1 /* 2 * Copyright (c) 2001, 2025, 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 "ci/ciFlatArrayKlass.hpp" 26 #include "ci/ciInlineKlass.hpp" 27 #include "ci/ciUtilities.hpp" 28 #include "classfile/javaClasses.hpp" 29 #include "ci/ciObjArray.hpp" 30 #include "asm/register.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "gc/shared/barrierSet.hpp" 33 #include "gc/shared/c2/barrierSetC2.hpp" 34 #include "interpreter/interpreter.hpp" 35 #include "memory/resourceArea.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/castnode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/graphKit.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/inlinetypenode.hpp" 42 #include "opto/intrinsicnode.hpp" 43 #include "opto/locknode.hpp" 44 #include "opto/machnode.hpp" 45 #include "opto/narrowptrnode.hpp" 46 #include "opto/opaquenode.hpp" 47 #include "opto/parse.hpp" 48 #include "opto/rootnode.hpp" 49 #include "opto/runtime.hpp" 50 #include "opto/subtypenode.hpp" 51 #include "runtime/deoptimization.hpp" 52 #include "runtime/sharedRuntime.hpp" 53 #include "utilities/bitMap.inline.hpp" 54 #include "utilities/powerOfTwo.hpp" 55 #include "utilities/growableArray.hpp" 56 57 //----------------------------GraphKit----------------------------------------- 58 // Main utility constructor. 59 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 60 : Phase(Phase::Parser), 61 _env(C->env()), 62 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()), 63 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 64 { 65 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled"); 66 _exceptions = jvms->map()->next_exception(); 67 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr); 68 set_jvms(jvms); 69 #ifdef ASSERT 70 if (_gvn.is_IterGVN() != nullptr) { 71 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used"); 72 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 73 _worklist_size = _gvn.C->igvn_worklist()->size(); 74 } 75 #endif 76 } 77 78 // Private constructor for parser. 79 GraphKit::GraphKit() 80 : Phase(Phase::Parser), 81 _env(C->env()), 82 _gvn(*C->initial_gvn()), 83 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 84 { 85 _exceptions = nullptr; 86 set_map(nullptr); 87 debug_only(_sp = -99); 88 debug_only(set_bci(-99)); 89 } 90 91 92 93 //---------------------------clean_stack--------------------------------------- 94 // Clear away rubbish from the stack area of the JVM state. 95 // This destroys any arguments that may be waiting on the stack. 96 void GraphKit::clean_stack(int from_sp) { 97 SafePointNode* map = this->map(); 98 JVMState* jvms = this->jvms(); 99 int stk_size = jvms->stk_size(); 100 int stkoff = jvms->stkoff(); 101 Node* top = this->top(); 102 for (int i = from_sp; i < stk_size; i++) { 103 if (map->in(stkoff + i) != top) { 104 map->set_req(stkoff + i, top); 105 } 106 } 107 } 108 109 110 //--------------------------------sync_jvms----------------------------------- 111 // Make sure our current jvms agrees with our parse state. 112 JVMState* GraphKit::sync_jvms() const { 113 JVMState* jvms = this->jvms(); 114 jvms->set_bci(bci()); // Record the new bci in the JVMState 115 jvms->set_sp(sp()); // Record the new sp in the JVMState 116 assert(jvms_in_sync(), "jvms is now in sync"); 117 return jvms; 118 } 119 120 //--------------------------------sync_jvms_for_reexecute--------------------- 121 // Make sure our current jvms agrees with our parse state. This version 122 // uses the reexecute_sp for reexecuting bytecodes. 123 JVMState* GraphKit::sync_jvms_for_reexecute() { 124 JVMState* jvms = this->jvms(); 125 jvms->set_bci(bci()); // Record the new bci in the JVMState 126 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 127 return jvms; 128 } 129 130 #ifdef ASSERT 131 bool GraphKit::jvms_in_sync() const { 132 Parse* parse = is_Parse(); 133 if (parse == nullptr) { 134 if (bci() != jvms()->bci()) return false; 135 if (sp() != (int)jvms()->sp()) return false; 136 return true; 137 } 138 if (jvms()->method() != parse->method()) return false; 139 if (jvms()->bci() != parse->bci()) return false; 140 int jvms_sp = jvms()->sp(); 141 if (jvms_sp != parse->sp()) return false; 142 int jvms_depth = jvms()->depth(); 143 if (jvms_depth != parse->depth()) return false; 144 return true; 145 } 146 147 // Local helper checks for special internal merge points 148 // used to accumulate and merge exception states. 149 // They are marked by the region's in(0) edge being the map itself. 150 // Such merge points must never "escape" into the parser at large, 151 // until they have been handed to gvn.transform. 152 static bool is_hidden_merge(Node* reg) { 153 if (reg == nullptr) return false; 154 if (reg->is_Phi()) { 155 reg = reg->in(0); 156 if (reg == nullptr) return false; 157 } 158 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root(); 159 } 160 161 void GraphKit::verify_map() const { 162 if (map() == nullptr) return; // null map is OK 163 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 164 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 165 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 166 } 167 168 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 169 assert(ex_map->next_exception() == nullptr, "not already part of a chain"); 170 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 171 } 172 #endif 173 174 //---------------------------stop_and_kill_map--------------------------------- 175 // Set _map to null, signalling a stop to further bytecode execution. 176 // First smash the current map's control to a constant, to mark it dead. 177 void GraphKit::stop_and_kill_map() { 178 SafePointNode* dead_map = stop(); 179 if (dead_map != nullptr) { 180 dead_map->disconnect_inputs(C); // Mark the map as killed. 181 assert(dead_map->is_killed(), "must be so marked"); 182 } 183 } 184 185 186 //--------------------------------stopped-------------------------------------- 187 // Tell if _map is null, or control is top. 188 bool GraphKit::stopped() { 189 if (map() == nullptr) return true; 190 else if (control() == top()) return true; 191 else return false; 192 } 193 194 195 //-----------------------------has_exception_handler---------------------------------- 196 // Tell if this method or any caller method has exception handlers. 197 bool GraphKit::has_exception_handler() { 198 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) { 199 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 200 return true; 201 } 202 } 203 return false; 204 } 205 206 //------------------------------save_ex_oop------------------------------------ 207 // Save an exception without blowing stack contents or other JVM state. 208 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 209 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 210 ex_map->add_req(ex_oop); 211 debug_only(verify_exception_state(ex_map)); 212 } 213 214 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 215 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 216 Node* ex_oop = ex_map->in(ex_map->req()-1); 217 if (clear_it) ex_map->del_req(ex_map->req()-1); 218 return ex_oop; 219 } 220 221 //-----------------------------saved_ex_oop------------------------------------ 222 // Recover a saved exception from its map. 223 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 224 return common_saved_ex_oop(ex_map, false); 225 } 226 227 //--------------------------clear_saved_ex_oop--------------------------------- 228 // Erase a previously saved exception from its map. 229 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 230 return common_saved_ex_oop(ex_map, true); 231 } 232 233 #ifdef ASSERT 234 //---------------------------has_saved_ex_oop---------------------------------- 235 // Erase a previously saved exception from its map. 236 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 237 return ex_map->req() == ex_map->jvms()->endoff()+1; 238 } 239 #endif 240 241 //-------------------------make_exception_state-------------------------------- 242 // Turn the current JVM state into an exception state, appending the ex_oop. 243 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 244 sync_jvms(); 245 SafePointNode* ex_map = stop(); // do not manipulate this map any more 246 set_saved_ex_oop(ex_map, ex_oop); 247 return ex_map; 248 } 249 250 251 //--------------------------add_exception_state-------------------------------- 252 // Add an exception to my list of exceptions. 253 void GraphKit::add_exception_state(SafePointNode* ex_map) { 254 if (ex_map == nullptr || ex_map->control() == top()) { 255 return; 256 } 257 #ifdef ASSERT 258 verify_exception_state(ex_map); 259 if (has_exceptions()) { 260 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 261 } 262 #endif 263 264 // If there is already an exception of exactly this type, merge with it. 265 // In particular, null-checks and other low-level exceptions common up here. 266 Node* ex_oop = saved_ex_oop(ex_map); 267 const Type* ex_type = _gvn.type(ex_oop); 268 if (ex_oop == top()) { 269 // No action needed. 270 return; 271 } 272 assert(ex_type->isa_instptr(), "exception must be an instance"); 273 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) { 274 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 275 // We check sp also because call bytecodes can generate exceptions 276 // both before and after arguments are popped! 277 if (ex_type2 == ex_type 278 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 279 combine_exception_states(ex_map, e2); 280 return; 281 } 282 } 283 284 // No pre-existing exception of the same type. Chain it on the list. 285 push_exception_state(ex_map); 286 } 287 288 //-----------------------add_exception_states_from----------------------------- 289 void GraphKit::add_exception_states_from(JVMState* jvms) { 290 SafePointNode* ex_map = jvms->map()->next_exception(); 291 if (ex_map != nullptr) { 292 jvms->map()->set_next_exception(nullptr); 293 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) { 294 next_map = ex_map->next_exception(); 295 ex_map->set_next_exception(nullptr); 296 add_exception_state(ex_map); 297 } 298 } 299 } 300 301 //-----------------------transfer_exceptions_into_jvms------------------------- 302 JVMState* GraphKit::transfer_exceptions_into_jvms() { 303 if (map() == nullptr) { 304 // We need a JVMS to carry the exceptions, but the map has gone away. 305 // Create a scratch JVMS, cloned from any of the exception states... 306 if (has_exceptions()) { 307 _map = _exceptions; 308 _map = clone_map(); 309 _map->set_next_exception(nullptr); 310 clear_saved_ex_oop(_map); 311 debug_only(verify_map()); 312 } else { 313 // ...or created from scratch 314 JVMState* jvms = new (C) JVMState(_method, nullptr); 315 jvms->set_bci(_bci); 316 jvms->set_sp(_sp); 317 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 318 set_jvms(jvms); 319 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 320 set_all_memory(top()); 321 while (map()->req() < jvms->endoff()) map()->add_req(top()); 322 } 323 // (This is a kludge, in case you didn't notice.) 324 set_control(top()); 325 } 326 JVMState* jvms = sync_jvms(); 327 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 328 jvms->map()->set_next_exception(_exceptions); 329 _exceptions = nullptr; // done with this set of exceptions 330 return jvms; 331 } 332 333 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 334 assert(is_hidden_merge(dstphi), "must be a special merge node"); 335 assert(is_hidden_merge(srcphi), "must be a special merge node"); 336 uint limit = srcphi->req(); 337 for (uint i = PhiNode::Input; i < limit; i++) { 338 dstphi->add_req(srcphi->in(i)); 339 } 340 } 341 static inline void add_one_req(Node* dstphi, Node* src) { 342 assert(is_hidden_merge(dstphi), "must be a special merge node"); 343 assert(!is_hidden_merge(src), "must not be a special merge node"); 344 dstphi->add_req(src); 345 } 346 347 //-----------------------combine_exception_states------------------------------ 348 // This helper function combines exception states by building phis on a 349 // specially marked state-merging region. These regions and phis are 350 // untransformed, and can build up gradually. The region is marked by 351 // having a control input of its exception map, rather than null. Such 352 // regions do not appear except in this function, and in use_exception_state. 353 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 354 if (failing_internal()) { 355 return; // dying anyway... 356 } 357 JVMState* ex_jvms = ex_map->_jvms; 358 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 359 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 360 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 361 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 362 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 363 assert(ex_map->req() == phi_map->req(), "matching maps"); 364 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 365 Node* hidden_merge_mark = root(); 366 Node* region = phi_map->control(); 367 MergeMemNode* phi_mem = phi_map->merged_memory(); 368 MergeMemNode* ex_mem = ex_map->merged_memory(); 369 if (region->in(0) != hidden_merge_mark) { 370 // The control input is not (yet) a specially-marked region in phi_map. 371 // Make it so, and build some phis. 372 region = new RegionNode(2); 373 _gvn.set_type(region, Type::CONTROL); 374 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 375 region->init_req(1, phi_map->control()); 376 phi_map->set_control(region); 377 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 378 record_for_igvn(io_phi); 379 _gvn.set_type(io_phi, Type::ABIO); 380 phi_map->set_i_o(io_phi); 381 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 382 Node* m = mms.memory(); 383 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 384 record_for_igvn(m_phi); 385 _gvn.set_type(m_phi, Type::MEMORY); 386 mms.set_memory(m_phi); 387 } 388 } 389 390 // Either or both of phi_map and ex_map might already be converted into phis. 391 Node* ex_control = ex_map->control(); 392 // if there is special marking on ex_map also, we add multiple edges from src 393 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 394 // how wide was the destination phi_map, originally? 395 uint orig_width = region->req(); 396 397 if (add_multiple) { 398 add_n_reqs(region, ex_control); 399 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 400 } else { 401 // ex_map has no merges, so we just add single edges everywhere 402 add_one_req(region, ex_control); 403 add_one_req(phi_map->i_o(), ex_map->i_o()); 404 } 405 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 406 if (mms.is_empty()) { 407 // get a copy of the base memory, and patch some inputs into it 408 const TypePtr* adr_type = mms.adr_type(C); 409 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 410 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 411 mms.set_memory(phi); 412 // Prepare to append interesting stuff onto the newly sliced phi: 413 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 414 } 415 // Append stuff from ex_map: 416 if (add_multiple) { 417 add_n_reqs(mms.memory(), mms.memory2()); 418 } else { 419 add_one_req(mms.memory(), mms.memory2()); 420 } 421 } 422 uint limit = ex_map->req(); 423 for (uint i = TypeFunc::Parms; i < limit; i++) { 424 // Skip everything in the JVMS after tos. (The ex_oop follows.) 425 if (i == tos) i = ex_jvms->monoff(); 426 Node* src = ex_map->in(i); 427 Node* dst = phi_map->in(i); 428 if (src != dst) { 429 PhiNode* phi; 430 if (dst->in(0) != region) { 431 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 432 record_for_igvn(phi); 433 _gvn.set_type(phi, phi->type()); 434 phi_map->set_req(i, dst); 435 // Prepare to append interesting stuff onto the new phi: 436 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 437 } else { 438 assert(dst->is_Phi(), "nobody else uses a hidden region"); 439 phi = dst->as_Phi(); 440 } 441 if (add_multiple && src->in(0) == ex_control) { 442 // Both are phis. 443 add_n_reqs(dst, src); 444 } else { 445 while (dst->req() < region->req()) add_one_req(dst, src); 446 } 447 const Type* srctype = _gvn.type(src); 448 if (phi->type() != srctype) { 449 const Type* dsttype = phi->type()->meet_speculative(srctype); 450 if (phi->type() != dsttype) { 451 phi->set_type(dsttype); 452 _gvn.set_type(phi, dsttype); 453 } 454 } 455 } 456 } 457 phi_map->merge_replaced_nodes_with(ex_map); 458 } 459 460 //--------------------------use_exception_state-------------------------------- 461 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 462 if (failing_internal()) { stop(); return top(); } 463 Node* region = phi_map->control(); 464 Node* hidden_merge_mark = root(); 465 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 466 Node* ex_oop = clear_saved_ex_oop(phi_map); 467 if (region->in(0) == hidden_merge_mark) { 468 // Special marking for internal ex-states. Process the phis now. 469 region->set_req(0, region); // now it's an ordinary region 470 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 471 // Note: Setting the jvms also sets the bci and sp. 472 set_control(_gvn.transform(region)); 473 uint tos = jvms()->stkoff() + sp(); 474 for (uint i = 1; i < tos; i++) { 475 Node* x = phi_map->in(i); 476 if (x->in(0) == region) { 477 assert(x->is_Phi(), "expected a special phi"); 478 phi_map->set_req(i, _gvn.transform(x)); 479 } 480 } 481 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 482 Node* x = mms.memory(); 483 if (x->in(0) == region) { 484 assert(x->is_Phi(), "nobody else uses a hidden region"); 485 mms.set_memory(_gvn.transform(x)); 486 } 487 } 488 if (ex_oop->in(0) == region) { 489 assert(ex_oop->is_Phi(), "expected a special phi"); 490 ex_oop = _gvn.transform(ex_oop); 491 } 492 } else { 493 set_jvms(phi_map->jvms()); 494 } 495 496 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 497 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 498 return ex_oop; 499 } 500 501 //---------------------------------java_bc------------------------------------- 502 Bytecodes::Code GraphKit::java_bc() const { 503 ciMethod* method = this->method(); 504 int bci = this->bci(); 505 if (method != nullptr && bci != InvocationEntryBci) 506 return method->java_code_at_bci(bci); 507 else 508 return Bytecodes::_illegal; 509 } 510 511 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 512 bool must_throw) { 513 // if the exception capability is set, then we will generate code 514 // to check the JavaThread.should_post_on_exceptions flag to see 515 // if we actually need to report exception events (for this 516 // thread). If we don't need to report exception events, we will 517 // take the normal fast path provided by add_exception_events. If 518 // exception event reporting is enabled for this thread, we will 519 // take the uncommon_trap in the BuildCutout below. 520 521 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 522 Node* jthread = _gvn.transform(new ThreadLocalNode()); 523 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 524 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 525 526 // Test the should_post_on_exceptions_flag vs. 0 527 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 528 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 529 530 // Branch to slow_path if should_post_on_exceptions_flag was true 531 { BuildCutout unless(this, tst, PROB_MAX); 532 // Do not try anything fancy if we're notifying the VM on every throw. 533 // Cf. case Bytecodes::_athrow in parse2.cpp. 534 uncommon_trap(reason, Deoptimization::Action_none, 535 (ciKlass*)nullptr, (char*)nullptr, must_throw); 536 } 537 538 } 539 540 //------------------------------builtin_throw---------------------------------- 541 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) { 542 bool must_throw = true; 543 544 // If this particular condition has not yet happened at this 545 // bytecode, then use the uncommon trap mechanism, and allow for 546 // a future recompilation if several traps occur here. 547 // If the throw is hot, try to use a more complicated inline mechanism 548 // which keeps execution inside the compiled code. 549 bool treat_throw_as_hot = false; 550 ciMethodData* md = method()->method_data(); 551 552 if (ProfileTraps) { 553 if (too_many_traps(reason)) { 554 treat_throw_as_hot = true; 555 } 556 // (If there is no MDO at all, assume it is early in 557 // execution, and that any deopts are part of the 558 // startup transient, and don't need to be remembered.) 559 560 // Also, if there is a local exception handler, treat all throws 561 // as hot if there has been at least one in this method. 562 if (C->trap_count(reason) != 0 563 && method()->method_data()->trap_count(reason) != 0 564 && has_exception_handler()) { 565 treat_throw_as_hot = true; 566 } 567 } 568 569 // If this throw happens frequently, an uncommon trap might cause 570 // a performance pothole. If there is a local exception handler, 571 // and if this particular bytecode appears to be deoptimizing often, 572 // let us handle the throw inline, with a preconstructed instance. 573 // Note: If the deopt count has blown up, the uncommon trap 574 // runtime is going to flush this nmethod, not matter what. 575 if (treat_throw_as_hot && method()->can_omit_stack_trace()) { 576 // If the throw is local, we use a pre-existing instance and 577 // punt on the backtrace. This would lead to a missing backtrace 578 // (a repeat of 4292742) if the backtrace object is ever asked 579 // for its backtrace. 580 // Fixing this remaining case of 4292742 requires some flavor of 581 // escape analysis. Leave that for the future. 582 ciInstance* ex_obj = nullptr; 583 switch (reason) { 584 case Deoptimization::Reason_null_check: 585 ex_obj = env()->NullPointerException_instance(); 586 break; 587 case Deoptimization::Reason_div0_check: 588 ex_obj = env()->ArithmeticException_instance(); 589 break; 590 case Deoptimization::Reason_range_check: 591 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 592 break; 593 case Deoptimization::Reason_class_check: 594 ex_obj = env()->ClassCastException_instance(); 595 break; 596 case Deoptimization::Reason_array_check: 597 ex_obj = env()->ArrayStoreException_instance(); 598 break; 599 default: 600 break; 601 } 602 if (failing()) { stop(); return; } // exception allocation might fail 603 if (ex_obj != nullptr) { 604 if (env()->jvmti_can_post_on_exceptions()) { 605 // check if we must post exception events, take uncommon trap if so 606 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 607 // here if should_post_on_exceptions is false 608 // continue on with the normal codegen 609 } 610 611 // Cheat with a preallocated exception object. 612 if (C->log() != nullptr) 613 C->log()->elem("hot_throw preallocated='1' reason='%s'", 614 Deoptimization::trap_reason_name(reason)); 615 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 616 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 617 618 // Clear the detail message of the preallocated exception object. 619 // Weblogic sometimes mutates the detail message of exceptions 620 // using reflection. 621 int offset = java_lang_Throwable::get_detailMessage_offset(); 622 const TypePtr* adr_typ = ex_con->add_offset(offset); 623 624 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 625 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 626 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 627 628 if (!method()->has_exception_handlers()) { 629 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway. 630 // This prevents issues with exception handling and late inlining. 631 set_sp(0); 632 clean_stack(0); 633 } 634 635 add_exception_state(make_exception_state(ex_node)); 636 return; 637 } 638 } 639 640 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 641 // It won't be much cheaper than bailing to the interp., since we'll 642 // have to pass up all the debug-info, and the runtime will have to 643 // create the stack trace. 644 645 // Usual case: Bail to interpreter. 646 // Reserve the right to recompile if we haven't seen anything yet. 647 648 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr; 649 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 650 if (treat_throw_as_hot 651 && (method()->method_data()->trap_recompiled_at(bci(), m) 652 || C->too_many_traps(reason))) { 653 // We cannot afford to take more traps here. Suffer in the interpreter. 654 if (C->log() != nullptr) 655 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 656 Deoptimization::trap_reason_name(reason), 657 C->trap_count(reason)); 658 action = Deoptimization::Action_none; 659 } 660 661 // "must_throw" prunes the JVM state to include only the stack, if there 662 // are no local exception handlers. This should cut down on register 663 // allocation time and code size, by drastically reducing the number 664 // of in-edges on the call to the uncommon trap. 665 666 uncommon_trap(reason, action, (ciKlass*)nullptr, (char*)nullptr, must_throw); 667 } 668 669 670 //----------------------------PreserveJVMState--------------------------------- 671 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 672 debug_only(kit->verify_map()); 673 _kit = kit; 674 _map = kit->map(); // preserve the map 675 _sp = kit->sp(); 676 kit->set_map(clone_map ? kit->clone_map() : nullptr); 677 #ifdef ASSERT 678 _bci = kit->bci(); 679 Parse* parser = kit->is_Parse(); 680 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo(); 681 _block = block; 682 #endif 683 } 684 PreserveJVMState::~PreserveJVMState() { 685 GraphKit* kit = _kit; 686 #ifdef ASSERT 687 assert(kit->bci() == _bci, "bci must not shift"); 688 Parse* parser = kit->is_Parse(); 689 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo(); 690 assert(block == _block, "block must not shift"); 691 #endif 692 kit->set_map(_map); 693 kit->set_sp(_sp); 694 } 695 696 697 //-----------------------------BuildCutout------------------------------------- 698 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 699 : PreserveJVMState(kit) 700 { 701 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 702 SafePointNode* outer_map = _map; // preserved map is caller's 703 SafePointNode* inner_map = kit->map(); 704 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 705 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 706 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 707 } 708 BuildCutout::~BuildCutout() { 709 GraphKit* kit = _kit; 710 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 711 } 712 713 //---------------------------PreserveReexecuteState---------------------------- 714 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 715 assert(!kit->stopped(), "must call stopped() before"); 716 _kit = kit; 717 _sp = kit->sp(); 718 _reexecute = kit->jvms()->_reexecute; 719 } 720 PreserveReexecuteState::~PreserveReexecuteState() { 721 if (_kit->stopped()) return; 722 _kit->jvms()->_reexecute = _reexecute; 723 _kit->set_sp(_sp); 724 } 725 726 //------------------------------clone_map-------------------------------------- 727 // Implementation of PreserveJVMState 728 // 729 // Only clone_map(...) here. If this function is only used in the 730 // PreserveJVMState class we may want to get rid of this extra 731 // function eventually and do it all there. 732 733 SafePointNode* GraphKit::clone_map() { 734 if (map() == nullptr) return nullptr; 735 736 // Clone the memory edge first 737 Node* mem = MergeMemNode::make(map()->memory()); 738 gvn().set_type_bottom(mem); 739 740 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 741 JVMState* jvms = this->jvms(); 742 JVMState* clonejvms = jvms->clone_shallow(C); 743 clonemap->set_memory(mem); 744 clonemap->set_jvms(clonejvms); 745 clonejvms->set_map(clonemap); 746 record_for_igvn(clonemap); 747 gvn().set_type_bottom(clonemap); 748 return clonemap; 749 } 750 751 //-----------------------------destruct_map_clone------------------------------ 752 // 753 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need 754 // to destruct/free/delete in the exact opposite order as clone_map(). 755 void GraphKit::destruct_map_clone(SafePointNode* sfp) { 756 if (sfp == nullptr) return; 757 758 Node* mem = sfp->memory(); 759 JVMState* jvms = sfp->jvms(); 760 761 if (jvms != nullptr) { 762 delete jvms; 763 } 764 765 remove_for_igvn(sfp); 766 gvn().clear_type(sfp); 767 sfp->destruct(&_gvn); 768 769 if (mem != nullptr) { 770 gvn().clear_type(mem); 771 mem->destruct(&_gvn); 772 } 773 } 774 775 //-----------------------------set_map_clone----------------------------------- 776 void GraphKit::set_map_clone(SafePointNode* m) { 777 _map = m; 778 _map = clone_map(); 779 _map->set_next_exception(nullptr); 780 debug_only(verify_map()); 781 } 782 783 784 //----------------------------kill_dead_locals--------------------------------- 785 // Detect any locals which are known to be dead, and force them to top. 786 void GraphKit::kill_dead_locals() { 787 // Consult the liveness information for the locals. If any 788 // of them are unused, then they can be replaced by top(). This 789 // should help register allocation time and cut down on the size 790 // of the deoptimization information. 791 792 // This call is made from many of the bytecode handling 793 // subroutines called from the Big Switch in do_one_bytecode. 794 // Every bytecode which might include a slow path is responsible 795 // for killing its dead locals. The more consistent we 796 // are about killing deads, the fewer useless phis will be 797 // constructed for them at various merge points. 798 799 // bci can be -1 (InvocationEntryBci). We return the entry 800 // liveness for the method. 801 802 if (method() == nullptr || method()->code_size() == 0) { 803 // We are building a graph for a call to a native method. 804 // All locals are live. 805 return; 806 } 807 808 ResourceMark rm; 809 810 // Consult the liveness information for the locals. If any 811 // of them are unused, then they can be replaced by top(). This 812 // should help register allocation time and cut down on the size 813 // of the deoptimization information. 814 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 815 816 int len = (int)live_locals.size(); 817 assert(len <= jvms()->loc_size(), "too many live locals"); 818 for (int local = 0; local < len; local++) { 819 if (!live_locals.at(local)) { 820 set_local(local, top()); 821 } 822 } 823 } 824 825 #ifdef ASSERT 826 //-------------------------dead_locals_are_killed------------------------------ 827 // Return true if all dead locals are set to top in the map. 828 // Used to assert "clean" debug info at various points. 829 bool GraphKit::dead_locals_are_killed() { 830 if (method() == nullptr || method()->code_size() == 0) { 831 // No locals need to be dead, so all is as it should be. 832 return true; 833 } 834 835 // Make sure somebody called kill_dead_locals upstream. 836 ResourceMark rm; 837 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) { 838 if (jvms->loc_size() == 0) continue; // no locals to consult 839 SafePointNode* map = jvms->map(); 840 ciMethod* method = jvms->method(); 841 int bci = jvms->bci(); 842 if (jvms == this->jvms()) { 843 bci = this->bci(); // it might not yet be synched 844 } 845 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 846 int len = (int)live_locals.size(); 847 if (!live_locals.is_valid() || len == 0) 848 // This method is trivial, or is poisoned by a breakpoint. 849 return true; 850 assert(len == jvms->loc_size(), "live map consistent with locals map"); 851 for (int local = 0; local < len; local++) { 852 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 853 if (PrintMiscellaneous && (Verbose || WizardMode)) { 854 tty->print_cr("Zombie local %d: ", local); 855 jvms->dump(); 856 } 857 return false; 858 } 859 } 860 } 861 return true; 862 } 863 864 #endif //ASSERT 865 866 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens. 867 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 868 ciMethod* cur_method = jvms->method(); 869 int cur_bci = jvms->bci(); 870 if (cur_method != nullptr && cur_bci != InvocationEntryBci) { 871 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 872 return Interpreter::bytecode_should_reexecute(code) || 873 (is_anewarray && (code == Bytecodes::_multianewarray)); 874 // Reexecute _multianewarray bytecode which was replaced with 875 // sequence of [a]newarray. See Parse::do_multianewarray(). 876 // 877 // Note: interpreter should not have it set since this optimization 878 // is limited by dimensions and guarded by flag so in some cases 879 // multianewarray() runtime calls will be generated and 880 // the bytecode should not be reexecutes (stack will not be reset). 881 } else { 882 return false; 883 } 884 } 885 886 // Helper function for adding JVMState and debug information to node 887 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 888 // Add the safepoint edges to the call (or other safepoint). 889 890 // Make sure dead locals are set to top. This 891 // should help register allocation time and cut down on the size 892 // of the deoptimization information. 893 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 894 895 // Walk the inline list to fill in the correct set of JVMState's 896 // Also fill in the associated edges for each JVMState. 897 898 // If the bytecode needs to be reexecuted we need to put 899 // the arguments back on the stack. 900 const bool should_reexecute = jvms()->should_reexecute(); 901 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 902 903 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 904 // undefined if the bci is different. This is normal for Parse but it 905 // should not happen for LibraryCallKit because only one bci is processed. 906 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 907 "in LibraryCallKit the reexecute bit should not change"); 908 909 // If we are guaranteed to throw, we can prune everything but the 910 // input to the current bytecode. 911 bool can_prune_locals = false; 912 uint stack_slots_not_pruned = 0; 913 int inputs = 0, depth = 0; 914 if (must_throw) { 915 assert(method() == youngest_jvms->method(), "sanity"); 916 if (compute_stack_effects(inputs, depth)) { 917 can_prune_locals = true; 918 stack_slots_not_pruned = inputs; 919 } 920 } 921 922 if (env()->should_retain_local_variables()) { 923 // At any safepoint, this method can get breakpointed, which would 924 // then require an immediate deoptimization. 925 can_prune_locals = false; // do not prune locals 926 stack_slots_not_pruned = 0; 927 } 928 929 // do not scribble on the input jvms 930 JVMState* out_jvms = youngest_jvms->clone_deep(C); 931 call->set_jvms(out_jvms); // Start jvms list for call node 932 933 // For a known set of bytecodes, the interpreter should reexecute them if 934 // deoptimization happens. We set the reexecute state for them here 935 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 936 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 937 #ifdef ASSERT 938 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects() 939 assert(method() == youngest_jvms->method(), "sanity"); 940 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc())); 941 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution"); 942 #endif // ASSERT 943 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 944 } 945 946 // Presize the call: 947 DEBUG_ONLY(uint non_debug_edges = call->req()); 948 call->add_req_batch(top(), youngest_jvms->debug_depth()); 949 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 950 951 // Set up edges so that the call looks like this: 952 // Call [state:] ctl io mem fptr retadr 953 // [parms:] parm0 ... parmN 954 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 955 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 956 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 957 // Note that caller debug info precedes callee debug info. 958 959 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 960 uint debug_ptr = call->req(); 961 962 // Loop over the map input edges associated with jvms, add them 963 // to the call node, & reset all offsets to match call node array. 964 965 JVMState* callee_jvms = nullptr; 966 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) { 967 uint debug_end = debug_ptr; 968 uint debug_start = debug_ptr - in_jvms->debug_size(); 969 debug_ptr = debug_start; // back up the ptr 970 971 uint p = debug_start; // walks forward in [debug_start, debug_end) 972 uint j, k, l; 973 SafePointNode* in_map = in_jvms->map(); 974 out_jvms->set_map(call); 975 976 if (can_prune_locals) { 977 assert(in_jvms->method() == out_jvms->method(), "sanity"); 978 // If the current throw can reach an exception handler in this JVMS, 979 // then we must keep everything live that can reach that handler. 980 // As a quick and dirty approximation, we look for any handlers at all. 981 if (in_jvms->method()->has_exception_handlers()) { 982 can_prune_locals = false; 983 } 984 } 985 986 // Add the Locals 987 k = in_jvms->locoff(); 988 l = in_jvms->loc_size(); 989 out_jvms->set_locoff(p); 990 if (!can_prune_locals) { 991 for (j = 0; j < l; j++) { 992 Node* val = in_map->in(k + j); 993 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state 994 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() && 995 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) && 996 val->bottom_type()->is_inlinetypeptr()) { 997 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver 998 } 999 call->set_req(p++, val); 1000 } 1001 } else { 1002 p += l; // already set to top above by add_req_batch 1003 } 1004 1005 // Add the Expression Stack 1006 k = in_jvms->stkoff(); 1007 l = in_jvms->sp(); 1008 out_jvms->set_stkoff(p); 1009 if (!can_prune_locals) { 1010 for (j = 0; j < l; j++) { 1011 Node* val = in_map->in(k + j); 1012 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state 1013 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() && 1014 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) && 1015 val->bottom_type()->is_inlinetypeptr()) { 1016 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver 1017 } 1018 call->set_req(p++, val); 1019 } 1020 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 1021 // Divide stack into {S0,...,S1}, where S0 is set to top. 1022 uint s1 = stack_slots_not_pruned; 1023 stack_slots_not_pruned = 0; // for next iteration 1024 if (s1 > l) s1 = l; 1025 uint s0 = l - s1; 1026 p += s0; // skip the tops preinstalled by add_req_batch 1027 for (j = s0; j < l; j++) 1028 call->set_req(p++, in_map->in(k+j)); 1029 } else { 1030 p += l; // already set to top above by add_req_batch 1031 } 1032 1033 // Add the Monitors 1034 k = in_jvms->monoff(); 1035 l = in_jvms->mon_size(); 1036 out_jvms->set_monoff(p); 1037 for (j = 0; j < l; j++) 1038 call->set_req(p++, in_map->in(k+j)); 1039 1040 // Copy any scalar object fields. 1041 k = in_jvms->scloff(); 1042 l = in_jvms->scl_size(); 1043 out_jvms->set_scloff(p); 1044 for (j = 0; j < l; j++) 1045 call->set_req(p++, in_map->in(k+j)); 1046 1047 // Finish the new jvms. 1048 out_jvms->set_endoff(p); 1049 1050 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 1051 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 1052 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 1053 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 1054 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 1055 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 1056 1057 // Update the two tail pointers in parallel. 1058 callee_jvms = out_jvms; 1059 out_jvms = out_jvms->caller(); 1060 in_jvms = in_jvms->caller(); 1061 } 1062 1063 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1064 1065 // Test the correctness of JVMState::debug_xxx accessors: 1066 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1067 assert(call->jvms()->debug_end() == call->req(), ""); 1068 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1069 } 1070 1071 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1072 Bytecodes::Code code = java_bc(); 1073 if (code == Bytecodes::_wide) { 1074 code = method()->java_code_at_bci(bci() + 1); 1075 } 1076 1077 if (code != Bytecodes::_illegal) { 1078 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1079 } 1080 1081 auto rsize = [&]() { 1082 assert(code != Bytecodes::_illegal, "code is illegal!"); 1083 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1084 return (rtype < T_CONFLICT) ? type2size[rtype] : 0; 1085 }; 1086 1087 switch (code) { 1088 case Bytecodes::_illegal: 1089 return false; 1090 1091 case Bytecodes::_ldc: 1092 case Bytecodes::_ldc_w: 1093 case Bytecodes::_ldc2_w: 1094 inputs = 0; 1095 break; 1096 1097 case Bytecodes::_dup: inputs = 1; break; 1098 case Bytecodes::_dup_x1: inputs = 2; break; 1099 case Bytecodes::_dup_x2: inputs = 3; break; 1100 case Bytecodes::_dup2: inputs = 2; break; 1101 case Bytecodes::_dup2_x1: inputs = 3; break; 1102 case Bytecodes::_dup2_x2: inputs = 4; break; 1103 case Bytecodes::_swap: inputs = 2; break; 1104 case Bytecodes::_arraylength: inputs = 1; break; 1105 1106 case Bytecodes::_getstatic: 1107 case Bytecodes::_putstatic: 1108 case Bytecodes::_getfield: 1109 case Bytecodes::_putfield: 1110 { 1111 bool ignored_will_link; 1112 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1113 int size = field->type()->size(); 1114 bool is_get = (depth >= 0), is_static = (depth & 1); 1115 inputs = (is_static ? 0 : 1); 1116 if (is_get) { 1117 depth = size - inputs; 1118 } else { 1119 inputs += size; // putxxx pops the value from the stack 1120 depth = - inputs; 1121 } 1122 } 1123 break; 1124 1125 case Bytecodes::_invokevirtual: 1126 case Bytecodes::_invokespecial: 1127 case Bytecodes::_invokestatic: 1128 case Bytecodes::_invokedynamic: 1129 case Bytecodes::_invokeinterface: 1130 { 1131 bool ignored_will_link; 1132 ciSignature* declared_signature = nullptr; 1133 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1134 assert(declared_signature != nullptr, "cannot be null"); 1135 inputs = declared_signature->arg_size_for_bc(code); 1136 int size = declared_signature->return_type()->size(); 1137 depth = size - inputs; 1138 } 1139 break; 1140 1141 case Bytecodes::_multianewarray: 1142 { 1143 ciBytecodeStream iter(method()); 1144 iter.reset_to_bci(bci()); 1145 iter.next(); 1146 inputs = iter.get_dimensions(); 1147 assert(rsize() == 1, ""); 1148 depth = 1 - inputs; 1149 } 1150 break; 1151 1152 case Bytecodes::_ireturn: 1153 case Bytecodes::_lreturn: 1154 case Bytecodes::_freturn: 1155 case Bytecodes::_dreturn: 1156 case Bytecodes::_areturn: 1157 assert(rsize() == -depth, ""); 1158 inputs = -depth; 1159 break; 1160 1161 case Bytecodes::_jsr: 1162 case Bytecodes::_jsr_w: 1163 inputs = 0; 1164 depth = 1; // S.B. depth=1, not zero 1165 break; 1166 1167 default: 1168 // bytecode produces a typed result 1169 inputs = rsize() - depth; 1170 assert(inputs >= 0, ""); 1171 break; 1172 } 1173 1174 #ifdef ASSERT 1175 // spot check 1176 int outputs = depth + inputs; 1177 assert(outputs >= 0, "sanity"); 1178 switch (code) { 1179 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1180 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1181 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1182 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1183 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1184 default: break; 1185 } 1186 #endif //ASSERT 1187 1188 return true; 1189 } 1190 1191 1192 1193 //------------------------------basic_plus_adr--------------------------------- 1194 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1195 // short-circuit a common case 1196 if (offset == intcon(0)) return ptr; 1197 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1198 } 1199 1200 Node* GraphKit::ConvI2L(Node* offset) { 1201 // short-circuit a common case 1202 jint offset_con = find_int_con(offset, Type::OffsetBot); 1203 if (offset_con != Type::OffsetBot) { 1204 return longcon((jlong) offset_con); 1205 } 1206 return _gvn.transform( new ConvI2LNode(offset)); 1207 } 1208 1209 Node* GraphKit::ConvI2UL(Node* offset) { 1210 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1211 if (offset_con != (juint) Type::OffsetBot) { 1212 return longcon((julong) offset_con); 1213 } 1214 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1215 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1216 return _gvn.transform( new AndLNode(conv, mask) ); 1217 } 1218 1219 Node* GraphKit::ConvL2I(Node* offset) { 1220 // short-circuit a common case 1221 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1222 if (offset_con != (jlong)Type::OffsetBot) { 1223 return intcon((int) offset_con); 1224 } 1225 return _gvn.transform( new ConvL2INode(offset)); 1226 } 1227 1228 //-------------------------load_object_klass----------------------------------- 1229 Node* GraphKit::load_object_klass(Node* obj) { 1230 // Special-case a fresh allocation to avoid building nodes: 1231 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1232 if (akls != nullptr) return akls; 1233 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1234 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 1235 } 1236 1237 //-------------------------load_array_length----------------------------------- 1238 Node* GraphKit::load_array_length(Node* array) { 1239 // Special-case a fresh allocation to avoid building nodes: 1240 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array); 1241 Node *alen; 1242 if (alloc == nullptr) { 1243 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1244 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS)); 1245 } else { 1246 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false); 1247 } 1248 return alen; 1249 } 1250 1251 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc, 1252 const TypeOopPtr* oop_type, 1253 bool replace_length_in_map) { 1254 Node* length = alloc->Ideal_length(); 1255 if (replace_length_in_map == false || map()->find_edge(length) >= 0) { 1256 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn); 1257 if (ccast != length) { 1258 // do not transform ccast here, it might convert to top node for 1259 // negative array length and break assumptions in parsing stage. 1260 _gvn.set_type_bottom(ccast); 1261 record_for_igvn(ccast); 1262 if (replace_length_in_map) { 1263 replace_in_map(length, ccast); 1264 } 1265 return ccast; 1266 } 1267 } 1268 return length; 1269 } 1270 1271 //------------------------------do_null_check---------------------------------- 1272 // Helper function to do a null pointer check. Returned value is 1273 // the incoming address with null casted away. You are allowed to use the 1274 // not-null value only if you are control dependent on the test. 1275 #ifndef PRODUCT 1276 extern uint explicit_null_checks_inserted, 1277 explicit_null_checks_elided; 1278 #endif 1279 Node* GraphKit::null_check_common(Node* value, BasicType type, 1280 // optional arguments for variations: 1281 bool assert_null, 1282 Node* *null_control, 1283 bool speculative, 1284 bool is_init_check) { 1285 assert(!assert_null || null_control == nullptr, "not both at once"); 1286 if (stopped()) return top(); 1287 NOT_PRODUCT(explicit_null_checks_inserted++); 1288 1289 if (value->is_InlineType()) { 1290 // Null checking a scalarized but nullable inline type. Check the IsInit 1291 // input instead of the oop input to avoid keeping buffer allocations alive. 1292 InlineTypeNode* vtptr = value->as_InlineType(); 1293 while (vtptr->get_oop()->is_InlineType()) { 1294 vtptr = vtptr->get_oop()->as_InlineType(); 1295 } 1296 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true); 1297 if (stopped()) { 1298 return top(); 1299 } 1300 if (assert_null) { 1301 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts) 1302 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass()); 1303 // replace_in_map(value, vtptr); 1304 // return vtptr; 1305 replace_in_map(value, null()); 1306 return null(); 1307 } 1308 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top()); 1309 return cast_not_null(value, do_replace_in_map); 1310 } 1311 1312 // Construct null check 1313 Node *chk = nullptr; 1314 switch(type) { 1315 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1316 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1317 case T_ARRAY : // fall through 1318 type = T_OBJECT; // simplify further tests 1319 case T_OBJECT : { 1320 const Type *t = _gvn.type( value ); 1321 1322 const TypeOopPtr* tp = t->isa_oopptr(); 1323 if (tp != nullptr && !tp->is_loaded() 1324 // Only for do_null_check, not any of its siblings: 1325 && !assert_null && null_control == nullptr) { 1326 // Usually, any field access or invocation on an unloaded oop type 1327 // will simply fail to link, since the statically linked class is 1328 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1329 // the static class is loaded but the sharper oop type is not. 1330 // Rather than checking for this obscure case in lots of places, 1331 // we simply observe that a null check on an unloaded class 1332 // will always be followed by a nonsense operation, so we 1333 // can just issue the uncommon trap here. 1334 // Our access to the unloaded class will only be correct 1335 // after it has been loaded and initialized, which requires 1336 // a trip through the interpreter. 1337 ciKlass* klass = tp->unloaded_klass(); 1338 #ifndef PRODUCT 1339 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); } 1340 #endif 1341 uncommon_trap(Deoptimization::Reason_unloaded, 1342 Deoptimization::Action_reinterpret, 1343 klass, "!loaded"); 1344 return top(); 1345 } 1346 1347 if (assert_null) { 1348 // See if the type is contained in NULL_PTR. 1349 // If so, then the value is already null. 1350 if (t->higher_equal(TypePtr::NULL_PTR)) { 1351 NOT_PRODUCT(explicit_null_checks_elided++); 1352 return value; // Elided null assert quickly! 1353 } 1354 } else { 1355 // See if mixing in the null pointer changes type. 1356 // If so, then the null pointer was not allowed in the original 1357 // type. In other words, "value" was not-null. 1358 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1359 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1360 NOT_PRODUCT(explicit_null_checks_elided++); 1361 return value; // Elided null check quickly! 1362 } 1363 } 1364 chk = new CmpPNode( value, null() ); 1365 break; 1366 } 1367 1368 default: 1369 fatal("unexpected type: %s", type2name(type)); 1370 } 1371 assert(chk != nullptr, "sanity check"); 1372 chk = _gvn.transform(chk); 1373 1374 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1375 BoolNode *btst = new BoolNode( chk, btest); 1376 Node *tst = _gvn.transform( btst ); 1377 1378 //----------- 1379 // if peephole optimizations occurred, a prior test existed. 1380 // If a prior test existed, maybe it dominates as we can avoid this test. 1381 if (tst != btst && type == T_OBJECT) { 1382 // At this point we want to scan up the CFG to see if we can 1383 // find an identical test (and so avoid this test altogether). 1384 Node *cfg = control(); 1385 int depth = 0; 1386 while( depth < 16 ) { // Limit search depth for speed 1387 if( cfg->Opcode() == Op_IfTrue && 1388 cfg->in(0)->in(1) == tst ) { 1389 // Found prior test. Use "cast_not_null" to construct an identical 1390 // CastPP (and hence hash to) as already exists for the prior test. 1391 // Return that casted value. 1392 if (assert_null) { 1393 replace_in_map(value, null()); 1394 return null(); // do not issue the redundant test 1395 } 1396 Node *oldcontrol = control(); 1397 set_control(cfg); 1398 Node *res = cast_not_null(value); 1399 set_control(oldcontrol); 1400 NOT_PRODUCT(explicit_null_checks_elided++); 1401 return res; 1402 } 1403 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1404 if (cfg == nullptr) break; // Quit at region nodes 1405 depth++; 1406 } 1407 } 1408 1409 //----------- 1410 // Branch to failure if null 1411 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1412 Deoptimization::DeoptReason reason; 1413 if (assert_null) { 1414 reason = Deoptimization::reason_null_assert(speculative); 1415 } else if (type == T_OBJECT || is_init_check) { 1416 reason = Deoptimization::reason_null_check(speculative); 1417 } else { 1418 reason = Deoptimization::Reason_div0_check; 1419 } 1420 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1421 // ciMethodData::has_trap_at will return a conservative -1 if any 1422 // must-be-null assertion has failed. This could cause performance 1423 // problems for a method after its first do_null_assert failure. 1424 // Consider using 'Reason_class_check' instead? 1425 1426 // To cause an implicit null check, we set the not-null probability 1427 // to the maximum (PROB_MAX). For an explicit check the probability 1428 // is set to a smaller value. 1429 if (null_control != nullptr || too_many_traps(reason)) { 1430 // probability is less likely 1431 ok_prob = PROB_LIKELY_MAG(3); 1432 } else if (!assert_null && 1433 (ImplicitNullCheckThreshold > 0) && 1434 method() != nullptr && 1435 (method()->method_data()->trap_count(reason) 1436 >= (uint)ImplicitNullCheckThreshold)) { 1437 ok_prob = PROB_LIKELY_MAG(3); 1438 } 1439 1440 if (null_control != nullptr) { 1441 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1442 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1443 set_control( _gvn.transform( new IfTrueNode(iff))); 1444 #ifndef PRODUCT 1445 if (null_true == top()) { 1446 explicit_null_checks_elided++; 1447 } 1448 #endif 1449 (*null_control) = null_true; 1450 } else { 1451 BuildCutout unless(this, tst, ok_prob); 1452 // Check for optimizer eliding test at parse time 1453 if (stopped()) { 1454 // Failure not possible; do not bother making uncommon trap. 1455 NOT_PRODUCT(explicit_null_checks_elided++); 1456 } else if (assert_null) { 1457 uncommon_trap(reason, 1458 Deoptimization::Action_make_not_entrant, 1459 nullptr, "assert_null"); 1460 } else { 1461 replace_in_map(value, zerocon(type)); 1462 builtin_throw(reason); 1463 } 1464 } 1465 1466 // Must throw exception, fall-thru not possible? 1467 if (stopped()) { 1468 return top(); // No result 1469 } 1470 1471 if (assert_null) { 1472 // Cast obj to null on this path. 1473 replace_in_map(value, zerocon(type)); 1474 return zerocon(type); 1475 } 1476 1477 // Cast obj to not-null on this path, if there is no null_control. 1478 // (If there is a null_control, a non-null value may come back to haunt us.) 1479 if (type == T_OBJECT) { 1480 Node* cast = cast_not_null(value, false); 1481 if (null_control == nullptr || (*null_control) == top()) 1482 replace_in_map(value, cast); 1483 value = cast; 1484 } 1485 1486 return value; 1487 } 1488 1489 //------------------------------cast_not_null---------------------------------- 1490 // Cast obj to not-null on this path 1491 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1492 if (obj->is_InlineType()) { 1493 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map); 1494 vt->as_InlineType()->set_is_init(_gvn); 1495 vt = _gvn.transform(vt); 1496 if (do_replace_in_map) { 1497 replace_in_map(obj, vt); 1498 } 1499 return vt; 1500 } 1501 const Type *t = _gvn.type(obj); 1502 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1503 // Object is already not-null? 1504 if( t == t_not_null ) return obj; 1505 1506 Node* cast = new CastPPNode(control(), obj,t_not_null); 1507 cast = _gvn.transform( cast ); 1508 1509 // Scan for instances of 'obj' in the current JVM mapping. 1510 // These instances are known to be not-null after the test. 1511 if (do_replace_in_map) 1512 replace_in_map(obj, cast); 1513 1514 return cast; // Return casted value 1515 } 1516 1517 // Sometimes in intrinsics, we implicitly know an object is not null 1518 // (there's no actual null check) so we can cast it to not null. In 1519 // the course of optimizations, the input to the cast can become null. 1520 // In that case that data path will die and we need the control path 1521 // to become dead as well to keep the graph consistent. So we have to 1522 // add a check for null for which one branch can't be taken. It uses 1523 // an OpaqueNotNull node that will cause the check to be removed after loop 1524 // opts so the test goes away and the compiled code doesn't execute a 1525 // useless check. 1526 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1527 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1528 return value; 1529 } 1530 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1531 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1532 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst)); 1533 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1534 _gvn.set_type(iff, iff->Value(&_gvn)); 1535 if (!tst->is_Con()) { 1536 record_for_igvn(iff); 1537 } 1538 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1539 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1540 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1541 C->root()->add_req(halt); 1542 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1543 set_control(if_t); 1544 return cast_not_null(value, do_replace_in_map); 1545 } 1546 1547 1548 //--------------------------replace_in_map------------------------------------- 1549 void GraphKit::replace_in_map(Node* old, Node* neww) { 1550 if (old == neww) { 1551 return; 1552 } 1553 1554 map()->replace_edge(old, neww); 1555 1556 // Note: This operation potentially replaces any edge 1557 // on the map. This includes locals, stack, and monitors 1558 // of the current (innermost) JVM state. 1559 1560 // don't let inconsistent types from profiling escape this 1561 // method 1562 1563 const Type* told = _gvn.type(old); 1564 const Type* tnew = _gvn.type(neww); 1565 1566 if (!tnew->higher_equal(told)) { 1567 return; 1568 } 1569 1570 map()->record_replaced_node(old, neww); 1571 } 1572 1573 1574 //============================================================================= 1575 //--------------------------------memory--------------------------------------- 1576 Node* GraphKit::memory(uint alias_idx) { 1577 MergeMemNode* mem = merged_memory(); 1578 Node* p = mem->memory_at(alias_idx); 1579 assert(p != mem->empty_memory(), "empty"); 1580 _gvn.set_type(p, Type::MEMORY); // must be mapped 1581 return p; 1582 } 1583 1584 //-----------------------------reset_memory------------------------------------ 1585 Node* GraphKit::reset_memory() { 1586 Node* mem = map()->memory(); 1587 // do not use this node for any more parsing! 1588 debug_only( map()->set_memory((Node*)nullptr) ); 1589 return _gvn.transform( mem ); 1590 } 1591 1592 //------------------------------set_all_memory--------------------------------- 1593 void GraphKit::set_all_memory(Node* newmem) { 1594 Node* mergemem = MergeMemNode::make(newmem); 1595 gvn().set_type_bottom(mergemem); 1596 map()->set_memory(mergemem); 1597 } 1598 1599 //------------------------------set_all_memory_call---------------------------- 1600 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1601 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1602 set_all_memory(newmem); 1603 } 1604 1605 //============================================================================= 1606 // 1607 // parser factory methods for MemNodes 1608 // 1609 // These are layered on top of the factory methods in LoadNode and StoreNode, 1610 // and integrate with the parser's memory state and _gvn engine. 1611 // 1612 1613 // factory methods in "int adr_idx" 1614 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1615 MemNode::MemOrd mo, 1616 LoadNode::ControlDependency control_dependency, 1617 bool require_atomic_access, 1618 bool unaligned, 1619 bool mismatched, 1620 bool unsafe, 1621 uint8_t barrier_data) { 1622 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr()); 1623 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1624 const TypePtr* adr_type = nullptr; // debug-mode-only argument 1625 debug_only(adr_type = C->get_adr_type(adr_idx)); 1626 Node* mem = memory(adr_idx); 1627 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data); 1628 ld = _gvn.transform(ld); 1629 1630 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1631 // Improve graph before escape analysis and boxing elimination. 1632 record_for_igvn(ld); 1633 if (ld->is_DecodeN()) { 1634 // Also record the actual load (LoadN) in case ld is DecodeN. In some 1635 // rare corner cases, ld->in(1) can be something other than LoadN (e.g., 1636 // a Phi). Recording such cases is still perfectly sound, but may be 1637 // unnecessary and result in some minor IGVN overhead. 1638 record_for_igvn(ld->in(1)); 1639 } 1640 } 1641 return ld; 1642 } 1643 1644 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1645 MemNode::MemOrd mo, 1646 bool require_atomic_access, 1647 bool unaligned, 1648 bool mismatched, 1649 bool unsafe, 1650 int barrier_data) { 1651 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr()); 1652 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1653 const TypePtr* adr_type = nullptr; 1654 debug_only(adr_type = C->get_adr_type(adr_idx)); 1655 Node *mem = memory(adr_idx); 1656 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access); 1657 if (unaligned) { 1658 st->as_Store()->set_unaligned_access(); 1659 } 1660 if (mismatched) { 1661 st->as_Store()->set_mismatched_access(); 1662 } 1663 if (unsafe) { 1664 st->as_Store()->set_unsafe_access(); 1665 } 1666 st->as_Store()->set_barrier_data(barrier_data); 1667 st = _gvn.transform(st); 1668 set_memory(st, adr_idx); 1669 // Back-to-back stores can only remove intermediate store with DU info 1670 // so push on worklist for optimizer. 1671 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1672 record_for_igvn(st); 1673 1674 return st; 1675 } 1676 1677 Node* GraphKit::access_store_at(Node* obj, 1678 Node* adr, 1679 const TypePtr* adr_type, 1680 Node* val, 1681 const Type* val_type, 1682 BasicType bt, 1683 DecoratorSet decorators, 1684 bool safe_for_replace, 1685 const InlineTypeNode* vt) { 1686 // Transformation of a value which could be null pointer (CastPP #null) 1687 // could be delayed during Parse (for example, in adjust_map_after_if()). 1688 // Execute transformation here to avoid barrier generation in such case. 1689 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1690 val = _gvn.makecon(TypePtr::NULL_PTR); 1691 } 1692 1693 if (stopped()) { 1694 return top(); // Dead path ? 1695 } 1696 1697 assert(val != nullptr, "not dead path"); 1698 if (val->is_InlineType()) { 1699 // Store to non-flat field. Buffer the inline type and make sure 1700 // the store is re-executed if the allocation triggers deoptimization. 1701 PreserveReexecuteState preexecs(this); 1702 jvms()->set_should_reexecute(true); 1703 val = val->as_InlineType()->buffer(this, safe_for_replace); 1704 } 1705 1706 C2AccessValuePtr addr(adr, adr_type); 1707 C2AccessValue value(val, val_type); 1708 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt); 1709 if (access.is_raw()) { 1710 return _barrier_set->BarrierSetC2::store_at(access, value); 1711 } else { 1712 return _barrier_set->store_at(access, value); 1713 } 1714 } 1715 1716 Node* GraphKit::access_load_at(Node* obj, // containing obj 1717 Node* adr, // actual address to store val at 1718 const TypePtr* adr_type, 1719 const Type* val_type, 1720 BasicType bt, 1721 DecoratorSet decorators, 1722 Node* ctl) { 1723 if (stopped()) { 1724 return top(); // Dead path ? 1725 } 1726 1727 C2AccessValuePtr addr(adr, adr_type); 1728 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl); 1729 if (access.is_raw()) { 1730 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1731 } else { 1732 return _barrier_set->load_at(access, val_type); 1733 } 1734 } 1735 1736 Node* GraphKit::access_load(Node* adr, // actual address to load val at 1737 const Type* val_type, 1738 BasicType bt, 1739 DecoratorSet decorators) { 1740 if (stopped()) { 1741 return top(); // Dead path ? 1742 } 1743 1744 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr()); 1745 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr); 1746 if (access.is_raw()) { 1747 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1748 } else { 1749 return _barrier_set->load_at(access, val_type); 1750 } 1751 } 1752 1753 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1754 Node* adr, 1755 const TypePtr* adr_type, 1756 int alias_idx, 1757 Node* expected_val, 1758 Node* new_val, 1759 const Type* value_type, 1760 BasicType bt, 1761 DecoratorSet decorators) { 1762 C2AccessValuePtr addr(adr, adr_type); 1763 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1764 bt, obj, addr, alias_idx); 1765 if (access.is_raw()) { 1766 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1767 } else { 1768 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1769 } 1770 } 1771 1772 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1773 Node* adr, 1774 const TypePtr* adr_type, 1775 int alias_idx, 1776 Node* expected_val, 1777 Node* new_val, 1778 const Type* value_type, 1779 BasicType bt, 1780 DecoratorSet decorators) { 1781 C2AccessValuePtr addr(adr, adr_type); 1782 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1783 bt, obj, addr, alias_idx); 1784 if (access.is_raw()) { 1785 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1786 } else { 1787 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1788 } 1789 } 1790 1791 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1792 Node* adr, 1793 const TypePtr* adr_type, 1794 int alias_idx, 1795 Node* new_val, 1796 const Type* value_type, 1797 BasicType bt, 1798 DecoratorSet decorators) { 1799 C2AccessValuePtr addr(adr, adr_type); 1800 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1801 bt, obj, addr, alias_idx); 1802 if (access.is_raw()) { 1803 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1804 } else { 1805 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1806 } 1807 } 1808 1809 Node* GraphKit::access_atomic_add_at(Node* obj, 1810 Node* adr, 1811 const TypePtr* adr_type, 1812 int alias_idx, 1813 Node* new_val, 1814 const Type* value_type, 1815 BasicType bt, 1816 DecoratorSet decorators) { 1817 C2AccessValuePtr addr(adr, adr_type); 1818 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1819 if (access.is_raw()) { 1820 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1821 } else { 1822 return _barrier_set->atomic_add_at(access, new_val, value_type); 1823 } 1824 } 1825 1826 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) { 1827 return _barrier_set->clone(this, src, dst, size, is_array); 1828 } 1829 1830 //-------------------------array_element_address------------------------- 1831 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1832 const TypeInt* sizetype, Node* ctrl) { 1833 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 1834 assert(!arytype->is_flat() || elembt == T_OBJECT, "element type of flat arrays are T_OBJECT"); 1835 uint shift; 1836 if (arytype->is_flat() && arytype->klass_is_exact()) { 1837 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different 1838 // value classes at runtime with a potentially different layout. The caller needs to fall back to call 1839 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that 1840 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even 1841 // though we don't need the address node in this case and throw it away again. 1842 shift = arytype->flat_log_elem_size(); 1843 } else { 1844 shift = exact_log2(type2aelembytes(elembt)); 1845 } 1846 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1847 1848 // short-circuit a common case (saves lots of confusing waste motion) 1849 jint idx_con = find_int_con(idx, -1); 1850 if (idx_con >= 0) { 1851 intptr_t offset = header + ((intptr_t)idx_con << shift); 1852 return basic_plus_adr(ary, offset); 1853 } 1854 1855 // must be correct type for alignment purposes 1856 Node* base = basic_plus_adr(ary, header); 1857 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1858 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1859 return basic_plus_adr(ary, base, scale); 1860 } 1861 1862 //-------------------------load_array_element------------------------- 1863 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) { 1864 const Type* elemtype = arytype->elem(); 1865 BasicType elembt = elemtype->array_element_basic_type(); 1866 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1867 if (elembt == T_NARROWOOP) { 1868 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1869 } 1870 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt, 1871 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0)); 1872 return ld; 1873 } 1874 1875 //-------------------------set_arguments_for_java_call------------------------- 1876 // Arguments (pre-popped from the stack) are taken from the JVMS. 1877 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) { 1878 PreserveReexecuteState preexecs(this); 1879 if (EnableValhalla) { 1880 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization. 1881 // At this point, the call hasn't been executed yet, so we will only ever execute the call once. 1882 jvms()->set_should_reexecute(true); 1883 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc()); 1884 inc_sp(arg_size); 1885 } 1886 // Add the call arguments 1887 const TypeTuple* domain = call->tf()->domain_sig(); 1888 uint nargs = domain->cnt(); 1889 int arg_num = 0; 1890 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1891 Node* arg = argument(i-TypeFunc::Parms); 1892 const Type* t = domain->field_at(i); 1893 // TODO 8284443 A static call to a mismatched method should still be scalarized 1894 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) { 1895 // We don't pass inline type arguments by reference but instead pass each field of the inline type 1896 if (!arg->is_InlineType()) { 1897 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type"); 1898 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free()); 1899 } 1900 InlineTypeNode* vt = arg->as_InlineType(); 1901 vt->pass_fields(this, call, idx, true, !t->maybe_null()); 1902 // If an inline type argument is passed as fields, attach the Method* to the call site 1903 // to be able to access the extended signature later via attached_method_before_pc(). 1904 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1905 call->set_override_symbolic_info(true); 1906 // Register an evol dependency on the callee method to make sure that this method is deoptimized and 1907 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched. 1908 C->dependencies()->assert_evol_method(call->method()); 1909 arg_num++; 1910 continue; 1911 } else if (arg->is_InlineType()) { 1912 // Pass inline type argument via oop to callee 1913 InlineTypeNode* inline_type = arg->as_InlineType(); 1914 const ciMethod* method = call->method(); 1915 ciInstanceKlass* holder = method->holder(); 1916 const bool is_receiver = (i == TypeFunc::Parms); 1917 const bool is_abstract_or_object_klass_constructor = method->is_object_constructor() && 1918 (holder->is_abstract() || holder->is_java_lang_Object()); 1919 const bool is_larval_receiver_on_super_constructor = is_receiver && is_abstract_or_object_klass_constructor; 1920 bool must_init_buffer = true; 1921 // We always need to buffer inline types when they are escaping. However, we can skip the actual initialization 1922 // of the buffer if the inline type is a larval because we are going to update the buffer anyway which requires 1923 // us to create a new one. But there is one special case where we are still required to initialize the buffer: 1924 // When we have a larval receiver invoked on an abstract (value class) constructor or the Object constructor (that 1925 // is not going to be inlined). After this call, the larval is completely initialized and thus not a larval anymore. 1926 // We therefore need to force an initialization of the buffer to not lose all the field writes so far in case the 1927 // buffer needs to be used (e.g. to read from when deoptimizing at runtime) or further updated in abstract super 1928 // value class constructors which could have more fields to be initialized. Note that we do not need to 1929 // initialize the buffer when invoking another constructor in the same class on a larval receiver because we 1930 // have not initialized any fields, yet (this is done completely by the other constructor call). 1931 if (inline_type->is_larval() && !is_larval_receiver_on_super_constructor) { 1932 must_init_buffer = false; 1933 } 1934 arg = inline_type->buffer(this, true, must_init_buffer); 1935 } 1936 if (t != Type::HALF) { 1937 arg_num++; 1938 } 1939 call->init_req(idx++, arg); 1940 } 1941 } 1942 1943 //---------------------------set_edges_for_java_call--------------------------- 1944 // Connect a newly created call into the current JVMS. 1945 // A return value node (if any) is returned from set_edges_for_java_call. 1946 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1947 1948 // Add the predefined inputs: 1949 call->init_req( TypeFunc::Control, control() ); 1950 call->init_req( TypeFunc::I_O , i_o() ); 1951 call->init_req( TypeFunc::Memory , reset_memory() ); 1952 call->init_req( TypeFunc::FramePtr, frameptr() ); 1953 call->init_req( TypeFunc::ReturnAdr, top() ); 1954 1955 add_safepoint_edges(call, must_throw); 1956 1957 Node* xcall = _gvn.transform(call); 1958 1959 if (xcall == top()) { 1960 set_control(top()); 1961 return; 1962 } 1963 assert(xcall == call, "call identity is stable"); 1964 1965 // Re-use the current map to produce the result. 1966 1967 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1968 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1969 set_all_memory_call(xcall, separate_io_proj); 1970 1971 //return xcall; // no need, caller already has it 1972 } 1973 1974 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1975 if (stopped()) return top(); // maybe the call folded up? 1976 1977 // Note: Since any out-of-line call can produce an exception, 1978 // we always insert an I_O projection from the call into the result. 1979 1980 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1981 1982 if (separate_io_proj) { 1983 // The caller requested separate projections be used by the fall 1984 // through and exceptional paths, so replace the projections for 1985 // the fall through path. 1986 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1987 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1988 } 1989 1990 // Capture the return value, if any. 1991 Node* ret; 1992 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) { 1993 ret = top(); 1994 } else if (call->tf()->returns_inline_type_as_fields()) { 1995 // Return of multiple values (inline type fields): we create a 1996 // InlineType node, each field is a projection from the call. 1997 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass(); 1998 uint base_input = TypeFunc::Parms; 1999 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false); 2000 } else { 2001 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 2002 ciType* t = call->method()->return_type(); 2003 if (t->is_klass()) { 2004 const Type* type = TypeOopPtr::make_from_klass(t->as_klass()); 2005 if (type->is_inlinetypeptr()) { 2006 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass(), type->inline_klass()->is_null_free()); 2007 } 2008 } 2009 } 2010 2011 // We just called the constructor on a value type receiver. Reload it from the buffer 2012 ciMethod* method = call->method(); 2013 if (method->is_object_constructor() && !method->holder()->is_java_lang_Object()) { 2014 InlineTypeNode* inline_type_receiver = call->in(TypeFunc::Parms)->isa_InlineType(); 2015 if (inline_type_receiver != nullptr) { 2016 assert(inline_type_receiver->is_larval(), "must be larval"); 2017 assert(inline_type_receiver->is_allocated(&gvn()), "larval must be buffered"); 2018 InlineTypeNode* reloaded = InlineTypeNode::make_from_oop(this, inline_type_receiver->get_oop(), 2019 inline_type_receiver->bottom_type()->inline_klass(), true); 2020 assert(!reloaded->is_larval(), "should not be larval anymore"); 2021 replace_in_map(inline_type_receiver, reloaded); 2022 } 2023 } 2024 2025 return ret; 2026 } 2027 2028 //--------------------set_predefined_input_for_runtime_call-------------------- 2029 // Reading and setting the memory state is way conservative here. 2030 // The real problem is that I am not doing real Type analysis on memory, 2031 // so I cannot distinguish card mark stores from other stores. Across a GC 2032 // point the Store Barrier and the card mark memory has to agree. I cannot 2033 // have a card mark store and its barrier split across the GC point from 2034 // either above or below. Here I get that to happen by reading ALL of memory. 2035 // A better answer would be to separate out card marks from other memory. 2036 // For now, return the input memory state, so that it can be reused 2037 // after the call, if this call has restricted memory effects. 2038 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 2039 // Set fixed predefined input arguments 2040 Node* memory = reset_memory(); 2041 Node* m = narrow_mem == nullptr ? memory : narrow_mem; 2042 call->init_req( TypeFunc::Control, control() ); 2043 call->init_req( TypeFunc::I_O, top() ); // does no i/o 2044 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 2045 call->init_req( TypeFunc::FramePtr, frameptr() ); 2046 call->init_req( TypeFunc::ReturnAdr, top() ); 2047 return memory; 2048 } 2049 2050 //-------------------set_predefined_output_for_runtime_call-------------------- 2051 // Set control and memory (not i_o) from the call. 2052 // If keep_mem is not null, use it for the output state, 2053 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 2054 // If hook_mem is null, this call produces no memory effects at all. 2055 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 2056 // then only that memory slice is taken from the call. 2057 // In the last case, we must put an appropriate memory barrier before 2058 // the call, so as to create the correct anti-dependencies on loads 2059 // preceding the call. 2060 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 2061 Node* keep_mem, 2062 const TypePtr* hook_mem) { 2063 // no i/o 2064 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 2065 if (keep_mem) { 2066 // First clone the existing memory state 2067 set_all_memory(keep_mem); 2068 if (hook_mem != nullptr) { 2069 // Make memory for the call 2070 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 2071 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 2072 // We also use hook_mem to extract specific effects from arraycopy stubs. 2073 set_memory(mem, hook_mem); 2074 } 2075 // ...else the call has NO memory effects. 2076 2077 // Make sure the call advertises its memory effects precisely. 2078 // This lets us build accurate anti-dependences in gcm.cpp. 2079 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 2080 "call node must be constructed correctly"); 2081 } else { 2082 assert(hook_mem == nullptr, ""); 2083 // This is not a "slow path" call; all memory comes from the call. 2084 set_all_memory_call(call); 2085 } 2086 } 2087 2088 // Keep track of MergeMems feeding into other MergeMems 2089 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 2090 if (!mem->is_MergeMem()) { 2091 return; 2092 } 2093 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 2094 Node* use = i.get(); 2095 if (use->is_MergeMem()) { 2096 wl.push(use); 2097 } 2098 } 2099 } 2100 2101 // Replace the call with the current state of the kit. 2102 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) { 2103 JVMState* ejvms = nullptr; 2104 if (has_exceptions()) { 2105 ejvms = transfer_exceptions_into_jvms(); 2106 } 2107 2108 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 2109 ReplacedNodes replaced_nodes_exception; 2110 Node* ex_ctl = top(); 2111 2112 SafePointNode* final_state = stop(); 2113 2114 // Find all the needed outputs of this call 2115 CallProjections* callprojs = call->extract_projections(true, do_asserts); 2116 2117 Unique_Node_List wl; 2118 Node* init_mem = call->in(TypeFunc::Memory); 2119 Node* final_mem = final_state->in(TypeFunc::Memory); 2120 Node* final_ctl = final_state->in(TypeFunc::Control); 2121 Node* final_io = final_state->in(TypeFunc::I_O); 2122 2123 // Replace all the old call edges with the edges from the inlining result 2124 if (callprojs->fallthrough_catchproj != nullptr) { 2125 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl); 2126 } 2127 if (callprojs->fallthrough_memproj != nullptr) { 2128 if (final_mem->is_MergeMem()) { 2129 // Parser's exits MergeMem was not transformed but may be optimized 2130 final_mem = _gvn.transform(final_mem); 2131 } 2132 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem); 2133 add_mergemem_users_to_worklist(wl, final_mem); 2134 } 2135 if (callprojs->fallthrough_ioproj != nullptr) { 2136 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io); 2137 } 2138 2139 // Replace the result with the new result if it exists and is used 2140 if (callprojs->resproj[0] != nullptr && result != nullptr) { 2141 // If the inlined code is dead, the result projections for an inline type returned as 2142 // fields have not been replaced. They will go away once the call is replaced by TOP below. 2143 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()), 2144 "unexpected number of results"); 2145 C->gvn_replace_by(callprojs->resproj[0], result); 2146 } 2147 2148 if (ejvms == nullptr) { 2149 // No exception edges to simply kill off those paths 2150 if (callprojs->catchall_catchproj != nullptr) { 2151 C->gvn_replace_by(callprojs->catchall_catchproj, C->top()); 2152 } 2153 if (callprojs->catchall_memproj != nullptr) { 2154 C->gvn_replace_by(callprojs->catchall_memproj, C->top()); 2155 } 2156 if (callprojs->catchall_ioproj != nullptr) { 2157 C->gvn_replace_by(callprojs->catchall_ioproj, C->top()); 2158 } 2159 // Replace the old exception object with top 2160 if (callprojs->exobj != nullptr) { 2161 C->gvn_replace_by(callprojs->exobj, C->top()); 2162 } 2163 } else { 2164 GraphKit ekit(ejvms); 2165 2166 // Load my combined exception state into the kit, with all phis transformed: 2167 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 2168 replaced_nodes_exception = ex_map->replaced_nodes(); 2169 2170 Node* ex_oop = ekit.use_exception_state(ex_map); 2171 2172 if (callprojs->catchall_catchproj != nullptr) { 2173 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control()); 2174 ex_ctl = ekit.control(); 2175 } 2176 if (callprojs->catchall_memproj != nullptr) { 2177 Node* ex_mem = ekit.reset_memory(); 2178 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem); 2179 add_mergemem_users_to_worklist(wl, ex_mem); 2180 } 2181 if (callprojs->catchall_ioproj != nullptr) { 2182 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o()); 2183 } 2184 2185 // Replace the old exception object with the newly created one 2186 if (callprojs->exobj != nullptr) { 2187 C->gvn_replace_by(callprojs->exobj, ex_oop); 2188 } 2189 } 2190 2191 // Disconnect the call from the graph 2192 call->disconnect_inputs(C); 2193 C->gvn_replace_by(call, C->top()); 2194 2195 // Clean up any MergeMems that feed other MergeMems since the 2196 // optimizer doesn't like that. 2197 while (wl.size() > 0) { 2198 _gvn.transform(wl.pop()); 2199 } 2200 2201 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) { 2202 replaced_nodes.apply(C, final_ctl); 2203 } 2204 if (!ex_ctl->is_top() && do_replaced_nodes) { 2205 replaced_nodes_exception.apply(C, ex_ctl); 2206 } 2207 } 2208 2209 2210 //------------------------------increment_counter------------------------------ 2211 // for statistics: increment a VM counter by 1 2212 2213 void GraphKit::increment_counter(address counter_addr) { 2214 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2215 increment_counter(adr1); 2216 } 2217 2218 void GraphKit::increment_counter(Node* counter_addr) { 2219 Node* ctrl = control(); 2220 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered); 2221 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1))); 2222 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered); 2223 } 2224 2225 2226 //------------------------------uncommon_trap---------------------------------- 2227 // Bail out to the interpreter in mid-method. Implemented by calling the 2228 // uncommon_trap blob. This helper function inserts a runtime call with the 2229 // right debug info. 2230 Node* GraphKit::uncommon_trap(int trap_request, 2231 ciKlass* klass, const char* comment, 2232 bool must_throw, 2233 bool keep_exact_action) { 2234 if (failing_internal()) { 2235 stop(); 2236 } 2237 if (stopped()) return nullptr; // trap reachable? 2238 2239 // Note: If ProfileTraps is true, and if a deopt. actually 2240 // occurs here, the runtime will make sure an MDO exists. There is 2241 // no need to call method()->ensure_method_data() at this point. 2242 2243 // Set the stack pointer to the right value for reexecution: 2244 set_sp(reexecute_sp()); 2245 2246 #ifdef ASSERT 2247 if (!must_throw) { 2248 // Make sure the stack has at least enough depth to execute 2249 // the current bytecode. 2250 int inputs, ignored_depth; 2251 if (compute_stack_effects(inputs, ignored_depth)) { 2252 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2253 Bytecodes::name(java_bc()), sp(), inputs); 2254 } 2255 } 2256 #endif 2257 2258 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2259 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2260 2261 switch (action) { 2262 case Deoptimization::Action_maybe_recompile: 2263 case Deoptimization::Action_reinterpret: 2264 // Temporary fix for 6529811 to allow virtual calls to be sure they 2265 // get the chance to go from mono->bi->mega 2266 if (!keep_exact_action && 2267 Deoptimization::trap_request_index(trap_request) < 0 && 2268 too_many_recompiles(reason)) { 2269 // This BCI is causing too many recompilations. 2270 if (C->log() != nullptr) { 2271 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2272 Deoptimization::trap_reason_name(reason), 2273 Deoptimization::trap_action_name(action)); 2274 } 2275 action = Deoptimization::Action_none; 2276 trap_request = Deoptimization::make_trap_request(reason, action); 2277 } else { 2278 C->set_trap_can_recompile(true); 2279 } 2280 break; 2281 case Deoptimization::Action_make_not_entrant: 2282 C->set_trap_can_recompile(true); 2283 break; 2284 case Deoptimization::Action_none: 2285 case Deoptimization::Action_make_not_compilable: 2286 break; 2287 default: 2288 #ifdef ASSERT 2289 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2290 #endif 2291 break; 2292 } 2293 2294 if (TraceOptoParse) { 2295 char buf[100]; 2296 tty->print_cr("Uncommon trap %s at bci:%d", 2297 Deoptimization::format_trap_request(buf, sizeof(buf), 2298 trap_request), bci()); 2299 } 2300 2301 CompileLog* log = C->log(); 2302 if (log != nullptr) { 2303 int kid = (klass == nullptr)? -1: log->identify(klass); 2304 log->begin_elem("uncommon_trap bci='%d'", bci()); 2305 char buf[100]; 2306 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2307 trap_request)); 2308 if (kid >= 0) log->print(" klass='%d'", kid); 2309 if (comment != nullptr) log->print(" comment='%s'", comment); 2310 log->end_elem(); 2311 } 2312 2313 // Make sure any guarding test views this path as very unlikely 2314 Node *i0 = control()->in(0); 2315 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test? 2316 IfNode *iff = i0->as_If(); 2317 float f = iff->_prob; // Get prob 2318 if (control()->Opcode() == Op_IfTrue) { 2319 if (f > PROB_UNLIKELY_MAG(4)) 2320 iff->_prob = PROB_MIN; 2321 } else { 2322 if (f < PROB_LIKELY_MAG(4)) 2323 iff->_prob = PROB_MAX; 2324 } 2325 } 2326 2327 // Clear out dead values from the debug info. 2328 kill_dead_locals(); 2329 2330 // Now insert the uncommon trap subroutine call 2331 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point(); 2332 const TypePtr* no_memory_effects = nullptr; 2333 // Pass the index of the class to be loaded 2334 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2335 (must_throw ? RC_MUST_THROW : 0), 2336 OptoRuntime::uncommon_trap_Type(), 2337 call_addr, "uncommon_trap", no_memory_effects, 2338 intcon(trap_request)); 2339 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2340 "must extract request correctly from the graph"); 2341 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2342 2343 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2344 // The debug info is the only real input to this call. 2345 2346 // Halt-and-catch fire here. The above call should never return! 2347 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2348 PRODUCT_ONLY(COMMA /*reachable*/false)); 2349 _gvn.set_type_bottom(halt); 2350 root()->add_req(halt); 2351 2352 stop_and_kill_map(); 2353 return call; 2354 } 2355 2356 2357 //--------------------------just_allocated_object------------------------------ 2358 // Report the object that was just allocated. 2359 // It must be the case that there are no intervening safepoints. 2360 // We use this to determine if an object is so "fresh" that 2361 // it does not require card marks. 2362 Node* GraphKit::just_allocated_object(Node* current_control) { 2363 Node* ctrl = current_control; 2364 // Object::<init> is invoked after allocation, most of invoke nodes 2365 // will be reduced, but a region node is kept in parse time, we check 2366 // the pattern and skip the region node if it degraded to a copy. 2367 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 && 2368 ctrl->as_Region()->is_copy()) { 2369 ctrl = ctrl->as_Region()->is_copy(); 2370 } 2371 if (C->recent_alloc_ctl() == ctrl) { 2372 return C->recent_alloc_obj(); 2373 } 2374 return nullptr; 2375 } 2376 2377 2378 /** 2379 * Record profiling data exact_kls for Node n with the type system so 2380 * that it can propagate it (speculation) 2381 * 2382 * @param n node that the type applies to 2383 * @param exact_kls type from profiling 2384 * @param maybe_null did profiling see null? 2385 * 2386 * @return node with improved type 2387 */ 2388 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2389 const Type* current_type = _gvn.type(n); 2390 assert(UseTypeSpeculation, "type speculation must be on"); 2391 2392 const TypePtr* speculative = current_type->speculative(); 2393 2394 // Should the klass from the profile be recorded in the speculative type? 2395 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2396 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces); 2397 const TypeOopPtr* xtype = tklass->as_instance_type(); 2398 assert(xtype->klass_is_exact(), "Should be exact"); 2399 // Any reason to believe n is not null (from this profiling or a previous one)? 2400 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2401 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2402 // record the new speculative type's depth 2403 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2404 speculative = speculative->with_inline_depth(jvms()->depth()); 2405 } else if (current_type->would_improve_ptr(ptr_kind)) { 2406 // Profiling report that null was never seen so we can change the 2407 // speculative type to non null ptr. 2408 if (ptr_kind == ProfileAlwaysNull) { 2409 speculative = TypePtr::NULL_PTR; 2410 } else { 2411 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2412 const TypePtr* ptr = TypePtr::NOTNULL; 2413 if (speculative != nullptr) { 2414 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2415 } else { 2416 speculative = ptr; 2417 } 2418 } 2419 } 2420 2421 if (speculative != current_type->speculative()) { 2422 // Build a type with a speculative type (what we think we know 2423 // about the type but will need a guard when we use it) 2424 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2425 // We're changing the type, we need a new CheckCast node to carry 2426 // the new type. The new type depends on the control: what 2427 // profiling tells us is only valid from here as far as we can 2428 // tell. 2429 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2430 cast = _gvn.transform(cast); 2431 replace_in_map(n, cast); 2432 n = cast; 2433 } 2434 2435 return n; 2436 } 2437 2438 /** 2439 * Record profiling data from receiver profiling at an invoke with the 2440 * type system so that it can propagate it (speculation) 2441 * 2442 * @param n receiver node 2443 * 2444 * @return node with improved type 2445 */ 2446 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2447 if (!UseTypeSpeculation) { 2448 return n; 2449 } 2450 ciKlass* exact_kls = profile_has_unique_klass(); 2451 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2452 if ((java_bc() == Bytecodes::_checkcast || 2453 java_bc() == Bytecodes::_instanceof || 2454 java_bc() == Bytecodes::_aastore) && 2455 method()->method_data()->is_mature()) { 2456 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2457 if (data != nullptr) { 2458 if (java_bc() == Bytecodes::_aastore) { 2459 ciKlass* array_type = nullptr; 2460 ciKlass* element_type = nullptr; 2461 ProfilePtrKind element_ptr = ProfileMaybeNull; 2462 bool flat_array = true; 2463 bool null_free_array = true; 2464 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 2465 exact_kls = element_type; 2466 ptr_kind = element_ptr; 2467 } else { 2468 if (!data->as_BitData()->null_seen()) { 2469 ptr_kind = ProfileNeverNull; 2470 } else { 2471 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2472 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2473 uint i = 0; 2474 for (; i < call->row_limit(); i++) { 2475 ciKlass* receiver = call->receiver(i); 2476 if (receiver != nullptr) { 2477 break; 2478 } 2479 } 2480 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2481 } 2482 } 2483 } 2484 } 2485 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2486 } 2487 2488 /** 2489 * Record profiling data from argument profiling at an invoke with the 2490 * type system so that it can propagate it (speculation) 2491 * 2492 * @param dest_method target method for the call 2493 * @param bc what invoke bytecode is this? 2494 */ 2495 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2496 if (!UseTypeSpeculation) { 2497 return; 2498 } 2499 const TypeFunc* tf = TypeFunc::make(dest_method); 2500 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2501 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2502 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2503 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2504 if (is_reference_type(targ->basic_type())) { 2505 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2506 ciKlass* better_type = nullptr; 2507 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2508 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2509 } 2510 i++; 2511 } 2512 } 2513 } 2514 2515 /** 2516 * Record profiling data from parameter profiling at an invoke with 2517 * the type system so that it can propagate it (speculation) 2518 */ 2519 void GraphKit::record_profiled_parameters_for_speculation() { 2520 if (!UseTypeSpeculation) { 2521 return; 2522 } 2523 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2524 if (_gvn.type(local(i))->isa_oopptr()) { 2525 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2526 ciKlass* better_type = nullptr; 2527 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2528 record_profile_for_speculation(local(i), better_type, ptr_kind); 2529 } 2530 j++; 2531 } 2532 } 2533 } 2534 2535 /** 2536 * Record profiling data from return value profiling at an invoke with 2537 * the type system so that it can propagate it (speculation) 2538 */ 2539 void GraphKit::record_profiled_return_for_speculation() { 2540 if (!UseTypeSpeculation) { 2541 return; 2542 } 2543 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2544 ciKlass* better_type = nullptr; 2545 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2546 // If profiling reports a single type for the return value, 2547 // feed it to the type system so it can propagate it as a 2548 // speculative type 2549 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2550 } 2551 } 2552 2553 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2554 if (Matcher::strict_fp_requires_explicit_rounding) { 2555 // (Note: TypeFunc::make has a cache that makes this fast.) 2556 const TypeFunc* tf = TypeFunc::make(dest_method); 2557 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2558 for (int j = 0; j < nargs; j++) { 2559 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2560 if (targ->basic_type() == T_DOUBLE) { 2561 // If any parameters are doubles, they must be rounded before 2562 // the call, dprecision_rounding does gvn.transform 2563 Node *arg = argument(j); 2564 arg = dprecision_rounding(arg); 2565 set_argument(j, arg); 2566 } 2567 } 2568 } 2569 } 2570 2571 // rounding for strict float precision conformance 2572 Node* GraphKit::precision_rounding(Node* n) { 2573 if (Matcher::strict_fp_requires_explicit_rounding) { 2574 #ifdef IA32 2575 if (UseSSE == 0) { 2576 return _gvn.transform(new RoundFloatNode(nullptr, n)); 2577 } 2578 #else 2579 Unimplemented(); 2580 #endif // IA32 2581 } 2582 return n; 2583 } 2584 2585 // rounding for strict double precision conformance 2586 Node* GraphKit::dprecision_rounding(Node *n) { 2587 if (Matcher::strict_fp_requires_explicit_rounding) { 2588 #ifdef IA32 2589 if (UseSSE < 2) { 2590 return _gvn.transform(new RoundDoubleNode(nullptr, n)); 2591 } 2592 #else 2593 Unimplemented(); 2594 #endif // IA32 2595 } 2596 return n; 2597 } 2598 2599 //============================================================================= 2600 // Generate a fast path/slow path idiom. Graph looks like: 2601 // [foo] indicates that 'foo' is a parameter 2602 // 2603 // [in] null 2604 // \ / 2605 // CmpP 2606 // Bool ne 2607 // If 2608 // / \ 2609 // True False-<2> 2610 // / | 2611 // / cast_not_null 2612 // Load | | ^ 2613 // [fast_test] | | 2614 // gvn to opt_test | | 2615 // / \ | <1> 2616 // True False | 2617 // | \\ | 2618 // [slow_call] \[fast_result] 2619 // Ctl Val \ \ 2620 // | \ \ 2621 // Catch <1> \ \ 2622 // / \ ^ \ \ 2623 // Ex No_Ex | \ \ 2624 // | \ \ | \ <2> \ 2625 // ... \ [slow_res] | | \ [null_result] 2626 // \ \--+--+--- | | 2627 // \ | / \ | / 2628 // --------Region Phi 2629 // 2630 //============================================================================= 2631 // Code is structured as a series of driver functions all called 'do_XXX' that 2632 // call a set of helper functions. Helper functions first, then drivers. 2633 2634 //------------------------------null_check_oop--------------------------------- 2635 // Null check oop. Set null-path control into Region in slot 3. 2636 // Make a cast-not-nullness use the other not-null control. Return cast. 2637 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2638 bool never_see_null, 2639 bool safe_for_replace, 2640 bool speculative) { 2641 // Initial null check taken path 2642 (*null_control) = top(); 2643 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2644 2645 // Generate uncommon_trap: 2646 if (never_see_null && (*null_control) != top()) { 2647 // If we see an unexpected null at a check-cast we record it and force a 2648 // recompile; the offending check-cast will be compiled to handle nulls. 2649 // If we see more than one offending BCI, then all checkcasts in the 2650 // method will be compiled to handle nulls. 2651 PreserveJVMState pjvms(this); 2652 set_control(*null_control); 2653 replace_in_map(value, null()); 2654 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2655 uncommon_trap(reason, 2656 Deoptimization::Action_make_not_entrant); 2657 (*null_control) = top(); // null path is dead 2658 } 2659 if ((*null_control) == top() && safe_for_replace) { 2660 replace_in_map(value, cast); 2661 } 2662 2663 // Cast away null-ness on the result 2664 return cast; 2665 } 2666 2667 //------------------------------opt_iff---------------------------------------- 2668 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2669 // Return slow-path control. 2670 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2671 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2672 2673 // Fast path taken; set region slot 2 2674 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2675 region->init_req(2,fast_taken); // Capture fast-control 2676 2677 // Fast path not-taken, i.e. slow path 2678 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2679 return slow_taken; 2680 } 2681 2682 //-----------------------------make_runtime_call------------------------------- 2683 Node* GraphKit::make_runtime_call(int flags, 2684 const TypeFunc* call_type, address call_addr, 2685 const char* call_name, 2686 const TypePtr* adr_type, 2687 // The following parms are all optional. 2688 // The first null ends the list. 2689 Node* parm0, Node* parm1, 2690 Node* parm2, Node* parm3, 2691 Node* parm4, Node* parm5, 2692 Node* parm6, Node* parm7) { 2693 assert(call_addr != nullptr, "must not call null targets"); 2694 2695 // Slow-path call 2696 bool is_leaf = !(flags & RC_NO_LEAF); 2697 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2698 if (call_name == nullptr) { 2699 assert(!is_leaf, "must supply name for leaf"); 2700 call_name = OptoRuntime::stub_name(call_addr); 2701 } 2702 CallNode* call; 2703 if (!is_leaf) { 2704 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type); 2705 } else if (flags & RC_NO_FP) { 2706 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2707 } else if (flags & RC_VECTOR){ 2708 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte; 2709 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits); 2710 } else { 2711 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2712 } 2713 2714 // The following is similar to set_edges_for_java_call, 2715 // except that the memory effects of the call are restricted to AliasIdxRaw. 2716 2717 // Slow path call has no side-effects, uses few values 2718 bool wide_in = !(flags & RC_NARROW_MEM); 2719 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2720 2721 Node* prev_mem = nullptr; 2722 if (wide_in) { 2723 prev_mem = set_predefined_input_for_runtime_call(call); 2724 } else { 2725 assert(!wide_out, "narrow in => narrow out"); 2726 Node* narrow_mem = memory(adr_type); 2727 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2728 } 2729 2730 // Hook each parm in order. Stop looking at the first null. 2731 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0); 2732 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1); 2733 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2); 2734 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3); 2735 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4); 2736 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5); 2737 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6); 2738 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7); 2739 /* close each nested if ===> */ } } } } } } } } 2740 assert(call->in(call->req()-1) != nullptr, "must initialize all parms"); 2741 2742 if (!is_leaf) { 2743 // Non-leaves can block and take safepoints: 2744 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2745 } 2746 // Non-leaves can throw exceptions: 2747 if (has_io) { 2748 call->set_req(TypeFunc::I_O, i_o()); 2749 } 2750 2751 if (flags & RC_UNCOMMON) { 2752 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2753 // (An "if" probability corresponds roughly to an unconditional count. 2754 // Sort of.) 2755 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2756 } 2757 2758 Node* c = _gvn.transform(call); 2759 assert(c == call, "cannot disappear"); 2760 2761 if (wide_out) { 2762 // Slow path call has full side-effects. 2763 set_predefined_output_for_runtime_call(call); 2764 } else { 2765 // Slow path call has few side-effects, and/or sets few values. 2766 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2767 } 2768 2769 if (has_io) { 2770 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2771 } 2772 return call; 2773 2774 } 2775 2776 // i2b 2777 Node* GraphKit::sign_extend_byte(Node* in) { 2778 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24))); 2779 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24))); 2780 } 2781 2782 // i2s 2783 Node* GraphKit::sign_extend_short(Node* in) { 2784 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16))); 2785 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16))); 2786 } 2787 2788 2789 //------------------------------merge_memory----------------------------------- 2790 // Merge memory from one path into the current memory state. 2791 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2792 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2793 Node* old_slice = mms.force_memory(); 2794 Node* new_slice = mms.memory2(); 2795 if (old_slice != new_slice) { 2796 PhiNode* phi; 2797 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2798 if (mms.is_empty()) { 2799 // clone base memory Phi's inputs for this memory slice 2800 assert(old_slice == mms.base_memory(), "sanity"); 2801 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C)); 2802 _gvn.set_type(phi, Type::MEMORY); 2803 for (uint i = 1; i < phi->req(); i++) { 2804 phi->init_req(i, old_slice->in(i)); 2805 } 2806 } else { 2807 phi = old_slice->as_Phi(); // Phi was generated already 2808 } 2809 } else { 2810 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2811 _gvn.set_type(phi, Type::MEMORY); 2812 } 2813 phi->set_req(new_path, new_slice); 2814 mms.set_memory(phi); 2815 } 2816 } 2817 } 2818 2819 //------------------------------make_slow_call_ex------------------------------ 2820 // Make the exception handler hookups for the slow call 2821 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2822 if (stopped()) return; 2823 2824 // Make a catch node with just two handlers: fall-through and catch-all 2825 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2826 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2827 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2828 _gvn.set_type_bottom(norm); 2829 C->record_for_igvn(norm); 2830 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2831 2832 { PreserveJVMState pjvms(this); 2833 set_control(excp); 2834 set_i_o(i_o); 2835 2836 if (excp != top()) { 2837 if (deoptimize) { 2838 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2839 uncommon_trap(Deoptimization::Reason_unhandled, 2840 Deoptimization::Action_none); 2841 } else { 2842 // Create an exception state also. 2843 // Use an exact type if the caller has a specific exception. 2844 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2845 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2846 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2847 } 2848 } 2849 } 2850 2851 // Get the no-exception control from the CatchNode. 2852 set_control(norm); 2853 } 2854 2855 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) { 2856 Node* cmp = nullptr; 2857 switch(bt) { 2858 case T_INT: cmp = new CmpINode(in1, in2); break; 2859 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2860 default: fatal("unexpected comparison type %s", type2name(bt)); 2861 } 2862 cmp = gvn.transform(cmp); 2863 Node* bol = gvn.transform(new BoolNode(cmp, test)); 2864 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2865 gvn.transform(iff); 2866 if (!bol->is_Con()) gvn.record_for_igvn(iff); 2867 return iff; 2868 } 2869 2870 //-------------------------------gen_subtype_check----------------------------- 2871 // Generate a subtyping check. Takes as input the subtype and supertype. 2872 // Returns 2 values: sets the default control() to the true path and returns 2873 // the false path. Only reads invariant memory; sets no (visible) memory. 2874 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2875 // but that's not exposed to the optimizer. This call also doesn't take in an 2876 // Object; if you wish to check an Object you need to load the Object's class 2877 // prior to coming here. 2878 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn, 2879 ciMethod* method, int bci) { 2880 Compile* C = gvn.C; 2881 if ((*ctrl)->is_top()) { 2882 return C->top(); 2883 } 2884 2885 // Fast check for identical types, perhaps identical constants. 2886 // The types can even be identical non-constants, in cases 2887 // involving Array.newInstance, Object.clone, etc. 2888 if (subklass == superklass) 2889 return C->top(); // false path is dead; no test needed. 2890 2891 if (gvn.type(superklass)->singleton()) { 2892 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2893 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr(); 2894 2895 // In the common case of an exact superklass, try to fold up the 2896 // test before generating code. You may ask, why not just generate 2897 // the code and then let it fold up? The answer is that the generated 2898 // code will necessarily include null checks, which do not always 2899 // completely fold away. If they are also needless, then they turn 2900 // into a performance loss. Example: 2901 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2902 // Here, the type of 'fa' is often exact, so the store check 2903 // of fa[1]=x will fold up, without testing the nullness of x. 2904 // 2905 // At macro expansion, we would have already folded the SubTypeCheckNode 2906 // being expanded here because we always perform the static sub type 2907 // check in SubTypeCheckNode::sub() regardless of whether 2908 // StressReflectiveCode is set or not. We can therefore skip this 2909 // static check when StressReflectiveCode is on. 2910 switch (C->static_subtype_check(superk, subk)) { 2911 case Compile::SSC_always_false: 2912 { 2913 Node* always_fail = *ctrl; 2914 *ctrl = gvn.C->top(); 2915 return always_fail; 2916 } 2917 case Compile::SSC_always_true: 2918 return C->top(); 2919 case Compile::SSC_easy_test: 2920 { 2921 // Just do a direct pointer compare and be done. 2922 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2923 *ctrl = gvn.transform(new IfTrueNode(iff)); 2924 return gvn.transform(new IfFalseNode(iff)); 2925 } 2926 case Compile::SSC_full_test: 2927 break; 2928 default: 2929 ShouldNotReachHere(); 2930 } 2931 } 2932 2933 // %%% Possible further optimization: Even if the superklass is not exact, 2934 // if the subklass is the unique subtype of the superklass, the check 2935 // will always succeed. We could leave a dependency behind to ensure this. 2936 2937 // First load the super-klass's check-offset 2938 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2939 Node* m = C->immutable_memory(); 2940 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2941 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2942 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int(); 2943 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con; 2944 bool might_be_cache = (chk_off_con == cacheoff_con); 2945 2946 // Load from the sub-klass's super-class display list, or a 1-word cache of 2947 // the secondary superclass list, or a failing value with a sentinel offset 2948 // if the super-klass is an interface or exceptionally deep in the Java 2949 // hierarchy and we have to scan the secondary superclass list the hard way. 2950 // Worst-case type is a little odd: null is allowed as a result (usually 2951 // klass loads can never produce a null). 2952 Node *chk_off_X = chk_off; 2953 #ifdef _LP64 2954 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X)); 2955 #endif 2956 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X)); 2957 // For some types like interfaces the following loadKlass is from a 1-word 2958 // cache which is mutable so can't use immutable memory. Other 2959 // types load from the super-class display table which is immutable. 2960 Node *kmem = C->immutable_memory(); 2961 // secondary_super_cache is not immutable but can be treated as such because: 2962 // - no ideal node writes to it in a way that could cause an 2963 // incorrect/missed optimization of the following Load. 2964 // - it's a cache so, worse case, not reading the latest value 2965 // wouldn't cause incorrect execution 2966 if (might_be_cache && mem != nullptr) { 2967 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem; 2968 } 2969 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL)); 2970 2971 // Compile speed common case: ARE a subtype and we canNOT fail 2972 if (superklass == nkls) { 2973 return C->top(); // false path is dead; no test needed. 2974 } 2975 2976 // Gather the various success & failures here 2977 RegionNode* r_not_subtype = new RegionNode(3); 2978 gvn.record_for_igvn(r_not_subtype); 2979 RegionNode* r_ok_subtype = new RegionNode(4); 2980 gvn.record_for_igvn(r_ok_subtype); 2981 2982 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the 2983 // SubTypeCheck node 2984 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) { 2985 ciCallProfile profile = method->call_profile_at_bci(bci); 2986 float total_prob = 0; 2987 for (int i = 0; profile.has_receiver(i); ++i) { 2988 float prob = profile.receiver_prob(i); 2989 total_prob += prob; 2990 } 2991 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) { 2992 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2993 for (int i = 0; profile.has_receiver(i); ++i) { 2994 ciKlass* klass = profile.receiver(i); 2995 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass); 2996 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t); 2997 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) { 2998 continue; 2999 } 3000 float prob = profile.receiver_prob(i); 3001 ConNode* klass_node = gvn.makecon(klass_t); 3002 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS); 3003 Node* iftrue = gvn.transform(new IfTrueNode(iff)); 3004 3005 if (result == Compile::SSC_always_true) { 3006 r_ok_subtype->add_req(iftrue); 3007 } else { 3008 assert(result == Compile::SSC_always_false, ""); 3009 r_not_subtype->add_req(iftrue); 3010 } 3011 *ctrl = gvn.transform(new IfFalseNode(iff)); 3012 } 3013 } 3014 } 3015 3016 // See if we get an immediate positive hit. Happens roughly 83% of the 3017 // time. Test to see if the value loaded just previously from the subklass 3018 // is exactly the superklass. 3019 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 3020 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1)); 3021 *ctrl = gvn.transform(new IfFalseNode(iff1)); 3022 3023 // Compile speed common case: Check for being deterministic right now. If 3024 // chk_off is a constant and not equal to cacheoff then we are NOT a 3025 // subklass. In this case we need exactly the 1 test above and we can 3026 // return those results immediately. 3027 if (!might_be_cache) { 3028 Node* not_subtype_ctrl = *ctrl; 3029 *ctrl = iftrue1; // We need exactly the 1 test above 3030 PhaseIterGVN* igvn = gvn.is_IterGVN(); 3031 if (igvn != nullptr) { 3032 igvn->remove_globally_dead_node(r_ok_subtype); 3033 igvn->remove_globally_dead_node(r_not_subtype); 3034 } 3035 return not_subtype_ctrl; 3036 } 3037 3038 r_ok_subtype->init_req(1, iftrue1); 3039 3040 // Check for immediate negative hit. Happens roughly 11% of the time (which 3041 // is roughly 63% of the remaining cases). Test to see if the loaded 3042 // check-offset points into the subklass display list or the 1-element 3043 // cache. If it points to the display (and NOT the cache) and the display 3044 // missed then it's not a subtype. 3045 Node *cacheoff = gvn.intcon(cacheoff_con); 3046 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 3047 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2))); 3048 *ctrl = gvn.transform(new IfFalseNode(iff2)); 3049 3050 // Check for self. Very rare to get here, but it is taken 1/3 the time. 3051 // No performance impact (too rare) but allows sharing of secondary arrays 3052 // which has some footprint reduction. 3053 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 3054 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3))); 3055 *ctrl = gvn.transform(new IfFalseNode(iff3)); 3056 3057 // -- Roads not taken here: -- 3058 // We could also have chosen to perform the self-check at the beginning 3059 // of this code sequence, as the assembler does. This would not pay off 3060 // the same way, since the optimizer, unlike the assembler, can perform 3061 // static type analysis to fold away many successful self-checks. 3062 // Non-foldable self checks work better here in second position, because 3063 // the initial primary superclass check subsumes a self-check for most 3064 // types. An exception would be a secondary type like array-of-interface, 3065 // which does not appear in its own primary supertype display. 3066 // Finally, we could have chosen to move the self-check into the 3067 // PartialSubtypeCheckNode, and from there out-of-line in a platform 3068 // dependent manner. But it is worthwhile to have the check here, 3069 // where it can be perhaps be optimized. The cost in code space is 3070 // small (register compare, branch). 3071 3072 // Now do a linear scan of the secondary super-klass array. Again, no real 3073 // performance impact (too rare) but it's gotta be done. 3074 // Since the code is rarely used, there is no penalty for moving it 3075 // out of line, and it can only improve I-cache density. 3076 // The decision to inline or out-of-line this final check is platform 3077 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 3078 Node* psc = gvn.transform( 3079 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 3080 3081 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 3082 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4))); 3083 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4))); 3084 3085 // Return false path; set default control to true path. 3086 *ctrl = gvn.transform(r_ok_subtype); 3087 return gvn.transform(r_not_subtype); 3088 } 3089 3090 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) { 3091 const Type* sub_t = _gvn.type(obj_or_subklass); 3092 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) { 3093 sub_t = TypeKlassPtr::make(sub_t->inline_klass()); 3094 obj_or_subklass = makecon(sub_t); 3095 } 3096 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over 3097 if (expand_subtype_check) { 3098 MergeMemNode* mem = merged_memory(); 3099 Node* ctrl = control(); 3100 Node* subklass = obj_or_subklass; 3101 if (!sub_t->isa_klassptr()) { 3102 subklass = load_object_klass(obj_or_subklass); 3103 } 3104 3105 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci()); 3106 set_control(ctrl); 3107 return n; 3108 } 3109 3110 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci())); 3111 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq)); 3112 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3113 set_control(_gvn.transform(new IfTrueNode(iff))); 3114 return _gvn.transform(new IfFalseNode(iff)); 3115 } 3116 3117 // Profile-driven exact type check: 3118 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 3119 float prob, Node* *casted_receiver) { 3120 assert(!klass->is_interface(), "no exact type check on interfaces"); 3121 Node* fail = top(); 3122 const Type* rec_t = _gvn.type(receiver); 3123 if (rec_t->is_inlinetypeptr()) { 3124 if (klass->equals(rec_t->inline_klass())) { 3125 (*casted_receiver) = receiver; // Always passes 3126 } else { 3127 (*casted_receiver) = top(); // Always fails 3128 fail = control(); 3129 set_control(top()); 3130 } 3131 return fail; 3132 } 3133 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces); 3134 Node* recv_klass = load_object_klass(receiver); 3135 fail = type_check(recv_klass, tklass, prob); 3136 3137 if (!stopped()) { 3138 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3139 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 3140 assert(recv_xtype->klass_is_exact(), ""); 3141 3142 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts 3143 // Subsume downstream occurrences of receiver with a cast to 3144 // recv_xtype, since now we know what the type will be. 3145 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 3146 Node* res = _gvn.transform(cast); 3147 if (recv_xtype->is_inlinetypeptr()) { 3148 assert(!gvn().type(res)->maybe_null(), "receiver should never be null"); 3149 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass()); 3150 } 3151 (*casted_receiver) = res; 3152 assert(!(*casted_receiver)->is_top(), "that path should be unreachable"); 3153 // (User must make the replace_in_map call.) 3154 } 3155 } 3156 3157 return fail; 3158 } 3159 3160 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 3161 float prob) { 3162 Node* want_klass = makecon(tklass); 3163 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass)); 3164 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3165 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 3166 set_control(_gvn.transform(new IfTrueNode (iff))); 3167 Node* fail = _gvn.transform(new IfFalseNode(iff)); 3168 return fail; 3169 } 3170 3171 //------------------------------subtype_check_receiver------------------------- 3172 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 3173 Node** casted_receiver) { 3174 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve(); 3175 Node* want_klass = makecon(tklass); 3176 3177 Node* slow_ctl = gen_subtype_check(receiver, want_klass); 3178 3179 // Ignore interface type information until interface types are properly tracked. 3180 if (!stopped() && !klass->is_interface()) { 3181 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3182 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 3183 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts 3184 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type)); 3185 if (recv_type->is_inlinetypeptr()) { 3186 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass()); 3187 } 3188 (*casted_receiver) = cast; 3189 } 3190 } 3191 3192 return slow_ctl; 3193 } 3194 3195 //------------------------------seems_never_null------------------------------- 3196 // Use null_seen information if it is available from the profile. 3197 // If we see an unexpected null at a type check we record it and force a 3198 // recompile; the offending check will be recompiled to handle nulls. 3199 // If we see several offending BCIs, then all checks in the 3200 // method will be recompiled. 3201 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 3202 speculating = !_gvn.type(obj)->speculative_maybe_null(); 3203 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 3204 if (UncommonNullCast // Cutout for this technique 3205 && obj != null() // And not the -Xcomp stupid case? 3206 && !too_many_traps(reason) 3207 ) { 3208 if (speculating) { 3209 return true; 3210 } 3211 if (data == nullptr) 3212 // Edge case: no mature data. Be optimistic here. 3213 return true; 3214 // If the profile has not seen a null, assume it won't happen. 3215 assert(java_bc() == Bytecodes::_checkcast || 3216 java_bc() == Bytecodes::_instanceof || 3217 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 3218 return !data->as_BitData()->null_seen(); 3219 } 3220 speculating = false; 3221 return false; 3222 } 3223 3224 void GraphKit::guard_klass_being_initialized(Node* klass) { 3225 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 3226 Node* adr = basic_plus_adr(top(), klass, init_state_off); 3227 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3228 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 3229 T_BYTE, MemNode::acquire); 3230 init_state = _gvn.transform(init_state); 3231 3232 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 3233 3234 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 3235 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3236 3237 { BuildCutout unless(this, tst, PROB_MAX); 3238 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 3239 } 3240 } 3241 3242 void GraphKit::guard_init_thread(Node* klass) { 3243 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 3244 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 3245 3246 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3247 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 3248 T_ADDRESS, MemNode::unordered); 3249 init_thread = _gvn.transform(init_thread); 3250 3251 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 3252 3253 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 3254 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3255 3256 { BuildCutout unless(this, tst, PROB_MAX); 3257 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 3258 } 3259 } 3260 3261 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3262 if (ik->is_being_initialized()) { 3263 if (C->needs_clinit_barrier(ik, context)) { 3264 Node* klass = makecon(TypeKlassPtr::make(ik)); 3265 guard_klass_being_initialized(klass); 3266 guard_init_thread(klass); 3267 insert_mem_bar(Op_MemBarCPUOrder); 3268 } 3269 } else if (ik->is_initialized()) { 3270 return; // no barrier needed 3271 } else { 3272 uncommon_trap(Deoptimization::Reason_uninitialized, 3273 Deoptimization::Action_reinterpret, 3274 nullptr); 3275 } 3276 } 3277 3278 //------------------------maybe_cast_profiled_receiver------------------------- 3279 // If the profile has seen exactly one type, narrow to exactly that type. 3280 // Subsequent type checks will always fold up. 3281 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3282 const TypeKlassPtr* require_klass, 3283 ciKlass* spec_klass, 3284 bool safe_for_replace) { 3285 if (!UseTypeProfile || !TypeProfileCasts) return nullptr; 3286 3287 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr); 3288 3289 // Make sure we haven't already deoptimized from this tactic. 3290 if (too_many_traps_or_recompiles(reason)) 3291 return nullptr; 3292 3293 // (No, this isn't a call, but it's enough like a virtual call 3294 // to use the same ciMethod accessor to get the profile info...) 3295 // If we have a speculative type use it instead of profiling (which 3296 // may not help us) 3297 ciKlass* exact_kls = spec_klass; 3298 if (exact_kls == nullptr) { 3299 if (java_bc() == Bytecodes::_aastore) { 3300 ciKlass* array_type = nullptr; 3301 ciKlass* element_type = nullptr; 3302 ProfilePtrKind element_ptr = ProfileMaybeNull; 3303 bool flat_array = true; 3304 bool null_free_array = true; 3305 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 3306 exact_kls = element_type; 3307 } else { 3308 exact_kls = profile_has_unique_klass(); 3309 } 3310 } 3311 if (exact_kls != nullptr) {// no cast failures here 3312 if (require_klass == nullptr || 3313 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) { 3314 // If we narrow the type to match what the type profile sees or 3315 // the speculative type, we can then remove the rest of the 3316 // cast. 3317 // This is a win, even if the exact_kls is very specific, 3318 // because downstream operations, such as method calls, 3319 // will often benefit from the sharper type. 3320 Node* exact_obj = not_null_obj; // will get updated in place... 3321 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3322 &exact_obj); 3323 { PreserveJVMState pjvms(this); 3324 set_control(slow_ctl); 3325 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3326 } 3327 if (safe_for_replace) { 3328 replace_in_map(not_null_obj, exact_obj); 3329 } 3330 return exact_obj; 3331 } 3332 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3333 } 3334 3335 return nullptr; 3336 } 3337 3338 /** 3339 * Cast obj to type and emit guard unless we had too many traps here 3340 * already 3341 * 3342 * @param obj node being casted 3343 * @param type type to cast the node to 3344 * @param not_null true if we know node cannot be null 3345 */ 3346 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3347 ciKlass* type, 3348 bool not_null) { 3349 if (stopped()) { 3350 return obj; 3351 } 3352 3353 // type is null if profiling tells us this object is always null 3354 if (type != nullptr) { 3355 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3356 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3357 3358 if (!too_many_traps_or_recompiles(null_reason) && 3359 !too_many_traps_or_recompiles(class_reason)) { 3360 Node* not_null_obj = nullptr; 3361 // not_null is true if we know the object is not null and 3362 // there's no need for a null check 3363 if (!not_null) { 3364 Node* null_ctl = top(); 3365 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3366 assert(null_ctl->is_top(), "no null control here"); 3367 } else { 3368 not_null_obj = obj; 3369 } 3370 3371 Node* exact_obj = not_null_obj; 3372 ciKlass* exact_kls = type; 3373 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3374 &exact_obj); 3375 { 3376 PreserveJVMState pjvms(this); 3377 set_control(slow_ctl); 3378 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3379 } 3380 replace_in_map(not_null_obj, exact_obj); 3381 obj = exact_obj; 3382 } 3383 } else { 3384 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3385 Node* exact_obj = null_assert(obj); 3386 replace_in_map(obj, exact_obj); 3387 obj = exact_obj; 3388 } 3389 } 3390 return obj; 3391 } 3392 3393 //-------------------------------gen_instanceof-------------------------------- 3394 // Generate an instance-of idiom. Used by both the instance-of bytecode 3395 // and the reflective instance-of call. 3396 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3397 kill_dead_locals(); // Benefit all the uncommon traps 3398 assert( !stopped(), "dead parse path should be checked in callers" ); 3399 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3400 "must check for not-null not-dead klass in callers"); 3401 3402 // Make the merge point 3403 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3404 RegionNode* region = new RegionNode(PATH_LIMIT); 3405 Node* phi = new PhiNode(region, TypeInt::BOOL); 3406 C->set_has_split_ifs(true); // Has chance for split-if optimization 3407 3408 ciProfileData* data = nullptr; 3409 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3410 data = method()->method_data()->bci_to_data(bci()); 3411 } 3412 bool speculative_not_null = false; 3413 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3414 && seems_never_null(obj, data, speculative_not_null)); 3415 3416 // Null check; get casted pointer; set region slot 3 3417 Node* null_ctl = top(); 3418 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3419 3420 // If not_null_obj is dead, only null-path is taken 3421 if (stopped()) { // Doing instance-of on a null? 3422 set_control(null_ctl); 3423 return intcon(0); 3424 } 3425 region->init_req(_null_path, null_ctl); 3426 phi ->init_req(_null_path, intcon(0)); // Set null path value 3427 if (null_ctl == top()) { 3428 // Do this eagerly, so that pattern matches like is_diamond_phi 3429 // will work even during parsing. 3430 assert(_null_path == PATH_LIMIT-1, "delete last"); 3431 region->del_req(_null_path); 3432 phi ->del_req(_null_path); 3433 } 3434 3435 // Do we know the type check always succeed? 3436 bool known_statically = false; 3437 if (_gvn.type(superklass)->singleton()) { 3438 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr(); 3439 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type(); 3440 if (subk != nullptr && subk->is_loaded()) { 3441 int static_res = C->static_subtype_check(superk, subk); 3442 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3443 } 3444 } 3445 3446 if (!known_statically) { 3447 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3448 // We may not have profiling here or it may not help us. If we 3449 // have a speculative type use it to perform an exact cast. 3450 ciKlass* spec_obj_type = obj_type->speculative_type(); 3451 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) { 3452 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace); 3453 if (stopped()) { // Profile disagrees with this path. 3454 set_control(null_ctl); // Null is the only remaining possibility. 3455 return intcon(0); 3456 } 3457 if (cast_obj != nullptr) { 3458 not_null_obj = cast_obj; 3459 } 3460 } 3461 } 3462 3463 // Generate the subtype check 3464 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3465 3466 // Plug in the success path to the general merge in slot 1. 3467 region->init_req(_obj_path, control()); 3468 phi ->init_req(_obj_path, intcon(1)); 3469 3470 // Plug in the failing path to the general merge in slot 2. 3471 region->init_req(_fail_path, not_subtype_ctrl); 3472 phi ->init_req(_fail_path, intcon(0)); 3473 3474 // Return final merged results 3475 set_control( _gvn.transform(region) ); 3476 record_for_igvn(region); 3477 3478 // If we know the type check always succeeds then we don't use the 3479 // profiling data at this bytecode. Don't lose it, feed it to the 3480 // type system as a speculative type. 3481 if (safe_for_replace) { 3482 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3483 replace_in_map(obj, casted_obj); 3484 } 3485 3486 return _gvn.transform(phi); 3487 } 3488 3489 //-------------------------------gen_checkcast--------------------------------- 3490 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3491 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3492 // uncommon-trap paths work. Adjust stack after this call. 3493 // If failure_control is supplied and not null, it is filled in with 3494 // the control edge for the cast failure. Otherwise, an appropriate 3495 // uncommon trap or exception is thrown. 3496 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) { 3497 kill_dead_locals(); // Benefit all the uncommon traps 3498 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr(); 3499 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve(); 3500 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type(); 3501 bool safe_for_replace = (failure_control == nullptr); 3502 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer"); 3503 3504 // Fast cutout: Check the case that the cast is vacuously true. 3505 // This detects the common cases where the test will short-circuit 3506 // away completely. We do this before we perform the null check, 3507 // because if the test is going to turn into zero code, we don't 3508 // want a residual null check left around. (Causes a slowdown, 3509 // for example, in some objArray manipulations, such as a[i]=a[j].) 3510 if (improved_klass_ptr_type->singleton()) { 3511 const TypeKlassPtr* kptr = nullptr; 3512 const Type* t = _gvn.type(obj); 3513 if (t->isa_oop_ptr()) { 3514 kptr = t->is_oopptr()->as_klass_type(); 3515 } else if (obj->is_InlineType()) { 3516 ciInlineKlass* vk = t->inline_klass(); 3517 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0)); 3518 } 3519 if (kptr != nullptr) { 3520 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) { 3521 case Compile::SSC_always_true: 3522 // If we know the type check always succeed then we don't use 3523 // the profiling data at this bytecode. Don't lose it, feed it 3524 // to the type system as a speculative type. 3525 obj = record_profiled_receiver_for_speculation(obj); 3526 if (null_free) { 3527 assert(safe_for_replace, "must be"); 3528 obj = null_check(obj); 3529 } 3530 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized"); 3531 return obj; 3532 case Compile::SSC_always_false: 3533 if (null_free) { 3534 assert(safe_for_replace, "must be"); 3535 obj = null_check(obj); 3536 } 3537 // It needs a null check because a null will *pass* the cast check. 3538 if (t->isa_oopptr() != nullptr && !t->is_oopptr()->maybe_null()) { 3539 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3540 Deoptimization::DeoptReason reason = is_aastore ? 3541 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3542 builtin_throw(reason); 3543 return top(); 3544 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3545 return null_assert(obj); 3546 } 3547 break; // Fall through to full check 3548 default: 3549 break; 3550 } 3551 } 3552 } 3553 3554 ciProfileData* data = nullptr; 3555 if (failure_control == nullptr) { // use MDO in regular case only 3556 assert(java_bc() == Bytecodes::_aastore || 3557 java_bc() == Bytecodes::_checkcast, 3558 "interpreter profiles type checks only for these BCs"); 3559 if (method()->method_data()->is_mature()) { 3560 data = method()->method_data()->bci_to_data(bci()); 3561 } 3562 } 3563 3564 // Make the merge point 3565 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3566 RegionNode* region = new RegionNode(PATH_LIMIT); 3567 Node* phi = new PhiNode(region, toop); 3568 _gvn.set_type(region, Type::CONTROL); 3569 _gvn.set_type(phi, toop); 3570 3571 C->set_has_split_ifs(true); // Has chance for split-if optimization 3572 3573 // Use null-cast information if it is available 3574 bool speculative_not_null = false; 3575 bool never_see_null = ((failure_control == nullptr) // regular case only 3576 && seems_never_null(obj, data, speculative_not_null)); 3577 3578 if (obj->is_InlineType()) { 3579 // Re-execute if buffering during triggers deoptimization 3580 PreserveReexecuteState preexecs(this); 3581 jvms()->set_should_reexecute(true); 3582 obj = obj->as_InlineType()->buffer(this, safe_for_replace); 3583 } 3584 3585 // Null check; get casted pointer; set region slot 3 3586 Node* null_ctl = top(); 3587 Node* not_null_obj = nullptr; 3588 if (null_free) { 3589 assert(safe_for_replace, "must be"); 3590 not_null_obj = null_check(obj); 3591 } else { 3592 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3593 } 3594 3595 // If not_null_obj is dead, only null-path is taken 3596 if (stopped()) { // Doing instance-of on a null? 3597 set_control(null_ctl); 3598 if (toop->is_inlinetypeptr()) { 3599 return InlineTypeNode::make_null(_gvn, toop->inline_klass()); 3600 } 3601 return null(); 3602 } 3603 region->init_req(_null_path, null_ctl); 3604 phi ->init_req(_null_path, null()); // Set null path value 3605 if (null_ctl == top()) { 3606 // Do this eagerly, so that pattern matches like is_diamond_phi 3607 // will work even during parsing. 3608 assert(_null_path == PATH_LIMIT-1, "delete last"); 3609 region->del_req(_null_path); 3610 phi ->del_req(_null_path); 3611 } 3612 3613 Node* cast_obj = nullptr; 3614 if (improved_klass_ptr_type->klass_is_exact()) { 3615 // The following optimization tries to statically cast the speculative type of the object 3616 // (for example obtained during profiling) to the type of the superklass and then do a 3617 // dynamic check that the type of the object is what we expect. To work correctly 3618 // for checkcast and aastore the type of superklass should be exact. 3619 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3620 // We may not have profiling here or it may not help us. If we have 3621 // a speculative type use it to perform an exact cast. 3622 ciKlass* spec_obj_type = obj_type->speculative_type(); 3623 if (spec_obj_type != nullptr || data != nullptr) { 3624 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace); 3625 if (cast_obj != nullptr) { 3626 if (failure_control != nullptr) // failure is now impossible 3627 (*failure_control) = top(); 3628 // adjust the type of the phi to the exact klass: 3629 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3630 } 3631 } 3632 } 3633 3634 if (cast_obj == nullptr) { 3635 // Generate the subtype check 3636 Node* improved_superklass = superklass; 3637 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) { 3638 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up. 3639 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top. 3640 // Additionally, the benefit would only be minor in non-constant cases. 3641 improved_superklass = makecon(improved_klass_ptr_type); 3642 } 3643 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass); 3644 // Plug in success path into the merge 3645 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3646 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3647 if (failure_control == nullptr) { 3648 if (not_subtype_ctrl != top()) { // If failure is possible 3649 PreserveJVMState pjvms(this); 3650 set_control(not_subtype_ctrl); 3651 Node* obj_klass = nullptr; 3652 if (not_null_obj->is_InlineType()) { 3653 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass())); 3654 } else { 3655 obj_klass = load_object_klass(not_null_obj); 3656 } 3657 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3658 Deoptimization::DeoptReason reason = is_aastore ? 3659 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3660 builtin_throw(reason); 3661 } 3662 } else { 3663 (*failure_control) = not_subtype_ctrl; 3664 } 3665 } 3666 3667 region->init_req(_obj_path, control()); 3668 phi ->init_req(_obj_path, cast_obj); 3669 3670 // A merge of null or Casted-NotNull obj 3671 Node* res = _gvn.transform(phi); 3672 3673 // Note I do NOT always 'replace_in_map(obj,result)' here. 3674 // if( tk->klass()->can_be_primary_super() ) 3675 // This means that if I successfully store an Object into an array-of-String 3676 // I 'forget' that the Object is really now known to be a String. I have to 3677 // do this because we don't have true union types for interfaces - if I store 3678 // a Baz into an array-of-Interface and then tell the optimizer it's an 3679 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3680 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3681 // replace_in_map( obj, res ); 3682 3683 // Return final merged results 3684 set_control( _gvn.transform(region) ); 3685 record_for_igvn(region); 3686 3687 bool not_inline = !toop->can_be_inline_type(); 3688 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flat_in_array()); 3689 if (EnableValhalla && not_flat_in_array) { 3690 // Check if obj has been loaded from an array 3691 obj = obj->isa_DecodeN() ? obj->in(1) : obj; 3692 Node* array = nullptr; 3693 if (obj->isa_Load()) { 3694 Node* address = obj->in(MemNode::Address); 3695 if (address->isa_AddP()) { 3696 array = address->as_AddP()->in(AddPNode::Base); 3697 } 3698 } else if (obj->is_Phi()) { 3699 Node* region = obj->in(0); 3700 // TODO make this more robust (see JDK-8231346) 3701 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) { 3702 IfNode* iff = region->in(2)->in(0)->isa_If(); 3703 if (iff != nullptr) { 3704 iff->is_flat_array_check(&_gvn, &array); 3705 } 3706 } 3707 } 3708 if (array != nullptr) { 3709 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr(); 3710 if (ary_t != nullptr && !ary_t->is_flat()) { 3711 if (!ary_t->is_not_null_free() && not_inline) { 3712 // Casting array element to a non-inline-type, mark array as not null-free. 3713 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free())); 3714 replace_in_map(array, cast); 3715 } else if (!ary_t->is_not_flat()) { 3716 // Casting array element to a non-flat type, mark array as not flat. 3717 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat())); 3718 replace_in_map(array, cast); 3719 } 3720 } 3721 } 3722 } 3723 3724 if (!stopped() && !res->is_InlineType()) { 3725 res = record_profiled_receiver_for_speculation(res); 3726 if (toop->is_inlinetypeptr()) { 3727 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null()); 3728 res = vt; 3729 if (safe_for_replace) { 3730 replace_in_map(obj, vt); 3731 replace_in_map(not_null_obj, vt); 3732 replace_in_map(res, vt); 3733 } 3734 } 3735 } 3736 return res; 3737 } 3738 3739 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) { 3740 // Load markword 3741 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3742 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3743 if (check_lock) { 3744 // Check if obj is locked 3745 Node* locked_bit = MakeConX(markWord::unlocked_value); 3746 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark)); 3747 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0))); 3748 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3749 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN); 3750 _gvn.transform(iff); 3751 Node* locked_region = new RegionNode(3); 3752 Node* mark_phi = new PhiNode(locked_region, TypeX_X); 3753 3754 // Unlocked: Use bits from mark word 3755 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff))); 3756 mark_phi->init_req(1, mark); 3757 3758 // Locked: Load prototype header from klass 3759 set_control(_gvn.transform(new IfFalseNode(iff))); 3760 // Make loads control dependent to make sure they are only executed if array is locked 3761 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 3762 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, control(), C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 3763 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset())); 3764 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered)); 3765 3766 locked_region->init_req(2, control()); 3767 mark_phi->init_req(2, proto); 3768 set_control(_gvn.transform(locked_region)); 3769 record_for_igvn(locked_region); 3770 3771 mark = mark_phi; 3772 } 3773 3774 // Now check if mark word bits are set 3775 Node* mask = MakeConX(mask_val); 3776 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask)); 3777 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN 3778 Node* cmp = _gvn.transform(new CmpXNode(masked, mask)); 3779 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne)); 3780 } 3781 3782 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) { 3783 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false); 3784 } 3785 3786 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) { 3787 // We can't use immutable memory here because the mark word is mutable. 3788 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the 3789 // check is moved out of loops (mainly to enable loop unswitching). 3790 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass)); 3791 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN 3792 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne)); 3793 } 3794 3795 Node* GraphKit::null_free_array_test(Node* array, bool null_free) { 3796 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free); 3797 } 3798 3799 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null 3800 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) { 3801 RegionNode* region = new RegionNode(3); 3802 Node* null_ctl = top(); 3803 null_check_oop(val, &null_ctl); 3804 if (null_ctl != top()) { 3805 PreserveJVMState pjvms(this); 3806 set_control(null_ctl); 3807 { 3808 // Deoptimize if null-free array 3809 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX); 3810 inc_sp(nargs); 3811 uncommon_trap(Deoptimization::Reason_null_check, 3812 Deoptimization::Action_none); 3813 } 3814 region->init_req(1, control()); 3815 } 3816 region->init_req(2, control()); 3817 set_control(_gvn.transform(region)); 3818 record_for_igvn(region); 3819 if (_gvn.type(val) == TypePtr::NULL_PTR) { 3820 // Since we were just successfully storing null, the array can't be null free. 3821 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 3822 ary_t = ary_t->cast_to_not_null_free(); 3823 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t)); 3824 if (safe_for_replace) { 3825 replace_in_map(ary, cast); 3826 } 3827 ary = cast; 3828 } 3829 return ary; 3830 } 3831 3832 //------------------------------next_monitor----------------------------------- 3833 // What number should be given to the next monitor? 3834 int GraphKit::next_monitor() { 3835 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3836 int next = current + C->sync_stack_slots(); 3837 // Keep the toplevel high water mark current: 3838 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3839 return current; 3840 } 3841 3842 //------------------------------insert_mem_bar--------------------------------- 3843 // Memory barrier to avoid floating things around 3844 // The membar serves as a pinch point between both control and all memory slices. 3845 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3846 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3847 mb->init_req(TypeFunc::Control, control()); 3848 mb->init_req(TypeFunc::Memory, reset_memory()); 3849 Node* membar = _gvn.transform(mb); 3850 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3851 set_all_memory_call(membar); 3852 return membar; 3853 } 3854 3855 //-------------------------insert_mem_bar_volatile---------------------------- 3856 // Memory barrier to avoid floating things around 3857 // The membar serves as a pinch point between both control and memory(alias_idx). 3858 // If you want to make a pinch point on all memory slices, do not use this 3859 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3860 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3861 // When Parse::do_put_xxx updates a volatile field, it appends a series 3862 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3863 // The first membar is on the same memory slice as the field store opcode. 3864 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3865 // All the other membars (for other volatile slices, including AliasIdxBot, 3866 // which stands for all unknown volatile slices) are control-dependent 3867 // on the first membar. This prevents later volatile loads or stores 3868 // from sliding up past the just-emitted store. 3869 3870 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3871 mb->set_req(TypeFunc::Control,control()); 3872 if (alias_idx == Compile::AliasIdxBot) { 3873 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3874 } else { 3875 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3876 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3877 } 3878 Node* membar = _gvn.transform(mb); 3879 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3880 if (alias_idx == Compile::AliasIdxBot) { 3881 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3882 } else { 3883 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3884 } 3885 return membar; 3886 } 3887 3888 //------------------------------shared_lock------------------------------------ 3889 // Emit locking code. 3890 FastLockNode* GraphKit::shared_lock(Node* obj) { 3891 // bci is either a monitorenter bc or InvocationEntryBci 3892 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3893 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3894 3895 if( !GenerateSynchronizationCode ) 3896 return nullptr; // Not locking things? 3897 3898 if (stopped()) // Dead monitor? 3899 return nullptr; 3900 3901 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3902 3903 // Box the stack location 3904 Node* box = new BoxLockNode(next_monitor()); 3905 // Check for bailout after new BoxLockNode 3906 if (failing()) { return nullptr; } 3907 box = _gvn.transform(box); 3908 Node* mem = reset_memory(); 3909 3910 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock(); 3911 3912 // Add monitor to debug info for the slow path. If we block inside the 3913 // slow path and de-opt, we need the monitor hanging around 3914 map()->push_monitor( flock ); 3915 3916 const TypeFunc *tf = LockNode::lock_type(); 3917 LockNode *lock = new LockNode(C, tf); 3918 3919 lock->init_req( TypeFunc::Control, control() ); 3920 lock->init_req( TypeFunc::Memory , mem ); 3921 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3922 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3923 lock->init_req( TypeFunc::ReturnAdr, top() ); 3924 3925 lock->init_req(TypeFunc::Parms + 0, obj); 3926 lock->init_req(TypeFunc::Parms + 1, box); 3927 lock->init_req(TypeFunc::Parms + 2, flock); 3928 add_safepoint_edges(lock); 3929 3930 lock = _gvn.transform( lock )->as_Lock(); 3931 3932 // lock has no side-effects, sets few values 3933 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3934 3935 insert_mem_bar(Op_MemBarAcquireLock); 3936 3937 // Add this to the worklist so that the lock can be eliminated 3938 record_for_igvn(lock); 3939 3940 #ifndef PRODUCT 3941 if (PrintLockStatistics) { 3942 // Update the counter for this lock. Don't bother using an atomic 3943 // operation since we don't require absolute accuracy. 3944 lock->create_lock_counter(map()->jvms()); 3945 increment_counter(lock->counter()->addr()); 3946 } 3947 #endif 3948 3949 return flock; 3950 } 3951 3952 3953 //------------------------------shared_unlock---------------------------------- 3954 // Emit unlocking code. 3955 void GraphKit::shared_unlock(Node* box, Node* obj) { 3956 // bci is either a monitorenter bc or InvocationEntryBci 3957 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3958 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3959 3960 if( !GenerateSynchronizationCode ) 3961 return; 3962 if (stopped()) { // Dead monitor? 3963 map()->pop_monitor(); // Kill monitor from debug info 3964 return; 3965 } 3966 assert(!obj->is_InlineType(), "should not unlock on inline type"); 3967 3968 // Memory barrier to avoid floating things down past the locked region 3969 insert_mem_bar(Op_MemBarReleaseLock); 3970 3971 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3972 UnlockNode *unlock = new UnlockNode(C, tf); 3973 #ifdef ASSERT 3974 unlock->set_dbg_jvms(sync_jvms()); 3975 #endif 3976 uint raw_idx = Compile::AliasIdxRaw; 3977 unlock->init_req( TypeFunc::Control, control() ); 3978 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3979 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3980 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3981 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3982 3983 unlock->init_req(TypeFunc::Parms + 0, obj); 3984 unlock->init_req(TypeFunc::Parms + 1, box); 3985 unlock = _gvn.transform(unlock)->as_Unlock(); 3986 3987 Node* mem = reset_memory(); 3988 3989 // unlock has no side-effects, sets few values 3990 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3991 3992 // Kill monitor from debug info 3993 map()->pop_monitor( ); 3994 } 3995 3996 //-------------------------------get_layout_helper----------------------------- 3997 // If the given klass is a constant or known to be an array, 3998 // fetch the constant layout helper value into constant_value 3999 // and return null. Otherwise, load the non-constant 4000 // layout helper value, and return the node which represents it. 4001 // This two-faced routine is useful because allocation sites 4002 // almost always feature constant types. 4003 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 4004 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr(); 4005 if (!StressReflectiveCode && klass_t != nullptr) { 4006 bool xklass = klass_t->klass_is_exact(); 4007 bool can_be_flat = false; 4008 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr(); 4009 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) { 4010 // Don't constant fold if the runtime type might be a flat array but the static type is not. 4011 const TypeOopPtr* elem = ary_type->elem()->make_oopptr(); 4012 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flat_in_array()); 4013 } 4014 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) { 4015 jint lhelper; 4016 if (klass_t->is_flat()) { 4017 lhelper = ary_type->flat_layout_helper(); 4018 } else if (klass_t->isa_aryklassptr()) { 4019 BasicType elem = ary_type->elem()->array_element_basic_type(); 4020 if (is_reference_type(elem, true)) { 4021 elem = T_OBJECT; 4022 } 4023 lhelper = Klass::array_layout_helper(elem); 4024 } else { 4025 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper(); 4026 } 4027 if (lhelper != Klass::_lh_neutral_value) { 4028 constant_value = lhelper; 4029 return (Node*) nullptr; 4030 } 4031 } 4032 } 4033 constant_value = Klass::_lh_neutral_value; // put in a known value 4034 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 4035 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered); 4036 } 4037 4038 // We just put in an allocate/initialize with a big raw-memory effect. 4039 // Hook selected additional alias categories on the initialization. 4040 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 4041 MergeMemNode* init_in_merge, 4042 Node* init_out_raw) { 4043 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 4044 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 4045 4046 Node* prevmem = kit.memory(alias_idx); 4047 init_in_merge->set_memory_at(alias_idx, prevmem); 4048 if (init_out_raw != nullptr) { 4049 kit.set_memory(init_out_raw, alias_idx); 4050 } 4051 } 4052 4053 //---------------------------set_output_for_allocation------------------------- 4054 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 4055 const TypeOopPtr* oop_type, 4056 bool deoptimize_on_exception) { 4057 int rawidx = Compile::AliasIdxRaw; 4058 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 4059 add_safepoint_edges(alloc); 4060 Node* allocx = _gvn.transform(alloc); 4061 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 4062 // create memory projection for i_o 4063 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 4064 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 4065 4066 // create a memory projection as for the normal control path 4067 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 4068 set_memory(malloc, rawidx); 4069 4070 // a normal slow-call doesn't change i_o, but an allocation does 4071 // we create a separate i_o projection for the normal control path 4072 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 4073 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 4074 4075 // put in an initialization barrier 4076 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 4077 rawoop)->as_Initialize(); 4078 assert(alloc->initialization() == init, "2-way macro link must work"); 4079 assert(init ->allocation() == alloc, "2-way macro link must work"); 4080 { 4081 // Extract memory strands which may participate in the new object's 4082 // initialization, and source them from the new InitializeNode. 4083 // This will allow us to observe initializations when they occur, 4084 // and link them properly (as a group) to the InitializeNode. 4085 assert(init->in(InitializeNode::Memory) == malloc, ""); 4086 MergeMemNode* minit_in = MergeMemNode::make(malloc); 4087 init->set_req(InitializeNode::Memory, minit_in); 4088 record_for_igvn(minit_in); // fold it up later, if possible 4089 _gvn.set_type(minit_in, Type::MEMORY); 4090 Node* minit_out = memory(rawidx); 4091 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 4092 // Add an edge in the MergeMem for the header fields so an access 4093 // to one of those has correct memory state 4094 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 4095 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 4096 if (oop_type->isa_aryptr()) { 4097 const TypeAryPtr* arytype = oop_type->is_aryptr(); 4098 if (arytype->is_flat()) { 4099 // Initially all flat array accesses share a single slice 4100 // but that changes after parsing. Prepare the memory graph so 4101 // it can optimize flat array accesses properly once they 4102 // don't share a single slice. 4103 assert(C->flat_accesses_share_alias(), "should be set at parse time"); 4104 C->set_flat_accesses_share_alias(false); 4105 ciInlineKlass* vk = arytype->elem()->inline_klass(); 4106 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) { 4107 ciField* field = vk->nonstatic_field_at(i); 4108 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize) 4109 continue; // do not bother to track really large numbers of fields 4110 int off_in_vt = field->offset_in_bytes() - vk->payload_offset(); 4111 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot); 4112 int fieldidx = C->get_alias_index(adr_type, true); 4113 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node 4114 // can result in per flat array field Phis to be created which confuses the logic of 4115 // Compile::adjust_flat_array_access_aliases(). 4116 hook_memory_on_init(*this, fieldidx, minit_in, nullptr); 4117 } 4118 C->set_flat_accesses_share_alias(true); 4119 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out); 4120 } else { 4121 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 4122 int elemidx = C->get_alias_index(telemref); 4123 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 4124 } 4125 } else if (oop_type->isa_instptr()) { 4126 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word 4127 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass(); 4128 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 4129 ciField* field = ik->nonstatic_field_at(i); 4130 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize) 4131 continue; // do not bother to track really large numbers of fields 4132 // Find (or create) the alias category for this field: 4133 int fieldidx = C->alias_type(field)->index(); 4134 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 4135 } 4136 } 4137 } 4138 4139 // Cast raw oop to the real thing... 4140 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 4141 javaoop = _gvn.transform(javaoop); 4142 C->set_recent_alloc(control(), javaoop); 4143 assert(just_allocated_object(control()) == javaoop, "just allocated"); 4144 4145 #ifdef ASSERT 4146 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 4147 assert(AllocateNode::Ideal_allocation(rawoop) == alloc, 4148 "Ideal_allocation works"); 4149 assert(AllocateNode::Ideal_allocation(javaoop) == alloc, 4150 "Ideal_allocation works"); 4151 if (alloc->is_AllocateArray()) { 4152 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(), 4153 "Ideal_allocation works"); 4154 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(), 4155 "Ideal_allocation works"); 4156 } else { 4157 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 4158 } 4159 } 4160 #endif //ASSERT 4161 4162 return javaoop; 4163 } 4164 4165 //---------------------------new_instance-------------------------------------- 4166 // This routine takes a klass_node which may be constant (for a static type) 4167 // or may be non-constant (for reflective code). It will work equally well 4168 // for either, and the graph will fold nicely if the optimizer later reduces 4169 // the type to a constant. 4170 // The optional arguments are for specialized use by intrinsics: 4171 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 4172 // - If 'return_size_val', report the total object size to the caller. 4173 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 4174 Node* GraphKit::new_instance(Node* klass_node, 4175 Node* extra_slow_test, 4176 Node* *return_size_val, 4177 bool deoptimize_on_exception, 4178 InlineTypeNode* inline_type_node) { 4179 // Compute size in doublewords 4180 // The size is always an integral number of doublewords, represented 4181 // as a positive bytewise size stored in the klass's layout_helper. 4182 // The layout_helper also encodes (in a low bit) the need for a slow path. 4183 jint layout_con = Klass::_lh_neutral_value; 4184 Node* layout_val = get_layout_helper(klass_node, layout_con); 4185 bool layout_is_con = (layout_val == nullptr); 4186 4187 if (extra_slow_test == nullptr) extra_slow_test = intcon(0); 4188 // Generate the initial go-slow test. It's either ALWAYS (return a 4189 // Node for 1) or NEVER (return a null) or perhaps (in the reflective 4190 // case) a computed value derived from the layout_helper. 4191 Node* initial_slow_test = nullptr; 4192 if (layout_is_con) { 4193 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4194 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 4195 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 4196 } else { // reflective case 4197 // This reflective path is used by Unsafe.allocateInstance. 4198 // (It may be stress-tested by specifying StressReflectiveCode.) 4199 // Basically, we want to get into the VM is there's an illegal argument. 4200 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 4201 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 4202 if (extra_slow_test != intcon(0)) { 4203 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 4204 } 4205 // (Macro-expander will further convert this to a Bool, if necessary.) 4206 } 4207 4208 // Find the size in bytes. This is easy; it's the layout_helper. 4209 // The size value must be valid even if the slow path is taken. 4210 Node* size = nullptr; 4211 if (layout_is_con) { 4212 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 4213 } else { // reflective case 4214 // This reflective path is used by clone and Unsafe.allocateInstance. 4215 size = ConvI2X(layout_val); 4216 4217 // Clear the low bits to extract layout_helper_size_in_bytes: 4218 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 4219 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 4220 size = _gvn.transform( new AndXNode(size, mask) ); 4221 } 4222 if (return_size_val != nullptr) { 4223 (*return_size_val) = size; 4224 } 4225 4226 // This is a precise notnull oop of the klass. 4227 // (Actually, it need not be precise if this is a reflective allocation.) 4228 // It's what we cast the result to. 4229 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 4230 if (!tklass) tklass = TypeInstKlassPtr::OBJECT; 4231 const TypeOopPtr* oop_type = tklass->as_instance_type(); 4232 4233 // Now generate allocation code 4234 4235 // The entire memory state is needed for slow path of the allocation 4236 // since GC and deoptimization can happen. 4237 Node *mem = reset_memory(); 4238 set_all_memory(mem); // Create new memory state 4239 4240 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 4241 control(), mem, i_o(), 4242 size, klass_node, 4243 initial_slow_test, inline_type_node); 4244 4245 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 4246 } 4247 4248 //-------------------------------new_array------------------------------------- 4249 // helper for newarray and anewarray 4250 // The 'length' parameter is (obviously) the length of the array. 4251 // The optional arguments are for specialized use by intrinsics: 4252 // - If 'return_size_val', report the non-padded array size (sum of header size 4253 // and array body) to the caller. 4254 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 4255 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 4256 Node* length, // number of array elements 4257 int nargs, // number of arguments to push back for uncommon trap 4258 Node* *return_size_val, 4259 bool deoptimize_on_exception) { 4260 jint layout_con = Klass::_lh_neutral_value; 4261 Node* layout_val = get_layout_helper(klass_node, layout_con); 4262 bool layout_is_con = (layout_val == nullptr); 4263 4264 if (!layout_is_con && !StressReflectiveCode && 4265 !too_many_traps(Deoptimization::Reason_class_check)) { 4266 // This is a reflective array creation site. 4267 // Optimistically assume that it is a subtype of Object[], 4268 // so that we can fold up all the address arithmetic. 4269 layout_con = Klass::array_layout_helper(T_OBJECT); 4270 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 4271 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 4272 { BuildCutout unless(this, bol_lh, PROB_MAX); 4273 inc_sp(nargs); 4274 uncommon_trap(Deoptimization::Reason_class_check, 4275 Deoptimization::Action_maybe_recompile); 4276 } 4277 layout_val = nullptr; 4278 layout_is_con = true; 4279 } 4280 4281 // Generate the initial go-slow test. Make sure we do not overflow 4282 // if length is huge (near 2Gig) or negative! We do not need 4283 // exact double-words here, just a close approximation of needed 4284 // double-words. We can't add any offset or rounding bits, lest we 4285 // take a size -1 of bytes and make it positive. Use an unsigned 4286 // compare, so negative sizes look hugely positive. 4287 int fast_size_limit = FastAllocateSizeLimit; 4288 if (layout_is_con) { 4289 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4290 // Increase the size limit if we have exact knowledge of array type. 4291 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 4292 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0); 4293 } 4294 4295 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 4296 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 4297 4298 // --- Size Computation --- 4299 // array_size = round_to_heap(array_header + (length << elem_shift)); 4300 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 4301 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 4302 // The rounding mask is strength-reduced, if possible. 4303 int round_mask = MinObjAlignmentInBytes - 1; 4304 Node* header_size = nullptr; 4305 // (T_BYTE has the weakest alignment and size restrictions...) 4306 if (layout_is_con) { 4307 int hsize = Klass::layout_helper_header_size(layout_con); 4308 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4309 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con); 4310 if ((round_mask & ~right_n_bits(eshift)) == 0) 4311 round_mask = 0; // strength-reduce it if it goes away completely 4312 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 4313 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 4314 assert(header_size_min <= hsize, "generic minimum is smallest"); 4315 header_size = intcon(hsize); 4316 } else { 4317 Node* hss = intcon(Klass::_lh_header_size_shift); 4318 Node* hsm = intcon(Klass::_lh_header_size_mask); 4319 header_size = _gvn.transform(new URShiftINode(layout_val, hss)); 4320 header_size = _gvn.transform(new AndINode(header_size, hsm)); 4321 } 4322 4323 Node* elem_shift = nullptr; 4324 if (layout_is_con) { 4325 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4326 if (eshift != 0) 4327 elem_shift = intcon(eshift); 4328 } else { 4329 // There is no need to mask or shift this value. 4330 // The semantics of LShiftINode include an implicit mask to 0x1F. 4331 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 4332 elem_shift = layout_val; 4333 } 4334 4335 // Transition to native address size for all offset calculations: 4336 Node* lengthx = ConvI2X(length); 4337 Node* headerx = ConvI2X(header_size); 4338 #ifdef _LP64 4339 { const TypeInt* tilen = _gvn.find_int_type(length); 4340 if (tilen != nullptr && tilen->_lo < 0) { 4341 // Add a manual constraint to a positive range. Cf. array_element_address. 4342 jint size_max = fast_size_limit; 4343 if (size_max > tilen->_hi) size_max = tilen->_hi; 4344 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 4345 4346 // Only do a narrow I2L conversion if the range check passed. 4347 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 4348 _gvn.transform(iff); 4349 RegionNode* region = new RegionNode(3); 4350 _gvn.set_type(region, Type::CONTROL); 4351 lengthx = new PhiNode(region, TypeLong::LONG); 4352 _gvn.set_type(lengthx, TypeLong::LONG); 4353 4354 // Range check passed. Use ConvI2L node with narrow type. 4355 Node* passed = IfFalse(iff); 4356 region->init_req(1, passed); 4357 // Make I2L conversion control dependent to prevent it from 4358 // floating above the range check during loop optimizations. 4359 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 4360 4361 // Range check failed. Use ConvI2L with wide type because length may be invalid. 4362 region->init_req(2, IfTrue(iff)); 4363 lengthx->init_req(2, ConvI2X(length)); 4364 4365 set_control(region); 4366 record_for_igvn(region); 4367 record_for_igvn(lengthx); 4368 } 4369 } 4370 #endif 4371 4372 // Combine header size and body size for the array copy part, then align (if 4373 // necessary) for the allocation part. This computation cannot overflow, 4374 // because it is used only in two places, one where the length is sharply 4375 // limited, and the other after a successful allocation. 4376 Node* abody = lengthx; 4377 if (elem_shift != nullptr) { 4378 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift)); 4379 } 4380 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody)); 4381 4382 if (return_size_val != nullptr) { 4383 // This is the size 4384 (*return_size_val) = non_rounded_size; 4385 } 4386 4387 Node* size = non_rounded_size; 4388 if (round_mask != 0) { 4389 Node* mask1 = MakeConX(round_mask); 4390 size = _gvn.transform(new AddXNode(size, mask1)); 4391 Node* mask2 = MakeConX(~round_mask); 4392 size = _gvn.transform(new AndXNode(size, mask2)); 4393 } 4394 // else if round_mask == 0, the size computation is self-rounding 4395 4396 // Now generate allocation code 4397 4398 // The entire memory state is needed for slow path of the allocation 4399 // since GC and deoptimization can happen. 4400 Node *mem = reset_memory(); 4401 set_all_memory(mem); // Create new memory state 4402 4403 if (initial_slow_test->is_Bool()) { 4404 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4405 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4406 } 4407 4408 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr(); 4409 const TypeOopPtr* ary_type = ary_klass->as_instance_type(); 4410 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 4411 4412 // Inline type array variants: 4413 // - null-ok: ciObjArrayKlass with is_elem_null_free() = false 4414 // - null-free: ciObjArrayKlass with is_elem_null_free() = true 4415 // - null-free, flat: ciFlatArrayKlass with is_elem_null_free() = true 4416 // Check if array is a null-free, non-flat inline type array 4417 // that needs to be initialized with the default inline type. 4418 Node* default_value = nullptr; 4419 Node* raw_default_value = nullptr; 4420 if (ary_ptr != nullptr && ary_ptr->klass_is_exact()) { 4421 // Array type is known 4422 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) { 4423 ciInlineKlass* vk = ary_ptr->elem()->inline_klass(); 4424 default_value = InlineTypeNode::default_oop(gvn(), vk); 4425 if (UseCompressedOops) { 4426 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops 4427 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop())); 4428 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value)); 4429 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32))); 4430 raw_default_value = _gvn.transform(new OrLNode(lower, upper)); 4431 } else { 4432 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value)); 4433 } 4434 } 4435 } 4436 4437 Node* valid_length_test = _gvn.intcon(1); 4438 if (ary_type->isa_aryptr()) { 4439 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type(); 4440 jint max = TypeAryPtr::max_array_length(bt); 4441 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max))); 4442 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le)); 4443 } 4444 4445 // Create the AllocateArrayNode and its result projections 4446 AllocateArrayNode* alloc 4447 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4448 control(), mem, i_o(), 4449 size, klass_node, 4450 initial_slow_test, 4451 length, valid_length_test, 4452 default_value, raw_default_value); 4453 // Cast to correct type. Note that the klass_node may be constant or not, 4454 // and in the latter case the actual array type will be inexact also. 4455 // (This happens via a non-constant argument to inline_native_newArray.) 4456 // In any case, the value of klass_node provides the desired array type. 4457 const TypeInt* length_type = _gvn.find_int_type(length); 4458 if (ary_type->isa_aryptr() && length_type != nullptr) { 4459 // Try to get a better type than POS for the size 4460 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4461 } 4462 4463 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4464 4465 array_ideal_length(alloc, ary_type, true); 4466 return javaoop; 4467 } 4468 4469 // The following "Ideal_foo" functions are placed here because they recognize 4470 // the graph shapes created by the functions immediately above. 4471 4472 //---------------------------Ideal_allocation---------------------------------- 4473 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4474 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) { 4475 if (ptr == nullptr) { // reduce dumb test in callers 4476 return nullptr; 4477 } 4478 4479 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4480 ptr = bs->step_over_gc_barrier(ptr); 4481 4482 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4483 ptr = ptr->in(1); 4484 if (ptr == nullptr) return nullptr; 4485 } 4486 // Return null for allocations with several casts: 4487 // j.l.reflect.Array.newInstance(jobject, jint) 4488 // Object.clone() 4489 // to keep more precise type from last cast. 4490 if (ptr->is_Proj()) { 4491 Node* allo = ptr->in(0); 4492 if (allo != nullptr && allo->is_Allocate()) { 4493 return allo->as_Allocate(); 4494 } 4495 } 4496 // Report failure to match. 4497 return nullptr; 4498 } 4499 4500 // Fancy version which also strips off an offset (and reports it to caller). 4501 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase, 4502 intptr_t& offset) { 4503 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4504 if (base == nullptr) return nullptr; 4505 return Ideal_allocation(base); 4506 } 4507 4508 // Trace Initialize <- Proj[Parm] <- Allocate 4509 AllocateNode* InitializeNode::allocation() { 4510 Node* rawoop = in(InitializeNode::RawAddress); 4511 if (rawoop->is_Proj()) { 4512 Node* alloc = rawoop->in(0); 4513 if (alloc->is_Allocate()) { 4514 return alloc->as_Allocate(); 4515 } 4516 } 4517 return nullptr; 4518 } 4519 4520 // Trace Allocate -> Proj[Parm] -> Initialize 4521 InitializeNode* AllocateNode::initialization() { 4522 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4523 if (rawoop == nullptr) return nullptr; 4524 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4525 Node* init = rawoop->fast_out(i); 4526 if (init->is_Initialize()) { 4527 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4528 return init->as_Initialize(); 4529 } 4530 } 4531 return nullptr; 4532 } 4533 4534 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path. 4535 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) { 4536 // Too many traps seen? 4537 if (too_many_traps(reason)) { 4538 #ifdef ASSERT 4539 if (TraceLoopPredicate) { 4540 int tc = C->trap_count(reason); 4541 tty->print("too many traps=%s tcount=%d in ", 4542 Deoptimization::trap_reason_name(reason), tc); 4543 method()->print(); // which method has too many predicate traps 4544 tty->cr(); 4545 } 4546 #endif 4547 // We cannot afford to take more traps here, 4548 // do not generate Parse Predicate. 4549 return; 4550 } 4551 4552 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn); 4553 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn)); 4554 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate)); 4555 { 4556 PreserveJVMState pjvms(this); 4557 set_control(if_false); 4558 inc_sp(nargs); 4559 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4560 } 4561 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate)); 4562 set_control(if_true); 4563 } 4564 4565 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All 4566 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate. 4567 void GraphKit::add_parse_predicates(int nargs) { 4568 if (UseLoopPredicate) { 4569 add_parse_predicate(Deoptimization::Reason_predicate, nargs); 4570 } 4571 if (UseProfiledLoopPredicate) { 4572 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs); 4573 } 4574 // Loop Limit Check Predicate should be near the loop. 4575 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs); 4576 } 4577 4578 void GraphKit::sync_kit(IdealKit& ideal) { 4579 set_all_memory(ideal.merged_memory()); 4580 set_i_o(ideal.i_o()); 4581 set_control(ideal.ctrl()); 4582 } 4583 4584 void GraphKit::final_sync(IdealKit& ideal) { 4585 // Final sync IdealKit and graphKit. 4586 sync_kit(ideal); 4587 } 4588 4589 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4590 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4591 Node* coder = load_String_coder(str, set_ctrl); 4592 // Divide length by 2 if coder is UTF16 4593 return _gvn.transform(new RShiftINode(len, coder)); 4594 } 4595 4596 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4597 int value_offset = java_lang_String::value_offset(); 4598 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4599 false, nullptr, Type::Offset(0)); 4600 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4601 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4602 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true), 4603 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4604 Node* p = basic_plus_adr(str, str, value_offset); 4605 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4606 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4607 return load; 4608 } 4609 4610 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4611 if (!CompactStrings) { 4612 return intcon(java_lang_String::CODER_UTF16); 4613 } 4614 int coder_offset = java_lang_String::coder_offset(); 4615 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4616 false, nullptr, Type::Offset(0)); 4617 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4618 4619 Node* p = basic_plus_adr(str, str, coder_offset); 4620 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4621 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4622 return load; 4623 } 4624 4625 void GraphKit::store_String_value(Node* str, Node* value) { 4626 int value_offset = java_lang_String::value_offset(); 4627 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4628 false, nullptr, Type::Offset(0)); 4629 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4630 4631 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4632 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4633 } 4634 4635 void GraphKit::store_String_coder(Node* str, Node* value) { 4636 int coder_offset = java_lang_String::coder_offset(); 4637 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4638 false, nullptr, Type::Offset(0)); 4639 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4640 4641 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4642 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4643 } 4644 4645 // Capture src and dst memory state with a MergeMemNode 4646 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4647 if (src_type == dst_type) { 4648 // Types are equal, we don't need a MergeMemNode 4649 return memory(src_type); 4650 } 4651 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4652 record_for_igvn(merge); // fold it up later, if possible 4653 int src_idx = C->get_alias_index(src_type); 4654 int dst_idx = C->get_alias_index(dst_type); 4655 merge->set_memory_at(src_idx, memory(src_idx)); 4656 merge->set_memory_at(dst_idx, memory(dst_idx)); 4657 return merge; 4658 } 4659 4660 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4661 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4662 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4663 // If input and output memory types differ, capture both states to preserve 4664 // the dependency between preceding and subsequent loads/stores. 4665 // For example, the following program: 4666 // StoreB 4667 // compress_string 4668 // LoadB 4669 // has this memory graph (use->def): 4670 // LoadB -> compress_string -> CharMem 4671 // ... -> StoreB -> ByteMem 4672 // The intrinsic hides the dependency between LoadB and StoreB, causing 4673 // the load to read from memory not containing the result of the StoreB. 4674 // The correct memory graph should look like this: 4675 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4676 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4677 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4678 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str))); 4679 set_memory(res_mem, TypeAryPtr::BYTES); 4680 return str; 4681 } 4682 4683 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4684 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4685 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4686 // Capture src and dst memory (see comment in 'compress_string'). 4687 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4688 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4689 set_memory(_gvn.transform(str), dst_type); 4690 } 4691 4692 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4693 /** 4694 * int i_char = start; 4695 * for (int i_byte = 0; i_byte < count; i_byte++) { 4696 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4697 * } 4698 */ 4699 add_parse_predicates(); 4700 C->set_has_loops(true); 4701 4702 RegionNode* head = new RegionNode(3); 4703 head->init_req(1, control()); 4704 gvn().set_type(head, Type::CONTROL); 4705 record_for_igvn(head); 4706 4707 Node* i_byte = new PhiNode(head, TypeInt::INT); 4708 i_byte->init_req(1, intcon(0)); 4709 gvn().set_type(i_byte, TypeInt::INT); 4710 record_for_igvn(i_byte); 4711 4712 Node* i_char = new PhiNode(head, TypeInt::INT); 4713 i_char->init_req(1, start); 4714 gvn().set_type(i_char, TypeInt::INT); 4715 record_for_igvn(i_char); 4716 4717 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4718 gvn().set_type(mem, Type::MEMORY); 4719 record_for_igvn(mem); 4720 set_control(head); 4721 set_memory(mem, TypeAryPtr::BYTES); 4722 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true); 4723 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4724 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false, 4725 false, true /* mismatched */); 4726 4727 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4728 head->init_req(2, IfTrue(iff)); 4729 mem->init_req(2, st); 4730 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4731 i_char->init_req(2, AddI(i_char, intcon(2))); 4732 4733 set_control(IfFalse(iff)); 4734 set_memory(st, TypeAryPtr::BYTES); 4735 } 4736 4737 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4738 if (!field->is_constant()) { 4739 return nullptr; // Field not marked as constant. 4740 } 4741 ciInstance* holder = nullptr; 4742 if (!field->is_static()) { 4743 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4744 if (const_oop != nullptr && const_oop->is_instance()) { 4745 holder = const_oop->as_instance(); 4746 } 4747 } 4748 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4749 /*is_unsigned_load=*/false); 4750 if (con_type != nullptr) { 4751 Node* con = makecon(con_type); 4752 if (field->type()->is_inlinetype()) { 4753 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free()); 4754 } else if (con_type->is_inlinetypeptr()) { 4755 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free()); 4756 } 4757 return con; 4758 } 4759 return nullptr; 4760 } 4761 4762 //---------------------------load_mirror_from_klass---------------------------- 4763 // Given a klass oop, load its java mirror (a java.lang.Class oop). 4764 Node* GraphKit::load_mirror_from_klass(Node* klass) { 4765 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 4766 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4767 // mirror = ((OopHandle)mirror)->resolve(); 4768 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 4769 } 4770 4771 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) { 4772 const Type* obj_type = obj->bottom_type(); 4773 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type); 4774 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) { 4775 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part 4776 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type)); 4777 obj = casted_obj; 4778 } 4779 if (sig_type->is_inlinetypeptr()) { 4780 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass(), !gvn().type(obj)->maybe_null()); 4781 } 4782 return obj; 4783 }