1 /* 2 * Copyright (c) 2001, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciUtilities.hpp" 27 #include "classfile/javaClasses.hpp" 28 #include "ci/ciObjArray.hpp" 29 #include "asm/register.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "gc/shared/barrierSet.hpp" 32 #include "gc/shared/c2/barrierSetC2.hpp" 33 #include "interpreter/interpreter.hpp" 34 #include "memory/resourceArea.hpp" 35 #include "opto/addnode.hpp" 36 #include "opto/castnode.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/graphKit.hpp" 39 #include "opto/idealKit.hpp" 40 #include "opto/intrinsicnode.hpp" 41 #include "opto/locknode.hpp" 42 #include "opto/machnode.hpp" 43 #include "opto/opaquenode.hpp" 44 #include "opto/parse.hpp" 45 #include "opto/rootnode.hpp" 46 #include "opto/runtime.hpp" 47 #include "opto/subtypenode.hpp" 48 #include "runtime/deoptimization.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 #include "utilities/bitMap.inline.hpp" 51 #include "utilities/powerOfTwo.hpp" 52 #include "utilities/growableArray.hpp" 53 54 //----------------------------GraphKit----------------------------------------- 55 // Main utility constructor. 56 GraphKit::GraphKit(JVMState* jvms) 57 : Phase(Phase::Parser), 58 _env(C->env()), 59 _gvn(*C->initial_gvn()), 60 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 61 { 62 _exceptions = jvms->map()->next_exception(); 63 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr); 64 set_jvms(jvms); 65 } 66 67 // Private constructor for parser. 68 GraphKit::GraphKit() 69 : Phase(Phase::Parser), 70 _env(C->env()), 71 _gvn(*C->initial_gvn()), 72 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 73 { 74 _exceptions = nullptr; 75 set_map(nullptr); 76 debug_only(_sp = -99); 77 debug_only(set_bci(-99)); 78 } 79 80 81 82 //---------------------------clean_stack--------------------------------------- 83 // Clear away rubbish from the stack area of the JVM state. 84 // This destroys any arguments that may be waiting on the stack. 85 void GraphKit::clean_stack(int from_sp) { 86 SafePointNode* map = this->map(); 87 JVMState* jvms = this->jvms(); 88 int stk_size = jvms->stk_size(); 89 int stkoff = jvms->stkoff(); 90 Node* top = this->top(); 91 for (int i = from_sp; i < stk_size; i++) { 92 if (map->in(stkoff + i) != top) { 93 map->set_req(stkoff + i, top); 94 } 95 } 96 } 97 98 99 //--------------------------------sync_jvms----------------------------------- 100 // Make sure our current jvms agrees with our parse state. 101 JVMState* GraphKit::sync_jvms() const { 102 JVMState* jvms = this->jvms(); 103 jvms->set_bci(bci()); // Record the new bci in the JVMState 104 jvms->set_sp(sp()); // Record the new sp in the JVMState 105 assert(jvms_in_sync(), "jvms is now in sync"); 106 return jvms; 107 } 108 109 //--------------------------------sync_jvms_for_reexecute--------------------- 110 // Make sure our current jvms agrees with our parse state. This version 111 // uses the reexecute_sp for reexecuting bytecodes. 112 JVMState* GraphKit::sync_jvms_for_reexecute() { 113 JVMState* jvms = this->jvms(); 114 jvms->set_bci(bci()); // Record the new bci in the JVMState 115 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 116 return jvms; 117 } 118 119 #ifdef ASSERT 120 bool GraphKit::jvms_in_sync() const { 121 Parse* parse = is_Parse(); 122 if (parse == nullptr) { 123 if (bci() != jvms()->bci()) return false; 124 if (sp() != (int)jvms()->sp()) return false; 125 return true; 126 } 127 if (jvms()->method() != parse->method()) return false; 128 if (jvms()->bci() != parse->bci()) return false; 129 int jvms_sp = jvms()->sp(); 130 if (jvms_sp != parse->sp()) return false; 131 int jvms_depth = jvms()->depth(); 132 if (jvms_depth != parse->depth()) return false; 133 return true; 134 } 135 136 // Local helper checks for special internal merge points 137 // used to accumulate and merge exception states. 138 // They are marked by the region's in(0) edge being the map itself. 139 // Such merge points must never "escape" into the parser at large, 140 // until they have been handed to gvn.transform. 141 static bool is_hidden_merge(Node* reg) { 142 if (reg == nullptr) return false; 143 if (reg->is_Phi()) { 144 reg = reg->in(0); 145 if (reg == nullptr) return false; 146 } 147 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root(); 148 } 149 150 void GraphKit::verify_map() const { 151 if (map() == nullptr) return; // null map is OK 152 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 153 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 154 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 155 } 156 157 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 158 assert(ex_map->next_exception() == nullptr, "not already part of a chain"); 159 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 160 } 161 #endif 162 163 //---------------------------stop_and_kill_map--------------------------------- 164 // Set _map to null, signalling a stop to further bytecode execution. 165 // First smash the current map's control to a constant, to mark it dead. 166 void GraphKit::stop_and_kill_map() { 167 SafePointNode* dead_map = stop(); 168 if (dead_map != nullptr) { 169 dead_map->disconnect_inputs(C); // Mark the map as killed. 170 assert(dead_map->is_killed(), "must be so marked"); 171 } 172 } 173 174 175 //--------------------------------stopped-------------------------------------- 176 // Tell if _map is null, or control is top. 177 bool GraphKit::stopped() { 178 if (map() == nullptr) return true; 179 else if (control() == top()) return true; 180 else return false; 181 } 182 183 184 //-----------------------------has_exception_handler---------------------------------- 185 // Tell if this method or any caller method has exception handlers. 186 bool GraphKit::has_exception_handler() { 187 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) { 188 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 189 return true; 190 } 191 } 192 return false; 193 } 194 195 //------------------------------save_ex_oop------------------------------------ 196 // Save an exception without blowing stack contents or other JVM state. 197 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 198 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 199 ex_map->add_req(ex_oop); 200 debug_only(verify_exception_state(ex_map)); 201 } 202 203 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 204 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 205 Node* ex_oop = ex_map->in(ex_map->req()-1); 206 if (clear_it) ex_map->del_req(ex_map->req()-1); 207 return ex_oop; 208 } 209 210 //-----------------------------saved_ex_oop------------------------------------ 211 // Recover a saved exception from its map. 212 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 213 return common_saved_ex_oop(ex_map, false); 214 } 215 216 //--------------------------clear_saved_ex_oop--------------------------------- 217 // Erase a previously saved exception from its map. 218 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 219 return common_saved_ex_oop(ex_map, true); 220 } 221 222 #ifdef ASSERT 223 //---------------------------has_saved_ex_oop---------------------------------- 224 // Erase a previously saved exception from its map. 225 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 226 return ex_map->req() == ex_map->jvms()->endoff()+1; 227 } 228 #endif 229 230 //-------------------------make_exception_state-------------------------------- 231 // Turn the current JVM state into an exception state, appending the ex_oop. 232 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 233 sync_jvms(); 234 SafePointNode* ex_map = stop(); // do not manipulate this map any more 235 set_saved_ex_oop(ex_map, ex_oop); 236 return ex_map; 237 } 238 239 240 //--------------------------add_exception_state-------------------------------- 241 // Add an exception to my list of exceptions. 242 void GraphKit::add_exception_state(SafePointNode* ex_map) { 243 if (ex_map == nullptr || ex_map->control() == top()) { 244 return; 245 } 246 #ifdef ASSERT 247 verify_exception_state(ex_map); 248 if (has_exceptions()) { 249 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 250 } 251 #endif 252 253 // If there is already an exception of exactly this type, merge with it. 254 // In particular, null-checks and other low-level exceptions common up here. 255 Node* ex_oop = saved_ex_oop(ex_map); 256 const Type* ex_type = _gvn.type(ex_oop); 257 if (ex_oop == top()) { 258 // No action needed. 259 return; 260 } 261 assert(ex_type->isa_instptr(), "exception must be an instance"); 262 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) { 263 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 264 // We check sp also because call bytecodes can generate exceptions 265 // both before and after arguments are popped! 266 if (ex_type2 == ex_type 267 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 268 combine_exception_states(ex_map, e2); 269 return; 270 } 271 } 272 273 // No pre-existing exception of the same type. Chain it on the list. 274 push_exception_state(ex_map); 275 } 276 277 //-----------------------add_exception_states_from----------------------------- 278 void GraphKit::add_exception_states_from(JVMState* jvms) { 279 SafePointNode* ex_map = jvms->map()->next_exception(); 280 if (ex_map != nullptr) { 281 jvms->map()->set_next_exception(nullptr); 282 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) { 283 next_map = ex_map->next_exception(); 284 ex_map->set_next_exception(nullptr); 285 add_exception_state(ex_map); 286 } 287 } 288 } 289 290 //-----------------------transfer_exceptions_into_jvms------------------------- 291 JVMState* GraphKit::transfer_exceptions_into_jvms() { 292 if (map() == nullptr) { 293 // We need a JVMS to carry the exceptions, but the map has gone away. 294 // Create a scratch JVMS, cloned from any of the exception states... 295 if (has_exceptions()) { 296 _map = _exceptions; 297 _map = clone_map(); 298 _map->set_next_exception(nullptr); 299 clear_saved_ex_oop(_map); 300 debug_only(verify_map()); 301 } else { 302 // ...or created from scratch 303 JVMState* jvms = new (C) JVMState(_method, nullptr); 304 jvms->set_bci(_bci); 305 jvms->set_sp(_sp); 306 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 307 set_jvms(jvms); 308 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 309 set_all_memory(top()); 310 while (map()->req() < jvms->endoff()) map()->add_req(top()); 311 } 312 // (This is a kludge, in case you didn't notice.) 313 set_control(top()); 314 } 315 JVMState* jvms = sync_jvms(); 316 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 317 jvms->map()->set_next_exception(_exceptions); 318 _exceptions = nullptr; // done with this set of exceptions 319 return jvms; 320 } 321 322 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 323 assert(is_hidden_merge(dstphi), "must be a special merge node"); 324 assert(is_hidden_merge(srcphi), "must be a special merge node"); 325 uint limit = srcphi->req(); 326 for (uint i = PhiNode::Input; i < limit; i++) { 327 dstphi->add_req(srcphi->in(i)); 328 } 329 } 330 static inline void add_one_req(Node* dstphi, Node* src) { 331 assert(is_hidden_merge(dstphi), "must be a special merge node"); 332 assert(!is_hidden_merge(src), "must not be a special merge node"); 333 dstphi->add_req(src); 334 } 335 336 //-----------------------combine_exception_states------------------------------ 337 // This helper function combines exception states by building phis on a 338 // specially marked state-merging region. These regions and phis are 339 // untransformed, and can build up gradually. The region is marked by 340 // having a control input of its exception map, rather than null. Such 341 // regions do not appear except in this function, and in use_exception_state. 342 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 343 if (failing()) return; // dying anyway... 344 JVMState* ex_jvms = ex_map->_jvms; 345 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 346 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 347 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 348 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 349 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 350 assert(ex_map->req() == phi_map->req(), "matching maps"); 351 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 352 Node* hidden_merge_mark = root(); 353 Node* region = phi_map->control(); 354 MergeMemNode* phi_mem = phi_map->merged_memory(); 355 MergeMemNode* ex_mem = ex_map->merged_memory(); 356 if (region->in(0) != hidden_merge_mark) { 357 // The control input is not (yet) a specially-marked region in phi_map. 358 // Make it so, and build some phis. 359 region = new RegionNode(2); 360 _gvn.set_type(region, Type::CONTROL); 361 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 362 region->init_req(1, phi_map->control()); 363 phi_map->set_control(region); 364 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 365 record_for_igvn(io_phi); 366 _gvn.set_type(io_phi, Type::ABIO); 367 phi_map->set_i_o(io_phi); 368 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 369 Node* m = mms.memory(); 370 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 371 record_for_igvn(m_phi); 372 _gvn.set_type(m_phi, Type::MEMORY); 373 mms.set_memory(m_phi); 374 } 375 } 376 377 // Either or both of phi_map and ex_map might already be converted into phis. 378 Node* ex_control = ex_map->control(); 379 // if there is special marking on ex_map also, we add multiple edges from src 380 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 381 // how wide was the destination phi_map, originally? 382 uint orig_width = region->req(); 383 384 if (add_multiple) { 385 add_n_reqs(region, ex_control); 386 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 387 } else { 388 // ex_map has no merges, so we just add single edges everywhere 389 add_one_req(region, ex_control); 390 add_one_req(phi_map->i_o(), ex_map->i_o()); 391 } 392 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 393 if (mms.is_empty()) { 394 // get a copy of the base memory, and patch some inputs into it 395 const TypePtr* adr_type = mms.adr_type(C); 396 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 397 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 398 mms.set_memory(phi); 399 // Prepare to append interesting stuff onto the newly sliced phi: 400 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 401 } 402 // Append stuff from ex_map: 403 if (add_multiple) { 404 add_n_reqs(mms.memory(), mms.memory2()); 405 } else { 406 add_one_req(mms.memory(), mms.memory2()); 407 } 408 } 409 uint limit = ex_map->req(); 410 for (uint i = TypeFunc::Parms; i < limit; i++) { 411 // Skip everything in the JVMS after tos. (The ex_oop follows.) 412 if (i == tos) i = ex_jvms->monoff(); 413 Node* src = ex_map->in(i); 414 Node* dst = phi_map->in(i); 415 if (src != dst) { 416 PhiNode* phi; 417 if (dst->in(0) != region) { 418 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 419 record_for_igvn(phi); 420 _gvn.set_type(phi, phi->type()); 421 phi_map->set_req(i, dst); 422 // Prepare to append interesting stuff onto the new phi: 423 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 424 } else { 425 assert(dst->is_Phi(), "nobody else uses a hidden region"); 426 phi = dst->as_Phi(); 427 } 428 if (add_multiple && src->in(0) == ex_control) { 429 // Both are phis. 430 add_n_reqs(dst, src); 431 } else { 432 while (dst->req() < region->req()) add_one_req(dst, src); 433 } 434 const Type* srctype = _gvn.type(src); 435 if (phi->type() != srctype) { 436 const Type* dsttype = phi->type()->meet_speculative(srctype); 437 if (phi->type() != dsttype) { 438 phi->set_type(dsttype); 439 _gvn.set_type(phi, dsttype); 440 } 441 } 442 } 443 } 444 phi_map->merge_replaced_nodes_with(ex_map); 445 } 446 447 //--------------------------use_exception_state-------------------------------- 448 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 449 if (failing()) { stop(); return top(); } 450 Node* region = phi_map->control(); 451 Node* hidden_merge_mark = root(); 452 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 453 Node* ex_oop = clear_saved_ex_oop(phi_map); 454 if (region->in(0) == hidden_merge_mark) { 455 // Special marking for internal ex-states. Process the phis now. 456 region->set_req(0, region); // now it's an ordinary region 457 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 458 // Note: Setting the jvms also sets the bci and sp. 459 set_control(_gvn.transform(region)); 460 uint tos = jvms()->stkoff() + sp(); 461 for (uint i = 1; i < tos; i++) { 462 Node* x = phi_map->in(i); 463 if (x->in(0) == region) { 464 assert(x->is_Phi(), "expected a special phi"); 465 phi_map->set_req(i, _gvn.transform(x)); 466 } 467 } 468 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 469 Node* x = mms.memory(); 470 if (x->in(0) == region) { 471 assert(x->is_Phi(), "nobody else uses a hidden region"); 472 mms.set_memory(_gvn.transform(x)); 473 } 474 } 475 if (ex_oop->in(0) == region) { 476 assert(ex_oop->is_Phi(), "expected a special phi"); 477 ex_oop = _gvn.transform(ex_oop); 478 } 479 } else { 480 set_jvms(phi_map->jvms()); 481 } 482 483 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 484 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 485 return ex_oop; 486 } 487 488 //---------------------------------java_bc------------------------------------- 489 Bytecodes::Code GraphKit::java_bc() const { 490 ciMethod* method = this->method(); 491 int bci = this->bci(); 492 if (method != nullptr && bci != InvocationEntryBci) 493 return method->java_code_at_bci(bci); 494 else 495 return Bytecodes::_illegal; 496 } 497 498 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 499 bool must_throw) { 500 // if the exception capability is set, then we will generate code 501 // to check the JavaThread.should_post_on_exceptions flag to see 502 // if we actually need to report exception events (for this 503 // thread). If we don't need to report exception events, we will 504 // take the normal fast path provided by add_exception_events. If 505 // exception event reporting is enabled for this thread, we will 506 // take the uncommon_trap in the BuildCutout below. 507 508 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 509 Node* jthread = _gvn.transform(new ThreadLocalNode()); 510 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 511 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 512 513 // Test the should_post_on_exceptions_flag vs. 0 514 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 515 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 516 517 // Branch to slow_path if should_post_on_exceptions_flag was true 518 { BuildCutout unless(this, tst, PROB_MAX); 519 // Do not try anything fancy if we're notifying the VM on every throw. 520 // Cf. case Bytecodes::_athrow in parse2.cpp. 521 uncommon_trap(reason, Deoptimization::Action_none, 522 (ciKlass*)nullptr, (char*)nullptr, must_throw); 523 } 524 525 } 526 527 //------------------------------builtin_throw---------------------------------- 528 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) { 529 bool must_throw = true; 530 531 // If this particular condition has not yet happened at this 532 // bytecode, then use the uncommon trap mechanism, and allow for 533 // a future recompilation if several traps occur here. 534 // If the throw is hot, try to use a more complicated inline mechanism 535 // which keeps execution inside the compiled code. 536 bool treat_throw_as_hot = false; 537 ciMethodData* md = method()->method_data(); 538 539 if (ProfileTraps) { 540 if (too_many_traps(reason)) { 541 treat_throw_as_hot = true; 542 } 543 // (If there is no MDO at all, assume it is early in 544 // execution, and that any deopts are part of the 545 // startup transient, and don't need to be remembered.) 546 547 // Also, if there is a local exception handler, treat all throws 548 // as hot if there has been at least one in this method. 549 if (C->trap_count(reason) != 0 550 && method()->method_data()->trap_count(reason) != 0 551 && has_exception_handler()) { 552 treat_throw_as_hot = true; 553 } 554 } 555 556 // If this throw happens frequently, an uncommon trap might cause 557 // a performance pothole. If there is a local exception handler, 558 // and if this particular bytecode appears to be deoptimizing often, 559 // let us handle the throw inline, with a preconstructed instance. 560 // Note: If the deopt count has blown up, the uncommon trap 561 // runtime is going to flush this nmethod, not matter what. 562 if (treat_throw_as_hot && method()->can_omit_stack_trace()) { 563 // If the throw is local, we use a pre-existing instance and 564 // punt on the backtrace. This would lead to a missing backtrace 565 // (a repeat of 4292742) if the backtrace object is ever asked 566 // for its backtrace. 567 // Fixing this remaining case of 4292742 requires some flavor of 568 // escape analysis. Leave that for the future. 569 ciInstance* ex_obj = nullptr; 570 switch (reason) { 571 case Deoptimization::Reason_null_check: 572 ex_obj = env()->NullPointerException_instance(); 573 break; 574 case Deoptimization::Reason_div0_check: 575 ex_obj = env()->ArithmeticException_instance(); 576 break; 577 case Deoptimization::Reason_range_check: 578 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 579 break; 580 case Deoptimization::Reason_class_check: 581 ex_obj = env()->ClassCastException_instance(); 582 break; 583 case Deoptimization::Reason_array_check: 584 ex_obj = env()->ArrayStoreException_instance(); 585 break; 586 default: 587 break; 588 } 589 if (failing()) { stop(); return; } // exception allocation might fail 590 if (ex_obj != nullptr) { 591 if (env()->jvmti_can_post_on_exceptions()) { 592 // check if we must post exception events, take uncommon trap if so 593 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 594 // here if should_post_on_exceptions is false 595 // continue on with the normal codegen 596 } 597 598 // Cheat with a preallocated exception object. 599 if (C->log() != nullptr) 600 C->log()->elem("hot_throw preallocated='1' reason='%s'", 601 Deoptimization::trap_reason_name(reason)); 602 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 603 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 604 605 // Clear the detail message of the preallocated exception object. 606 // Weblogic sometimes mutates the detail message of exceptions 607 // using reflection. 608 int offset = java_lang_Throwable::get_detailMessage_offset(); 609 const TypePtr* adr_typ = ex_con->add_offset(offset); 610 611 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 612 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 613 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 614 615 if (!method()->has_exception_handlers()) { 616 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway. 617 // This prevents issues with exception handling and late inlining. 618 set_sp(0); 619 clean_stack(0); 620 } 621 622 add_exception_state(make_exception_state(ex_node)); 623 return; 624 } 625 } 626 627 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 628 // It won't be much cheaper than bailing to the interp., since we'll 629 // have to pass up all the debug-info, and the runtime will have to 630 // create the stack trace. 631 632 // Usual case: Bail to interpreter. 633 // Reserve the right to recompile if we haven't seen anything yet. 634 635 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr; 636 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 637 if (treat_throw_as_hot 638 && (method()->method_data()->trap_recompiled_at(bci(), m) 639 || C->too_many_traps(reason))) { 640 // We cannot afford to take more traps here. Suffer in the interpreter. 641 if (C->log() != nullptr) 642 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 643 Deoptimization::trap_reason_name(reason), 644 C->trap_count(reason)); 645 action = Deoptimization::Action_none; 646 } 647 648 // "must_throw" prunes the JVM state to include only the stack, if there 649 // are no local exception handlers. This should cut down on register 650 // allocation time and code size, by drastically reducing the number 651 // of in-edges on the call to the uncommon trap. 652 653 uncommon_trap(reason, action, (ciKlass*)nullptr, (char*)nullptr, must_throw); 654 } 655 656 657 //----------------------------PreserveJVMState--------------------------------- 658 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 659 debug_only(kit->verify_map()); 660 _kit = kit; 661 _map = kit->map(); // preserve the map 662 _sp = kit->sp(); 663 kit->set_map(clone_map ? kit->clone_map() : nullptr); 664 #ifdef ASSERT 665 _bci = kit->bci(); 666 Parse* parser = kit->is_Parse(); 667 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo(); 668 _block = block; 669 #endif 670 } 671 PreserveJVMState::~PreserveJVMState() { 672 GraphKit* kit = _kit; 673 #ifdef ASSERT 674 assert(kit->bci() == _bci, "bci must not shift"); 675 Parse* parser = kit->is_Parse(); 676 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo(); 677 assert(block == _block, "block must not shift"); 678 #endif 679 kit->set_map(_map); 680 kit->set_sp(_sp); 681 } 682 683 684 //-----------------------------BuildCutout------------------------------------- 685 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 686 : PreserveJVMState(kit) 687 { 688 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 689 SafePointNode* outer_map = _map; // preserved map is caller's 690 SafePointNode* inner_map = kit->map(); 691 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 692 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 693 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 694 } 695 BuildCutout::~BuildCutout() { 696 GraphKit* kit = _kit; 697 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 698 } 699 700 //---------------------------PreserveReexecuteState---------------------------- 701 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 702 assert(!kit->stopped(), "must call stopped() before"); 703 _kit = kit; 704 _sp = kit->sp(); 705 _reexecute = kit->jvms()->_reexecute; 706 } 707 PreserveReexecuteState::~PreserveReexecuteState() { 708 if (_kit->stopped()) return; 709 _kit->jvms()->_reexecute = _reexecute; 710 _kit->set_sp(_sp); 711 } 712 713 //------------------------------clone_map-------------------------------------- 714 // Implementation of PreserveJVMState 715 // 716 // Only clone_map(...) here. If this function is only used in the 717 // PreserveJVMState class we may want to get rid of this extra 718 // function eventually and do it all there. 719 720 SafePointNode* GraphKit::clone_map() { 721 if (map() == nullptr) return nullptr; 722 723 // Clone the memory edge first 724 Node* mem = MergeMemNode::make(map()->memory()); 725 gvn().set_type_bottom(mem); 726 727 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 728 JVMState* jvms = this->jvms(); 729 JVMState* clonejvms = jvms->clone_shallow(C); 730 clonemap->set_memory(mem); 731 clonemap->set_jvms(clonejvms); 732 clonejvms->set_map(clonemap); 733 record_for_igvn(clonemap); 734 gvn().set_type_bottom(clonemap); 735 return clonemap; 736 } 737 738 //-----------------------------destruct_map_clone------------------------------ 739 // 740 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need 741 // to destruct/free/delete in the exact opposite order as clone_map(). 742 void GraphKit::destruct_map_clone(SafePointNode* sfp) { 743 if (sfp == nullptr) return; 744 745 Node* mem = sfp->memory(); 746 JVMState* jvms = sfp->jvms(); 747 748 if (jvms != nullptr) { 749 delete jvms; 750 } 751 752 remove_for_igvn(sfp); 753 gvn().clear_type(sfp); 754 sfp->destruct(&_gvn); 755 756 if (mem != nullptr) { 757 gvn().clear_type(mem); 758 mem->destruct(&_gvn); 759 } 760 } 761 762 //-----------------------------set_map_clone----------------------------------- 763 void GraphKit::set_map_clone(SafePointNode* m) { 764 _map = m; 765 _map = clone_map(); 766 _map->set_next_exception(nullptr); 767 debug_only(verify_map()); 768 } 769 770 771 //----------------------------kill_dead_locals--------------------------------- 772 // Detect any locals which are known to be dead, and force them to top. 773 void GraphKit::kill_dead_locals() { 774 // Consult the liveness information for the locals. If any 775 // of them are unused, then they can be replaced by top(). This 776 // should help register allocation time and cut down on the size 777 // of the deoptimization information. 778 779 // This call is made from many of the bytecode handling 780 // subroutines called from the Big Switch in do_one_bytecode. 781 // Every bytecode which might include a slow path is responsible 782 // for killing its dead locals. The more consistent we 783 // are about killing deads, the fewer useless phis will be 784 // constructed for them at various merge points. 785 786 // bci can be -1 (InvocationEntryBci). We return the entry 787 // liveness for the method. 788 789 if (method() == nullptr || method()->code_size() == 0) { 790 // We are building a graph for a call to a native method. 791 // All locals are live. 792 return; 793 } 794 795 ResourceMark rm; 796 797 // Consult the liveness information for the locals. If any 798 // of them are unused, then they can be replaced by top(). This 799 // should help register allocation time and cut down on the size 800 // of the deoptimization information. 801 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 802 803 int len = (int)live_locals.size(); 804 assert(len <= jvms()->loc_size(), "too many live locals"); 805 for (int local = 0; local < len; local++) { 806 if (!live_locals.at(local)) { 807 set_local(local, top()); 808 } 809 } 810 } 811 812 #ifdef ASSERT 813 //-------------------------dead_locals_are_killed------------------------------ 814 // Return true if all dead locals are set to top in the map. 815 // Used to assert "clean" debug info at various points. 816 bool GraphKit::dead_locals_are_killed() { 817 if (method() == nullptr || method()->code_size() == 0) { 818 // No locals need to be dead, so all is as it should be. 819 return true; 820 } 821 822 // Make sure somebody called kill_dead_locals upstream. 823 ResourceMark rm; 824 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) { 825 if (jvms->loc_size() == 0) continue; // no locals to consult 826 SafePointNode* map = jvms->map(); 827 ciMethod* method = jvms->method(); 828 int bci = jvms->bci(); 829 if (jvms == this->jvms()) { 830 bci = this->bci(); // it might not yet be synched 831 } 832 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 833 int len = (int)live_locals.size(); 834 if (!live_locals.is_valid() || len == 0) 835 // This method is trivial, or is poisoned by a breakpoint. 836 return true; 837 assert(len == jvms->loc_size(), "live map consistent with locals map"); 838 for (int local = 0; local < len; local++) { 839 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 840 if (PrintMiscellaneous && (Verbose || WizardMode)) { 841 tty->print_cr("Zombie local %d: ", local); 842 jvms->dump(); 843 } 844 return false; 845 } 846 } 847 } 848 return true; 849 } 850 851 #endif //ASSERT 852 853 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens. 854 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 855 ciMethod* cur_method = jvms->method(); 856 int cur_bci = jvms->bci(); 857 if (cur_method != nullptr && cur_bci != InvocationEntryBci) { 858 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 859 return Interpreter::bytecode_should_reexecute(code) || 860 (is_anewarray && code == Bytecodes::_multianewarray); 861 // Reexecute _multianewarray bytecode which was replaced with 862 // sequence of [a]newarray. See Parse::do_multianewarray(). 863 // 864 // Note: interpreter should not have it set since this optimization 865 // is limited by dimensions and guarded by flag so in some cases 866 // multianewarray() runtime calls will be generated and 867 // the bytecode should not be reexecutes (stack will not be reset). 868 } else { 869 return false; 870 } 871 } 872 873 // Helper function for adding JVMState and debug information to node 874 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 875 // Add the safepoint edges to the call (or other safepoint). 876 877 // Make sure dead locals are set to top. This 878 // should help register allocation time and cut down on the size 879 // of the deoptimization information. 880 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 881 882 // Walk the inline list to fill in the correct set of JVMState's 883 // Also fill in the associated edges for each JVMState. 884 885 // If the bytecode needs to be reexecuted we need to put 886 // the arguments back on the stack. 887 const bool should_reexecute = jvms()->should_reexecute(); 888 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 889 890 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 891 // undefined if the bci is different. This is normal for Parse but it 892 // should not happen for LibraryCallKit because only one bci is processed. 893 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 894 "in LibraryCallKit the reexecute bit should not change"); 895 896 // If we are guaranteed to throw, we can prune everything but the 897 // input to the current bytecode. 898 bool can_prune_locals = false; 899 uint stack_slots_not_pruned = 0; 900 int inputs = 0, depth = 0; 901 if (must_throw) { 902 assert(method() == youngest_jvms->method(), "sanity"); 903 if (compute_stack_effects(inputs, depth)) { 904 can_prune_locals = true; 905 stack_slots_not_pruned = inputs; 906 } 907 } 908 909 if (env()->should_retain_local_variables()) { 910 // At any safepoint, this method can get breakpointed, which would 911 // then require an immediate deoptimization. 912 can_prune_locals = false; // do not prune locals 913 stack_slots_not_pruned = 0; 914 } 915 916 // do not scribble on the input jvms 917 JVMState* out_jvms = youngest_jvms->clone_deep(C); 918 call->set_jvms(out_jvms); // Start jvms list for call node 919 920 // For a known set of bytecodes, the interpreter should reexecute them if 921 // deoptimization happens. We set the reexecute state for them here 922 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 923 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 924 #ifdef ASSERT 925 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects() 926 assert(method() == youngest_jvms->method(), "sanity"); 927 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc())); 928 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution"); 929 #endif // ASSERT 930 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 931 } 932 933 // Presize the call: 934 DEBUG_ONLY(uint non_debug_edges = call->req()); 935 call->add_req_batch(top(), youngest_jvms->debug_depth()); 936 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 937 938 // Set up edges so that the call looks like this: 939 // Call [state:] ctl io mem fptr retadr 940 // [parms:] parm0 ... parmN 941 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 942 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 943 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 944 // Note that caller debug info precedes callee debug info. 945 946 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 947 uint debug_ptr = call->req(); 948 949 // Loop over the map input edges associated with jvms, add them 950 // to the call node, & reset all offsets to match call node array. 951 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) { 952 uint debug_end = debug_ptr; 953 uint debug_start = debug_ptr - in_jvms->debug_size(); 954 debug_ptr = debug_start; // back up the ptr 955 956 uint p = debug_start; // walks forward in [debug_start, debug_end) 957 uint j, k, l; 958 SafePointNode* in_map = in_jvms->map(); 959 out_jvms->set_map(call); 960 961 if (can_prune_locals) { 962 assert(in_jvms->method() == out_jvms->method(), "sanity"); 963 // If the current throw can reach an exception handler in this JVMS, 964 // then we must keep everything live that can reach that handler. 965 // As a quick and dirty approximation, we look for any handlers at all. 966 if (in_jvms->method()->has_exception_handlers()) { 967 can_prune_locals = false; 968 } 969 } 970 971 // Add the Locals 972 k = in_jvms->locoff(); 973 l = in_jvms->loc_size(); 974 out_jvms->set_locoff(p); 975 if (!can_prune_locals) { 976 for (j = 0; j < l; j++) 977 call->set_req(p++, in_map->in(k+j)); 978 } else { 979 p += l; // already set to top above by add_req_batch 980 } 981 982 // Add the Expression Stack 983 k = in_jvms->stkoff(); 984 l = in_jvms->sp(); 985 out_jvms->set_stkoff(p); 986 if (!can_prune_locals) { 987 for (j = 0; j < l; j++) 988 call->set_req(p++, in_map->in(k+j)); 989 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 990 // Divide stack into {S0,...,S1}, where S0 is set to top. 991 uint s1 = stack_slots_not_pruned; 992 stack_slots_not_pruned = 0; // for next iteration 993 if (s1 > l) s1 = l; 994 uint s0 = l - s1; 995 p += s0; // skip the tops preinstalled by add_req_batch 996 for (j = s0; j < l; j++) 997 call->set_req(p++, in_map->in(k+j)); 998 } else { 999 p += l; // already set to top above by add_req_batch 1000 } 1001 1002 // Add the Monitors 1003 k = in_jvms->monoff(); 1004 l = in_jvms->mon_size(); 1005 out_jvms->set_monoff(p); 1006 for (j = 0; j < l; j++) 1007 call->set_req(p++, in_map->in(k+j)); 1008 1009 // Copy any scalar object fields. 1010 k = in_jvms->scloff(); 1011 l = in_jvms->scl_size(); 1012 out_jvms->set_scloff(p); 1013 for (j = 0; j < l; j++) 1014 call->set_req(p++, in_map->in(k+j)); 1015 1016 // Finish the new jvms. 1017 out_jvms->set_endoff(p); 1018 1019 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 1020 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 1021 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 1022 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 1023 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 1024 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 1025 1026 // Update the two tail pointers in parallel. 1027 out_jvms = out_jvms->caller(); 1028 in_jvms = in_jvms->caller(); 1029 } 1030 1031 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1032 1033 // Test the correctness of JVMState::debug_xxx accessors: 1034 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1035 assert(call->jvms()->debug_end() == call->req(), ""); 1036 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1037 } 1038 1039 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1040 Bytecodes::Code code = java_bc(); 1041 if (code == Bytecodes::_wide) { 1042 code = method()->java_code_at_bci(bci() + 1); 1043 } 1044 1045 if (code != Bytecodes::_illegal) { 1046 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1047 } 1048 1049 auto rsize = [&]() { 1050 assert(code != Bytecodes::_illegal, "code is illegal!"); 1051 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1052 return (rtype < T_CONFLICT) ? type2size[rtype] : 0; 1053 }; 1054 1055 switch (code) { 1056 case Bytecodes::_illegal: 1057 return false; 1058 1059 case Bytecodes::_ldc: 1060 case Bytecodes::_ldc_w: 1061 case Bytecodes::_ldc2_w: 1062 inputs = 0; 1063 break; 1064 1065 case Bytecodes::_dup: inputs = 1; break; 1066 case Bytecodes::_dup_x1: inputs = 2; break; 1067 case Bytecodes::_dup_x2: inputs = 3; break; 1068 case Bytecodes::_dup2: inputs = 2; break; 1069 case Bytecodes::_dup2_x1: inputs = 3; break; 1070 case Bytecodes::_dup2_x2: inputs = 4; break; 1071 case Bytecodes::_swap: inputs = 2; break; 1072 case Bytecodes::_arraylength: inputs = 1; break; 1073 1074 case Bytecodes::_getstatic: 1075 case Bytecodes::_putstatic: 1076 case Bytecodes::_getfield: 1077 case Bytecodes::_putfield: 1078 { 1079 bool ignored_will_link; 1080 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1081 int size = field->type()->size(); 1082 bool is_get = (depth >= 0), is_static = (depth & 1); 1083 inputs = (is_static ? 0 : 1); 1084 if (is_get) { 1085 depth = size - inputs; 1086 } else { 1087 inputs += size; // putxxx pops the value from the stack 1088 depth = - inputs; 1089 } 1090 } 1091 break; 1092 1093 case Bytecodes::_invokevirtual: 1094 case Bytecodes::_invokespecial: 1095 case Bytecodes::_invokestatic: 1096 case Bytecodes::_invokedynamic: 1097 case Bytecodes::_invokeinterface: 1098 { 1099 bool ignored_will_link; 1100 ciSignature* declared_signature = nullptr; 1101 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1102 assert(declared_signature != nullptr, "cannot be null"); 1103 inputs = declared_signature->arg_size_for_bc(code); 1104 int size = declared_signature->return_type()->size(); 1105 depth = size - inputs; 1106 } 1107 break; 1108 1109 case Bytecodes::_multianewarray: 1110 { 1111 ciBytecodeStream iter(method()); 1112 iter.reset_to_bci(bci()); 1113 iter.next(); 1114 inputs = iter.get_dimensions(); 1115 assert(rsize() == 1, ""); 1116 depth = 1 - inputs; 1117 } 1118 break; 1119 1120 case Bytecodes::_ireturn: 1121 case Bytecodes::_lreturn: 1122 case Bytecodes::_freturn: 1123 case Bytecodes::_dreturn: 1124 case Bytecodes::_areturn: 1125 assert(rsize() == -depth, ""); 1126 inputs = -depth; 1127 break; 1128 1129 case Bytecodes::_jsr: 1130 case Bytecodes::_jsr_w: 1131 inputs = 0; 1132 depth = 1; // S.B. depth=1, not zero 1133 break; 1134 1135 default: 1136 // bytecode produces a typed result 1137 inputs = rsize() - depth; 1138 assert(inputs >= 0, ""); 1139 break; 1140 } 1141 1142 #ifdef ASSERT 1143 // spot check 1144 int outputs = depth + inputs; 1145 assert(outputs >= 0, "sanity"); 1146 switch (code) { 1147 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1148 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1149 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1150 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1151 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1152 default: break; 1153 } 1154 #endif //ASSERT 1155 1156 return true; 1157 } 1158 1159 1160 1161 //------------------------------basic_plus_adr--------------------------------- 1162 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1163 // short-circuit a common case 1164 if (offset == intcon(0)) return ptr; 1165 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1166 } 1167 1168 Node* GraphKit::ConvI2L(Node* offset) { 1169 // short-circuit a common case 1170 jint offset_con = find_int_con(offset, Type::OffsetBot); 1171 if (offset_con != Type::OffsetBot) { 1172 return longcon((jlong) offset_con); 1173 } 1174 return _gvn.transform( new ConvI2LNode(offset)); 1175 } 1176 1177 Node* GraphKit::ConvI2UL(Node* offset) { 1178 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1179 if (offset_con != (juint) Type::OffsetBot) { 1180 return longcon((julong) offset_con); 1181 } 1182 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1183 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1184 return _gvn.transform( new AndLNode(conv, mask) ); 1185 } 1186 1187 Node* GraphKit::ConvL2I(Node* offset) { 1188 // short-circuit a common case 1189 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1190 if (offset_con != (jlong)Type::OffsetBot) { 1191 return intcon((int) offset_con); 1192 } 1193 return _gvn.transform( new ConvL2INode(offset)); 1194 } 1195 1196 //-------------------------load_object_klass----------------------------------- 1197 Node* GraphKit::load_object_klass(Node* obj) { 1198 // Special-case a fresh allocation to avoid building nodes: 1199 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1200 if (akls != nullptr) return akls; 1201 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1202 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1203 } 1204 1205 //-------------------------load_array_length----------------------------------- 1206 Node* GraphKit::load_array_length(Node* array) { 1207 // Special-case a fresh allocation to avoid building nodes: 1208 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array); 1209 Node *alen; 1210 if (alloc == nullptr) { 1211 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1212 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1213 } else { 1214 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false); 1215 } 1216 return alen; 1217 } 1218 1219 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc, 1220 const TypeOopPtr* oop_type, 1221 bool replace_length_in_map) { 1222 Node* length = alloc->Ideal_length(); 1223 if (replace_length_in_map == false || map()->find_edge(length) >= 0) { 1224 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn); 1225 if (ccast != length) { 1226 // do not transform ccast here, it might convert to top node for 1227 // negative array length and break assumptions in parsing stage. 1228 _gvn.set_type_bottom(ccast); 1229 record_for_igvn(ccast); 1230 if (replace_length_in_map) { 1231 replace_in_map(length, ccast); 1232 } 1233 return ccast; 1234 } 1235 } 1236 return length; 1237 } 1238 1239 //------------------------------do_null_check---------------------------------- 1240 // Helper function to do a null pointer check. Returned value is 1241 // the incoming address with null casted away. You are allowed to use the 1242 // not-null value only if you are control dependent on the test. 1243 #ifndef PRODUCT 1244 extern uint explicit_null_checks_inserted, 1245 explicit_null_checks_elided; 1246 #endif 1247 Node* GraphKit::null_check_common(Node* value, BasicType type, 1248 // optional arguments for variations: 1249 bool assert_null, 1250 Node* *null_control, 1251 bool speculative) { 1252 assert(!assert_null || null_control == nullptr, "not both at once"); 1253 if (stopped()) return top(); 1254 NOT_PRODUCT(explicit_null_checks_inserted++); 1255 1256 // Construct null check 1257 Node *chk = nullptr; 1258 switch(type) { 1259 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1260 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1261 case T_ARRAY : // fall through 1262 type = T_OBJECT; // simplify further tests 1263 case T_OBJECT : { 1264 const Type *t = _gvn.type( value ); 1265 1266 const TypeOopPtr* tp = t->isa_oopptr(); 1267 if (tp != nullptr && !tp->is_loaded() 1268 // Only for do_null_check, not any of its siblings: 1269 && !assert_null && null_control == nullptr) { 1270 // Usually, any field access or invocation on an unloaded oop type 1271 // will simply fail to link, since the statically linked class is 1272 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1273 // the static class is loaded but the sharper oop type is not. 1274 // Rather than checking for this obscure case in lots of places, 1275 // we simply observe that a null check on an unloaded class 1276 // will always be followed by a nonsense operation, so we 1277 // can just issue the uncommon trap here. 1278 // Our access to the unloaded class will only be correct 1279 // after it has been loaded and initialized, which requires 1280 // a trip through the interpreter. 1281 ciKlass* klass = tp->unloaded_klass(); 1282 #ifndef PRODUCT 1283 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); } 1284 #endif 1285 uncommon_trap(Deoptimization::Reason_unloaded, 1286 Deoptimization::Action_reinterpret, 1287 klass, "!loaded"); 1288 return top(); 1289 } 1290 1291 if (assert_null) { 1292 // See if the type is contained in NULL_PTR. 1293 // If so, then the value is already null. 1294 if (t->higher_equal(TypePtr::NULL_PTR)) { 1295 NOT_PRODUCT(explicit_null_checks_elided++); 1296 return value; // Elided null assert quickly! 1297 } 1298 } else { 1299 // See if mixing in the null pointer changes type. 1300 // If so, then the null pointer was not allowed in the original 1301 // type. In other words, "value" was not-null. 1302 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1303 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1304 NOT_PRODUCT(explicit_null_checks_elided++); 1305 return value; // Elided null check quickly! 1306 } 1307 } 1308 chk = new CmpPNode( value, null() ); 1309 break; 1310 } 1311 1312 default: 1313 fatal("unexpected type: %s", type2name(type)); 1314 } 1315 assert(chk != nullptr, "sanity check"); 1316 chk = _gvn.transform(chk); 1317 1318 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1319 BoolNode *btst = new BoolNode( chk, btest); 1320 Node *tst = _gvn.transform( btst ); 1321 1322 //----------- 1323 // if peephole optimizations occurred, a prior test existed. 1324 // If a prior test existed, maybe it dominates as we can avoid this test. 1325 if (tst != btst && type == T_OBJECT) { 1326 // At this point we want to scan up the CFG to see if we can 1327 // find an identical test (and so avoid this test altogether). 1328 Node *cfg = control(); 1329 int depth = 0; 1330 while( depth < 16 ) { // Limit search depth for speed 1331 if( cfg->Opcode() == Op_IfTrue && 1332 cfg->in(0)->in(1) == tst ) { 1333 // Found prior test. Use "cast_not_null" to construct an identical 1334 // CastPP (and hence hash to) as already exists for the prior test. 1335 // Return that casted value. 1336 if (assert_null) { 1337 replace_in_map(value, null()); 1338 return null(); // do not issue the redundant test 1339 } 1340 Node *oldcontrol = control(); 1341 set_control(cfg); 1342 Node *res = cast_not_null(value); 1343 set_control(oldcontrol); 1344 NOT_PRODUCT(explicit_null_checks_elided++); 1345 return res; 1346 } 1347 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1348 if (cfg == nullptr) break; // Quit at region nodes 1349 depth++; 1350 } 1351 } 1352 1353 //----------- 1354 // Branch to failure if null 1355 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1356 Deoptimization::DeoptReason reason; 1357 if (assert_null) { 1358 reason = Deoptimization::reason_null_assert(speculative); 1359 } else if (type == T_OBJECT) { 1360 reason = Deoptimization::reason_null_check(speculative); 1361 } else { 1362 reason = Deoptimization::Reason_div0_check; 1363 } 1364 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1365 // ciMethodData::has_trap_at will return a conservative -1 if any 1366 // must-be-null assertion has failed. This could cause performance 1367 // problems for a method after its first do_null_assert failure. 1368 // Consider using 'Reason_class_check' instead? 1369 1370 // To cause an implicit null check, we set the not-null probability 1371 // to the maximum (PROB_MAX). For an explicit check the probability 1372 // is set to a smaller value. 1373 if (null_control != nullptr || too_many_traps(reason)) { 1374 // probability is less likely 1375 ok_prob = PROB_LIKELY_MAG(3); 1376 } else if (!assert_null && 1377 (ImplicitNullCheckThreshold > 0) && 1378 method() != nullptr && 1379 (method()->method_data()->trap_count(reason) 1380 >= (uint)ImplicitNullCheckThreshold)) { 1381 ok_prob = PROB_LIKELY_MAG(3); 1382 } 1383 1384 if (null_control != nullptr) { 1385 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1386 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1387 set_control( _gvn.transform( new IfTrueNode(iff))); 1388 #ifndef PRODUCT 1389 if (null_true == top()) { 1390 explicit_null_checks_elided++; 1391 } 1392 #endif 1393 (*null_control) = null_true; 1394 } else { 1395 BuildCutout unless(this, tst, ok_prob); 1396 // Check for optimizer eliding test at parse time 1397 if (stopped()) { 1398 // Failure not possible; do not bother making uncommon trap. 1399 NOT_PRODUCT(explicit_null_checks_elided++); 1400 } else if (assert_null) { 1401 uncommon_trap(reason, 1402 Deoptimization::Action_make_not_entrant, 1403 nullptr, "assert_null"); 1404 } else { 1405 replace_in_map(value, zerocon(type)); 1406 builtin_throw(reason); 1407 } 1408 } 1409 1410 // Must throw exception, fall-thru not possible? 1411 if (stopped()) { 1412 return top(); // No result 1413 } 1414 1415 if (assert_null) { 1416 // Cast obj to null on this path. 1417 replace_in_map(value, zerocon(type)); 1418 return zerocon(type); 1419 } 1420 1421 // Cast obj to not-null on this path, if there is no null_control. 1422 // (If there is a null_control, a non-null value may come back to haunt us.) 1423 if (type == T_OBJECT) { 1424 Node* cast = cast_not_null(value, false); 1425 if (null_control == nullptr || (*null_control) == top()) 1426 replace_in_map(value, cast); 1427 value = cast; 1428 } 1429 1430 return value; 1431 } 1432 1433 1434 //------------------------------cast_not_null---------------------------------- 1435 // Cast obj to not-null on this path 1436 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1437 const Type *t = _gvn.type(obj); 1438 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1439 // Object is already not-null? 1440 if( t == t_not_null ) return obj; 1441 1442 Node* cast = new CastPPNode(control(), obj,t_not_null); 1443 cast = _gvn.transform( cast ); 1444 1445 // Scan for instances of 'obj' in the current JVM mapping. 1446 // These instances are known to be not-null after the test. 1447 if (do_replace_in_map) 1448 replace_in_map(obj, cast); 1449 1450 return cast; // Return casted value 1451 } 1452 1453 // Sometimes in intrinsics, we implicitly know an object is not null 1454 // (there's no actual null check) so we can cast it to not null. In 1455 // the course of optimizations, the input to the cast can become null. 1456 // In that case that data path will die and we need the control path 1457 // to become dead as well to keep the graph consistent. So we have to 1458 // add a check for null for which one branch can't be taken. It uses 1459 // an Opaque4 node that will cause the check to be removed after loop 1460 // opts so the test goes away and the compiled code doesn't execute a 1461 // useless check. 1462 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1463 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1464 return value; 1465 } 1466 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1467 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1468 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1469 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1470 _gvn.set_type(iff, iff->Value(&_gvn)); 1471 if (!tst->is_Con()) { 1472 record_for_igvn(iff); 1473 } 1474 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1475 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1476 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1477 C->root()->add_req(halt); 1478 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1479 set_control(if_t); 1480 return cast_not_null(value, do_replace_in_map); 1481 } 1482 1483 1484 //--------------------------replace_in_map------------------------------------- 1485 void GraphKit::replace_in_map(Node* old, Node* neww) { 1486 if (old == neww) { 1487 return; 1488 } 1489 1490 map()->replace_edge(old, neww); 1491 1492 // Note: This operation potentially replaces any edge 1493 // on the map. This includes locals, stack, and monitors 1494 // of the current (innermost) JVM state. 1495 1496 // don't let inconsistent types from profiling escape this 1497 // method 1498 1499 const Type* told = _gvn.type(old); 1500 const Type* tnew = _gvn.type(neww); 1501 1502 if (!tnew->higher_equal(told)) { 1503 return; 1504 } 1505 1506 map()->record_replaced_node(old, neww); 1507 } 1508 1509 1510 //============================================================================= 1511 //--------------------------------memory--------------------------------------- 1512 Node* GraphKit::memory(uint alias_idx) { 1513 MergeMemNode* mem = merged_memory(); 1514 Node* p = mem->memory_at(alias_idx); 1515 assert(p != mem->empty_memory(), "empty"); 1516 _gvn.set_type(p, Type::MEMORY); // must be mapped 1517 return p; 1518 } 1519 1520 //-----------------------------reset_memory------------------------------------ 1521 Node* GraphKit::reset_memory() { 1522 Node* mem = map()->memory(); 1523 // do not use this node for any more parsing! 1524 debug_only( map()->set_memory((Node*)nullptr) ); 1525 return _gvn.transform( mem ); 1526 } 1527 1528 //------------------------------set_all_memory--------------------------------- 1529 void GraphKit::set_all_memory(Node* newmem) { 1530 Node* mergemem = MergeMemNode::make(newmem); 1531 gvn().set_type_bottom(mergemem); 1532 map()->set_memory(mergemem); 1533 } 1534 1535 //------------------------------set_all_memory_call---------------------------- 1536 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1537 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1538 set_all_memory(newmem); 1539 } 1540 1541 //============================================================================= 1542 // 1543 // parser factory methods for MemNodes 1544 // 1545 // These are layered on top of the factory methods in LoadNode and StoreNode, 1546 // and integrate with the parser's memory state and _gvn engine. 1547 // 1548 1549 // factory methods in "int adr_idx" 1550 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1551 int adr_idx, 1552 MemNode::MemOrd mo, 1553 LoadNode::ControlDependency control_dependency, 1554 bool require_atomic_access, 1555 bool unaligned, 1556 bool mismatched, 1557 bool unsafe, 1558 uint8_t barrier_data) { 1559 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1560 const TypePtr* adr_type = nullptr; // debug-mode-only argument 1561 debug_only(adr_type = C->get_adr_type(adr_idx)); 1562 Node* mem = memory(adr_idx); 1563 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data); 1564 ld = _gvn.transform(ld); 1565 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1566 // Improve graph before escape analysis and boxing elimination. 1567 record_for_igvn(ld); 1568 if (ld->is_DecodeN()) { 1569 // Also record the actual load (LoadN) in case ld is DecodeN. In some 1570 // rare corner cases, ld->in(1) can be something other than LoadN (e.g., 1571 // a Phi). Recording such cases is still perfectly sound, but may be 1572 // unnecessary and result in some minor IGVN overhead. 1573 record_for_igvn(ld->in(1)); 1574 } 1575 } 1576 return ld; 1577 } 1578 1579 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1580 int adr_idx, 1581 MemNode::MemOrd mo, 1582 bool require_atomic_access, 1583 bool unaligned, 1584 bool mismatched, 1585 bool unsafe, 1586 int barrier_data) { 1587 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1588 const TypePtr* adr_type = nullptr; 1589 debug_only(adr_type = C->get_adr_type(adr_idx)); 1590 Node *mem = memory(adr_idx); 1591 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access); 1592 if (unaligned) { 1593 st->as_Store()->set_unaligned_access(); 1594 } 1595 if (mismatched) { 1596 st->as_Store()->set_mismatched_access(); 1597 } 1598 if (unsafe) { 1599 st->as_Store()->set_unsafe_access(); 1600 } 1601 st->as_Store()->set_barrier_data(barrier_data); 1602 st = _gvn.transform(st); 1603 set_memory(st, adr_idx); 1604 // Back-to-back stores can only remove intermediate store with DU info 1605 // so push on worklist for optimizer. 1606 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1607 record_for_igvn(st); 1608 1609 return st; 1610 } 1611 1612 Node* GraphKit::access_store_at(Node* obj, 1613 Node* adr, 1614 const TypePtr* adr_type, 1615 Node* val, 1616 const Type* val_type, 1617 BasicType bt, 1618 DecoratorSet decorators) { 1619 // Transformation of a value which could be null pointer (CastPP #null) 1620 // could be delayed during Parse (for example, in adjust_map_after_if()). 1621 // Execute transformation here to avoid barrier generation in such case. 1622 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1623 val = _gvn.makecon(TypePtr::NULL_PTR); 1624 } 1625 1626 if (stopped()) { 1627 return top(); // Dead path ? 1628 } 1629 1630 assert(val != nullptr, "not dead path"); 1631 1632 C2AccessValuePtr addr(adr, adr_type); 1633 C2AccessValue value(val, val_type); 1634 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1635 if (access.is_raw()) { 1636 return _barrier_set->BarrierSetC2::store_at(access, value); 1637 } else { 1638 return _barrier_set->store_at(access, value); 1639 } 1640 } 1641 1642 Node* GraphKit::access_load_at(Node* obj, // containing obj 1643 Node* adr, // actual address to store val at 1644 const TypePtr* adr_type, 1645 const Type* val_type, 1646 BasicType bt, 1647 DecoratorSet decorators) { 1648 if (stopped()) { 1649 return top(); // Dead path ? 1650 } 1651 1652 C2AccessValuePtr addr(adr, adr_type); 1653 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr); 1654 if (access.is_raw()) { 1655 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1656 } else { 1657 return _barrier_set->load_at(access, val_type); 1658 } 1659 } 1660 1661 Node* GraphKit::access_load(Node* adr, // actual address to load val at 1662 const Type* val_type, 1663 BasicType bt, 1664 DecoratorSet decorators) { 1665 if (stopped()) { 1666 return top(); // Dead path ? 1667 } 1668 1669 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr()); 1670 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr); 1671 if (access.is_raw()) { 1672 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1673 } else { 1674 return _barrier_set->load_at(access, val_type); 1675 } 1676 } 1677 1678 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1679 Node* adr, 1680 const TypePtr* adr_type, 1681 int alias_idx, 1682 Node* expected_val, 1683 Node* new_val, 1684 const Type* value_type, 1685 BasicType bt, 1686 DecoratorSet decorators) { 1687 C2AccessValuePtr addr(adr, adr_type); 1688 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1689 bt, obj, addr, alias_idx); 1690 if (access.is_raw()) { 1691 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1692 } else { 1693 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1694 } 1695 } 1696 1697 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1698 Node* adr, 1699 const TypePtr* adr_type, 1700 int alias_idx, 1701 Node* expected_val, 1702 Node* new_val, 1703 const Type* value_type, 1704 BasicType bt, 1705 DecoratorSet decorators) { 1706 C2AccessValuePtr addr(adr, adr_type); 1707 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1708 bt, obj, addr, alias_idx); 1709 if (access.is_raw()) { 1710 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1711 } else { 1712 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1713 } 1714 } 1715 1716 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1717 Node* adr, 1718 const TypePtr* adr_type, 1719 int alias_idx, 1720 Node* new_val, 1721 const Type* value_type, 1722 BasicType bt, 1723 DecoratorSet decorators) { 1724 C2AccessValuePtr addr(adr, adr_type); 1725 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1726 bt, obj, addr, alias_idx); 1727 if (access.is_raw()) { 1728 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1729 } else { 1730 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1731 } 1732 } 1733 1734 Node* GraphKit::access_atomic_add_at(Node* obj, 1735 Node* adr, 1736 const TypePtr* adr_type, 1737 int alias_idx, 1738 Node* new_val, 1739 const Type* value_type, 1740 BasicType bt, 1741 DecoratorSet decorators) { 1742 C2AccessValuePtr addr(adr, adr_type); 1743 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1744 if (access.is_raw()) { 1745 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1746 } else { 1747 return _barrier_set->atomic_add_at(access, new_val, value_type); 1748 } 1749 } 1750 1751 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) { 1752 return _barrier_set->clone(this, src, dst, size, is_array); 1753 } 1754 1755 //-------------------------array_element_address------------------------- 1756 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1757 const TypeInt* sizetype, Node* ctrl) { 1758 uint shift = exact_log2(type2aelembytes(elembt)); 1759 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1760 1761 // short-circuit a common case (saves lots of confusing waste motion) 1762 jint idx_con = find_int_con(idx, -1); 1763 if (idx_con >= 0) { 1764 intptr_t offset = header + ((intptr_t)idx_con << shift); 1765 return basic_plus_adr(ary, offset); 1766 } 1767 1768 // must be correct type for alignment purposes 1769 Node* base = basic_plus_adr(ary, header); 1770 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1771 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1772 return basic_plus_adr(ary, base, scale); 1773 } 1774 1775 //-------------------------load_array_element------------------------- 1776 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) { 1777 const Type* elemtype = arytype->elem(); 1778 BasicType elembt = elemtype->array_element_basic_type(); 1779 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1780 if (elembt == T_NARROWOOP) { 1781 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1782 } 1783 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt, 1784 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0)); 1785 return ld; 1786 } 1787 1788 //-------------------------set_arguments_for_java_call------------------------- 1789 // Arguments (pre-popped from the stack) are taken from the JVMS. 1790 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1791 // Add the call arguments: 1792 uint nargs = call->method()->arg_size(); 1793 for (uint i = 0; i < nargs; i++) { 1794 Node* arg = argument(i); 1795 call->init_req(i + TypeFunc::Parms, arg); 1796 } 1797 } 1798 1799 //---------------------------set_edges_for_java_call--------------------------- 1800 // Connect a newly created call into the current JVMS. 1801 // A return value node (if any) is returned from set_edges_for_java_call. 1802 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1803 1804 // Add the predefined inputs: 1805 call->init_req( TypeFunc::Control, control() ); 1806 call->init_req( TypeFunc::I_O , i_o() ); 1807 call->init_req( TypeFunc::Memory , reset_memory() ); 1808 call->init_req( TypeFunc::FramePtr, frameptr() ); 1809 call->init_req( TypeFunc::ReturnAdr, top() ); 1810 1811 add_safepoint_edges(call, must_throw); 1812 1813 Node* xcall = _gvn.transform(call); 1814 1815 if (xcall == top()) { 1816 set_control(top()); 1817 return; 1818 } 1819 assert(xcall == call, "call identity is stable"); 1820 1821 // Re-use the current map to produce the result. 1822 1823 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1824 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1825 set_all_memory_call(xcall, separate_io_proj); 1826 1827 //return xcall; // no need, caller already has it 1828 } 1829 1830 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1831 if (stopped()) return top(); // maybe the call folded up? 1832 1833 // Capture the return value, if any. 1834 Node* ret; 1835 if (call->method() == nullptr || 1836 call->method()->return_type()->basic_type() == T_VOID) 1837 ret = top(); 1838 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1839 1840 // Note: Since any out-of-line call can produce an exception, 1841 // we always insert an I_O projection from the call into the result. 1842 1843 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1844 1845 if (separate_io_proj) { 1846 // The caller requested separate projections be used by the fall 1847 // through and exceptional paths, so replace the projections for 1848 // the fall through path. 1849 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1850 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1851 } 1852 return ret; 1853 } 1854 1855 //--------------------set_predefined_input_for_runtime_call-------------------- 1856 // Reading and setting the memory state is way conservative here. 1857 // The real problem is that I am not doing real Type analysis on memory, 1858 // so I cannot distinguish card mark stores from other stores. Across a GC 1859 // point the Store Barrier and the card mark memory has to agree. I cannot 1860 // have a card mark store and its barrier split across the GC point from 1861 // either above or below. Here I get that to happen by reading ALL of memory. 1862 // A better answer would be to separate out card marks from other memory. 1863 // For now, return the input memory state, so that it can be reused 1864 // after the call, if this call has restricted memory effects. 1865 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1866 // Set fixed predefined input arguments 1867 Node* memory = reset_memory(); 1868 Node* m = narrow_mem == nullptr ? memory : narrow_mem; 1869 call->init_req( TypeFunc::Control, control() ); 1870 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1871 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1872 call->init_req( TypeFunc::FramePtr, frameptr() ); 1873 call->init_req( TypeFunc::ReturnAdr, top() ); 1874 return memory; 1875 } 1876 1877 //-------------------set_predefined_output_for_runtime_call-------------------- 1878 // Set control and memory (not i_o) from the call. 1879 // If keep_mem is not null, use it for the output state, 1880 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1881 // If hook_mem is null, this call produces no memory effects at all. 1882 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1883 // then only that memory slice is taken from the call. 1884 // In the last case, we must put an appropriate memory barrier before 1885 // the call, so as to create the correct anti-dependencies on loads 1886 // preceding the call. 1887 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1888 Node* keep_mem, 1889 const TypePtr* hook_mem) { 1890 // no i/o 1891 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1892 if (keep_mem) { 1893 // First clone the existing memory state 1894 set_all_memory(keep_mem); 1895 if (hook_mem != nullptr) { 1896 // Make memory for the call 1897 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1898 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1899 // We also use hook_mem to extract specific effects from arraycopy stubs. 1900 set_memory(mem, hook_mem); 1901 } 1902 // ...else the call has NO memory effects. 1903 1904 // Make sure the call advertises its memory effects precisely. 1905 // This lets us build accurate anti-dependences in gcm.cpp. 1906 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1907 "call node must be constructed correctly"); 1908 } else { 1909 assert(hook_mem == nullptr, ""); 1910 // This is not a "slow path" call; all memory comes from the call. 1911 set_all_memory_call(call); 1912 } 1913 } 1914 1915 // Keep track of MergeMems feeding into other MergeMems 1916 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 1917 if (!mem->is_MergeMem()) { 1918 return; 1919 } 1920 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 1921 Node* use = i.get(); 1922 if (use->is_MergeMem()) { 1923 wl.push(use); 1924 } 1925 } 1926 } 1927 1928 // Replace the call with the current state of the kit. 1929 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1930 JVMState* ejvms = nullptr; 1931 if (has_exceptions()) { 1932 ejvms = transfer_exceptions_into_jvms(); 1933 } 1934 1935 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1936 ReplacedNodes replaced_nodes_exception; 1937 Node* ex_ctl = top(); 1938 1939 SafePointNode* final_state = stop(); 1940 1941 // Find all the needed outputs of this call 1942 CallProjections callprojs; 1943 call->extract_projections(&callprojs, true); 1944 1945 Unique_Node_List wl; 1946 Node* init_mem = call->in(TypeFunc::Memory); 1947 Node* final_mem = final_state->in(TypeFunc::Memory); 1948 Node* final_ctl = final_state->in(TypeFunc::Control); 1949 Node* final_io = final_state->in(TypeFunc::I_O); 1950 1951 // Replace all the old call edges with the edges from the inlining result 1952 if (callprojs.fallthrough_catchproj != nullptr) { 1953 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1954 } 1955 if (callprojs.fallthrough_memproj != nullptr) { 1956 if (final_mem->is_MergeMem()) { 1957 // Parser's exits MergeMem was not transformed but may be optimized 1958 final_mem = _gvn.transform(final_mem); 1959 } 1960 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1961 add_mergemem_users_to_worklist(wl, final_mem); 1962 } 1963 if (callprojs.fallthrough_ioproj != nullptr) { 1964 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1965 } 1966 1967 // Replace the result with the new result if it exists and is used 1968 if (callprojs.resproj != nullptr && result != nullptr) { 1969 C->gvn_replace_by(callprojs.resproj, result); 1970 } 1971 1972 if (ejvms == nullptr) { 1973 // No exception edges to simply kill off those paths 1974 if (callprojs.catchall_catchproj != nullptr) { 1975 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1976 } 1977 if (callprojs.catchall_memproj != nullptr) { 1978 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1979 } 1980 if (callprojs.catchall_ioproj != nullptr) { 1981 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1982 } 1983 // Replace the old exception object with top 1984 if (callprojs.exobj != nullptr) { 1985 C->gvn_replace_by(callprojs.exobj, C->top()); 1986 } 1987 } else { 1988 GraphKit ekit(ejvms); 1989 1990 // Load my combined exception state into the kit, with all phis transformed: 1991 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1992 replaced_nodes_exception = ex_map->replaced_nodes(); 1993 1994 Node* ex_oop = ekit.use_exception_state(ex_map); 1995 1996 if (callprojs.catchall_catchproj != nullptr) { 1997 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1998 ex_ctl = ekit.control(); 1999 } 2000 if (callprojs.catchall_memproj != nullptr) { 2001 Node* ex_mem = ekit.reset_memory(); 2002 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem); 2003 add_mergemem_users_to_worklist(wl, ex_mem); 2004 } 2005 if (callprojs.catchall_ioproj != nullptr) { 2006 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 2007 } 2008 2009 // Replace the old exception object with the newly created one 2010 if (callprojs.exobj != nullptr) { 2011 C->gvn_replace_by(callprojs.exobj, ex_oop); 2012 } 2013 } 2014 2015 // Disconnect the call from the graph 2016 call->disconnect_inputs(C); 2017 C->gvn_replace_by(call, C->top()); 2018 2019 // Clean up any MergeMems that feed other MergeMems since the 2020 // optimizer doesn't like that. 2021 while (wl.size() > 0) { 2022 _gvn.transform(wl.pop()); 2023 } 2024 2025 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) { 2026 replaced_nodes.apply(C, final_ctl); 2027 } 2028 if (!ex_ctl->is_top() && do_replaced_nodes) { 2029 replaced_nodes_exception.apply(C, ex_ctl); 2030 } 2031 } 2032 2033 2034 //------------------------------increment_counter------------------------------ 2035 // for statistics: increment a VM counter by 1 2036 2037 void GraphKit::increment_counter(address counter_addr) { 2038 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2039 increment_counter(adr1); 2040 } 2041 2042 void GraphKit::increment_counter(Node* counter_addr) { 2043 int adr_type = Compile::AliasIdxRaw; 2044 Node* ctrl = control(); 2045 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered); 2046 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1))); 2047 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered); 2048 } 2049 2050 2051 //------------------------------uncommon_trap---------------------------------- 2052 // Bail out to the interpreter in mid-method. Implemented by calling the 2053 // uncommon_trap blob. This helper function inserts a runtime call with the 2054 // right debug info. 2055 Node* GraphKit::uncommon_trap(int trap_request, 2056 ciKlass* klass, const char* comment, 2057 bool must_throw, 2058 bool keep_exact_action) { 2059 if (failing()) stop(); 2060 if (stopped()) return nullptr; // trap reachable? 2061 2062 // Note: If ProfileTraps is true, and if a deopt. actually 2063 // occurs here, the runtime will make sure an MDO exists. There is 2064 // no need to call method()->ensure_method_data() at this point. 2065 2066 // Set the stack pointer to the right value for reexecution: 2067 set_sp(reexecute_sp()); 2068 2069 #ifdef ASSERT 2070 if (!must_throw) { 2071 // Make sure the stack has at least enough depth to execute 2072 // the current bytecode. 2073 int inputs, ignored_depth; 2074 if (compute_stack_effects(inputs, ignored_depth)) { 2075 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2076 Bytecodes::name(java_bc()), sp(), inputs); 2077 } 2078 } 2079 #endif 2080 2081 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2082 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2083 2084 switch (action) { 2085 case Deoptimization::Action_maybe_recompile: 2086 case Deoptimization::Action_reinterpret: 2087 // Temporary fix for 6529811 to allow virtual calls to be sure they 2088 // get the chance to go from mono->bi->mega 2089 if (!keep_exact_action && 2090 Deoptimization::trap_request_index(trap_request) < 0 && 2091 too_many_recompiles(reason)) { 2092 // This BCI is causing too many recompilations. 2093 if (C->log() != nullptr) { 2094 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2095 Deoptimization::trap_reason_name(reason), 2096 Deoptimization::trap_action_name(action)); 2097 } 2098 action = Deoptimization::Action_none; 2099 trap_request = Deoptimization::make_trap_request(reason, action); 2100 } else { 2101 C->set_trap_can_recompile(true); 2102 } 2103 break; 2104 case Deoptimization::Action_make_not_entrant: 2105 C->set_trap_can_recompile(true); 2106 break; 2107 case Deoptimization::Action_none: 2108 case Deoptimization::Action_make_not_compilable: 2109 break; 2110 default: 2111 #ifdef ASSERT 2112 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2113 #endif 2114 break; 2115 } 2116 2117 if (TraceOptoParse) { 2118 char buf[100]; 2119 tty->print_cr("Uncommon trap %s at bci:%d", 2120 Deoptimization::format_trap_request(buf, sizeof(buf), 2121 trap_request), bci()); 2122 } 2123 2124 CompileLog* log = C->log(); 2125 if (log != nullptr) { 2126 int kid = (klass == nullptr)? -1: log->identify(klass); 2127 log->begin_elem("uncommon_trap bci='%d'", bci()); 2128 char buf[100]; 2129 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2130 trap_request)); 2131 if (kid >= 0) log->print(" klass='%d'", kid); 2132 if (comment != nullptr) log->print(" comment='%s'", comment); 2133 log->end_elem(); 2134 } 2135 2136 // Make sure any guarding test views this path as very unlikely 2137 Node *i0 = control()->in(0); 2138 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test? 2139 IfNode *iff = i0->as_If(); 2140 float f = iff->_prob; // Get prob 2141 if (control()->Opcode() == Op_IfTrue) { 2142 if (f > PROB_UNLIKELY_MAG(4)) 2143 iff->_prob = PROB_MIN; 2144 } else { 2145 if (f < PROB_LIKELY_MAG(4)) 2146 iff->_prob = PROB_MAX; 2147 } 2148 } 2149 2150 // Clear out dead values from the debug info. 2151 kill_dead_locals(); 2152 2153 // Now insert the uncommon trap subroutine call 2154 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2155 const TypePtr* no_memory_effects = nullptr; 2156 // Pass the index of the class to be loaded 2157 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2158 (must_throw ? RC_MUST_THROW : 0), 2159 OptoRuntime::uncommon_trap_Type(), 2160 call_addr, "uncommon_trap", no_memory_effects, 2161 intcon(trap_request)); 2162 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2163 "must extract request correctly from the graph"); 2164 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2165 2166 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2167 // The debug info is the only real input to this call. 2168 2169 // Halt-and-catch fire here. The above call should never return! 2170 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2171 PRODUCT_ONLY(COMMA /*reachable*/false)); 2172 _gvn.set_type_bottom(halt); 2173 root()->add_req(halt); 2174 2175 stop_and_kill_map(); 2176 return call; 2177 } 2178 2179 2180 //--------------------------just_allocated_object------------------------------ 2181 // Report the object that was just allocated. 2182 // It must be the case that there are no intervening safepoints. 2183 // We use this to determine if an object is so "fresh" that 2184 // it does not require card marks. 2185 Node* GraphKit::just_allocated_object(Node* current_control) { 2186 Node* ctrl = current_control; 2187 // Object::<init> is invoked after allocation, most of invoke nodes 2188 // will be reduced, but a region node is kept in parse time, we check 2189 // the pattern and skip the region node if it degraded to a copy. 2190 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 && 2191 ctrl->as_Region()->is_copy()) { 2192 ctrl = ctrl->as_Region()->is_copy(); 2193 } 2194 if (C->recent_alloc_ctl() == ctrl) { 2195 return C->recent_alloc_obj(); 2196 } 2197 return nullptr; 2198 } 2199 2200 2201 /** 2202 * Record profiling data exact_kls for Node n with the type system so 2203 * that it can propagate it (speculation) 2204 * 2205 * @param n node that the type applies to 2206 * @param exact_kls type from profiling 2207 * @param maybe_null did profiling see null? 2208 * 2209 * @return node with improved type 2210 */ 2211 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2212 const Type* current_type = _gvn.type(n); 2213 assert(UseTypeSpeculation, "type speculation must be on"); 2214 2215 const TypePtr* speculative = current_type->speculative(); 2216 2217 // Should the klass from the profile be recorded in the speculative type? 2218 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2219 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces); 2220 const TypeOopPtr* xtype = tklass->as_instance_type(); 2221 assert(xtype->klass_is_exact(), "Should be exact"); 2222 // Any reason to believe n is not null (from this profiling or a previous one)? 2223 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2224 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2225 // record the new speculative type's depth 2226 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2227 speculative = speculative->with_inline_depth(jvms()->depth()); 2228 } else if (current_type->would_improve_ptr(ptr_kind)) { 2229 // Profiling report that null was never seen so we can change the 2230 // speculative type to non null ptr. 2231 if (ptr_kind == ProfileAlwaysNull) { 2232 speculative = TypePtr::NULL_PTR; 2233 } else { 2234 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2235 const TypePtr* ptr = TypePtr::NOTNULL; 2236 if (speculative != nullptr) { 2237 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2238 } else { 2239 speculative = ptr; 2240 } 2241 } 2242 } 2243 2244 if (speculative != current_type->speculative()) { 2245 // Build a type with a speculative type (what we think we know 2246 // about the type but will need a guard when we use it) 2247 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2248 // We're changing the type, we need a new CheckCast node to carry 2249 // the new type. The new type depends on the control: what 2250 // profiling tells us is only valid from here as far as we can 2251 // tell. 2252 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2253 cast = _gvn.transform(cast); 2254 replace_in_map(n, cast); 2255 n = cast; 2256 } 2257 2258 return n; 2259 } 2260 2261 /** 2262 * Record profiling data from receiver profiling at an invoke with the 2263 * type system so that it can propagate it (speculation) 2264 * 2265 * @param n receiver node 2266 * 2267 * @return node with improved type 2268 */ 2269 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2270 if (!UseTypeSpeculation) { 2271 return n; 2272 } 2273 ciKlass* exact_kls = profile_has_unique_klass(); 2274 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2275 if ((java_bc() == Bytecodes::_checkcast || 2276 java_bc() == Bytecodes::_instanceof || 2277 java_bc() == Bytecodes::_aastore) && 2278 method()->method_data()->is_mature()) { 2279 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2280 if (data != nullptr) { 2281 if (!data->as_BitData()->null_seen()) { 2282 ptr_kind = ProfileNeverNull; 2283 } else { 2284 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2285 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2286 uint i = 0; 2287 for (; i < call->row_limit(); i++) { 2288 ciKlass* receiver = call->receiver(i); 2289 if (receiver != nullptr) { 2290 break; 2291 } 2292 } 2293 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2294 } 2295 } 2296 } 2297 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2298 } 2299 2300 /** 2301 * Record profiling data from argument profiling at an invoke with the 2302 * type system so that it can propagate it (speculation) 2303 * 2304 * @param dest_method target method for the call 2305 * @param bc what invoke bytecode is this? 2306 */ 2307 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2308 if (!UseTypeSpeculation) { 2309 return; 2310 } 2311 const TypeFunc* tf = TypeFunc::make(dest_method); 2312 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2313 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2314 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2315 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2316 if (is_reference_type(targ->basic_type())) { 2317 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2318 ciKlass* better_type = nullptr; 2319 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2320 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2321 } 2322 i++; 2323 } 2324 } 2325 } 2326 2327 /** 2328 * Record profiling data from parameter profiling at an invoke with 2329 * the type system so that it can propagate it (speculation) 2330 */ 2331 void GraphKit::record_profiled_parameters_for_speculation() { 2332 if (!UseTypeSpeculation) { 2333 return; 2334 } 2335 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2336 if (_gvn.type(local(i))->isa_oopptr()) { 2337 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2338 ciKlass* better_type = nullptr; 2339 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2340 record_profile_for_speculation(local(i), better_type, ptr_kind); 2341 } 2342 j++; 2343 } 2344 } 2345 } 2346 2347 /** 2348 * Record profiling data from return value profiling at an invoke with 2349 * the type system so that it can propagate it (speculation) 2350 */ 2351 void GraphKit::record_profiled_return_for_speculation() { 2352 if (!UseTypeSpeculation) { 2353 return; 2354 } 2355 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2356 ciKlass* better_type = nullptr; 2357 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2358 // If profiling reports a single type for the return value, 2359 // feed it to the type system so it can propagate it as a 2360 // speculative type 2361 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2362 } 2363 } 2364 2365 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2366 if (Matcher::strict_fp_requires_explicit_rounding) { 2367 // (Note: TypeFunc::make has a cache that makes this fast.) 2368 const TypeFunc* tf = TypeFunc::make(dest_method); 2369 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2370 for (int j = 0; j < nargs; j++) { 2371 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2372 if (targ->basic_type() == T_DOUBLE) { 2373 // If any parameters are doubles, they must be rounded before 2374 // the call, dprecision_rounding does gvn.transform 2375 Node *arg = argument(j); 2376 arg = dprecision_rounding(arg); 2377 set_argument(j, arg); 2378 } 2379 } 2380 } 2381 } 2382 2383 // rounding for strict float precision conformance 2384 Node* GraphKit::precision_rounding(Node* n) { 2385 if (Matcher::strict_fp_requires_explicit_rounding) { 2386 #ifdef IA32 2387 if (UseSSE == 0) { 2388 return _gvn.transform(new RoundFloatNode(0, n)); 2389 } 2390 #else 2391 Unimplemented(); 2392 #endif // IA32 2393 } 2394 return n; 2395 } 2396 2397 // rounding for strict double precision conformance 2398 Node* GraphKit::dprecision_rounding(Node *n) { 2399 if (Matcher::strict_fp_requires_explicit_rounding) { 2400 #ifdef IA32 2401 if (UseSSE < 2) { 2402 return _gvn.transform(new RoundDoubleNode(0, n)); 2403 } 2404 #else 2405 Unimplemented(); 2406 #endif // IA32 2407 } 2408 return n; 2409 } 2410 2411 //============================================================================= 2412 // Generate a fast path/slow path idiom. Graph looks like: 2413 // [foo] indicates that 'foo' is a parameter 2414 // 2415 // [in] null 2416 // \ / 2417 // CmpP 2418 // Bool ne 2419 // If 2420 // / \ 2421 // True False-<2> 2422 // / | 2423 // / cast_not_null 2424 // Load | | ^ 2425 // [fast_test] | | 2426 // gvn to opt_test | | 2427 // / \ | <1> 2428 // True False | 2429 // | \\ | 2430 // [slow_call] \[fast_result] 2431 // Ctl Val \ \ 2432 // | \ \ 2433 // Catch <1> \ \ 2434 // / \ ^ \ \ 2435 // Ex No_Ex | \ \ 2436 // | \ \ | \ <2> \ 2437 // ... \ [slow_res] | | \ [null_result] 2438 // \ \--+--+--- | | 2439 // \ | / \ | / 2440 // --------Region Phi 2441 // 2442 //============================================================================= 2443 // Code is structured as a series of driver functions all called 'do_XXX' that 2444 // call a set of helper functions. Helper functions first, then drivers. 2445 2446 //------------------------------null_check_oop--------------------------------- 2447 // Null check oop. Set null-path control into Region in slot 3. 2448 // Make a cast-not-nullness use the other not-null control. Return cast. 2449 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2450 bool never_see_null, 2451 bool safe_for_replace, 2452 bool speculative) { 2453 // Initial null check taken path 2454 (*null_control) = top(); 2455 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2456 2457 // Generate uncommon_trap: 2458 if (never_see_null && (*null_control) != top()) { 2459 // If we see an unexpected null at a check-cast we record it and force a 2460 // recompile; the offending check-cast will be compiled to handle nulls. 2461 // If we see more than one offending BCI, then all checkcasts in the 2462 // method will be compiled to handle nulls. 2463 PreserveJVMState pjvms(this); 2464 set_control(*null_control); 2465 replace_in_map(value, null()); 2466 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2467 uncommon_trap(reason, 2468 Deoptimization::Action_make_not_entrant); 2469 (*null_control) = top(); // null path is dead 2470 } 2471 if ((*null_control) == top() && safe_for_replace) { 2472 replace_in_map(value, cast); 2473 } 2474 2475 // Cast away null-ness on the result 2476 return cast; 2477 } 2478 2479 //------------------------------opt_iff---------------------------------------- 2480 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2481 // Return slow-path control. 2482 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2483 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2484 2485 // Fast path taken; set region slot 2 2486 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2487 region->init_req(2,fast_taken); // Capture fast-control 2488 2489 // Fast path not-taken, i.e. slow path 2490 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2491 return slow_taken; 2492 } 2493 2494 //-----------------------------make_runtime_call------------------------------- 2495 Node* GraphKit::make_runtime_call(int flags, 2496 const TypeFunc* call_type, address call_addr, 2497 const char* call_name, 2498 const TypePtr* adr_type, 2499 // The following parms are all optional. 2500 // The first null ends the list. 2501 Node* parm0, Node* parm1, 2502 Node* parm2, Node* parm3, 2503 Node* parm4, Node* parm5, 2504 Node* parm6, Node* parm7) { 2505 assert(call_addr != nullptr, "must not call null targets"); 2506 2507 // Slow-path call 2508 bool is_leaf = !(flags & RC_NO_LEAF); 2509 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2510 if (call_name == nullptr) { 2511 assert(!is_leaf, "must supply name for leaf"); 2512 call_name = OptoRuntime::stub_name(call_addr); 2513 } 2514 CallNode* call; 2515 if (!is_leaf) { 2516 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type); 2517 } else if (flags & RC_NO_FP) { 2518 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2519 } else if (flags & RC_VECTOR){ 2520 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte; 2521 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits); 2522 } else { 2523 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2524 } 2525 2526 // The following is similar to set_edges_for_java_call, 2527 // except that the memory effects of the call are restricted to AliasIdxRaw. 2528 2529 // Slow path call has no side-effects, uses few values 2530 bool wide_in = !(flags & RC_NARROW_MEM); 2531 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2532 2533 Node* prev_mem = nullptr; 2534 if (wide_in) { 2535 prev_mem = set_predefined_input_for_runtime_call(call); 2536 } else { 2537 assert(!wide_out, "narrow in => narrow out"); 2538 Node* narrow_mem = memory(adr_type); 2539 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2540 } 2541 2542 // Hook each parm in order. Stop looking at the first null. 2543 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0); 2544 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1); 2545 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2); 2546 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3); 2547 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4); 2548 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5); 2549 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6); 2550 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7); 2551 /* close each nested if ===> */ } } } } } } } } 2552 assert(call->in(call->req()-1) != nullptr, "must initialize all parms"); 2553 2554 if (!is_leaf) { 2555 // Non-leaves can block and take safepoints: 2556 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2557 } 2558 // Non-leaves can throw exceptions: 2559 if (has_io) { 2560 call->set_req(TypeFunc::I_O, i_o()); 2561 } 2562 2563 if (flags & RC_UNCOMMON) { 2564 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2565 // (An "if" probability corresponds roughly to an unconditional count. 2566 // Sort of.) 2567 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2568 } 2569 2570 Node* c = _gvn.transform(call); 2571 assert(c == call, "cannot disappear"); 2572 2573 if (wide_out) { 2574 // Slow path call has full side-effects. 2575 set_predefined_output_for_runtime_call(call); 2576 } else { 2577 // Slow path call has few side-effects, and/or sets few values. 2578 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2579 } 2580 2581 if (has_io) { 2582 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2583 } 2584 return call; 2585 2586 } 2587 2588 // i2b 2589 Node* GraphKit::sign_extend_byte(Node* in) { 2590 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24))); 2591 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24))); 2592 } 2593 2594 // i2s 2595 Node* GraphKit::sign_extend_short(Node* in) { 2596 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16))); 2597 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16))); 2598 } 2599 2600 //------------------------------merge_memory----------------------------------- 2601 // Merge memory from one path into the current memory state. 2602 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2603 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2604 Node* old_slice = mms.force_memory(); 2605 Node* new_slice = mms.memory2(); 2606 if (old_slice != new_slice) { 2607 PhiNode* phi; 2608 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2609 if (mms.is_empty()) { 2610 // clone base memory Phi's inputs for this memory slice 2611 assert(old_slice == mms.base_memory(), "sanity"); 2612 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C)); 2613 _gvn.set_type(phi, Type::MEMORY); 2614 for (uint i = 1; i < phi->req(); i++) { 2615 phi->init_req(i, old_slice->in(i)); 2616 } 2617 } else { 2618 phi = old_slice->as_Phi(); // Phi was generated already 2619 } 2620 } else { 2621 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2622 _gvn.set_type(phi, Type::MEMORY); 2623 } 2624 phi->set_req(new_path, new_slice); 2625 mms.set_memory(phi); 2626 } 2627 } 2628 } 2629 2630 //------------------------------make_slow_call_ex------------------------------ 2631 // Make the exception handler hookups for the slow call 2632 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2633 if (stopped()) return; 2634 2635 // Make a catch node with just two handlers: fall-through and catch-all 2636 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2637 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2638 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2639 _gvn.set_type_bottom(norm); 2640 C->record_for_igvn(norm); 2641 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2642 2643 { PreserveJVMState pjvms(this); 2644 set_control(excp); 2645 set_i_o(i_o); 2646 2647 if (excp != top()) { 2648 if (deoptimize) { 2649 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2650 uncommon_trap(Deoptimization::Reason_unhandled, 2651 Deoptimization::Action_none); 2652 } else { 2653 // Create an exception state also. 2654 // Use an exact type if the caller has a specific exception. 2655 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2656 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2657 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2658 } 2659 } 2660 } 2661 2662 // Get the no-exception control from the CatchNode. 2663 set_control(norm); 2664 } 2665 2666 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) { 2667 Node* cmp = nullptr; 2668 switch(bt) { 2669 case T_INT: cmp = new CmpINode(in1, in2); break; 2670 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2671 default: fatal("unexpected comparison type %s", type2name(bt)); 2672 } 2673 cmp = gvn.transform(cmp); 2674 Node* bol = gvn.transform(new BoolNode(cmp, test)); 2675 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2676 gvn.transform(iff); 2677 if (!bol->is_Con()) gvn.record_for_igvn(iff); 2678 return iff; 2679 } 2680 2681 //-------------------------------gen_subtype_check----------------------------- 2682 // Generate a subtyping check. Takes as input the subtype and supertype. 2683 // Returns 2 values: sets the default control() to the true path and returns 2684 // the false path. Only reads invariant memory; sets no (visible) memory. 2685 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2686 // but that's not exposed to the optimizer. This call also doesn't take in an 2687 // Object; if you wish to check an Object you need to load the Object's class 2688 // prior to coming here. 2689 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn, 2690 ciMethod* method, int bci) { 2691 Compile* C = gvn.C; 2692 if ((*ctrl)->is_top()) { 2693 return C->top(); 2694 } 2695 2696 // Fast check for identical types, perhaps identical constants. 2697 // The types can even be identical non-constants, in cases 2698 // involving Array.newInstance, Object.clone, etc. 2699 if (subklass == superklass) 2700 return C->top(); // false path is dead; no test needed. 2701 2702 if (gvn.type(superklass)->singleton()) { 2703 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2704 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr(); 2705 2706 // In the common case of an exact superklass, try to fold up the 2707 // test before generating code. You may ask, why not just generate 2708 // the code and then let it fold up? The answer is that the generated 2709 // code will necessarily include null checks, which do not always 2710 // completely fold away. If they are also needless, then they turn 2711 // into a performance loss. Example: 2712 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2713 // Here, the type of 'fa' is often exact, so the store check 2714 // of fa[1]=x will fold up, without testing the nullness of x. 2715 // 2716 // Do not skip the static sub type check with StressReflectiveCode during 2717 // parsing (i.e. with ExpandSubTypeCheckAtParseTime) because the 2718 // associated CheckCastNodePP could already be folded when the type 2719 // system can prove it's an impossible type. Therefore, we should also 2720 // do the static sub type check here to ensure control is folded as well. 2721 // Otherwise, the graph is left in a broken state. 2722 // At macro expansion, we would have already folded the SubTypeCheckNode 2723 // being expanded here because we always perform the static sub type 2724 // check in SubTypeCheckNode::sub() regardless of whether 2725 // StressReflectiveCode is set or not. 2726 switch (C->static_subtype_check(superk, subk, !ExpandSubTypeCheckAtParseTime)) { 2727 case Compile::SSC_always_false: 2728 { 2729 Node* always_fail = *ctrl; 2730 *ctrl = gvn.C->top(); 2731 return always_fail; 2732 } 2733 case Compile::SSC_always_true: 2734 return C->top(); 2735 case Compile::SSC_easy_test: 2736 { 2737 // Just do a direct pointer compare and be done. 2738 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2739 *ctrl = gvn.transform(new IfTrueNode(iff)); 2740 return gvn.transform(new IfFalseNode(iff)); 2741 } 2742 case Compile::SSC_full_test: 2743 break; 2744 default: 2745 ShouldNotReachHere(); 2746 } 2747 } 2748 2749 // %%% Possible further optimization: Even if the superklass is not exact, 2750 // if the subklass is the unique subtype of the superklass, the check 2751 // will always succeed. We could leave a dependency behind to ensure this. 2752 2753 // First load the super-klass's check-offset 2754 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2755 Node* m = C->immutable_memory(); 2756 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2757 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2758 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int(); 2759 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con; 2760 bool might_be_cache = (chk_off_con == cacheoff_con); 2761 2762 // Load from the sub-klass's super-class display list, or a 1-word cache of 2763 // the secondary superclass list, or a failing value with a sentinel offset 2764 // if the super-klass is an interface or exceptionally deep in the Java 2765 // hierarchy and we have to scan the secondary superclass list the hard way. 2766 // Worst-case type is a little odd: null is allowed as a result (usually 2767 // klass loads can never produce a null). 2768 Node *chk_off_X = chk_off; 2769 #ifdef _LP64 2770 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X)); 2771 #endif 2772 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X)); 2773 // For some types like interfaces the following loadKlass is from a 1-word 2774 // cache which is mutable so can't use immutable memory. Other 2775 // types load from the super-class display table which is immutable. 2776 Node *kmem = C->immutable_memory(); 2777 // secondary_super_cache is not immutable but can be treated as such because: 2778 // - no ideal node writes to it in a way that could cause an 2779 // incorrect/missed optimization of the following Load. 2780 // - it's a cache so, worse case, not reading the latest value 2781 // wouldn't cause incorrect execution 2782 if (might_be_cache && mem != nullptr) { 2783 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem; 2784 } 2785 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL)); 2786 2787 // Compile speed common case: ARE a subtype and we canNOT fail 2788 if (superklass == nkls) { 2789 return C->top(); // false path is dead; no test needed. 2790 } 2791 2792 // Gather the various success & failures here 2793 RegionNode* r_not_subtype = new RegionNode(3); 2794 gvn.record_for_igvn(r_not_subtype); 2795 RegionNode* r_ok_subtype = new RegionNode(4); 2796 gvn.record_for_igvn(r_ok_subtype); 2797 2798 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the 2799 // SubTypeCheck node 2800 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) { 2801 ciCallProfile profile = method->call_profile_at_bci(bci); 2802 float total_prob = 0; 2803 for (int i = 0; profile.has_receiver(i); ++i) { 2804 float prob = profile.receiver_prob(i); 2805 total_prob += prob; 2806 } 2807 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) { 2808 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2809 for (int i = 0; profile.has_receiver(i); ++i) { 2810 ciKlass* klass = profile.receiver(i); 2811 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass); 2812 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t); 2813 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) { 2814 continue; 2815 } 2816 float prob = profile.receiver_prob(i); 2817 ConNode* klass_node = gvn.makecon(klass_t); 2818 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS); 2819 Node* iftrue = gvn.transform(new IfTrueNode(iff)); 2820 2821 if (result == Compile::SSC_always_true) { 2822 r_ok_subtype->add_req(iftrue); 2823 } else { 2824 assert(result == Compile::SSC_always_false, ""); 2825 r_not_subtype->add_req(iftrue); 2826 } 2827 *ctrl = gvn.transform(new IfFalseNode(iff)); 2828 } 2829 } 2830 } 2831 2832 // See if we get an immediate positive hit. Happens roughly 83% of the 2833 // time. Test to see if the value loaded just previously from the subklass 2834 // is exactly the superklass. 2835 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2836 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1)); 2837 *ctrl = gvn.transform(new IfFalseNode(iff1)); 2838 2839 // Compile speed common case: Check for being deterministic right now. If 2840 // chk_off is a constant and not equal to cacheoff then we are NOT a 2841 // subklass. In this case we need exactly the 1 test above and we can 2842 // return those results immediately. 2843 if (!might_be_cache) { 2844 Node* not_subtype_ctrl = *ctrl; 2845 *ctrl = iftrue1; // We need exactly the 1 test above 2846 PhaseIterGVN* igvn = gvn.is_IterGVN(); 2847 if (igvn != nullptr) { 2848 igvn->remove_globally_dead_node(r_ok_subtype); 2849 igvn->remove_globally_dead_node(r_not_subtype); 2850 } 2851 return not_subtype_ctrl; 2852 } 2853 2854 r_ok_subtype->init_req(1, iftrue1); 2855 2856 // Check for immediate negative hit. Happens roughly 11% of the time (which 2857 // is roughly 63% of the remaining cases). Test to see if the loaded 2858 // check-offset points into the subklass display list or the 1-element 2859 // cache. If it points to the display (and NOT the cache) and the display 2860 // missed then it's not a subtype. 2861 Node *cacheoff = gvn.intcon(cacheoff_con); 2862 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2863 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2))); 2864 *ctrl = gvn.transform(new IfFalseNode(iff2)); 2865 2866 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2867 // No performance impact (too rare) but allows sharing of secondary arrays 2868 // which has some footprint reduction. 2869 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2870 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3))); 2871 *ctrl = gvn.transform(new IfFalseNode(iff3)); 2872 2873 // -- Roads not taken here: -- 2874 // We could also have chosen to perform the self-check at the beginning 2875 // of this code sequence, as the assembler does. This would not pay off 2876 // the same way, since the optimizer, unlike the assembler, can perform 2877 // static type analysis to fold away many successful self-checks. 2878 // Non-foldable self checks work better here in second position, because 2879 // the initial primary superclass check subsumes a self-check for most 2880 // types. An exception would be a secondary type like array-of-interface, 2881 // which does not appear in its own primary supertype display. 2882 // Finally, we could have chosen to move the self-check into the 2883 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2884 // dependent manner. But it is worthwhile to have the check here, 2885 // where it can be perhaps be optimized. The cost in code space is 2886 // small (register compare, branch). 2887 2888 // Now do a linear scan of the secondary super-klass array. Again, no real 2889 // performance impact (too rare) but it's gotta be done. 2890 // Since the code is rarely used, there is no penalty for moving it 2891 // out of line, and it can only improve I-cache density. 2892 // The decision to inline or out-of-line this final check is platform 2893 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2894 Node* psc = gvn.transform( 2895 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2896 2897 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2898 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4))); 2899 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4))); 2900 2901 // Return false path; set default control to true path. 2902 *ctrl = gvn.transform(r_ok_subtype); 2903 return gvn.transform(r_not_subtype); 2904 } 2905 2906 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) { 2907 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over 2908 ExpandSubTypeCheckAtParseTime; // forced expansion 2909 if (expand_subtype_check) { 2910 MergeMemNode* mem = merged_memory(); 2911 Node* ctrl = control(); 2912 Node* subklass = obj_or_subklass; 2913 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) { 2914 subklass = load_object_klass(obj_or_subklass); 2915 } 2916 2917 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci()); 2918 set_control(ctrl); 2919 return n; 2920 } 2921 2922 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci())); 2923 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq)); 2924 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2925 set_control(_gvn.transform(new IfTrueNode(iff))); 2926 return _gvn.transform(new IfFalseNode(iff)); 2927 } 2928 2929 // Profile-driven exact type check: 2930 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2931 float prob, 2932 Node* *casted_receiver) { 2933 assert(!klass->is_interface(), "no exact type check on interfaces"); 2934 2935 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces); 2936 Node* recv_klass = load_object_klass(receiver); 2937 Node* want_klass = makecon(tklass); 2938 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass)); 2939 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 2940 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2941 set_control( _gvn.transform(new IfTrueNode (iff))); 2942 Node* fail = _gvn.transform(new IfFalseNode(iff)); 2943 2944 if (!stopped()) { 2945 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 2946 const TypeOopPtr* recvx_type = tklass->as_instance_type(); 2947 assert(recvx_type->klass_is_exact(), ""); 2948 2949 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts 2950 // Subsume downstream occurrences of receiver with a cast to 2951 // recv_xtype, since now we know what the type will be. 2952 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type); 2953 (*casted_receiver) = _gvn.transform(cast); 2954 assert(!(*casted_receiver)->is_top(), "that path should be unreachable"); 2955 // (User must make the replace_in_map call.) 2956 } 2957 } 2958 2959 return fail; 2960 } 2961 2962 //------------------------------subtype_check_receiver------------------------- 2963 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 2964 Node** casted_receiver) { 2965 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve(); 2966 Node* want_klass = makecon(tklass); 2967 2968 Node* slow_ctl = gen_subtype_check(receiver, want_klass); 2969 2970 // Ignore interface type information until interface types are properly tracked. 2971 if (!stopped() && !klass->is_interface()) { 2972 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 2973 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 2974 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts 2975 Node* cast = new CheckCastPPNode(control(), receiver, recv_type); 2976 (*casted_receiver) = _gvn.transform(cast); 2977 } 2978 } 2979 2980 return slow_ctl; 2981 } 2982 2983 //------------------------------seems_never_null------------------------------- 2984 // Use null_seen information if it is available from the profile. 2985 // If we see an unexpected null at a type check we record it and force a 2986 // recompile; the offending check will be recompiled to handle nulls. 2987 // If we see several offending BCIs, then all checks in the 2988 // method will be recompiled. 2989 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2990 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2991 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2992 if (UncommonNullCast // Cutout for this technique 2993 && obj != null() // And not the -Xcomp stupid case? 2994 && !too_many_traps(reason) 2995 ) { 2996 if (speculating) { 2997 return true; 2998 } 2999 if (data == nullptr) 3000 // Edge case: no mature data. Be optimistic here. 3001 return true; 3002 // If the profile has not seen a null, assume it won't happen. 3003 assert(java_bc() == Bytecodes::_checkcast || 3004 java_bc() == Bytecodes::_instanceof || 3005 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 3006 return !data->as_BitData()->null_seen(); 3007 } 3008 speculating = false; 3009 return false; 3010 } 3011 3012 void GraphKit::guard_klass_being_initialized(Node* klass) { 3013 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 3014 Node* adr = basic_plus_adr(top(), klass, init_state_off); 3015 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3016 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 3017 T_BYTE, MemNode::unordered); 3018 init_state = _gvn.transform(init_state); 3019 3020 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 3021 3022 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 3023 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3024 3025 { BuildCutout unless(this, tst, PROB_MAX); 3026 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 3027 } 3028 } 3029 3030 void GraphKit::guard_init_thread(Node* klass) { 3031 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 3032 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 3033 3034 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3035 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 3036 T_ADDRESS, MemNode::unordered); 3037 init_thread = _gvn.transform(init_thread); 3038 3039 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 3040 3041 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 3042 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3043 3044 { BuildCutout unless(this, tst, PROB_MAX); 3045 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 3046 } 3047 } 3048 3049 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3050 if (ik->is_being_initialized()) { 3051 if (C->needs_clinit_barrier(ik, context)) { 3052 Node* klass = makecon(TypeKlassPtr::make(ik)); 3053 guard_klass_being_initialized(klass); 3054 guard_init_thread(klass); 3055 insert_mem_bar(Op_MemBarCPUOrder); 3056 } 3057 } else if (ik->is_initialized()) { 3058 return; // no barrier needed 3059 } else { 3060 uncommon_trap(Deoptimization::Reason_uninitialized, 3061 Deoptimization::Action_reinterpret, 3062 nullptr); 3063 } 3064 } 3065 3066 //------------------------maybe_cast_profiled_receiver------------------------- 3067 // If the profile has seen exactly one type, narrow to exactly that type. 3068 // Subsequent type checks will always fold up. 3069 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3070 const TypeKlassPtr* require_klass, 3071 ciKlass* spec_klass, 3072 bool safe_for_replace) { 3073 if (!UseTypeProfile || !TypeProfileCasts) return nullptr; 3074 3075 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr); 3076 3077 // Make sure we haven't already deoptimized from this tactic. 3078 if (too_many_traps_or_recompiles(reason)) 3079 return nullptr; 3080 3081 // (No, this isn't a call, but it's enough like a virtual call 3082 // to use the same ciMethod accessor to get the profile info...) 3083 // If we have a speculative type use it instead of profiling (which 3084 // may not help us) 3085 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass; 3086 if (exact_kls != nullptr) {// no cast failures here 3087 if (require_klass == nullptr || 3088 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) { 3089 // If we narrow the type to match what the type profile sees or 3090 // the speculative type, we can then remove the rest of the 3091 // cast. 3092 // This is a win, even if the exact_kls is very specific, 3093 // because downstream operations, such as method calls, 3094 // will often benefit from the sharper type. 3095 Node* exact_obj = not_null_obj; // will get updated in place... 3096 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3097 &exact_obj); 3098 { PreserveJVMState pjvms(this); 3099 set_control(slow_ctl); 3100 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3101 } 3102 if (safe_for_replace) { 3103 replace_in_map(not_null_obj, exact_obj); 3104 } 3105 return exact_obj; 3106 } 3107 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3108 } 3109 3110 return nullptr; 3111 } 3112 3113 /** 3114 * Cast obj to type and emit guard unless we had too many traps here 3115 * already 3116 * 3117 * @param obj node being casted 3118 * @param type type to cast the node to 3119 * @param not_null true if we know node cannot be null 3120 */ 3121 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3122 ciKlass* type, 3123 bool not_null) { 3124 if (stopped()) { 3125 return obj; 3126 } 3127 3128 // type is null if profiling tells us this object is always null 3129 if (type != nullptr) { 3130 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3131 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3132 3133 if (!too_many_traps_or_recompiles(null_reason) && 3134 !too_many_traps_or_recompiles(class_reason)) { 3135 Node* not_null_obj = nullptr; 3136 // not_null is true if we know the object is not null and 3137 // there's no need for a null check 3138 if (!not_null) { 3139 Node* null_ctl = top(); 3140 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3141 assert(null_ctl->is_top(), "no null control here"); 3142 } else { 3143 not_null_obj = obj; 3144 } 3145 3146 Node* exact_obj = not_null_obj; 3147 ciKlass* exact_kls = type; 3148 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3149 &exact_obj); 3150 { 3151 PreserveJVMState pjvms(this); 3152 set_control(slow_ctl); 3153 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3154 } 3155 replace_in_map(not_null_obj, exact_obj); 3156 obj = exact_obj; 3157 } 3158 } else { 3159 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3160 Node* exact_obj = null_assert(obj); 3161 replace_in_map(obj, exact_obj); 3162 obj = exact_obj; 3163 } 3164 } 3165 return obj; 3166 } 3167 3168 //-------------------------------gen_instanceof-------------------------------- 3169 // Generate an instance-of idiom. Used by both the instance-of bytecode 3170 // and the reflective instance-of call. 3171 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3172 kill_dead_locals(); // Benefit all the uncommon traps 3173 assert( !stopped(), "dead parse path should be checked in callers" ); 3174 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3175 "must check for not-null not-dead klass in callers"); 3176 3177 // Make the merge point 3178 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3179 RegionNode* region = new RegionNode(PATH_LIMIT); 3180 Node* phi = new PhiNode(region, TypeInt::BOOL); 3181 C->set_has_split_ifs(true); // Has chance for split-if optimization 3182 3183 ciProfileData* data = nullptr; 3184 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3185 data = method()->method_data()->bci_to_data(bci()); 3186 } 3187 bool speculative_not_null = false; 3188 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3189 && seems_never_null(obj, data, speculative_not_null)); 3190 3191 // Null check; get casted pointer; set region slot 3 3192 Node* null_ctl = top(); 3193 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3194 3195 // If not_null_obj is dead, only null-path is taken 3196 if (stopped()) { // Doing instance-of on a null? 3197 set_control(null_ctl); 3198 return intcon(0); 3199 } 3200 region->init_req(_null_path, null_ctl); 3201 phi ->init_req(_null_path, intcon(0)); // Set null path value 3202 if (null_ctl == top()) { 3203 // Do this eagerly, so that pattern matches like is_diamond_phi 3204 // will work even during parsing. 3205 assert(_null_path == PATH_LIMIT-1, "delete last"); 3206 region->del_req(_null_path); 3207 phi ->del_req(_null_path); 3208 } 3209 3210 // Do we know the type check always succeed? 3211 bool known_statically = false; 3212 if (_gvn.type(superklass)->singleton()) { 3213 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr(); 3214 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type(); 3215 if (subk->is_loaded()) { 3216 int static_res = C->static_subtype_check(superk, subk); 3217 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3218 } 3219 } 3220 3221 if (!known_statically) { 3222 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3223 // We may not have profiling here or it may not help us. If we 3224 // have a speculative type use it to perform an exact cast. 3225 ciKlass* spec_obj_type = obj_type->speculative_type(); 3226 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) { 3227 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace); 3228 if (stopped()) { // Profile disagrees with this path. 3229 set_control(null_ctl); // Null is the only remaining possibility. 3230 return intcon(0); 3231 } 3232 if (cast_obj != nullptr) { 3233 not_null_obj = cast_obj; 3234 } 3235 } 3236 } 3237 3238 // Generate the subtype check 3239 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3240 3241 // Plug in the success path to the general merge in slot 1. 3242 region->init_req(_obj_path, control()); 3243 phi ->init_req(_obj_path, intcon(1)); 3244 3245 // Plug in the failing path to the general merge in slot 2. 3246 region->init_req(_fail_path, not_subtype_ctrl); 3247 phi ->init_req(_fail_path, intcon(0)); 3248 3249 // Return final merged results 3250 set_control( _gvn.transform(region) ); 3251 record_for_igvn(region); 3252 3253 // If we know the type check always succeeds then we don't use the 3254 // profiling data at this bytecode. Don't lose it, feed it to the 3255 // type system as a speculative type. 3256 if (safe_for_replace) { 3257 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3258 replace_in_map(obj, casted_obj); 3259 } 3260 3261 return _gvn.transform(phi); 3262 } 3263 3264 //-------------------------------gen_checkcast--------------------------------- 3265 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3266 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3267 // uncommon-trap paths work. Adjust stack after this call. 3268 // If failure_control is supplied and not null, it is filled in with 3269 // the control edge for the cast failure. Otherwise, an appropriate 3270 // uncommon trap or exception is thrown. 3271 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3272 Node* *failure_control) { 3273 kill_dead_locals(); // Benefit all the uncommon traps 3274 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr(); 3275 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve(); 3276 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type(); 3277 3278 // Fast cutout: Check the case that the cast is vacuously true. 3279 // This detects the common cases where the test will short-circuit 3280 // away completely. We do this before we perform the null check, 3281 // because if the test is going to turn into zero code, we don't 3282 // want a residual null check left around. (Causes a slowdown, 3283 // for example, in some objArray manipulations, such as a[i]=a[j].) 3284 if (improved_klass_ptr_type->singleton()) { 3285 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3286 if (objtp != nullptr) { 3287 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) { 3288 case Compile::SSC_always_true: 3289 // If we know the type check always succeed then we don't use 3290 // the profiling data at this bytecode. Don't lose it, feed it 3291 // to the type system as a speculative type. 3292 return record_profiled_receiver_for_speculation(obj); 3293 case Compile::SSC_always_false: 3294 // It needs a null check because a null will *pass* the cast check. 3295 // A non-null value will always produce an exception. 3296 if (!objtp->maybe_null()) { 3297 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3298 Deoptimization::DeoptReason reason = is_aastore ? 3299 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3300 builtin_throw(reason); 3301 return top(); 3302 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3303 return null_assert(obj); 3304 } 3305 break; // Fall through to full check 3306 default: 3307 break; 3308 } 3309 } 3310 } 3311 3312 ciProfileData* data = nullptr; 3313 bool safe_for_replace = false; 3314 if (failure_control == nullptr) { // use MDO in regular case only 3315 assert(java_bc() == Bytecodes::_aastore || 3316 java_bc() == Bytecodes::_checkcast, 3317 "interpreter profiles type checks only for these BCs"); 3318 data = method()->method_data()->bci_to_data(bci()); 3319 safe_for_replace = true; 3320 } 3321 3322 // Make the merge point 3323 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3324 RegionNode* region = new RegionNode(PATH_LIMIT); 3325 Node* phi = new PhiNode(region, toop); 3326 C->set_has_split_ifs(true); // Has chance for split-if optimization 3327 3328 // Use null-cast information if it is available 3329 bool speculative_not_null = false; 3330 bool never_see_null = ((failure_control == nullptr) // regular case only 3331 && seems_never_null(obj, data, speculative_not_null)); 3332 3333 // Null check; get casted pointer; set region slot 3 3334 Node* null_ctl = top(); 3335 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3336 3337 // If not_null_obj is dead, only null-path is taken 3338 if (stopped()) { // Doing instance-of on a null? 3339 set_control(null_ctl); 3340 return null(); 3341 } 3342 region->init_req(_null_path, null_ctl); 3343 phi ->init_req(_null_path, null()); // Set null path value 3344 if (null_ctl == top()) { 3345 // Do this eagerly, so that pattern matches like is_diamond_phi 3346 // will work even during parsing. 3347 assert(_null_path == PATH_LIMIT-1, "delete last"); 3348 region->del_req(_null_path); 3349 phi ->del_req(_null_path); 3350 } 3351 3352 Node* cast_obj = nullptr; 3353 if (improved_klass_ptr_type->klass_is_exact()) { 3354 // The following optimization tries to statically cast the speculative type of the object 3355 // (for example obtained during profiling) to the type of the superklass and then do a 3356 // dynamic check that the type of the object is what we expect. To work correctly 3357 // for checkcast and aastore the type of superklass should be exact. 3358 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3359 // We may not have profiling here or it may not help us. If we have 3360 // a speculative type use it to perform an exact cast. 3361 ciKlass* spec_obj_type = obj_type->speculative_type(); 3362 if (spec_obj_type != nullptr || data != nullptr) { 3363 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace); 3364 if (cast_obj != nullptr) { 3365 if (failure_control != nullptr) // failure is now impossible 3366 (*failure_control) = top(); 3367 // adjust the type of the phi to the exact klass: 3368 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3369 } 3370 } 3371 } 3372 3373 if (cast_obj == nullptr) { 3374 // Generate the subtype check 3375 Node* improved_superklass = superklass; 3376 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) { 3377 improved_superklass = makecon(improved_klass_ptr_type); 3378 } 3379 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass); 3380 3381 // Plug in success path into the merge 3382 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3383 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3384 if (failure_control == nullptr) { 3385 if (not_subtype_ctrl != top()) { // If failure is possible 3386 PreserveJVMState pjvms(this); 3387 set_control(not_subtype_ctrl); 3388 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3389 Deoptimization::DeoptReason reason = is_aastore ? 3390 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3391 builtin_throw(reason); 3392 } 3393 } else { 3394 (*failure_control) = not_subtype_ctrl; 3395 } 3396 } 3397 3398 region->init_req(_obj_path, control()); 3399 phi ->init_req(_obj_path, cast_obj); 3400 3401 // A merge of null or Casted-NotNull obj 3402 Node* res = _gvn.transform(phi); 3403 3404 // Note I do NOT always 'replace_in_map(obj,result)' here. 3405 // if( tk->klass()->can_be_primary_super() ) 3406 // This means that if I successfully store an Object into an array-of-String 3407 // I 'forget' that the Object is really now known to be a String. I have to 3408 // do this because we don't have true union types for interfaces - if I store 3409 // a Baz into an array-of-Interface and then tell the optimizer it's an 3410 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3411 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3412 // replace_in_map( obj, res ); 3413 3414 // Return final merged results 3415 set_control( _gvn.transform(region) ); 3416 record_for_igvn(region); 3417 3418 return record_profiled_receiver_for_speculation(res); 3419 } 3420 3421 //------------------------------next_monitor----------------------------------- 3422 // What number should be given to the next monitor? 3423 int GraphKit::next_monitor() { 3424 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3425 int next = current + C->sync_stack_slots(); 3426 // Keep the toplevel high water mark current: 3427 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3428 return current; 3429 } 3430 3431 //------------------------------insert_mem_bar--------------------------------- 3432 // Memory barrier to avoid floating things around 3433 // The membar serves as a pinch point between both control and all memory slices. 3434 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3435 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3436 mb->init_req(TypeFunc::Control, control()); 3437 mb->init_req(TypeFunc::Memory, reset_memory()); 3438 Node* membar = _gvn.transform(mb); 3439 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3440 set_all_memory_call(membar); 3441 return membar; 3442 } 3443 3444 //-------------------------insert_mem_bar_volatile---------------------------- 3445 // Memory barrier to avoid floating things around 3446 // The membar serves as a pinch point between both control and memory(alias_idx). 3447 // If you want to make a pinch point on all memory slices, do not use this 3448 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3449 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3450 // When Parse::do_put_xxx updates a volatile field, it appends a series 3451 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3452 // The first membar is on the same memory slice as the field store opcode. 3453 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3454 // All the other membars (for other volatile slices, including AliasIdxBot, 3455 // which stands for all unknown volatile slices) are control-dependent 3456 // on the first membar. This prevents later volatile loads or stores 3457 // from sliding up past the just-emitted store. 3458 3459 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3460 mb->set_req(TypeFunc::Control,control()); 3461 if (alias_idx == Compile::AliasIdxBot) { 3462 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3463 } else { 3464 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3465 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3466 } 3467 Node* membar = _gvn.transform(mb); 3468 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3469 if (alias_idx == Compile::AliasIdxBot) { 3470 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3471 } else { 3472 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3473 } 3474 return membar; 3475 } 3476 3477 //------------------------------shared_lock------------------------------------ 3478 // Emit locking code. 3479 FastLockNode* GraphKit::shared_lock(Node* obj) { 3480 // bci is either a monitorenter bc or InvocationEntryBci 3481 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3482 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3483 3484 if( !GenerateSynchronizationCode ) 3485 return nullptr; // Not locking things? 3486 if (stopped()) // Dead monitor? 3487 return nullptr; 3488 3489 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3490 3491 // Box the stack location 3492 Node* box = new BoxLockNode(next_monitor()); 3493 // Check for bailout after new BoxLockNode 3494 if (failing()) { return nullptr; } 3495 box = _gvn.transform(box); 3496 Node* mem = reset_memory(); 3497 3498 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3499 3500 // Create the rtm counters for this fast lock if needed. 3501 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3502 3503 // Add monitor to debug info for the slow path. If we block inside the 3504 // slow path and de-opt, we need the monitor hanging around 3505 map()->push_monitor( flock ); 3506 3507 const TypeFunc *tf = LockNode::lock_type(); 3508 LockNode *lock = new LockNode(C, tf); 3509 3510 lock->init_req( TypeFunc::Control, control() ); 3511 lock->init_req( TypeFunc::Memory , mem ); 3512 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3513 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3514 lock->init_req( TypeFunc::ReturnAdr, top() ); 3515 3516 lock->init_req(TypeFunc::Parms + 0, obj); 3517 lock->init_req(TypeFunc::Parms + 1, box); 3518 lock->init_req(TypeFunc::Parms + 2, flock); 3519 add_safepoint_edges(lock); 3520 3521 lock = _gvn.transform( lock )->as_Lock(); 3522 3523 // lock has no side-effects, sets few values 3524 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3525 3526 insert_mem_bar(Op_MemBarAcquireLock); 3527 3528 // Add this to the worklist so that the lock can be eliminated 3529 record_for_igvn(lock); 3530 3531 #ifndef PRODUCT 3532 if (PrintLockStatistics) { 3533 // Update the counter for this lock. Don't bother using an atomic 3534 // operation since we don't require absolute accuracy. 3535 lock->create_lock_counter(map()->jvms()); 3536 increment_counter(lock->counter()->addr()); 3537 } 3538 #endif 3539 3540 return flock; 3541 } 3542 3543 3544 //------------------------------shared_unlock---------------------------------- 3545 // Emit unlocking code. 3546 void GraphKit::shared_unlock(Node* box, Node* obj) { 3547 // bci is either a monitorenter bc or InvocationEntryBci 3548 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3549 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3550 3551 if( !GenerateSynchronizationCode ) 3552 return; 3553 if (stopped()) { // Dead monitor? 3554 map()->pop_monitor(); // Kill monitor from debug info 3555 return; 3556 } 3557 3558 // Memory barrier to avoid floating things down past the locked region 3559 insert_mem_bar(Op_MemBarReleaseLock); 3560 3561 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3562 UnlockNode *unlock = new UnlockNode(C, tf); 3563 #ifdef ASSERT 3564 unlock->set_dbg_jvms(sync_jvms()); 3565 #endif 3566 uint raw_idx = Compile::AliasIdxRaw; 3567 unlock->init_req( TypeFunc::Control, control() ); 3568 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3569 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3570 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3571 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3572 3573 unlock->init_req(TypeFunc::Parms + 0, obj); 3574 unlock->init_req(TypeFunc::Parms + 1, box); 3575 unlock = _gvn.transform(unlock)->as_Unlock(); 3576 3577 Node* mem = reset_memory(); 3578 3579 // unlock has no side-effects, sets few values 3580 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3581 3582 // Kill monitor from debug info 3583 map()->pop_monitor( ); 3584 } 3585 3586 //-------------------------------get_layout_helper----------------------------- 3587 // If the given klass is a constant or known to be an array, 3588 // fetch the constant layout helper value into constant_value 3589 // and return null. Otherwise, load the non-constant 3590 // layout helper value, and return the node which represents it. 3591 // This two-faced routine is useful because allocation sites 3592 // almost always feature constant types. 3593 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3594 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr(); 3595 if (!StressReflectiveCode && klass_t != nullptr) { 3596 bool xklass = klass_t->klass_is_exact(); 3597 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) { 3598 jint lhelper; 3599 if (klass_t->isa_aryklassptr()) { 3600 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type(); 3601 if (is_reference_type(elem, true)) { 3602 elem = T_OBJECT; 3603 } 3604 lhelper = Klass::array_layout_helper(elem); 3605 } else { 3606 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper(); 3607 } 3608 if (lhelper != Klass::_lh_neutral_value) { 3609 constant_value = lhelper; 3610 return (Node*) nullptr; 3611 } 3612 } 3613 } 3614 constant_value = Klass::_lh_neutral_value; // put in a known value 3615 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3616 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3617 } 3618 3619 // We just put in an allocate/initialize with a big raw-memory effect. 3620 // Hook selected additional alias categories on the initialization. 3621 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3622 MergeMemNode* init_in_merge, 3623 Node* init_out_raw) { 3624 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3625 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3626 3627 Node* prevmem = kit.memory(alias_idx); 3628 init_in_merge->set_memory_at(alias_idx, prevmem); 3629 kit.set_memory(init_out_raw, alias_idx); 3630 } 3631 3632 //---------------------------set_output_for_allocation------------------------- 3633 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3634 const TypeOopPtr* oop_type, 3635 bool deoptimize_on_exception) { 3636 int rawidx = Compile::AliasIdxRaw; 3637 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3638 add_safepoint_edges(alloc); 3639 Node* allocx = _gvn.transform(alloc); 3640 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3641 // create memory projection for i_o 3642 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3643 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3644 3645 // create a memory projection as for the normal control path 3646 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3647 set_memory(malloc, rawidx); 3648 3649 // a normal slow-call doesn't change i_o, but an allocation does 3650 // we create a separate i_o projection for the normal control path 3651 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3652 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3653 3654 // put in an initialization barrier 3655 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3656 rawoop)->as_Initialize(); 3657 assert(alloc->initialization() == init, "2-way macro link must work"); 3658 assert(init ->allocation() == alloc, "2-way macro link must work"); 3659 { 3660 // Extract memory strands which may participate in the new object's 3661 // initialization, and source them from the new InitializeNode. 3662 // This will allow us to observe initializations when they occur, 3663 // and link them properly (as a group) to the InitializeNode. 3664 assert(init->in(InitializeNode::Memory) == malloc, ""); 3665 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3666 init->set_req(InitializeNode::Memory, minit_in); 3667 record_for_igvn(minit_in); // fold it up later, if possible 3668 Node* minit_out = memory(rawidx); 3669 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3670 // Add an edge in the MergeMem for the header fields so an access 3671 // to one of those has correct memory state 3672 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 3673 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 3674 if (oop_type->isa_aryptr()) { 3675 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3676 int elemidx = C->get_alias_index(telemref); 3677 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3678 } else if (oop_type->isa_instptr()) { 3679 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass(); 3680 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3681 ciField* field = ik->nonstatic_field_at(i); 3682 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize) 3683 continue; // do not bother to track really large numbers of fields 3684 // Find (or create) the alias category for this field: 3685 int fieldidx = C->alias_type(field)->index(); 3686 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3687 } 3688 } 3689 } 3690 3691 // Cast raw oop to the real thing... 3692 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3693 javaoop = _gvn.transform(javaoop); 3694 C->set_recent_alloc(control(), javaoop); 3695 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3696 3697 #ifdef ASSERT 3698 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3699 assert(AllocateNode::Ideal_allocation(rawoop) == alloc, 3700 "Ideal_allocation works"); 3701 assert(AllocateNode::Ideal_allocation(javaoop) == alloc, 3702 "Ideal_allocation works"); 3703 if (alloc->is_AllocateArray()) { 3704 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(), 3705 "Ideal_allocation works"); 3706 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(), 3707 "Ideal_allocation works"); 3708 } else { 3709 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3710 } 3711 } 3712 #endif //ASSERT 3713 3714 return javaoop; 3715 } 3716 3717 //---------------------------new_instance-------------------------------------- 3718 // This routine takes a klass_node which may be constant (for a static type) 3719 // or may be non-constant (for reflective code). It will work equally well 3720 // for either, and the graph will fold nicely if the optimizer later reduces 3721 // the type to a constant. 3722 // The optional arguments are for specialized use by intrinsics: 3723 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3724 // - If 'return_size_val', report the total object size to the caller. 3725 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3726 Node* GraphKit::new_instance(Node* klass_node, 3727 Node* extra_slow_test, 3728 Node* *return_size_val, 3729 bool deoptimize_on_exception) { 3730 // Compute size in doublewords 3731 // The size is always an integral number of doublewords, represented 3732 // as a positive bytewise size stored in the klass's layout_helper. 3733 // The layout_helper also encodes (in a low bit) the need for a slow path. 3734 jint layout_con = Klass::_lh_neutral_value; 3735 Node* layout_val = get_layout_helper(klass_node, layout_con); 3736 int layout_is_con = (layout_val == nullptr); 3737 3738 if (extra_slow_test == nullptr) extra_slow_test = intcon(0); 3739 // Generate the initial go-slow test. It's either ALWAYS (return a 3740 // Node for 1) or NEVER (return a null) or perhaps (in the reflective 3741 // case) a computed value derived from the layout_helper. 3742 Node* initial_slow_test = nullptr; 3743 if (layout_is_con) { 3744 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3745 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3746 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3747 } else { // reflective case 3748 // This reflective path is used by Unsafe.allocateInstance. 3749 // (It may be stress-tested by specifying StressReflectiveCode.) 3750 // Basically, we want to get into the VM is there's an illegal argument. 3751 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3752 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3753 if (extra_slow_test != intcon(0)) { 3754 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3755 } 3756 // (Macro-expander will further convert this to a Bool, if necessary.) 3757 } 3758 3759 // Find the size in bytes. This is easy; it's the layout_helper. 3760 // The size value must be valid even if the slow path is taken. 3761 Node* size = nullptr; 3762 if (layout_is_con) { 3763 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3764 } else { // reflective case 3765 // This reflective path is used by clone and Unsafe.allocateInstance. 3766 size = ConvI2X(layout_val); 3767 3768 // Clear the low bits to extract layout_helper_size_in_bytes: 3769 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3770 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3771 size = _gvn.transform( new AndXNode(size, mask) ); 3772 } 3773 if (return_size_val != nullptr) { 3774 (*return_size_val) = size; 3775 } 3776 3777 // This is a precise notnull oop of the klass. 3778 // (Actually, it need not be precise if this is a reflective allocation.) 3779 // It's what we cast the result to. 3780 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3781 if (!tklass) tklass = TypeInstKlassPtr::OBJECT; 3782 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3783 3784 // Now generate allocation code 3785 3786 // The entire memory state is needed for slow path of the allocation 3787 // since GC and deoptimization can happened. 3788 Node *mem = reset_memory(); 3789 set_all_memory(mem); // Create new memory state 3790 3791 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3792 control(), mem, i_o(), 3793 size, klass_node, 3794 initial_slow_test); 3795 3796 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3797 } 3798 3799 //-------------------------------new_array------------------------------------- 3800 // helper for both newarray and anewarray 3801 // The 'length' parameter is (obviously) the length of the array. 3802 // The optional arguments are for specialized use by intrinsics: 3803 // - If 'return_size_val', report the non-padded array size (sum of header size 3804 // and array body) to the caller. 3805 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3806 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3807 Node* length, // number of array elements 3808 int nargs, // number of arguments to push back for uncommon trap 3809 Node* *return_size_val, 3810 bool deoptimize_on_exception) { 3811 jint layout_con = Klass::_lh_neutral_value; 3812 Node* layout_val = get_layout_helper(klass_node, layout_con); 3813 int layout_is_con = (layout_val == nullptr); 3814 3815 if (!layout_is_con && !StressReflectiveCode && 3816 !too_many_traps(Deoptimization::Reason_class_check)) { 3817 // This is a reflective array creation site. 3818 // Optimistically assume that it is a subtype of Object[], 3819 // so that we can fold up all the address arithmetic. 3820 layout_con = Klass::array_layout_helper(T_OBJECT); 3821 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3822 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3823 { BuildCutout unless(this, bol_lh, PROB_MAX); 3824 inc_sp(nargs); 3825 uncommon_trap(Deoptimization::Reason_class_check, 3826 Deoptimization::Action_maybe_recompile); 3827 } 3828 layout_val = nullptr; 3829 layout_is_con = true; 3830 } 3831 3832 // Generate the initial go-slow test. Make sure we do not overflow 3833 // if length is huge (near 2Gig) or negative! We do not need 3834 // exact double-words here, just a close approximation of needed 3835 // double-words. We can't add any offset or rounding bits, lest we 3836 // take a size -1 of bytes and make it positive. Use an unsigned 3837 // compare, so negative sizes look hugely positive. 3838 int fast_size_limit = FastAllocateSizeLimit; 3839 if (layout_is_con) { 3840 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3841 // Increase the size limit if we have exact knowledge of array type. 3842 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3843 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3844 } 3845 3846 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3847 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3848 3849 // --- Size Computation --- 3850 // array_size = round_to_heap(array_header + (length << elem_shift)); 3851 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3852 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3853 // The rounding mask is strength-reduced, if possible. 3854 int round_mask = MinObjAlignmentInBytes - 1; 3855 Node* header_size = nullptr; 3856 // (T_BYTE has the weakest alignment and size restrictions...) 3857 if (layout_is_con) { 3858 int hsize = Klass::layout_helper_header_size(layout_con); 3859 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3860 if ((round_mask & ~right_n_bits(eshift)) == 0) 3861 round_mask = 0; // strength-reduce it if it goes away completely 3862 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3863 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3864 assert(header_size_min <= hsize, "generic minimum is smallest"); 3865 header_size = intcon(hsize); 3866 } else { 3867 Node* hss = intcon(Klass::_lh_header_size_shift); 3868 Node* hsm = intcon(Klass::_lh_header_size_mask); 3869 header_size = _gvn.transform(new URShiftINode(layout_val, hss)); 3870 header_size = _gvn.transform(new AndINode(header_size, hsm)); 3871 } 3872 3873 Node* elem_shift = nullptr; 3874 if (layout_is_con) { 3875 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3876 if (eshift != 0) 3877 elem_shift = intcon(eshift); 3878 } else { 3879 // There is no need to mask or shift this value. 3880 // The semantics of LShiftINode include an implicit mask to 0x1F. 3881 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3882 elem_shift = layout_val; 3883 } 3884 3885 // Transition to native address size for all offset calculations: 3886 Node* lengthx = ConvI2X(length); 3887 Node* headerx = ConvI2X(header_size); 3888 #ifdef _LP64 3889 { const TypeInt* tilen = _gvn.find_int_type(length); 3890 if (tilen != nullptr && tilen->_lo < 0) { 3891 // Add a manual constraint to a positive range. Cf. array_element_address. 3892 jint size_max = fast_size_limit; 3893 if (size_max > tilen->_hi) size_max = tilen->_hi; 3894 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3895 3896 // Only do a narrow I2L conversion if the range check passed. 3897 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3898 _gvn.transform(iff); 3899 RegionNode* region = new RegionNode(3); 3900 _gvn.set_type(region, Type::CONTROL); 3901 lengthx = new PhiNode(region, TypeLong::LONG); 3902 _gvn.set_type(lengthx, TypeLong::LONG); 3903 3904 // Range check passed. Use ConvI2L node with narrow type. 3905 Node* passed = IfFalse(iff); 3906 region->init_req(1, passed); 3907 // Make I2L conversion control dependent to prevent it from 3908 // floating above the range check during loop optimizations. 3909 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3910 3911 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3912 region->init_req(2, IfTrue(iff)); 3913 lengthx->init_req(2, ConvI2X(length)); 3914 3915 set_control(region); 3916 record_for_igvn(region); 3917 record_for_igvn(lengthx); 3918 } 3919 } 3920 #endif 3921 3922 // Combine header size and body size for the array copy part, then align (if 3923 // necessary) for the allocation part. This computation cannot overflow, 3924 // because it is used only in two places, one where the length is sharply 3925 // limited, and the other after a successful allocation. 3926 Node* abody = lengthx; 3927 if (elem_shift != nullptr) { 3928 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift)); 3929 } 3930 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody)); 3931 3932 if (return_size_val != nullptr) { 3933 // This is the size 3934 (*return_size_val) = non_rounded_size; 3935 } 3936 3937 Node* size = non_rounded_size; 3938 if (round_mask != 0) { 3939 Node* mask1 = MakeConX(round_mask); 3940 size = _gvn.transform(new AddXNode(size, mask1)); 3941 Node* mask2 = MakeConX(~round_mask); 3942 size = _gvn.transform(new AndXNode(size, mask2)); 3943 } 3944 // else if round_mask == 0, the size computation is self-rounding 3945 3946 // Now generate allocation code 3947 3948 // The entire memory state is needed for slow path of the allocation 3949 // since GC and deoptimization can happened. 3950 Node *mem = reset_memory(); 3951 set_all_memory(mem); // Create new memory state 3952 3953 if (initial_slow_test->is_Bool()) { 3954 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3955 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3956 } 3957 3958 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3959 Node* valid_length_test = _gvn.intcon(1); 3960 if (ary_type->isa_aryptr()) { 3961 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type(); 3962 jint max = TypeAryPtr::max_array_length(bt); 3963 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max))); 3964 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le)); 3965 } 3966 3967 // Create the AllocateArrayNode and its result projections 3968 AllocateArrayNode* alloc 3969 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3970 control(), mem, i_o(), 3971 size, klass_node, 3972 initial_slow_test, 3973 length, valid_length_test); 3974 3975 // Cast to correct type. Note that the klass_node may be constant or not, 3976 // and in the latter case the actual array type will be inexact also. 3977 // (This happens via a non-constant argument to inline_native_newArray.) 3978 // In any case, the value of klass_node provides the desired array type. 3979 const TypeInt* length_type = _gvn.find_int_type(length); 3980 if (ary_type->isa_aryptr() && length_type != nullptr) { 3981 // Try to get a better type than POS for the size 3982 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3983 } 3984 3985 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3986 3987 array_ideal_length(alloc, ary_type, true); 3988 return javaoop; 3989 } 3990 3991 // The following "Ideal_foo" functions are placed here because they recognize 3992 // the graph shapes created by the functions immediately above. 3993 3994 //---------------------------Ideal_allocation---------------------------------- 3995 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3996 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) { 3997 if (ptr == nullptr) { // reduce dumb test in callers 3998 return nullptr; 3999 } 4000 4001 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4002 ptr = bs->step_over_gc_barrier(ptr); 4003 4004 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4005 ptr = ptr->in(1); 4006 if (ptr == nullptr) return nullptr; 4007 } 4008 // Return null for allocations with several casts: 4009 // j.l.reflect.Array.newInstance(jobject, jint) 4010 // Object.clone() 4011 // to keep more precise type from last cast. 4012 if (ptr->is_Proj()) { 4013 Node* allo = ptr->in(0); 4014 if (allo != nullptr && allo->is_Allocate()) { 4015 return allo->as_Allocate(); 4016 } 4017 } 4018 // Report failure to match. 4019 return nullptr; 4020 } 4021 4022 // Fancy version which also strips off an offset (and reports it to caller). 4023 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase, 4024 intptr_t& offset) { 4025 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4026 if (base == nullptr) return nullptr; 4027 return Ideal_allocation(base); 4028 } 4029 4030 // Trace Initialize <- Proj[Parm] <- Allocate 4031 AllocateNode* InitializeNode::allocation() { 4032 Node* rawoop = in(InitializeNode::RawAddress); 4033 if (rawoop->is_Proj()) { 4034 Node* alloc = rawoop->in(0); 4035 if (alloc->is_Allocate()) { 4036 return alloc->as_Allocate(); 4037 } 4038 } 4039 return nullptr; 4040 } 4041 4042 // Trace Allocate -> Proj[Parm] -> Initialize 4043 InitializeNode* AllocateNode::initialization() { 4044 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4045 if (rawoop == nullptr) return nullptr; 4046 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4047 Node* init = rawoop->fast_out(i); 4048 if (init->is_Initialize()) { 4049 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4050 return init->as_Initialize(); 4051 } 4052 } 4053 return nullptr; 4054 } 4055 4056 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path. 4057 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) { 4058 // Too many traps seen? 4059 if (too_many_traps(reason)) { 4060 #ifdef ASSERT 4061 if (TraceLoopPredicate) { 4062 int tc = C->trap_count(reason); 4063 tty->print("too many traps=%s tcount=%d in ", 4064 Deoptimization::trap_reason_name(reason), tc); 4065 method()->print(); // which method has too many predicate traps 4066 tty->cr(); 4067 } 4068 #endif 4069 // We cannot afford to take more traps here, 4070 // do not generate Parse Predicate. 4071 return; 4072 } 4073 4074 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn); 4075 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn)); 4076 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate)); 4077 { 4078 PreserveJVMState pjvms(this); 4079 set_control(if_false); 4080 inc_sp(nargs); 4081 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4082 } 4083 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate)); 4084 set_control(if_true); 4085 } 4086 4087 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All 4088 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate. 4089 void GraphKit::add_parse_predicates(int nargs) { 4090 if (UseLoopPredicate) { 4091 add_parse_predicate(Deoptimization::Reason_predicate, nargs); 4092 } 4093 if (UseProfiledLoopPredicate) { 4094 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs); 4095 } 4096 // Loop Limit Check Predicate should be near the loop. 4097 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs); 4098 } 4099 4100 void GraphKit::sync_kit(IdealKit& ideal) { 4101 set_all_memory(ideal.merged_memory()); 4102 set_i_o(ideal.i_o()); 4103 set_control(ideal.ctrl()); 4104 } 4105 4106 void GraphKit::final_sync(IdealKit& ideal) { 4107 // Final sync IdealKit and graphKit. 4108 sync_kit(ideal); 4109 } 4110 4111 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4112 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4113 Node* coder = load_String_coder(str, set_ctrl); 4114 // Divide length by 2 if coder is UTF16 4115 return _gvn.transform(new RShiftINode(len, coder)); 4116 } 4117 4118 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4119 int value_offset = java_lang_String::value_offset(); 4120 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4121 false, nullptr, 0); 4122 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4123 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4124 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4125 ciTypeArrayKlass::make(T_BYTE), true, 0); 4126 Node* p = basic_plus_adr(str, str, value_offset); 4127 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4128 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4129 return load; 4130 } 4131 4132 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4133 if (!CompactStrings) { 4134 return intcon(java_lang_String::CODER_UTF16); 4135 } 4136 int coder_offset = java_lang_String::coder_offset(); 4137 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4138 false, nullptr, 0); 4139 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4140 4141 Node* p = basic_plus_adr(str, str, coder_offset); 4142 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4143 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4144 return load; 4145 } 4146 4147 void GraphKit::store_String_value(Node* str, Node* value) { 4148 int value_offset = java_lang_String::value_offset(); 4149 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4150 false, nullptr, 0); 4151 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4152 4153 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4154 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4155 } 4156 4157 void GraphKit::store_String_coder(Node* str, Node* value) { 4158 int coder_offset = java_lang_String::coder_offset(); 4159 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4160 false, nullptr, 0); 4161 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4162 4163 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4164 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4165 } 4166 4167 // Capture src and dst memory state with a MergeMemNode 4168 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4169 if (src_type == dst_type) { 4170 // Types are equal, we don't need a MergeMemNode 4171 return memory(src_type); 4172 } 4173 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4174 record_for_igvn(merge); // fold it up later, if possible 4175 int src_idx = C->get_alias_index(src_type); 4176 int dst_idx = C->get_alias_index(dst_type); 4177 merge->set_memory_at(src_idx, memory(src_idx)); 4178 merge->set_memory_at(dst_idx, memory(dst_idx)); 4179 return merge; 4180 } 4181 4182 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4183 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4184 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4185 // If input and output memory types differ, capture both states to preserve 4186 // the dependency between preceding and subsequent loads/stores. 4187 // For example, the following program: 4188 // StoreB 4189 // compress_string 4190 // LoadB 4191 // has this memory graph (use->def): 4192 // LoadB -> compress_string -> CharMem 4193 // ... -> StoreB -> ByteMem 4194 // The intrinsic hides the dependency between LoadB and StoreB, causing 4195 // the load to read from memory not containing the result of the StoreB. 4196 // The correct memory graph should look like this: 4197 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4198 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4199 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4200 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str))); 4201 set_memory(res_mem, TypeAryPtr::BYTES); 4202 return str; 4203 } 4204 4205 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4206 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4207 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4208 // Capture src and dst memory (see comment in 'compress_string'). 4209 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4210 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4211 set_memory(_gvn.transform(str), dst_type); 4212 } 4213 4214 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4215 /** 4216 * int i_char = start; 4217 * for (int i_byte = 0; i_byte < count; i_byte++) { 4218 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4219 * } 4220 */ 4221 add_parse_predicates(); 4222 C->set_has_loops(true); 4223 4224 RegionNode* head = new RegionNode(3); 4225 head->init_req(1, control()); 4226 gvn().set_type(head, Type::CONTROL); 4227 record_for_igvn(head); 4228 4229 Node* i_byte = new PhiNode(head, TypeInt::INT); 4230 i_byte->init_req(1, intcon(0)); 4231 gvn().set_type(i_byte, TypeInt::INT); 4232 record_for_igvn(i_byte); 4233 4234 Node* i_char = new PhiNode(head, TypeInt::INT); 4235 i_char->init_req(1, start); 4236 gvn().set_type(i_char, TypeInt::INT); 4237 record_for_igvn(i_char); 4238 4239 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4240 gvn().set_type(mem, Type::MEMORY); 4241 record_for_igvn(mem); 4242 set_control(head); 4243 set_memory(mem, TypeAryPtr::BYTES); 4244 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true); 4245 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4246 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4247 false, false, true /* mismatched */); 4248 4249 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4250 head->init_req(2, IfTrue(iff)); 4251 mem->init_req(2, st); 4252 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4253 i_char->init_req(2, AddI(i_char, intcon(2))); 4254 4255 set_control(IfFalse(iff)); 4256 set_memory(st, TypeAryPtr::BYTES); 4257 } 4258 4259 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4260 if (!field->is_constant()) { 4261 return nullptr; // Field not marked as constant. 4262 } 4263 ciInstance* holder = nullptr; 4264 if (!field->is_static()) { 4265 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4266 if (const_oop != nullptr && const_oop->is_instance()) { 4267 holder = const_oop->as_instance(); 4268 } 4269 } 4270 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4271 /*is_unsigned_load=*/false); 4272 if (con_type != nullptr) { 4273 return makecon(con_type); 4274 } 4275 return nullptr; 4276 } 4277 4278 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) { 4279 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr(); 4280 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type); 4281 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) { 4282 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part 4283 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type)); 4284 return casted_obj; 4285 } 4286 return obj; 4287 }