 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 2850 // - If 'raw_mem_only', do not cast the result to an oop. 2851 // - If 'return_size_val', report the the total object size to the caller. 2852 Node* GraphKit::new_instance(Node* klass_node, 2853 Node* extra_slow_test, 2854 bool raw_mem_only, // affect only raw memory 2855 Node* *return_size_val) { 2856 // Compute size in doublewords 2857 // The size is always an integral number of doublewords, represented 2858 // as a positive bytewise size stored in the klass's layout_helper. 2859 // The layout_helper also encodes (in a low bit) the need for a slow path. 2860 jint layout_con = Klass::_lh_neutral_value; 2861 Node* layout_val = get_layout_helper(klass_node, layout_con); 2862 int layout_is_con = (layout_val == NULL); 2863 2864 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 2865 // Generate the initial go-slow test. It's either ALWAYS (return a 2866 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 2867 // case) a computed value derived from the layout_helper. 2868 Node* initial_slow_test = NULL; 2869 if (layout_is_con) { 2870 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2871 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 2872 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 2873 2874 } else { // reflective case 2875 // This reflective path is used by Unsafe.allocateInstance. 2876 // (It may be stress-tested by specifying StressReflectiveCode.) 2877 // Basically, we want to get into the VM is there's an illegal argument. 2878 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 2879 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); 2880 if (extra_slow_test != intcon(0)) { 2881 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); 2882 } 2883 // (Macro-expander will further convert this to a Bool, if necessary.) 2884 } 2885 2886 // Find the size in bytes. This is easy; it's the layout_helper. 2887 // The size value must be valid even if the slow path is taken. 2888 Node* size = NULL; 2889 if (layout_is_con) { 2890 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 2891 } else { // reflective case 2892 // This reflective path is used by clone and Unsafe.allocateInstance. 2893 size = ConvI2X(layout_val); 2894 2895 // Clear the low bits to extract layout_helper_size_in_bytes: 2896 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 2897 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 2898 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); 2899 } 2900 if (return_size_val != NULL) { 2901 (*return_size_val) = size; 2902 } 2903 2904 // This is a precise notnull oop of the klass. 2905 // (Actually, it need not be precise if this is a reflective allocation.) 2906 // It's what we cast the result to. 2907 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 2908 if (!tklass) tklass = TypeKlassPtr::OBJECT; 2909 const TypeOopPtr* oop_type = tklass->as_instance_type(); 2910 2911 // Now generate allocation code 2912 2913 // The entire memory state is needed for slow path of the allocation 2914 // since GC and deoptimization can happened. 2915 Node *mem = reset_memory(); 2916 set_all_memory(mem); // Create new memory state 2917 2918 AllocateNode* alloc 2919 = new (C, AllocateNode::ParmLimit) 2920 AllocateNode(C, AllocateNode::alloc_type(), 2921 control(), mem, i_o(), 2922 size, klass_node, 2923 initial_slow_test); 2924 2925 return set_output_for_allocation(alloc, oop_type, raw_mem_only); 2926 } 2927 2928 //-------------------------------new_array------------------------------------- 2929 // helper for both newarray and anewarray 2930 // The 'length' parameter is (obviously) the length of the array. 2931 // See comments on new_instance for the meaning of the other arguments. 2932 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 2933 Node* length, // number of array elements 2934 int nargs, // number of arguments to push back for uncommon trap 2935 bool raw_mem_only, // affect only raw memory 2936 Node* *return_size_val) { 2937 jint layout_con = Klass::_lh_neutral_value; 2938 Node* layout_val = get_layout_helper(klass_node, layout_con); 2939 int layout_is_con = (layout_val == NULL); 2940 2941 if (!layout_is_con && !StressReflectiveCode && 2942 !too_many_traps(Deoptimization::Reason_class_check)) { 2943 // This is a reflective array creation site. 2944 // Optimistically assume that it is a subtype of Object[], 2945 // so that we can fold up all the address arithmetic. 2946 layout_con = Klass::array_layout_helper(T_OBJECT); 2947 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); 2948 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); 2949 { BuildCutout unless(this, bol_lh, PROB_MAX); 2950 _sp += nargs; 2951 uncommon_trap(Deoptimization::Reason_class_check, 2952 Deoptimization::Action_maybe_recompile); 2953 } 2954 layout_val = NULL; 2955 layout_is_con = true; 2956 } 2957 2958 // Generate the initial go-slow test. Make sure we do not overflow 2959 // if length is huge (near 2Gig) or negative! We do not need 2960 // exact double-words here, just a close approximation of needed 2961 // double-words. We can't add any offset or rounding bits, lest we 2962 // take a size -1 of bytes and make it positive. Use an unsigned 2963 // compare, so negative sizes look hugely positive. 2964 int fast_size_limit = FastAllocateSizeLimit; 2965 if (layout_is_con) { 2966 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2967 // Increase the size limit if we have exact knowledge of array type. 2968 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 2969 fast_size_limit <<= (LogBytesPerLong - log2_esize); 2970 } 2971 2972 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); 2973 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 2974 if (initial_slow_test->is_Bool()) { 2975 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 2976 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 2977 } 2978 2979 // --- Size Computation --- 2980 // array_size = round_to_heap(array_header + (length << elem_shift)); 2981 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 2982 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 2983 // The rounding mask is strength-reduced, if possible. 2984 int round_mask = MinObjAlignmentInBytes - 1; 2985 Node* header_size = NULL; 2986 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 2987 // (T_BYTE has the weakest alignment and size restrictions...) 2988 if (layout_is_con) { 2989 int hsize = Klass::layout_helper_header_size(layout_con); 2990 int eshift = Klass::layout_helper_log2_element_size(layout_con); 2991 BasicType etype = Klass::layout_helper_element_type(layout_con); 2992 if ((round_mask & ~right_n_bits(eshift)) == 0) 2993 round_mask = 0; // strength-reduce it if it goes away completely 2994 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 2995 assert(header_size_min <= hsize, "generic minimum is smallest"); 2996 header_size_min = hsize; 2997 header_size = intcon(hsize + round_mask); 2998 } else { 2999 Node* hss = intcon(Klass::_lh_header_size_shift); 3000 Node* hsm = intcon(Klass::_lh_header_size_mask); 3001 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); 3002 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); 3003 Node* mask = intcon(round_mask); 3004 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); 3005 } 3006 3007 Node* elem_shift = NULL; 3008 if (layout_is_con) { 3009 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3010 if (eshift != 0) 3011 elem_shift = intcon(eshift); 3012 } else { 3013 // There is no need to mask or shift this value. 3014 // The semantics of LShiftINode include an implicit mask to 0x1F. 3015 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3016 elem_shift = layout_val; 3017 } 3018 3019 // Transition to native address size for all offset calculations: 3020 Node* lengthx = ConvI2X(length); 3021 Node* headerx = ConvI2X(header_size); 3022 #ifdef _LP64 3023 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3024 if (tllen != NULL && tllen->_lo < 0) { 3025 // Add a manual constraint to a positive range. Cf. array_element_address. 3026 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3027 if (size_max > tllen->_hi) size_max = tllen->_hi; 3028 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3029 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); 3030 } 3031 } 3032 #endif 3033 3034 // Combine header size (plus rounding) and body size. Then round down. 3035 // This computation cannot overflow, because it is used only in two 3036 // places, one where the length is sharply limited, and the other 3037 // after a successful allocation. 3038 Node* abody = lengthx; 3039 if (elem_shift != NULL) 3040 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); 3041 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); 3042 if (round_mask != 0) { 3043 Node* mask = MakeConX(~round_mask); 3044 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); 3045 } 3046 // else if round_mask == 0, the size computation is self-rounding 3047 3048 if (return_size_val != NULL) { 3049 // This is the size 3050 (*return_size_val) = size; 3051 } 3052 3053 // Now generate allocation code 3054 3055 // The entire memory state is needed for slow path of the allocation 3056 // since GC and deoptimization can happened. 3057 Node *mem = reset_memory(); 3058 set_all_memory(mem); // Create new memory state 3059 3060 // Create the AllocateArrayNode and its result projections 3061 AllocateArrayNode* alloc 3062 = new (C, AllocateArrayNode::ParmLimit) 3063 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3064 control(), mem, i_o(), 3065 size, klass_node, 3066 initial_slow_test, 3067 length); 3068 3069 // Cast to correct type. Note that the klass_node may be constant or not, 3070 // and in the latter case the actual array type will be inexact also. 3071 // (This happens via a non-constant argument to inline_native_newArray.) 3072 // In any case, the value of klass_node provides the desired array type. 3073 const TypeInt* length_type = _gvn.find_int_type(length); 3074 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3075 if (ary_type->isa_aryptr() && length_type != NULL) { 3076 // Try to get a better type than POS for the size 3077 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3078 } 3079 3080 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); 3081 3082 // Cast length on remaining path to be as narrow as possible 3083 if (map()->find_edge(length) >= 0) { 3084 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3085 if (ccast != length) { 3086 _gvn.set_type_bottom(ccast); 3087 record_for_igvn(ccast); 3088 replace_in_map(length, ccast); 3089 } 3090 } 3091 3092 return javaoop; 3093 } 3094 3095 // The following "Ideal_foo" functions are placed here because they recognize 3096 // the graph shapes created by the functions immediately above. 3097 3098 //---------------------------Ideal_allocation---------------------------------- 3099 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3100 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3101 if (ptr == NULL) { // reduce dumb test in callers 3102 return NULL; 3103 } 3104 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3105 ptr = ptr->in(1); 3106 if (ptr == NULL) return NULL; 3107 } 3108 if (ptr->is_Proj()) { 3109 Node* allo = ptr->in(0); 3110 if (allo != NULL && allo->is_Allocate()) { 3111 return allo->as_Allocate(); 3112 } 3113 } 3114 // Report failure to match. 3115 return NULL; 3116 } 3117 3118 // Fancy version which also strips off an offset (and reports it to caller). 3119 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3120 intptr_t& offset) { 3121 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3122 if (base == NULL) return NULL; 3123 return Ideal_allocation(base, phase); 3124 } 3125 3126 // Trace Initialize <- Proj[Parm] <- Allocate 3127 AllocateNode* InitializeNode::allocation() { 3128 Node* rawoop = in(InitializeNode::RawAddress); 3129 if (rawoop->is_Proj()) { 3130 Node* alloc = rawoop->in(0); 3131 if (alloc->is_Allocate()) { 3132 return alloc->as_Allocate(); 3133 } 3134 } 3135 return NULL; 3136 } 3137 3138 // Trace Allocate -> Proj[Parm] -> Initialize 3139 InitializeNode* AllocateNode::initialization() { 3140 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3141 if (rawoop == NULL) return NULL; 3142 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3143 Node* init = rawoop->fast_out(i); 3144 if (init->is_Initialize()) { 3145 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3146 return init->as_Initialize(); 3147 } 3148 } 3149 return NULL; 3150 } 3151 3152 //----------------------------- store barriers ---------------------------- 3153 #define __ ideal. 3154 3155 void GraphKit::sync_kit(IdealKit& ideal) { 3156 // Final sync IdealKit and graphKit. 3157 __ drain_delay_transform(); 3158 set_all_memory(__ merged_memory()); 3159 set_control(__ ctrl()); 3160 } 3161 3162 // vanilla/CMS post barrier 3163 // Insert a write-barrier store. This is to let generational GC work; we have 3164 // to flag all oop-stores before the next GC point. 3165 void GraphKit::write_barrier_post(Node* oop_store, 3166 Node* obj, 3167 Node* adr, 3168 uint adr_idx, 3169 Node* val, 3170 bool use_precise) { 3171 // No store check needed if we're storing a NULL or an old object 3172 // (latter case is probably a string constant). The concurrent 3173 // mark sweep garbage collector, however, needs to have all nonNull 3174 // oop updates flagged via card-marks. 3175 if (val != NULL && val->is_Con()) { 3176 // must be either an oop or NULL 3177 const Type* t = val->bottom_type(); 3178 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3179 // stores of null never (?) need barriers 3180 return; 3181 ciObject* con = t->is_oopptr()->const_oop(); 3182 if (con != NULL 3183 && con->is_perm() 3184 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 3185 // no store barrier needed, because no old-to-new ref created 3186 return; 3187 } 3188 3189 if (!use_precise) { 3190 // All card marks for a (non-array) instance are in one place: 3191 adr = obj; 3192 } 3193 // (Else it's an array (or unknown), and we want more precise card marks.) 3194 assert(adr != NULL, ""); 3195 3196 IdealKit ideal(gvn(), control(), merged_memory(), true); 3197 3198 // Convert the pointer to an int prior to doing math on it 3199 Node* cast = __ CastPX(__ ctrl(), adr); 3200 3201 // Divide by card size 3202 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3203 "Only one we handle so far."); 3204 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3205 3206 // Combine card table base and card offset 3207 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3208 3209 // Get the alias_index for raw card-mark memory 3210 int adr_type = Compile::AliasIdxRaw; 3211 // Smash zero into card 3212 Node* zero = __ ConI(0); 3213 BasicType bt = T_BYTE; 3214 if( !UseConcMarkSweepGC ) { 3215 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3216 } else { 3217 // Specialized path for CM store barrier 3218 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3219 } 3220 3221 // Final sync IdealKit and GraphKit. 3222 sync_kit(ideal); 3223 } 3224 3225 // G1 pre/post barriers 3226 void GraphKit::g1_write_barrier_pre(Node* obj, 3227 Node* adr, 3228 uint alias_idx, 3229 Node* val, 3230 const TypeOopPtr* val_type, 3231 BasicType bt) { 3232 IdealKit ideal(gvn(), control(), merged_memory(), true); 3233 3234 Node* tls = __ thread(); // ThreadLocalStorage 3235 3236 Node* no_ctrl = NULL; 3237 Node* no_base = __ top(); 3238 Node* zero = __ ConI(0); 3239 3240 float likely = PROB_LIKELY(0.999); 3241 float unlikely = PROB_UNLIKELY(0.999); 3242 3243 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3244 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3245 3246 // Offsets into the thread 3247 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3248 PtrQueue::byte_offset_of_active()); 3249 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3250 PtrQueue::byte_offset_of_index()); 3251 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3252 PtrQueue::byte_offset_of_buf()); 3253 // Now the actual pointers into the thread 3254 3255 // set_control( ctl); 3256 3257 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3258 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3259 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3260 3261 // Now some of the values 3262 3263 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3264 3265 // if (!marking) 3266 __ if_then(marking, BoolTest::ne, zero); { 3267 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3268 3269 const Type* t1 = adr->bottom_type(); 3270 const Type* t2 = val->bottom_type(); 3271 3272 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx); 3273 // if (orig != NULL) 3274 __ if_then(orig, BoolTest::ne, null()); { 3275 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3276 3277 // load original value 3278 // alias_idx correct?? 3279 3280 // is the queue for this thread full? 3281 __ if_then(index, BoolTest::ne, zero, likely); { 3282 3283 // decrement the index 3284 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3285 Node* next_indexX = next_index; 3286 #ifdef _LP64 3287 // We could refine the type for what it's worth 3288 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3289 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3290 #endif 3291 3292 // Now get the buffer location we will log the original value into and store it 3293 Node *log_addr = __ AddP(no_base, buffer, next_indexX); 3294 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw); 3295 3296 // update the index 3297 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3298 3299 } __ else_(); { 3300 3301 // logging buffer is full, call the runtime 3302 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3303 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls); 3304 } __ end_if(); // (!index) 3305 } __ end_if(); // (orig != NULL) 3306 } __ end_if(); // (!marking) 3307 3308 // Final sync IdealKit and GraphKit. 3309 sync_kit(ideal); 3310 } 3311 3312 // 3313 // Update the card table and add card address to the queue 3314 // 3315 void GraphKit::g1_mark_card(IdealKit& ideal, 3316 Node* card_adr, 3317 Node* oop_store, 3318 uint oop_alias_idx, 3319 Node* index, 3320 Node* index_adr, 3321 Node* buffer, 3322 const TypeFunc* tf) { 3323 3324 Node* zero = __ ConI(0); 3325 Node* no_base = __ top(); 3326 BasicType card_bt = T_BYTE; 3327 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3328 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3329 3330 // Now do the queue work 3331 __ if_then(index, BoolTest::ne, zero); { 3332 3333 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3334 Node* next_indexX = next_index; 3335 #ifdef _LP64 3336 // We could refine the type for what it's worth 3337 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3338 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3339 #endif // _LP64 3340 Node* log_addr = __ AddP(no_base, buffer, next_indexX); 3341 3342 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3343 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3344 3345 } __ else_(); { 3346 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3347 } __ end_if(); 3348 3349 } 3350 3351 void GraphKit::g1_write_barrier_post(Node* oop_store, 3352 Node* obj, 3353 Node* adr, 3354 uint alias_idx, 3355 Node* val, 3356 BasicType bt, 3357 bool use_precise) { 3358 // If we are writing a NULL then we need no post barrier 3359 3360 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3361 // Must be NULL 3362 const Type* t = val->bottom_type(); 3363 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3364 // No post barrier if writing NULLx 3365 return; 3366 } 3367 3368 if (!use_precise) { 3369 // All card marks for a (non-array) instance are in one place: 3370 adr = obj; 3371 } 3372 // (Else it's an array (or unknown), and we want more precise card marks.) 3373 assert(adr != NULL, ""); 3374 3375 IdealKit ideal(gvn(), control(), merged_memory(), true); 3376 3377 Node* tls = __ thread(); // ThreadLocalStorage 3378 3379 Node* no_ctrl = NULL; 3380 Node* no_base = __ top(); 3381 float likely = PROB_LIKELY(0.999); 3382 float unlikely = PROB_UNLIKELY(0.999); 3383 Node* zero = __ ConI(0); 3384 Node* zeroX = __ ConX(0); 3385 3386 // Get the alias_index for raw card-mark memory 3387 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3388 3389 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3390 3391 // Offsets into the thread 3392 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3393 PtrQueue::byte_offset_of_index()); 3394 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3395 PtrQueue::byte_offset_of_buf()); 3396 3397 // Pointers into the thread 3398 3399 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3400 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3401 3402 // Now some values 3403 3404 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3405 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3406 3407 3408 // Convert the store obj pointer to an int prior to doing math on it 3409 // Must use ctrl to prevent "integerized oop" existing across safepoint 3410 Node* cast = __ CastPX(__ ctrl(), adr); 3411 3412 // Divide pointer by card size 3413 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3414 3415 // Combine card table base and card offset 3416 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3417 3418 // If we know the value being stored does it cross regions? 3419 3420 if (val != NULL) { 3421 // Does the store cause us to cross regions? 3422 3423 // Should be able to do an unsigned compare of region_size instead of 3424 // and extra shift. Do we have an unsigned compare?? 3425 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3426 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3427 3428 // if (xor_res == 0) same region so skip 3429 __ if_then(xor_res, BoolTest::ne, zeroX); { 3430 3431 // No barrier if we are storing a NULL 3432 __ if_then(val, BoolTest::ne, null(), unlikely); { 3433 3434 // Ok must mark the card if not already dirty 3435 3436 // load the original value of the card 3437 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3438 3439 __ if_then(card_val, BoolTest::ne, zero); { 3440 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3441 } __ end_if(); 3442 } __ end_if(); 3443 } __ end_if(); 3444 } else { 3445 // Object.clone() instrinsic uses this path. 3446 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3447 } 3448 3449 // Final sync IdealKit and GraphKit. 3450 sync_kit(ideal); 3451 } 3452 #undef __ 3453