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 if (CheckExceptionEventsNeeded) { 464 // if the exception capability is set, then we will generate a 465 // runtime call to see if we actually need to report exception 466 // events. If we don't need to report exception events, we will 467 // take the normal fast path provided by add_exception_events. 468 // If exception event reporting is enabled, we will take the 469 // uncommon_trap in the BuildCutout below. 470 471 // call C++ runtime routine to determine whether we need to 472 // report exception events 473 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::must_post_exception_events_Type(), 474 CAST_FROM_FN_PTR(address, JvmtiExport::must_post_exception_events_jrt_leaf), 475 "must_post_exception_events", NULL //no memory effects 476 ); 477 478 Node* need_exc_bool = _gvn.transform(new (C, 1) ProjNode(c, TypeFunc::Parms + 0)); 479 480 // Test the result returned by the runtime call vs. 0 481 Node* zero = _gvn.intcon(0); 482 Node* chk = _gvn.transform( new (C, 3) CmpINode(need_exc_bool, zero) ); 483 BoolTest::mask btesteq = BoolTest::eq; 484 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, btesteq) ); 485 486 // Branch to slow_path if runtime call returned true 487 { BuildCutout unless(this, tst, PROB_MAX); 488 // Do not try anything fancy if we're notifying the VM on every throw. 489 // Cf. case Bytecodes::_athrow in parse2.cpp. 490 uncommon_trap(reason, Deoptimization::Action_none, 491 (ciKlass*)NULL, (char*)NULL, must_throw); 492 } 493 // here if the runtime call returned false 494 // note we do not return here, but continue on with the normal codegen 495 } 496 else { 497 // CheckExceptionEventsNeeded is false 498 // Do not try anything fancy if we're notifying the VM on every throw. 499 // Cf. case Bytecodes::_athrow in parse2.cpp. 500 uncommon_trap(reason, Deoptimization::Action_none, 501 (ciKlass*)NULL, (char*)NULL, must_throw); 502 return; 503 } 504 } 505 506 // If this particular condition has not yet happened at this 507 // bytecode, then use the uncommon trap mechanism, and allow for 508 // a future recompilation if several traps occur here. 509 // If the throw is hot, try to use a more complicated inline mechanism 510 // which keeps execution inside the compiled code. 511 bool treat_throw_as_hot = false; 512 ciMethodData* md = method()->method_data(); 513 514 if (ProfileTraps) { 515 if (too_many_traps(reason)) { 516 treat_throw_as_hot = true; 517 } 518 // (If there is no MDO at all, assume it is early in 519 // execution, and that any deopts are part of the 520 // startup transient, and don't need to be remembered.) 521 522 // Also, if there is a local exception handler, treat all throws 523 // as hot if there has been at least one in this method. 524 if (C->trap_count(reason) != 0 525 && method()->method_data()->trap_count(reason) != 0 526 && has_ex_handler()) { 527 treat_throw_as_hot = true; 528 } 529 } 530 531 // If this throw happens frequently, an uncommon trap might cause 532 // a performance pothole. If there is a local exception handler, 533 // and if this particular bytecode appears to be deoptimizing often, 534 // let us handle the throw inline, with a preconstructed instance. 535 // Note: If the deopt count has blown up, the uncommon trap 536 // runtime is going to flush this nmethod, not matter what. 537 if (treat_throw_as_hot 538 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 539 // If the throw is local, we use a pre-existing instance and 540 // punt on the backtrace. This would lead to a missing backtrace 541 // (a repeat of 4292742) if the backtrace object is ever asked 542 // for its backtrace. 543 // Fixing this remaining case of 4292742 requires some flavor of 544 // escape analysis. Leave that for the future. 545 ciInstance* ex_obj = NULL; 546 switch (reason) { 547 case Deoptimization::Reason_null_check: 548 ex_obj = env()->NullPointerException_instance(); 549 break; 550 case Deoptimization::Reason_div0_check: 551 ex_obj = env()->ArithmeticException_instance(); 552 break; 553 case Deoptimization::Reason_range_check: 554 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 555 break; 556 case Deoptimization::Reason_class_check: 557 if (java_bc() == Bytecodes::_aastore) { 558 ex_obj = env()->ArrayStoreException_instance(); 559 } else { 560 ex_obj = env()->ClassCastException_instance(); 561 } 562 break; 563 } 564 if (failing()) { stop(); return; } // exception allocation might fail 565 if (ex_obj != NULL) { 566 // Cheat with a preallocated exception object. 567 if (C->log() != NULL) 568 C->log()->elem("hot_throw preallocated='1' reason='%s'", 569 Deoptimization::trap_reason_name(reason)); 570 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 571 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 572 573 // Clear the detail message of the preallocated exception object. 574 // Weblogic sometimes mutates the detail message of exceptions 575 // using reflection. 576 int offset = java_lang_Throwable::get_detailMessage_offset(); 577 const TypePtr* adr_typ = ex_con->add_offset(offset); 578 579 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 580 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT); 581 582 add_exception_state(make_exception_state(ex_node)); 583 return; 584 } 585 } 586 587 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 588 // It won't be much cheaper than bailing to the interp., since we'll 589 // have to pass up all the debug-info, and the runtime will have to 590 // create the stack trace. 591 592 // Usual case: Bail to interpreter. 593 // Reserve the right to recompile if we haven't seen anything yet. 594 595 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 596 if (treat_throw_as_hot 597 && (method()->method_data()->trap_recompiled_at(bci()) 598 || C->too_many_traps(reason))) { 599 // We cannot afford to take more traps here. Suffer in the interpreter. 600 if (C->log() != NULL) 601 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 602 Deoptimization::trap_reason_name(reason), 603 C->trap_count(reason)); 604 action = Deoptimization::Action_none; 605 } 606 607 // "must_throw" prunes the JVM state to include only the stack, if there 608 // are no local exception handlers. This should cut down on register 609 // allocation time and code size, by drastically reducing the number 610 // of in-edges on the call to the uncommon trap. 611 612 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 613 } 614 615 616 //----------------------------PreserveJVMState--------------------------------- 617 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 618 debug_only(kit->verify_map()); 619 _kit = kit; 620 _map = kit->map(); // preserve the map 621 _sp = kit->sp(); 622 kit->set_map(clone_map ? kit->clone_map() : NULL); 623 #ifdef ASSERT 624 _bci = kit->bci(); 625 Parse* parser = kit->is_Parse(); 626 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 627 _block = block; 628 #endif 629 } 630 PreserveJVMState::~PreserveJVMState() { 631 GraphKit* kit = _kit; 632 #ifdef ASSERT 633 assert(kit->bci() == _bci, "bci must not shift"); 634 Parse* parser = kit->is_Parse(); 635 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 636 assert(block == _block, "block must not shift"); 637 #endif 638 kit->set_map(_map); 639 kit->set_sp(_sp); 640 } 641 642 643 //-----------------------------BuildCutout------------------------------------- 644 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 645 : PreserveJVMState(kit) 646 { 647 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 648 SafePointNode* outer_map = _map; // preserved map is caller's 649 SafePointNode* inner_map = kit->map(); 650 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 651 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) )); 652 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) )); 653 } 654 BuildCutout::~BuildCutout() { 655 GraphKit* kit = _kit; 656 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 657 } 658 659 //---------------------------PreserveReexecuteState---------------------------- 660 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 661 assert(!kit->stopped(), "must call stopped() before"); 662 _kit = kit; 663 _sp = kit->sp(); 664 _reexecute = kit->jvms()->_reexecute; 665 } 666 PreserveReexecuteState::~PreserveReexecuteState() { 667 if (_kit->stopped()) return; 668 _kit->jvms()->_reexecute = _reexecute; 669 _kit->set_sp(_sp); 670 } 671 672 //------------------------------clone_map-------------------------------------- 673 // Implementation of PreserveJVMState 674 // 675 // Only clone_map(...) here. If this function is only used in the 676 // PreserveJVMState class we may want to get rid of this extra 677 // function eventually and do it all there. 678 679 SafePointNode* GraphKit::clone_map() { 680 if (map() == NULL) return NULL; 681 682 // Clone the memory edge first 683 Node* mem = MergeMemNode::make(C, map()->memory()); 684 gvn().set_type_bottom(mem); 685 686 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 687 JVMState* jvms = this->jvms(); 688 JVMState* clonejvms = jvms->clone_shallow(C); 689 clonemap->set_memory(mem); 690 clonemap->set_jvms(clonejvms); 691 clonejvms->set_map(clonemap); 692 record_for_igvn(clonemap); 693 gvn().set_type_bottom(clonemap); 694 return clonemap; 695 } 696 697 698 //-----------------------------set_map_clone----------------------------------- 699 void GraphKit::set_map_clone(SafePointNode* m) { 700 _map = m; 701 _map = clone_map(); 702 _map->set_next_exception(NULL); 703 debug_only(verify_map()); 704 } 705 706 707 //----------------------------kill_dead_locals--------------------------------- 708 // Detect any locals which are known to be dead, and force them to top. 709 void GraphKit::kill_dead_locals() { 710 // Consult the liveness information for the locals. If any 711 // of them are unused, then they can be replaced by top(). This 712 // should help register allocation time and cut down on the size 713 // of the deoptimization information. 714 715 // This call is made from many of the bytecode handling 716 // subroutines called from the Big Switch in do_one_bytecode. 717 // Every bytecode which might include a slow path is responsible 718 // for killing its dead locals. The more consistent we 719 // are about killing deads, the fewer useless phis will be 720 // constructed for them at various merge points. 721 722 // bci can be -1 (InvocationEntryBci). We return the entry 723 // liveness for the method. 724 725 if (method() == NULL || method()->code_size() == 0) { 726 // We are building a graph for a call to a native method. 727 // All locals are live. 728 return; 729 } 730 731 ResourceMark rm; 732 733 // Consult the liveness information for the locals. If any 734 // of them are unused, then they can be replaced by top(). This 735 // should help register allocation time and cut down on the size 736 // of the deoptimization information. 737 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 738 739 int len = (int)live_locals.size(); 740 assert(len <= jvms()->loc_size(), "too many live locals"); 741 for (int local = 0; local < len; local++) { 742 if (!live_locals.at(local)) { 743 set_local(local, top()); 744 } 745 } 746 } 747 748 #ifdef ASSERT 749 //-------------------------dead_locals_are_killed------------------------------ 750 // Return true if all dead locals are set to top in the map. 751 // Used to assert "clean" debug info at various points. 752 bool GraphKit::dead_locals_are_killed() { 753 if (method() == NULL || method()->code_size() == 0) { 754 // No locals need to be dead, so all is as it should be. 755 return true; 756 } 757 758 // Make sure somebody called kill_dead_locals upstream. 759 ResourceMark rm; 760 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 761 if (jvms->loc_size() == 0) continue; // no locals to consult 762 SafePointNode* map = jvms->map(); 763 ciMethod* method = jvms->method(); 764 int bci = jvms->bci(); 765 if (jvms == this->jvms()) { 766 bci = this->bci(); // it might not yet be synched 767 } 768 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 769 int len = (int)live_locals.size(); 770 if (!live_locals.is_valid() || len == 0) 771 // This method is trivial, or is poisoned by a breakpoint. 772 return true; 773 assert(len == jvms->loc_size(), "live map consistent with locals map"); 774 for (int local = 0; local < len; local++) { 775 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 776 if (PrintMiscellaneous && (Verbose || WizardMode)) { 777 tty->print_cr("Zombie local %d: ", local); 778 jvms->dump(); 779 } 780 return false; 781 } 782 } 783 } 784 return true; 785 } 786 787 #endif //ASSERT 788 789 // Helper function for enforcing certain bytecodes to reexecute if 790 // deoptimization happens 791 static bool should_reexecute_implied_by_bytecode(JVMState *jvms) { 792 ciMethod* cur_method = jvms->method(); 793 int cur_bci = jvms->bci(); 794 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 795 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 796 return Interpreter::bytecode_should_reexecute(code); 797 } else 798 return false; 799 } 800 801 // Helper function for adding JVMState and debug information to node 802 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 803 // Add the safepoint edges to the call (or other safepoint). 804 805 // Make sure dead locals are set to top. This 806 // should help register allocation time and cut down on the size 807 // of the deoptimization information. 808 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 809 810 // Walk the inline list to fill in the correct set of JVMState's 811 // Also fill in the associated edges for each JVMState. 812 813 JVMState* youngest_jvms = sync_jvms(); 814 815 // Do we need debug info here? If it is a SafePoint and this method 816 // cannot de-opt, then we do NOT need any debug info. 817 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint); 818 819 // If we are guaranteed to throw, we can prune everything but the 820 // input to the current bytecode. 821 bool can_prune_locals = false; 822 uint stack_slots_not_pruned = 0; 823 int inputs = 0, depth = 0; 824 if (must_throw) { 825 assert(method() == youngest_jvms->method(), "sanity"); 826 if (compute_stack_effects(inputs, depth)) { 827 can_prune_locals = true; 828 stack_slots_not_pruned = inputs; 829 } 830 } 831 832 if (env()->jvmti_can_examine_or_deopt_anywhere()) { 833 // At any safepoint, this method can get breakpointed, which would 834 // then require an immediate deoptimization. 835 full_info = true; 836 can_prune_locals = false; // do not prune locals 837 stack_slots_not_pruned = 0; 838 } 839 840 // do not scribble on the input jvms 841 JVMState* out_jvms = youngest_jvms->clone_deep(C); 842 call->set_jvms(out_jvms); // Start jvms list for call node 843 844 // For a known set of bytecodes, the interpreter should reexecute them if 845 // deoptimization happens. We set the reexecute state for them here 846 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 847 should_reexecute_implied_by_bytecode(out_jvms)) { 848 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 849 } 850 851 // Presize the call: 852 debug_only(uint non_debug_edges = call->req()); 853 call->add_req_batch(top(), youngest_jvms->debug_depth()); 854 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 855 856 // Set up edges so that the call looks like this: 857 // Call [state:] ctl io mem fptr retadr 858 // [parms:] parm0 ... parmN 859 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 860 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 861 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 862 // Note that caller debug info precedes callee debug info. 863 864 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 865 uint debug_ptr = call->req(); 866 867 // Loop over the map input edges associated with jvms, add them 868 // to the call node, & reset all offsets to match call node array. 869 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 870 uint debug_end = debug_ptr; 871 uint debug_start = debug_ptr - in_jvms->debug_size(); 872 debug_ptr = debug_start; // back up the ptr 873 874 uint p = debug_start; // walks forward in [debug_start, debug_end) 875 uint j, k, l; 876 SafePointNode* in_map = in_jvms->map(); 877 out_jvms->set_map(call); 878 879 if (can_prune_locals) { 880 assert(in_jvms->method() == out_jvms->method(), "sanity"); 881 // If the current throw can reach an exception handler in this JVMS, 882 // then we must keep everything live that can reach that handler. 883 // As a quick and dirty approximation, we look for any handlers at all. 884 if (in_jvms->method()->has_exception_handlers()) { 885 can_prune_locals = false; 886 } 887 } 888 889 // Add the Locals 890 k = in_jvms->locoff(); 891 l = in_jvms->loc_size(); 892 out_jvms->set_locoff(p); 893 if (full_info && !can_prune_locals) { 894 for (j = 0; j < l; j++) 895 call->set_req(p++, in_map->in(k+j)); 896 } else { 897 p += l; // already set to top above by add_req_batch 898 } 899 900 // Add the Expression Stack 901 k = in_jvms->stkoff(); 902 l = in_jvms->sp(); 903 out_jvms->set_stkoff(p); 904 if (full_info && !can_prune_locals) { 905 for (j = 0; j < l; j++) 906 call->set_req(p++, in_map->in(k+j)); 907 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 908 // Divide stack into {S0,...,S1}, where S0 is set to top. 909 uint s1 = stack_slots_not_pruned; 910 stack_slots_not_pruned = 0; // for next iteration 911 if (s1 > l) s1 = l; 912 uint s0 = l - s1; 913 p += s0; // skip the tops preinstalled by add_req_batch 914 for (j = s0; j < l; j++) 915 call->set_req(p++, in_map->in(k+j)); 916 } else { 917 p += l; // already set to top above by add_req_batch 918 } 919 920 // Add the Monitors 921 k = in_jvms->monoff(); 922 l = in_jvms->mon_size(); 923 out_jvms->set_monoff(p); 924 for (j = 0; j < l; j++) 925 call->set_req(p++, in_map->in(k+j)); 926 927 // Copy any scalar object fields. 928 k = in_jvms->scloff(); 929 l = in_jvms->scl_size(); 930 out_jvms->set_scloff(p); 931 for (j = 0; j < l; j++) 932 call->set_req(p++, in_map->in(k+j)); 933 934 // Finish the new jvms. 935 out_jvms->set_endoff(p); 936 937 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 938 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 939 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 940 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 941 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 942 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 943 944 // Update the two tail pointers in parallel. 945 out_jvms = out_jvms->caller(); 946 in_jvms = in_jvms->caller(); 947 } 948 949 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 950 951 // Test the correctness of JVMState::debug_xxx accessors: 952 assert(call->jvms()->debug_start() == non_debug_edges, ""); 953 assert(call->jvms()->debug_end() == call->req(), ""); 954 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 955 } 956 957 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 958 Bytecodes::Code code = java_bc(); 959 if (code == Bytecodes::_wide) { 960 code = method()->java_code_at_bci(bci() + 1); 961 } 962 963 BasicType rtype = T_ILLEGAL; 964 int rsize = 0; 965 966 if (code != Bytecodes::_illegal) { 967 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 968 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 969 if (rtype < T_CONFLICT) 970 rsize = type2size[rtype]; 971 } 972 973 switch (code) { 974 case Bytecodes::_illegal: 975 return false; 976 977 case Bytecodes::_ldc: 978 case Bytecodes::_ldc_w: 979 case Bytecodes::_ldc2_w: 980 inputs = 0; 981 break; 982 983 case Bytecodes::_dup: inputs = 1; break; 984 case Bytecodes::_dup_x1: inputs = 2; break; 985 case Bytecodes::_dup_x2: inputs = 3; break; 986 case Bytecodes::_dup2: inputs = 2; break; 987 case Bytecodes::_dup2_x1: inputs = 3; break; 988 case Bytecodes::_dup2_x2: inputs = 4; break; 989 case Bytecodes::_swap: inputs = 2; break; 990 case Bytecodes::_arraylength: inputs = 1; break; 991 992 case Bytecodes::_getstatic: 993 case Bytecodes::_putstatic: 994 case Bytecodes::_getfield: 995 case Bytecodes::_putfield: 996 { 997 bool is_get = (depth >= 0), is_static = (depth & 1); 998 bool ignore; 999 ciBytecodeStream iter(method()); 1000 iter.reset_to_bci(bci()); 1001 iter.next(); 1002 ciField* field = iter.get_field(ignore); 1003 int size = field->type()->size(); 1004 inputs = (is_static ? 0 : 1); 1005 if (is_get) { 1006 depth = size - inputs; 1007 } else { 1008 inputs += size; // putxxx pops the value from the stack 1009 depth = - inputs; 1010 } 1011 } 1012 break; 1013 1014 case Bytecodes::_invokevirtual: 1015 case Bytecodes::_invokespecial: 1016 case Bytecodes::_invokestatic: 1017 case Bytecodes::_invokedynamic: 1018 case Bytecodes::_invokeinterface: 1019 { 1020 bool is_static = (depth == 0); 1021 bool ignore; 1022 ciBytecodeStream iter(method()); 1023 iter.reset_to_bci(bci()); 1024 iter.next(); 1025 ciMethod* method = iter.get_method(ignore); 1026 inputs = method->arg_size_no_receiver(); 1027 if (!is_static) inputs += 1; 1028 int size = method->return_type()->size(); 1029 depth = size - inputs; 1030 } 1031 break; 1032 1033 case Bytecodes::_multianewarray: 1034 { 1035 ciBytecodeStream iter(method()); 1036 iter.reset_to_bci(bci()); 1037 iter.next(); 1038 inputs = iter.get_dimensions(); 1039 assert(rsize == 1, ""); 1040 depth = rsize - inputs; 1041 } 1042 break; 1043 1044 case Bytecodes::_ireturn: 1045 case Bytecodes::_lreturn: 1046 case Bytecodes::_freturn: 1047 case Bytecodes::_dreturn: 1048 case Bytecodes::_areturn: 1049 assert(rsize = -depth, ""); 1050 inputs = rsize; 1051 break; 1052 1053 case Bytecodes::_jsr: 1054 case Bytecodes::_jsr_w: 1055 inputs = 0; 1056 depth = 1; // S.B. depth=1, not zero 1057 break; 1058 1059 default: 1060 // bytecode produces a typed result 1061 inputs = rsize - depth; 1062 assert(inputs >= 0, ""); 1063 break; 1064 } 1065 1066 #ifdef ASSERT 1067 // spot check 1068 int outputs = depth + inputs; 1069 assert(outputs >= 0, "sanity"); 1070 switch (code) { 1071 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1072 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1073 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1074 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1075 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1076 } 1077 #endif //ASSERT 1078 1079 return true; 1080 } 1081 1082 1083 1084 //------------------------------basic_plus_adr--------------------------------- 1085 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1086 // short-circuit a common case 1087 if (offset == intcon(0)) return ptr; 1088 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); 1089 } 1090 1091 Node* GraphKit::ConvI2L(Node* offset) { 1092 // short-circuit a common case 1093 jint offset_con = find_int_con(offset, Type::OffsetBot); 1094 if (offset_con != Type::OffsetBot) { 1095 return longcon((long) offset_con); 1096 } 1097 return _gvn.transform( new (C, 2) ConvI2LNode(offset)); 1098 } 1099 Node* GraphKit::ConvL2I(Node* offset) { 1100 // short-circuit a common case 1101 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1102 if (offset_con != (jlong)Type::OffsetBot) { 1103 return intcon((int) offset_con); 1104 } 1105 return _gvn.transform( new (C, 2) ConvL2INode(offset)); 1106 } 1107 1108 //-------------------------load_object_klass----------------------------------- 1109 Node* GraphKit::load_object_klass(Node* obj) { 1110 // Special-case a fresh allocation to avoid building nodes: 1111 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1112 if (akls != NULL) return akls; 1113 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1114 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1115 } 1116 1117 //-------------------------load_array_length----------------------------------- 1118 Node* GraphKit::load_array_length(Node* array) { 1119 // Special-case a fresh allocation to avoid building nodes: 1120 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1121 Node *alen; 1122 if (alloc == NULL) { 1123 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1124 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1125 } else { 1126 alen = alloc->Ideal_length(); 1127 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1128 if (ccast != alen) { 1129 alen = _gvn.transform(ccast); 1130 } 1131 } 1132 return alen; 1133 } 1134 1135 //------------------------------do_null_check---------------------------------- 1136 // Helper function to do a NULL pointer check. Returned value is 1137 // the incoming address with NULL casted away. You are allowed to use the 1138 // not-null value only if you are control dependent on the test. 1139 extern int explicit_null_checks_inserted, 1140 explicit_null_checks_elided; 1141 Node* GraphKit::null_check_common(Node* value, BasicType type, 1142 // optional arguments for variations: 1143 bool assert_null, 1144 Node* *null_control) { 1145 assert(!assert_null || null_control == NULL, "not both at once"); 1146 if (stopped()) return top(); 1147 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1148 // For some performance testing, we may wish to suppress null checking. 1149 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1150 return value; 1151 } 1152 explicit_null_checks_inserted++; 1153 1154 // Construct NULL check 1155 Node *chk = NULL; 1156 switch(type) { 1157 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1158 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; 1159 case T_ARRAY : // fall through 1160 type = T_OBJECT; // simplify further tests 1161 case T_OBJECT : { 1162 const Type *t = _gvn.type( value ); 1163 1164 const TypeOopPtr* tp = t->isa_oopptr(); 1165 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1166 // Only for do_null_check, not any of its siblings: 1167 && !assert_null && null_control == NULL) { 1168 // Usually, any field access or invocation on an unloaded oop type 1169 // will simply fail to link, since the statically linked class is 1170 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1171 // the static class is loaded but the sharper oop type is not. 1172 // Rather than checking for this obscure case in lots of places, 1173 // we simply observe that a null check on an unloaded class 1174 // will always be followed by a nonsense operation, so we 1175 // can just issue the uncommon trap here. 1176 // Our access to the unloaded class will only be correct 1177 // after it has been loaded and initialized, which requires 1178 // a trip through the interpreter. 1179 #ifndef PRODUCT 1180 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1181 #endif 1182 uncommon_trap(Deoptimization::Reason_unloaded, 1183 Deoptimization::Action_reinterpret, 1184 tp->klass(), "!loaded"); 1185 return top(); 1186 } 1187 1188 if (assert_null) { 1189 // See if the type is contained in NULL_PTR. 1190 // If so, then the value is already null. 1191 if (t->higher_equal(TypePtr::NULL_PTR)) { 1192 explicit_null_checks_elided++; 1193 return value; // Elided null assert quickly! 1194 } 1195 } else { 1196 // See if mixing in the NULL pointer changes type. 1197 // If so, then the NULL pointer was not allowed in the original 1198 // type. In other words, "value" was not-null. 1199 if (t->meet(TypePtr::NULL_PTR) != t) { 1200 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1201 explicit_null_checks_elided++; 1202 return value; // Elided null check quickly! 1203 } 1204 } 1205 chk = new (C, 3) CmpPNode( value, null() ); 1206 break; 1207 } 1208 1209 default : ShouldNotReachHere(); 1210 } 1211 assert(chk != NULL, "sanity check"); 1212 chk = _gvn.transform(chk); 1213 1214 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1215 BoolNode *btst = new (C, 2) BoolNode( chk, btest); 1216 Node *tst = _gvn.transform( btst ); 1217 1218 //----------- 1219 // if peephole optimizations occurred, a prior test existed. 1220 // If a prior test existed, maybe it dominates as we can avoid this test. 1221 if (tst != btst && type == T_OBJECT) { 1222 // At this point we want to scan up the CFG to see if we can 1223 // find an identical test (and so avoid this test altogether). 1224 Node *cfg = control(); 1225 int depth = 0; 1226 while( depth < 16 ) { // Limit search depth for speed 1227 if( cfg->Opcode() == Op_IfTrue && 1228 cfg->in(0)->in(1) == tst ) { 1229 // Found prior test. Use "cast_not_null" to construct an identical 1230 // CastPP (and hence hash to) as already exists for the prior test. 1231 // Return that casted value. 1232 if (assert_null) { 1233 replace_in_map(value, null()); 1234 return null(); // do not issue the redundant test 1235 } 1236 Node *oldcontrol = control(); 1237 set_control(cfg); 1238 Node *res = cast_not_null(value); 1239 set_control(oldcontrol); 1240 explicit_null_checks_elided++; 1241 return res; 1242 } 1243 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1244 if (cfg == NULL) break; // Quit at region nodes 1245 depth++; 1246 } 1247 } 1248 1249 //----------- 1250 // Branch to failure if null 1251 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1252 Deoptimization::DeoptReason reason; 1253 if (assert_null) 1254 reason = Deoptimization::Reason_null_assert; 1255 else if (type == T_OBJECT) 1256 reason = Deoptimization::Reason_null_check; 1257 else 1258 reason = Deoptimization::Reason_div0_check; 1259 1260 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1261 // ciMethodData::has_trap_at will return a conservative -1 if any 1262 // must-be-null assertion has failed. This could cause performance 1263 // problems for a method after its first do_null_assert failure. 1264 // Consider using 'Reason_class_check' instead? 1265 1266 // To cause an implicit null check, we set the not-null probability 1267 // to the maximum (PROB_MAX). For an explicit check the probability 1268 // is set to a smaller value. 1269 if (null_control != NULL || too_many_traps(reason)) { 1270 // probability is less likely 1271 ok_prob = PROB_LIKELY_MAG(3); 1272 } else if (!assert_null && 1273 (ImplicitNullCheckThreshold > 0) && 1274 method() != NULL && 1275 (method()->method_data()->trap_count(reason) 1276 >= (uint)ImplicitNullCheckThreshold)) { 1277 ok_prob = PROB_LIKELY_MAG(3); 1278 } 1279 1280 if (null_control != NULL) { 1281 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1282 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); 1283 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); 1284 if (null_true == top()) 1285 explicit_null_checks_elided++; 1286 (*null_control) = null_true; 1287 } else { 1288 BuildCutout unless(this, tst, ok_prob); 1289 // Check for optimizer eliding test at parse time 1290 if (stopped()) { 1291 // Failure not possible; do not bother making uncommon trap. 1292 explicit_null_checks_elided++; 1293 } else if (assert_null) { 1294 uncommon_trap(reason, 1295 Deoptimization::Action_make_not_entrant, 1296 NULL, "assert_null"); 1297 } else { 1298 replace_in_map(value, zerocon(type)); 1299 builtin_throw(reason); 1300 } 1301 } 1302 1303 // Must throw exception, fall-thru not possible? 1304 if (stopped()) { 1305 return top(); // No result 1306 } 1307 1308 if (assert_null) { 1309 // Cast obj to null on this path. 1310 replace_in_map(value, zerocon(type)); 1311 return zerocon(type); 1312 } 1313 1314 // Cast obj to not-null on this path, if there is no null_control. 1315 // (If there is a null_control, a non-null value may come back to haunt us.) 1316 if (type == T_OBJECT) { 1317 Node* cast = cast_not_null(value, false); 1318 if (null_control == NULL || (*null_control) == top()) 1319 replace_in_map(value, cast); 1320 value = cast; 1321 } 1322 1323 return value; 1324 } 1325 1326 1327 //------------------------------cast_not_null---------------------------------- 1328 // Cast obj to not-null on this path 1329 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1330 const Type *t = _gvn.type(obj); 1331 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1332 // Object is already not-null? 1333 if( t == t_not_null ) return obj; 1334 1335 Node *cast = new (C, 2) CastPPNode(obj,t_not_null); 1336 cast->init_req(0, control()); 1337 cast = _gvn.transform( cast ); 1338 1339 // Scan for instances of 'obj' in the current JVM mapping. 1340 // These instances are known to be not-null after the test. 1341 if (do_replace_in_map) 1342 replace_in_map(obj, cast); 1343 1344 return cast; // Return casted value 1345 } 1346 1347 1348 //--------------------------replace_in_map------------------------------------- 1349 void GraphKit::replace_in_map(Node* old, Node* neww) { 1350 this->map()->replace_edge(old, neww); 1351 1352 // Note: This operation potentially replaces any edge 1353 // on the map. This includes locals, stack, and monitors 1354 // of the current (innermost) JVM state. 1355 1356 // We can consider replacing in caller maps. 1357 // The idea would be that an inlined function's null checks 1358 // can be shared with the entire inlining tree. 1359 // The expense of doing this is that the PreserveJVMState class 1360 // would have to preserve caller states too, with a deep copy. 1361 } 1362 1363 1364 1365 //============================================================================= 1366 //--------------------------------memory--------------------------------------- 1367 Node* GraphKit::memory(uint alias_idx) { 1368 MergeMemNode* mem = merged_memory(); 1369 Node* p = mem->memory_at(alias_idx); 1370 _gvn.set_type(p, Type::MEMORY); // must be mapped 1371 return p; 1372 } 1373 1374 //-----------------------------reset_memory------------------------------------ 1375 Node* GraphKit::reset_memory() { 1376 Node* mem = map()->memory(); 1377 // do not use this node for any more parsing! 1378 debug_only( map()->set_memory((Node*)NULL) ); 1379 return _gvn.transform( mem ); 1380 } 1381 1382 //------------------------------set_all_memory--------------------------------- 1383 void GraphKit::set_all_memory(Node* newmem) { 1384 Node* mergemem = MergeMemNode::make(C, newmem); 1385 gvn().set_type_bottom(mergemem); 1386 map()->set_memory(mergemem); 1387 } 1388 1389 //------------------------------set_all_memory_call---------------------------- 1390 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1391 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1392 set_all_memory(newmem); 1393 } 1394 1395 //============================================================================= 1396 // 1397 // parser factory methods for MemNodes 1398 // 1399 // These are layered on top of the factory methods in LoadNode and StoreNode, 1400 // and integrate with the parser's memory state and _gvn engine. 1401 // 1402 1403 // factory methods in "int adr_idx" 1404 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1405 int adr_idx, 1406 bool require_atomic_access) { 1407 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1408 const TypePtr* adr_type = NULL; // debug-mode-only argument 1409 debug_only(adr_type = C->get_adr_type(adr_idx)); 1410 Node* mem = memory(adr_idx); 1411 Node* ld; 1412 if (require_atomic_access && bt == T_LONG) { 1413 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); 1414 } else { 1415 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); 1416 } 1417 return _gvn.transform(ld); 1418 } 1419 1420 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1421 int adr_idx, 1422 bool require_atomic_access) { 1423 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1424 const TypePtr* adr_type = NULL; 1425 debug_only(adr_type = C->get_adr_type(adr_idx)); 1426 Node *mem = memory(adr_idx); 1427 Node* st; 1428 if (require_atomic_access && bt == T_LONG) { 1429 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); 1430 } else { 1431 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); 1432 } 1433 st = _gvn.transform(st); 1434 set_memory(st, adr_idx); 1435 // Back-to-back stores can only remove intermediate store with DU info 1436 // so push on worklist for optimizer. 1437 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1438 record_for_igvn(st); 1439 1440 return st; 1441 } 1442 1443 1444 void GraphKit::pre_barrier(Node* ctl, 1445 Node* obj, 1446 Node* adr, 1447 uint adr_idx, 1448 Node* val, 1449 const TypeOopPtr* val_type, 1450 BasicType bt) { 1451 BarrierSet* bs = Universe::heap()->barrier_set(); 1452 set_control(ctl); 1453 switch (bs->kind()) { 1454 case BarrierSet::G1SATBCT: 1455 case BarrierSet::G1SATBCTLogging: 1456 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt); 1457 break; 1458 1459 case BarrierSet::CardTableModRef: 1460 case BarrierSet::CardTableExtension: 1461 case BarrierSet::ModRef: 1462 break; 1463 1464 case BarrierSet::Other: 1465 default : 1466 ShouldNotReachHere(); 1467 1468 } 1469 } 1470 1471 void GraphKit::post_barrier(Node* ctl, 1472 Node* store, 1473 Node* obj, 1474 Node* adr, 1475 uint adr_idx, 1476 Node* val, 1477 BasicType bt, 1478 bool use_precise) { 1479 BarrierSet* bs = Universe::heap()->barrier_set(); 1480 set_control(ctl); 1481 switch (bs->kind()) { 1482 case BarrierSet::G1SATBCT: 1483 case BarrierSet::G1SATBCTLogging: 1484 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1485 break; 1486 1487 case BarrierSet::CardTableModRef: 1488 case BarrierSet::CardTableExtension: 1489 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1490 break; 1491 1492 case BarrierSet::ModRef: 1493 break; 1494 1495 case BarrierSet::Other: 1496 default : 1497 ShouldNotReachHere(); 1498 1499 } 1500 } 1501 1502 Node* GraphKit::store_oop(Node* ctl, 1503 Node* obj, 1504 Node* adr, 1505 const TypePtr* adr_type, 1506 Node* val, 1507 const TypeOopPtr* val_type, 1508 BasicType bt, 1509 bool use_precise) { 1510 1511 set_control(ctl); 1512 if (stopped()) return top(); // Dead path ? 1513 1514 assert(bt == T_OBJECT, "sanity"); 1515 assert(val != NULL, "not dead path"); 1516 uint adr_idx = C->get_alias_index(adr_type); 1517 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1518 1519 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt); 1520 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1521 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1522 return store; 1523 } 1524 1525 // Could be an array or object we don't know at compile time (unsafe ref.) 1526 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1527 Node* obj, // containing obj 1528 Node* adr, // actual adress to store val at 1529 const TypePtr* adr_type, 1530 Node* val, 1531 BasicType bt) { 1532 Compile::AliasType* at = C->alias_type(adr_type); 1533 const TypeOopPtr* val_type = NULL; 1534 if (adr_type->isa_instptr()) { 1535 if (at->field() != NULL) { 1536 // known field. This code is a copy of the do_put_xxx logic. 1537 ciField* field = at->field(); 1538 if (!field->type()->is_loaded()) { 1539 val_type = TypeInstPtr::BOTTOM; 1540 } else { 1541 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1542 } 1543 } 1544 } else if (adr_type->isa_aryptr()) { 1545 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1546 } 1547 if (val_type == NULL) { 1548 val_type = TypeInstPtr::BOTTOM; 1549 } 1550 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); 1551 } 1552 1553 1554 //-------------------------array_element_address------------------------- 1555 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1556 const TypeInt* sizetype) { 1557 uint shift = exact_log2(type2aelembytes(elembt)); 1558 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1559 1560 // short-circuit a common case (saves lots of confusing waste motion) 1561 jint idx_con = find_int_con(idx, -1); 1562 if (idx_con >= 0) { 1563 intptr_t offset = header + ((intptr_t)idx_con << shift); 1564 return basic_plus_adr(ary, offset); 1565 } 1566 1567 // must be correct type for alignment purposes 1568 Node* base = basic_plus_adr(ary, header); 1569 #ifdef _LP64 1570 // The scaled index operand to AddP must be a clean 64-bit value. 1571 // Java allows a 32-bit int to be incremented to a negative 1572 // value, which appears in a 64-bit register as a large 1573 // positive number. Using that large positive number as an 1574 // operand in pointer arithmetic has bad consequences. 1575 // On the other hand, 32-bit overflow is rare, and the possibility 1576 // can often be excluded, if we annotate the ConvI2L node with 1577 // a type assertion that its value is known to be a small positive 1578 // number. (The prior range check has ensured this.) 1579 // This assertion is used by ConvI2LNode::Ideal. 1580 int index_max = max_jint - 1; // array size is max_jint, index is one less 1581 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1582 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1583 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); 1584 #endif 1585 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); 1586 return basic_plus_adr(ary, base, scale); 1587 } 1588 1589 //-------------------------load_array_element------------------------- 1590 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1591 const Type* elemtype = arytype->elem(); 1592 BasicType elembt = elemtype->array_element_basic_type(); 1593 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1594 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1595 return ld; 1596 } 1597 1598 //-------------------------set_arguments_for_java_call------------------------- 1599 // Arguments (pre-popped from the stack) are taken from the JVMS. 1600 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1601 // Add the call arguments: 1602 uint nargs = call->method()->arg_size(); 1603 for (uint i = 0; i < nargs; i++) { 1604 Node* arg = argument(i); 1605 call->init_req(i + TypeFunc::Parms, arg); 1606 } 1607 } 1608 1609 //---------------------------set_edges_for_java_call--------------------------- 1610 // Connect a newly created call into the current JVMS. 1611 // A return value node (if any) is returned from set_edges_for_java_call. 1612 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1613 1614 // Add the predefined inputs: 1615 call->init_req( TypeFunc::Control, control() ); 1616 call->init_req( TypeFunc::I_O , i_o() ); 1617 call->init_req( TypeFunc::Memory , reset_memory() ); 1618 call->init_req( TypeFunc::FramePtr, frameptr() ); 1619 call->init_req( TypeFunc::ReturnAdr, top() ); 1620 1621 add_safepoint_edges(call, must_throw); 1622 1623 Node* xcall = _gvn.transform(call); 1624 1625 if (xcall == top()) { 1626 set_control(top()); 1627 return; 1628 } 1629 assert(xcall == call, "call identity is stable"); 1630 1631 // Re-use the current map to produce the result. 1632 1633 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); 1634 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1635 set_all_memory_call(xcall, separate_io_proj); 1636 1637 //return xcall; // no need, caller already has it 1638 } 1639 1640 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1641 if (stopped()) return top(); // maybe the call folded up? 1642 1643 // Capture the return value, if any. 1644 Node* ret; 1645 if (call->method() == NULL || 1646 call->method()->return_type()->basic_type() == T_VOID) 1647 ret = top(); 1648 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 1649 1650 // Note: Since any out-of-line call can produce an exception, 1651 // we always insert an I_O projection from the call into the result. 1652 1653 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1654 1655 if (separate_io_proj) { 1656 // The caller requested separate projections be used by the fall 1657 // through and exceptional paths, so replace the projections for 1658 // the fall through path. 1659 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) )); 1660 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) )); 1661 } 1662 return ret; 1663 } 1664 1665 //--------------------set_predefined_input_for_runtime_call-------------------- 1666 // Reading and setting the memory state is way conservative here. 1667 // The real problem is that I am not doing real Type analysis on memory, 1668 // so I cannot distinguish card mark stores from other stores. Across a GC 1669 // point the Store Barrier and the card mark memory has to agree. I cannot 1670 // have a card mark store and its barrier split across the GC point from 1671 // either above or below. Here I get that to happen by reading ALL of memory. 1672 // A better answer would be to separate out card marks from other memory. 1673 // For now, return the input memory state, so that it can be reused 1674 // after the call, if this call has restricted memory effects. 1675 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1676 // Set fixed predefined input arguments 1677 Node* memory = reset_memory(); 1678 call->init_req( TypeFunc::Control, control() ); 1679 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1680 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1681 call->init_req( TypeFunc::FramePtr, frameptr() ); 1682 call->init_req( TypeFunc::ReturnAdr, top() ); 1683 return memory; 1684 } 1685 1686 //-------------------set_predefined_output_for_runtime_call-------------------- 1687 // Set control and memory (not i_o) from the call. 1688 // If keep_mem is not NULL, use it for the output state, 1689 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1690 // If hook_mem is NULL, this call produces no memory effects at all. 1691 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1692 // then only that memory slice is taken from the call. 1693 // In the last case, we must put an appropriate memory barrier before 1694 // the call, so as to create the correct anti-dependencies on loads 1695 // preceding the call. 1696 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1697 Node* keep_mem, 1698 const TypePtr* hook_mem) { 1699 // no i/o 1700 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); 1701 if (keep_mem) { 1702 // First clone the existing memory state 1703 set_all_memory(keep_mem); 1704 if (hook_mem != NULL) { 1705 // Make memory for the call 1706 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1707 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1708 // We also use hook_mem to extract specific effects from arraycopy stubs. 1709 set_memory(mem, hook_mem); 1710 } 1711 // ...else the call has NO memory effects. 1712 1713 // Make sure the call advertises its memory effects precisely. 1714 // This lets us build accurate anti-dependences in gcm.cpp. 1715 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1716 "call node must be constructed correctly"); 1717 } else { 1718 assert(hook_mem == NULL, ""); 1719 // This is not a "slow path" call; all memory comes from the call. 1720 set_all_memory_call(call); 1721 } 1722 } 1723 1724 1725 // Replace the call with the current state of the kit. 1726 void GraphKit::replace_call(CallNode* call, Node* result) { 1727 JVMState* ejvms = NULL; 1728 if (has_exceptions()) { 1729 ejvms = transfer_exceptions_into_jvms(); 1730 } 1731 1732 SafePointNode* final_state = stop(); 1733 1734 // Find all the needed outputs of this call 1735 CallProjections callprojs; 1736 call->extract_projections(&callprojs, true); 1737 1738 // Replace all the old call edges with the edges from the inlining result 1739 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control)); 1740 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory)); 1741 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O)); 1742 1743 // Replace the result with the new result if it exists and is used 1744 if (callprojs.resproj != NULL && result != NULL) { 1745 C->gvn_replace_by(callprojs.resproj, result); 1746 } 1747 1748 if (ejvms == NULL) { 1749 // No exception edges to simply kill off those paths 1750 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1751 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1752 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1753 1754 // Replace the old exception object with top 1755 if (callprojs.exobj != NULL) { 1756 C->gvn_replace_by(callprojs.exobj, C->top()); 1757 } 1758 } else { 1759 GraphKit ekit(ejvms); 1760 1761 // Load my combined exception state into the kit, with all phis transformed: 1762 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1763 1764 Node* ex_oop = ekit.use_exception_state(ex_map); 1765 1766 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1767 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1768 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1769 1770 // Replace the old exception object with the newly created one 1771 if (callprojs.exobj != NULL) { 1772 C->gvn_replace_by(callprojs.exobj, ex_oop); 1773 } 1774 } 1775 1776 // Disconnect the call from the graph 1777 call->disconnect_inputs(NULL); 1778 C->gvn_replace_by(call, C->top()); 1779 } 1780 1781 1782 //------------------------------increment_counter------------------------------ 1783 // for statistics: increment a VM counter by 1 1784 1785 void GraphKit::increment_counter(address counter_addr) { 1786 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1787 increment_counter(adr1); 1788 } 1789 1790 void GraphKit::increment_counter(Node* counter_addr) { 1791 int adr_type = Compile::AliasIdxRaw; 1792 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 1793 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 1794 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); 1795 } 1796 1797 1798 //------------------------------uncommon_trap---------------------------------- 1799 // Bail out to the interpreter in mid-method. Implemented by calling the 1800 // uncommon_trap blob. This helper function inserts a runtime call with the 1801 // right debug info. 1802 void GraphKit::uncommon_trap(int trap_request, 1803 ciKlass* klass, const char* comment, 1804 bool must_throw, 1805 bool keep_exact_action) { 1806 if (failing()) stop(); 1807 if (stopped()) return; // trap reachable? 1808 1809 // Note: If ProfileTraps is true, and if a deopt. actually 1810 // occurs here, the runtime will make sure an MDO exists. There is 1811 // no need to call method()->build_method_data() at this point. 1812 1813 #ifdef ASSERT 1814 if (!must_throw) { 1815 // Make sure the stack has at least enough depth to execute 1816 // the current bytecode. 1817 int inputs, ignore; 1818 if (compute_stack_effects(inputs, ignore)) { 1819 assert(sp() >= inputs, "must have enough JVMS stack to execute"); 1820 // It is a frequent error in library_call.cpp to issue an 1821 // uncommon trap with the _sp value already popped. 1822 } 1823 } 1824 #endif 1825 1826 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1827 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1828 1829 switch (action) { 1830 case Deoptimization::Action_maybe_recompile: 1831 case Deoptimization::Action_reinterpret: 1832 // Temporary fix for 6529811 to allow virtual calls to be sure they 1833 // get the chance to go from mono->bi->mega 1834 if (!keep_exact_action && 1835 Deoptimization::trap_request_index(trap_request) < 0 && 1836 too_many_recompiles(reason)) { 1837 // This BCI is causing too many recompilations. 1838 action = Deoptimization::Action_none; 1839 trap_request = Deoptimization::make_trap_request(reason, action); 1840 } else { 1841 C->set_trap_can_recompile(true); 1842 } 1843 break; 1844 case Deoptimization::Action_make_not_entrant: 1845 C->set_trap_can_recompile(true); 1846 break; 1847 #ifdef ASSERT 1848 case Deoptimization::Action_none: 1849 case Deoptimization::Action_make_not_compilable: 1850 break; 1851 default: 1852 assert(false, "bad action"); 1853 #endif 1854 } 1855 1856 if (TraceOptoParse) { 1857 char buf[100]; 1858 tty->print_cr("Uncommon trap %s at bci:%d", 1859 Deoptimization::format_trap_request(buf, sizeof(buf), 1860 trap_request), bci()); 1861 } 1862 1863 CompileLog* log = C->log(); 1864 if (log != NULL) { 1865 int kid = (klass == NULL)? -1: log->identify(klass); 1866 log->begin_elem("uncommon_trap bci='%d'", bci()); 1867 char buf[100]; 1868 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1869 trap_request)); 1870 if (kid >= 0) log->print(" klass='%d'", kid); 1871 if (comment != NULL) log->print(" comment='%s'", comment); 1872 log->end_elem(); 1873 } 1874 1875 // Make sure any guarding test views this path as very unlikely 1876 Node *i0 = control()->in(0); 1877 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1878 IfNode *iff = i0->as_If(); 1879 float f = iff->_prob; // Get prob 1880 if (control()->Opcode() == Op_IfTrue) { 1881 if (f > PROB_UNLIKELY_MAG(4)) 1882 iff->_prob = PROB_MIN; 1883 } else { 1884 if (f < PROB_LIKELY_MAG(4)) 1885 iff->_prob = PROB_MAX; 1886 } 1887 } 1888 1889 // Clear out dead values from the debug info. 1890 kill_dead_locals(); 1891 1892 // Now insert the uncommon trap subroutine call 1893 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); 1894 const TypePtr* no_memory_effects = NULL; 1895 // Pass the index of the class to be loaded 1896 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1897 (must_throw ? RC_MUST_THROW : 0), 1898 OptoRuntime::uncommon_trap_Type(), 1899 call_addr, "uncommon_trap", no_memory_effects, 1900 intcon(trap_request)); 1901 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 1902 "must extract request correctly from the graph"); 1903 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 1904 1905 call->set_req(TypeFunc::ReturnAdr, returnadr()); 1906 // The debug info is the only real input to this call. 1907 1908 // Halt-and-catch fire here. The above call should never return! 1909 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); 1910 _gvn.set_type_bottom(halt); 1911 root()->add_req(halt); 1912 1913 stop_and_kill_map(); 1914 } 1915 1916 1917 //--------------------------just_allocated_object------------------------------ 1918 // Report the object that was just allocated. 1919 // It must be the case that there are no intervening safepoints. 1920 // We use this to determine if an object is so "fresh" that 1921 // it does not require card marks. 1922 Node* GraphKit::just_allocated_object(Node* current_control) { 1923 if (C->recent_alloc_ctl() == current_control) 1924 return C->recent_alloc_obj(); 1925 return NULL; 1926 } 1927 1928 1929 void GraphKit::round_double_arguments(ciMethod* dest_method) { 1930 // (Note: TypeFunc::make has a cache that makes this fast.) 1931 const TypeFunc* tf = TypeFunc::make(dest_method); 1932 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 1933 for (int j = 0; j < nargs; j++) { 1934 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 1935 if( targ->basic_type() == T_DOUBLE ) { 1936 // If any parameters are doubles, they must be rounded before 1937 // the call, dstore_rounding does gvn.transform 1938 Node *arg = argument(j); 1939 arg = dstore_rounding(arg); 1940 set_argument(j, arg); 1941 } 1942 } 1943 } 1944 1945 void GraphKit::round_double_result(ciMethod* dest_method) { 1946 // A non-strict method may return a double value which has an extended 1947 // exponent, but this must not be visible in a caller which is 'strict' 1948 // If a strict caller invokes a non-strict callee, round a double result 1949 1950 BasicType result_type = dest_method->return_type()->basic_type(); 1951 assert( method() != NULL, "must have caller context"); 1952 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 1953 // Destination method's return value is on top of stack 1954 // dstore_rounding() does gvn.transform 1955 Node *result = pop_pair(); 1956 result = dstore_rounding(result); 1957 push_pair(result); 1958 } 1959 } 1960 1961 // rounding for strict float precision conformance 1962 Node* GraphKit::precision_rounding(Node* n) { 1963 return UseStrictFP && _method->flags().is_strict() 1964 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 1965 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) 1966 : n; 1967 } 1968 1969 // rounding for strict double precision conformance 1970 Node* GraphKit::dprecision_rounding(Node *n) { 1971 return UseStrictFP && _method->flags().is_strict() 1972 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 1973 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1974 : n; 1975 } 1976 1977 // rounding for non-strict double stores 1978 Node* GraphKit::dstore_rounding(Node* n) { 1979 return Matcher::strict_fp_requires_explicit_rounding 1980 && UseSSE <= 1 1981 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1982 : n; 1983 } 1984 1985 //============================================================================= 1986 // Generate a fast path/slow path idiom. Graph looks like: 1987 // [foo] indicates that 'foo' is a parameter 1988 // 1989 // [in] NULL 1990 // \ / 1991 // CmpP 1992 // Bool ne 1993 // If 1994 // / \ 1995 // True False-<2> 1996 // / | 1997 // / cast_not_null 1998 // Load | | ^ 1999 // [fast_test] | | 2000 // gvn to opt_test | | 2001 // / \ | <1> 2002 // True False | 2003 // | \\ | 2004 // [slow_call] \[fast_result] 2005 // Ctl Val \ \ 2006 // | \ \ 2007 // Catch <1> \ \ 2008 // / \ ^ \ \ 2009 // Ex No_Ex | \ \ 2010 // | \ \ | \ <2> \ 2011 // ... \ [slow_res] | | \ [null_result] 2012 // \ \--+--+--- | | 2013 // \ | / \ | / 2014 // --------Region Phi 2015 // 2016 //============================================================================= 2017 // Code is structured as a series of driver functions all called 'do_XXX' that 2018 // call a set of helper functions. Helper functions first, then drivers. 2019 2020 //------------------------------null_check_oop--------------------------------- 2021 // Null check oop. Set null-path control into Region in slot 3. 2022 // Make a cast-not-nullness use the other not-null control. Return cast. 2023 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2024 bool never_see_null) { 2025 // Initial NULL check taken path 2026 (*null_control) = top(); 2027 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2028 2029 // Generate uncommon_trap: 2030 if (never_see_null && (*null_control) != top()) { 2031 // If we see an unexpected null at a check-cast we record it and force a 2032 // recompile; the offending check-cast will be compiled to handle NULLs. 2033 // If we see more than one offending BCI, then all checkcasts in the 2034 // method will be compiled to handle NULLs. 2035 PreserveJVMState pjvms(this); 2036 set_control(*null_control); 2037 replace_in_map(value, null()); 2038 uncommon_trap(Deoptimization::Reason_null_check, 2039 Deoptimization::Action_make_not_entrant); 2040 (*null_control) = top(); // NULL path is dead 2041 } 2042 2043 // Cast away null-ness on the result 2044 return cast; 2045 } 2046 2047 //------------------------------opt_iff---------------------------------------- 2048 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2049 // Return slow-path control. 2050 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2051 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2052 2053 // Fast path taken; set region slot 2 2054 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); 2055 region->init_req(2,fast_taken); // Capture fast-control 2056 2057 // Fast path not-taken, i.e. slow path 2058 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); 2059 return slow_taken; 2060 } 2061 2062 //-----------------------------make_runtime_call------------------------------- 2063 Node* GraphKit::make_runtime_call(int flags, 2064 const TypeFunc* call_type, address call_addr, 2065 const char* call_name, 2066 const TypePtr* adr_type, 2067 // The following parms are all optional. 2068 // The first NULL ends the list. 2069 Node* parm0, Node* parm1, 2070 Node* parm2, Node* parm3, 2071 Node* parm4, Node* parm5, 2072 Node* parm6, Node* parm7) { 2073 // Slow-path call 2074 int size = call_type->domain()->cnt(); 2075 bool is_leaf = !(flags & RC_NO_LEAF); 2076 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2077 if (call_name == NULL) { 2078 assert(!is_leaf, "must supply name for leaf"); 2079 call_name = OptoRuntime::stub_name(call_addr); 2080 } 2081 CallNode* call; 2082 if (!is_leaf) { 2083 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, 2084 bci(), adr_type); 2085 } else if (flags & RC_NO_FP) { 2086 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2087 } else { 2088 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); 2089 } 2090 2091 // The following is similar to set_edges_for_java_call, 2092 // except that the memory effects of the call are restricted to AliasIdxRaw. 2093 2094 // Slow path call has no side-effects, uses few values 2095 bool wide_in = !(flags & RC_NARROW_MEM); 2096 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2097 2098 Node* prev_mem = NULL; 2099 if (wide_in) { 2100 prev_mem = set_predefined_input_for_runtime_call(call); 2101 } else { 2102 assert(!wide_out, "narrow in => narrow out"); 2103 Node* narrow_mem = memory(adr_type); 2104 prev_mem = reset_memory(); 2105 map()->set_memory(narrow_mem); 2106 set_predefined_input_for_runtime_call(call); 2107 } 2108 2109 // Hook each parm in order. Stop looking at the first NULL. 2110 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2111 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2112 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2113 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2114 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2115 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2116 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2117 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2118 /* close each nested if ===> */ } } } } } } } } 2119 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2120 2121 if (!is_leaf) { 2122 // Non-leaves can block and take safepoints: 2123 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2124 } 2125 // Non-leaves can throw exceptions: 2126 if (has_io) { 2127 call->set_req(TypeFunc::I_O, i_o()); 2128 } 2129 2130 if (flags & RC_UNCOMMON) { 2131 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2132 // (An "if" probability corresponds roughly to an unconditional count. 2133 // Sort of.) 2134 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2135 } 2136 2137 Node* c = _gvn.transform(call); 2138 assert(c == call, "cannot disappear"); 2139 2140 if (wide_out) { 2141 // Slow path call has full side-effects. 2142 set_predefined_output_for_runtime_call(call); 2143 } else { 2144 // Slow path call has few side-effects, and/or sets few values. 2145 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2146 } 2147 2148 if (has_io) { 2149 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); 2150 } 2151 return call; 2152 2153 } 2154 2155 //------------------------------merge_memory----------------------------------- 2156 // Merge memory from one path into the current memory state. 2157 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2158 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2159 Node* old_slice = mms.force_memory(); 2160 Node* new_slice = mms.memory2(); 2161 if (old_slice != new_slice) { 2162 PhiNode* phi; 2163 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2164 phi = new_slice->as_Phi(); 2165 #ifdef ASSERT 2166 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2167 old_slice = old_slice->in(new_path); 2168 // Caller is responsible for ensuring that any pre-existing 2169 // phis are already aware of old memory. 2170 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2171 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2172 #endif 2173 mms.set_memory(phi); 2174 } else { 2175 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2176 _gvn.set_type(phi, Type::MEMORY); 2177 phi->set_req(new_path, new_slice); 2178 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2179 } 2180 } 2181 } 2182 } 2183 2184 //------------------------------make_slow_call_ex------------------------------ 2185 // Make the exception handler hookups for the slow call 2186 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2187 if (stopped()) return; 2188 2189 // Make a catch node with just two handlers: fall-through and catch-all 2190 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2191 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); 2192 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2193 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2194 2195 { PreserveJVMState pjvms(this); 2196 set_control(excp); 2197 set_i_o(i_o); 2198 2199 if (excp != top()) { 2200 // Create an exception state also. 2201 // Use an exact type if the caller has specified a specific exception. 2202 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2203 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); 2204 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2205 } 2206 } 2207 2208 // Get the no-exception control from the CatchNode. 2209 set_control(norm); 2210 } 2211 2212 2213 //-------------------------------gen_subtype_check----------------------------- 2214 // Generate a subtyping check. Takes as input the subtype and supertype. 2215 // Returns 2 values: sets the default control() to the true path and returns 2216 // the false path. Only reads invariant memory; sets no (visible) memory. 2217 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2218 // but that's not exposed to the optimizer. This call also doesn't take in an 2219 // Object; if you wish to check an Object you need to load the Object's class 2220 // prior to coming here. 2221 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2222 // Fast check for identical types, perhaps identical constants. 2223 // The types can even be identical non-constants, in cases 2224 // involving Array.newInstance, Object.clone, etc. 2225 if (subklass == superklass) 2226 return top(); // false path is dead; no test needed. 2227 2228 if (_gvn.type(superklass)->singleton()) { 2229 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2230 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2231 2232 // In the common case of an exact superklass, try to fold up the 2233 // test before generating code. You may ask, why not just generate 2234 // the code and then let it fold up? The answer is that the generated 2235 // code will necessarily include null checks, which do not always 2236 // completely fold away. If they are also needless, then they turn 2237 // into a performance loss. Example: 2238 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2239 // Here, the type of 'fa' is often exact, so the store check 2240 // of fa[1]=x will fold up, without testing the nullness of x. 2241 switch (static_subtype_check(superk, subk)) { 2242 case SSC_always_false: 2243 { 2244 Node* always_fail = control(); 2245 set_control(top()); 2246 return always_fail; 2247 } 2248 case SSC_always_true: 2249 return top(); 2250 case SSC_easy_test: 2251 { 2252 // Just do a direct pointer compare and be done. 2253 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); 2254 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2255 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2256 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); 2257 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2258 } 2259 case SSC_full_test: 2260 break; 2261 default: 2262 ShouldNotReachHere(); 2263 } 2264 } 2265 2266 // %%% Possible further optimization: Even if the superklass is not exact, 2267 // if the subklass is the unique subtype of the superklass, the check 2268 // will always succeed. We could leave a dependency behind to ensure this. 2269 2270 // First load the super-klass's check-offset 2271 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); 2272 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2273 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); 2274 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2275 2276 // Load from the sub-klass's super-class display list, or a 1-word cache of 2277 // the secondary superclass list, or a failing value with a sentinel offset 2278 // if the super-klass is an interface or exceptionally deep in the Java 2279 // hierarchy and we have to scan the secondary superclass list the hard way. 2280 // Worst-case type is a little odd: NULL is allowed as a result (usually 2281 // klass loads can never produce a NULL). 2282 Node *chk_off_X = ConvI2X(chk_off); 2283 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); 2284 // For some types like interfaces the following loadKlass is from a 1-word 2285 // cache which is mutable so can't use immutable memory. Other 2286 // types load from the super-class display table which is immutable. 2287 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2288 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2289 2290 // Compile speed common case: ARE a subtype and we canNOT fail 2291 if( superklass == nkls ) 2292 return top(); // false path is dead; no test needed. 2293 2294 // See if we get an immediate positive hit. Happens roughly 83% of the 2295 // time. Test to see if the value loaded just previously from the subklass 2296 // is exactly the superklass. 2297 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); 2298 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); 2299 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2300 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); 2301 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); 2302 2303 // Compile speed common case: Check for being deterministic right now. If 2304 // chk_off is a constant and not equal to cacheoff then we are NOT a 2305 // subklass. In this case we need exactly the 1 test above and we can 2306 // return those results immediately. 2307 if (!might_be_cache) { 2308 Node* not_subtype_ctrl = control(); 2309 set_control(iftrue1); // We need exactly the 1 test above 2310 return not_subtype_ctrl; 2311 } 2312 2313 // Gather the various success & failures here 2314 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); 2315 record_for_igvn(r_ok_subtype); 2316 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); 2317 record_for_igvn(r_not_subtype); 2318 2319 r_ok_subtype->init_req(1, iftrue1); 2320 2321 // Check for immediate negative hit. Happens roughly 11% of the time (which 2322 // is roughly 63% of the remaining cases). Test to see if the loaded 2323 // check-offset points into the subklass display list or the 1-element 2324 // cache. If it points to the display (and NOT the cache) and the display 2325 // missed then it's not a subtype. 2326 Node *cacheoff = _gvn.intcon(cacheoff_con); 2327 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); 2328 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); 2329 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2330 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); 2331 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); 2332 2333 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2334 // No performance impact (too rare) but allows sharing of secondary arrays 2335 // which has some footprint reduction. 2336 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); 2337 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); 2338 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2339 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); 2340 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); 2341 2342 // -- Roads not taken here: -- 2343 // We could also have chosen to perform the self-check at the beginning 2344 // of this code sequence, as the assembler does. This would not pay off 2345 // the same way, since the optimizer, unlike the assembler, can perform 2346 // static type analysis to fold away many successful self-checks. 2347 // Non-foldable self checks work better here in second position, because 2348 // the initial primary superclass check subsumes a self-check for most 2349 // types. An exception would be a secondary type like array-of-interface, 2350 // which does not appear in its own primary supertype display. 2351 // Finally, we could have chosen to move the self-check into the 2352 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2353 // dependent manner. But it is worthwhile to have the check here, 2354 // where it can be perhaps be optimized. The cost in code space is 2355 // small (register compare, branch). 2356 2357 // Now do a linear scan of the secondary super-klass array. Again, no real 2358 // performance impact (too rare) but it's gotta be done. 2359 // Since the code is rarely used, there is no penalty for moving it 2360 // out of line, and it can only improve I-cache density. 2361 // The decision to inline or out-of-line this final check is platform 2362 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2363 Node* psc = _gvn.transform( 2364 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2365 2366 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); 2367 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); 2368 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2369 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); 2370 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); 2371 2372 // Return false path; set default control to true path. 2373 set_control( _gvn.transform(r_ok_subtype) ); 2374 return _gvn.transform(r_not_subtype); 2375 } 2376 2377 //----------------------------static_subtype_check----------------------------- 2378 // Shortcut important common cases when superklass is exact: 2379 // (0) superklass is java.lang.Object (can occur in reflective code) 2380 // (1) subklass is already limited to a subtype of superklass => always ok 2381 // (2) subklass does not overlap with superklass => always fail 2382 // (3) superklass has NO subtypes and we can check with a simple compare. 2383 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2384 if (StressReflectiveCode) { 2385 return SSC_full_test; // Let caller generate the general case. 2386 } 2387 2388 if (superk == env()->Object_klass()) { 2389 return SSC_always_true; // (0) this test cannot fail 2390 } 2391 2392 ciType* superelem = superk; 2393 if (superelem->is_array_klass()) 2394 superelem = superelem->as_array_klass()->base_element_type(); 2395 2396 if (!subk->is_interface()) { // cannot trust static interface types yet 2397 if (subk->is_subtype_of(superk)) { 2398 return SSC_always_true; // (1) false path dead; no dynamic test needed 2399 } 2400 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2401 !superk->is_subtype_of(subk)) { 2402 return SSC_always_false; 2403 } 2404 } 2405 2406 // If casting to an instance klass, it must have no subtypes 2407 if (superk->is_interface()) { 2408 // Cannot trust interfaces yet. 2409 // %%% S.B. superk->nof_implementors() == 1 2410 } else if (superelem->is_instance_klass()) { 2411 ciInstanceKlass* ik = superelem->as_instance_klass(); 2412 if (!ik->has_subklass() && !ik->is_interface()) { 2413 if (!ik->is_final()) { 2414 // Add a dependency if there is a chance of a later subclass. 2415 C->dependencies()->assert_leaf_type(ik); 2416 } 2417 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2418 } 2419 } else { 2420 // A primitive array type has no subtypes. 2421 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2422 } 2423 2424 return SSC_full_test; 2425 } 2426 2427 // Profile-driven exact type check: 2428 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2429 float prob, 2430 Node* *casted_receiver) { 2431 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2432 Node* recv_klass = load_object_klass(receiver); 2433 Node* want_klass = makecon(tklass); 2434 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); 2435 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2436 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2437 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); 2438 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2439 2440 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2441 assert(recv_xtype->klass_is_exact(), ""); 2442 2443 // Subsume downstream occurrences of receiver with a cast to 2444 // recv_xtype, since now we know what the type will be. 2445 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); 2446 (*casted_receiver) = _gvn.transform(cast); 2447 // (User must make the replace_in_map call.) 2448 2449 return fail; 2450 } 2451 2452 2453 //-------------------------------gen_instanceof-------------------------------- 2454 // Generate an instance-of idiom. Used by both the instance-of bytecode 2455 // and the reflective instance-of call. 2456 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { 2457 C->set_has_split_ifs(true); // Has chance for split-if optimization 2458 assert( !stopped(), "dead parse path should be checked in callers" ); 2459 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2460 "must check for not-null not-dead klass in callers"); 2461 2462 // Make the merge point 2463 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2464 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2465 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2466 C->set_has_split_ifs(true); // Has chance for split-if optimization 2467 2468 // Null check; get casted pointer; set region slot 3 2469 Node* null_ctl = top(); 2470 Node* not_null_obj = null_check_oop(subobj, &null_ctl); 2471 2472 // If not_null_obj is dead, only null-path is taken 2473 if (stopped()) { // Doing instance-of on a NULL? 2474 set_control(null_ctl); 2475 return intcon(0); 2476 } 2477 region->init_req(_null_path, null_ctl); 2478 phi ->init_req(_null_path, intcon(0)); // Set null path value 2479 2480 // Load the object's klass 2481 Node* obj_klass = load_object_klass(not_null_obj); 2482 2483 // Generate the subtype check 2484 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2485 2486 // Plug in the success path to the general merge in slot 1. 2487 region->init_req(_obj_path, control()); 2488 phi ->init_req(_obj_path, intcon(1)); 2489 2490 // Plug in the failing path to the general merge in slot 2. 2491 region->init_req(_fail_path, not_subtype_ctrl); 2492 phi ->init_req(_fail_path, intcon(0)); 2493 2494 // Return final merged results 2495 set_control( _gvn.transform(region) ); 2496 record_for_igvn(region); 2497 return _gvn.transform(phi); 2498 } 2499 2500 //-------------------------------gen_checkcast--------------------------------- 2501 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2502 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2503 // uncommon-trap paths work. Adjust stack after this call. 2504 // If failure_control is supplied and not null, it is filled in with 2505 // the control edge for the cast failure. Otherwise, an appropriate 2506 // uncommon trap or exception is thrown. 2507 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2508 Node* *failure_control) { 2509 kill_dead_locals(); // Benefit all the uncommon traps 2510 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2511 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2512 2513 // Fast cutout: Check the case that the cast is vacuously true. 2514 // This detects the common cases where the test will short-circuit 2515 // away completely. We do this before we perform the null check, 2516 // because if the test is going to turn into zero code, we don't 2517 // want a residual null check left around. (Causes a slowdown, 2518 // for example, in some objArray manipulations, such as a[i]=a[j].) 2519 if (tk->singleton()) { 2520 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2521 if (objtp != NULL && objtp->klass() != NULL) { 2522 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2523 case SSC_always_true: 2524 return obj; 2525 case SSC_always_false: 2526 // It needs a null check because a null will *pass* the cast check. 2527 // A non-null value will always produce an exception. 2528 return do_null_assert(obj, T_OBJECT); 2529 } 2530 } 2531 } 2532 2533 ciProfileData* data = NULL; 2534 if (failure_control == NULL) { // use MDO in regular case only 2535 assert(java_bc() == Bytecodes::_aastore || 2536 java_bc() == Bytecodes::_checkcast, 2537 "interpreter profiles type checks only for these BCs"); 2538 data = method()->method_data()->bci_to_data(bci()); 2539 } 2540 2541 // Make the merge point 2542 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2543 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2544 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); 2545 C->set_has_split_ifs(true); // Has chance for split-if optimization 2546 2547 // Use null-cast information if it is available 2548 bool never_see_null = false; 2549 // If we see an unexpected null at a check-cast we record it and force a 2550 // recompile; the offending check-cast will be compiled to handle NULLs. 2551 // If we see several offending BCIs, then all checkcasts in the 2552 // method will be compiled to handle NULLs. 2553 if (UncommonNullCast // Cutout for this technique 2554 && failure_control == NULL // regular case 2555 && obj != null() // And not the -Xcomp stupid case? 2556 && !too_many_traps(Deoptimization::Reason_null_check)) { 2557 // Finally, check the "null_seen" bit from the interpreter. 2558 if (data == NULL || !data->as_BitData()->null_seen()) { 2559 never_see_null = true; 2560 } 2561 } 2562 2563 // Null check; get casted pointer; set region slot 3 2564 Node* null_ctl = top(); 2565 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2566 2567 // If not_null_obj is dead, only null-path is taken 2568 if (stopped()) { // Doing instance-of on a NULL? 2569 set_control(null_ctl); 2570 return null(); 2571 } 2572 region->init_req(_null_path, null_ctl); 2573 phi ->init_req(_null_path, null()); // Set null path value 2574 2575 Node* cast_obj = NULL; // the casted version of the object 2576 2577 // If the profile has seen exactly one type, narrow to that type. 2578 // (The subsequent subtype check will always fold up.) 2579 if (UseTypeProfile && TypeProfileCasts && data != NULL && 2580 // Counter has never been decremented (due to cast failure). 2581 // ...This is a reasonable thing to expect. It is true of 2582 // all casts inserted by javac to implement generic types. 2583 data->as_CounterData()->count() >= 0 && 2584 !too_many_traps(Deoptimization::Reason_class_check)) { 2585 // (No, this isn't a call, but it's enough like a virtual call 2586 // to use the same ciMethod accessor to get the profile info...) 2587 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2588 if (profile.count() >= 0 && // no cast failures here 2589 profile.has_receiver(0) && 2590 profile.morphism() == 1) { 2591 ciKlass* exact_kls = profile.receiver(0); 2592 int ssc = static_subtype_check(tk->klass(), exact_kls); 2593 if (ssc == SSC_always_true) { 2594 // If we narrow the type to match what the type profile sees, 2595 // we can then remove the rest of the cast. 2596 // This is a win, even if the exact_kls is very specific, 2597 // because downstream operations, such as method calls, 2598 // will often benefit from the sharper type. 2599 Node* exact_obj = not_null_obj; // will get updated in place... 2600 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2601 &exact_obj); 2602 { PreserveJVMState pjvms(this); 2603 set_control(slow_ctl); 2604 uncommon_trap(Deoptimization::Reason_class_check, 2605 Deoptimization::Action_maybe_recompile); 2606 } 2607 if (failure_control != NULL) // failure is now impossible 2608 (*failure_control) = top(); 2609 replace_in_map(not_null_obj, exact_obj); 2610 // adjust the type of the phi to the exact klass: 2611 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); 2612 cast_obj = exact_obj; 2613 } 2614 // assert(cast_obj != NULL)... except maybe the profile lied to us. 2615 } 2616 } 2617 2618 if (cast_obj == NULL) { 2619 // Load the object's klass 2620 Node* obj_klass = load_object_klass(not_null_obj); 2621 2622 // Generate the subtype check 2623 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2624 2625 // Plug in success path into the merge 2626 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), 2627 not_null_obj, toop)); 2628 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2629 if (failure_control == NULL) { 2630 if (not_subtype_ctrl != top()) { // If failure is possible 2631 PreserveJVMState pjvms(this); 2632 set_control(not_subtype_ctrl); 2633 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2634 } 2635 } else { 2636 (*failure_control) = not_subtype_ctrl; 2637 } 2638 } 2639 2640 region->init_req(_obj_path, control()); 2641 phi ->init_req(_obj_path, cast_obj); 2642 2643 // A merge of NULL or Casted-NotNull obj 2644 Node* res = _gvn.transform(phi); 2645 2646 // Note I do NOT always 'replace_in_map(obj,result)' here. 2647 // if( tk->klass()->can_be_primary_super() ) 2648 // This means that if I successfully store an Object into an array-of-String 2649 // I 'forget' that the Object is really now known to be a String. I have to 2650 // do this because we don't have true union types for interfaces - if I store 2651 // a Baz into an array-of-Interface and then tell the optimizer it's an 2652 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2653 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2654 // replace_in_map( obj, res ); 2655 2656 // Return final merged results 2657 set_control( _gvn.transform(region) ); 2658 record_for_igvn(region); 2659 return res; 2660 } 2661 2662 //------------------------------next_monitor----------------------------------- 2663 // What number should be given to the next monitor? 2664 int GraphKit::next_monitor() { 2665 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2666 int next = current + C->sync_stack_slots(); 2667 // Keep the toplevel high water mark current: 2668 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2669 return current; 2670 } 2671 2672 //------------------------------insert_mem_bar--------------------------------- 2673 // Memory barrier to avoid floating things around 2674 // The membar serves as a pinch point between both control and all memory slices. 2675 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2676 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2677 mb->init_req(TypeFunc::Control, control()); 2678 mb->init_req(TypeFunc::Memory, reset_memory()); 2679 Node* membar = _gvn.transform(mb); 2680 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); 2681 set_all_memory_call(membar); 2682 return membar; 2683 } 2684 2685 //-------------------------insert_mem_bar_volatile---------------------------- 2686 // Memory barrier to avoid floating things around 2687 // The membar serves as a pinch point between both control and memory(alias_idx). 2688 // If you want to make a pinch point on all memory slices, do not use this 2689 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2690 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2691 // When Parse::do_put_xxx updates a volatile field, it appends a series 2692 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2693 // The first membar is on the same memory slice as the field store opcode. 2694 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2695 // All the other membars (for other volatile slices, including AliasIdxBot, 2696 // which stands for all unknown volatile slices) are control-dependent 2697 // on the first membar. This prevents later volatile loads or stores 2698 // from sliding up past the just-emitted store. 2699 2700 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2701 mb->set_req(TypeFunc::Control,control()); 2702 if (alias_idx == Compile::AliasIdxBot) { 2703 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2704 } else { 2705 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2706 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2707 } 2708 Node* membar = _gvn.transform(mb); 2709 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); 2710 if (alias_idx == Compile::AliasIdxBot) { 2711 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); 2712 } else { 2713 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2714 } 2715 return membar; 2716 } 2717 2718 //------------------------------shared_lock------------------------------------ 2719 // Emit locking code. 2720 FastLockNode* GraphKit::shared_lock(Node* obj) { 2721 // bci is either a monitorenter bc or InvocationEntryBci 2722 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2723 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2724 2725 if( !GenerateSynchronizationCode ) 2726 return NULL; // Not locking things? 2727 if (stopped()) // Dead monitor? 2728 return NULL; 2729 2730 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2731 2732 // Box the stack location 2733 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); 2734 Node* mem = reset_memory(); 2735 2736 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); 2737 if (PrintPreciseBiasedLockingStatistics) { 2738 // Create the counters for this fast lock. 2739 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2740 } 2741 // Add monitor to debug info for the slow path. If we block inside the 2742 // slow path and de-opt, we need the monitor hanging around 2743 map()->push_monitor( flock ); 2744 2745 const TypeFunc *tf = LockNode::lock_type(); 2746 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); 2747 2748 lock->init_req( TypeFunc::Control, control() ); 2749 lock->init_req( TypeFunc::Memory , mem ); 2750 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2751 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2752 lock->init_req( TypeFunc::ReturnAdr, top() ); 2753 2754 lock->init_req(TypeFunc::Parms + 0, obj); 2755 lock->init_req(TypeFunc::Parms + 1, box); 2756 lock->init_req(TypeFunc::Parms + 2, flock); 2757 add_safepoint_edges(lock); 2758 2759 lock = _gvn.transform( lock )->as_Lock(); 2760 2761 // lock has no side-effects, sets few values 2762 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2763 2764 insert_mem_bar(Op_MemBarAcquire); 2765 2766 // Add this to the worklist so that the lock can be eliminated 2767 record_for_igvn(lock); 2768 2769 #ifndef PRODUCT 2770 if (PrintLockStatistics) { 2771 // Update the counter for this lock. Don't bother using an atomic 2772 // operation since we don't require absolute accuracy. 2773 lock->create_lock_counter(map()->jvms()); 2774 int adr_type = Compile::AliasIdxRaw; 2775 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); 2776 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 2777 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 2778 store_to_memory(control(), counter_addr, incr, T_INT, adr_type); 2779 } 2780 #endif 2781 2782 return flock; 2783 } 2784 2785 2786 //------------------------------shared_unlock---------------------------------- 2787 // Emit unlocking code. 2788 void GraphKit::shared_unlock(Node* box, Node* obj) { 2789 // bci is either a monitorenter bc or InvocationEntryBci 2790 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2791 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2792 2793 if( !GenerateSynchronizationCode ) 2794 return; 2795 if (stopped()) { // Dead monitor? 2796 map()->pop_monitor(); // Kill monitor from debug info 2797 return; 2798 } 2799 2800 // Memory barrier to avoid floating things down past the locked region 2801 insert_mem_bar(Op_MemBarRelease); 2802 2803 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2804 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); 2805 uint raw_idx = Compile::AliasIdxRaw; 2806 unlock->init_req( TypeFunc::Control, control() ); 2807 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2808 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2809 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2810 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2811 2812 unlock->init_req(TypeFunc::Parms + 0, obj); 2813 unlock->init_req(TypeFunc::Parms + 1, box); 2814 unlock = _gvn.transform(unlock)->as_Unlock(); 2815 2816 Node* mem = reset_memory(); 2817 2818 // unlock has no side-effects, sets few values 2819 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 2820 2821 // Kill monitor from debug info 2822 map()->pop_monitor( ); 2823 } 2824 2825 //-------------------------------get_layout_helper----------------------------- 2826 // If the given klass is a constant or known to be an array, 2827 // fetch the constant layout helper value into constant_value 2828 // and return (Node*)NULL. Otherwise, load the non-constant 2829 // layout helper value, and return the node which represents it. 2830 // This two-faced routine is useful because allocation sites 2831 // almost always feature constant types. 2832 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 2833 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 2834 if (!StressReflectiveCode && inst_klass != NULL) { 2835 ciKlass* klass = inst_klass->klass(); 2836 bool xklass = inst_klass->klass_is_exact(); 2837 if (xklass || klass->is_array_klass()) { 2838 jint lhelper = klass->layout_helper(); 2839 if (lhelper != Klass::_lh_neutral_value) { 2840 constant_value = lhelper; 2841 return (Node*) NULL; 2842 } 2843 } 2844 } 2845 constant_value = Klass::_lh_neutral_value; // put in a known value 2846 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); 2847 return make_load(NULL, lhp, TypeInt::INT, T_INT); 2848 } 2849 2850 // We just put in an allocate/initialize with a big raw-memory effect. 2851 // Hook selected additional alias categories on the initialization. 2852 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 2853 MergeMemNode* init_in_merge, 2854 Node* init_out_raw) { 2855 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 2856 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 2857 2858 Node* prevmem = kit.memory(alias_idx); 2859 init_in_merge->set_memory_at(alias_idx, prevmem); 2860 kit.set_memory(init_out_raw, alias_idx); 2861 } 2862 2863 //---------------------------set_output_for_allocation------------------------- 2864 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 2865 const TypeOopPtr* oop_type, 2866 bool raw_mem_only) { 2867 int rawidx = Compile::AliasIdxRaw; 2868 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 2869 add_safepoint_edges(alloc); 2870 Node* allocx = _gvn.transform(alloc); 2871 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); 2872 // create memory projection for i_o 2873 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 2874 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); 2875 2876 // create a memory projection as for the normal control path 2877 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); 2878 set_memory(malloc, rawidx); 2879 2880 // a normal slow-call doesn't change i_o, but an allocation does 2881 // we create a separate i_o projection for the normal control path 2882 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); 2883 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); 2884 2885 // put in an initialization barrier 2886 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 2887 rawoop)->as_Initialize(); 2888 assert(alloc->initialization() == init, "2-way macro link must work"); 2889 assert(init ->allocation() == alloc, "2-way macro link must work"); 2890 if (ReduceFieldZeroing && !raw_mem_only) { 2891 // Extract memory strands which may participate in the new object's 2892 // initialization, and source them from the new InitializeNode. 2893 // This will allow us to observe initializations when they occur, 2894 // and link them properly (as a group) to the InitializeNode. 2895 assert(init->in(InitializeNode::Memory) == malloc, ""); 2896 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 2897 init->set_req(InitializeNode::Memory, minit_in); 2898 record_for_igvn(minit_in); // fold it up later, if possible 2899 Node* minit_out = memory(rawidx); 2900 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 2901 if (oop_type->isa_aryptr()) { 2902 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 2903 int elemidx = C->get_alias_index(telemref); 2904 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 2905 } else if (oop_type->isa_instptr()) { 2906 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 2907 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 2908 ciField* field = ik->nonstatic_field_at(i); 2909 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 2910 continue; // do not bother to track really large numbers of fields 2911 // Find (or create) the alias category for this field: 2912 int fieldidx = C->alias_type(field)->index(); 2913 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 2914 } 2915 } 2916 } 2917 2918 // Cast raw oop to the real thing... 2919 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); 2920 javaoop = _gvn.transform(javaoop); 2921 C->set_recent_alloc(control(), javaoop); 2922 assert(just_allocated_object(control()) == javaoop, "just allocated"); 2923 2924 #ifdef ASSERT 2925 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 2926 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 2927 "Ideal_allocation works"); 2928 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 2929 "Ideal_allocation works"); 2930 if (alloc->is_AllocateArray()) { 2931 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 2932 "Ideal_allocation works"); 2933 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 2934 "Ideal_allocation works"); 2935 } else { 2936 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 2937 } 2938 } 2939 #endif //ASSERT 2940 2941 return javaoop; 2942 } 2943 2944 //---------------------------new_instance-------------------------------------- 2945 // This routine takes a klass_node which may be constant (for a static type) 2946 // or may be non-constant (for reflective code). It will work equally well 2947 // for either, and the graph will fold nicely if the optimizer later reduces 2948 // the type to a constant. 2949 // The optional arguments are for specialized use by intrinsics: 2950 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 2951 // - If 'raw_mem_only', do not cast the result to an oop. 2952 // - If 'return_size_val', report the the total object size to the caller. 2953 Node* GraphKit::new_instance(Node* klass_node, 2954 Node* extra_slow_test, 2955 bool raw_mem_only, // affect only raw memory 2956 Node* *return_size_val) { 2957 // Compute size in doublewords 2958 // The size is always an integral number of doublewords, represented 2959 // as a positive bytewise size stored in the klass's layout_helper. 2960 // The layout_helper also encodes (in a low bit) the need for a slow path. 2961 jint layout_con = Klass::_lh_neutral_value; 2962 Node* layout_val = get_layout_helper(klass_node, layout_con); 2963 int layout_is_con = (layout_val == NULL); 2964 2965 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 2966 // Generate the initial go-slow test. It's either ALWAYS (return a 2967 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 2968 // case) a computed value derived from the layout_helper. 2969 Node* initial_slow_test = NULL; 2970 if (layout_is_con) { 2971 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2972 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 2973 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 2974 2975 } else { // reflective case 2976 // This reflective path is used by Unsafe.allocateInstance. 2977 // (It may be stress-tested by specifying StressReflectiveCode.) 2978 // Basically, we want to get into the VM is there's an illegal argument. 2979 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 2980 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); 2981 if (extra_slow_test != intcon(0)) { 2982 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); 2983 } 2984 // (Macro-expander will further convert this to a Bool, if necessary.) 2985 } 2986 2987 // Find the size in bytes. This is easy; it's the layout_helper. 2988 // The size value must be valid even if the slow path is taken. 2989 Node* size = NULL; 2990 if (layout_is_con) { 2991 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 2992 } else { // reflective case 2993 // This reflective path is used by clone and Unsafe.allocateInstance. 2994 size = ConvI2X(layout_val); 2995 2996 // Clear the low bits to extract layout_helper_size_in_bytes: 2997 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 2998 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 2999 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); 3000 } 3001 if (return_size_val != NULL) { 3002 (*return_size_val) = size; 3003 } 3004 3005 // This is a precise notnull oop of the klass. 3006 // (Actually, it need not be precise if this is a reflective allocation.) 3007 // It's what we cast the result to. 3008 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3009 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3010 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3011 3012 // Now generate allocation code 3013 3014 // The entire memory state is needed for slow path of the allocation 3015 // since GC and deoptimization can happened. 3016 Node *mem = reset_memory(); 3017 set_all_memory(mem); // Create new memory state 3018 3019 AllocateNode* alloc 3020 = new (C, AllocateNode::ParmLimit) 3021 AllocateNode(C, AllocateNode::alloc_type(), 3022 control(), mem, i_o(), 3023 size, klass_node, 3024 initial_slow_test); 3025 3026 return set_output_for_allocation(alloc, oop_type, raw_mem_only); 3027 } 3028 3029 //-------------------------------new_array------------------------------------- 3030 // helper for both newarray and anewarray 3031 // The 'length' parameter is (obviously) the length of the array. 3032 // See comments on new_instance for the meaning of the other arguments. 3033 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3034 Node* length, // number of array elements 3035 int nargs, // number of arguments to push back for uncommon trap 3036 bool raw_mem_only, // affect only raw memory 3037 Node* *return_size_val) { 3038 jint layout_con = Klass::_lh_neutral_value; 3039 Node* layout_val = get_layout_helper(klass_node, layout_con); 3040 int layout_is_con = (layout_val == NULL); 3041 3042 if (!layout_is_con && !StressReflectiveCode && 3043 !too_many_traps(Deoptimization::Reason_class_check)) { 3044 // This is a reflective array creation site. 3045 // Optimistically assume that it is a subtype of Object[], 3046 // so that we can fold up all the address arithmetic. 3047 layout_con = Klass::array_layout_helper(T_OBJECT); 3048 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); 3049 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); 3050 { BuildCutout unless(this, bol_lh, PROB_MAX); 3051 _sp += nargs; 3052 uncommon_trap(Deoptimization::Reason_class_check, 3053 Deoptimization::Action_maybe_recompile); 3054 } 3055 layout_val = NULL; 3056 layout_is_con = true; 3057 } 3058 3059 // Generate the initial go-slow test. Make sure we do not overflow 3060 // if length is huge (near 2Gig) or negative! We do not need 3061 // exact double-words here, just a close approximation of needed 3062 // double-words. We can't add any offset or rounding bits, lest we 3063 // take a size -1 of bytes and make it positive. Use an unsigned 3064 // compare, so negative sizes look hugely positive. 3065 int fast_size_limit = FastAllocateSizeLimit; 3066 if (layout_is_con) { 3067 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3068 // Increase the size limit if we have exact knowledge of array type. 3069 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3070 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3071 } 3072 3073 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); 3074 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3075 if (initial_slow_test->is_Bool()) { 3076 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3077 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3078 } 3079 3080 // --- Size Computation --- 3081 // array_size = round_to_heap(array_header + (length << elem_shift)); 3082 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3083 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3084 // The rounding mask is strength-reduced, if possible. 3085 int round_mask = MinObjAlignmentInBytes - 1; 3086 Node* header_size = NULL; 3087 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3088 // (T_BYTE has the weakest alignment and size restrictions...) 3089 if (layout_is_con) { 3090 int hsize = Klass::layout_helper_header_size(layout_con); 3091 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3092 BasicType etype = Klass::layout_helper_element_type(layout_con); 3093 if ((round_mask & ~right_n_bits(eshift)) == 0) 3094 round_mask = 0; // strength-reduce it if it goes away completely 3095 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3096 assert(header_size_min <= hsize, "generic minimum is smallest"); 3097 header_size_min = hsize; 3098 header_size = intcon(hsize + round_mask); 3099 } else { 3100 Node* hss = intcon(Klass::_lh_header_size_shift); 3101 Node* hsm = intcon(Klass::_lh_header_size_mask); 3102 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); 3103 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); 3104 Node* mask = intcon(round_mask); 3105 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); 3106 } 3107 3108 Node* elem_shift = NULL; 3109 if (layout_is_con) { 3110 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3111 if (eshift != 0) 3112 elem_shift = intcon(eshift); 3113 } else { 3114 // There is no need to mask or shift this value. 3115 // The semantics of LShiftINode include an implicit mask to 0x1F. 3116 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3117 elem_shift = layout_val; 3118 } 3119 3120 // Transition to native address size for all offset calculations: 3121 Node* lengthx = ConvI2X(length); 3122 Node* headerx = ConvI2X(header_size); 3123 #ifdef _LP64 3124 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3125 if (tllen != NULL && tllen->_lo < 0) { 3126 // Add a manual constraint to a positive range. Cf. array_element_address. 3127 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3128 if (size_max > tllen->_hi) size_max = tllen->_hi; 3129 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3130 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); 3131 } 3132 } 3133 #endif 3134 3135 // Combine header size (plus rounding) and body size. Then round down. 3136 // This computation cannot overflow, because it is used only in two 3137 // places, one where the length is sharply limited, and the other 3138 // after a successful allocation. 3139 Node* abody = lengthx; 3140 if (elem_shift != NULL) 3141 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); 3142 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); 3143 if (round_mask != 0) { 3144 Node* mask = MakeConX(~round_mask); 3145 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); 3146 } 3147 // else if round_mask == 0, the size computation is self-rounding 3148 3149 if (return_size_val != NULL) { 3150 // This is the size 3151 (*return_size_val) = size; 3152 } 3153 3154 // Now generate allocation code 3155 3156 // The entire memory state is needed for slow path of the allocation 3157 // since GC and deoptimization can happened. 3158 Node *mem = reset_memory(); 3159 set_all_memory(mem); // Create new memory state 3160 3161 // Create the AllocateArrayNode and its result projections 3162 AllocateArrayNode* alloc 3163 = new (C, AllocateArrayNode::ParmLimit) 3164 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3165 control(), mem, i_o(), 3166 size, klass_node, 3167 initial_slow_test, 3168 length); 3169 3170 // Cast to correct type. Note that the klass_node may be constant or not, 3171 // and in the latter case the actual array type will be inexact also. 3172 // (This happens via a non-constant argument to inline_native_newArray.) 3173 // In any case, the value of klass_node provides the desired array type. 3174 const TypeInt* length_type = _gvn.find_int_type(length); 3175 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3176 if (ary_type->isa_aryptr() && length_type != NULL) { 3177 // Try to get a better type than POS for the size 3178 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3179 } 3180 3181 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); 3182 3183 // Cast length on remaining path to be as narrow as possible 3184 if (map()->find_edge(length) >= 0) { 3185 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3186 if (ccast != length) { 3187 _gvn.set_type_bottom(ccast); 3188 record_for_igvn(ccast); 3189 replace_in_map(length, ccast); 3190 } 3191 } 3192 3193 return javaoop; 3194 } 3195 3196 // The following "Ideal_foo" functions are placed here because they recognize 3197 // the graph shapes created by the functions immediately above. 3198 3199 //---------------------------Ideal_allocation---------------------------------- 3200 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3201 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3202 if (ptr == NULL) { // reduce dumb test in callers 3203 return NULL; 3204 } 3205 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3206 ptr = ptr->in(1); 3207 if (ptr == NULL) return NULL; 3208 } 3209 if (ptr->is_Proj()) { 3210 Node* allo = ptr->in(0); 3211 if (allo != NULL && allo->is_Allocate()) { 3212 return allo->as_Allocate(); 3213 } 3214 } 3215 // Report failure to match. 3216 return NULL; 3217 } 3218 3219 // Fancy version which also strips off an offset (and reports it to caller). 3220 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3221 intptr_t& offset) { 3222 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3223 if (base == NULL) return NULL; 3224 return Ideal_allocation(base, phase); 3225 } 3226 3227 // Trace Initialize <- Proj[Parm] <- Allocate 3228 AllocateNode* InitializeNode::allocation() { 3229 Node* rawoop = in(InitializeNode::RawAddress); 3230 if (rawoop->is_Proj()) { 3231 Node* alloc = rawoop->in(0); 3232 if (alloc->is_Allocate()) { 3233 return alloc->as_Allocate(); 3234 } 3235 } 3236 return NULL; 3237 } 3238 3239 // Trace Allocate -> Proj[Parm] -> Initialize 3240 InitializeNode* AllocateNode::initialization() { 3241 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3242 if (rawoop == NULL) return NULL; 3243 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3244 Node* init = rawoop->fast_out(i); 3245 if (init->is_Initialize()) { 3246 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3247 return init->as_Initialize(); 3248 } 3249 } 3250 return NULL; 3251 } 3252 3253 //----------------------------- store barriers ---------------------------- 3254 #define __ ideal. 3255 3256 void GraphKit::sync_kit(IdealKit& ideal) { 3257 // Final sync IdealKit and graphKit. 3258 __ drain_delay_transform(); 3259 set_all_memory(__ merged_memory()); 3260 set_control(__ ctrl()); 3261 } 3262 3263 // vanilla/CMS post barrier 3264 // Insert a write-barrier store. This is to let generational GC work; we have 3265 // to flag all oop-stores before the next GC point. 3266 void GraphKit::write_barrier_post(Node* oop_store, 3267 Node* obj, 3268 Node* adr, 3269 uint adr_idx, 3270 Node* val, 3271 bool use_precise) { 3272 // No store check needed if we're storing a NULL or an old object 3273 // (latter case is probably a string constant). The concurrent 3274 // mark sweep garbage collector, however, needs to have all nonNull 3275 // oop updates flagged via card-marks. 3276 if (val != NULL && val->is_Con()) { 3277 // must be either an oop or NULL 3278 const Type* t = val->bottom_type(); 3279 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3280 // stores of null never (?) need barriers 3281 return; 3282 ciObject* con = t->is_oopptr()->const_oop(); 3283 if (con != NULL 3284 && con->is_perm() 3285 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 3286 // no store barrier needed, because no old-to-new ref created 3287 return; 3288 } 3289 3290 if (use_ReduceInitialCardMarks() 3291 && obj == just_allocated_object(control())) { 3292 // We can skip marks on a freshly-allocated object in Eden. 3293 // Keep this code in sync with maybe_defer_card_mark() in runtime.cpp. 3294 // That routine informs GC to take appropriate compensating steps 3295 // so as to make this card-mark elision safe. 3296 return; 3297 } 3298 3299 if (!use_precise) { 3300 // All card marks for a (non-array) instance are in one place: 3301 adr = obj; 3302 } 3303 // (Else it's an array (or unknown), and we want more precise card marks.) 3304 assert(adr != NULL, ""); 3305 3306 IdealKit ideal(gvn(), control(), merged_memory(), true); 3307 3308 // Convert the pointer to an int prior to doing math on it 3309 Node* cast = __ CastPX(__ ctrl(), adr); 3310 3311 // Divide by card size 3312 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3313 "Only one we handle so far."); 3314 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3315 3316 // Combine card table base and card offset 3317 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3318 3319 // Get the alias_index for raw card-mark memory 3320 int adr_type = Compile::AliasIdxRaw; 3321 // Smash zero into card 3322 Node* zero = __ ConI(0); 3323 BasicType bt = T_BYTE; 3324 if( !UseConcMarkSweepGC ) { 3325 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3326 } else { 3327 // Specialized path for CM store barrier 3328 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3329 } 3330 3331 // Final sync IdealKit and GraphKit. 3332 sync_kit(ideal); 3333 } 3334 3335 // G1 pre/post barriers 3336 void GraphKit::g1_write_barrier_pre(Node* obj, 3337 Node* adr, 3338 uint alias_idx, 3339 Node* val, 3340 const TypeOopPtr* val_type, 3341 BasicType bt) { 3342 IdealKit ideal(gvn(), control(), merged_memory(), true); 3343 3344 Node* tls = __ thread(); // ThreadLocalStorage 3345 3346 Node* no_ctrl = NULL; 3347 Node* no_base = __ top(); 3348 Node* zero = __ ConI(0); 3349 3350 float likely = PROB_LIKELY(0.999); 3351 float unlikely = PROB_UNLIKELY(0.999); 3352 3353 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3354 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3355 3356 // Offsets into the thread 3357 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3358 PtrQueue::byte_offset_of_active()); 3359 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3360 PtrQueue::byte_offset_of_index()); 3361 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3362 PtrQueue::byte_offset_of_buf()); 3363 // Now the actual pointers into the thread 3364 3365 // set_control( ctl); 3366 3367 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3368 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3369 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3370 3371 // Now some of the values 3372 3373 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3374 3375 // if (!marking) 3376 __ if_then(marking, BoolTest::ne, zero); { 3377 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3378 3379 const Type* t1 = adr->bottom_type(); 3380 const Type* t2 = val->bottom_type(); 3381 3382 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx); 3383 // if (orig != NULL) 3384 __ if_then(orig, BoolTest::ne, null()); { 3385 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3386 3387 // load original value 3388 // alias_idx correct?? 3389 3390 // is the queue for this thread full? 3391 __ if_then(index, BoolTest::ne, zero, likely); { 3392 3393 // decrement the index 3394 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3395 Node* next_indexX = next_index; 3396 #ifdef _LP64 3397 // We could refine the type for what it's worth 3398 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3399 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3400 #endif 3401 3402 // Now get the buffer location we will log the original value into and store it 3403 Node *log_addr = __ AddP(no_base, buffer, next_indexX); 3404 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw); 3405 3406 // update the index 3407 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3408 3409 } __ else_(); { 3410 3411 // logging buffer is full, call the runtime 3412 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3413 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls); 3414 } __ end_if(); // (!index) 3415 } __ end_if(); // (orig != NULL) 3416 } __ end_if(); // (!marking) 3417 3418 // Final sync IdealKit and GraphKit. 3419 sync_kit(ideal); 3420 } 3421 3422 // 3423 // Update the card table and add card address to the queue 3424 // 3425 void GraphKit::g1_mark_card(IdealKit& ideal, 3426 Node* card_adr, 3427 Node* oop_store, 3428 uint oop_alias_idx, 3429 Node* index, 3430 Node* index_adr, 3431 Node* buffer, 3432 const TypeFunc* tf) { 3433 3434 Node* zero = __ ConI(0); 3435 Node* no_base = __ top(); 3436 BasicType card_bt = T_BYTE; 3437 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3438 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3439 3440 // Now do the queue work 3441 __ if_then(index, BoolTest::ne, zero); { 3442 3443 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3444 Node* next_indexX = next_index; 3445 #ifdef _LP64 3446 // We could refine the type for what it's worth 3447 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3448 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3449 #endif // _LP64 3450 Node* log_addr = __ AddP(no_base, buffer, next_indexX); 3451 3452 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3453 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3454 3455 } __ else_(); { 3456 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3457 } __ end_if(); 3458 3459 } 3460 3461 void GraphKit::g1_write_barrier_post(Node* oop_store, 3462 Node* obj, 3463 Node* adr, 3464 uint alias_idx, 3465 Node* val, 3466 BasicType bt, 3467 bool use_precise) { 3468 // If we are writing a NULL then we need no post barrier 3469 3470 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3471 // Must be NULL 3472 const Type* t = val->bottom_type(); 3473 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3474 // No post barrier if writing NULLx 3475 return; 3476 } 3477 3478 if (!use_precise) { 3479 // All card marks for a (non-array) instance are in one place: 3480 adr = obj; 3481 } 3482 // (Else it's an array (or unknown), and we want more precise card marks.) 3483 assert(adr != NULL, ""); 3484 3485 IdealKit ideal(gvn(), control(), merged_memory(), true); 3486 3487 Node* tls = __ thread(); // ThreadLocalStorage 3488 3489 Node* no_ctrl = NULL; 3490 Node* no_base = __ top(); 3491 float likely = PROB_LIKELY(0.999); 3492 float unlikely = PROB_UNLIKELY(0.999); 3493 Node* zero = __ ConI(0); 3494 Node* zeroX = __ ConX(0); 3495 3496 // Get the alias_index for raw card-mark memory 3497 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3498 3499 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3500 3501 // Offsets into the thread 3502 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3503 PtrQueue::byte_offset_of_index()); 3504 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3505 PtrQueue::byte_offset_of_buf()); 3506 3507 // Pointers into the thread 3508 3509 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3510 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3511 3512 // Now some values 3513 3514 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3515 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3516 3517 3518 // Convert the store obj pointer to an int prior to doing math on it 3519 // Must use ctrl to prevent "integerized oop" existing across safepoint 3520 Node* cast = __ CastPX(__ ctrl(), adr); 3521 3522 // Divide pointer by card size 3523 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3524 3525 // Combine card table base and card offset 3526 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3527 3528 // If we know the value being stored does it cross regions? 3529 3530 if (val != NULL) { 3531 // Does the store cause us to cross regions? 3532 3533 // Should be able to do an unsigned compare of region_size instead of 3534 // and extra shift. Do we have an unsigned compare?? 3535 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3536 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3537 3538 // if (xor_res == 0) same region so skip 3539 __ if_then(xor_res, BoolTest::ne, zeroX); { 3540 3541 // No barrier if we are storing a NULL 3542 __ if_then(val, BoolTest::ne, null(), unlikely); { 3543 3544 // Ok must mark the card if not already dirty 3545 3546 // load the original value of the card 3547 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3548 3549 __ if_then(card_val, BoolTest::ne, zero); { 3550 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3551 } __ end_if(); 3552 } __ end_if(); 3553 } __ end_if(); 3554 } else { 3555 // Object.clone() instrinsic uses this path. 3556 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3557 } 3558 3559 // Final sync IdealKit and GraphKit. 3560 sync_kit(ideal); 3561 } 3562 #undef __