1 /* 2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "ci/ciValueKlass.hpp" 28 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 29 #include "gc/g1/heapRegion.hpp" 30 #include "gc/shared/barrierSet.hpp" 31 #include "gc/shared/cardTableModRefBS.hpp" 32 #include "gc/shared/collectedHeap.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/graphKit.hpp" 38 #include "opto/idealKit.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/locknode.hpp" 41 #include "opto/machnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/valuetypenode.hpp" 47 #include "runtime/deoptimization.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 50 //----------------------------GraphKit----------------------------------------- 51 // Main utility constructor. 52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 53 : Phase(Phase::Parser), 54 _env(C->env()), 55 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()) 56 { 57 _exceptions = jvms->map()->next_exception(); 58 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 59 set_jvms(jvms); 60 #ifdef ASSERT 61 if (_gvn.is_IterGVN() != NULL) { 62 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used"); 63 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 64 _worklist_size = _gvn.C->for_igvn()->size(); 65 } 66 #endif 67 } 68 69 // Private constructor for parser. 70 GraphKit::GraphKit() 71 : Phase(Phase::Parser), 72 _env(C->env()), 73 _gvn(*C->initial_gvn()) 74 { 75 _exceptions = NULL; 76 set_map(NULL); 77 debug_only(_sp = -99); 78 debug_only(set_bci(-99)); 79 } 80 81 82 83 //---------------------------clean_stack--------------------------------------- 84 // Clear away rubbish from the stack area of the JVM state. 85 // This destroys any arguments that may be waiting on the stack. 86 void GraphKit::clean_stack(int from_sp) { 87 SafePointNode* map = this->map(); 88 JVMState* jvms = this->jvms(); 89 int stk_size = jvms->stk_size(); 90 int stkoff = jvms->stkoff(); 91 Node* top = this->top(); 92 for (int i = from_sp; i < stk_size; i++) { 93 if (map->in(stkoff + i) != top) { 94 map->set_req(stkoff + i, top); 95 } 96 } 97 } 98 99 100 //--------------------------------sync_jvms----------------------------------- 101 // Make sure our current jvms agrees with our parse state. 102 JVMState* GraphKit::sync_jvms() const { 103 JVMState* jvms = this->jvms(); 104 jvms->set_bci(bci()); // Record the new bci in the JVMState 105 jvms->set_sp(sp()); // Record the new sp in the JVMState 106 assert(jvms_in_sync(), "jvms is now in sync"); 107 return jvms; 108 } 109 110 //--------------------------------sync_jvms_for_reexecute--------------------- 111 // Make sure our current jvms agrees with our parse state. This version 112 // uses the reexecute_sp for reexecuting bytecodes. 113 JVMState* GraphKit::sync_jvms_for_reexecute() { 114 JVMState* jvms = this->jvms(); 115 jvms->set_bci(bci()); // Record the new bci in the JVMState 116 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 117 return jvms; 118 } 119 120 #ifdef ASSERT 121 bool GraphKit::jvms_in_sync() const { 122 Parse* parse = is_Parse(); 123 if (parse == NULL) { 124 if (bci() != jvms()->bci()) return false; 125 if (sp() != (int)jvms()->sp()) return false; 126 return true; 127 } 128 if (jvms()->method() != parse->method()) return false; 129 if (jvms()->bci() != parse->bci()) return false; 130 int jvms_sp = jvms()->sp(); 131 if (jvms_sp != parse->sp()) return false; 132 int jvms_depth = jvms()->depth(); 133 if (jvms_depth != parse->depth()) return false; 134 return true; 135 } 136 137 // Local helper checks for special internal merge points 138 // used to accumulate and merge exception states. 139 // They are marked by the region's in(0) edge being the map itself. 140 // Such merge points must never "escape" into the parser at large, 141 // until they have been handed to gvn.transform. 142 static bool is_hidden_merge(Node* reg) { 143 if (reg == NULL) return false; 144 if (reg->is_Phi()) { 145 reg = reg->in(0); 146 if (reg == NULL) return false; 147 } 148 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 149 } 150 151 void GraphKit::verify_map() const { 152 if (map() == NULL) return; // null map is OK 153 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 154 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 155 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 156 } 157 158 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 159 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 160 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 161 } 162 #endif 163 164 //---------------------------stop_and_kill_map--------------------------------- 165 // Set _map to NULL, signalling a stop to further bytecode execution. 166 // First smash the current map's control to a constant, to mark it dead. 167 void GraphKit::stop_and_kill_map() { 168 SafePointNode* dead_map = stop(); 169 if (dead_map != NULL) { 170 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 171 assert(dead_map->is_killed(), "must be so marked"); 172 } 173 } 174 175 176 //--------------------------------stopped-------------------------------------- 177 // Tell if _map is NULL, or control is top. 178 bool GraphKit::stopped() { 179 if (map() == NULL) return true; 180 else if (control() == top()) return true; 181 else return false; 182 } 183 184 185 //-----------------------------has_ex_handler---------------------------------- 186 // Tell if this method or any caller method has exception handlers. 187 bool GraphKit::has_ex_handler() { 188 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 189 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 190 return true; 191 } 192 } 193 return false; 194 } 195 196 //------------------------------save_ex_oop------------------------------------ 197 // Save an exception without blowing stack contents or other JVM state. 198 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 199 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 200 ex_map->add_req(ex_oop); 201 debug_only(verify_exception_state(ex_map)); 202 } 203 204 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 205 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 206 Node* ex_oop = ex_map->in(ex_map->req()-1); 207 if (clear_it) ex_map->del_req(ex_map->req()-1); 208 return ex_oop; 209 } 210 211 //-----------------------------saved_ex_oop------------------------------------ 212 // Recover a saved exception from its map. 213 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 214 return common_saved_ex_oop(ex_map, false); 215 } 216 217 //--------------------------clear_saved_ex_oop--------------------------------- 218 // Erase a previously saved exception from its map. 219 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 220 return common_saved_ex_oop(ex_map, true); 221 } 222 223 #ifdef ASSERT 224 //---------------------------has_saved_ex_oop---------------------------------- 225 // Erase a previously saved exception from its map. 226 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 227 return ex_map->req() == ex_map->jvms()->endoff()+1; 228 } 229 #endif 230 231 //-------------------------make_exception_state-------------------------------- 232 // Turn the current JVM state into an exception state, appending the ex_oop. 233 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 234 sync_jvms(); 235 SafePointNode* ex_map = stop(); // do not manipulate this map any more 236 set_saved_ex_oop(ex_map, ex_oop); 237 return ex_map; 238 } 239 240 241 //--------------------------add_exception_state-------------------------------- 242 // Add an exception to my list of exceptions. 243 void GraphKit::add_exception_state(SafePointNode* ex_map) { 244 if (ex_map == NULL || ex_map->control() == top()) { 245 return; 246 } 247 #ifdef ASSERT 248 verify_exception_state(ex_map); 249 if (has_exceptions()) { 250 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 251 } 252 #endif 253 254 // If there is already an exception of exactly this type, merge with it. 255 // In particular, null-checks and other low-level exceptions common up here. 256 Node* ex_oop = saved_ex_oop(ex_map); 257 const Type* ex_type = _gvn.type(ex_oop); 258 if (ex_oop == top()) { 259 // No action needed. 260 return; 261 } 262 assert(ex_type->isa_instptr(), "exception must be an instance"); 263 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 264 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 265 // We check sp also because call bytecodes can generate exceptions 266 // both before and after arguments are popped! 267 if (ex_type2 == ex_type 268 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 269 combine_exception_states(ex_map, e2); 270 return; 271 } 272 } 273 274 // No pre-existing exception of the same type. Chain it on the list. 275 push_exception_state(ex_map); 276 } 277 278 //-----------------------add_exception_states_from----------------------------- 279 void GraphKit::add_exception_states_from(JVMState* jvms) { 280 SafePointNode* ex_map = jvms->map()->next_exception(); 281 if (ex_map != NULL) { 282 jvms->map()->set_next_exception(NULL); 283 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 284 next_map = ex_map->next_exception(); 285 ex_map->set_next_exception(NULL); 286 add_exception_state(ex_map); 287 } 288 } 289 } 290 291 //-----------------------transfer_exceptions_into_jvms------------------------- 292 JVMState* GraphKit::transfer_exceptions_into_jvms() { 293 if (map() == NULL) { 294 // We need a JVMS to carry the exceptions, but the map has gone away. 295 // Create a scratch JVMS, cloned from any of the exception states... 296 if (has_exceptions()) { 297 _map = _exceptions; 298 _map = clone_map(); 299 _map->set_next_exception(NULL); 300 clear_saved_ex_oop(_map); 301 debug_only(verify_map()); 302 } else { 303 // ...or created from scratch 304 JVMState* jvms = new (C) JVMState(_method, NULL); 305 jvms->set_bci(_bci); 306 jvms->set_sp(_sp); 307 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 308 set_jvms(jvms); 309 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 310 set_all_memory(top()); 311 while (map()->req() < jvms->endoff()) map()->add_req(top()); 312 } 313 // (This is a kludge, in case you didn't notice.) 314 set_control(top()); 315 } 316 JVMState* jvms = sync_jvms(); 317 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 318 jvms->map()->set_next_exception(_exceptions); 319 _exceptions = NULL; // done with this set of exceptions 320 return jvms; 321 } 322 323 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 324 assert(is_hidden_merge(dstphi), "must be a special merge node"); 325 assert(is_hidden_merge(srcphi), "must be a special merge node"); 326 uint limit = srcphi->req(); 327 for (uint i = PhiNode::Input; i < limit; i++) { 328 dstphi->add_req(srcphi->in(i)); 329 } 330 } 331 static inline void add_one_req(Node* dstphi, Node* src) { 332 assert(is_hidden_merge(dstphi), "must be a special merge node"); 333 assert(!is_hidden_merge(src), "must not be a special merge node"); 334 dstphi->add_req(src); 335 } 336 337 //-----------------------combine_exception_states------------------------------ 338 // This helper function combines exception states by building phis on a 339 // specially marked state-merging region. These regions and phis are 340 // untransformed, and can build up gradually. The region is marked by 341 // having a control input of its exception map, rather than NULL. Such 342 // regions do not appear except in this function, and in use_exception_state. 343 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 344 if (failing()) return; // dying anyway... 345 JVMState* ex_jvms = ex_map->_jvms; 346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 351 assert(ex_map->req() == phi_map->req(), "matching maps"); 352 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 353 Node* hidden_merge_mark = root(); 354 Node* region = phi_map->control(); 355 MergeMemNode* phi_mem = phi_map->merged_memory(); 356 MergeMemNode* ex_mem = ex_map->merged_memory(); 357 if (region->in(0) != hidden_merge_mark) { 358 // The control input is not (yet) a specially-marked region in phi_map. 359 // Make it so, and build some phis. 360 region = new RegionNode(2); 361 _gvn.set_type(region, Type::CONTROL); 362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 363 region->init_req(1, phi_map->control()); 364 phi_map->set_control(region); 365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 366 record_for_igvn(io_phi); 367 _gvn.set_type(io_phi, Type::ABIO); 368 phi_map->set_i_o(io_phi); 369 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 370 Node* m = mms.memory(); 371 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 372 record_for_igvn(m_phi); 373 _gvn.set_type(m_phi, Type::MEMORY); 374 mms.set_memory(m_phi); 375 } 376 } 377 378 // Either or both of phi_map and ex_map might already be converted into phis. 379 Node* ex_control = ex_map->control(); 380 // if there is special marking on ex_map also, we add multiple edges from src 381 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 382 // how wide was the destination phi_map, originally? 383 uint orig_width = region->req(); 384 385 if (add_multiple) { 386 add_n_reqs(region, ex_control); 387 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 388 } else { 389 // ex_map has no merges, so we just add single edges everywhere 390 add_one_req(region, ex_control); 391 add_one_req(phi_map->i_o(), ex_map->i_o()); 392 } 393 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 394 if (mms.is_empty()) { 395 // get a copy of the base memory, and patch some inputs into it 396 const TypePtr* adr_type = mms.adr_type(C); 397 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 398 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 399 mms.set_memory(phi); 400 // Prepare to append interesting stuff onto the newly sliced phi: 401 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 402 } 403 // Append stuff from ex_map: 404 if (add_multiple) { 405 add_n_reqs(mms.memory(), mms.memory2()); 406 } else { 407 add_one_req(mms.memory(), mms.memory2()); 408 } 409 } 410 uint limit = ex_map->req(); 411 for (uint i = TypeFunc::Parms; i < limit; i++) { 412 // Skip everything in the JVMS after tos. (The ex_oop follows.) 413 if (i == tos) i = ex_jvms->monoff(); 414 Node* src = ex_map->in(i); 415 Node* dst = phi_map->in(i); 416 if (src != dst) { 417 PhiNode* phi; 418 if (dst->in(0) != region) { 419 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 420 record_for_igvn(phi); 421 _gvn.set_type(phi, phi->type()); 422 phi_map->set_req(i, dst); 423 // Prepare to append interesting stuff onto the new phi: 424 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 425 } else { 426 assert(dst->is_Phi(), "nobody else uses a hidden region"); 427 phi = dst->as_Phi(); 428 } 429 if (add_multiple && src->in(0) == ex_control) { 430 // Both are phis. 431 add_n_reqs(dst, src); 432 } else { 433 while (dst->req() < region->req()) add_one_req(dst, src); 434 } 435 const Type* srctype = _gvn.type(src); 436 if (phi->type() != srctype) { 437 const Type* dsttype = phi->type()->meet_speculative(srctype); 438 if (phi->type() != dsttype) { 439 phi->set_type(dsttype); 440 _gvn.set_type(phi, dsttype); 441 } 442 } 443 } 444 } 445 phi_map->merge_replaced_nodes_with(ex_map); 446 } 447 448 //--------------------------use_exception_state-------------------------------- 449 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 450 if (failing()) { stop(); return top(); } 451 Node* region = phi_map->control(); 452 Node* hidden_merge_mark = root(); 453 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 454 Node* ex_oop = clear_saved_ex_oop(phi_map); 455 if (region->in(0) == hidden_merge_mark) { 456 // Special marking for internal ex-states. Process the phis now. 457 region->set_req(0, region); // now it's an ordinary region 458 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 459 // Note: Setting the jvms also sets the bci and sp. 460 set_control(_gvn.transform(region)); 461 uint tos = jvms()->stkoff() + sp(); 462 for (uint i = 1; i < tos; i++) { 463 Node* x = phi_map->in(i); 464 if (x->in(0) == region) { 465 assert(x->is_Phi(), "expected a special phi"); 466 phi_map->set_req(i, _gvn.transform(x)); 467 } 468 } 469 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 470 Node* x = mms.memory(); 471 if (x->in(0) == region) { 472 assert(x->is_Phi(), "nobody else uses a hidden region"); 473 mms.set_memory(_gvn.transform(x)); 474 } 475 } 476 if (ex_oop->in(0) == region) { 477 assert(ex_oop->is_Phi(), "expected a special phi"); 478 ex_oop = _gvn.transform(ex_oop); 479 } 480 } else { 481 set_jvms(phi_map->jvms()); 482 } 483 484 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 485 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 486 return ex_oop; 487 } 488 489 //---------------------------------java_bc------------------------------------- 490 Bytecodes::Code GraphKit::java_bc() const { 491 ciMethod* method = this->method(); 492 int bci = this->bci(); 493 if (method != NULL && bci != InvocationEntryBci) 494 return method->java_code_at_bci(bci); 495 else 496 return Bytecodes::_illegal; 497 } 498 499 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 500 bool must_throw) { 501 // if the exception capability is set, then we will generate code 502 // to check the JavaThread.should_post_on_exceptions flag to see 503 // if we actually need to report exception events (for this 504 // thread). If we don't need to report exception events, we will 505 // take the normal fast path provided by add_exception_events. If 506 // exception event reporting is enabled for this thread, we will 507 // take the uncommon_trap in the BuildCutout below. 508 509 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 510 Node* jthread = _gvn.transform(new ThreadLocalNode()); 511 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 512 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 513 514 // Test the should_post_on_exceptions_flag vs. 0 515 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 516 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 517 518 // Branch to slow_path if should_post_on_exceptions_flag was true 519 { BuildCutout unless(this, tst, PROB_MAX); 520 // Do not try anything fancy if we're notifying the VM on every throw. 521 // Cf. case Bytecodes::_athrow in parse2.cpp. 522 uncommon_trap(reason, Deoptimization::Action_none, 523 (ciKlass*)NULL, (char*)NULL, must_throw); 524 } 525 526 } 527 528 //------------------------------builtin_throw---------------------------------- 529 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 530 bool must_throw = true; 531 532 if (env()->jvmti_can_post_on_exceptions()) { 533 // check if we must post exception events, take uncommon trap if so 534 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 535 // here if should_post_on_exceptions is false 536 // continue on with the normal codegen 537 } 538 539 // If this particular condition has not yet happened at this 540 // bytecode, then use the uncommon trap mechanism, and allow for 541 // a future recompilation if several traps occur here. 542 // If the throw is hot, try to use a more complicated inline mechanism 543 // which keeps execution inside the compiled code. 544 bool treat_throw_as_hot = false; 545 ciMethodData* md = method()->method_data(); 546 547 if (ProfileTraps) { 548 if (too_many_traps(reason)) { 549 treat_throw_as_hot = true; 550 } 551 // (If there is no MDO at all, assume it is early in 552 // execution, and that any deopts are part of the 553 // startup transient, and don't need to be remembered.) 554 555 // Also, if there is a local exception handler, treat all throws 556 // as hot if there has been at least one in this method. 557 if (C->trap_count(reason) != 0 558 && method()->method_data()->trap_count(reason) != 0 559 && has_ex_handler()) { 560 treat_throw_as_hot = true; 561 } 562 } 563 564 // If this throw happens frequently, an uncommon trap might cause 565 // a performance pothole. If there is a local exception handler, 566 // and if this particular bytecode appears to be deoptimizing often, 567 // let us handle the throw inline, with a preconstructed instance. 568 // Note: If the deopt count has blown up, the uncommon trap 569 // runtime is going to flush this nmethod, not matter what. 570 if (treat_throw_as_hot 571 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 572 // If the throw is local, we use a pre-existing instance and 573 // punt on the backtrace. This would lead to a missing backtrace 574 // (a repeat of 4292742) if the backtrace object is ever asked 575 // for its backtrace. 576 // Fixing this remaining case of 4292742 requires some flavor of 577 // escape analysis. Leave that for the future. 578 ciInstance* ex_obj = NULL; 579 switch (reason) { 580 case Deoptimization::Reason_null_check: 581 ex_obj = env()->NullPointerException_instance(); 582 break; 583 case Deoptimization::Reason_div0_check: 584 ex_obj = env()->ArithmeticException_instance(); 585 break; 586 case Deoptimization::Reason_range_check: 587 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 588 break; 589 case Deoptimization::Reason_class_check: 590 if (java_bc() == Bytecodes::_aastore || java_bc() == Bytecodes::_vastore) { 591 ex_obj = env()->ArrayStoreException_instance(); 592 } else { 593 ex_obj = env()->ClassCastException_instance(); 594 } 595 break; 596 default: 597 break; 598 } 599 if (failing()) { stop(); return; } // exception allocation might fail 600 if (ex_obj != NULL) { 601 // Cheat with a preallocated exception object. 602 if (C->log() != NULL) 603 C->log()->elem("hot_throw preallocated='1' reason='%s'", 604 Deoptimization::trap_reason_name(reason)); 605 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 606 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 607 608 // Clear the detail message of the preallocated exception object. 609 // Weblogic sometimes mutates the detail message of exceptions 610 // using reflection. 611 int offset = java_lang_Throwable::get_detailMessage_offset(); 612 const TypePtr* adr_typ = ex_con->add_offset(offset); 613 614 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 615 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 616 // Conservatively release stores of object references. 617 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 618 619 add_exception_state(make_exception_state(ex_node)); 620 return; 621 } 622 } 623 624 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 625 // It won't be much cheaper than bailing to the interp., since we'll 626 // have to pass up all the debug-info, and the runtime will have to 627 // create the stack trace. 628 629 // Usual case: Bail to interpreter. 630 // Reserve the right to recompile if we haven't seen anything yet. 631 632 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 633 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 634 if (treat_throw_as_hot 635 && (method()->method_data()->trap_recompiled_at(bci(), m) 636 || C->too_many_traps(reason))) { 637 // We cannot afford to take more traps here. Suffer in the interpreter. 638 if (C->log() != NULL) 639 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 640 Deoptimization::trap_reason_name(reason), 641 C->trap_count(reason)); 642 action = Deoptimization::Action_none; 643 } 644 645 // "must_throw" prunes the JVM state to include only the stack, if there 646 // are no local exception handlers. This should cut down on register 647 // allocation time and code size, by drastically reducing the number 648 // of in-edges on the call to the uncommon trap. 649 650 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 651 } 652 653 654 //----------------------------PreserveJVMState--------------------------------- 655 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 656 debug_only(kit->verify_map()); 657 _kit = kit; 658 _map = kit->map(); // preserve the map 659 _sp = kit->sp(); 660 kit->set_map(clone_map ? kit->clone_map() : NULL); 661 #ifdef ASSERT 662 _bci = kit->bci(); 663 Parse* parser = kit->is_Parse(); 664 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 665 _block = block; 666 #endif 667 } 668 PreserveJVMState::~PreserveJVMState() { 669 GraphKit* kit = _kit; 670 #ifdef ASSERT 671 assert(kit->bci() == _bci, "bci must not shift"); 672 Parse* parser = kit->is_Parse(); 673 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 674 assert(block == _block, "block must not shift"); 675 #endif 676 kit->set_map(_map); 677 kit->set_sp(_sp); 678 } 679 680 681 //-----------------------------BuildCutout------------------------------------- 682 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 683 : PreserveJVMState(kit) 684 { 685 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 686 SafePointNode* outer_map = _map; // preserved map is caller's 687 SafePointNode* inner_map = kit->map(); 688 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 689 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 690 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 691 } 692 BuildCutout::~BuildCutout() { 693 GraphKit* kit = _kit; 694 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 695 } 696 697 //---------------------------PreserveReexecuteState---------------------------- 698 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 699 assert(!kit->stopped(), "must call stopped() before"); 700 _kit = kit; 701 _sp = kit->sp(); 702 _reexecute = kit->jvms()->_reexecute; 703 } 704 PreserveReexecuteState::~PreserveReexecuteState() { 705 if (_kit->stopped()) return; 706 _kit->jvms()->_reexecute = _reexecute; 707 _kit->set_sp(_sp); 708 } 709 710 //------------------------------clone_map-------------------------------------- 711 // Implementation of PreserveJVMState 712 // 713 // Only clone_map(...) here. If this function is only used in the 714 // PreserveJVMState class we may want to get rid of this extra 715 // function eventually and do it all there. 716 717 SafePointNode* GraphKit::clone_map() { 718 if (map() == NULL) return NULL; 719 720 // Clone the memory edge first 721 Node* mem = MergeMemNode::make(map()->memory()); 722 gvn().set_type_bottom(mem); 723 724 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 725 JVMState* jvms = this->jvms(); 726 JVMState* clonejvms = jvms->clone_shallow(C); 727 clonemap->set_memory(mem); 728 clonemap->set_jvms(clonejvms); 729 clonejvms->set_map(clonemap); 730 record_for_igvn(clonemap); 731 gvn().set_type_bottom(clonemap); 732 return clonemap; 733 } 734 735 736 //-----------------------------set_map_clone----------------------------------- 737 void GraphKit::set_map_clone(SafePointNode* m) { 738 _map = m; 739 _map = clone_map(); 740 _map->set_next_exception(NULL); 741 debug_only(verify_map()); 742 } 743 744 745 //----------------------------kill_dead_locals--------------------------------- 746 // Detect any locals which are known to be dead, and force them to top. 747 void GraphKit::kill_dead_locals() { 748 // Consult the liveness information for the locals. If any 749 // of them are unused, then they can be replaced by top(). This 750 // should help register allocation time and cut down on the size 751 // of the deoptimization information. 752 753 // This call is made from many of the bytecode handling 754 // subroutines called from the Big Switch in do_one_bytecode. 755 // Every bytecode which might include a slow path is responsible 756 // for killing its dead locals. The more consistent we 757 // are about killing deads, the fewer useless phis will be 758 // constructed for them at various merge points. 759 760 // bci can be -1 (InvocationEntryBci). We return the entry 761 // liveness for the method. 762 763 if (method() == NULL || method()->code_size() == 0) { 764 // We are building a graph for a call to a native method. 765 // All locals are live. 766 return; 767 } 768 769 ResourceMark rm; 770 771 // Consult the liveness information for the locals. If any 772 // of them are unused, then they can be replaced by top(). This 773 // should help register allocation time and cut down on the size 774 // of the deoptimization information. 775 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 776 777 int len = (int)live_locals.size(); 778 assert(len <= jvms()->loc_size(), "too many live locals"); 779 for (int local = 0; local < len; local++) { 780 if (!live_locals.at(local)) { 781 set_local(local, top()); 782 } 783 } 784 } 785 786 #ifdef ASSERT 787 //-------------------------dead_locals_are_killed------------------------------ 788 // Return true if all dead locals are set to top in the map. 789 // Used to assert "clean" debug info at various points. 790 bool GraphKit::dead_locals_are_killed() { 791 if (method() == NULL || method()->code_size() == 0) { 792 // No locals need to be dead, so all is as it should be. 793 return true; 794 } 795 796 // Make sure somebody called kill_dead_locals upstream. 797 ResourceMark rm; 798 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 799 if (jvms->loc_size() == 0) continue; // no locals to consult 800 SafePointNode* map = jvms->map(); 801 ciMethod* method = jvms->method(); 802 int bci = jvms->bci(); 803 if (jvms == this->jvms()) { 804 bci = this->bci(); // it might not yet be synched 805 } 806 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 807 int len = (int)live_locals.size(); 808 if (!live_locals.is_valid() || len == 0) 809 // This method is trivial, or is poisoned by a breakpoint. 810 return true; 811 assert(len == jvms->loc_size(), "live map consistent with locals map"); 812 for (int local = 0; local < len; local++) { 813 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 814 if (PrintMiscellaneous && (Verbose || WizardMode)) { 815 tty->print_cr("Zombie local %d: ", local); 816 jvms->dump(); 817 } 818 return false; 819 } 820 } 821 } 822 return true; 823 } 824 825 #endif //ASSERT 826 827 // Helper function for adding JVMState and debug information to node 828 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 829 // Add the safepoint edges to the call (or other safepoint). 830 831 // Make sure dead locals are set to top. This 832 // should help register allocation time and cut down on the size 833 // of the deoptimization information. 834 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 835 836 // Walk the inline list to fill in the correct set of JVMState's 837 // Also fill in the associated edges for each JVMState. 838 839 // If the bytecode needs to be reexecuted we need to put 840 // the arguments back on the stack. 841 const bool should_reexecute = jvms()->should_reexecute(); 842 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 843 844 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 845 // undefined if the bci is different. This is normal for Parse but it 846 // should not happen for LibraryCallKit because only one bci is processed. 847 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 848 "in LibraryCallKit the reexecute bit should not change"); 849 850 // If we are guaranteed to throw, we can prune everything but the 851 // input to the current bytecode. 852 bool can_prune_locals = false; 853 uint stack_slots_not_pruned = 0; 854 int inputs = 0, depth = 0; 855 if (must_throw) { 856 assert(method() == youngest_jvms->method(), "sanity"); 857 if (compute_stack_effects(inputs, depth)) { 858 can_prune_locals = true; 859 stack_slots_not_pruned = inputs; 860 } 861 } 862 863 if (env()->should_retain_local_variables()) { 864 // At any safepoint, this method can get breakpointed, which would 865 // then require an immediate deoptimization. 866 can_prune_locals = false; // do not prune locals 867 stack_slots_not_pruned = 0; 868 } 869 870 C->add_safepoint_edges(call, youngest_jvms, can_prune_locals, stack_slots_not_pruned); 871 } 872 873 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 874 Bytecodes::Code code = java_bc(); 875 if (code == Bytecodes::_wide) { 876 code = method()->java_code_at_bci(bci() + 1); 877 } 878 879 BasicType rtype = T_ILLEGAL; 880 int rsize = 0; 881 882 if (code != Bytecodes::_illegal) { 883 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 884 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 885 if (rtype < T_CONFLICT) 886 rsize = type2size[rtype]; 887 } 888 889 switch (code) { 890 case Bytecodes::_illegal: 891 return false; 892 893 case Bytecodes::_ldc: 894 case Bytecodes::_ldc_w: 895 case Bytecodes::_ldc2_w: 896 inputs = 0; 897 break; 898 899 case Bytecodes::_dup: inputs = 1; break; 900 case Bytecodes::_dup_x1: inputs = 2; break; 901 case Bytecodes::_dup_x2: inputs = 3; break; 902 case Bytecodes::_dup2: inputs = 2; break; 903 case Bytecodes::_dup2_x1: inputs = 3; break; 904 case Bytecodes::_dup2_x2: inputs = 4; break; 905 case Bytecodes::_swap: inputs = 2; break; 906 case Bytecodes::_arraylength: inputs = 1; break; 907 908 case Bytecodes::_getstatic: 909 case Bytecodes::_putstatic: 910 case Bytecodes::_getfield: 911 case Bytecodes::_putfield: 912 { 913 bool ignored_will_link; 914 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 915 int size = field->type()->size(); 916 bool is_get = (depth >= 0), is_static = (depth & 1); 917 inputs = (is_static ? 0 : 1); 918 if (is_get) { 919 depth = size - inputs; 920 } else { 921 inputs += size; // putxxx pops the value from the stack 922 depth = - inputs; 923 } 924 } 925 break; 926 927 case Bytecodes::_invokevirtual: 928 case Bytecodes::_invokespecial: 929 case Bytecodes::_invokestatic: 930 case Bytecodes::_invokedynamic: 931 case Bytecodes::_invokeinterface: 932 { 933 bool ignored_will_link; 934 ciSignature* declared_signature = NULL; 935 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 936 assert(declared_signature != NULL, "cannot be null"); 937 inputs = declared_signature->arg_size_for_bc(code); 938 int size = declared_signature->return_type()->size(); 939 depth = size - inputs; 940 } 941 break; 942 943 case Bytecodes::_multianewarray: 944 { 945 ciBytecodeStream iter(method()); 946 iter.reset_to_bci(bci()); 947 iter.next(); 948 inputs = iter.get_dimensions(); 949 assert(rsize == 1, ""); 950 depth = rsize - inputs; 951 } 952 break; 953 954 case Bytecodes::_vwithfield: { 955 bool ignored_will_link; 956 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 957 int size = field->type()->size(); 958 inputs = size+1; 959 depth = rsize - inputs; 960 break; 961 } 962 963 case Bytecodes::_ireturn: 964 case Bytecodes::_lreturn: 965 case Bytecodes::_freturn: 966 case Bytecodes::_dreturn: 967 case Bytecodes::_areturn: 968 case Bytecodes::_vreturn: 969 assert(rsize == -depth, ""); 970 inputs = rsize; 971 break; 972 973 case Bytecodes::_jsr: 974 case Bytecodes::_jsr_w: 975 inputs = 0; 976 depth = 1; // S.B. depth=1, not zero 977 break; 978 979 default: 980 // bytecode produces a typed result 981 inputs = rsize - depth; 982 assert(inputs >= 0, ""); 983 break; 984 } 985 986 #ifdef ASSERT 987 // spot check 988 int outputs = depth + inputs; 989 assert(outputs >= 0, "sanity"); 990 switch (code) { 991 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 992 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 993 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 994 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 995 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 996 default: break; 997 } 998 #endif //ASSERT 999 1000 return true; 1001 } 1002 1003 1004 1005 //------------------------------basic_plus_adr--------------------------------- 1006 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1007 // short-circuit a common case 1008 if (offset == intcon(0)) return ptr; 1009 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1010 } 1011 1012 Node* GraphKit::ConvI2L(Node* offset) { 1013 // short-circuit a common case 1014 jint offset_con = find_int_con(offset, Type::OffsetBot); 1015 if (offset_con != Type::OffsetBot) { 1016 return longcon((jlong) offset_con); 1017 } 1018 return _gvn.transform( new ConvI2LNode(offset)); 1019 } 1020 1021 Node* GraphKit::ConvI2UL(Node* offset) { 1022 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1023 if (offset_con != (juint) Type::OffsetBot) { 1024 return longcon((julong) offset_con); 1025 } 1026 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1027 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1028 return _gvn.transform( new AndLNode(conv, mask) ); 1029 } 1030 1031 Node* GraphKit::ConvL2I(Node* offset) { 1032 // short-circuit a common case 1033 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1034 if (offset_con != (jlong)Type::OffsetBot) { 1035 return intcon((int) offset_con); 1036 } 1037 return _gvn.transform( new ConvL2INode(offset)); 1038 } 1039 1040 //-------------------------load_object_klass----------------------------------- 1041 Node* GraphKit::load_object_klass(Node* obj) { 1042 // Special-case a fresh allocation to avoid building nodes: 1043 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1044 if (akls != NULL) return akls; 1045 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1046 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1047 } 1048 1049 //-------------------------load_array_length----------------------------------- 1050 Node* GraphKit::load_array_length(Node* array) { 1051 // Special-case a fresh allocation to avoid building nodes: 1052 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1053 Node *alen; 1054 if (alloc == NULL) { 1055 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1056 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1057 } else { 1058 alen = alloc->Ideal_length(); 1059 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1060 if (ccast != alen) { 1061 alen = _gvn.transform(ccast); 1062 } 1063 } 1064 return alen; 1065 } 1066 1067 //------------------------------do_null_check---------------------------------- 1068 // Helper function to do a NULL pointer check. Returned value is 1069 // the incoming address with NULL casted away. You are allowed to use the 1070 // not-null value only if you are control dependent on the test. 1071 #ifndef PRODUCT 1072 extern int explicit_null_checks_inserted, 1073 explicit_null_checks_elided; 1074 #endif 1075 Node* GraphKit::null_check_common(Node* value, BasicType type, 1076 // optional arguments for variations: 1077 bool assert_null, 1078 Node* *null_control, 1079 bool speculative) { 1080 assert(!assert_null || null_control == NULL, "not both at once"); 1081 if (stopped()) return top(); 1082 NOT_PRODUCT(explicit_null_checks_inserted++); 1083 1084 // Construct NULL check 1085 Node *chk = NULL; 1086 switch(type) { 1087 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1088 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1089 case T_VALUETYPE : // fall through 1090 case T_ARRAY : // fall through 1091 type = T_OBJECT; // simplify further tests 1092 case T_OBJECT : { 1093 const Type *t = _gvn.type( value ); 1094 1095 const TypeOopPtr* tp = t->isa_oopptr(); 1096 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1097 // Only for do_null_check, not any of its siblings: 1098 && !assert_null && null_control == NULL) { 1099 // Usually, any field access or invocation on an unloaded oop type 1100 // will simply fail to link, since the statically linked class is 1101 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1102 // the static class is loaded but the sharper oop type is not. 1103 // Rather than checking for this obscure case in lots of places, 1104 // we simply observe that a null check on an unloaded class 1105 // will always be followed by a nonsense operation, so we 1106 // can just issue the uncommon trap here. 1107 // Our access to the unloaded class will only be correct 1108 // after it has been loaded and initialized, which requires 1109 // a trip through the interpreter. 1110 #ifndef PRODUCT 1111 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1112 #endif 1113 uncommon_trap(Deoptimization::Reason_unloaded, 1114 Deoptimization::Action_reinterpret, 1115 tp->klass(), "!loaded"); 1116 return top(); 1117 } 1118 1119 if (assert_null) { 1120 // See if the type is contained in NULL_PTR. 1121 // If so, then the value is already null. 1122 if (t->higher_equal(TypePtr::NULL_PTR)) { 1123 NOT_PRODUCT(explicit_null_checks_elided++); 1124 return value; // Elided null assert quickly! 1125 } 1126 } else { 1127 // See if mixing in the NULL pointer changes type. 1128 // If so, then the NULL pointer was not allowed in the original 1129 // type. In other words, "value" was not-null. 1130 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1131 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1132 NOT_PRODUCT(explicit_null_checks_elided++); 1133 return value; // Elided null check quickly! 1134 } 1135 } 1136 chk = new CmpPNode( value, null() ); 1137 break; 1138 } 1139 1140 default: 1141 fatal("unexpected type: %s", type2name(type)); 1142 } 1143 assert(chk != NULL, "sanity check"); 1144 chk = _gvn.transform(chk); 1145 1146 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1147 BoolNode *btst = new BoolNode( chk, btest); 1148 Node *tst = _gvn.transform( btst ); 1149 1150 //----------- 1151 // if peephole optimizations occurred, a prior test existed. 1152 // If a prior test existed, maybe it dominates as we can avoid this test. 1153 if (tst != btst && type == T_OBJECT) { 1154 // At this point we want to scan up the CFG to see if we can 1155 // find an identical test (and so avoid this test altogether). 1156 Node *cfg = control(); 1157 int depth = 0; 1158 while( depth < 16 ) { // Limit search depth for speed 1159 if( cfg->Opcode() == Op_IfTrue && 1160 cfg->in(0)->in(1) == tst ) { 1161 // Found prior test. Use "cast_not_null" to construct an identical 1162 // CastPP (and hence hash to) as already exists for the prior test. 1163 // Return that casted value. 1164 if (assert_null) { 1165 replace_in_map(value, null()); 1166 return null(); // do not issue the redundant test 1167 } 1168 Node *oldcontrol = control(); 1169 set_control(cfg); 1170 Node *res = cast_not_null(value); 1171 set_control(oldcontrol); 1172 NOT_PRODUCT(explicit_null_checks_elided++); 1173 return res; 1174 } 1175 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1176 if (cfg == NULL) break; // Quit at region nodes 1177 depth++; 1178 } 1179 } 1180 1181 //----------- 1182 // Branch to failure if null 1183 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1184 Deoptimization::DeoptReason reason; 1185 if (assert_null) { 1186 reason = Deoptimization::reason_null_assert(speculative); 1187 } else if (type == T_OBJECT) { 1188 reason = Deoptimization::reason_null_check(speculative); 1189 } else { 1190 reason = Deoptimization::Reason_div0_check; 1191 } 1192 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1193 // ciMethodData::has_trap_at will return a conservative -1 if any 1194 // must-be-null assertion has failed. This could cause performance 1195 // problems for a method after its first do_null_assert failure. 1196 // Consider using 'Reason_class_check' instead? 1197 1198 // To cause an implicit null check, we set the not-null probability 1199 // to the maximum (PROB_MAX). For an explicit check the probability 1200 // is set to a smaller value. 1201 if (null_control != NULL || too_many_traps(reason)) { 1202 // probability is less likely 1203 ok_prob = PROB_LIKELY_MAG(3); 1204 } else if (!assert_null && 1205 (ImplicitNullCheckThreshold > 0) && 1206 method() != NULL && 1207 (method()->method_data()->trap_count(reason) 1208 >= (uint)ImplicitNullCheckThreshold)) { 1209 ok_prob = PROB_LIKELY_MAG(3); 1210 } 1211 1212 if (null_control != NULL) { 1213 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1214 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1215 set_control( _gvn.transform( new IfTrueNode(iff))); 1216 #ifndef PRODUCT 1217 if (null_true == top()) { 1218 explicit_null_checks_elided++; 1219 } 1220 #endif 1221 (*null_control) = null_true; 1222 } else { 1223 BuildCutout unless(this, tst, ok_prob); 1224 // Check for optimizer eliding test at parse time 1225 if (stopped()) { 1226 // Failure not possible; do not bother making uncommon trap. 1227 NOT_PRODUCT(explicit_null_checks_elided++); 1228 } else if (assert_null) { 1229 uncommon_trap(reason, 1230 Deoptimization::Action_make_not_entrant, 1231 NULL, "assert_null"); 1232 } else { 1233 replace_in_map(value, zerocon(type)); 1234 builtin_throw(reason); 1235 } 1236 } 1237 1238 // Must throw exception, fall-thru not possible? 1239 if (stopped()) { 1240 return top(); // No result 1241 } 1242 1243 if (assert_null) { 1244 // Cast obj to null on this path. 1245 replace_in_map(value, zerocon(type)); 1246 return zerocon(type); 1247 } 1248 1249 // Cast obj to not-null on this path, if there is no null_control. 1250 // (If there is a null_control, a non-null value may come back to haunt us.) 1251 if (type == T_OBJECT) { 1252 Node* cast = cast_not_null(value, false); 1253 if (null_control == NULL || (*null_control) == top()) 1254 replace_in_map(value, cast); 1255 value = cast; 1256 } 1257 1258 return value; 1259 } 1260 1261 1262 //------------------------------cast_not_null---------------------------------- 1263 // Cast obj to not-null on this path 1264 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1265 const Type *t = _gvn.type(obj); 1266 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1267 // Object is already not-null? 1268 if( t == t_not_null ) return obj; 1269 1270 Node *cast = new CastPPNode(obj,t_not_null); 1271 cast->init_req(0, control()); 1272 cast = _gvn.transform( cast ); 1273 1274 // Scan for instances of 'obj' in the current JVM mapping. 1275 // These instances are known to be not-null after the test. 1276 if (do_replace_in_map) 1277 replace_in_map(obj, cast); 1278 1279 return cast; // Return casted value 1280 } 1281 1282 // Sometimes in intrinsics, we implicitly know an object is not null 1283 // (there's no actual null check) so we can cast it to not null. In 1284 // the course of optimizations, the input to the cast can become null. 1285 // In that case that data path will die and we need the control path 1286 // to become dead as well to keep the graph consistent. So we have to 1287 // add a check for null for which one branch can't be taken. It uses 1288 // an Opaque4 node that will cause the check to be removed after loop 1289 // opts so the test goes away and the compiled code doesn't execute a 1290 // useless check. 1291 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1292 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1293 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1294 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1295 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1296 _gvn.set_type(iff, iff->Value(&_gvn)); 1297 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1298 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1299 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1300 C->root()->add_req(halt); 1301 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1302 set_control(if_t); 1303 return cast_not_null(value, do_replace_in_map); 1304 } 1305 1306 1307 //--------------------------replace_in_map------------------------------------- 1308 void GraphKit::replace_in_map(Node* old, Node* neww) { 1309 if (old == neww) { 1310 return; 1311 } 1312 1313 map()->replace_edge(old, neww); 1314 1315 // Note: This operation potentially replaces any edge 1316 // on the map. This includes locals, stack, and monitors 1317 // of the current (innermost) JVM state. 1318 1319 // don't let inconsistent types from profiling escape this 1320 // method 1321 1322 const Type* told = _gvn.type(old); 1323 const Type* tnew = _gvn.type(neww); 1324 1325 if (!tnew->higher_equal(told)) { 1326 return; 1327 } 1328 1329 map()->record_replaced_node(old, neww); 1330 } 1331 1332 1333 //============================================================================= 1334 //--------------------------------memory--------------------------------------- 1335 Node* GraphKit::memory(uint alias_idx) { 1336 MergeMemNode* mem = merged_memory(); 1337 Node* p = mem->memory_at(alias_idx); 1338 _gvn.set_type(p, Type::MEMORY); // must be mapped 1339 return p; 1340 } 1341 1342 //-----------------------------reset_memory------------------------------------ 1343 Node* GraphKit::reset_memory() { 1344 Node* mem = map()->memory(); 1345 // do not use this node for any more parsing! 1346 debug_only( map()->set_memory((Node*)NULL) ); 1347 return _gvn.transform( mem ); 1348 } 1349 1350 //------------------------------set_all_memory--------------------------------- 1351 void GraphKit::set_all_memory(Node* newmem) { 1352 Node* mergemem = MergeMemNode::make(newmem); 1353 gvn().set_type_bottom(mergemem); 1354 map()->set_memory(mergemem); 1355 } 1356 1357 //------------------------------set_all_memory_call---------------------------- 1358 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1359 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1360 set_all_memory(newmem); 1361 } 1362 1363 //============================================================================= 1364 // 1365 // parser factory methods for MemNodes 1366 // 1367 // These are layered on top of the factory methods in LoadNode and StoreNode, 1368 // and integrate with the parser's memory state and _gvn engine. 1369 // 1370 1371 // factory methods in "int adr_idx" 1372 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1373 int adr_idx, 1374 MemNode::MemOrd mo, 1375 LoadNode::ControlDependency control_dependency, 1376 bool require_atomic_access, 1377 bool unaligned, 1378 bool mismatched) { 1379 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1380 const TypePtr* adr_type = NULL; // debug-mode-only argument 1381 debug_only(adr_type = C->get_adr_type(adr_idx)); 1382 Node* mem = memory(adr_idx); 1383 Node* ld; 1384 if (require_atomic_access && bt == T_LONG) { 1385 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1386 } else if (require_atomic_access && bt == T_DOUBLE) { 1387 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1388 } else { 1389 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1390 } 1391 ld = _gvn.transform(ld); 1392 if (bt == T_VALUETYPE) { 1393 // Loading a non-flattened value type from memory requires a null check. 1394 ld = ValueTypeNode::make_from_oop(this, ld, true /* null check */); 1395 } else if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1396 // Improve graph before escape analysis and boxing elimination. 1397 record_for_igvn(ld); 1398 } 1399 return ld; 1400 } 1401 1402 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1403 int adr_idx, 1404 MemNode::MemOrd mo, 1405 bool require_atomic_access, 1406 bool unaligned, 1407 bool mismatched) { 1408 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1409 const TypePtr* adr_type = NULL; 1410 debug_only(adr_type = C->get_adr_type(adr_idx)); 1411 Node *mem = memory(adr_idx); 1412 Node* st; 1413 if (require_atomic_access && bt == T_LONG) { 1414 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1415 } else if (require_atomic_access && bt == T_DOUBLE) { 1416 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1417 } else { 1418 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1419 } 1420 if (unaligned) { 1421 st->as_Store()->set_unaligned_access(); 1422 } 1423 if (mismatched) { 1424 st->as_Store()->set_mismatched_access(); 1425 } 1426 st = _gvn.transform(st); 1427 set_memory(st, adr_idx); 1428 // Back-to-back stores can only remove intermediate store with DU info 1429 // so push on worklist for optimizer. 1430 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1431 record_for_igvn(st); 1432 1433 return st; 1434 } 1435 1436 1437 void GraphKit::pre_barrier(bool do_load, 1438 Node* ctl, 1439 Node* obj, 1440 Node* adr, 1441 uint adr_idx, 1442 Node* val, 1443 const TypeOopPtr* val_type, 1444 Node* pre_val, 1445 BasicType bt) { 1446 1447 BarrierSet* bs = Universe::heap()->barrier_set(); 1448 set_control(ctl); 1449 switch (bs->kind()) { 1450 case BarrierSet::G1SATBCTLogging: 1451 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1452 break; 1453 1454 case BarrierSet::CardTableForRS: 1455 case BarrierSet::CardTableExtension: 1456 case BarrierSet::ModRef: 1457 break; 1458 1459 default : 1460 ShouldNotReachHere(); 1461 1462 } 1463 } 1464 1465 bool GraphKit::can_move_pre_barrier() const { 1466 BarrierSet* bs = Universe::heap()->barrier_set(); 1467 switch (bs->kind()) { 1468 case BarrierSet::G1SATBCTLogging: 1469 return true; // Can move it if no safepoint 1470 1471 case BarrierSet::CardTableForRS: 1472 case BarrierSet::CardTableExtension: 1473 case BarrierSet::ModRef: 1474 return true; // There is no pre-barrier 1475 1476 default : 1477 ShouldNotReachHere(); 1478 } 1479 return false; 1480 } 1481 1482 void GraphKit::post_barrier(Node* ctl, 1483 Node* store, 1484 Node* obj, 1485 Node* adr, 1486 uint adr_idx, 1487 Node* val, 1488 BasicType bt, 1489 bool use_precise) { 1490 BarrierSet* bs = Universe::heap()->barrier_set(); 1491 set_control(ctl); 1492 switch (bs->kind()) { 1493 case BarrierSet::G1SATBCTLogging: 1494 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1495 break; 1496 1497 case BarrierSet::CardTableForRS: 1498 case BarrierSet::CardTableExtension: 1499 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1500 break; 1501 1502 case BarrierSet::ModRef: 1503 break; 1504 1505 default : 1506 ShouldNotReachHere(); 1507 1508 } 1509 } 1510 1511 Node* GraphKit::store_oop(Node* ctl, 1512 Node* obj, 1513 Node* adr, 1514 const TypePtr* adr_type, 1515 Node* val, 1516 const TypeOopPtr* val_type, 1517 BasicType bt, 1518 bool use_precise, 1519 MemNode::MemOrd mo, 1520 bool mismatched) { 1521 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1522 // could be delayed during Parse (for example, in adjust_map_after_if()). 1523 // Execute transformation here to avoid barrier generation in such case. 1524 if (_gvn.type(val) == TypePtr::NULL_PTR) 1525 val = _gvn.makecon(TypePtr::NULL_PTR); 1526 1527 set_control(ctl); 1528 if (stopped()) return top(); // Dead path ? 1529 1530 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity"); 1531 assert(val != NULL, "not dead path"); 1532 uint adr_idx = C->get_alias_index(adr_type); 1533 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1534 1535 if (val->is_ValueType()) { 1536 // Allocate value type and get oop 1537 val = val->as_ValueType()->allocate(this)->get_oop(); 1538 } 1539 1540 pre_barrier(true /* do_load */, 1541 control(), obj, adr, adr_idx, val, val_type, 1542 NULL /* pre_val */, 1543 bt); 1544 1545 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1546 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1547 return store; 1548 } 1549 1550 // Could be an array or object we don't know at compile time (unsafe ref.) 1551 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1552 Node* obj, // containing obj 1553 Node* adr, // actual adress to store val at 1554 const TypePtr* adr_type, 1555 Node* val, 1556 BasicType bt, 1557 MemNode::MemOrd mo, 1558 bool mismatched) { 1559 Compile::AliasType* at = C->alias_type(adr_type); 1560 const TypeOopPtr* val_type = NULL; 1561 if (adr_type->isa_instptr()) { 1562 if (at->field() != NULL) { 1563 // known field. This code is a copy of the do_put_xxx logic. 1564 ciField* field = at->field(); 1565 if (!field->type()->is_loaded()) { 1566 val_type = TypeInstPtr::BOTTOM; 1567 } else { 1568 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1569 } 1570 } 1571 } else if (adr_type->isa_aryptr()) { 1572 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1573 } 1574 if (val_type == NULL) { 1575 val_type = TypeInstPtr::BOTTOM; 1576 } 1577 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1578 } 1579 1580 1581 //-------------------------array_element_address------------------------- 1582 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1583 const TypeInt* sizetype, Node* ctrl) { 1584 uint shift = exact_log2(type2aelembytes(elembt)); 1585 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1586 if (arytype_klass->is_value_array_klass()) { 1587 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass(); 1588 shift = vak->log2_element_size(); 1589 } 1590 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1591 1592 // short-circuit a common case (saves lots of confusing waste motion) 1593 jint idx_con = find_int_con(idx, -1); 1594 if (idx_con >= 0) { 1595 intptr_t offset = header + ((intptr_t)idx_con << shift); 1596 return basic_plus_adr(ary, offset); 1597 } 1598 1599 // must be correct type for alignment purposes 1600 Node* base = basic_plus_adr(ary, header); 1601 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1602 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1603 return basic_plus_adr(ary, base, scale); 1604 } 1605 1606 //-------------------------load_array_element------------------------- 1607 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1608 const Type* elemtype = arytype->elem(); 1609 BasicType elembt = elemtype->array_element_basic_type(); 1610 assert(elembt != T_VALUETYPE, "value types are not supported by this method"); 1611 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1612 if (elembt == T_NARROWOOP) { 1613 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1614 } 1615 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1616 return ld; 1617 } 1618 1619 //-------------------------set_arguments_for_java_call------------------------- 1620 // Arguments (pre-popped from the stack) are taken from the JVMS. 1621 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1622 // Add the call arguments: 1623 const TypeTuple* domain = call->tf()->domain_sig(); 1624 uint nargs = domain->cnt(); 1625 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1626 Node* arg = argument(i-TypeFunc::Parms); 1627 if (ValueTypePassFieldsAsArgs) { 1628 if (arg->is_ValueType()) { 1629 ValueTypeNode* vt = arg->as_ValueType(); 1630 if (!domain->field_at(i)->is_valuetypeptr()->is__Value()) { 1631 // We don't pass value type arguments by reference but instead 1632 // pass each field of the value type 1633 idx += vt->pass_fields(call, idx, *this); 1634 // If a value type argument is passed as fields, attach the Method* to the call site 1635 // to be able to access the extended signature later via attached_method_before_pc(). 1636 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1637 call->set_override_symbolic_info(true); 1638 } else { 1639 arg = arg->as_ValueType()->allocate(this)->get_oop(); 1640 call->init_req(idx, arg); 1641 idx++; 1642 } 1643 } else { 1644 call->init_req(idx, arg); 1645 idx++; 1646 } 1647 } else { 1648 if (arg->is_ValueType()) { 1649 // Pass value type argument via oop to callee 1650 arg = arg->as_ValueType()->allocate(this)->get_oop(); 1651 } 1652 call->init_req(i, arg); 1653 } 1654 } 1655 } 1656 1657 //---------------------------set_edges_for_java_call--------------------------- 1658 // Connect a newly created call into the current JVMS. 1659 // A return value node (if any) is returned from set_edges_for_java_call. 1660 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1661 1662 // Add the predefined inputs: 1663 call->init_req( TypeFunc::Control, control() ); 1664 call->init_req( TypeFunc::I_O , i_o() ); 1665 call->init_req( TypeFunc::Memory , reset_memory() ); 1666 call->init_req( TypeFunc::FramePtr, frameptr() ); 1667 call->init_req( TypeFunc::ReturnAdr, top() ); 1668 1669 add_safepoint_edges(call, must_throw); 1670 1671 Node* xcall = _gvn.transform(call); 1672 1673 if (xcall == top()) { 1674 set_control(top()); 1675 return; 1676 } 1677 assert(xcall == call, "call identity is stable"); 1678 1679 // Re-use the current map to produce the result. 1680 1681 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1682 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1683 set_all_memory_call(xcall, separate_io_proj); 1684 1685 //return xcall; // no need, caller already has it 1686 } 1687 1688 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1689 if (stopped()) return top(); // maybe the call folded up? 1690 1691 // Note: Since any out-of-line call can produce an exception, 1692 // we always insert an I_O projection from the call into the result. 1693 1694 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1695 1696 if (separate_io_proj) { 1697 // The caller requested separate projections be used by the fall 1698 // through and exceptional paths, so replace the projections for 1699 // the fall through path. 1700 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1701 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1702 } 1703 1704 // Capture the return value, if any. 1705 Node* ret; 1706 if (call->method() == NULL || 1707 call->method()->return_type()->basic_type() == T_VOID) { 1708 ret = top(); 1709 } else { 1710 if (!call->tf()->returns_value_type_as_fields()) { 1711 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1712 } else { 1713 // Return of multiple values (value type fields): we create a 1714 // ValueType node, each field is a projection from the call. 1715 const TypeTuple* range_sig = call->tf()->range_sig(); 1716 const Type* t = range_sig->field_at(TypeFunc::Parms); 1717 assert(t->isa_valuetypeptr(), "only value types for multiple return values"); 1718 ciValueKlass* vk = t->is_valuetypeptr()->value_klass(); 1719 Node* ctl = control(); 1720 ret = ValueTypeNode::make_from_multi(_gvn, ctl, merged_memory(), call, vk, TypeFunc::Parms+1, false); 1721 set_control(ctl); 1722 } 1723 } 1724 1725 return ret; 1726 } 1727 1728 //--------------------set_predefined_input_for_runtime_call-------------------- 1729 // Reading and setting the memory state is way conservative here. 1730 // The real problem is that I am not doing real Type analysis on memory, 1731 // so I cannot distinguish card mark stores from other stores. Across a GC 1732 // point the Store Barrier and the card mark memory has to agree. I cannot 1733 // have a card mark store and its barrier split across the GC point from 1734 // either above or below. Here I get that to happen by reading ALL of memory. 1735 // A better answer would be to separate out card marks from other memory. 1736 // For now, return the input memory state, so that it can be reused 1737 // after the call, if this call has restricted memory effects. 1738 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1739 // Set fixed predefined input arguments 1740 Node* memory = reset_memory(); 1741 call->init_req( TypeFunc::Control, control() ); 1742 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1743 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1744 call->init_req( TypeFunc::FramePtr, frameptr() ); 1745 call->init_req( TypeFunc::ReturnAdr, top() ); 1746 return memory; 1747 } 1748 1749 //-------------------set_predefined_output_for_runtime_call-------------------- 1750 // Set control and memory (not i_o) from the call. 1751 // If keep_mem is not NULL, use it for the output state, 1752 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1753 // If hook_mem is NULL, this call produces no memory effects at all. 1754 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1755 // then only that memory slice is taken from the call. 1756 // In the last case, we must put an appropriate memory barrier before 1757 // the call, so as to create the correct anti-dependencies on loads 1758 // preceding the call. 1759 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1760 Node* keep_mem, 1761 const TypePtr* hook_mem) { 1762 // no i/o 1763 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1764 if (keep_mem) { 1765 // First clone the existing memory state 1766 set_all_memory(keep_mem); 1767 if (hook_mem != NULL) { 1768 // Make memory for the call 1769 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1770 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1771 // We also use hook_mem to extract specific effects from arraycopy stubs. 1772 set_memory(mem, hook_mem); 1773 } 1774 // ...else the call has NO memory effects. 1775 1776 // Make sure the call advertises its memory effects precisely. 1777 // This lets us build accurate anti-dependences in gcm.cpp. 1778 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1779 "call node must be constructed correctly"); 1780 } else { 1781 assert(hook_mem == NULL, ""); 1782 // This is not a "slow path" call; all memory comes from the call. 1783 set_all_memory_call(call); 1784 } 1785 } 1786 1787 1788 // Replace the call with the current state of the kit. 1789 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1790 JVMState* ejvms = NULL; 1791 if (has_exceptions()) { 1792 ejvms = transfer_exceptions_into_jvms(); 1793 } 1794 1795 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1796 ReplacedNodes replaced_nodes_exception; 1797 Node* ex_ctl = top(); 1798 1799 SafePointNode* final_state = stop(); 1800 1801 // Find all the needed outputs of this call 1802 CallProjections callprojs; 1803 call->extract_projections(&callprojs, true); 1804 1805 Node* init_mem = call->in(TypeFunc::Memory); 1806 Node* final_mem = final_state->in(TypeFunc::Memory); 1807 Node* final_ctl = final_state->in(TypeFunc::Control); 1808 Node* final_io = final_state->in(TypeFunc::I_O); 1809 1810 // Replace all the old call edges with the edges from the inlining result 1811 if (callprojs.fallthrough_catchproj != NULL) { 1812 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1813 } 1814 if (callprojs.fallthrough_memproj != NULL) { 1815 if (final_mem->is_MergeMem()) { 1816 // Parser's exits MergeMem was not transformed but may be optimized 1817 final_mem = _gvn.transform(final_mem); 1818 } 1819 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1820 } 1821 if (callprojs.fallthrough_ioproj != NULL) { 1822 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1823 } 1824 1825 // Replace the result with the new result if it exists and is used 1826 if (callprojs.resproj != NULL && result != NULL) { 1827 C->gvn_replace_by(callprojs.resproj, result); 1828 } 1829 1830 if (ejvms == NULL) { 1831 // No exception edges to simply kill off those paths 1832 if (callprojs.catchall_catchproj != NULL) { 1833 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1834 } 1835 if (callprojs.catchall_memproj != NULL) { 1836 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1837 } 1838 if (callprojs.catchall_ioproj != NULL) { 1839 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1840 } 1841 // Replace the old exception object with top 1842 if (callprojs.exobj != NULL) { 1843 C->gvn_replace_by(callprojs.exobj, C->top()); 1844 } 1845 } else { 1846 GraphKit ekit(ejvms); 1847 1848 // Load my combined exception state into the kit, with all phis transformed: 1849 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1850 replaced_nodes_exception = ex_map->replaced_nodes(); 1851 1852 Node* ex_oop = ekit.use_exception_state(ex_map); 1853 1854 if (callprojs.catchall_catchproj != NULL) { 1855 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1856 ex_ctl = ekit.control(); 1857 } 1858 if (callprojs.catchall_memproj != NULL) { 1859 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1860 } 1861 if (callprojs.catchall_ioproj != NULL) { 1862 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1863 } 1864 1865 // Replace the old exception object with the newly created one 1866 if (callprojs.exobj != NULL) { 1867 C->gvn_replace_by(callprojs.exobj, ex_oop); 1868 } 1869 } 1870 1871 // Disconnect the call from the graph 1872 call->disconnect_inputs(NULL, C); 1873 C->gvn_replace_by(call, C->top()); 1874 1875 // Clean up any MergeMems that feed other MergeMems since the 1876 // optimizer doesn't like that. 1877 if (final_mem->is_MergeMem()) { 1878 Node_List wl; 1879 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1880 Node* m = i.get(); 1881 if (m->is_MergeMem() && !wl.contains(m)) { 1882 wl.push(m); 1883 } 1884 } 1885 while (wl.size() > 0) { 1886 _gvn.transform(wl.pop()); 1887 } 1888 } 1889 1890 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1891 replaced_nodes.apply(C, final_ctl); 1892 } 1893 if (!ex_ctl->is_top() && do_replaced_nodes) { 1894 replaced_nodes_exception.apply(C, ex_ctl); 1895 } 1896 } 1897 1898 1899 //------------------------------increment_counter------------------------------ 1900 // for statistics: increment a VM counter by 1 1901 1902 void GraphKit::increment_counter(address counter_addr) { 1903 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1904 increment_counter(adr1); 1905 } 1906 1907 void GraphKit::increment_counter(Node* counter_addr) { 1908 int adr_type = Compile::AliasIdxRaw; 1909 Node* ctrl = control(); 1910 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1911 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1912 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1913 } 1914 1915 1916 //------------------------------uncommon_trap---------------------------------- 1917 // Bail out to the interpreter in mid-method. Implemented by calling the 1918 // uncommon_trap blob. This helper function inserts a runtime call with the 1919 // right debug info. 1920 void GraphKit::uncommon_trap(int trap_request, 1921 ciKlass* klass, const char* comment, 1922 bool must_throw, 1923 bool keep_exact_action) { 1924 if (failing()) stop(); 1925 if (stopped()) return; // trap reachable? 1926 1927 // Note: If ProfileTraps is true, and if a deopt. actually 1928 // occurs here, the runtime will make sure an MDO exists. There is 1929 // no need to call method()->ensure_method_data() at this point. 1930 1931 // Set the stack pointer to the right value for reexecution: 1932 set_sp(reexecute_sp()); 1933 1934 #ifdef ASSERT 1935 if (!must_throw) { 1936 // Make sure the stack has at least enough depth to execute 1937 // the current bytecode. 1938 int inputs, ignored_depth; 1939 if (compute_stack_effects(inputs, ignored_depth)) { 1940 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1941 Bytecodes::name(java_bc()), sp(), inputs); 1942 } 1943 } 1944 #endif 1945 1946 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1947 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1948 1949 switch (action) { 1950 case Deoptimization::Action_maybe_recompile: 1951 case Deoptimization::Action_reinterpret: 1952 // Temporary fix for 6529811 to allow virtual calls to be sure they 1953 // get the chance to go from mono->bi->mega 1954 if (!keep_exact_action && 1955 Deoptimization::trap_request_index(trap_request) < 0 && 1956 too_many_recompiles(reason)) { 1957 // This BCI is causing too many recompilations. 1958 if (C->log() != NULL) { 1959 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1960 Deoptimization::trap_reason_name(reason), 1961 Deoptimization::trap_action_name(action)); 1962 } 1963 action = Deoptimization::Action_none; 1964 trap_request = Deoptimization::make_trap_request(reason, action); 1965 } else { 1966 C->set_trap_can_recompile(true); 1967 } 1968 break; 1969 case Deoptimization::Action_make_not_entrant: 1970 C->set_trap_can_recompile(true); 1971 break; 1972 case Deoptimization::Action_none: 1973 case Deoptimization::Action_make_not_compilable: 1974 break; 1975 default: 1976 #ifdef ASSERT 1977 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 1978 #endif 1979 break; 1980 } 1981 1982 if (TraceOptoParse) { 1983 char buf[100]; 1984 tty->print_cr("Uncommon trap %s at bci:%d", 1985 Deoptimization::format_trap_request(buf, sizeof(buf), 1986 trap_request), bci()); 1987 } 1988 1989 CompileLog* log = C->log(); 1990 if (log != NULL) { 1991 int kid = (klass == NULL)? -1: log->identify(klass); 1992 log->begin_elem("uncommon_trap bci='%d'", bci()); 1993 char buf[100]; 1994 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1995 trap_request)); 1996 if (kid >= 0) log->print(" klass='%d'", kid); 1997 if (comment != NULL) log->print(" comment='%s'", comment); 1998 log->end_elem(); 1999 } 2000 2001 // Make sure any guarding test views this path as very unlikely 2002 Node *i0 = control()->in(0); 2003 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2004 IfNode *iff = i0->as_If(); 2005 float f = iff->_prob; // Get prob 2006 if (control()->Opcode() == Op_IfTrue) { 2007 if (f > PROB_UNLIKELY_MAG(4)) 2008 iff->_prob = PROB_MIN; 2009 } else { 2010 if (f < PROB_LIKELY_MAG(4)) 2011 iff->_prob = PROB_MAX; 2012 } 2013 } 2014 2015 // Clear out dead values from the debug info. 2016 kill_dead_locals(); 2017 2018 // Now insert the uncommon trap subroutine call 2019 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2020 const TypePtr* no_memory_effects = NULL; 2021 // Pass the index of the class to be loaded 2022 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2023 (must_throw ? RC_MUST_THROW : 0), 2024 OptoRuntime::uncommon_trap_Type(), 2025 call_addr, "uncommon_trap", no_memory_effects, 2026 intcon(trap_request)); 2027 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2028 "must extract request correctly from the graph"); 2029 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2030 2031 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2032 // The debug info is the only real input to this call. 2033 2034 // Halt-and-catch fire here. The above call should never return! 2035 HaltNode* halt = new HaltNode(control(), frameptr()); 2036 _gvn.set_type_bottom(halt); 2037 root()->add_req(halt); 2038 2039 stop_and_kill_map(); 2040 } 2041 2042 2043 //--------------------------just_allocated_object------------------------------ 2044 // Report the object that was just allocated. 2045 // It must be the case that there are no intervening safepoints. 2046 // We use this to determine if an object is so "fresh" that 2047 // it does not require card marks. 2048 Node* GraphKit::just_allocated_object(Node* current_control) { 2049 if (C->recent_alloc_ctl() == current_control) 2050 return C->recent_alloc_obj(); 2051 return NULL; 2052 } 2053 2054 2055 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2056 // (Note: TypeFunc::make has a cache that makes this fast.) 2057 const TypeFunc* tf = TypeFunc::make(dest_method); 2058 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2059 for (int j = 0; j < nargs; j++) { 2060 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2061 if( targ->basic_type() == T_DOUBLE ) { 2062 // If any parameters are doubles, they must be rounded before 2063 // the call, dstore_rounding does gvn.transform 2064 Node *arg = argument(j); 2065 arg = dstore_rounding(arg); 2066 set_argument(j, arg); 2067 } 2068 } 2069 } 2070 2071 /** 2072 * Record profiling data exact_kls for Node n with the type system so 2073 * that it can propagate it (speculation) 2074 * 2075 * @param n node that the type applies to 2076 * @param exact_kls type from profiling 2077 * @param maybe_null did profiling see null? 2078 * 2079 * @return node with improved type 2080 */ 2081 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2082 const Type* current_type = _gvn.type(n); 2083 assert(UseTypeSpeculation, "type speculation must be on"); 2084 2085 const TypePtr* speculative = current_type->speculative(); 2086 2087 // Should the klass from the profile be recorded in the speculative type? 2088 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2089 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2090 const TypeOopPtr* xtype = tklass->as_instance_type(); 2091 assert(xtype->klass_is_exact(), "Should be exact"); 2092 // Any reason to believe n is not null (from this profiling or a previous one)? 2093 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2094 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2095 // record the new speculative type's depth 2096 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2097 speculative = speculative->with_inline_depth(jvms()->depth()); 2098 } else if (current_type->would_improve_ptr(ptr_kind)) { 2099 // Profiling report that null was never seen so we can change the 2100 // speculative type to non null ptr. 2101 if (ptr_kind == ProfileAlwaysNull) { 2102 speculative = TypePtr::NULL_PTR; 2103 } else { 2104 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2105 const TypePtr* ptr = TypePtr::NOTNULL; 2106 if (speculative != NULL) { 2107 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2108 } else { 2109 speculative = ptr; 2110 } 2111 } 2112 } 2113 2114 if (speculative != current_type->speculative()) { 2115 // Build a type with a speculative type (what we think we know 2116 // about the type but will need a guard when we use it) 2117 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2118 // We're changing the type, we need a new CheckCast node to carry 2119 // the new type. The new type depends on the control: what 2120 // profiling tells us is only valid from here as far as we can 2121 // tell. 2122 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2123 cast = _gvn.transform(cast); 2124 replace_in_map(n, cast); 2125 n = cast; 2126 } 2127 2128 return n; 2129 } 2130 2131 /** 2132 * Record profiling data from receiver profiling at an invoke with the 2133 * type system so that it can propagate it (speculation) 2134 * 2135 * @param n receiver node 2136 * 2137 * @return node with improved type 2138 */ 2139 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2140 if (!UseTypeSpeculation) { 2141 return n; 2142 } 2143 ciKlass* exact_kls = profile_has_unique_klass(); 2144 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2145 if ((java_bc() == Bytecodes::_checkcast || 2146 java_bc() == Bytecodes::_instanceof || 2147 java_bc() == Bytecodes::_aastore) && 2148 method()->method_data()->is_mature()) { 2149 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2150 if (data != NULL) { 2151 if (!data->as_BitData()->null_seen()) { 2152 ptr_kind = ProfileNeverNull; 2153 } else { 2154 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2155 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2156 uint i = 0; 2157 for (; i < call->row_limit(); i++) { 2158 ciKlass* receiver = call->receiver(i); 2159 if (receiver != NULL) { 2160 break; 2161 } 2162 } 2163 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2164 } 2165 } 2166 } 2167 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2168 } 2169 2170 /** 2171 * Record profiling data from argument profiling at an invoke with the 2172 * type system so that it can propagate it (speculation) 2173 * 2174 * @param dest_method target method for the call 2175 * @param bc what invoke bytecode is this? 2176 */ 2177 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2178 if (!UseTypeSpeculation) { 2179 return; 2180 } 2181 const TypeFunc* tf = TypeFunc::make(dest_method); 2182 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2183 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2184 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2185 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2186 if (targ->isa_oopptr() && !targ->isa_valuetypeptr()) { 2187 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2188 ciKlass* better_type = NULL; 2189 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2190 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2191 } 2192 i++; 2193 } 2194 } 2195 } 2196 2197 /** 2198 * Record profiling data from parameter profiling at an invoke with 2199 * the type system so that it can propagate it (speculation) 2200 */ 2201 void GraphKit::record_profiled_parameters_for_speculation() { 2202 if (!UseTypeSpeculation) { 2203 return; 2204 } 2205 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2206 if (_gvn.type(local(i))->isa_oopptr()) { 2207 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2208 ciKlass* better_type = NULL; 2209 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2210 record_profile_for_speculation(local(i), better_type, ptr_kind); 2211 } 2212 j++; 2213 } 2214 } 2215 } 2216 2217 /** 2218 * Record profiling data from return value profiling at an invoke with 2219 * the type system so that it can propagate it (speculation) 2220 */ 2221 void GraphKit::record_profiled_return_for_speculation() { 2222 if (!UseTypeSpeculation) { 2223 return; 2224 } 2225 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2226 ciKlass* better_type = NULL; 2227 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2228 // If profiling reports a single type for the return value, 2229 // feed it to the type system so it can propagate it as a 2230 // speculative type 2231 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2232 } 2233 } 2234 2235 void GraphKit::round_double_result(ciMethod* dest_method) { 2236 // A non-strict method may return a double value which has an extended 2237 // exponent, but this must not be visible in a caller which is 'strict' 2238 // If a strict caller invokes a non-strict callee, round a double result 2239 2240 BasicType result_type = dest_method->return_type()->basic_type(); 2241 assert( method() != NULL, "must have caller context"); 2242 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2243 // Destination method's return value is on top of stack 2244 // dstore_rounding() does gvn.transform 2245 Node *result = pop_pair(); 2246 result = dstore_rounding(result); 2247 push_pair(result); 2248 } 2249 } 2250 2251 // rounding for strict float precision conformance 2252 Node* GraphKit::precision_rounding(Node* n) { 2253 return UseStrictFP && _method->flags().is_strict() 2254 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2255 ? _gvn.transform( new RoundFloatNode(0, n) ) 2256 : n; 2257 } 2258 2259 // rounding for strict double precision conformance 2260 Node* GraphKit::dprecision_rounding(Node *n) { 2261 return UseStrictFP && _method->flags().is_strict() 2262 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2263 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2264 : n; 2265 } 2266 2267 // rounding for non-strict double stores 2268 Node* GraphKit::dstore_rounding(Node* n) { 2269 return Matcher::strict_fp_requires_explicit_rounding 2270 && UseSSE <= 1 2271 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2272 : n; 2273 } 2274 2275 //============================================================================= 2276 // Generate a fast path/slow path idiom. Graph looks like: 2277 // [foo] indicates that 'foo' is a parameter 2278 // 2279 // [in] NULL 2280 // \ / 2281 // CmpP 2282 // Bool ne 2283 // If 2284 // / \ 2285 // True False-<2> 2286 // / | 2287 // / cast_not_null 2288 // Load | | ^ 2289 // [fast_test] | | 2290 // gvn to opt_test | | 2291 // / \ | <1> 2292 // True False | 2293 // | \\ | 2294 // [slow_call] \[fast_result] 2295 // Ctl Val \ \ 2296 // | \ \ 2297 // Catch <1> \ \ 2298 // / \ ^ \ \ 2299 // Ex No_Ex | \ \ 2300 // | \ \ | \ <2> \ 2301 // ... \ [slow_res] | | \ [null_result] 2302 // \ \--+--+--- | | 2303 // \ | / \ | / 2304 // --------Region Phi 2305 // 2306 //============================================================================= 2307 // Code is structured as a series of driver functions all called 'do_XXX' that 2308 // call a set of helper functions. Helper functions first, then drivers. 2309 2310 //------------------------------null_check_oop--------------------------------- 2311 // Null check oop. Set null-path control into Region in slot 3. 2312 // Make a cast-not-nullness use the other not-null control. Return cast. 2313 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2314 bool never_see_null, 2315 bool safe_for_replace, 2316 bool speculative) { 2317 // Initial NULL check taken path 2318 (*null_control) = top(); 2319 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2320 2321 // Generate uncommon_trap: 2322 if (never_see_null && (*null_control) != top()) { 2323 // If we see an unexpected null at a check-cast we record it and force a 2324 // recompile; the offending check-cast will be compiled to handle NULLs. 2325 // If we see more than one offending BCI, then all checkcasts in the 2326 // method will be compiled to handle NULLs. 2327 PreserveJVMState pjvms(this); 2328 set_control(*null_control); 2329 replace_in_map(value, null()); 2330 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2331 uncommon_trap(reason, 2332 Deoptimization::Action_make_not_entrant); 2333 (*null_control) = top(); // NULL path is dead 2334 } 2335 if ((*null_control) == top() && safe_for_replace) { 2336 replace_in_map(value, cast); 2337 } 2338 2339 // Cast away null-ness on the result 2340 return cast; 2341 } 2342 2343 //------------------------------opt_iff---------------------------------------- 2344 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2345 // Return slow-path control. 2346 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2347 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2348 2349 // Fast path taken; set region slot 2 2350 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2351 region->init_req(2,fast_taken); // Capture fast-control 2352 2353 // Fast path not-taken, i.e. slow path 2354 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2355 return slow_taken; 2356 } 2357 2358 //-----------------------------make_runtime_call------------------------------- 2359 Node* GraphKit::make_runtime_call(int flags, 2360 const TypeFunc* call_type, address call_addr, 2361 const char* call_name, 2362 const TypePtr* adr_type, 2363 // The following parms are all optional. 2364 // The first NULL ends the list. 2365 Node* parm0, Node* parm1, 2366 Node* parm2, Node* parm3, 2367 Node* parm4, Node* parm5, 2368 Node* parm6, Node* parm7) { 2369 // Slow-path call 2370 bool is_leaf = !(flags & RC_NO_LEAF); 2371 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2372 if (call_name == NULL) { 2373 assert(!is_leaf, "must supply name for leaf"); 2374 call_name = OptoRuntime::stub_name(call_addr); 2375 } 2376 CallNode* call; 2377 if (!is_leaf) { 2378 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2379 bci(), adr_type); 2380 } else if (flags & RC_NO_FP) { 2381 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2382 } else { 2383 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2384 } 2385 2386 // The following is similar to set_edges_for_java_call, 2387 // except that the memory effects of the call are restricted to AliasIdxRaw. 2388 2389 // Slow path call has no side-effects, uses few values 2390 bool wide_in = !(flags & RC_NARROW_MEM); 2391 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2392 2393 Node* prev_mem = NULL; 2394 if (wide_in) { 2395 prev_mem = set_predefined_input_for_runtime_call(call); 2396 } else { 2397 assert(!wide_out, "narrow in => narrow out"); 2398 Node* narrow_mem = memory(adr_type); 2399 prev_mem = reset_memory(); 2400 map()->set_memory(narrow_mem); 2401 set_predefined_input_for_runtime_call(call); 2402 } 2403 2404 // Hook each parm in order. Stop looking at the first NULL. 2405 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2406 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2407 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2408 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2409 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2410 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2411 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2412 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2413 /* close each nested if ===> */ } } } } } } } } 2414 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2415 2416 if (!is_leaf) { 2417 // Non-leaves can block and take safepoints: 2418 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2419 } 2420 // Non-leaves can throw exceptions: 2421 if (has_io) { 2422 call->set_req(TypeFunc::I_O, i_o()); 2423 } 2424 2425 if (flags & RC_UNCOMMON) { 2426 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2427 // (An "if" probability corresponds roughly to an unconditional count. 2428 // Sort of.) 2429 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2430 } 2431 2432 Node* c = _gvn.transform(call); 2433 assert(c == call, "cannot disappear"); 2434 2435 if (wide_out) { 2436 // Slow path call has full side-effects. 2437 set_predefined_output_for_runtime_call(call); 2438 } else { 2439 // Slow path call has few side-effects, and/or sets few values. 2440 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2441 } 2442 2443 if (has_io) { 2444 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2445 } 2446 return call; 2447 2448 } 2449 2450 //------------------------------merge_memory----------------------------------- 2451 // Merge memory from one path into the current memory state. 2452 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2453 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2454 Node* old_slice = mms.force_memory(); 2455 Node* new_slice = mms.memory2(); 2456 if (old_slice != new_slice) { 2457 PhiNode* phi; 2458 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2459 if (mms.is_empty()) { 2460 // clone base memory Phi's inputs for this memory slice 2461 assert(old_slice == mms.base_memory(), "sanity"); 2462 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2463 _gvn.set_type(phi, Type::MEMORY); 2464 for (uint i = 1; i < phi->req(); i++) { 2465 phi->init_req(i, old_slice->in(i)); 2466 } 2467 } else { 2468 phi = old_slice->as_Phi(); // Phi was generated already 2469 } 2470 } else { 2471 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2472 _gvn.set_type(phi, Type::MEMORY); 2473 } 2474 phi->set_req(new_path, new_slice); 2475 mms.set_memory(phi); 2476 } 2477 } 2478 } 2479 2480 //------------------------------make_slow_call_ex------------------------------ 2481 // Make the exception handler hookups for the slow call 2482 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2483 if (stopped()) return; 2484 2485 // Make a catch node with just two handlers: fall-through and catch-all 2486 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2487 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2488 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2489 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2490 2491 { PreserveJVMState pjvms(this); 2492 set_control(excp); 2493 set_i_o(i_o); 2494 2495 if (excp != top()) { 2496 if (deoptimize) { 2497 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2498 uncommon_trap(Deoptimization::Reason_unhandled, 2499 Deoptimization::Action_none); 2500 } else { 2501 // Create an exception state also. 2502 // Use an exact type if the caller has specified a specific exception. 2503 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2504 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2505 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2506 } 2507 } 2508 } 2509 2510 // Get the no-exception control from the CatchNode. 2511 set_control(norm); 2512 } 2513 2514 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2515 Node* cmp = NULL; 2516 switch(bt) { 2517 case T_INT: cmp = new CmpINode(in1, in2); break; 2518 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2519 default: fatal("unexpected comparison type %s", type2name(bt)); 2520 } 2521 gvn->transform(cmp); 2522 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2523 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2524 gvn->transform(iff); 2525 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2526 return iff; 2527 } 2528 2529 2530 //-------------------------------gen_subtype_check----------------------------- 2531 // Generate a subtyping check. Takes as input the subtype and supertype. 2532 // Returns 2 values: sets the default control() to the true path and returns 2533 // the false path. Only reads invariant memory; sets no (visible) memory. 2534 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2535 // but that's not exposed to the optimizer. This call also doesn't take in an 2536 // Object; if you wish to check an Object you need to load the Object's class 2537 // prior to coming here. 2538 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2539 Compile* C = gvn->C; 2540 2541 if ((*ctrl)->is_top()) { 2542 return C->top(); 2543 } 2544 2545 // Fast check for identical types, perhaps identical constants. 2546 // The types can even be identical non-constants, in cases 2547 // involving Array.newInstance, Object.clone, etc. 2548 if (subklass == superklass) 2549 return C->top(); // false path is dead; no test needed. 2550 2551 if (gvn->type(superklass)->singleton()) { 2552 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2553 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2554 2555 // In the common case of an exact superklass, try to fold up the 2556 // test before generating code. You may ask, why not just generate 2557 // the code and then let it fold up? The answer is that the generated 2558 // code will necessarily include null checks, which do not always 2559 // completely fold away. If they are also needless, then they turn 2560 // into a performance loss. Example: 2561 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2562 // Here, the type of 'fa' is often exact, so the store check 2563 // of fa[1]=x will fold up, without testing the nullness of x. 2564 switch (C->static_subtype_check(superk, subk)) { 2565 case Compile::SSC_always_false: 2566 { 2567 Node* always_fail = *ctrl; 2568 *ctrl = gvn->C->top(); 2569 return always_fail; 2570 } 2571 case Compile::SSC_always_true: 2572 return C->top(); 2573 case Compile::SSC_easy_test: 2574 { 2575 // Just do a direct pointer compare and be done. 2576 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2577 *ctrl = gvn->transform(new IfTrueNode(iff)); 2578 return gvn->transform(new IfFalseNode(iff)); 2579 } 2580 case Compile::SSC_full_test: 2581 break; 2582 default: 2583 ShouldNotReachHere(); 2584 } 2585 } 2586 2587 // %%% Possible further optimization: Even if the superklass is not exact, 2588 // if the subklass is the unique subtype of the superklass, the check 2589 // will always succeed. We could leave a dependency behind to ensure this. 2590 2591 // First load the super-klass's check-offset 2592 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2593 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2594 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2595 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2596 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2597 2598 // Load from the sub-klass's super-class display list, or a 1-word cache of 2599 // the secondary superclass list, or a failing value with a sentinel offset 2600 // if the super-klass is an interface or exceptionally deep in the Java 2601 // hierarchy and we have to scan the secondary superclass list the hard way. 2602 // Worst-case type is a little odd: NULL is allowed as a result (usually 2603 // klass loads can never produce a NULL). 2604 Node *chk_off_X = chk_off; 2605 #ifdef _LP64 2606 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2607 #endif 2608 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2609 // For some types like interfaces the following loadKlass is from a 1-word 2610 // cache which is mutable so can't use immutable memory. Other 2611 // types load from the super-class display table which is immutable. 2612 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2613 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2614 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM)); 2615 2616 // Compile speed common case: ARE a subtype and we canNOT fail 2617 if( superklass == nkls ) 2618 return C->top(); // false path is dead; no test needed. 2619 2620 // See if we get an immediate positive hit. Happens roughly 83% of the 2621 // time. Test to see if the value loaded just previously from the subklass 2622 // is exactly the superklass. 2623 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2624 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2625 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2626 2627 // Compile speed common case: Check for being deterministic right now. If 2628 // chk_off is a constant and not equal to cacheoff then we are NOT a 2629 // subklass. In this case we need exactly the 1 test above and we can 2630 // return those results immediately. 2631 if (!might_be_cache) { 2632 Node* not_subtype_ctrl = *ctrl; 2633 *ctrl = iftrue1; // We need exactly the 1 test above 2634 return not_subtype_ctrl; 2635 } 2636 2637 // Gather the various success & failures here 2638 RegionNode *r_ok_subtype = new RegionNode(4); 2639 gvn->record_for_igvn(r_ok_subtype); 2640 RegionNode *r_not_subtype = new RegionNode(3); 2641 gvn->record_for_igvn(r_not_subtype); 2642 2643 r_ok_subtype->init_req(1, iftrue1); 2644 2645 // Check for immediate negative hit. Happens roughly 11% of the time (which 2646 // is roughly 63% of the remaining cases). Test to see if the loaded 2647 // check-offset points into the subklass display list or the 1-element 2648 // cache. If it points to the display (and NOT the cache) and the display 2649 // missed then it's not a subtype. 2650 Node *cacheoff = gvn->intcon(cacheoff_con); 2651 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2652 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2653 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2654 2655 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2656 // No performance impact (too rare) but allows sharing of secondary arrays 2657 // which has some footprint reduction. 2658 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2659 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2660 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2661 2662 // -- Roads not taken here: -- 2663 // We could also have chosen to perform the self-check at the beginning 2664 // of this code sequence, as the assembler does. This would not pay off 2665 // the same way, since the optimizer, unlike the assembler, can perform 2666 // static type analysis to fold away many successful self-checks. 2667 // Non-foldable self checks work better here in second position, because 2668 // the initial primary superclass check subsumes a self-check for most 2669 // types. An exception would be a secondary type like array-of-interface, 2670 // which does not appear in its own primary supertype display. 2671 // Finally, we could have chosen to move the self-check into the 2672 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2673 // dependent manner. But it is worthwhile to have the check here, 2674 // where it can be perhaps be optimized. The cost in code space is 2675 // small (register compare, branch). 2676 2677 // Now do a linear scan of the secondary super-klass array. Again, no real 2678 // performance impact (too rare) but it's gotta be done. 2679 // Since the code is rarely used, there is no penalty for moving it 2680 // out of line, and it can only improve I-cache density. 2681 // The decision to inline or out-of-line this final check is platform 2682 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2683 Node* psc = gvn->transform( 2684 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2685 2686 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2687 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2688 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2689 2690 // Return false path; set default control to true path. 2691 *ctrl = gvn->transform(r_ok_subtype); 2692 return gvn->transform(r_not_subtype); 2693 } 2694 2695 // Profile-driven exact type check: 2696 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2697 float prob, 2698 Node* *casted_receiver) { 2699 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2700 Node* recv_klass = load_object_klass(receiver); 2701 Node* fail = type_check(recv_klass, tklass, prob); 2702 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2703 assert(recv_xtype->klass_is_exact(), ""); 2704 2705 // Subsume downstream occurrences of receiver with a cast to 2706 // recv_xtype, since now we know what the type will be. 2707 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2708 (*casted_receiver) = _gvn.transform(cast); 2709 // (User must make the replace_in_map call.) 2710 2711 return fail; 2712 } 2713 2714 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 2715 float prob) { 2716 //const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2717 Node* want_klass = makecon(tklass); 2718 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass)); 2719 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2720 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2721 set_control( _gvn.transform( new IfTrueNode (iff))); 2722 Node* fail = _gvn.transform( new IfFalseNode(iff)); 2723 return fail; 2724 } 2725 2726 2727 //------------------------------seems_never_null------------------------------- 2728 // Use null_seen information if it is available from the profile. 2729 // If we see an unexpected null at a type check we record it and force a 2730 // recompile; the offending check will be recompiled to handle NULLs. 2731 // If we see several offending BCIs, then all checks in the 2732 // method will be recompiled. 2733 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2734 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2735 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2736 if (UncommonNullCast // Cutout for this technique 2737 && obj != null() // And not the -Xcomp stupid case? 2738 && !too_many_traps(reason) 2739 ) { 2740 if (speculating) { 2741 return true; 2742 } 2743 if (data == NULL) 2744 // Edge case: no mature data. Be optimistic here. 2745 return true; 2746 // If the profile has not seen a null, assume it won't happen. 2747 assert(java_bc() == Bytecodes::_checkcast || 2748 java_bc() == Bytecodes::_instanceof || 2749 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2750 return !data->as_BitData()->null_seen(); 2751 } 2752 speculating = false; 2753 return false; 2754 } 2755 2756 //------------------------maybe_cast_profiled_receiver------------------------- 2757 // If the profile has seen exactly one type, narrow to exactly that type. 2758 // Subsequent type checks will always fold up. 2759 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2760 ciKlass* require_klass, 2761 ciKlass* spec_klass, 2762 bool safe_for_replace) { 2763 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2764 2765 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2766 2767 // Make sure we haven't already deoptimized from this tactic. 2768 if (too_many_traps(reason) || too_many_recompiles(reason)) 2769 return NULL; 2770 2771 // (No, this isn't a call, but it's enough like a virtual call 2772 // to use the same ciMethod accessor to get the profile info...) 2773 // If we have a speculative type use it instead of profiling (which 2774 // may not help us) 2775 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2776 if (exact_kls != NULL) {// no cast failures here 2777 if (require_klass == NULL || 2778 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2779 // If we narrow the type to match what the type profile sees or 2780 // the speculative type, we can then remove the rest of the 2781 // cast. 2782 // This is a win, even if the exact_kls is very specific, 2783 // because downstream operations, such as method calls, 2784 // will often benefit from the sharper type. 2785 Node* exact_obj = not_null_obj; // will get updated in place... 2786 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2787 &exact_obj); 2788 { PreserveJVMState pjvms(this); 2789 set_control(slow_ctl); 2790 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2791 } 2792 if (safe_for_replace) { 2793 replace_in_map(not_null_obj, exact_obj); 2794 } 2795 return exact_obj; 2796 } 2797 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2798 } 2799 2800 return NULL; 2801 } 2802 2803 /** 2804 * Cast obj to type and emit guard unless we had too many traps here 2805 * already 2806 * 2807 * @param obj node being casted 2808 * @param type type to cast the node to 2809 * @param not_null true if we know node cannot be null 2810 */ 2811 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2812 ciKlass* type, 2813 bool not_null) { 2814 if (stopped()) { 2815 return obj; 2816 } 2817 2818 // type == NULL if profiling tells us this object is always null 2819 if (type != NULL) { 2820 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2821 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2822 2823 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2824 !too_many_traps(class_reason) && 2825 !too_many_recompiles(class_reason)) { 2826 Node* not_null_obj = NULL; 2827 // not_null is true if we know the object is not null and 2828 // there's no need for a null check 2829 if (!not_null) { 2830 Node* null_ctl = top(); 2831 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2832 assert(null_ctl->is_top(), "no null control here"); 2833 } else { 2834 not_null_obj = obj; 2835 } 2836 2837 Node* exact_obj = not_null_obj; 2838 ciKlass* exact_kls = type; 2839 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2840 &exact_obj); 2841 { 2842 PreserveJVMState pjvms(this); 2843 set_control(slow_ctl); 2844 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2845 } 2846 replace_in_map(not_null_obj, exact_obj); 2847 obj = exact_obj; 2848 } 2849 } else { 2850 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2851 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2852 Node* exact_obj = null_assert(obj); 2853 replace_in_map(obj, exact_obj); 2854 obj = exact_obj; 2855 } 2856 } 2857 return obj; 2858 } 2859 2860 //-------------------------------gen_instanceof-------------------------------- 2861 // Generate an instance-of idiom. Used by both the instance-of bytecode 2862 // and the reflective instance-of call. 2863 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2864 kill_dead_locals(); // Benefit all the uncommon traps 2865 assert( !stopped(), "dead parse path should be checked in callers" ); 2866 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2867 "must check for not-null not-dead klass in callers"); 2868 2869 // Make the merge point 2870 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2871 RegionNode* region = new RegionNode(PATH_LIMIT); 2872 Node* phi = new PhiNode(region, TypeInt::BOOL); 2873 C->set_has_split_ifs(true); // Has chance for split-if optimization 2874 2875 ciProfileData* data = NULL; 2876 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2877 data = method()->method_data()->bci_to_data(bci()); 2878 } 2879 bool speculative_not_null = false; 2880 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2881 && seems_never_null(obj, data, speculative_not_null)); 2882 2883 // Null check; get casted pointer; set region slot 3 2884 Node* null_ctl = top(); 2885 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2886 2887 // If not_null_obj is dead, only null-path is taken 2888 if (stopped()) { // Doing instance-of on a NULL? 2889 set_control(null_ctl); 2890 return intcon(0); 2891 } 2892 region->init_req(_null_path, null_ctl); 2893 phi ->init_req(_null_path, intcon(0)); // Set null path value 2894 if (null_ctl == top()) { 2895 // Do this eagerly, so that pattern matches like is_diamond_phi 2896 // will work even during parsing. 2897 assert(_null_path == PATH_LIMIT-1, "delete last"); 2898 region->del_req(_null_path); 2899 phi ->del_req(_null_path); 2900 } 2901 2902 // Do we know the type check always succeed? 2903 bool known_statically = false; 2904 if (_gvn.type(superklass)->singleton()) { 2905 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2906 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2907 if (subk != NULL && subk->is_loaded()) { 2908 int static_res = C->static_subtype_check(superk, subk); 2909 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2910 } 2911 } 2912 2913 if (!known_statically) { 2914 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2915 // We may not have profiling here or it may not help us. If we 2916 // have a speculative type use it to perform an exact cast. 2917 ciKlass* spec_obj_type = obj_type->speculative_type(); 2918 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2919 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2920 if (stopped()) { // Profile disagrees with this path. 2921 set_control(null_ctl); // Null is the only remaining possibility. 2922 return intcon(0); 2923 } 2924 if (cast_obj != NULL) { 2925 not_null_obj = cast_obj; 2926 } 2927 } 2928 } 2929 2930 // Load the object's klass 2931 Node* obj_klass = load_object_klass(not_null_obj); 2932 2933 // Generate the subtype check 2934 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2935 2936 // Plug in the success path to the general merge in slot 1. 2937 region->init_req(_obj_path, control()); 2938 phi ->init_req(_obj_path, intcon(1)); 2939 2940 // Plug in the failing path to the general merge in slot 2. 2941 region->init_req(_fail_path, not_subtype_ctrl); 2942 phi ->init_req(_fail_path, intcon(0)); 2943 2944 // Return final merged results 2945 set_control( _gvn.transform(region) ); 2946 record_for_igvn(region); 2947 2948 // If we know the type check always succeeds then we don't use the 2949 // profiling data at this bytecode. Don't lose it, feed it to the 2950 // type system as a speculative type. 2951 if (safe_for_replace) { 2952 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 2953 replace_in_map(obj, casted_obj); 2954 } 2955 2956 return _gvn.transform(phi); 2957 } 2958 2959 //-------------------------------gen_checkcast--------------------------------- 2960 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2961 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2962 // uncommon-trap paths work. Adjust stack after this call. 2963 // If failure_control is supplied and not null, it is filled in with 2964 // the control edge for the cast failure. Otherwise, an appropriate 2965 // uncommon trap or exception is thrown. 2966 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2967 Node* *failure_control) { 2968 kill_dead_locals(); // Benefit all the uncommon traps 2969 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2970 assert(tk->is_loaded(), "must be loaded"); 2971 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2972 2973 // Fast cutout: Check the case that the cast is vacuously true. 2974 // This detects the common cases where the test will short-circuit 2975 // away completely. We do this before we perform the null check, 2976 // because if the test is going to turn into zero code, we don't 2977 // want a residual null check left around. (Causes a slowdown, 2978 // for example, in some objArray manipulations, such as a[i]=a[j].) 2979 if (tk->singleton()) { 2980 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2981 if (objtp != NULL && objtp->klass() != NULL) { 2982 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2983 case Compile::SSC_always_true: 2984 // If we know the type check always succeed then we don't use 2985 // the profiling data at this bytecode. Don't lose it, feed it 2986 // to the type system as a speculative type. 2987 return record_profiled_receiver_for_speculation(obj); 2988 case Compile::SSC_always_false: 2989 // It needs a null check because a null will *pass* the cast check. 2990 // A non-null value will always produce an exception. 2991 return null_assert(obj); 2992 } 2993 } 2994 } 2995 2996 ciProfileData* data = NULL; 2997 bool safe_for_replace = false; 2998 if (failure_control == NULL) { // use MDO in regular case only 2999 assert(java_bc() == Bytecodes::_aastore || 3000 java_bc() == Bytecodes::_checkcast, 3001 "interpreter profiles type checks only for these BCs"); 3002 data = method()->method_data()->bci_to_data(bci()); 3003 safe_for_replace = true; 3004 } 3005 3006 // Make the merge point 3007 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3008 RegionNode* region = new RegionNode(PATH_LIMIT); 3009 Node* phi = new PhiNode(region, toop); 3010 C->set_has_split_ifs(true); // Has chance for split-if optimization 3011 3012 // Use null-cast information if it is available 3013 bool speculative_not_null = false; 3014 bool never_see_null = ((failure_control == NULL) // regular case only 3015 && seems_never_null(obj, data, speculative_not_null)); 3016 3017 // Null check; get casted pointer; set region slot 3 3018 Node* null_ctl = top(); 3019 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3020 3021 // If not_null_obj is dead, only null-path is taken 3022 if (stopped()) { // Doing instance-of on a NULL? 3023 set_control(null_ctl); 3024 return null(); 3025 } 3026 region->init_req(_null_path, null_ctl); 3027 phi ->init_req(_null_path, null()); // Set null path value 3028 if (null_ctl == top()) { 3029 // Do this eagerly, so that pattern matches like is_diamond_phi 3030 // will work even during parsing. 3031 assert(_null_path == PATH_LIMIT-1, "delete last"); 3032 region->del_req(_null_path); 3033 phi ->del_req(_null_path); 3034 } 3035 3036 Node* cast_obj = NULL; 3037 if (tk->klass_is_exact()) { 3038 // The following optimization tries to statically cast the speculative type of the object 3039 // (for example obtained during profiling) to the type of the superklass and then do a 3040 // dynamic check that the type of the object is what we expect. To work correctly 3041 // for checkcast and aastore the type of superklass should be exact. 3042 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3043 // We may not have profiling here or it may not help us. If we have 3044 // a speculative type use it to perform an exact cast. 3045 ciKlass* spec_obj_type = obj_type->speculative_type(); 3046 if (spec_obj_type != NULL || data != NULL) { 3047 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3048 if (cast_obj != NULL) { 3049 if (failure_control != NULL) // failure is now impossible 3050 (*failure_control) = top(); 3051 // adjust the type of the phi to the exact klass: 3052 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3053 } 3054 } 3055 } 3056 3057 if (cast_obj == NULL) { 3058 // Load the object's klass 3059 Node* obj_klass = load_object_klass(not_null_obj); 3060 3061 // Generate the subtype check 3062 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3063 3064 // Plug in success path into the merge 3065 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3066 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3067 if (failure_control == NULL) { 3068 if (not_subtype_ctrl != top()) { // If failure is possible 3069 PreserveJVMState pjvms(this); 3070 set_control(not_subtype_ctrl); 3071 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3072 } 3073 } else { 3074 (*failure_control) = not_subtype_ctrl; 3075 } 3076 } 3077 3078 region->init_req(_obj_path, control()); 3079 phi ->init_req(_obj_path, cast_obj); 3080 3081 // A merge of NULL or Casted-NotNull obj 3082 Node* res = _gvn.transform(phi); 3083 3084 // Note I do NOT always 'replace_in_map(obj,result)' here. 3085 // if( tk->klass()->can_be_primary_super() ) 3086 // This means that if I successfully store an Object into an array-of-String 3087 // I 'forget' that the Object is really now known to be a String. I have to 3088 // do this because we don't have true union types for interfaces - if I store 3089 // a Baz into an array-of-Interface and then tell the optimizer it's an 3090 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3091 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3092 // replace_in_map( obj, res ); 3093 3094 // Return final merged results 3095 set_control( _gvn.transform(region) ); 3096 record_for_igvn(region); 3097 3098 return record_profiled_receiver_for_speculation(res); 3099 } 3100 3101 //------------------------------next_monitor----------------------------------- 3102 // What number should be given to the next monitor? 3103 int GraphKit::next_monitor() { 3104 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3105 int next = current + C->sync_stack_slots(); 3106 // Keep the toplevel high water mark current: 3107 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3108 return current; 3109 } 3110 3111 //------------------------------insert_mem_bar--------------------------------- 3112 // Memory barrier to avoid floating things around 3113 // The membar serves as a pinch point between both control and all memory slices. 3114 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3115 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3116 mb->init_req(TypeFunc::Control, control()); 3117 mb->init_req(TypeFunc::Memory, reset_memory()); 3118 Node* membar = _gvn.transform(mb); 3119 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3120 set_all_memory_call(membar); 3121 return membar; 3122 } 3123 3124 //-------------------------insert_mem_bar_volatile---------------------------- 3125 // Memory barrier to avoid floating things around 3126 // The membar serves as a pinch point between both control and memory(alias_idx). 3127 // If you want to make a pinch point on all memory slices, do not use this 3128 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3129 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3130 // When Parse::do_put_xxx updates a volatile field, it appends a series 3131 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3132 // The first membar is on the same memory slice as the field store opcode. 3133 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3134 // All the other membars (for other volatile slices, including AliasIdxBot, 3135 // which stands for all unknown volatile slices) are control-dependent 3136 // on the first membar. This prevents later volatile loads or stores 3137 // from sliding up past the just-emitted store. 3138 3139 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3140 mb->set_req(TypeFunc::Control,control()); 3141 if (alias_idx == Compile::AliasIdxBot) { 3142 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3143 } else { 3144 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3145 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3146 } 3147 Node* membar = _gvn.transform(mb); 3148 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3149 if (alias_idx == Compile::AliasIdxBot) { 3150 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3151 } else { 3152 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3153 } 3154 return membar; 3155 } 3156 3157 //------------------------------shared_lock------------------------------------ 3158 // Emit locking code. 3159 FastLockNode* GraphKit::shared_lock(Node* obj) { 3160 // bci is either a monitorenter bc or InvocationEntryBci 3161 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3162 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3163 3164 if( !GenerateSynchronizationCode ) 3165 return NULL; // Not locking things? 3166 if (stopped()) // Dead monitor? 3167 return NULL; 3168 3169 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3170 3171 // Box the stack location 3172 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3173 Node* mem = reset_memory(); 3174 3175 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3176 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3177 // Create the counters for this fast lock. 3178 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3179 } 3180 3181 // Create the rtm counters for this fast lock if needed. 3182 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3183 3184 // Add monitor to debug info for the slow path. If we block inside the 3185 // slow path and de-opt, we need the monitor hanging around 3186 map()->push_monitor( flock ); 3187 3188 const TypeFunc *tf = LockNode::lock_type(); 3189 LockNode *lock = new LockNode(C, tf); 3190 3191 lock->init_req( TypeFunc::Control, control() ); 3192 lock->init_req( TypeFunc::Memory , mem ); 3193 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3194 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3195 lock->init_req( TypeFunc::ReturnAdr, top() ); 3196 3197 lock->init_req(TypeFunc::Parms + 0, obj); 3198 lock->init_req(TypeFunc::Parms + 1, box); 3199 lock->init_req(TypeFunc::Parms + 2, flock); 3200 add_safepoint_edges(lock); 3201 3202 lock = _gvn.transform( lock )->as_Lock(); 3203 3204 // lock has no side-effects, sets few values 3205 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3206 3207 insert_mem_bar(Op_MemBarAcquireLock); 3208 3209 // Add this to the worklist so that the lock can be eliminated 3210 record_for_igvn(lock); 3211 3212 #ifndef PRODUCT 3213 if (PrintLockStatistics) { 3214 // Update the counter for this lock. Don't bother using an atomic 3215 // operation since we don't require absolute accuracy. 3216 lock->create_lock_counter(map()->jvms()); 3217 increment_counter(lock->counter()->addr()); 3218 } 3219 #endif 3220 3221 return flock; 3222 } 3223 3224 3225 //------------------------------shared_unlock---------------------------------- 3226 // Emit unlocking code. 3227 void GraphKit::shared_unlock(Node* box, Node* obj) { 3228 // bci is either a monitorenter bc or InvocationEntryBci 3229 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3230 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3231 3232 if( !GenerateSynchronizationCode ) 3233 return; 3234 if (stopped()) { // Dead monitor? 3235 map()->pop_monitor(); // Kill monitor from debug info 3236 return; 3237 } 3238 3239 // Memory barrier to avoid floating things down past the locked region 3240 insert_mem_bar(Op_MemBarReleaseLock); 3241 3242 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3243 UnlockNode *unlock = new UnlockNode(C, tf); 3244 #ifdef ASSERT 3245 unlock->set_dbg_jvms(sync_jvms()); 3246 #endif 3247 uint raw_idx = Compile::AliasIdxRaw; 3248 unlock->init_req( TypeFunc::Control, control() ); 3249 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3250 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3251 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3252 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3253 3254 unlock->init_req(TypeFunc::Parms + 0, obj); 3255 unlock->init_req(TypeFunc::Parms + 1, box); 3256 unlock = _gvn.transform(unlock)->as_Unlock(); 3257 3258 Node* mem = reset_memory(); 3259 3260 // unlock has no side-effects, sets few values 3261 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3262 3263 // Kill monitor from debug info 3264 map()->pop_monitor( ); 3265 } 3266 3267 //-------------------------------get_layout_helper----------------------------- 3268 // If the given klass is a constant or known to be an array, 3269 // fetch the constant layout helper value into constant_value 3270 // and return (Node*)NULL. Otherwise, load the non-constant 3271 // layout helper value, and return the node which represents it. 3272 // This two-faced routine is useful because allocation sites 3273 // almost always feature constant types. 3274 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3275 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3276 if (!StressReflectiveCode && inst_klass != NULL) { 3277 ciKlass* klass = inst_klass->klass(); 3278 assert(klass != NULL, "klass should not be NULL"); 3279 bool xklass = inst_klass->klass_is_exact(); 3280 if (xklass || klass->is_array_klass()) { 3281 jint lhelper = klass->layout_helper(); 3282 if (lhelper != Klass::_lh_neutral_value) { 3283 constant_value = lhelper; 3284 return (Node*) NULL; 3285 } 3286 } 3287 } 3288 constant_value = Klass::_lh_neutral_value; // put in a known value 3289 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3290 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3291 } 3292 3293 // We just put in an allocate/initialize with a big raw-memory effect. 3294 // Hook selected additional alias categories on the initialization. 3295 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3296 MergeMemNode* init_in_merge, 3297 Node* init_out_raw) { 3298 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3299 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3300 3301 Node* prevmem = kit.memory(alias_idx); 3302 init_in_merge->set_memory_at(alias_idx, prevmem); 3303 kit.set_memory(init_out_raw, alias_idx); 3304 } 3305 3306 //---------------------------set_output_for_allocation------------------------- 3307 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3308 const TypeOopPtr* oop_type, 3309 bool deoptimize_on_exception) { 3310 int rawidx = Compile::AliasIdxRaw; 3311 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3312 add_safepoint_edges(alloc); 3313 Node* allocx = _gvn.transform(alloc); 3314 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3315 // create memory projection for i_o 3316 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3317 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3318 3319 // create a memory projection as for the normal control path 3320 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3321 set_memory(malloc, rawidx); 3322 3323 // a normal slow-call doesn't change i_o, but an allocation does 3324 // we create a separate i_o projection for the normal control path 3325 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3326 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3327 3328 // put in an initialization barrier 3329 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3330 rawoop)->as_Initialize(); 3331 assert(alloc->initialization() == init, "2-way macro link must work"); 3332 assert(init ->allocation() == alloc, "2-way macro link must work"); 3333 { 3334 // Extract memory strands which may participate in the new object's 3335 // initialization, and source them from the new InitializeNode. 3336 // This will allow us to observe initializations when they occur, 3337 // and link them properly (as a group) to the InitializeNode. 3338 assert(init->in(InitializeNode::Memory) == malloc, ""); 3339 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3340 init->set_req(InitializeNode::Memory, minit_in); 3341 record_for_igvn(minit_in); // fold it up later, if possible 3342 Node* minit_out = memory(rawidx); 3343 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3344 if (oop_type->isa_aryptr()) { 3345 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3346 int elemidx = C->get_alias_index(telemref); 3347 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3348 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) { 3349 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3350 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3351 ciField* field = ik->nonstatic_field_at(i); 3352 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3353 continue; // do not bother to track really large numbers of fields 3354 // Find (or create) the alias category for this field: 3355 int fieldidx = C->alias_type(field)->index(); 3356 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3357 } 3358 } 3359 } 3360 3361 // Cast raw oop to the real thing... 3362 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3363 javaoop = _gvn.transform(javaoop); 3364 C->set_recent_alloc(control(), javaoop); 3365 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3366 3367 #ifdef ASSERT 3368 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3369 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3370 "Ideal_allocation works"); 3371 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3372 "Ideal_allocation works"); 3373 if (alloc->is_AllocateArray()) { 3374 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3375 "Ideal_allocation works"); 3376 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3377 "Ideal_allocation works"); 3378 } else { 3379 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3380 } 3381 } 3382 #endif //ASSERT 3383 3384 return javaoop; 3385 } 3386 3387 //---------------------------new_instance-------------------------------------- 3388 // This routine takes a klass_node which may be constant (for a static type) 3389 // or may be non-constant (for reflective code). It will work equally well 3390 // for either, and the graph will fold nicely if the optimizer later reduces 3391 // the type to a constant. 3392 // The optional arguments are for specialized use by intrinsics: 3393 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3394 // - If 'return_size_val', report the the total object size to the caller. 3395 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3396 Node* GraphKit::new_instance(Node* klass_node, 3397 Node* extra_slow_test, 3398 Node* *return_size_val, 3399 bool deoptimize_on_exception, 3400 ValueTypeBaseNode* value_node) { 3401 // Compute size in doublewords 3402 // The size is always an integral number of doublewords, represented 3403 // as a positive bytewise size stored in the klass's layout_helper. 3404 // The layout_helper also encodes (in a low bit) the need for a slow path. 3405 jint layout_con = Klass::_lh_neutral_value; 3406 Node* layout_val = get_layout_helper(klass_node, layout_con); 3407 bool layout_is_con = (layout_val == NULL); 3408 3409 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3410 // Generate the initial go-slow test. It's either ALWAYS (return a 3411 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3412 // case) a computed value derived from the layout_helper. 3413 Node* initial_slow_test = NULL; 3414 if (layout_is_con) { 3415 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3416 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3417 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3418 } else { // reflective case 3419 // This reflective path is used by Unsafe.allocateInstance. 3420 // (It may be stress-tested by specifying StressReflectiveCode.) 3421 // Basically, we want to get into the VM is there's an illegal argument. 3422 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3423 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3424 if (extra_slow_test != intcon(0)) { 3425 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3426 } 3427 // (Macro-expander will further convert this to a Bool, if necessary.) 3428 } 3429 3430 // Find the size in bytes. This is easy; it's the layout_helper. 3431 // The size value must be valid even if the slow path is taken. 3432 Node* size = NULL; 3433 if (layout_is_con) { 3434 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3435 } else { // reflective case 3436 // This reflective path is used by clone and Unsafe.allocateInstance. 3437 size = ConvI2X(layout_val); 3438 3439 // Clear the low bits to extract layout_helper_size_in_bytes: 3440 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3441 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3442 size = _gvn.transform( new AndXNode(size, mask) ); 3443 } 3444 if (return_size_val != NULL) { 3445 (*return_size_val) = size; 3446 } 3447 3448 // This is a precise notnull oop of the klass. 3449 // (Actually, it need not be precise if this is a reflective allocation.) 3450 // It's what we cast the result to. 3451 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3452 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3453 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3454 3455 // Now generate allocation code 3456 3457 // The entire memory state is needed for slow path of the allocation 3458 // since GC and deoptimization can happen. 3459 Node *mem = reset_memory(); 3460 set_all_memory(mem); // Create new memory state 3461 3462 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3463 control(), mem, i_o(), 3464 size, klass_node, 3465 initial_slow_test, value_node); 3466 3467 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3468 } 3469 3470 //-------------------------------new_array------------------------------------- 3471 // helper for newarray and anewarray 3472 // The 'length' parameter is (obviously) the length of the array. 3473 // See comments on new_instance for the meaning of the other arguments. 3474 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3475 Node* length, // number of array elements 3476 int nargs, // number of arguments to push back for uncommon trap 3477 Node* *return_size_val, 3478 bool deoptimize_on_exception) { 3479 jint layout_con = Klass::_lh_neutral_value; 3480 Node* layout_val = get_layout_helper(klass_node, layout_con); 3481 bool layout_is_con = (layout_val == NULL); 3482 3483 if (!layout_is_con && !StressReflectiveCode && 3484 !too_many_traps(Deoptimization::Reason_class_check)) { 3485 // This is a reflective array creation site. 3486 // Optimistically assume that it is a subtype of Object[], 3487 // so that we can fold up all the address arithmetic. 3488 layout_con = Klass::array_layout_helper(T_OBJECT); 3489 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3490 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3491 { BuildCutout unless(this, bol_lh, PROB_MAX); 3492 inc_sp(nargs); 3493 uncommon_trap(Deoptimization::Reason_class_check, 3494 Deoptimization::Action_maybe_recompile); 3495 } 3496 layout_val = NULL; 3497 layout_is_con = true; 3498 } 3499 3500 // Generate the initial go-slow test. Make sure we do not overflow 3501 // if length is huge (near 2Gig) or negative! We do not need 3502 // exact double-words here, just a close approximation of needed 3503 // double-words. We can't add any offset or rounding bits, lest we 3504 // take a size -1 of bytes and make it positive. Use an unsigned 3505 // compare, so negative sizes look hugely positive. 3506 int fast_size_limit = FastAllocateSizeLimit; 3507 if (layout_is_con) { 3508 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3509 // Increase the size limit if we have exact knowledge of array type. 3510 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3511 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3512 } 3513 3514 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3515 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3516 3517 // --- Size Computation --- 3518 // array_size = round_to_heap(array_header + (length << elem_shift)); 3519 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3520 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3521 // The rounding mask is strength-reduced, if possible. 3522 int round_mask = MinObjAlignmentInBytes - 1; 3523 Node* header_size = NULL; 3524 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3525 // (T_BYTE has the weakest alignment and size restrictions...) 3526 if (layout_is_con) { 3527 int hsize = Klass::layout_helper_header_size(layout_con); 3528 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3529 BasicType etype = Klass::layout_helper_element_type(layout_con); 3530 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con); 3531 if ((round_mask & ~right_n_bits(eshift)) == 0) 3532 round_mask = 0; // strength-reduce it if it goes away completely 3533 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3534 assert(header_size_min <= hsize, "generic minimum is smallest"); 3535 header_size_min = hsize; 3536 header_size = intcon(hsize + round_mask); 3537 } else { 3538 Node* hss = intcon(Klass::_lh_header_size_shift); 3539 Node* hsm = intcon(Klass::_lh_header_size_mask); 3540 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3541 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3542 Node* mask = intcon(round_mask); 3543 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3544 } 3545 3546 Node* elem_shift = NULL; 3547 if (layout_is_con) { 3548 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3549 if (eshift != 0) 3550 elem_shift = intcon(eshift); 3551 } else { 3552 // There is no need to mask or shift this value. 3553 // The semantics of LShiftINode include an implicit mask to 0x1F. 3554 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3555 elem_shift = layout_val; 3556 } 3557 3558 // Transition to native address size for all offset calculations: 3559 Node* lengthx = ConvI2X(length); 3560 Node* headerx = ConvI2X(header_size); 3561 #ifdef _LP64 3562 { const TypeInt* tilen = _gvn.find_int_type(length); 3563 if (tilen != NULL && tilen->_lo < 0) { 3564 // Add a manual constraint to a positive range. Cf. array_element_address. 3565 jint size_max = fast_size_limit; 3566 if (size_max > tilen->_hi) size_max = tilen->_hi; 3567 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3568 3569 // Only do a narrow I2L conversion if the range check passed. 3570 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3571 _gvn.transform(iff); 3572 RegionNode* region = new RegionNode(3); 3573 _gvn.set_type(region, Type::CONTROL); 3574 lengthx = new PhiNode(region, TypeLong::LONG); 3575 _gvn.set_type(lengthx, TypeLong::LONG); 3576 3577 // Range check passed. Use ConvI2L node with narrow type. 3578 Node* passed = IfFalse(iff); 3579 region->init_req(1, passed); 3580 // Make I2L conversion control dependent to prevent it from 3581 // floating above the range check during loop optimizations. 3582 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3583 3584 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3585 region->init_req(2, IfTrue(iff)); 3586 lengthx->init_req(2, ConvI2X(length)); 3587 3588 set_control(region); 3589 record_for_igvn(region); 3590 record_for_igvn(lengthx); 3591 } 3592 } 3593 #endif 3594 3595 // Combine header size (plus rounding) and body size. Then round down. 3596 // This computation cannot overflow, because it is used only in two 3597 // places, one where the length is sharply limited, and the other 3598 // after a successful allocation. 3599 Node* abody = lengthx; 3600 if (elem_shift != NULL) 3601 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3602 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3603 if (round_mask != 0) { 3604 Node* mask = MakeConX(~round_mask); 3605 size = _gvn.transform( new AndXNode(size, mask) ); 3606 } 3607 // else if round_mask == 0, the size computation is self-rounding 3608 3609 if (return_size_val != NULL) { 3610 // This is the size 3611 (*return_size_val) = size; 3612 } 3613 3614 // Now generate allocation code 3615 3616 // The entire memory state is needed for slow path of the allocation 3617 // since GC and deoptimization can happen. 3618 Node *mem = reset_memory(); 3619 set_all_memory(mem); // Create new memory state 3620 3621 if (initial_slow_test->is_Bool()) { 3622 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3623 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3624 } 3625 3626 // Create the AllocateArrayNode and its result projections 3627 AllocateArrayNode* alloc 3628 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3629 control(), mem, i_o(), 3630 size, klass_node, 3631 initial_slow_test, 3632 length); 3633 3634 // Cast to correct type. Note that the klass_node may be constant or not, 3635 // and in the latter case the actual array type will be inexact also. 3636 // (This happens via a non-constant argument to inline_native_newArray.) 3637 // In any case, the value of klass_node provides the desired array type. 3638 const TypeInt* length_type = _gvn.find_int_type(length); 3639 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3640 if (ary_type->isa_aryptr() && length_type != NULL) { 3641 // Try to get a better type than POS for the size 3642 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3643 } 3644 3645 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3646 3647 // Cast length on remaining path to be as narrow as possible 3648 if (map()->find_edge(length) >= 0) { 3649 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3650 if (ccast != length) { 3651 _gvn.set_type_bottom(ccast); 3652 record_for_igvn(ccast); 3653 replace_in_map(length, ccast); 3654 } 3655 } 3656 3657 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 3658 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL; 3659 if (elem_klass != NULL && elem_klass->is_valuetype()) { 3660 ciValueKlass* vk = elem_klass->as_value_klass(); 3661 if (!vk->flatten_array()) { 3662 // Non-flattened value type arrays need to be initialized with default value type oops 3663 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs); 3664 InitializeNode* init = alloc->initialization(); 3665 init->set_complete_with_arraycopy(); 3666 } 3667 } 3668 3669 return javaoop; 3670 } 3671 3672 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) { 3673 // Check for zero length 3674 Node* null_ctl = top(); 3675 null_check_common(length, T_INT, false, &null_ctl, false); 3676 if (stopped()) { 3677 set_control(null_ctl); // Always zero 3678 return; 3679 } 3680 3681 RegionNode* res_ctl = new RegionNode(3); 3682 gvn().set_type(res_ctl, Type::CONTROL); 3683 record_for_igvn(res_ctl); 3684 3685 // Length is zero: don't execute initialization loop 3686 res_ctl->init_req(1, null_ctl); 3687 PhiNode* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO); 3688 PhiNode* res_mem = PhiNode::make(res_ctl, merged_memory(), Type::MEMORY, TypePtr::BOTTOM); 3689 gvn().set_type(res_io, Type::ABIO); 3690 gvn().set_type(res_mem, Type::MEMORY); 3691 record_for_igvn(res_io); 3692 record_for_igvn(res_mem); 3693 3694 // Length is non-zero: execute a loop that initializes the array with the default value type 3695 Node* oop = ValueTypeNode::make_default(gvn(), vk); 3696 oop = oop->as_ValueType()->allocate(this)->get_oop(); 3697 3698 add_predicate(nargs); 3699 RegionNode* loop = new RegionNode(3); 3700 loop->init_req(1, control()); 3701 PhiNode* index = PhiNode::make(loop, intcon(0), TypeInt::INT); 3702 PhiNode* mem = PhiNode::make(loop, reset_memory(), Type::MEMORY, TypePtr::BOTTOM); 3703 3704 gvn().set_type(loop, Type::CONTROL); 3705 gvn().set_type(index, TypeInt::INT); 3706 gvn().set_type(mem, Type::MEMORY); 3707 record_for_igvn(loop); 3708 record_for_igvn(index); 3709 record_for_igvn(mem); 3710 3711 // Loop body: initialize array element at 'index' 3712 set_control(loop); 3713 set_all_memory(mem); 3714 Node* adr = array_element_address(array, index, T_OBJECT); 3715 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk); 3716 store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_VALUETYPE, MemNode::release); 3717 3718 // Check if we need to execute another loop iteration 3719 length = SubI(length, intcon(1)); 3720 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 3721 3722 // Continue with next iteration 3723 loop->init_req(2, IfTrue(iff)); 3724 index->init_req(2, AddI(index, intcon(1))); 3725 mem->init_req(2, merged_memory()); 3726 3727 // Exit loop 3728 res_ctl->init_req(2, IfFalse(iff)); 3729 res_io->set_req(2, i_o()); 3730 res_mem->set_req(2, reset_memory()); 3731 3732 // Set merged control, IO and memory 3733 set_control(res_ctl); 3734 set_i_o(res_io); 3735 set_all_memory(res_mem); 3736 } 3737 3738 // The following "Ideal_foo" functions are placed here because they recognize 3739 // the graph shapes created by the functions immediately above. 3740 3741 //---------------------------Ideal_allocation---------------------------------- 3742 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3743 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3744 if (ptr == NULL) { // reduce dumb test in callers 3745 return NULL; 3746 } 3747 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3748 ptr = ptr->in(1); 3749 if (ptr == NULL) return NULL; 3750 } 3751 // Return NULL for allocations with several casts: 3752 // j.l.reflect.Array.newInstance(jobject, jint) 3753 // Object.clone() 3754 // to keep more precise type from last cast. 3755 if (ptr->is_Proj()) { 3756 Node* allo = ptr->in(0); 3757 if (allo != NULL && allo->is_Allocate()) { 3758 return allo->as_Allocate(); 3759 } 3760 } 3761 // Report failure to match. 3762 return NULL; 3763 } 3764 3765 // Fancy version which also strips off an offset (and reports it to caller). 3766 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3767 intptr_t& offset) { 3768 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3769 if (base == NULL) return NULL; 3770 return Ideal_allocation(base, phase); 3771 } 3772 3773 // Trace Initialize <- Proj[Parm] <- Allocate 3774 AllocateNode* InitializeNode::allocation() { 3775 Node* rawoop = in(InitializeNode::RawAddress); 3776 if (rawoop->is_Proj()) { 3777 Node* alloc = rawoop->in(0); 3778 if (alloc->is_Allocate()) { 3779 return alloc->as_Allocate(); 3780 } 3781 } 3782 return NULL; 3783 } 3784 3785 // Trace Allocate -> Proj[Parm] -> Initialize 3786 InitializeNode* AllocateNode::initialization() { 3787 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3788 if (rawoop == NULL) return NULL; 3789 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3790 Node* init = rawoop->fast_out(i); 3791 if (init->is_Initialize()) { 3792 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3793 return init->as_Initialize(); 3794 } 3795 } 3796 return NULL; 3797 } 3798 3799 //----------------------------- loop predicates --------------------------- 3800 3801 //------------------------------add_predicate_impl---------------------------- 3802 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3803 // Too many traps seen? 3804 if (too_many_traps(reason)) { 3805 #ifdef ASSERT 3806 if (TraceLoopPredicate) { 3807 int tc = C->trap_count(reason); 3808 tty->print("too many traps=%s tcount=%d in ", 3809 Deoptimization::trap_reason_name(reason), tc); 3810 method()->print(); // which method has too many predicate traps 3811 tty->cr(); 3812 } 3813 #endif 3814 // We cannot afford to take more traps here, 3815 // do not generate predicate. 3816 return; 3817 } 3818 3819 Node *cont = _gvn.intcon(1); 3820 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3821 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3822 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3823 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3824 C->add_predicate_opaq(opq); 3825 { 3826 PreserveJVMState pjvms(this); 3827 set_control(iffalse); 3828 inc_sp(nargs); 3829 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3830 } 3831 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3832 set_control(iftrue); 3833 } 3834 3835 //------------------------------add_predicate--------------------------------- 3836 void GraphKit::add_predicate(int nargs) { 3837 if (UseLoopPredicate) { 3838 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3839 } 3840 // loop's limit check predicate should be near the loop. 3841 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3842 } 3843 3844 //----------------------------- store barriers ---------------------------- 3845 #define __ ideal. 3846 3847 void GraphKit::sync_kit(IdealKit& ideal) { 3848 set_all_memory(__ merged_memory()); 3849 set_i_o(__ i_o()); 3850 set_control(__ ctrl()); 3851 } 3852 3853 void GraphKit::final_sync(IdealKit& ideal) { 3854 // Final sync IdealKit and graphKit. 3855 sync_kit(ideal); 3856 } 3857 3858 Node* GraphKit::byte_map_base_node() { 3859 // Get base of card map 3860 CardTableModRefBS* ct = 3861 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3862 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3863 if (ct->byte_map_base != NULL) { 3864 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3865 } else { 3866 return null(); 3867 } 3868 } 3869 3870 // vanilla/CMS post barrier 3871 // Insert a write-barrier store. This is to let generational GC work; we have 3872 // to flag all oop-stores before the next GC point. 3873 void GraphKit::write_barrier_post(Node* oop_store, 3874 Node* obj, 3875 Node* adr, 3876 uint adr_idx, 3877 Node* val, 3878 bool use_precise) { 3879 // No store check needed if we're storing a NULL or an old object 3880 // (latter case is probably a string constant). The concurrent 3881 // mark sweep garbage collector, however, needs to have all nonNull 3882 // oop updates flagged via card-marks. 3883 if (val != NULL && val->is_Con()) { 3884 // must be either an oop or NULL 3885 const Type* t = val->bottom_type(); 3886 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3887 // stores of null never (?) need barriers 3888 return; 3889 } 3890 3891 if (use_ReduceInitialCardMarks() 3892 && obj == just_allocated_object(control())) { 3893 // We can skip marks on a freshly-allocated object in Eden. 3894 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3895 // That routine informs GC to take appropriate compensating steps, 3896 // upon a slow-path allocation, so as to make this card-mark 3897 // elision safe. 3898 return; 3899 } 3900 3901 if (!use_precise) { 3902 // All card marks for a (non-array) instance are in one place: 3903 adr = obj; 3904 } 3905 // (Else it's an array (or unknown), and we want more precise card marks.) 3906 assert(adr != NULL, ""); 3907 3908 IdealKit ideal(this, true); 3909 3910 // Convert the pointer to an int prior to doing math on it 3911 Node* cast = __ CastPX(__ ctrl(), adr); 3912 3913 // Divide by card size 3914 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3915 "Only one we handle so far."); 3916 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3917 3918 // Combine card table base and card offset 3919 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3920 3921 // Get the alias_index for raw card-mark memory 3922 int adr_type = Compile::AliasIdxRaw; 3923 Node* zero = __ ConI(0); // Dirty card value 3924 BasicType bt = T_BYTE; 3925 3926 if (UseConcMarkSweepGC && UseCondCardMark) { 3927 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3928 __ sync_kit(this); 3929 } 3930 3931 if (UseCondCardMark) { 3932 // The classic GC reference write barrier is typically implemented 3933 // as a store into the global card mark table. Unfortunately 3934 // unconditional stores can result in false sharing and excessive 3935 // coherence traffic as well as false transactional aborts. 3936 // UseCondCardMark enables MP "polite" conditional card mark 3937 // stores. In theory we could relax the load from ctrl() to 3938 // no_ctrl, but that doesn't buy much latitude. 3939 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3940 __ if_then(card_val, BoolTest::ne, zero); 3941 } 3942 3943 // Smash zero into card 3944 if( !UseConcMarkSweepGC ) { 3945 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3946 } else { 3947 // Specialized path for CM store barrier 3948 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3949 } 3950 3951 if (UseCondCardMark) { 3952 __ end_if(); 3953 } 3954 3955 // Final sync IdealKit and GraphKit. 3956 final_sync(ideal); 3957 } 3958 /* 3959 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3960 * required by SATB to make sure all objects live at the start of the 3961 * marking are kept alive, all reference updates need to any previous 3962 * reference stored before writing. 3963 * 3964 * If the previous value is NULL there is no need to save the old value. 3965 * References that are NULL are filtered during runtime by the barrier 3966 * code to avoid unnecessary queuing. 3967 * 3968 * However in the case of newly allocated objects it might be possible to 3969 * prove that the reference about to be overwritten is NULL during compile 3970 * time and avoid adding the barrier code completely. 3971 * 3972 * The compiler needs to determine that the object in which a field is about 3973 * to be written is newly allocated, and that no prior store to the same field 3974 * has happened since the allocation. 3975 * 3976 * Returns true if the pre-barrier can be removed 3977 */ 3978 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3979 BasicType bt, uint adr_idx) { 3980 intptr_t offset = 0; 3981 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3982 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3983 3984 if (offset == Type::OffsetBot) { 3985 return false; // cannot unalias unless there are precise offsets 3986 } 3987 3988 if (alloc == NULL) { 3989 return false; // No allocation found 3990 } 3991 3992 intptr_t size_in_bytes = type2aelembytes(bt); 3993 3994 Node* mem = memory(adr_idx); // start searching here... 3995 3996 for (int cnt = 0; cnt < 50; cnt++) { 3997 3998 if (mem->is_Store()) { 3999 4000 Node* st_adr = mem->in(MemNode::Address); 4001 intptr_t st_offset = 0; 4002 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 4003 4004 if (st_base == NULL) { 4005 break; // inscrutable pointer 4006 } 4007 4008 // Break we have found a store with same base and offset as ours so break 4009 if (st_base == base && st_offset == offset) { 4010 break; 4011 } 4012 4013 if (st_offset != offset && st_offset != Type::OffsetBot) { 4014 const int MAX_STORE = BytesPerLong; 4015 if (st_offset >= offset + size_in_bytes || 4016 st_offset <= offset - MAX_STORE || 4017 st_offset <= offset - mem->as_Store()->memory_size()) { 4018 // Success: The offsets are provably independent. 4019 // (You may ask, why not just test st_offset != offset and be done? 4020 // The answer is that stores of different sizes can co-exist 4021 // in the same sequence of RawMem effects. We sometimes initialize 4022 // a whole 'tile' of array elements with a single jint or jlong.) 4023 mem = mem->in(MemNode::Memory); 4024 continue; // advance through independent store memory 4025 } 4026 } 4027 4028 if (st_base != base 4029 && MemNode::detect_ptr_independence(base, alloc, st_base, 4030 AllocateNode::Ideal_allocation(st_base, phase), 4031 phase)) { 4032 // Success: The bases are provably independent. 4033 mem = mem->in(MemNode::Memory); 4034 continue; // advance through independent store memory 4035 } 4036 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4037 4038 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4039 AllocateNode* st_alloc = st_init->allocation(); 4040 4041 // Make sure that we are looking at the same allocation site. 4042 // The alloc variable is guaranteed to not be null here from earlier check. 4043 if (alloc == st_alloc) { 4044 // Check that the initialization is storing NULL so that no previous store 4045 // has been moved up and directly write a reference 4046 Node* captured_store = st_init->find_captured_store(offset, 4047 type2aelembytes(T_OBJECT), 4048 phase); 4049 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4050 return true; 4051 } 4052 } 4053 } 4054 4055 // Unless there is an explicit 'continue', we must bail out here, 4056 // because 'mem' is an inscrutable memory state (e.g., a call). 4057 break; 4058 } 4059 4060 return false; 4061 } 4062 4063 // G1 pre/post barriers 4064 void GraphKit::g1_write_barrier_pre(bool do_load, 4065 Node* obj, 4066 Node* adr, 4067 uint alias_idx, 4068 Node* val, 4069 const TypeOopPtr* val_type, 4070 Node* pre_val, 4071 BasicType bt) { 4072 4073 // Some sanity checks 4074 // Note: val is unused in this routine. 4075 4076 if (do_load) { 4077 // We need to generate the load of the previous value 4078 assert(obj != NULL, "must have a base"); 4079 assert(adr != NULL, "where are loading from?"); 4080 assert(pre_val == NULL, "loaded already?"); 4081 assert(val_type != NULL, "need a type"); 4082 4083 if (use_ReduceInitialCardMarks() 4084 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4085 return; 4086 } 4087 4088 } else { 4089 // In this case both val_type and alias_idx are unused. 4090 assert(pre_val != NULL, "must be loaded already"); 4091 // Nothing to be done if pre_val is null. 4092 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4093 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4094 } 4095 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here"); 4096 4097 IdealKit ideal(this, true); 4098 4099 Node* tls = __ thread(); // ThreadLocalStorage 4100 4101 Node* no_ctrl = NULL; 4102 Node* no_base = __ top(); 4103 Node* zero = __ ConI(0); 4104 Node* zeroX = __ ConX(0); 4105 4106 float likely = PROB_LIKELY(0.999); 4107 float unlikely = PROB_UNLIKELY(0.999); 4108 4109 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4110 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4111 4112 // Offsets into the thread 4113 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4114 SATBMarkQueue::byte_offset_of_active()); 4115 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4116 SATBMarkQueue::byte_offset_of_index()); 4117 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4118 SATBMarkQueue::byte_offset_of_buf()); 4119 4120 // Now the actual pointers into the thread 4121 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4122 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4123 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4124 4125 // Now some of the values 4126 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4127 4128 // if (!marking) 4129 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4130 BasicType index_bt = TypeX_X->basic_type(); 4131 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4132 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4133 4134 if (do_load) { 4135 // load original value 4136 // alias_idx correct?? 4137 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4138 } 4139 4140 // if (pre_val != NULL) 4141 __ if_then(pre_val, BoolTest::ne, null()); { 4142 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4143 4144 // is the queue for this thread full? 4145 __ if_then(index, BoolTest::ne, zeroX, likely); { 4146 4147 // decrement the index 4148 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4149 4150 // Now get the buffer location we will log the previous value into and store it 4151 Node *log_addr = __ AddP(no_base, buffer, next_index); 4152 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4153 // update the index 4154 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4155 4156 } __ else_(); { 4157 4158 // logging buffer is full, call the runtime 4159 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4160 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4161 } __ end_if(); // (!index) 4162 } __ end_if(); // (pre_val != NULL) 4163 } __ end_if(); // (!marking) 4164 4165 // Final sync IdealKit and GraphKit. 4166 final_sync(ideal); 4167 } 4168 4169 /* 4170 * G1 similar to any GC with a Young Generation requires a way to keep track of 4171 * references from Old Generation to Young Generation to make sure all live 4172 * objects are found. G1 also requires to keep track of object references 4173 * between different regions to enable evacuation of old regions, which is done 4174 * as part of mixed collections. References are tracked in remembered sets and 4175 * is continuously updated as reference are written to with the help of the 4176 * post-barrier. 4177 * 4178 * To reduce the number of updates to the remembered set the post-barrier 4179 * filters updates to fields in objects located in the Young Generation, 4180 * the same region as the reference, when the NULL is being written or 4181 * if the card is already marked as dirty by an earlier write. 4182 * 4183 * Under certain circumstances it is possible to avoid generating the 4184 * post-barrier completely if it is possible during compile time to prove 4185 * the object is newly allocated and that no safepoint exists between the 4186 * allocation and the store. 4187 * 4188 * In the case of slow allocation the allocation code must handle the barrier 4189 * as part of the allocation in the case the allocated object is not located 4190 * in the nursery, this would happen for humongous objects. This is similar to 4191 * how CMS is required to handle this case, see the comments for the method 4192 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4193 * A deferred card mark is required for these objects and handled in the above 4194 * mentioned methods. 4195 * 4196 * Returns true if the post barrier can be removed 4197 */ 4198 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4199 Node* adr) { 4200 intptr_t offset = 0; 4201 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4202 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4203 4204 if (offset == Type::OffsetBot) { 4205 return false; // cannot unalias unless there are precise offsets 4206 } 4207 4208 if (alloc == NULL) { 4209 return false; // No allocation found 4210 } 4211 4212 // Start search from Store node 4213 Node* mem = store->in(MemNode::Control); 4214 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4215 4216 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4217 AllocateNode* st_alloc = st_init->allocation(); 4218 4219 // Make sure we are looking at the same allocation 4220 if (alloc == st_alloc) { 4221 return true; 4222 } 4223 } 4224 4225 return false; 4226 } 4227 4228 // 4229 // Update the card table and add card address to the queue 4230 // 4231 void GraphKit::g1_mark_card(IdealKit& ideal, 4232 Node* card_adr, 4233 Node* oop_store, 4234 uint oop_alias_idx, 4235 Node* index, 4236 Node* index_adr, 4237 Node* buffer, 4238 const TypeFunc* tf) { 4239 4240 Node* zero = __ ConI(0); 4241 Node* zeroX = __ ConX(0); 4242 Node* no_base = __ top(); 4243 BasicType card_bt = T_BYTE; 4244 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4245 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4246 4247 // Now do the queue work 4248 __ if_then(index, BoolTest::ne, zeroX); { 4249 4250 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4251 Node* log_addr = __ AddP(no_base, buffer, next_index); 4252 4253 // Order, see storeCM. 4254 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4255 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4256 4257 } __ else_(); { 4258 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4259 } __ end_if(); 4260 4261 } 4262 4263 void GraphKit::g1_write_barrier_post(Node* oop_store, 4264 Node* obj, 4265 Node* adr, 4266 uint alias_idx, 4267 Node* val, 4268 BasicType bt, 4269 bool use_precise) { 4270 // If we are writing a NULL then we need no post barrier 4271 4272 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4273 // Must be NULL 4274 const Type* t = val->bottom_type(); 4275 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4276 // No post barrier if writing NULLx 4277 return; 4278 } 4279 4280 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4281 // We can skip marks on a freshly-allocated object in Eden. 4282 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4283 // That routine informs GC to take appropriate compensating steps, 4284 // upon a slow-path allocation, so as to make this card-mark 4285 // elision safe. 4286 return; 4287 } 4288 4289 if (use_ReduceInitialCardMarks() 4290 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4291 return; 4292 } 4293 4294 if (!use_precise) { 4295 // All card marks for a (non-array) instance are in one place: 4296 adr = obj; 4297 } 4298 // (Else it's an array (or unknown), and we want more precise card marks.) 4299 assert(adr != NULL, ""); 4300 4301 IdealKit ideal(this, true); 4302 4303 Node* tls = __ thread(); // ThreadLocalStorage 4304 4305 Node* no_base = __ top(); 4306 float likely = PROB_LIKELY(0.999); 4307 float unlikely = PROB_UNLIKELY(0.999); 4308 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4309 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4310 Node* zeroX = __ ConX(0); 4311 4312 // Get the alias_index for raw card-mark memory 4313 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4314 4315 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4316 4317 // Offsets into the thread 4318 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4319 DirtyCardQueue::byte_offset_of_index()); 4320 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4321 DirtyCardQueue::byte_offset_of_buf()); 4322 4323 // Pointers into the thread 4324 4325 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4326 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4327 4328 // Now some values 4329 // Use ctrl to avoid hoisting these values past a safepoint, which could 4330 // potentially reset these fields in the JavaThread. 4331 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4332 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4333 4334 // Convert the store obj pointer to an int prior to doing math on it 4335 // Must use ctrl to prevent "integerized oop" existing across safepoint 4336 Node* cast = __ CastPX(__ ctrl(), adr); 4337 4338 // Divide pointer by card size 4339 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4340 4341 // Combine card table base and card offset 4342 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4343 4344 // If we know the value being stored does it cross regions? 4345 4346 if (val != NULL) { 4347 // Does the store cause us to cross regions? 4348 4349 // Should be able to do an unsigned compare of region_size instead of 4350 // and extra shift. Do we have an unsigned compare?? 4351 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4352 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4353 4354 // if (xor_res == 0) same region so skip 4355 __ if_then(xor_res, BoolTest::ne, zeroX); { 4356 4357 // No barrier if we are storing a NULL 4358 __ if_then(val, BoolTest::ne, null(), unlikely); { 4359 4360 // Ok must mark the card if not already dirty 4361 4362 // load the original value of the card 4363 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4364 4365 __ if_then(card_val, BoolTest::ne, young_card); { 4366 sync_kit(ideal); 4367 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4368 insert_mem_bar(Op_MemBarVolatile, oop_store); 4369 __ sync_kit(this); 4370 4371 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4372 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4373 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4374 } __ end_if(); 4375 } __ end_if(); 4376 } __ end_if(); 4377 } __ end_if(); 4378 } else { 4379 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4380 // We don't need a barrier here if the destination is a newly allocated object 4381 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4382 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4383 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4384 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4385 __ if_then(card_val, BoolTest::ne, young_card); { 4386 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4387 } __ end_if(); 4388 } 4389 4390 // Final sync IdealKit and GraphKit. 4391 final_sync(ideal); 4392 } 4393 #undef __ 4394 4395 4396 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4397 Node* len = load_array_length(load_String_value(ctrl, str)); 4398 Node* coder = load_String_coder(ctrl, str); 4399 // Divide length by 2 if coder is UTF16 4400 return _gvn.transform(new RShiftINode(len, coder)); 4401 } 4402 4403 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4404 int value_offset = java_lang_String::value_offset_in_bytes(); 4405 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4406 false, NULL, Type::Offset(0)); 4407 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4408 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4409 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4410 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4411 int value_field_idx = C->get_alias_index(value_field_type); 4412 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4413 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4414 // String.value field is known to be @Stable. 4415 if (UseImplicitStableValues) { 4416 load = cast_array_to_stable(load, value_type); 4417 } 4418 return load; 4419 } 4420 4421 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4422 if (!CompactStrings) { 4423 return intcon(java_lang_String::CODER_UTF16); 4424 } 4425 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4426 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4427 false, NULL, Type::Offset(0)); 4428 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4429 int coder_field_idx = C->get_alias_index(coder_field_type); 4430 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4431 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4432 } 4433 4434 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4435 int value_offset = java_lang_String::value_offset_in_bytes(); 4436 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4437 false, NULL, Type::Offset(0)); 4438 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4439 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4440 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4441 } 4442 4443 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4444 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4445 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4446 false, NULL, Type::Offset(0)); 4447 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4448 int coder_field_idx = C->get_alias_index(coder_field_type); 4449 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4450 value, T_BYTE, coder_field_idx, MemNode::unordered); 4451 } 4452 4453 // Capture src and dst memory state with a MergeMemNode 4454 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4455 if (src_type == dst_type) { 4456 // Types are equal, we don't need a MergeMemNode 4457 return memory(src_type); 4458 } 4459 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4460 record_for_igvn(merge); // fold it up later, if possible 4461 int src_idx = C->get_alias_index(src_type); 4462 int dst_idx = C->get_alias_index(dst_type); 4463 merge->set_memory_at(src_idx, memory(src_idx)); 4464 merge->set_memory_at(dst_idx, memory(dst_idx)); 4465 return merge; 4466 } 4467 4468 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4469 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4470 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4471 // If input and output memory types differ, capture both states to preserve 4472 // the dependency between preceding and subsequent loads/stores. 4473 // For example, the following program: 4474 // StoreB 4475 // compress_string 4476 // LoadB 4477 // has this memory graph (use->def): 4478 // LoadB -> compress_string -> CharMem 4479 // ... -> StoreB -> ByteMem 4480 // The intrinsic hides the dependency between LoadB and StoreB, causing 4481 // the load to read from memory not containing the result of the StoreB. 4482 // The correct memory graph should look like this: 4483 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4484 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4485 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4486 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4487 set_memory(res_mem, TypeAryPtr::BYTES); 4488 return str; 4489 } 4490 4491 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4492 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4493 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4494 // Capture src and dst memory (see comment in 'compress_string'). 4495 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4496 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4497 set_memory(_gvn.transform(str), dst_type); 4498 } 4499 4500 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4501 /** 4502 * int i_char = start; 4503 * for (int i_byte = 0; i_byte < count; i_byte++) { 4504 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4505 * } 4506 */ 4507 add_predicate(); 4508 RegionNode* head = new RegionNode(3); 4509 head->init_req(1, control()); 4510 gvn().set_type(head, Type::CONTROL); 4511 record_for_igvn(head); 4512 4513 Node* i_byte = new PhiNode(head, TypeInt::INT); 4514 i_byte->init_req(1, intcon(0)); 4515 gvn().set_type(i_byte, TypeInt::INT); 4516 record_for_igvn(i_byte); 4517 4518 Node* i_char = new PhiNode(head, TypeInt::INT); 4519 i_char->init_req(1, start); 4520 gvn().set_type(i_char, TypeInt::INT); 4521 record_for_igvn(i_char); 4522 4523 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4524 gvn().set_type(mem, Type::MEMORY); 4525 record_for_igvn(mem); 4526 set_control(head); 4527 set_memory(mem, TypeAryPtr::BYTES); 4528 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4529 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4530 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4531 false, false, true /* mismatched */); 4532 4533 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4534 head->init_req(2, IfTrue(iff)); 4535 mem->init_req(2, st); 4536 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4537 i_char->init_req(2, AddI(i_char, intcon(2))); 4538 4539 set_control(IfFalse(iff)); 4540 set_memory(st, TypeAryPtr::BYTES); 4541 } 4542 4543 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4544 if (!field->is_constant()) { 4545 return NULL; // Field not marked as constant. 4546 } 4547 ciInstance* holder = NULL; 4548 if (!field->is_static()) { 4549 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4550 if (const_oop != NULL && const_oop->is_instance()) { 4551 holder = const_oop->as_instance(); 4552 } 4553 } 4554 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4555 /*is_unsigned_load=*/false); 4556 if (con_type != NULL) { 4557 Node* con = makecon(con_type); 4558 if (field->layout_type() == T_VALUETYPE) { 4559 // Load value type from constant oop 4560 con = ValueTypeNode::make_from_oop(this, con); 4561 } 4562 return con; 4563 } 4564 return NULL; 4565 } 4566 4567 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4568 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4569 // assumption of CCP analysis. 4570 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4571 }