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