1 /* 2 * Copyright (c) 2001, 2023, 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 PEAState& as = youngest_jvms->alloc_state(); 1032 backfill_materialized(call, TypeFunc::Parms, call->req(), as); 1033 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1034 1035 // Test the correctness of JVMState::debug_xxx accessors: 1036 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1037 assert(call->jvms()->debug_end() == call->req(), ""); 1038 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1039 } 1040 1041 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1042 Bytecodes::Code code = java_bc(); 1043 if (code == Bytecodes::_wide) { 1044 code = method()->java_code_at_bci(bci() + 1); 1045 } 1046 1047 if (code != Bytecodes::_illegal) { 1048 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1049 } 1050 1051 auto rsize = [&]() { 1052 assert(code != Bytecodes::_illegal, "code is illegal!"); 1053 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1054 return (rtype < T_CONFLICT) ? type2size[rtype] : 0; 1055 }; 1056 1057 switch (code) { 1058 case Bytecodes::_illegal: 1059 return false; 1060 1061 case Bytecodes::_ldc: 1062 case Bytecodes::_ldc_w: 1063 case Bytecodes::_ldc2_w: 1064 inputs = 0; 1065 break; 1066 1067 case Bytecodes::_dup: inputs = 1; break; 1068 case Bytecodes::_dup_x1: inputs = 2; break; 1069 case Bytecodes::_dup_x2: inputs = 3; break; 1070 case Bytecodes::_dup2: inputs = 2; break; 1071 case Bytecodes::_dup2_x1: inputs = 3; break; 1072 case Bytecodes::_dup2_x2: inputs = 4; break; 1073 case Bytecodes::_swap: inputs = 2; break; 1074 case Bytecodes::_arraylength: inputs = 1; break; 1075 1076 case Bytecodes::_getstatic: 1077 case Bytecodes::_putstatic: 1078 case Bytecodes::_getfield: 1079 case Bytecodes::_putfield: 1080 { 1081 bool ignored_will_link; 1082 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1083 int size = field->type()->size(); 1084 bool is_get = (depth >= 0), is_static = (depth & 1); 1085 inputs = (is_static ? 0 : 1); 1086 if (is_get) { 1087 depth = size - inputs; 1088 } else { 1089 inputs += size; // putxxx pops the value from the stack 1090 depth = - inputs; 1091 } 1092 } 1093 break; 1094 1095 case Bytecodes::_invokevirtual: 1096 case Bytecodes::_invokespecial: 1097 case Bytecodes::_invokestatic: 1098 case Bytecodes::_invokedynamic: 1099 case Bytecodes::_invokeinterface: 1100 { 1101 bool ignored_will_link; 1102 ciSignature* declared_signature = nullptr; 1103 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1104 assert(declared_signature != nullptr, "cannot be null"); 1105 inputs = declared_signature->arg_size_for_bc(code); 1106 int size = declared_signature->return_type()->size(); 1107 depth = size - inputs; 1108 } 1109 break; 1110 1111 case Bytecodes::_multianewarray: 1112 { 1113 ciBytecodeStream iter(method()); 1114 iter.reset_to_bci(bci()); 1115 iter.next(); 1116 inputs = iter.get_dimensions(); 1117 assert(rsize() == 1, ""); 1118 depth = 1 - inputs; 1119 } 1120 break; 1121 1122 case Bytecodes::_ireturn: 1123 case Bytecodes::_lreturn: 1124 case Bytecodes::_freturn: 1125 case Bytecodes::_dreturn: 1126 case Bytecodes::_areturn: 1127 assert(rsize() == -depth, ""); 1128 inputs = -depth; 1129 break; 1130 1131 case Bytecodes::_jsr: 1132 case Bytecodes::_jsr_w: 1133 inputs = 0; 1134 depth = 1; // S.B. depth=1, not zero 1135 break; 1136 1137 default: 1138 // bytecode produces a typed result 1139 inputs = rsize() - depth; 1140 assert(inputs >= 0, ""); 1141 break; 1142 } 1143 1144 #ifdef ASSERT 1145 // spot check 1146 int outputs = depth + inputs; 1147 assert(outputs >= 0, "sanity"); 1148 switch (code) { 1149 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1150 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1151 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1152 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1153 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1154 default: break; 1155 } 1156 #endif //ASSERT 1157 1158 return true; 1159 } 1160 1161 1162 1163 //------------------------------basic_plus_adr--------------------------------- 1164 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1165 // short-circuit a common case 1166 if (offset == intcon(0)) return ptr; 1167 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1168 } 1169 1170 Node* GraphKit::ConvI2L(Node* offset) { 1171 // short-circuit a common case 1172 jint offset_con = find_int_con(offset, Type::OffsetBot); 1173 if (offset_con != Type::OffsetBot) { 1174 return longcon((jlong) offset_con); 1175 } 1176 return _gvn.transform( new ConvI2LNode(offset)); 1177 } 1178 1179 Node* GraphKit::ConvI2UL(Node* offset) { 1180 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1181 if (offset_con != (juint) Type::OffsetBot) { 1182 return longcon((julong) offset_con); 1183 } 1184 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1185 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1186 return _gvn.transform( new AndLNode(conv, mask) ); 1187 } 1188 1189 Node* GraphKit::ConvL2I(Node* offset) { 1190 // short-circuit a common case 1191 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1192 if (offset_con != (jlong)Type::OffsetBot) { 1193 return intcon((int) offset_con); 1194 } 1195 return _gvn.transform( new ConvL2INode(offset)); 1196 } 1197 1198 //-------------------------load_object_klass----------------------------------- 1199 Node* GraphKit::load_object_klass(Node* obj) { 1200 // Special-case a fresh allocation to avoid building nodes: 1201 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1202 if (akls != nullptr) return akls; 1203 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1204 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1205 } 1206 1207 //-------------------------load_array_length----------------------------------- 1208 Node* GraphKit::load_array_length(Node* array) { 1209 // Special-case a fresh allocation to avoid building nodes: 1210 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array); 1211 Node *alen; 1212 if (alloc == nullptr) { 1213 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1214 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1215 } else { 1216 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false); 1217 } 1218 return alen; 1219 } 1220 1221 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc, 1222 const TypeOopPtr* oop_type, 1223 bool replace_length_in_map) { 1224 Node* length = alloc->Ideal_length(); 1225 if (replace_length_in_map == false || map()->find_edge(length) >= 0) { 1226 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn); 1227 if (ccast != length) { 1228 // do not transform ccast here, it might convert to top node for 1229 // negative array length and break assumptions in parsing stage. 1230 _gvn.set_type_bottom(ccast); 1231 record_for_igvn(ccast); 1232 if (replace_length_in_map) { 1233 replace_in_map(length, ccast); 1234 } 1235 return ccast; 1236 } 1237 } 1238 return length; 1239 } 1240 1241 //------------------------------do_null_check---------------------------------- 1242 // Helper function to do a null pointer check. Returned value is 1243 // the incoming address with null casted away. You are allowed to use the 1244 // not-null value only if you are control dependent on the test. 1245 #ifndef PRODUCT 1246 extern uint explicit_null_checks_inserted, 1247 explicit_null_checks_elided; 1248 #endif 1249 Node* GraphKit::null_check_common(Node* value, BasicType type, 1250 // optional arguments for variations: 1251 bool assert_null, 1252 Node* *null_control, 1253 bool speculative) { 1254 assert(!assert_null || null_control == nullptr, "not both at once"); 1255 if (stopped()) return top(); 1256 NOT_PRODUCT(explicit_null_checks_inserted++); 1257 1258 // Construct null check 1259 Node *chk = nullptr; 1260 switch(type) { 1261 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1262 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1263 case T_ARRAY : // fall through 1264 type = T_OBJECT; // simplify further tests 1265 case T_OBJECT : { 1266 const Type *t = _gvn.type( value ); 1267 1268 const TypeOopPtr* tp = t->isa_oopptr(); 1269 if (tp != nullptr && !tp->is_loaded() 1270 // Only for do_null_check, not any of its siblings: 1271 && !assert_null && null_control == nullptr) { 1272 // Usually, any field access or invocation on an unloaded oop type 1273 // will simply fail to link, since the statically linked class is 1274 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1275 // the static class is loaded but the sharper oop type is not. 1276 // Rather than checking for this obscure case in lots of places, 1277 // we simply observe that a null check on an unloaded class 1278 // will always be followed by a nonsense operation, so we 1279 // can just issue the uncommon trap here. 1280 // Our access to the unloaded class will only be correct 1281 // after it has been loaded and initialized, which requires 1282 // a trip through the interpreter. 1283 ciKlass* klass = tp->unloaded_klass(); 1284 #ifndef PRODUCT 1285 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); } 1286 #endif 1287 uncommon_trap(Deoptimization::Reason_unloaded, 1288 Deoptimization::Action_reinterpret, 1289 klass, "!loaded"); 1290 return top(); 1291 } 1292 1293 if (assert_null) { 1294 // See if the type is contained in NULL_PTR. 1295 // If so, then the value is already null. 1296 if (t->higher_equal(TypePtr::NULL_PTR)) { 1297 NOT_PRODUCT(explicit_null_checks_elided++); 1298 return value; // Elided null assert quickly! 1299 } 1300 } else { 1301 // See if mixing in the null pointer changes type. 1302 // If so, then the null pointer was not allowed in the original 1303 // type. In other words, "value" was not-null. 1304 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1305 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1306 NOT_PRODUCT(explicit_null_checks_elided++); 1307 return value; // Elided null check quickly! 1308 } 1309 } 1310 chk = new CmpPNode( value, null() ); 1311 break; 1312 } 1313 1314 default: 1315 fatal("unexpected type: %s", type2name(type)); 1316 } 1317 assert(chk != nullptr, "sanity check"); 1318 chk = _gvn.transform(chk); 1319 1320 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1321 BoolNode *btst = new BoolNode( chk, btest); 1322 Node *tst = _gvn.transform( btst ); 1323 1324 //----------- 1325 // if peephole optimizations occurred, a prior test existed. 1326 // If a prior test existed, maybe it dominates as we can avoid this test. 1327 if (tst != btst && type == T_OBJECT) { 1328 // At this point we want to scan up the CFG to see if we can 1329 // find an identical test (and so avoid this test altogether). 1330 Node *cfg = control(); 1331 int depth = 0; 1332 while( depth < 16 ) { // Limit search depth for speed 1333 if( cfg->Opcode() == Op_IfTrue && 1334 cfg->in(0)->in(1) == tst ) { 1335 // Found prior test. Use "cast_not_null" to construct an identical 1336 // CastPP (and hence hash to) as already exists for the prior test. 1337 // Return that casted value. 1338 if (assert_null) { 1339 replace_in_map(value, null()); 1340 return null(); // do not issue the redundant test 1341 } 1342 Node *oldcontrol = control(); 1343 set_control(cfg); 1344 Node *res = cast_not_null(value); 1345 set_control(oldcontrol); 1346 NOT_PRODUCT(explicit_null_checks_elided++); 1347 return res; 1348 } 1349 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1350 if (cfg == nullptr) break; // Quit at region nodes 1351 depth++; 1352 } 1353 } 1354 1355 //----------- 1356 // Branch to failure if null 1357 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1358 Deoptimization::DeoptReason reason; 1359 if (assert_null) { 1360 reason = Deoptimization::reason_null_assert(speculative); 1361 } else if (type == T_OBJECT) { 1362 reason = Deoptimization::reason_null_check(speculative); 1363 } else { 1364 reason = Deoptimization::Reason_div0_check; 1365 } 1366 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1367 // ciMethodData::has_trap_at will return a conservative -1 if any 1368 // must-be-null assertion has failed. This could cause performance 1369 // problems for a method after its first do_null_assert failure. 1370 // Consider using 'Reason_class_check' instead? 1371 1372 // To cause an implicit null check, we set the not-null probability 1373 // to the maximum (PROB_MAX). For an explicit check the probability 1374 // is set to a smaller value. 1375 if (null_control != nullptr || too_many_traps(reason)) { 1376 // probability is less likely 1377 ok_prob = PROB_LIKELY_MAG(3); 1378 } else if (!assert_null && 1379 (ImplicitNullCheckThreshold > 0) && 1380 method() != nullptr && 1381 (method()->method_data()->trap_count(reason) 1382 >= (uint)ImplicitNullCheckThreshold)) { 1383 ok_prob = PROB_LIKELY_MAG(3); 1384 } 1385 1386 if (null_control != nullptr) { 1387 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1388 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1389 set_control( _gvn.transform( new IfTrueNode(iff))); 1390 #ifndef PRODUCT 1391 if (null_true == top()) { 1392 explicit_null_checks_elided++; 1393 } 1394 #endif 1395 (*null_control) = null_true; 1396 } else { 1397 BuildCutout unless(this, tst, ok_prob); 1398 // Check for optimizer eliding test at parse time 1399 if (stopped()) { 1400 // Failure not possible; do not bother making uncommon trap. 1401 NOT_PRODUCT(explicit_null_checks_elided++); 1402 } else if (assert_null) { 1403 uncommon_trap(reason, 1404 Deoptimization::Action_make_not_entrant, 1405 nullptr, "assert_null"); 1406 } else { 1407 replace_in_map(value, zerocon(type)); 1408 builtin_throw(reason); 1409 } 1410 } 1411 1412 // Must throw exception, fall-thru not possible? 1413 if (stopped()) { 1414 return top(); // No result 1415 } 1416 1417 if (assert_null) { 1418 // Cast obj to null on this path. 1419 replace_in_map(value, zerocon(type)); 1420 return zerocon(type); 1421 } 1422 1423 // Cast obj to not-null on this path, if there is no null_control. 1424 // (If there is a null_control, a non-null value may come back to haunt us.) 1425 if (type == T_OBJECT) { 1426 Node* cast = cast_not_null(value, false); 1427 if (null_control == nullptr || (*null_control) == top()) 1428 replace_in_map(value, cast); 1429 value = cast; 1430 } 1431 1432 return value; 1433 } 1434 1435 1436 //------------------------------cast_not_null---------------------------------- 1437 // Cast obj to not-null on this path 1438 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1439 const Type *t = _gvn.type(obj); 1440 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1441 // Object is already not-null? 1442 if( t == t_not_null ) return obj; 1443 1444 Node* cast = new CastPPNode(control(), obj,t_not_null); 1445 cast = _gvn.transform( cast ); 1446 1447 // Scan for instances of 'obj' in the current JVM mapping. 1448 // These instances are known to be not-null after the test. 1449 if (do_replace_in_map) 1450 replace_in_map(obj, cast); 1451 1452 return cast; // Return casted value 1453 } 1454 1455 // Sometimes in intrinsics, we implicitly know an object is not null 1456 // (there's no actual null check) so we can cast it to not null. In 1457 // the course of optimizations, the input to the cast can become null. 1458 // In that case that data path will die and we need the control path 1459 // to become dead as well to keep the graph consistent. So we have to 1460 // add a check for null for which one branch can't be taken. It uses 1461 // an Opaque4 node that will cause the check to be removed after loop 1462 // opts so the test goes away and the compiled code doesn't execute a 1463 // useless check. 1464 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1465 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1466 return value; 1467 } 1468 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1469 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1470 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1471 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1472 _gvn.set_type(iff, iff->Value(&_gvn)); 1473 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1474 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1475 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1476 C->root()->add_req(halt); 1477 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1478 set_control(if_t); 1479 return cast_not_null(value, do_replace_in_map); 1480 } 1481 1482 1483 //--------------------------replace_in_map------------------------------------- 1484 void GraphKit::replace_in_map(Node* old, Node* neww) { 1485 if (old == neww) { 1486 return; 1487 } 1488 1489 map()->replace_edge(old, neww); 1490 1491 // Note: This operation potentially replaces any edge 1492 // on the map. This includes locals, stack, and monitors 1493 // of the current (innermost) JVM state. 1494 1495 // don't let inconsistent types from profiling escape this 1496 // method 1497 1498 const Type* told = _gvn.type(old); 1499 const Type* tnew = _gvn.type(neww); 1500 1501 if (!tnew->higher_equal(told)) { 1502 return; 1503 } 1504 1505 map()->record_replaced_node(old, neww); 1506 } 1507 1508 1509 //============================================================================= 1510 //--------------------------------memory--------------------------------------- 1511 Node* GraphKit::memory(uint alias_idx) { 1512 MergeMemNode* mem = merged_memory(); 1513 Node* p = mem->memory_at(alias_idx); 1514 assert(p != mem->empty_memory(), "empty"); 1515 _gvn.set_type(p, Type::MEMORY); // must be mapped 1516 return p; 1517 } 1518 1519 //-----------------------------reset_memory------------------------------------ 1520 Node* GraphKit::reset_memory() { 1521 Node* mem = map()->memory(); 1522 // do not use this node for any more parsing! 1523 debug_only( map()->set_memory((Node*)nullptr) ); 1524 return _gvn.transform( mem ); 1525 } 1526 1527 //------------------------------set_all_memory--------------------------------- 1528 void GraphKit::set_all_memory(Node* newmem) { 1529 Node* mergemem = MergeMemNode::make(newmem); 1530 gvn().set_type_bottom(mergemem); 1531 map()->set_memory(mergemem); 1532 } 1533 1534 //------------------------------set_all_memory_call---------------------------- 1535 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1536 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1537 set_all_memory(newmem); 1538 } 1539 1540 //============================================================================= 1541 // 1542 // parser factory methods for MemNodes 1543 // 1544 // These are layered on top of the factory methods in LoadNode and StoreNode, 1545 // and integrate with the parser's memory state and _gvn engine. 1546 // 1547 1548 // factory methods in "int adr_idx" 1549 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1550 int adr_idx, 1551 MemNode::MemOrd mo, 1552 LoadNode::ControlDependency control_dependency, 1553 bool require_atomic_access, 1554 bool unaligned, 1555 bool mismatched, 1556 bool unsafe, 1557 uint8_t barrier_data) { 1558 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1559 const TypePtr* adr_type = nullptr; // debug-mode-only argument 1560 debug_only(adr_type = C->get_adr_type(adr_idx)); 1561 Node* mem = memory(adr_idx); 1562 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data); 1563 ld = _gvn.transform(ld); 1564 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1565 // Improve graph before escape analysis and boxing elimination. 1566 record_for_igvn(ld); 1567 if (ld->is_DecodeN()) { 1568 // Also record the actual load (LoadN) in case ld is DecodeN 1569 assert(ld->in(1)->Opcode() == Op_LoadN, "Assumption invalid: input to DecodeN is not LoadN"); 1570 record_for_igvn(ld->in(1)); 1571 } 1572 } 1573 return ld; 1574 } 1575 1576 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1577 int adr_idx, 1578 MemNode::MemOrd mo, 1579 bool require_atomic_access, 1580 bool unaligned, 1581 bool mismatched, 1582 bool unsafe, 1583 int barrier_data) { 1584 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1585 const TypePtr* adr_type = nullptr; 1586 debug_only(adr_type = C->get_adr_type(adr_idx)); 1587 Node *mem = memory(adr_idx); 1588 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access); 1589 if (unaligned) { 1590 st->as_Store()->set_unaligned_access(); 1591 } 1592 if (mismatched) { 1593 st->as_Store()->set_mismatched_access(); 1594 } 1595 if (unsafe) { 1596 st->as_Store()->set_unsafe_access(); 1597 } 1598 st->as_Store()->set_barrier_data(barrier_data); 1599 st = _gvn.transform(st); 1600 set_memory(st, adr_idx); 1601 // Back-to-back stores can only remove intermediate store with DU info 1602 // so push on worklist for optimizer. 1603 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1604 record_for_igvn(st); 1605 1606 return st; 1607 } 1608 1609 Node* GraphKit::access_store_at(Node* obj, 1610 Node* adr, 1611 const TypePtr* adr_type, 1612 Node* val, 1613 const Type* val_type, 1614 BasicType bt, 1615 DecoratorSet decorators) { 1616 // Transformation of a value which could be null pointer (CastPP #null) 1617 // could be delayed during Parse (for example, in adjust_map_after_if()). 1618 // Execute transformation here to avoid barrier generation in such case. 1619 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1620 val = _gvn.makecon(TypePtr::NULL_PTR); 1621 } 1622 1623 if (stopped()) { 1624 return top(); // Dead path ? 1625 } 1626 1627 assert(val != nullptr, "not dead path"); 1628 1629 C2AccessValuePtr addr(adr, adr_type); 1630 C2AccessValue value(val, val_type); 1631 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1632 if (access.is_raw()) { 1633 return _barrier_set->BarrierSetC2::store_at(access, value); 1634 } else { 1635 return _barrier_set->store_at(access, value); 1636 } 1637 } 1638 1639 Node* GraphKit::access_load_at(Node* obj, // containing obj 1640 Node* adr, // actual address to store val at 1641 const TypePtr* adr_type, 1642 const Type* val_type, 1643 BasicType bt, 1644 DecoratorSet decorators) { 1645 if (stopped()) { 1646 return top(); // Dead path ? 1647 } 1648 1649 C2AccessValuePtr addr(adr, adr_type); 1650 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr); 1651 if (access.is_raw()) { 1652 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1653 } else { 1654 return _barrier_set->load_at(access, val_type); 1655 } 1656 } 1657 1658 Node* GraphKit::access_load(Node* adr, // actual address to load val at 1659 const Type* val_type, 1660 BasicType bt, 1661 DecoratorSet decorators) { 1662 if (stopped()) { 1663 return top(); // Dead path ? 1664 } 1665 1666 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr()); 1667 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr); 1668 if (access.is_raw()) { 1669 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1670 } else { 1671 return _barrier_set->load_at(access, val_type); 1672 } 1673 } 1674 1675 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1676 Node* adr, 1677 const TypePtr* adr_type, 1678 int alias_idx, 1679 Node* expected_val, 1680 Node* new_val, 1681 const Type* value_type, 1682 BasicType bt, 1683 DecoratorSet decorators) { 1684 C2AccessValuePtr addr(adr, adr_type); 1685 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1686 bt, obj, addr, alias_idx); 1687 if (access.is_raw()) { 1688 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1689 } else { 1690 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1691 } 1692 } 1693 1694 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1695 Node* adr, 1696 const TypePtr* adr_type, 1697 int alias_idx, 1698 Node* expected_val, 1699 Node* new_val, 1700 const Type* value_type, 1701 BasicType bt, 1702 DecoratorSet decorators) { 1703 C2AccessValuePtr addr(adr, adr_type); 1704 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1705 bt, obj, addr, alias_idx); 1706 if (access.is_raw()) { 1707 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1708 } else { 1709 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1710 } 1711 } 1712 1713 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1714 Node* adr, 1715 const TypePtr* adr_type, 1716 int alias_idx, 1717 Node* new_val, 1718 const Type* value_type, 1719 BasicType bt, 1720 DecoratorSet decorators) { 1721 C2AccessValuePtr addr(adr, adr_type); 1722 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1723 bt, obj, addr, alias_idx); 1724 if (access.is_raw()) { 1725 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1726 } else { 1727 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1728 } 1729 } 1730 1731 Node* GraphKit::access_atomic_add_at(Node* obj, 1732 Node* adr, 1733 const TypePtr* adr_type, 1734 int alias_idx, 1735 Node* new_val, 1736 const Type* value_type, 1737 BasicType bt, 1738 DecoratorSet decorators) { 1739 C2AccessValuePtr addr(adr, adr_type); 1740 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1741 if (access.is_raw()) { 1742 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1743 } else { 1744 return _barrier_set->atomic_add_at(access, new_val, value_type); 1745 } 1746 } 1747 1748 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) { 1749 return _barrier_set->clone(this, src, dst, size, is_array); 1750 } 1751 1752 //-------------------------array_element_address------------------------- 1753 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1754 const TypeInt* sizetype, Node* ctrl) { 1755 uint shift = exact_log2(type2aelembytes(elembt)); 1756 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1757 1758 // short-circuit a common case (saves lots of confusing waste motion) 1759 jint idx_con = find_int_con(idx, -1); 1760 if (idx_con >= 0) { 1761 intptr_t offset = header + ((intptr_t)idx_con << shift); 1762 return basic_plus_adr(ary, offset); 1763 } 1764 1765 // must be correct type for alignment purposes 1766 Node* base = basic_plus_adr(ary, header); 1767 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1768 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1769 return basic_plus_adr(ary, base, scale); 1770 } 1771 1772 //-------------------------load_array_element------------------------- 1773 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) { 1774 const Type* elemtype = arytype->elem(); 1775 BasicType elembt = elemtype->array_element_basic_type(); 1776 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1777 if (elembt == T_NARROWOOP) { 1778 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1779 } 1780 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt, 1781 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0)); 1782 return ld; 1783 } 1784 1785 //-------------------------set_arguments_for_java_call------------------------- 1786 // Arguments (pre-popped from the stack) are taken from the JVMS. 1787 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1788 // Add the call arguments: 1789 uint nargs = call->method()->arg_size(); 1790 for (uint i = 0; i < nargs; i++) { 1791 Node* arg = argument(i); 1792 call->init_req(i + TypeFunc::Parms, arg); 1793 } 1794 } 1795 1796 //---------------------------set_edges_for_java_call--------------------------- 1797 // Connect a newly created call into the current JVMS. 1798 // A return value node (if any) is returned from set_edges_for_java_call. 1799 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1800 1801 // Add the predefined inputs: 1802 call->init_req( TypeFunc::Control, control() ); 1803 call->init_req( TypeFunc::I_O , i_o() ); 1804 call->init_req( TypeFunc::Memory , reset_memory() ); 1805 call->init_req( TypeFunc::FramePtr, frameptr() ); 1806 call->init_req( TypeFunc::ReturnAdr, top() ); 1807 1808 add_safepoint_edges(call, must_throw); 1809 1810 Node* xcall = _gvn.transform(call); 1811 1812 if (xcall == top()) { 1813 set_control(top()); 1814 return; 1815 } 1816 assert(xcall == call, "call identity is stable"); 1817 1818 // Re-use the current map to produce the result. 1819 1820 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1821 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1822 set_all_memory_call(xcall, separate_io_proj); 1823 1824 //return xcall; // no need, caller already has it 1825 } 1826 1827 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1828 if (stopped()) return top(); // maybe the call folded up? 1829 1830 // Capture the return value, if any. 1831 Node* ret; 1832 if (call->method() == nullptr || 1833 call->method()->return_type()->basic_type() == T_VOID) 1834 ret = top(); 1835 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1836 1837 // Note: Since any out-of-line call can produce an exception, 1838 // we always insert an I_O projection from the call into the result. 1839 1840 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1841 1842 if (separate_io_proj) { 1843 // The caller requested separate projections be used by the fall 1844 // through and exceptional paths, so replace the projections for 1845 // the fall through path. 1846 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1847 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1848 } 1849 return ret; 1850 } 1851 1852 //--------------------set_predefined_input_for_runtime_call-------------------- 1853 // Reading and setting the memory state is way conservative here. 1854 // The real problem is that I am not doing real Type analysis on memory, 1855 // so I cannot distinguish card mark stores from other stores. Across a GC 1856 // point the Store Barrier and the card mark memory has to agree. I cannot 1857 // have a card mark store and its barrier split across the GC point from 1858 // either above or below. Here I get that to happen by reading ALL of memory. 1859 // A better answer would be to separate out card marks from other memory. 1860 // For now, return the input memory state, so that it can be reused 1861 // after the call, if this call has restricted memory effects. 1862 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1863 // Set fixed predefined input arguments 1864 Node* memory = reset_memory(); 1865 Node* m = narrow_mem == nullptr ? memory : narrow_mem; 1866 call->init_req( TypeFunc::Control, control() ); 1867 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1868 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1869 call->init_req( TypeFunc::FramePtr, frameptr() ); 1870 call->init_req( TypeFunc::ReturnAdr, top() ); 1871 return memory; 1872 } 1873 1874 //-------------------set_predefined_output_for_runtime_call-------------------- 1875 // Set control and memory (not i_o) from the call. 1876 // If keep_mem is not null, use it for the output state, 1877 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1878 // If hook_mem is null, this call produces no memory effects at all. 1879 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1880 // then only that memory slice is taken from the call. 1881 // In the last case, we must put an appropriate memory barrier before 1882 // the call, so as to create the correct anti-dependencies on loads 1883 // preceding the call. 1884 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1885 Node* keep_mem, 1886 const TypePtr* hook_mem) { 1887 // no i/o 1888 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1889 if (keep_mem) { 1890 // First clone the existing memory state 1891 set_all_memory(keep_mem); 1892 if (hook_mem != nullptr) { 1893 // Make memory for the call 1894 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1895 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1896 // We also use hook_mem to extract specific effects from arraycopy stubs. 1897 set_memory(mem, hook_mem); 1898 } 1899 // ...else the call has NO memory effects. 1900 1901 // Make sure the call advertises its memory effects precisely. 1902 // This lets us build accurate anti-dependences in gcm.cpp. 1903 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1904 "call node must be constructed correctly"); 1905 } else { 1906 assert(hook_mem == nullptr, ""); 1907 // This is not a "slow path" call; all memory comes from the call. 1908 set_all_memory_call(call); 1909 } 1910 } 1911 1912 // Keep track of MergeMems feeding into other MergeMems 1913 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 1914 if (!mem->is_MergeMem()) { 1915 return; 1916 } 1917 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 1918 Node* use = i.get(); 1919 if (use->is_MergeMem()) { 1920 wl.push(use); 1921 } 1922 } 1923 } 1924 1925 // Replace the call with the current state of the kit. 1926 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1927 JVMState* ejvms = nullptr; 1928 if (has_exceptions()) { 1929 ejvms = transfer_exceptions_into_jvms(); 1930 } 1931 1932 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1933 ReplacedNodes replaced_nodes_exception; 1934 Node* ex_ctl = top(); 1935 1936 SafePointNode* final_state = stop(); 1937 1938 // Find all the needed outputs of this call 1939 CallProjections callprojs; 1940 call->extract_projections(&callprojs, true); 1941 1942 Unique_Node_List wl; 1943 Node* init_mem = call->in(TypeFunc::Memory); 1944 Node* final_mem = final_state->in(TypeFunc::Memory); 1945 Node* final_ctl = final_state->in(TypeFunc::Control); 1946 Node* final_io = final_state->in(TypeFunc::I_O); 1947 1948 // Replace all the old call edges with the edges from the inlining result 1949 if (callprojs.fallthrough_catchproj != nullptr) { 1950 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1951 } 1952 if (callprojs.fallthrough_memproj != nullptr) { 1953 if (final_mem->is_MergeMem()) { 1954 // Parser's exits MergeMem was not transformed but may be optimized 1955 final_mem = _gvn.transform(final_mem); 1956 } 1957 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1958 add_mergemem_users_to_worklist(wl, final_mem); 1959 } 1960 if (callprojs.fallthrough_ioproj != nullptr) { 1961 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1962 } 1963 1964 // Replace the result with the new result if it exists and is used 1965 if (callprojs.resproj != nullptr && result != nullptr) { 1966 C->gvn_replace_by(callprojs.resproj, result); 1967 } 1968 1969 if (ejvms == nullptr) { 1970 // No exception edges to simply kill off those paths 1971 if (callprojs.catchall_catchproj != nullptr) { 1972 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1973 } 1974 if (callprojs.catchall_memproj != nullptr) { 1975 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1976 } 1977 if (callprojs.catchall_ioproj != nullptr) { 1978 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1979 } 1980 // Replace the old exception object with top 1981 if (callprojs.exobj != nullptr) { 1982 C->gvn_replace_by(callprojs.exobj, C->top()); 1983 } 1984 } else { 1985 GraphKit ekit(ejvms); 1986 1987 // Load my combined exception state into the kit, with all phis transformed: 1988 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1989 replaced_nodes_exception = ex_map->replaced_nodes(); 1990 1991 Node* ex_oop = ekit.use_exception_state(ex_map); 1992 1993 if (callprojs.catchall_catchproj != nullptr) { 1994 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1995 ex_ctl = ekit.control(); 1996 } 1997 if (callprojs.catchall_memproj != nullptr) { 1998 Node* ex_mem = ekit.reset_memory(); 1999 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem); 2000 add_mergemem_users_to_worklist(wl, ex_mem); 2001 } 2002 if (callprojs.catchall_ioproj != nullptr) { 2003 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 2004 } 2005 2006 // Replace the old exception object with the newly created one 2007 if (callprojs.exobj != nullptr) { 2008 C->gvn_replace_by(callprojs.exobj, ex_oop); 2009 } 2010 } 2011 2012 // Disconnect the call from the graph 2013 call->disconnect_inputs(C); 2014 C->gvn_replace_by(call, C->top()); 2015 2016 // Clean up any MergeMems that feed other MergeMems since the 2017 // optimizer doesn't like that. 2018 while (wl.size() > 0) { 2019 _gvn.transform(wl.pop()); 2020 } 2021 2022 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) { 2023 replaced_nodes.apply(C, final_ctl); 2024 } 2025 if (!ex_ctl->is_top() && do_replaced_nodes) { 2026 replaced_nodes_exception.apply(C, ex_ctl); 2027 } 2028 } 2029 2030 2031 //------------------------------increment_counter------------------------------ 2032 // for statistics: increment a VM counter by 1 2033 2034 void GraphKit::increment_counter(address counter_addr) { 2035 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2036 increment_counter(adr1); 2037 } 2038 2039 void GraphKit::increment_counter(Node* counter_addr) { 2040 int adr_type = Compile::AliasIdxRaw; 2041 Node* ctrl = control(); 2042 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered); 2043 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1))); 2044 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered); 2045 } 2046 2047 2048 //------------------------------uncommon_trap---------------------------------- 2049 // Bail out to the interpreter in mid-method. Implemented by calling the 2050 // uncommon_trap blob. This helper function inserts a runtime call with the 2051 // right debug info. 2052 Node* GraphKit::uncommon_trap(int trap_request, 2053 ciKlass* klass, const char* comment, 2054 bool must_throw, 2055 bool keep_exact_action) { 2056 if (failing()) stop(); 2057 if (stopped()) return nullptr; // trap reachable? 2058 2059 // Note: If ProfileTraps is true, and if a deopt. actually 2060 // occurs here, the runtime will make sure an MDO exists. There is 2061 // no need to call method()->ensure_method_data() at this point. 2062 2063 // Set the stack pointer to the right value for reexecution: 2064 set_sp(reexecute_sp()); 2065 2066 #ifdef ASSERT 2067 if (!must_throw) { 2068 // Make sure the stack has at least enough depth to execute 2069 // the current bytecode. 2070 int inputs, ignored_depth; 2071 if (compute_stack_effects(inputs, ignored_depth)) { 2072 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2073 Bytecodes::name(java_bc()), sp(), inputs); 2074 } 2075 } 2076 #endif 2077 2078 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2079 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2080 2081 switch (action) { 2082 case Deoptimization::Action_maybe_recompile: 2083 case Deoptimization::Action_reinterpret: 2084 // Temporary fix for 6529811 to allow virtual calls to be sure they 2085 // get the chance to go from mono->bi->mega 2086 if (!keep_exact_action && 2087 Deoptimization::trap_request_index(trap_request) < 0 && 2088 too_many_recompiles(reason)) { 2089 // This BCI is causing too many recompilations. 2090 if (C->log() != nullptr) { 2091 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2092 Deoptimization::trap_reason_name(reason), 2093 Deoptimization::trap_action_name(action)); 2094 } 2095 action = Deoptimization::Action_none; 2096 trap_request = Deoptimization::make_trap_request(reason, action); 2097 } else { 2098 C->set_trap_can_recompile(true); 2099 } 2100 break; 2101 case Deoptimization::Action_make_not_entrant: 2102 C->set_trap_can_recompile(true); 2103 break; 2104 case Deoptimization::Action_none: 2105 case Deoptimization::Action_make_not_compilable: 2106 break; 2107 default: 2108 #ifdef ASSERT 2109 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2110 #endif 2111 break; 2112 } 2113 2114 if (TraceOptoParse) { 2115 char buf[100]; 2116 tty->print_cr("Uncommon trap %s at bci:%d", 2117 Deoptimization::format_trap_request(buf, sizeof(buf), 2118 trap_request), bci()); 2119 } 2120 2121 CompileLog* log = C->log(); 2122 if (log != nullptr) { 2123 int kid = (klass == nullptr)? -1: log->identify(klass); 2124 log->begin_elem("uncommon_trap bci='%d'", bci()); 2125 char buf[100]; 2126 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2127 trap_request)); 2128 if (kid >= 0) log->print(" klass='%d'", kid); 2129 if (comment != nullptr) log->print(" comment='%s'", comment); 2130 log->end_elem(); 2131 } 2132 2133 // Make sure any guarding test views this path as very unlikely 2134 Node *i0 = control()->in(0); 2135 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test? 2136 IfNode *iff = i0->as_If(); 2137 float f = iff->_prob; // Get prob 2138 if (control()->Opcode() == Op_IfTrue) { 2139 if (f > PROB_UNLIKELY_MAG(4)) 2140 iff->_prob = PROB_MIN; 2141 } else { 2142 if (f < PROB_LIKELY_MAG(4)) 2143 iff->_prob = PROB_MAX; 2144 } 2145 } 2146 2147 // Clear out dead values from the debug info. 2148 kill_dead_locals(); 2149 2150 // Now insert the uncommon trap subroutine call 2151 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2152 const TypePtr* no_memory_effects = nullptr; 2153 // Pass the index of the class to be loaded 2154 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2155 (must_throw ? RC_MUST_THROW : 0), 2156 OptoRuntime::uncommon_trap_Type(), 2157 call_addr, "uncommon_trap", no_memory_effects, 2158 intcon(trap_request)); 2159 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2160 "must extract request correctly from the graph"); 2161 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2162 2163 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2164 // The debug info is the only real input to this call. 2165 2166 // Halt-and-catch fire here. The above call should never return! 2167 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2168 PRODUCT_ONLY(COMMA /*reachable*/false)); 2169 _gvn.set_type_bottom(halt); 2170 root()->add_req(halt); 2171 2172 stop_and_kill_map(); 2173 return call; 2174 } 2175 2176 2177 //--------------------------just_allocated_object------------------------------ 2178 // Report the object that was just allocated. 2179 // It must be the case that there are no intervening safepoints. 2180 // We use this to determine if an object is so "fresh" that 2181 // it does not require card marks. 2182 Node* GraphKit::just_allocated_object(Node* current_control) { 2183 Node* ctrl = current_control; 2184 // Object::<init> is invoked after allocation, most of invoke nodes 2185 // will be reduced, but a region node is kept in parse time, we check 2186 // the pattern and skip the region node if it degraded to a copy. 2187 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 && 2188 ctrl->as_Region()->is_copy()) { 2189 ctrl = ctrl->as_Region()->is_copy(); 2190 } 2191 if (C->recent_alloc_ctl() == ctrl) { 2192 return C->recent_alloc_obj(); 2193 } 2194 return nullptr; 2195 } 2196 2197 2198 /** 2199 * Record profiling data exact_kls for Node n with the type system so 2200 * that it can propagate it (speculation) 2201 * 2202 * @param n node that the type applies to 2203 * @param exact_kls type from profiling 2204 * @param maybe_null did profiling see null? 2205 * 2206 * @return node with improved type 2207 */ 2208 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2209 const Type* current_type = _gvn.type(n); 2210 assert(UseTypeSpeculation, "type speculation must be on"); 2211 2212 const TypePtr* speculative = current_type->speculative(); 2213 2214 // Should the klass from the profile be recorded in the speculative type? 2215 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2216 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces); 2217 const TypeOopPtr* xtype = tklass->as_instance_type(); 2218 assert(xtype->klass_is_exact(), "Should be exact"); 2219 // Any reason to believe n is not null (from this profiling or a previous one)? 2220 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2221 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2222 // record the new speculative type's depth 2223 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2224 speculative = speculative->with_inline_depth(jvms()->depth()); 2225 } else if (current_type->would_improve_ptr(ptr_kind)) { 2226 // Profiling report that null was never seen so we can change the 2227 // speculative type to non null ptr. 2228 if (ptr_kind == ProfileAlwaysNull) { 2229 speculative = TypePtr::NULL_PTR; 2230 } else { 2231 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2232 const TypePtr* ptr = TypePtr::NOTNULL; 2233 if (speculative != nullptr) { 2234 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2235 } else { 2236 speculative = ptr; 2237 } 2238 } 2239 } 2240 2241 if (speculative != current_type->speculative()) { 2242 // Build a type with a speculative type (what we think we know 2243 // about the type but will need a guard when we use it) 2244 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2245 // We're changing the type, we need a new CheckCast node to carry 2246 // the new type. The new type depends on the control: what 2247 // profiling tells us is only valid from here as far as we can 2248 // tell. 2249 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2250 cast = _gvn.transform(cast); 2251 replace_in_map(n, cast); 2252 n = cast; 2253 } 2254 2255 return n; 2256 } 2257 2258 /** 2259 * Record profiling data from receiver profiling at an invoke with the 2260 * type system so that it can propagate it (speculation) 2261 * 2262 * @param n receiver node 2263 * 2264 * @return node with improved type 2265 */ 2266 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2267 if (!UseTypeSpeculation) { 2268 return n; 2269 } 2270 ciKlass* exact_kls = profile_has_unique_klass(); 2271 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2272 if ((java_bc() == Bytecodes::_checkcast || 2273 java_bc() == Bytecodes::_instanceof || 2274 java_bc() == Bytecodes::_aastore) && 2275 method()->method_data()->is_mature()) { 2276 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2277 if (data != nullptr) { 2278 if (!data->as_BitData()->null_seen()) { 2279 ptr_kind = ProfileNeverNull; 2280 } else { 2281 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2282 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2283 uint i = 0; 2284 for (; i < call->row_limit(); i++) { 2285 ciKlass* receiver = call->receiver(i); 2286 if (receiver != nullptr) { 2287 break; 2288 } 2289 } 2290 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2291 } 2292 } 2293 } 2294 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2295 } 2296 2297 /** 2298 * Record profiling data from argument profiling at an invoke with the 2299 * type system so that it can propagate it (speculation) 2300 * 2301 * @param dest_method target method for the call 2302 * @param bc what invoke bytecode is this? 2303 */ 2304 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2305 if (!UseTypeSpeculation) { 2306 return; 2307 } 2308 const TypeFunc* tf = TypeFunc::make(dest_method); 2309 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2310 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2311 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2312 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2313 if (is_reference_type(targ->basic_type())) { 2314 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2315 ciKlass* better_type = nullptr; 2316 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2317 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2318 } 2319 i++; 2320 } 2321 } 2322 } 2323 2324 /** 2325 * Record profiling data from parameter profiling at an invoke with 2326 * the type system so that it can propagate it (speculation) 2327 */ 2328 void GraphKit::record_profiled_parameters_for_speculation() { 2329 if (!UseTypeSpeculation) { 2330 return; 2331 } 2332 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2333 if (_gvn.type(local(i))->isa_oopptr()) { 2334 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2335 ciKlass* better_type = nullptr; 2336 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2337 record_profile_for_speculation(local(i), better_type, ptr_kind); 2338 } 2339 j++; 2340 } 2341 } 2342 } 2343 2344 /** 2345 * Record profiling data from return value profiling at an invoke with 2346 * the type system so that it can propagate it (speculation) 2347 */ 2348 void GraphKit::record_profiled_return_for_speculation() { 2349 if (!UseTypeSpeculation) { 2350 return; 2351 } 2352 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2353 ciKlass* better_type = nullptr; 2354 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2355 // If profiling reports a single type for the return value, 2356 // feed it to the type system so it can propagate it as a 2357 // speculative type 2358 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2359 } 2360 } 2361 2362 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2363 if (Matcher::strict_fp_requires_explicit_rounding) { 2364 // (Note: TypeFunc::make has a cache that makes this fast.) 2365 const TypeFunc* tf = TypeFunc::make(dest_method); 2366 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2367 for (int j = 0; j < nargs; j++) { 2368 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2369 if (targ->basic_type() == T_DOUBLE) { 2370 // If any parameters are doubles, they must be rounded before 2371 // the call, dprecision_rounding does gvn.transform 2372 Node *arg = argument(j); 2373 arg = dprecision_rounding(arg); 2374 set_argument(j, arg); 2375 } 2376 } 2377 } 2378 } 2379 2380 // rounding for strict float precision conformance 2381 Node* GraphKit::precision_rounding(Node* n) { 2382 if (Matcher::strict_fp_requires_explicit_rounding) { 2383 #ifdef IA32 2384 if (UseSSE == 0) { 2385 return _gvn.transform(new RoundFloatNode(0, n)); 2386 } 2387 #else 2388 Unimplemented(); 2389 #endif // IA32 2390 } 2391 return n; 2392 } 2393 2394 // rounding for strict double precision conformance 2395 Node* GraphKit::dprecision_rounding(Node *n) { 2396 if (Matcher::strict_fp_requires_explicit_rounding) { 2397 #ifdef IA32 2398 if (UseSSE < 2) { 2399 return _gvn.transform(new RoundDoubleNode(0, n)); 2400 } 2401 #else 2402 Unimplemented(); 2403 #endif // IA32 2404 } 2405 return n; 2406 } 2407 2408 //============================================================================= 2409 // Generate a fast path/slow path idiom. Graph looks like: 2410 // [foo] indicates that 'foo' is a parameter 2411 // 2412 // [in] null 2413 // \ / 2414 // CmpP 2415 // Bool ne 2416 // If 2417 // / \ 2418 // True False-<2> 2419 // / | 2420 // / cast_not_null 2421 // Load | | ^ 2422 // [fast_test] | | 2423 // gvn to opt_test | | 2424 // / \ | <1> 2425 // True False | 2426 // | \\ | 2427 // [slow_call] \[fast_result] 2428 // Ctl Val \ \ 2429 // | \ \ 2430 // Catch <1> \ \ 2431 // / \ ^ \ \ 2432 // Ex No_Ex | \ \ 2433 // | \ \ | \ <2> \ 2434 // ... \ [slow_res] | | \ [null_result] 2435 // \ \--+--+--- | | 2436 // \ | / \ | / 2437 // --------Region Phi 2438 // 2439 //============================================================================= 2440 // Code is structured as a series of driver functions all called 'do_XXX' that 2441 // call a set of helper functions. Helper functions first, then drivers. 2442 2443 //------------------------------null_check_oop--------------------------------- 2444 // Null check oop. Set null-path control into Region in slot 3. 2445 // Make a cast-not-nullness use the other not-null control. Return cast. 2446 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2447 bool never_see_null, 2448 bool safe_for_replace, 2449 bool speculative) { 2450 // Initial null check taken path 2451 (*null_control) = top(); 2452 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2453 2454 // Generate uncommon_trap: 2455 if (never_see_null && (*null_control) != top()) { 2456 // If we see an unexpected null at a check-cast we record it and force a 2457 // recompile; the offending check-cast will be compiled to handle nulls. 2458 // If we see more than one offending BCI, then all checkcasts in the 2459 // method will be compiled to handle nulls. 2460 PreserveJVMState pjvms(this); 2461 set_control(*null_control); 2462 replace_in_map(value, null()); 2463 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2464 uncommon_trap(reason, 2465 Deoptimization::Action_make_not_entrant); 2466 (*null_control) = top(); // null path is dead 2467 } 2468 if ((*null_control) == top() && safe_for_replace) { 2469 replace_in_map(value, cast); 2470 } 2471 2472 // Cast away null-ness on the result 2473 return cast; 2474 } 2475 2476 //------------------------------opt_iff---------------------------------------- 2477 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2478 // Return slow-path control. 2479 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2480 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2481 2482 // Fast path taken; set region slot 2 2483 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2484 region->init_req(2,fast_taken); // Capture fast-control 2485 2486 // Fast path not-taken, i.e. slow path 2487 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2488 return slow_taken; 2489 } 2490 2491 //-----------------------------make_runtime_call------------------------------- 2492 Node* GraphKit::make_runtime_call(int flags, 2493 const TypeFunc* call_type, address call_addr, 2494 const char* call_name, 2495 const TypePtr* adr_type, 2496 // The following parms are all optional. 2497 // The first null ends the list. 2498 Node* parm0, Node* parm1, 2499 Node* parm2, Node* parm3, 2500 Node* parm4, Node* parm5, 2501 Node* parm6, Node* parm7) { 2502 assert(call_addr != nullptr, "must not call null targets"); 2503 2504 // Slow-path call 2505 bool is_leaf = !(flags & RC_NO_LEAF); 2506 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2507 if (call_name == nullptr) { 2508 assert(!is_leaf, "must supply name for leaf"); 2509 call_name = OptoRuntime::stub_name(call_addr); 2510 } 2511 CallNode* call; 2512 if (!is_leaf) { 2513 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type); 2514 } else if (flags & RC_NO_FP) { 2515 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2516 } else if (flags & RC_VECTOR){ 2517 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte; 2518 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits); 2519 } else { 2520 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2521 } 2522 2523 // The following is similar to set_edges_for_java_call, 2524 // except that the memory effects of the call are restricted to AliasIdxRaw. 2525 2526 // Slow path call has no side-effects, uses few values 2527 bool wide_in = !(flags & RC_NARROW_MEM); 2528 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2529 2530 Node* prev_mem = nullptr; 2531 if (wide_in) { 2532 prev_mem = set_predefined_input_for_runtime_call(call); 2533 } else { 2534 assert(!wide_out, "narrow in => narrow out"); 2535 Node* narrow_mem = memory(adr_type); 2536 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2537 } 2538 2539 // Hook each parm in order. Stop looking at the first null. 2540 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0); 2541 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1); 2542 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2); 2543 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3); 2544 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4); 2545 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5); 2546 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6); 2547 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7); 2548 /* close each nested if ===> */ } } } } } } } } 2549 assert(call->in(call->req()-1) != nullptr, "must initialize all parms"); 2550 2551 if (!is_leaf) { 2552 // Non-leaves can block and take safepoints: 2553 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2554 } 2555 // Non-leaves can throw exceptions: 2556 if (has_io) { 2557 call->set_req(TypeFunc::I_O, i_o()); 2558 } 2559 2560 if (flags & RC_UNCOMMON) { 2561 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2562 // (An "if" probability corresponds roughly to an unconditional count. 2563 // Sort of.) 2564 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2565 } 2566 2567 Node* c = _gvn.transform(call); 2568 assert(c == call, "cannot disappear"); 2569 2570 if (wide_out) { 2571 // Slow path call has full side-effects. 2572 set_predefined_output_for_runtime_call(call); 2573 } else { 2574 // Slow path call has few side-effects, and/or sets few values. 2575 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2576 } 2577 2578 if (has_io) { 2579 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2580 } 2581 return call; 2582 2583 } 2584 2585 // i2b 2586 Node* GraphKit::sign_extend_byte(Node* in) { 2587 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24))); 2588 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24))); 2589 } 2590 2591 // i2s 2592 Node* GraphKit::sign_extend_short(Node* in) { 2593 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16))); 2594 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16))); 2595 } 2596 2597 //------------------------------merge_memory----------------------------------- 2598 // Merge memory from one path into the current memory state. 2599 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2600 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2601 Node* old_slice = mms.force_memory(); 2602 Node* new_slice = mms.memory2(); 2603 if (old_slice != new_slice) { 2604 PhiNode* phi; 2605 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2606 if (mms.is_empty()) { 2607 // clone base memory Phi's inputs for this memory slice 2608 assert(old_slice == mms.base_memory(), "sanity"); 2609 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C)); 2610 _gvn.set_type(phi, Type::MEMORY); 2611 for (uint i = 1; i < phi->req(); i++) { 2612 phi->init_req(i, old_slice->in(i)); 2613 } 2614 } else { 2615 phi = old_slice->as_Phi(); // Phi was generated already 2616 } 2617 } else { 2618 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2619 _gvn.set_type(phi, Type::MEMORY); 2620 } 2621 phi->set_req(new_path, new_slice); 2622 mms.set_memory(phi); 2623 } 2624 } 2625 } 2626 2627 //------------------------------make_slow_call_ex------------------------------ 2628 // Make the exception handler hookups for the slow call 2629 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2630 if (stopped()) return; 2631 2632 // Make a catch node with just two handlers: fall-through and catch-all 2633 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2634 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2635 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2636 _gvn.set_type_bottom(norm); 2637 C->record_for_igvn(norm); 2638 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2639 2640 { PreserveJVMState pjvms(this); 2641 set_control(excp); 2642 set_i_o(i_o); 2643 2644 if (excp != top()) { 2645 if (deoptimize) { 2646 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2647 uncommon_trap(Deoptimization::Reason_unhandled, 2648 Deoptimization::Action_none); 2649 } else { 2650 // Create an exception state also. 2651 // Use an exact type if the caller has a specific exception. 2652 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2653 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2654 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2655 } 2656 } 2657 } 2658 2659 // Get the no-exception control from the CatchNode. 2660 set_control(norm); 2661 } 2662 2663 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) { 2664 Node* cmp = nullptr; 2665 switch(bt) { 2666 case T_INT: cmp = new CmpINode(in1, in2); break; 2667 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2668 default: fatal("unexpected comparison type %s", type2name(bt)); 2669 } 2670 cmp = gvn.transform(cmp); 2671 Node* bol = gvn.transform(new BoolNode(cmp, test)); 2672 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2673 gvn.transform(iff); 2674 if (!bol->is_Con()) gvn.record_for_igvn(iff); 2675 return iff; 2676 } 2677 2678 //-------------------------------gen_subtype_check----------------------------- 2679 // Generate a subtyping check. Takes as input the subtype and supertype. 2680 // Returns 2 values: sets the default control() to the true path and returns 2681 // the false path. Only reads invariant memory; sets no (visible) memory. 2682 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2683 // but that's not exposed to the optimizer. This call also doesn't take in an 2684 // Object; if you wish to check an Object you need to load the Object's class 2685 // prior to coming here. 2686 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn, 2687 ciMethod* method, int bci) { 2688 Compile* C = gvn.C; 2689 if ((*ctrl)->is_top()) { 2690 return C->top(); 2691 } 2692 2693 // Fast check for identical types, perhaps identical constants. 2694 // The types can even be identical non-constants, in cases 2695 // involving Array.newInstance, Object.clone, etc. 2696 if (subklass == superklass) 2697 return C->top(); // false path is dead; no test needed. 2698 2699 if (gvn.type(superklass)->singleton()) { 2700 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2701 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr(); 2702 2703 // In the common case of an exact superklass, try to fold up the 2704 // test before generating code. You may ask, why not just generate 2705 // the code and then let it fold up? The answer is that the generated 2706 // code will necessarily include null checks, which do not always 2707 // completely fold away. If they are also needless, then they turn 2708 // into a performance loss. Example: 2709 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2710 // Here, the type of 'fa' is often exact, so the store check 2711 // of fa[1]=x will fold up, without testing the nullness of x. 2712 switch (C->static_subtype_check(superk, subk)) { 2713 case Compile::SSC_always_false: 2714 { 2715 Node* always_fail = *ctrl; 2716 *ctrl = gvn.C->top(); 2717 return always_fail; 2718 } 2719 case Compile::SSC_always_true: 2720 return C->top(); 2721 case Compile::SSC_easy_test: 2722 { 2723 // Just do a direct pointer compare and be done. 2724 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2725 *ctrl = gvn.transform(new IfTrueNode(iff)); 2726 return gvn.transform(new IfFalseNode(iff)); 2727 } 2728 case Compile::SSC_full_test: 2729 break; 2730 default: 2731 ShouldNotReachHere(); 2732 } 2733 } 2734 2735 // %%% Possible further optimization: Even if the superklass is not exact, 2736 // if the subklass is the unique subtype of the superklass, the check 2737 // will always succeed. We could leave a dependency behind to ensure this. 2738 2739 // First load the super-klass's check-offset 2740 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2741 Node* m = C->immutable_memory(); 2742 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2743 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2744 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int(); 2745 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con; 2746 bool might_be_cache = (chk_off_con == cacheoff_con); 2747 2748 // Load from the sub-klass's super-class display list, or a 1-word cache of 2749 // the secondary superclass list, or a failing value with a sentinel offset 2750 // if the super-klass is an interface or exceptionally deep in the Java 2751 // hierarchy and we have to scan the secondary superclass list the hard way. 2752 // Worst-case type is a little odd: null is allowed as a result (usually 2753 // klass loads can never produce a null). 2754 Node *chk_off_X = chk_off; 2755 #ifdef _LP64 2756 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X)); 2757 #endif 2758 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X)); 2759 // For some types like interfaces the following loadKlass is from a 1-word 2760 // cache which is mutable so can't use immutable memory. Other 2761 // types load from the super-class display table which is immutable. 2762 Node *kmem = C->immutable_memory(); 2763 // secondary_super_cache is not immutable but can be treated as such because: 2764 // - no ideal node writes to it in a way that could cause an 2765 // incorrect/missed optimization of the following Load. 2766 // - it's a cache so, worse case, not reading the latest value 2767 // wouldn't cause incorrect execution 2768 if (might_be_cache && mem != nullptr) { 2769 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem; 2770 } 2771 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL)); 2772 2773 // Compile speed common case: ARE a subtype and we canNOT fail 2774 if (superklass == nkls) { 2775 return C->top(); // false path is dead; no test needed. 2776 } 2777 2778 // Gather the various success & failures here 2779 RegionNode* r_not_subtype = new RegionNode(3); 2780 gvn.record_for_igvn(r_not_subtype); 2781 RegionNode* r_ok_subtype = new RegionNode(4); 2782 gvn.record_for_igvn(r_ok_subtype); 2783 2784 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the 2785 // SubTypeCheck node 2786 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) { 2787 ciCallProfile profile = method->call_profile_at_bci(bci); 2788 float total_prob = 0; 2789 for (int i = 0; profile.has_receiver(i); ++i) { 2790 float prob = profile.receiver_prob(i); 2791 total_prob += prob; 2792 } 2793 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) { 2794 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2795 for (int i = 0; profile.has_receiver(i); ++i) { 2796 ciKlass* klass = profile.receiver(i); 2797 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass); 2798 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t); 2799 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) { 2800 continue; 2801 } 2802 float prob = profile.receiver_prob(i); 2803 ConNode* klass_node = gvn.makecon(klass_t); 2804 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS); 2805 Node* iftrue = gvn.transform(new IfTrueNode(iff)); 2806 2807 if (result == Compile::SSC_always_true) { 2808 r_ok_subtype->add_req(iftrue); 2809 } else { 2810 assert(result == Compile::SSC_always_false, ""); 2811 r_not_subtype->add_req(iftrue); 2812 } 2813 *ctrl = gvn.transform(new IfFalseNode(iff)); 2814 } 2815 } 2816 } 2817 2818 // See if we get an immediate positive hit. Happens roughly 83% of the 2819 // time. Test to see if the value loaded just previously from the subklass 2820 // is exactly the superklass. 2821 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2822 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1)); 2823 *ctrl = gvn.transform(new IfFalseNode(iff1)); 2824 2825 // Compile speed common case: Check for being deterministic right now. If 2826 // chk_off is a constant and not equal to cacheoff then we are NOT a 2827 // subklass. In this case we need exactly the 1 test above and we can 2828 // return those results immediately. 2829 if (!might_be_cache) { 2830 Node* not_subtype_ctrl = *ctrl; 2831 *ctrl = iftrue1; // We need exactly the 1 test above 2832 PhaseIterGVN* igvn = gvn.is_IterGVN(); 2833 if (igvn != nullptr) { 2834 igvn->remove_globally_dead_node(r_ok_subtype); 2835 igvn->remove_globally_dead_node(r_not_subtype); 2836 } 2837 return not_subtype_ctrl; 2838 } 2839 2840 r_ok_subtype->init_req(1, iftrue1); 2841 2842 // Check for immediate negative hit. Happens roughly 11% of the time (which 2843 // is roughly 63% of the remaining cases). Test to see if the loaded 2844 // check-offset points into the subklass display list or the 1-element 2845 // cache. If it points to the display (and NOT the cache) and the display 2846 // missed then it's not a subtype. 2847 Node *cacheoff = gvn.intcon(cacheoff_con); 2848 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2849 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2))); 2850 *ctrl = gvn.transform(new IfFalseNode(iff2)); 2851 2852 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2853 // No performance impact (too rare) but allows sharing of secondary arrays 2854 // which has some footprint reduction. 2855 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2856 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3))); 2857 *ctrl = gvn.transform(new IfFalseNode(iff3)); 2858 2859 // -- Roads not taken here: -- 2860 // We could also have chosen to perform the self-check at the beginning 2861 // of this code sequence, as the assembler does. This would not pay off 2862 // the same way, since the optimizer, unlike the assembler, can perform 2863 // static type analysis to fold away many successful self-checks. 2864 // Non-foldable self checks work better here in second position, because 2865 // the initial primary superclass check subsumes a self-check for most 2866 // types. An exception would be a secondary type like array-of-interface, 2867 // which does not appear in its own primary supertype display. 2868 // Finally, we could have chosen to move the self-check into the 2869 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2870 // dependent manner. But it is worthwhile to have the check here, 2871 // where it can be perhaps be optimized. The cost in code space is 2872 // small (register compare, branch). 2873 2874 // Now do a linear scan of the secondary super-klass array. Again, no real 2875 // performance impact (too rare) but it's gotta be done. 2876 // Since the code is rarely used, there is no penalty for moving it 2877 // out of line, and it can only improve I-cache density. 2878 // The decision to inline or out-of-line this final check is platform 2879 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2880 Node* psc = gvn.transform( 2881 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2882 2883 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2884 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4))); 2885 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4))); 2886 2887 // Return false path; set default control to true path. 2888 *ctrl = gvn.transform(r_ok_subtype); 2889 return gvn.transform(r_not_subtype); 2890 } 2891 2892 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) { 2893 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over 2894 ExpandSubTypeCheckAtParseTime; // forced expansion 2895 if (expand_subtype_check) { 2896 MergeMemNode* mem = merged_memory(); 2897 Node* ctrl = control(); 2898 Node* subklass = obj_or_subklass; 2899 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) { 2900 subklass = load_object_klass(obj_or_subklass); 2901 } 2902 2903 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci()); 2904 set_control(ctrl); 2905 return n; 2906 } 2907 2908 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci())); 2909 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq)); 2910 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2911 set_control(_gvn.transform(new IfTrueNode(iff))); 2912 return _gvn.transform(new IfFalseNode(iff)); 2913 } 2914 2915 // Profile-driven exact type check: 2916 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2917 float prob, 2918 Node* *casted_receiver) { 2919 assert(!klass->is_interface(), "no exact type check on interfaces"); 2920 2921 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces); 2922 Node* recv_klass = load_object_klass(receiver); 2923 Node* want_klass = makecon(tklass); 2924 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass)); 2925 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 2926 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2927 set_control( _gvn.transform(new IfTrueNode (iff))); 2928 Node* fail = _gvn.transform(new IfFalseNode(iff)); 2929 2930 if (!stopped()) { 2931 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 2932 const TypeOopPtr* recvx_type = tklass->as_instance_type(); 2933 assert(recvx_type->klass_is_exact(), ""); 2934 2935 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts 2936 // Subsume downstream occurrences of receiver with a cast to 2937 // recv_xtype, since now we know what the type will be. 2938 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type); 2939 (*casted_receiver) = _gvn.transform(cast); 2940 assert(!(*casted_receiver)->is_top(), "that path should be unreachable"); 2941 // (User must make the replace_in_map call.) 2942 } 2943 } 2944 2945 return fail; 2946 } 2947 2948 //------------------------------subtype_check_receiver------------------------- 2949 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 2950 Node** casted_receiver) { 2951 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve(); 2952 Node* want_klass = makecon(tklass); 2953 2954 Node* slow_ctl = gen_subtype_check(receiver, want_klass); 2955 2956 // Ignore interface type information until interface types are properly tracked. 2957 if (!stopped() && !klass->is_interface()) { 2958 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 2959 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 2960 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts 2961 Node* cast = new CheckCastPPNode(control(), receiver, recv_type); 2962 (*casted_receiver) = _gvn.transform(cast); 2963 } 2964 } 2965 2966 return slow_ctl; 2967 } 2968 2969 //------------------------------seems_never_null------------------------------- 2970 // Use null_seen information if it is available from the profile. 2971 // If we see an unexpected null at a type check we record it and force a 2972 // recompile; the offending check will be recompiled to handle nulls. 2973 // If we see several offending BCIs, then all checks in the 2974 // method will be recompiled. 2975 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2976 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2977 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2978 if (UncommonNullCast // Cutout for this technique 2979 && obj != null() // And not the -Xcomp stupid case? 2980 && !too_many_traps(reason) 2981 ) { 2982 if (speculating) { 2983 return true; 2984 } 2985 if (data == nullptr) 2986 // Edge case: no mature data. Be optimistic here. 2987 return true; 2988 // If the profile has not seen a null, assume it won't happen. 2989 assert(java_bc() == Bytecodes::_checkcast || 2990 java_bc() == Bytecodes::_instanceof || 2991 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2992 return !data->as_BitData()->null_seen(); 2993 } 2994 speculating = false; 2995 return false; 2996 } 2997 2998 void GraphKit::guard_klass_being_initialized(Node* klass) { 2999 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 3000 Node* adr = basic_plus_adr(top(), klass, init_state_off); 3001 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3002 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 3003 T_BYTE, MemNode::unordered); 3004 init_state = _gvn.transform(init_state); 3005 3006 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 3007 3008 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 3009 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3010 3011 { BuildCutout unless(this, tst, PROB_MAX); 3012 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 3013 } 3014 } 3015 3016 void GraphKit::guard_init_thread(Node* klass) { 3017 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 3018 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 3019 3020 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3021 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 3022 T_ADDRESS, MemNode::unordered); 3023 init_thread = _gvn.transform(init_thread); 3024 3025 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 3026 3027 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 3028 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3029 3030 { BuildCutout unless(this, tst, PROB_MAX); 3031 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 3032 } 3033 } 3034 3035 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3036 if (ik->is_being_initialized()) { 3037 if (C->needs_clinit_barrier(ik, context)) { 3038 Node* klass = makecon(TypeKlassPtr::make(ik)); 3039 guard_klass_being_initialized(klass); 3040 guard_init_thread(klass); 3041 insert_mem_bar(Op_MemBarCPUOrder); 3042 } 3043 } else if (ik->is_initialized()) { 3044 return; // no barrier needed 3045 } else { 3046 uncommon_trap(Deoptimization::Reason_uninitialized, 3047 Deoptimization::Action_reinterpret, 3048 nullptr); 3049 } 3050 } 3051 3052 //------------------------maybe_cast_profiled_receiver------------------------- 3053 // If the profile has seen exactly one type, narrow to exactly that type. 3054 // Subsequent type checks will always fold up. 3055 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3056 const TypeKlassPtr* require_klass, 3057 ciKlass* spec_klass, 3058 bool safe_for_replace) { 3059 if (!UseTypeProfile || !TypeProfileCasts) return nullptr; 3060 3061 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr); 3062 3063 // Make sure we haven't already deoptimized from this tactic. 3064 if (too_many_traps_or_recompiles(reason)) 3065 return nullptr; 3066 3067 // (No, this isn't a call, but it's enough like a virtual call 3068 // to use the same ciMethod accessor to get the profile info...) 3069 // If we have a speculative type use it instead of profiling (which 3070 // may not help us) 3071 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass; 3072 if (exact_kls != nullptr) {// no cast failures here 3073 if (require_klass == nullptr || 3074 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) { 3075 // If we narrow the type to match what the type profile sees or 3076 // the speculative type, we can then remove the rest of the 3077 // cast. 3078 // This is a win, even if the exact_kls is very specific, 3079 // because downstream operations, such as method calls, 3080 // will often benefit from the sharper type. 3081 Node* exact_obj = not_null_obj; // will get updated in place... 3082 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3083 &exact_obj); 3084 { PreserveJVMState pjvms(this); 3085 set_control(slow_ctl); 3086 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3087 } 3088 if (safe_for_replace) { 3089 replace_in_map(not_null_obj, exact_obj); 3090 } 3091 return exact_obj; 3092 } 3093 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3094 } 3095 3096 return nullptr; 3097 } 3098 3099 /** 3100 * Cast obj to type and emit guard unless we had too many traps here 3101 * already 3102 * 3103 * @param obj node being casted 3104 * @param type type to cast the node to 3105 * @param not_null true if we know node cannot be null 3106 */ 3107 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3108 ciKlass* type, 3109 bool not_null) { 3110 if (stopped()) { 3111 return obj; 3112 } 3113 3114 // type is null if profiling tells us this object is always null 3115 if (type != nullptr) { 3116 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3117 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3118 3119 if (!too_many_traps_or_recompiles(null_reason) && 3120 !too_many_traps_or_recompiles(class_reason)) { 3121 Node* not_null_obj = nullptr; 3122 // not_null is true if we know the object is not null and 3123 // there's no need for a null check 3124 if (!not_null) { 3125 Node* null_ctl = top(); 3126 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3127 assert(null_ctl->is_top(), "no null control here"); 3128 } else { 3129 not_null_obj = obj; 3130 } 3131 3132 Node* exact_obj = not_null_obj; 3133 ciKlass* exact_kls = type; 3134 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3135 &exact_obj); 3136 { 3137 PreserveJVMState pjvms(this); 3138 set_control(slow_ctl); 3139 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3140 } 3141 replace_in_map(not_null_obj, exact_obj); 3142 obj = exact_obj; 3143 } 3144 } else { 3145 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3146 Node* exact_obj = null_assert(obj); 3147 replace_in_map(obj, exact_obj); 3148 obj = exact_obj; 3149 } 3150 } 3151 return obj; 3152 } 3153 3154 //-------------------------------gen_instanceof-------------------------------- 3155 // Generate an instance-of idiom. Used by both the instance-of bytecode 3156 // and the reflective instance-of call. 3157 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3158 kill_dead_locals(); // Benefit all the uncommon traps 3159 assert( !stopped(), "dead parse path should be checked in callers" ); 3160 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3161 "must check for not-null not-dead klass in callers"); 3162 3163 // Make the merge point 3164 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3165 RegionNode* region = new RegionNode(PATH_LIMIT); 3166 Node* phi = new PhiNode(region, TypeInt::BOOL); 3167 C->set_has_split_ifs(true); // Has chance for split-if optimization 3168 3169 ciProfileData* data = nullptr; 3170 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3171 data = method()->method_data()->bci_to_data(bci()); 3172 } 3173 bool speculative_not_null = false; 3174 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3175 && seems_never_null(obj, data, speculative_not_null)); 3176 3177 // Null check; get casted pointer; set region slot 3 3178 Node* null_ctl = top(); 3179 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3180 3181 // If not_null_obj is dead, only null-path is taken 3182 if (stopped()) { // Doing instance-of on a null? 3183 set_control(null_ctl); 3184 return intcon(0); 3185 } 3186 region->init_req(_null_path, null_ctl); 3187 phi ->init_req(_null_path, intcon(0)); // Set null path value 3188 if (null_ctl == top()) { 3189 // Do this eagerly, so that pattern matches like is_diamond_phi 3190 // will work even during parsing. 3191 assert(_null_path == PATH_LIMIT-1, "delete last"); 3192 region->del_req(_null_path); 3193 phi ->del_req(_null_path); 3194 } 3195 3196 // Do we know the type check always succeed? 3197 bool known_statically = false; 3198 if (_gvn.type(superklass)->singleton()) { 3199 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr(); 3200 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type(); 3201 if (subk->is_loaded()) { 3202 int static_res = C->static_subtype_check(superk, subk); 3203 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3204 } 3205 } 3206 3207 if (!known_statically) { 3208 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3209 // We may not have profiling here or it may not help us. If we 3210 // have a speculative type use it to perform an exact cast. 3211 ciKlass* spec_obj_type = obj_type->speculative_type(); 3212 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) { 3213 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace); 3214 if (stopped()) { // Profile disagrees with this path. 3215 set_control(null_ctl); // Null is the only remaining possibility. 3216 return intcon(0); 3217 } 3218 if (cast_obj != nullptr) { 3219 not_null_obj = cast_obj; 3220 } 3221 } 3222 } 3223 3224 // Generate the subtype check 3225 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3226 3227 // Plug in the success path to the general merge in slot 1. 3228 region->init_req(_obj_path, control()); 3229 phi ->init_req(_obj_path, intcon(1)); 3230 3231 // Plug in the failing path to the general merge in slot 2. 3232 region->init_req(_fail_path, not_subtype_ctrl); 3233 phi ->init_req(_fail_path, intcon(0)); 3234 3235 // Return final merged results 3236 set_control( _gvn.transform(region) ); 3237 record_for_igvn(region); 3238 3239 // If we know the type check always succeeds then we don't use the 3240 // profiling data at this bytecode. Don't lose it, feed it to the 3241 // type system as a speculative type. 3242 if (safe_for_replace) { 3243 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3244 replace_in_map(obj, casted_obj); 3245 } 3246 3247 return _gvn.transform(phi); 3248 } 3249 3250 //-------------------------------gen_checkcast--------------------------------- 3251 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3252 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3253 // uncommon-trap paths work. Adjust stack after this call. 3254 // If failure_control is supplied and not null, it is filled in with 3255 // the control edge for the cast failure. Otherwise, an appropriate 3256 // uncommon trap or exception is thrown. 3257 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3258 Node* *failure_control) { 3259 kill_dead_locals(); // Benefit all the uncommon traps 3260 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr()->try_improve(); 3261 const TypeOopPtr *toop = tk->cast_to_exactness(false)->as_instance_type(); 3262 3263 // Fast cutout: Check the case that the cast is vacuously true. 3264 // This detects the common cases where the test will short-circuit 3265 // away completely. We do this before we perform the null check, 3266 // because if the test is going to turn into zero code, we don't 3267 // want a residual null check left around. (Causes a slowdown, 3268 // for example, in some objArray manipulations, such as a[i]=a[j].) 3269 if (tk->singleton()) { 3270 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3271 if (objtp != nullptr) { 3272 switch (C->static_subtype_check(tk, objtp->as_klass_type())) { 3273 case Compile::SSC_always_true: 3274 // If we know the type check always succeed then we don't use 3275 // the profiling data at this bytecode. Don't lose it, feed it 3276 // to the type system as a speculative type. 3277 return record_profiled_receiver_for_speculation(obj); 3278 case Compile::SSC_always_false: 3279 // It needs a null check because a null will *pass* the cast check. 3280 // A non-null value will always produce an exception. 3281 if (!objtp->maybe_null()) { 3282 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3283 Deoptimization::DeoptReason reason = is_aastore ? 3284 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3285 builtin_throw(reason); 3286 return top(); 3287 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3288 return null_assert(obj); 3289 } 3290 break; // Fall through to full check 3291 default: 3292 break; 3293 } 3294 } 3295 } 3296 3297 ciProfileData* data = nullptr; 3298 bool safe_for_replace = false; 3299 if (failure_control == nullptr) { // use MDO in regular case only 3300 assert(java_bc() == Bytecodes::_aastore || 3301 java_bc() == Bytecodes::_checkcast, 3302 "interpreter profiles type checks only for these BCs"); 3303 data = method()->method_data()->bci_to_data(bci()); 3304 safe_for_replace = true; 3305 } 3306 3307 // Make the merge point 3308 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3309 RegionNode* region = new RegionNode(PATH_LIMIT); 3310 Node* phi = new PhiNode(region, toop); 3311 C->set_has_split_ifs(true); // Has chance for split-if optimization 3312 3313 // Use null-cast information if it is available 3314 bool speculative_not_null = false; 3315 bool never_see_null = ((failure_control == nullptr) // regular case only 3316 && seems_never_null(obj, data, speculative_not_null)); 3317 3318 // Null check; get casted pointer; set region slot 3 3319 Node* null_ctl = top(); 3320 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3321 3322 // If not_null_obj is dead, only null-path is taken 3323 if (stopped()) { // Doing instance-of on a null? 3324 set_control(null_ctl); 3325 return null(); 3326 } 3327 region->init_req(_null_path, null_ctl); 3328 phi ->init_req(_null_path, null()); // Set null path value 3329 if (null_ctl == top()) { 3330 // Do this eagerly, so that pattern matches like is_diamond_phi 3331 // will work even during parsing. 3332 assert(_null_path == PATH_LIMIT-1, "delete last"); 3333 region->del_req(_null_path); 3334 phi ->del_req(_null_path); 3335 } 3336 3337 Node* cast_obj = nullptr; 3338 if (tk->klass_is_exact()) { 3339 // The following optimization tries to statically cast the speculative type of the object 3340 // (for example obtained during profiling) to the type of the superklass and then do a 3341 // dynamic check that the type of the object is what we expect. To work correctly 3342 // for checkcast and aastore the type of superklass should be exact. 3343 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3344 // We may not have profiling here or it may not help us. If we have 3345 // a speculative type use it to perform an exact cast. 3346 ciKlass* spec_obj_type = obj_type->speculative_type(); 3347 if (spec_obj_type != nullptr || data != nullptr) { 3348 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk, spec_obj_type, safe_for_replace); 3349 if (cast_obj != nullptr) { 3350 if (failure_control != nullptr) // failure is now impossible 3351 (*failure_control) = top(); 3352 // adjust the type of the phi to the exact klass: 3353 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3354 } 3355 } 3356 } 3357 3358 if (cast_obj == nullptr) { 3359 // Generate the subtype check 3360 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass ); 3361 3362 // Plug in success path into the merge 3363 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3364 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3365 if (failure_control == nullptr) { 3366 if (not_subtype_ctrl != top()) { // If failure is possible 3367 PreserveJVMState pjvms(this); 3368 set_control(not_subtype_ctrl); 3369 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3370 Deoptimization::DeoptReason reason = is_aastore ? 3371 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3372 builtin_throw(reason); 3373 } 3374 } else { 3375 (*failure_control) = not_subtype_ctrl; 3376 } 3377 } 3378 3379 region->init_req(_obj_path, control()); 3380 phi ->init_req(_obj_path, cast_obj); 3381 3382 // A merge of null or Casted-NotNull obj 3383 Node* res = _gvn.transform(phi); 3384 3385 // Note I do NOT always 'replace_in_map(obj,result)' here. 3386 // if( tk->klass()->can_be_primary_super() ) 3387 // This means that if I successfully store an Object into an array-of-String 3388 // I 'forget' that the Object is really now known to be a String. I have to 3389 // do this because we don't have true union types for interfaces - if I store 3390 // a Baz into an array-of-Interface and then tell the optimizer it's an 3391 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3392 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3393 // replace_in_map( obj, res ); 3394 3395 // Return final merged results 3396 set_control( _gvn.transform(region) ); 3397 record_for_igvn(region); 3398 3399 return record_profiled_receiver_for_speculation(res); 3400 } 3401 3402 //------------------------------next_monitor----------------------------------- 3403 // What number should be given to the next monitor? 3404 int GraphKit::next_monitor() { 3405 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3406 int next = current + C->sync_stack_slots(); 3407 // Keep the toplevel high water mark current: 3408 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3409 return current; 3410 } 3411 3412 //------------------------------insert_mem_bar--------------------------------- 3413 // Memory barrier to avoid floating things around 3414 // The membar serves as a pinch point between both control and all memory slices. 3415 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3416 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3417 mb->init_req(TypeFunc::Control, control()); 3418 mb->init_req(TypeFunc::Memory, reset_memory()); 3419 Node* membar = _gvn.transform(mb); 3420 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3421 set_all_memory_call(membar); 3422 return membar; 3423 } 3424 3425 //-------------------------insert_mem_bar_volatile---------------------------- 3426 // Memory barrier to avoid floating things around 3427 // The membar serves as a pinch point between both control and memory(alias_idx). 3428 // If you want to make a pinch point on all memory slices, do not use this 3429 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3430 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3431 // When Parse::do_put_xxx updates a volatile field, it appends a series 3432 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3433 // The first membar is on the same memory slice as the field store opcode. 3434 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3435 // All the other membars (for other volatile slices, including AliasIdxBot, 3436 // which stands for all unknown volatile slices) are control-dependent 3437 // on the first membar. This prevents later volatile loads or stores 3438 // from sliding up past the just-emitted store. 3439 3440 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3441 mb->set_req(TypeFunc::Control,control()); 3442 if (alias_idx == Compile::AliasIdxBot) { 3443 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3444 } else { 3445 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3446 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3447 } 3448 Node* membar = _gvn.transform(mb); 3449 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3450 if (alias_idx == Compile::AliasIdxBot) { 3451 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3452 } else { 3453 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3454 } 3455 return membar; 3456 } 3457 3458 //------------------------------shared_lock------------------------------------ 3459 // Emit locking code. 3460 FastLockNode* GraphKit::shared_lock(Node* obj) { 3461 // bci is either a monitorenter bc or InvocationEntryBci 3462 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3463 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3464 3465 if( !GenerateSynchronizationCode ) 3466 return nullptr; // Not locking things? 3467 if (stopped()) // Dead monitor? 3468 return nullptr; 3469 3470 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3471 3472 // Box the stack location 3473 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3474 Node* mem = reset_memory(); 3475 3476 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3477 3478 // Create the rtm counters for this fast lock if needed. 3479 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3480 3481 // Add monitor to debug info for the slow path. If we block inside the 3482 // slow path and de-opt, we need the monitor hanging around 3483 map()->push_monitor( flock ); 3484 3485 const TypeFunc *tf = LockNode::lock_type(); 3486 LockNode *lock = new LockNode(C, tf); 3487 3488 lock->init_req( TypeFunc::Control, control() ); 3489 lock->init_req( TypeFunc::Memory , mem ); 3490 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3491 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3492 lock->init_req( TypeFunc::ReturnAdr, top() ); 3493 3494 lock->init_req(TypeFunc::Parms + 0, obj); 3495 lock->init_req(TypeFunc::Parms + 1, box); 3496 lock->init_req(TypeFunc::Parms + 2, flock); 3497 add_safepoint_edges(lock); 3498 3499 lock = _gvn.transform( lock )->as_Lock(); 3500 3501 // lock has no side-effects, sets few values 3502 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3503 3504 insert_mem_bar(Op_MemBarAcquireLock); 3505 3506 // Add this to the worklist so that the lock can be eliminated 3507 record_for_igvn(lock); 3508 3509 #ifndef PRODUCT 3510 if (PrintLockStatistics) { 3511 // Update the counter for this lock. Don't bother using an atomic 3512 // operation since we don't require absolute accuracy. 3513 lock->create_lock_counter(map()->jvms()); 3514 increment_counter(lock->counter()->addr()); 3515 } 3516 #endif 3517 3518 return flock; 3519 } 3520 // Clone LockNode for PEA materialization. 3521 // LockNode is a safepoint, so it's location sensitive. We can't just clone it. 3522 // We create it using the current JVMState and mark bytecode non-reexecute. 3523 // 3524 // It is worthy noting that we reuse BoxNode. It represents the slot on stack. 3525 // PhaseMacroExpand::mark_eliminated_box() can't eliminate BoxLockNode and its 3526 // associated AbstractLockNodes until share prove that the 'eliminated' BoxLockNode 3527 // is exclusive. 3528 void GraphKit::clone_shared_lock(Node* box, Node* obj) { 3529 kill_dead_locals(); 3530 3531 Node* mem = reset_memory(); 3532 3533 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3534 const TypeFunc *tf = LockNode::lock_type(); 3535 LockNode *lock = new LockNode(C, tf); 3536 3537 lock->set_req(TypeFunc::Control, control()); 3538 lock->set_req(TypeFunc::Memory , mem); 3539 lock->set_req(TypeFunc::I_O , top()) ; // does no i/o 3540 lock->set_req(TypeFunc::FramePtr, frameptr()); 3541 lock->set_req(TypeFunc::ReturnAdr, top()); 3542 3543 lock->set_req(TypeFunc::Parms + 0, obj); 3544 lock->set_req(TypeFunc::Parms + 1, box); 3545 lock->set_req(TypeFunc::Parms + 2, flock); 3546 3547 // we can't reexecute current bc. it's not monitorenter. 3548 jvms()->set_should_reexecute(false); 3549 add_safepoint_edges(lock); 3550 3551 lock = _gvn.transform( lock )->as_Lock(); 3552 3553 // lock has no side-effects, sets few values 3554 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3555 3556 insert_mem_bar(Op_MemBarAcquireLock); 3557 3558 // Add this to the worklist so that the lock can be eliminated 3559 record_for_igvn(lock); 3560 } 3561 3562 //------------------------------shared_unlock---------------------------------- 3563 // Emit unlocking code. 3564 void GraphKit::shared_unlock(Node* box, Node* obj, bool preserve_monitor) { 3565 // bci is either a monitorenter bc or InvocationEntryBci 3566 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3567 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3568 3569 if( !GenerateSynchronizationCode ) 3570 return; 3571 if (stopped()) { // Dead monitor? 3572 if (!preserve_monitor) { 3573 map()->pop_monitor(); // Kill monitor from debug info 3574 } 3575 return; 3576 } 3577 3578 // Memory barrier to avoid floating things down past the locked region 3579 insert_mem_bar(Op_MemBarReleaseLock); 3580 3581 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3582 UnlockNode *unlock = new UnlockNode(C, tf); 3583 #ifdef ASSERT 3584 unlock->set_dbg_jvms(sync_jvms()); 3585 #endif 3586 uint raw_idx = Compile::AliasIdxRaw; 3587 unlock->init_req( TypeFunc::Control, control() ); 3588 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3589 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3590 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3591 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3592 3593 unlock->init_req(TypeFunc::Parms + 0, obj); 3594 unlock->init_req(TypeFunc::Parms + 1, box); 3595 unlock = _gvn.transform(unlock)->as_Unlock(); 3596 3597 Node* mem = reset_memory(); 3598 3599 // unlock has no side-effects, sets few values 3600 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3601 if (!preserve_monitor) { 3602 // Kill monitor from debug info 3603 map()->pop_monitor(); 3604 } 3605 } 3606 3607 //-------------------------------get_layout_helper----------------------------- 3608 // If the given klass is a constant or known to be an array, 3609 // fetch the constant layout helper value into constant_value 3610 // and return null. Otherwise, load the non-constant 3611 // layout helper value, and return the node which represents it. 3612 // This two-faced routine is useful because allocation sites 3613 // almost always feature constant types. 3614 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3615 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr(); 3616 if (!StressReflectiveCode && klass_t != nullptr) { 3617 bool xklass = klass_t->klass_is_exact(); 3618 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) { 3619 jint lhelper; 3620 if (klass_t->isa_aryklassptr()) { 3621 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type(); 3622 if (is_reference_type(elem, true)) { 3623 elem = T_OBJECT; 3624 } 3625 lhelper = Klass::array_layout_helper(elem); 3626 } else { 3627 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper(); 3628 } 3629 if (lhelper != Klass::_lh_neutral_value) { 3630 constant_value = lhelper; 3631 return (Node*) nullptr; 3632 } 3633 } 3634 } 3635 constant_value = Klass::_lh_neutral_value; // put in a known value 3636 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3637 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3638 } 3639 3640 // We just put in an allocate/initialize with a big raw-memory effect. 3641 // Hook selected additional alias categories on the initialization. 3642 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3643 MergeMemNode* init_in_merge, 3644 Node* init_out_raw) { 3645 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3646 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3647 3648 Node* prevmem = kit.memory(alias_idx); 3649 init_in_merge->set_memory_at(alias_idx, prevmem); 3650 kit.set_memory(init_out_raw, alias_idx); 3651 } 3652 Node* GraphKit::set_output_for_allocation_common(AllocateNode* alloc, 3653 const TypeOopPtr* oop_type, 3654 bool deoptimize_on_exception) { 3655 int rawidx = Compile::AliasIdxRaw; 3656 Node* allocx = _gvn.transform(alloc); 3657 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3658 // create memory projection for i_o 3659 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3660 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3661 3662 // create a memory projection as for the normal control path 3663 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3664 set_memory(malloc, rawidx); 3665 3666 // a normal slow-call doesn't change i_o, but an allocation does 3667 // we create a separate i_o projection for the normal control path 3668 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3669 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3670 3671 // put in an initialization barrier 3672 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3673 rawoop)->as_Initialize(); 3674 assert(alloc->initialization() == init, "2-way macro link must work"); 3675 assert(init ->allocation() == alloc, "2-way macro link must work"); 3676 { 3677 // Extract memory strands which may participate in the new object's 3678 // initialization, and source them from the new InitializeNode. 3679 // This will allow us to observe initializations when they occur, 3680 // and link them properly (as a group) to the InitializeNode. 3681 assert(init->in(InitializeNode::Memory) == malloc, ""); 3682 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3683 init->set_req(InitializeNode::Memory, minit_in); 3684 record_for_igvn(minit_in); // fold it up later, if possible 3685 Node* minit_out = memory(rawidx); 3686 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3687 // Add an edge in the MergeMem for the header fields so an access 3688 // to one of those has correct memory state 3689 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 3690 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 3691 if (oop_type->isa_aryptr()) { 3692 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3693 int elemidx = C->get_alias_index(telemref); 3694 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3695 } else if (oop_type->isa_instptr()) { 3696 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass(); 3697 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3698 ciField* field = ik->nonstatic_field_at(i); 3699 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize) 3700 continue; // do not bother to track really large numbers of fields 3701 // Find (or create) the alias category for this field: 3702 int fieldidx = C->alias_type(field)->index(); 3703 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3704 } 3705 } 3706 } 3707 3708 // Cast raw oop to the real thing... 3709 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3710 javaoop = _gvn.transform(javaoop); 3711 C->set_recent_alloc(control(), javaoop); 3712 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3713 3714 #ifdef ASSERT 3715 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3716 assert(AllocateNode::Ideal_allocation(rawoop) == alloc, 3717 "Ideal_allocation works"); 3718 assert(AllocateNode::Ideal_allocation(javaoop) == alloc, 3719 "Ideal_allocation works"); 3720 if (alloc->is_AllocateArray()) { 3721 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(), 3722 "Ideal_allocation works"); 3723 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(), 3724 "Ideal_allocation works"); 3725 } else { 3726 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3727 } 3728 } 3729 #endif //ASSERT 3730 3731 return javaoop; 3732 } 3733 3734 //---------------------------set_output_for_allocation------------------------- 3735 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3736 const TypeOopPtr* oop_type, 3737 bool deoptimize_on_exception) { 3738 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3739 add_safepoint_edges(alloc); 3740 return set_output_for_allocation_common(alloc, oop_type, deoptimize_on_exception); 3741 } 3742 3743 // 3744 // Position-Agnostic Materialization 3745 // ------------------------------------- 3746 // When PEA materializes a virtual object, it emits a cluster of nodes in the current position. 3747 // Unlike ordinary "floating" nodes, an AllocateNode is a subclass of SafePointNode so it is 3748 // dependent on JVMState. The JVMState of current position may not fit for an AllocateNode. 3749 // 3750 // To ensure we can safely embed nodes into the curren position, we have the following 3751 // measures: 3752 // 3753 // 1. Debug edges and JVMState of the cloned AllocateNode are not from current GraphKit. 3754 // We copy them from the original AllocateNode instead. 3755 // 3756 // 2. Choose deoptimization on exception. A real exception may be mistakenly dispatched to 3757 // the exception handler in current context. 3758 // 3759 Node* GraphKit::materialize_object(AllocateNode* alloc, const TypeOopPtr* oop_type) { 3760 Node *mem = reset_memory(); 3761 AllocateNode* allocx = new AllocateNode(C, alloc->tf(), control(), mem, i_o(), 3762 alloc->in(AllocateNode::AllocSize), 3763 alloc->in(AllocateNode::KlassNode), 3764 alloc->in(AllocateNode::InitialTest)); 3765 allocx->set_req(TypeFunc::FramePtr, frameptr()); 3766 3767 JVMState* out_jvms = alloc->jvms()->clone_shallow(C); 3768 out_jvms->bind_map(allocx); 3769 // copy all debuginfo edges from the original AllocateNode 3770 for (uint i=allocx->req(); i < alloc->req(); ++i) { 3771 allocx->add_req(alloc->in(i)); 3772 } 3773 3774 JVMState* jvms = sync_jvms(); 3775 // We can not use PreserveJVMState here because 'this' is a Parse. We would fail in jvms_in_sync(). 3776 GraphKit kit(allocx->jvms()); 3777 kit.set_map_clone(allocx); 3778 3779 Node* objx = kit.set_output_for_allocation_common(allocx, oop_type, true /*deoptimize_on_ex*/); 3780 3781 // copy back compile-time state to 'this'. 3782 set_jvms(jvms); 3783 set_control(kit.control()); 3784 set_i_o(kit.i_o()); 3785 set_all_memory(kit.merged_memory()); 3786 3787 return objx; 3788 } 3789 3790 //---------------------------new_instance-------------------------------------- 3791 // This routine takes a klass_node which may be constant (for a static type) 3792 // or may be non-constant (for reflective code). It will work equally well 3793 // for either, and the graph will fold nicely if the optimizer later reduces 3794 // the type to a constant. 3795 // The optional arguments are for specialized use by intrinsics: 3796 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3797 // - If 'return_size_val', report the total object size to the caller. 3798 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3799 Node* GraphKit::new_instance(Node* klass_node, 3800 Node* extra_slow_test, 3801 Node* *return_size_val, 3802 bool deoptimize_on_exception) { 3803 // Compute size in doublewords 3804 // The size is always an integral number of doublewords, represented 3805 // as a positive bytewise size stored in the klass's layout_helper. 3806 // The layout_helper also encodes (in a low bit) the need for a slow path. 3807 jint layout_con = Klass::_lh_neutral_value; 3808 Node* layout_val = get_layout_helper(klass_node, layout_con); 3809 int layout_is_con = (layout_val == nullptr); 3810 3811 if (extra_slow_test == nullptr) extra_slow_test = intcon(0); 3812 // Generate the initial go-slow test. It's either ALWAYS (return a 3813 // Node for 1) or NEVER (return a null) or perhaps (in the reflective 3814 // case) a computed value derived from the layout_helper. 3815 Node* initial_slow_test = nullptr; 3816 if (layout_is_con) { 3817 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3818 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3819 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3820 } else { // reflective case 3821 // This reflective path is used by Unsafe.allocateInstance. 3822 // (It may be stress-tested by specifying StressReflectiveCode.) 3823 // Basically, we want to get into the VM is there's an illegal argument. 3824 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3825 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3826 if (extra_slow_test != intcon(0)) { 3827 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3828 } 3829 // (Macro-expander will further convert this to a Bool, if necessary.) 3830 } 3831 3832 // Find the size in bytes. This is easy; it's the layout_helper. 3833 // The size value must be valid even if the slow path is taken. 3834 Node* size = nullptr; 3835 if (layout_is_con) { 3836 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3837 } else { // reflective case 3838 // This reflective path is used by clone and Unsafe.allocateInstance. 3839 size = ConvI2X(layout_val); 3840 3841 // Clear the low bits to extract layout_helper_size_in_bytes: 3842 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3843 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3844 size = _gvn.transform( new AndXNode(size, mask) ); 3845 } 3846 if (return_size_val != nullptr) { 3847 (*return_size_val) = size; 3848 } 3849 3850 // This is a precise notnull oop of the klass. 3851 // (Actually, it need not be precise if this is a reflective allocation.) 3852 // It's what we cast the result to. 3853 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3854 if (!tklass) tklass = TypeInstKlassPtr::OBJECT; 3855 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3856 3857 // Now generate allocation code 3858 3859 // The entire memory state is needed for slow path of the allocation 3860 // since GC and deoptimization can happened. 3861 Node *mem = reset_memory(); 3862 set_all_memory(mem); // Create new memory state 3863 3864 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3865 control(), mem, i_o(), 3866 size, klass_node, 3867 initial_slow_test); 3868 3869 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3870 } 3871 3872 //-------------------------------new_array------------------------------------- 3873 // helper for both newarray and anewarray 3874 // The 'length' parameter is (obviously) the length of the array. 3875 // The optional arguments are for specialized use by intrinsics: 3876 // - If 'return_size_val', report the non-padded array size (sum of header size 3877 // and array body) to the caller. 3878 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3879 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3880 Node* length, // number of array elements 3881 int nargs, // number of arguments to push back for uncommon trap 3882 Node* *return_size_val, 3883 bool deoptimize_on_exception) { 3884 jint layout_con = Klass::_lh_neutral_value; 3885 Node* layout_val = get_layout_helper(klass_node, layout_con); 3886 int layout_is_con = (layout_val == nullptr); 3887 3888 if (!layout_is_con && !StressReflectiveCode && 3889 !too_many_traps(Deoptimization::Reason_class_check)) { 3890 // This is a reflective array creation site. 3891 // Optimistically assume that it is a subtype of Object[], 3892 // so that we can fold up all the address arithmetic. 3893 layout_con = Klass::array_layout_helper(T_OBJECT); 3894 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3895 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3896 { BuildCutout unless(this, bol_lh, PROB_MAX); 3897 inc_sp(nargs); 3898 uncommon_trap(Deoptimization::Reason_class_check, 3899 Deoptimization::Action_maybe_recompile); 3900 } 3901 layout_val = nullptr; 3902 layout_is_con = true; 3903 } 3904 3905 // Generate the initial go-slow test. Make sure we do not overflow 3906 // if length is huge (near 2Gig) or negative! We do not need 3907 // exact double-words here, just a close approximation of needed 3908 // double-words. We can't add any offset or rounding bits, lest we 3909 // take a size -1 of bytes and make it positive. Use an unsigned 3910 // compare, so negative sizes look hugely positive. 3911 int fast_size_limit = FastAllocateSizeLimit; 3912 if (layout_is_con) { 3913 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3914 // Increase the size limit if we have exact knowledge of array type. 3915 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3916 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3917 } 3918 3919 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3920 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3921 3922 // --- Size Computation --- 3923 // array_size = round_to_heap(array_header + (length << elem_shift)); 3924 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3925 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3926 // The rounding mask is strength-reduced, if possible. 3927 int round_mask = MinObjAlignmentInBytes - 1; 3928 Node* header_size = nullptr; 3929 // (T_BYTE has the weakest alignment and size restrictions...) 3930 if (layout_is_con) { 3931 int hsize = Klass::layout_helper_header_size(layout_con); 3932 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3933 if ((round_mask & ~right_n_bits(eshift)) == 0) 3934 round_mask = 0; // strength-reduce it if it goes away completely 3935 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3936 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3937 assert(header_size_min <= hsize, "generic minimum is smallest"); 3938 header_size = intcon(hsize); 3939 } else { 3940 Node* hss = intcon(Klass::_lh_header_size_shift); 3941 Node* hsm = intcon(Klass::_lh_header_size_mask); 3942 header_size = _gvn.transform(new URShiftINode(layout_val, hss)); 3943 header_size = _gvn.transform(new AndINode(header_size, hsm)); 3944 } 3945 3946 Node* elem_shift = nullptr; 3947 if (layout_is_con) { 3948 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3949 if (eshift != 0) 3950 elem_shift = intcon(eshift); 3951 } else { 3952 // There is no need to mask or shift this value. 3953 // The semantics of LShiftINode include an implicit mask to 0x1F. 3954 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3955 elem_shift = layout_val; 3956 } 3957 3958 // Transition to native address size for all offset calculations: 3959 Node* lengthx = ConvI2X(length); 3960 Node* headerx = ConvI2X(header_size); 3961 #ifdef _LP64 3962 { const TypeInt* tilen = _gvn.find_int_type(length); 3963 if (tilen != nullptr && tilen->_lo < 0) { 3964 // Add a manual constraint to a positive range. Cf. array_element_address. 3965 jint size_max = fast_size_limit; 3966 if (size_max > tilen->_hi) size_max = tilen->_hi; 3967 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3968 3969 // Only do a narrow I2L conversion if the range check passed. 3970 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3971 _gvn.transform(iff); 3972 RegionNode* region = new RegionNode(3); 3973 _gvn.set_type(region, Type::CONTROL); 3974 lengthx = new PhiNode(region, TypeLong::LONG); 3975 _gvn.set_type(lengthx, TypeLong::LONG); 3976 3977 // Range check passed. Use ConvI2L node with narrow type. 3978 Node* passed = IfFalse(iff); 3979 region->init_req(1, passed); 3980 // Make I2L conversion control dependent to prevent it from 3981 // floating above the range check during loop optimizations. 3982 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3983 3984 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3985 region->init_req(2, IfTrue(iff)); 3986 lengthx->init_req(2, ConvI2X(length)); 3987 3988 set_control(region); 3989 record_for_igvn(region); 3990 record_for_igvn(lengthx); 3991 } 3992 } 3993 #endif 3994 3995 // Combine header size and body size for the array copy part, then align (if 3996 // necessary) for the allocation part. This computation cannot overflow, 3997 // because it is used only in two places, one where the length is sharply 3998 // limited, and the other after a successful allocation. 3999 Node* abody = lengthx; 4000 if (elem_shift != nullptr) { 4001 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift)); 4002 } 4003 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody)); 4004 4005 if (return_size_val != nullptr) { 4006 // This is the size 4007 (*return_size_val) = non_rounded_size; 4008 } 4009 4010 Node* size = non_rounded_size; 4011 if (round_mask != 0) { 4012 Node* mask1 = MakeConX(round_mask); 4013 size = _gvn.transform(new AddXNode(size, mask1)); 4014 Node* mask2 = MakeConX(~round_mask); 4015 size = _gvn.transform(new AndXNode(size, mask2)); 4016 } 4017 // else if round_mask == 0, the size computation is self-rounding 4018 4019 // Now generate allocation code 4020 4021 // The entire memory state is needed for slow path of the allocation 4022 // since GC and deoptimization can happened. 4023 Node *mem = reset_memory(); 4024 set_all_memory(mem); // Create new memory state 4025 4026 if (initial_slow_test->is_Bool()) { 4027 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4028 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4029 } 4030 4031 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 4032 Node* valid_length_test = _gvn.intcon(1); 4033 if (ary_type->isa_aryptr()) { 4034 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type(); 4035 jint max = TypeAryPtr::max_array_length(bt); 4036 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max))); 4037 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le)); 4038 } 4039 4040 // Create the AllocateArrayNode and its result projections 4041 AllocateArrayNode* alloc 4042 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4043 control(), mem, i_o(), 4044 size, klass_node, 4045 initial_slow_test, 4046 length, valid_length_test); 4047 4048 // Cast to correct type. Note that the klass_node may be constant or not, 4049 // and in the latter case the actual array type will be inexact also. 4050 // (This happens via a non-constant argument to inline_native_newArray.) 4051 // In any case, the value of klass_node provides the desired array type. 4052 const TypeInt* length_type = _gvn.find_int_type(length); 4053 if (ary_type->isa_aryptr() && length_type != nullptr) { 4054 // Try to get a better type than POS for the size 4055 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4056 } 4057 4058 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4059 4060 array_ideal_length(alloc, ary_type, true); 4061 return javaoop; 4062 } 4063 4064 // The following "Ideal_foo" functions are placed here because they recognize 4065 // the graph shapes created by the functions immediately above. 4066 4067 //---------------------------Ideal_allocation---------------------------------- 4068 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4069 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) { 4070 if (ptr == nullptr) { // reduce dumb test in callers 4071 return nullptr; 4072 } 4073 4074 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4075 ptr = bs->step_over_gc_barrier(ptr); 4076 4077 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4078 ptr = ptr->in(1); 4079 if (ptr == nullptr) return nullptr; 4080 } 4081 // Return null for allocations with several casts: 4082 // j.l.reflect.Array.newInstance(jobject, jint) 4083 // Object.clone() 4084 // to keep more precise type from last cast. 4085 if (ptr->is_Proj()) { 4086 Node* allo = ptr->in(0); 4087 if (allo != nullptr && allo->is_Allocate()) { 4088 return allo->as_Allocate(); 4089 } 4090 } 4091 // Report failure to match. 4092 return nullptr; 4093 } 4094 4095 // Fancy version which also strips off an offset (and reports it to caller). 4096 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase, 4097 intptr_t& offset) { 4098 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4099 if (base == nullptr) return nullptr; 4100 return Ideal_allocation(base); 4101 } 4102 4103 // Trace Initialize <- Proj[Parm] <- Allocate 4104 AllocateNode* InitializeNode::allocation() { 4105 Node* rawoop = in(InitializeNode::RawAddress); 4106 if (rawoop->is_Proj()) { 4107 Node* alloc = rawoop->in(0); 4108 if (alloc->is_Allocate()) { 4109 return alloc->as_Allocate(); 4110 } 4111 } 4112 return nullptr; 4113 } 4114 4115 // Trace Allocate -> Proj[Parm] -> Initialize 4116 InitializeNode* AllocateNode::initialization() { 4117 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4118 if (rawoop == nullptr) return nullptr; 4119 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4120 Node* init = rawoop->fast_out(i); 4121 if (init->is_Initialize()) { 4122 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4123 return init->as_Initialize(); 4124 } 4125 } 4126 return nullptr; 4127 } 4128 4129 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path. 4130 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) { 4131 // Too many traps seen? 4132 if (too_many_traps(reason)) { 4133 #ifdef ASSERT 4134 if (TraceLoopPredicate) { 4135 int tc = C->trap_count(reason); 4136 tty->print("too many traps=%s tcount=%d in ", 4137 Deoptimization::trap_reason_name(reason), tc); 4138 method()->print(); // which method has too many predicate traps 4139 tty->cr(); 4140 } 4141 #endif 4142 // We cannot afford to take more traps here, 4143 // do not generate Parse Predicate. 4144 return; 4145 } 4146 4147 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn); 4148 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn)); 4149 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate)); 4150 { 4151 PreserveJVMState pjvms(this); 4152 set_control(if_false); 4153 inc_sp(nargs); 4154 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4155 } 4156 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate)); 4157 set_control(if_true); 4158 } 4159 4160 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All 4161 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate. 4162 void GraphKit::add_parse_predicates(int nargs) { 4163 if (UseLoopPredicate) { 4164 add_parse_predicate(Deoptimization::Reason_predicate, nargs); 4165 } 4166 if (UseProfiledLoopPredicate) { 4167 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs); 4168 } 4169 // Loop Limit Check Predicate should be near the loop. 4170 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs); 4171 } 4172 4173 void GraphKit::sync_kit(IdealKit& ideal) { 4174 set_all_memory(ideal.merged_memory()); 4175 set_i_o(ideal.i_o()); 4176 set_control(ideal.ctrl()); 4177 } 4178 4179 void GraphKit::final_sync(IdealKit& ideal) { 4180 // Final sync IdealKit and graphKit. 4181 sync_kit(ideal); 4182 } 4183 4184 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4185 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4186 Node* coder = load_String_coder(str, set_ctrl); 4187 // Divide length by 2 if coder is UTF16 4188 return _gvn.transform(new RShiftINode(len, coder)); 4189 } 4190 4191 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4192 int value_offset = java_lang_String::value_offset(); 4193 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4194 false, nullptr, 0); 4195 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4196 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4197 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4198 ciTypeArrayKlass::make(T_BYTE), true, 0); 4199 Node* p = basic_plus_adr(str, str, value_offset); 4200 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4201 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4202 return load; 4203 } 4204 4205 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4206 if (!CompactStrings) { 4207 return intcon(java_lang_String::CODER_UTF16); 4208 } 4209 int coder_offset = java_lang_String::coder_offset(); 4210 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4211 false, nullptr, 0); 4212 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4213 4214 Node* p = basic_plus_adr(str, str, coder_offset); 4215 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4216 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4217 return load; 4218 } 4219 4220 void GraphKit::store_String_value(Node* str, Node* value) { 4221 int value_offset = java_lang_String::value_offset(); 4222 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4223 false, nullptr, 0); 4224 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4225 4226 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4227 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4228 } 4229 4230 void GraphKit::store_String_coder(Node* str, Node* value) { 4231 int coder_offset = java_lang_String::coder_offset(); 4232 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4233 false, nullptr, 0); 4234 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4235 4236 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4237 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4238 } 4239 4240 // Capture src and dst memory state with a MergeMemNode 4241 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4242 if (src_type == dst_type) { 4243 // Types are equal, we don't need a MergeMemNode 4244 return memory(src_type); 4245 } 4246 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4247 record_for_igvn(merge); // fold it up later, if possible 4248 int src_idx = C->get_alias_index(src_type); 4249 int dst_idx = C->get_alias_index(dst_type); 4250 merge->set_memory_at(src_idx, memory(src_idx)); 4251 merge->set_memory_at(dst_idx, memory(dst_idx)); 4252 return merge; 4253 } 4254 4255 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4256 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4257 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4258 // If input and output memory types differ, capture both states to preserve 4259 // the dependency between preceding and subsequent loads/stores. 4260 // For example, the following program: 4261 // StoreB 4262 // compress_string 4263 // LoadB 4264 // has this memory graph (use->def): 4265 // LoadB -> compress_string -> CharMem 4266 // ... -> StoreB -> ByteMem 4267 // The intrinsic hides the dependency between LoadB and StoreB, causing 4268 // the load to read from memory not containing the result of the StoreB. 4269 // The correct memory graph should look like this: 4270 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4271 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4272 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4273 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str))); 4274 set_memory(res_mem, TypeAryPtr::BYTES); 4275 return str; 4276 } 4277 4278 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4279 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4280 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4281 // Capture src and dst memory (see comment in 'compress_string'). 4282 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4283 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4284 set_memory(_gvn.transform(str), dst_type); 4285 } 4286 4287 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4288 /** 4289 * int i_char = start; 4290 * for (int i_byte = 0; i_byte < count; i_byte++) { 4291 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4292 * } 4293 */ 4294 add_parse_predicates(); 4295 C->set_has_loops(true); 4296 4297 RegionNode* head = new RegionNode(3); 4298 head->init_req(1, control()); 4299 gvn().set_type(head, Type::CONTROL); 4300 record_for_igvn(head); 4301 4302 Node* i_byte = new PhiNode(head, TypeInt::INT); 4303 i_byte->init_req(1, intcon(0)); 4304 gvn().set_type(i_byte, TypeInt::INT); 4305 record_for_igvn(i_byte); 4306 4307 Node* i_char = new PhiNode(head, TypeInt::INT); 4308 i_char->init_req(1, start); 4309 gvn().set_type(i_char, TypeInt::INT); 4310 record_for_igvn(i_char); 4311 4312 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4313 gvn().set_type(mem, Type::MEMORY); 4314 record_for_igvn(mem); 4315 set_control(head); 4316 set_memory(mem, TypeAryPtr::BYTES); 4317 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true); 4318 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4319 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4320 false, false, true /* mismatched */); 4321 4322 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4323 head->init_req(2, IfTrue(iff)); 4324 mem->init_req(2, st); 4325 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4326 i_char->init_req(2, AddI(i_char, intcon(2))); 4327 4328 set_control(IfFalse(iff)); 4329 set_memory(st, TypeAryPtr::BYTES); 4330 } 4331 4332 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4333 if (!field->is_constant()) { 4334 return nullptr; // Field not marked as constant. 4335 } 4336 ciInstance* holder = nullptr; 4337 if (!field->is_static()) { 4338 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4339 if (const_oop != nullptr && const_oop->is_instance()) { 4340 holder = const_oop->as_instance(); 4341 } 4342 } 4343 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4344 /*is_unsigned_load=*/false); 4345 if (con_type != nullptr) { 4346 return makecon(con_type); 4347 } 4348 return nullptr; 4349 } 4350 4351 void GraphKit::backfill_materialized(SafePointNode* map, uint begin, uint end, PEAState& as){ 4352 bool printed = false; 4353 4354 for (uint i = begin; i < end; ++i) { 4355 Node* t = map->in(i); 4356 4357 if (t != nullptr && t->is_CheckCastPP()) { 4358 AllocateNode* alloc = AllocateNode::Ideal_allocation(t); 4359 4360 if (as.contains(alloc)) { 4361 Node* neww = as.get_materialized_value(alloc); 4362 if (neww != nullptr && neww != t) { 4363 #ifndef PRODUCT 4364 if (PEAVerbose) { 4365 if (!printed) { 4366 map->dump(); 4367 printed = true; 4368 } 4369 tty->print_cr("[PEA] replace %d with node %d", i, neww->_idx); 4370 } 4371 #endif 4372 map->set_req(i, neww); 4373 } 4374 } 4375 } 4376 } 4377 }