1 /* 2 * Copyright (c) 2001, 2025, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "asm/register.hpp" 26 #include "ci/ciFlatArrayKlass.hpp" 27 #include "ci/ciInlineKlass.hpp" 28 #include "ci/ciObjArray.hpp" 29 #include "ci/ciUtilities.hpp" 30 #include "classfile/javaClasses.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "gc/shared/barrierSet.hpp" 33 #include "gc/shared/c2/barrierSetC2.hpp" 34 #include "interpreter/interpreter.hpp" 35 #include "memory/resourceArea.hpp" 36 #include "oops/flatArrayKlass.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/castnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/graphKit.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/inlinetypenode.hpp" 43 #include "opto/intrinsicnode.hpp" 44 #include "opto/locknode.hpp" 45 #include "opto/machnode.hpp" 46 #include "opto/narrowptrnode.hpp" 47 #include "opto/opaquenode.hpp" 48 #include "opto/parse.hpp" 49 #include "opto/rootnode.hpp" 50 #include "opto/runtime.hpp" 51 #include "opto/subtypenode.hpp" 52 #include "runtime/deoptimization.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "runtime/stubRoutines.hpp" 55 #include "utilities/bitMap.inline.hpp" 56 #include "utilities/growableArray.hpp" 57 #include "utilities/powerOfTwo.hpp" 58 59 //----------------------------GraphKit----------------------------------------- 60 // Main utility constructor. 61 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 62 : Phase(Phase::Parser), 63 _env(C->env()), 64 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()), 65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 66 { 67 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled"); 68 _exceptions = jvms->map()->next_exception(); 69 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr); 70 set_jvms(jvms); 71 #ifdef ASSERT 72 if (_gvn.is_IterGVN() != nullptr) { 73 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used"); 74 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 75 _worklist_size = _gvn.C->igvn_worklist()->size(); 76 } 77 #endif 78 } 79 80 // Private constructor for parser. 81 GraphKit::GraphKit() 82 : Phase(Phase::Parser), 83 _env(C->env()), 84 _gvn(*C->initial_gvn()), 85 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 86 { 87 _exceptions = nullptr; 88 set_map(nullptr); 89 DEBUG_ONLY(_sp = -99); 90 DEBUG_ONLY(set_bci(-99)); 91 } 92 93 94 95 //---------------------------clean_stack--------------------------------------- 96 // Clear away rubbish from the stack area of the JVM state. 97 // This destroys any arguments that may be waiting on the stack. 98 void GraphKit::clean_stack(int from_sp) { 99 SafePointNode* map = this->map(); 100 JVMState* jvms = this->jvms(); 101 int stk_size = jvms->stk_size(); 102 int stkoff = jvms->stkoff(); 103 Node* top = this->top(); 104 for (int i = from_sp; i < stk_size; i++) { 105 if (map->in(stkoff + i) != top) { 106 map->set_req(stkoff + i, top); 107 } 108 } 109 } 110 111 112 //--------------------------------sync_jvms----------------------------------- 113 // Make sure our current jvms agrees with our parse state. 114 JVMState* GraphKit::sync_jvms() const { 115 JVMState* jvms = this->jvms(); 116 jvms->set_bci(bci()); // Record the new bci in the JVMState 117 jvms->set_sp(sp()); // Record the new sp in the JVMState 118 assert(jvms_in_sync(), "jvms is now in sync"); 119 return jvms; 120 } 121 122 //--------------------------------sync_jvms_for_reexecute--------------------- 123 // Make sure our current jvms agrees with our parse state. This version 124 // uses the reexecute_sp for reexecuting bytecodes. 125 JVMState* GraphKit::sync_jvms_for_reexecute() { 126 JVMState* jvms = this->jvms(); 127 jvms->set_bci(bci()); // Record the new bci in the JVMState 128 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 129 return jvms; 130 } 131 132 #ifdef ASSERT 133 bool GraphKit::jvms_in_sync() const { 134 Parse* parse = is_Parse(); 135 if (parse == nullptr) { 136 if (bci() != jvms()->bci()) return false; 137 if (sp() != (int)jvms()->sp()) return false; 138 return true; 139 } 140 if (jvms()->method() != parse->method()) return false; 141 if (jvms()->bci() != parse->bci()) return false; 142 int jvms_sp = jvms()->sp(); 143 if (jvms_sp != parse->sp()) return false; 144 int jvms_depth = jvms()->depth(); 145 if (jvms_depth != parse->depth()) return false; 146 return true; 147 } 148 149 // Local helper checks for special internal merge points 150 // used to accumulate and merge exception states. 151 // They are marked by the region's in(0) edge being the map itself. 152 // Such merge points must never "escape" into the parser at large, 153 // until they have been handed to gvn.transform. 154 static bool is_hidden_merge(Node* reg) { 155 if (reg == nullptr) return false; 156 if (reg->is_Phi()) { 157 reg = reg->in(0); 158 if (reg == nullptr) return false; 159 } 160 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root(); 161 } 162 163 void GraphKit::verify_map() const { 164 if (map() == nullptr) return; // null map is OK 165 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 166 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 167 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 168 } 169 170 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 171 assert(ex_map->next_exception() == nullptr, "not already part of a chain"); 172 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 173 } 174 #endif 175 176 //---------------------------stop_and_kill_map--------------------------------- 177 // Set _map to null, signalling a stop to further bytecode execution. 178 // First smash the current map's control to a constant, to mark it dead. 179 void GraphKit::stop_and_kill_map() { 180 SafePointNode* dead_map = stop(); 181 if (dead_map != nullptr) { 182 dead_map->disconnect_inputs(C); // Mark the map as killed. 183 assert(dead_map->is_killed(), "must be so marked"); 184 } 185 } 186 187 188 //--------------------------------stopped-------------------------------------- 189 // Tell if _map is null, or control is top. 190 bool GraphKit::stopped() { 191 if (map() == nullptr) return true; 192 else if (control() == top()) return true; 193 else return false; 194 } 195 196 197 //-----------------------------has_exception_handler---------------------------------- 198 // Tell if this method or any caller method has exception handlers. 199 bool GraphKit::has_exception_handler() { 200 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) { 201 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 202 return true; 203 } 204 } 205 return false; 206 } 207 208 //------------------------------save_ex_oop------------------------------------ 209 // Save an exception without blowing stack contents or other JVM state. 210 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 211 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 212 ex_map->add_req(ex_oop); 213 DEBUG_ONLY(verify_exception_state(ex_map)); 214 } 215 216 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 217 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 218 Node* ex_oop = ex_map->in(ex_map->req()-1); 219 if (clear_it) ex_map->del_req(ex_map->req()-1); 220 return ex_oop; 221 } 222 223 //-----------------------------saved_ex_oop------------------------------------ 224 // Recover a saved exception from its map. 225 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 226 return common_saved_ex_oop(ex_map, false); 227 } 228 229 //--------------------------clear_saved_ex_oop--------------------------------- 230 // Erase a previously saved exception from its map. 231 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 232 return common_saved_ex_oop(ex_map, true); 233 } 234 235 #ifdef ASSERT 236 //---------------------------has_saved_ex_oop---------------------------------- 237 // Erase a previously saved exception from its map. 238 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 239 return ex_map->req() == ex_map->jvms()->endoff()+1; 240 } 241 #endif 242 243 //-------------------------make_exception_state-------------------------------- 244 // Turn the current JVM state into an exception state, appending the ex_oop. 245 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 246 sync_jvms(); 247 SafePointNode* ex_map = stop(); // do not manipulate this map any more 248 set_saved_ex_oop(ex_map, ex_oop); 249 return ex_map; 250 } 251 252 253 //--------------------------add_exception_state-------------------------------- 254 // Add an exception to my list of exceptions. 255 void GraphKit::add_exception_state(SafePointNode* ex_map) { 256 if (ex_map == nullptr || ex_map->control() == top()) { 257 return; 258 } 259 #ifdef ASSERT 260 verify_exception_state(ex_map); 261 if (has_exceptions()) { 262 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 263 } 264 #endif 265 266 // If there is already an exception of exactly this type, merge with it. 267 // In particular, null-checks and other low-level exceptions common up here. 268 Node* ex_oop = saved_ex_oop(ex_map); 269 const Type* ex_type = _gvn.type(ex_oop); 270 if (ex_oop == top()) { 271 // No action needed. 272 return; 273 } 274 assert(ex_type->isa_instptr(), "exception must be an instance"); 275 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) { 276 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 277 // We check sp also because call bytecodes can generate exceptions 278 // both before and after arguments are popped! 279 if (ex_type2 == ex_type 280 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 281 combine_exception_states(ex_map, e2); 282 return; 283 } 284 } 285 286 // No pre-existing exception of the same type. Chain it on the list. 287 push_exception_state(ex_map); 288 } 289 290 //-----------------------add_exception_states_from----------------------------- 291 void GraphKit::add_exception_states_from(JVMState* jvms) { 292 SafePointNode* ex_map = jvms->map()->next_exception(); 293 if (ex_map != nullptr) { 294 jvms->map()->set_next_exception(nullptr); 295 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) { 296 next_map = ex_map->next_exception(); 297 ex_map->set_next_exception(nullptr); 298 add_exception_state(ex_map); 299 } 300 } 301 } 302 303 //-----------------------transfer_exceptions_into_jvms------------------------- 304 JVMState* GraphKit::transfer_exceptions_into_jvms() { 305 if (map() == nullptr) { 306 // We need a JVMS to carry the exceptions, but the map has gone away. 307 // Create a scratch JVMS, cloned from any of the exception states... 308 if (has_exceptions()) { 309 _map = _exceptions; 310 _map = clone_map(); 311 _map->set_next_exception(nullptr); 312 clear_saved_ex_oop(_map); 313 DEBUG_ONLY(verify_map()); 314 } else { 315 // ...or created from scratch 316 JVMState* jvms = new (C) JVMState(_method, nullptr); 317 jvms->set_bci(_bci); 318 jvms->set_sp(_sp); 319 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 320 set_jvms(jvms); 321 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 322 set_all_memory(top()); 323 while (map()->req() < jvms->endoff()) map()->add_req(top()); 324 } 325 // (This is a kludge, in case you didn't notice.) 326 set_control(top()); 327 } 328 JVMState* jvms = sync_jvms(); 329 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 330 jvms->map()->set_next_exception(_exceptions); 331 _exceptions = nullptr; // done with this set of exceptions 332 return jvms; 333 } 334 335 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 336 assert(is_hidden_merge(dstphi), "must be a special merge node"); 337 assert(is_hidden_merge(srcphi), "must be a special merge node"); 338 uint limit = srcphi->req(); 339 for (uint i = PhiNode::Input; i < limit; i++) { 340 dstphi->add_req(srcphi->in(i)); 341 } 342 } 343 static inline void add_one_req(Node* dstphi, Node* src) { 344 assert(is_hidden_merge(dstphi), "must be a special merge node"); 345 assert(!is_hidden_merge(src), "must not be a special merge node"); 346 dstphi->add_req(src); 347 } 348 349 //-----------------------combine_exception_states------------------------------ 350 // This helper function combines exception states by building phis on a 351 // specially marked state-merging region. These regions and phis are 352 // untransformed, and can build up gradually. The region is marked by 353 // having a control input of its exception map, rather than null. Such 354 // regions do not appear except in this function, and in use_exception_state. 355 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 356 if (failing_internal()) { 357 return; // dying anyway... 358 } 359 JVMState* ex_jvms = ex_map->_jvms; 360 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 361 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 362 // TODO 8325632 Re-enable 363 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 364 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 365 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 366 assert(ex_map->req() == phi_map->req(), "matching maps"); 367 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 368 Node* hidden_merge_mark = root(); 369 Node* region = phi_map->control(); 370 MergeMemNode* phi_mem = phi_map->merged_memory(); 371 MergeMemNode* ex_mem = ex_map->merged_memory(); 372 if (region->in(0) != hidden_merge_mark) { 373 // The control input is not (yet) a specially-marked region in phi_map. 374 // Make it so, and build some phis. 375 region = new RegionNode(2); 376 _gvn.set_type(region, Type::CONTROL); 377 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 378 region->init_req(1, phi_map->control()); 379 phi_map->set_control(region); 380 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 381 record_for_igvn(io_phi); 382 _gvn.set_type(io_phi, Type::ABIO); 383 phi_map->set_i_o(io_phi); 384 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 385 Node* m = mms.memory(); 386 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 387 record_for_igvn(m_phi); 388 _gvn.set_type(m_phi, Type::MEMORY); 389 mms.set_memory(m_phi); 390 } 391 } 392 393 // Either or both of phi_map and ex_map might already be converted into phis. 394 Node* ex_control = ex_map->control(); 395 // if there is special marking on ex_map also, we add multiple edges from src 396 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 397 // how wide was the destination phi_map, originally? 398 uint orig_width = region->req(); 399 400 if (add_multiple) { 401 add_n_reqs(region, ex_control); 402 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 403 } else { 404 // ex_map has no merges, so we just add single edges everywhere 405 add_one_req(region, ex_control); 406 add_one_req(phi_map->i_o(), ex_map->i_o()); 407 } 408 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 409 if (mms.is_empty()) { 410 // get a copy of the base memory, and patch some inputs into it 411 const TypePtr* adr_type = mms.adr_type(C); 412 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 413 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 414 mms.set_memory(phi); 415 // Prepare to append interesting stuff onto the newly sliced phi: 416 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 417 } 418 // Append stuff from ex_map: 419 if (add_multiple) { 420 add_n_reqs(mms.memory(), mms.memory2()); 421 } else { 422 add_one_req(mms.memory(), mms.memory2()); 423 } 424 } 425 uint limit = ex_map->req(); 426 for (uint i = TypeFunc::Parms; i < limit; i++) { 427 // Skip everything in the JVMS after tos. (The ex_oop follows.) 428 if (i == tos) i = ex_jvms->monoff(); 429 Node* src = ex_map->in(i); 430 Node* dst = phi_map->in(i); 431 if (src != dst) { 432 PhiNode* phi; 433 if (dst->in(0) != region) { 434 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 435 record_for_igvn(phi); 436 _gvn.set_type(phi, phi->type()); 437 phi_map->set_req(i, dst); 438 // Prepare to append interesting stuff onto the new phi: 439 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 440 } else { 441 assert(dst->is_Phi(), "nobody else uses a hidden region"); 442 phi = dst->as_Phi(); 443 } 444 if (add_multiple && src->in(0) == ex_control) { 445 // Both are phis. 446 add_n_reqs(dst, src); 447 } else { 448 while (dst->req() < region->req()) add_one_req(dst, src); 449 } 450 const Type* srctype = _gvn.type(src); 451 if (phi->type() != srctype) { 452 const Type* dsttype = phi->type()->meet_speculative(srctype); 453 if (phi->type() != dsttype) { 454 phi->set_type(dsttype); 455 _gvn.set_type(phi, dsttype); 456 } 457 } 458 } 459 } 460 phi_map->merge_replaced_nodes_with(ex_map); 461 } 462 463 //--------------------------use_exception_state-------------------------------- 464 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 465 if (failing_internal()) { stop(); return top(); } 466 Node* region = phi_map->control(); 467 Node* hidden_merge_mark = root(); 468 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 469 Node* ex_oop = clear_saved_ex_oop(phi_map); 470 if (region->in(0) == hidden_merge_mark) { 471 // Special marking for internal ex-states. Process the phis now. 472 region->set_req(0, region); // now it's an ordinary region 473 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 474 // Note: Setting the jvms also sets the bci and sp. 475 set_control(_gvn.transform(region)); 476 uint tos = jvms()->stkoff() + sp(); 477 for (uint i = 1; i < tos; i++) { 478 Node* x = phi_map->in(i); 479 if (x->in(0) == region) { 480 assert(x->is_Phi(), "expected a special phi"); 481 phi_map->set_req(i, _gvn.transform(x)); 482 } 483 } 484 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 485 Node* x = mms.memory(); 486 if (x->in(0) == region) { 487 assert(x->is_Phi(), "nobody else uses a hidden region"); 488 mms.set_memory(_gvn.transform(x)); 489 } 490 } 491 if (ex_oop->in(0) == region) { 492 assert(ex_oop->is_Phi(), "expected a special phi"); 493 ex_oop = _gvn.transform(ex_oop); 494 } 495 } else { 496 set_jvms(phi_map->jvms()); 497 } 498 499 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 500 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 501 return ex_oop; 502 } 503 504 //---------------------------------java_bc------------------------------------- 505 Bytecodes::Code GraphKit::java_bc() const { 506 ciMethod* method = this->method(); 507 int bci = this->bci(); 508 if (method != nullptr && bci != InvocationEntryBci) 509 return method->java_code_at_bci(bci); 510 else 511 return Bytecodes::_illegal; 512 } 513 514 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 515 bool must_throw) { 516 // if the exception capability is set, then we will generate code 517 // to check the JavaThread.should_post_on_exceptions flag to see 518 // if we actually need to report exception events (for this 519 // thread). If we don't need to report exception events, we will 520 // take the normal fast path provided by add_exception_events. If 521 // exception event reporting is enabled for this thread, we will 522 // take the uncommon_trap in the BuildCutout below. 523 524 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 525 Node* jthread = _gvn.transform(new ThreadLocalNode()); 526 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 527 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 528 529 // Test the should_post_on_exceptions_flag vs. 0 530 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 531 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 532 533 // Branch to slow_path if should_post_on_exceptions_flag was true 534 { BuildCutout unless(this, tst, PROB_MAX); 535 // Do not try anything fancy if we're notifying the VM on every throw. 536 // Cf. case Bytecodes::_athrow in parse2.cpp. 537 uncommon_trap(reason, Deoptimization::Action_none, 538 (ciKlass*)nullptr, (char*)nullptr, must_throw); 539 } 540 541 } 542 543 //------------------------------builtin_throw---------------------------------- 544 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) { 545 builtin_throw(reason, builtin_throw_exception(reason), /*allow_too_many_traps*/ true); 546 } 547 548 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, 549 ciInstance* ex_obj, 550 bool allow_too_many_traps) { 551 // If this throw happens frequently, an uncommon trap might cause 552 // a performance pothole. If there is a local exception handler, 553 // and if this particular bytecode appears to be deoptimizing often, 554 // let us handle the throw inline, with a preconstructed instance. 555 // Note: If the deopt count has blown up, the uncommon trap 556 // runtime is going to flush this nmethod, not matter what. 557 if (is_builtin_throw_hot(reason)) { 558 if (method()->can_omit_stack_trace() && ex_obj != nullptr) { 559 // If the throw is local, we use a pre-existing instance and 560 // punt on the backtrace. This would lead to a missing backtrace 561 // (a repeat of 4292742) if the backtrace object is ever asked 562 // for its backtrace. 563 // Fixing this remaining case of 4292742 requires some flavor of 564 // escape analysis. Leave that for the future. 565 if (env()->jvmti_can_post_on_exceptions()) { 566 // check if we must post exception events, take uncommon trap if so 567 uncommon_trap_if_should_post_on_exceptions(reason, true /*must_throw*/); 568 // here if should_post_on_exceptions is false 569 // continue on with the normal codegen 570 } 571 572 // Cheat with a preallocated exception object. 573 if (C->log() != nullptr) 574 C->log()->elem("hot_throw preallocated='1' reason='%s'", 575 Deoptimization::trap_reason_name(reason)); 576 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 577 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 578 579 // Clear the detail message of the preallocated exception object. 580 // Weblogic sometimes mutates the detail message of exceptions 581 // using reflection. 582 int offset = java_lang_Throwable::get_detailMessage_offset(); 583 const TypePtr* adr_typ = ex_con->add_offset(offset); 584 585 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 586 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 587 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 588 589 if (!method()->has_exception_handlers()) { 590 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway. 591 // This prevents issues with exception handling and late inlining. 592 set_sp(0); 593 clean_stack(0); 594 } 595 596 add_exception_state(make_exception_state(ex_node)); 597 return; 598 } else if (builtin_throw_too_many_traps(reason, ex_obj)) { 599 // We cannot afford to take too many traps here. Suffer in the interpreter instead. 600 assert(allow_too_many_traps, "not allowed"); 601 if (C->log() != nullptr) { 602 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 603 Deoptimization::trap_reason_name(reason), 604 C->trap_count(reason)); 605 } 606 uncommon_trap(reason, Deoptimization::Action_none, 607 (ciKlass*) nullptr, (char*) nullptr, 608 true /*must_throw*/); 609 return; 610 } 611 } 612 613 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 614 // It won't be much cheaper than bailing to the interp., since we'll 615 // have to pass up all the debug-info, and the runtime will have to 616 // create the stack trace. 617 618 // Usual case: Bail to interpreter. 619 // Reserve the right to recompile if we haven't seen anything yet. 620 621 // "must_throw" prunes the JVM state to include only the stack, if there 622 // are no local exception handlers. This should cut down on register 623 // allocation time and code size, by drastically reducing the number 624 // of in-edges on the call to the uncommon trap. 625 uncommon_trap(reason, Deoptimization::Action_maybe_recompile, 626 (ciKlass*) nullptr, (char*) nullptr, 627 true /*must_throw*/); 628 } 629 630 bool GraphKit::is_builtin_throw_hot(Deoptimization::DeoptReason reason) { 631 // If this particular condition has not yet happened at this 632 // bytecode, then use the uncommon trap mechanism, and allow for 633 // a future recompilation if several traps occur here. 634 // If the throw is hot, try to use a more complicated inline mechanism 635 // which keeps execution inside the compiled code. 636 if (ProfileTraps) { 637 if (too_many_traps(reason)) { 638 return true; 639 } 640 // (If there is no MDO at all, assume it is early in 641 // execution, and that any deopts are part of the 642 // startup transient, and don't need to be remembered.) 643 644 // Also, if there is a local exception handler, treat all throws 645 // as hot if there has been at least one in this method. 646 if (C->trap_count(reason) != 0 && 647 method()->method_data()->trap_count(reason) != 0 && 648 has_exception_handler()) { 649 return true; 650 } 651 } 652 return false; 653 } 654 655 bool GraphKit::builtin_throw_too_many_traps(Deoptimization::DeoptReason reason, 656 ciInstance* ex_obj) { 657 if (is_builtin_throw_hot(reason)) { 658 if (method()->can_omit_stack_trace() && ex_obj != nullptr) { 659 return false; // no traps; throws preallocated exception instead 660 } 661 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr; 662 if (method()->method_data()->trap_recompiled_at(bci(), m) || 663 C->too_many_traps(reason)) { 664 return true; 665 } 666 } 667 return false; 668 } 669 670 ciInstance* GraphKit::builtin_throw_exception(Deoptimization::DeoptReason reason) const { 671 // Preallocated exception objects to use when we don't need the backtrace. 672 switch (reason) { 673 case Deoptimization::Reason_null_check: 674 return env()->NullPointerException_instance(); 675 case Deoptimization::Reason_div0_check: 676 return env()->ArithmeticException_instance(); 677 case Deoptimization::Reason_range_check: 678 return env()->ArrayIndexOutOfBoundsException_instance(); 679 case Deoptimization::Reason_class_check: 680 return env()->ClassCastException_instance(); 681 case Deoptimization::Reason_array_check: 682 return env()->ArrayStoreException_instance(); 683 default: 684 return nullptr; 685 } 686 } 687 688 //----------------------------PreserveJVMState--------------------------------- 689 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 690 DEBUG_ONLY(kit->verify_map()); 691 _kit = kit; 692 _map = kit->map(); // preserve the map 693 _sp = kit->sp(); 694 kit->set_map(clone_map ? kit->clone_map() : nullptr); 695 #ifdef ASSERT 696 _bci = kit->bci(); 697 Parse* parser = kit->is_Parse(); 698 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo(); 699 _block = block; 700 #endif 701 } 702 PreserveJVMState::~PreserveJVMState() { 703 GraphKit* kit = _kit; 704 #ifdef ASSERT 705 assert(kit->bci() == _bci, "bci must not shift"); 706 Parse* parser = kit->is_Parse(); 707 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo(); 708 assert(block == _block, "block must not shift"); 709 #endif 710 kit->set_map(_map); 711 kit->set_sp(_sp); 712 } 713 714 715 //-----------------------------BuildCutout------------------------------------- 716 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 717 : PreserveJVMState(kit) 718 { 719 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 720 SafePointNode* outer_map = _map; // preserved map is caller's 721 SafePointNode* inner_map = kit->map(); 722 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 723 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 724 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 725 } 726 BuildCutout::~BuildCutout() { 727 GraphKit* kit = _kit; 728 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 729 } 730 731 //---------------------------PreserveReexecuteState---------------------------- 732 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 733 assert(!kit->stopped(), "must call stopped() before"); 734 _kit = kit; 735 _sp = kit->sp(); 736 _reexecute = kit->jvms()->_reexecute; 737 } 738 PreserveReexecuteState::~PreserveReexecuteState() { 739 if (_kit->stopped()) return; 740 _kit->jvms()->_reexecute = _reexecute; 741 _kit->set_sp(_sp); 742 } 743 744 //------------------------------clone_map-------------------------------------- 745 // Implementation of PreserveJVMState 746 // 747 // Only clone_map(...) here. If this function is only used in the 748 // PreserveJVMState class we may want to get rid of this extra 749 // function eventually and do it all there. 750 751 SafePointNode* GraphKit::clone_map() { 752 if (map() == nullptr) return nullptr; 753 754 // Clone the memory edge first 755 Node* mem = MergeMemNode::make(map()->memory()); 756 gvn().set_type_bottom(mem); 757 758 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 759 JVMState* jvms = this->jvms(); 760 JVMState* clonejvms = jvms->clone_shallow(C); 761 clonemap->set_memory(mem); 762 clonemap->set_jvms(clonejvms); 763 clonejvms->set_map(clonemap); 764 record_for_igvn(clonemap); 765 gvn().set_type_bottom(clonemap); 766 return clonemap; 767 } 768 769 //-----------------------------destruct_map_clone------------------------------ 770 // 771 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need 772 // to destruct/free/delete in the exact opposite order as clone_map(). 773 void GraphKit::destruct_map_clone(SafePointNode* sfp) { 774 if (sfp == nullptr) return; 775 776 Node* mem = sfp->memory(); 777 JVMState* jvms = sfp->jvms(); 778 779 if (jvms != nullptr) { 780 delete jvms; 781 } 782 783 remove_for_igvn(sfp); 784 gvn().clear_type(sfp); 785 sfp->destruct(&_gvn); 786 787 if (mem != nullptr) { 788 gvn().clear_type(mem); 789 mem->destruct(&_gvn); 790 } 791 } 792 793 //-----------------------------set_map_clone----------------------------------- 794 void GraphKit::set_map_clone(SafePointNode* m) { 795 _map = m; 796 _map = clone_map(); 797 _map->set_next_exception(nullptr); 798 DEBUG_ONLY(verify_map()); 799 } 800 801 802 //----------------------------kill_dead_locals--------------------------------- 803 // Detect any locals which are known to be dead, and force them to top. 804 void GraphKit::kill_dead_locals() { 805 // Consult the liveness information for the locals. If any 806 // of them are unused, then they can be replaced by top(). This 807 // should help register allocation time and cut down on the size 808 // of the deoptimization information. 809 810 // This call is made from many of the bytecode handling 811 // subroutines called from the Big Switch in do_one_bytecode. 812 // Every bytecode which might include a slow path is responsible 813 // for killing its dead locals. The more consistent we 814 // are about killing deads, the fewer useless phis will be 815 // constructed for them at various merge points. 816 817 // bci can be -1 (InvocationEntryBci). We return the entry 818 // liveness for the method. 819 820 if (method() == nullptr || method()->code_size() == 0) { 821 // We are building a graph for a call to a native method. 822 // All locals are live. 823 return; 824 } 825 826 ResourceMark rm; 827 828 // Consult the liveness information for the locals. If any 829 // of them are unused, then they can be replaced by top(). This 830 // should help register allocation time and cut down on the size 831 // of the deoptimization information. 832 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 833 834 int len = (int)live_locals.size(); 835 assert(len <= jvms()->loc_size(), "too many live locals"); 836 for (int local = 0; local < len; local++) { 837 if (!live_locals.at(local)) { 838 set_local(local, top()); 839 } 840 } 841 } 842 843 #ifdef ASSERT 844 //-------------------------dead_locals_are_killed------------------------------ 845 // Return true if all dead locals are set to top in the map. 846 // Used to assert "clean" debug info at various points. 847 bool GraphKit::dead_locals_are_killed() { 848 if (method() == nullptr || method()->code_size() == 0) { 849 // No locals need to be dead, so all is as it should be. 850 return true; 851 } 852 853 // Make sure somebody called kill_dead_locals upstream. 854 ResourceMark rm; 855 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) { 856 if (jvms->loc_size() == 0) continue; // no locals to consult 857 SafePointNode* map = jvms->map(); 858 ciMethod* method = jvms->method(); 859 int bci = jvms->bci(); 860 if (jvms == this->jvms()) { 861 bci = this->bci(); // it might not yet be synched 862 } 863 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 864 int len = (int)live_locals.size(); 865 if (!live_locals.is_valid() || len == 0) 866 // This method is trivial, or is poisoned by a breakpoint. 867 return true; 868 assert(len == jvms->loc_size(), "live map consistent with locals map"); 869 for (int local = 0; local < len; local++) { 870 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 871 if (PrintMiscellaneous && (Verbose || WizardMode)) { 872 tty->print_cr("Zombie local %d: ", local); 873 jvms->dump(); 874 } 875 return false; 876 } 877 } 878 } 879 return true; 880 } 881 882 #endif //ASSERT 883 884 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens. 885 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 886 ciMethod* cur_method = jvms->method(); 887 int cur_bci = jvms->bci(); 888 if (cur_method != nullptr && cur_bci != InvocationEntryBci) { 889 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 890 return Interpreter::bytecode_should_reexecute(code) || 891 (is_anewarray && (code == Bytecodes::_multianewarray)); 892 // Reexecute _multianewarray bytecode which was replaced with 893 // sequence of [a]newarray. See Parse::do_multianewarray(). 894 // 895 // Note: interpreter should not have it set since this optimization 896 // is limited by dimensions and guarded by flag so in some cases 897 // multianewarray() runtime calls will be generated and 898 // the bytecode should not be reexecutes (stack will not be reset). 899 } else { 900 return false; 901 } 902 } 903 904 // Helper function for adding JVMState and debug information to node 905 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 906 // Add the safepoint edges to the call (or other safepoint). 907 908 // Make sure dead locals are set to top. This 909 // should help register allocation time and cut down on the size 910 // of the deoptimization information. 911 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 912 913 // Walk the inline list to fill in the correct set of JVMState's 914 // Also fill in the associated edges for each JVMState. 915 916 // If the bytecode needs to be reexecuted we need to put 917 // the arguments back on the stack. 918 const bool should_reexecute = jvms()->should_reexecute(); 919 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 920 921 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 922 // undefined if the bci is different. This is normal for Parse but it 923 // should not happen for LibraryCallKit because only one bci is processed. 924 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 925 "in LibraryCallKit the reexecute bit should not change"); 926 927 // If we are guaranteed to throw, we can prune everything but the 928 // input to the current bytecode. 929 bool can_prune_locals = false; 930 uint stack_slots_not_pruned = 0; 931 int inputs = 0, depth = 0; 932 if (must_throw) { 933 assert(method() == youngest_jvms->method(), "sanity"); 934 if (compute_stack_effects(inputs, depth)) { 935 can_prune_locals = true; 936 stack_slots_not_pruned = inputs; 937 } 938 } 939 940 if (env()->should_retain_local_variables()) { 941 // At any safepoint, this method can get breakpointed, which would 942 // then require an immediate deoptimization. 943 can_prune_locals = false; // do not prune locals 944 stack_slots_not_pruned = 0; 945 } 946 947 // do not scribble on the input jvms 948 JVMState* out_jvms = youngest_jvms->clone_deep(C); 949 call->set_jvms(out_jvms); // Start jvms list for call node 950 951 // For a known set of bytecodes, the interpreter should reexecute them if 952 // deoptimization happens. We set the reexecute state for them here 953 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 954 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 955 #ifdef ASSERT 956 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects() 957 assert(method() == youngest_jvms->method(), "sanity"); 958 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc())); 959 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution"); 960 #endif // ASSERT 961 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 962 } 963 964 // Presize the call: 965 DEBUG_ONLY(uint non_debug_edges = call->req()); 966 call->add_req_batch(top(), youngest_jvms->debug_depth()); 967 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 968 969 // Set up edges so that the call looks like this: 970 // Call [state:] ctl io mem fptr retadr 971 // [parms:] parm0 ... parmN 972 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 973 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 974 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 975 // Note that caller debug info precedes callee debug info. 976 977 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 978 uint debug_ptr = call->req(); 979 980 // Loop over the map input edges associated with jvms, add them 981 // to the call node, & reset all offsets to match call node array. 982 983 JVMState* callee_jvms = nullptr; 984 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) { 985 uint debug_end = debug_ptr; 986 uint debug_start = debug_ptr - in_jvms->debug_size(); 987 debug_ptr = debug_start; // back up the ptr 988 989 uint p = debug_start; // walks forward in [debug_start, debug_end) 990 uint j, k, l; 991 SafePointNode* in_map = in_jvms->map(); 992 out_jvms->set_map(call); 993 994 if (can_prune_locals) { 995 assert(in_jvms->method() == out_jvms->method(), "sanity"); 996 // If the current throw can reach an exception handler in this JVMS, 997 // then we must keep everything live that can reach that handler. 998 // As a quick and dirty approximation, we look for any handlers at all. 999 if (in_jvms->method()->has_exception_handlers()) { 1000 can_prune_locals = false; 1001 } 1002 } 1003 1004 // Add the Locals 1005 k = in_jvms->locoff(); 1006 l = in_jvms->loc_size(); 1007 out_jvms->set_locoff(p); 1008 if (!can_prune_locals) { 1009 for (j = 0; j < l; j++) { 1010 call->set_req(p++, in_map->in(k + j)); 1011 } 1012 } else { 1013 p += l; // already set to top above by add_req_batch 1014 } 1015 1016 // Add the Expression Stack 1017 k = in_jvms->stkoff(); 1018 l = in_jvms->sp(); 1019 out_jvms->set_stkoff(p); 1020 if (!can_prune_locals) { 1021 for (j = 0; j < l; j++) { 1022 call->set_req(p++, in_map->in(k + j)); 1023 } 1024 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 1025 // Divide stack into {S0,...,S1}, where S0 is set to top. 1026 uint s1 = stack_slots_not_pruned; 1027 stack_slots_not_pruned = 0; // for next iteration 1028 if (s1 > l) s1 = l; 1029 uint s0 = l - s1; 1030 p += s0; // skip the tops preinstalled by add_req_batch 1031 for (j = s0; j < l; j++) 1032 call->set_req(p++, in_map->in(k+j)); 1033 } else { 1034 p += l; // already set to top above by add_req_batch 1035 } 1036 1037 // Add the Monitors 1038 k = in_jvms->monoff(); 1039 l = in_jvms->mon_size(); 1040 out_jvms->set_monoff(p); 1041 for (j = 0; j < l; j++) 1042 call->set_req(p++, in_map->in(k+j)); 1043 1044 // Copy any scalar object fields. 1045 k = in_jvms->scloff(); 1046 l = in_jvms->scl_size(); 1047 out_jvms->set_scloff(p); 1048 for (j = 0; j < l; j++) 1049 call->set_req(p++, in_map->in(k+j)); 1050 1051 // Finish the new jvms. 1052 out_jvms->set_endoff(p); 1053 1054 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 1055 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 1056 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 1057 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 1058 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 1059 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 1060 1061 // Update the two tail pointers in parallel. 1062 callee_jvms = out_jvms; 1063 out_jvms = out_jvms->caller(); 1064 in_jvms = in_jvms->caller(); 1065 } 1066 1067 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1068 1069 // Test the correctness of JVMState::debug_xxx accessors: 1070 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1071 assert(call->jvms()->debug_end() == call->req(), ""); 1072 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1073 } 1074 1075 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1076 Bytecodes::Code code = java_bc(); 1077 if (code == Bytecodes::_wide) { 1078 code = method()->java_code_at_bci(bci() + 1); 1079 } 1080 1081 if (code != Bytecodes::_illegal) { 1082 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1083 } 1084 1085 auto rsize = [&]() { 1086 assert(code != Bytecodes::_illegal, "code is illegal!"); 1087 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1088 return (rtype < T_CONFLICT) ? type2size[rtype] : 0; 1089 }; 1090 1091 switch (code) { 1092 case Bytecodes::_illegal: 1093 return false; 1094 1095 case Bytecodes::_ldc: 1096 case Bytecodes::_ldc_w: 1097 case Bytecodes::_ldc2_w: 1098 inputs = 0; 1099 break; 1100 1101 case Bytecodes::_dup: inputs = 1; break; 1102 case Bytecodes::_dup_x1: inputs = 2; break; 1103 case Bytecodes::_dup_x2: inputs = 3; break; 1104 case Bytecodes::_dup2: inputs = 2; break; 1105 case Bytecodes::_dup2_x1: inputs = 3; break; 1106 case Bytecodes::_dup2_x2: inputs = 4; break; 1107 case Bytecodes::_swap: inputs = 2; break; 1108 case Bytecodes::_arraylength: inputs = 1; break; 1109 1110 case Bytecodes::_getstatic: 1111 case Bytecodes::_putstatic: 1112 case Bytecodes::_getfield: 1113 case Bytecodes::_putfield: 1114 { 1115 bool ignored_will_link; 1116 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1117 int size = field->type()->size(); 1118 bool is_get = (depth >= 0), is_static = (depth & 1); 1119 inputs = (is_static ? 0 : 1); 1120 if (is_get) { 1121 depth = size - inputs; 1122 } else { 1123 inputs += size; // putxxx pops the value from the stack 1124 depth = - inputs; 1125 } 1126 } 1127 break; 1128 1129 case Bytecodes::_invokevirtual: 1130 case Bytecodes::_invokespecial: 1131 case Bytecodes::_invokestatic: 1132 case Bytecodes::_invokedynamic: 1133 case Bytecodes::_invokeinterface: 1134 { 1135 bool ignored_will_link; 1136 ciSignature* declared_signature = nullptr; 1137 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1138 assert(declared_signature != nullptr, "cannot be null"); 1139 inputs = declared_signature->arg_size_for_bc(code); 1140 int size = declared_signature->return_type()->size(); 1141 depth = size - inputs; 1142 } 1143 break; 1144 1145 case Bytecodes::_multianewarray: 1146 { 1147 ciBytecodeStream iter(method()); 1148 iter.reset_to_bci(bci()); 1149 iter.next(); 1150 inputs = iter.get_dimensions(); 1151 assert(rsize() == 1, ""); 1152 depth = 1 - inputs; 1153 } 1154 break; 1155 1156 case Bytecodes::_ireturn: 1157 case Bytecodes::_lreturn: 1158 case Bytecodes::_freturn: 1159 case Bytecodes::_dreturn: 1160 case Bytecodes::_areturn: 1161 assert(rsize() == -depth, ""); 1162 inputs = -depth; 1163 break; 1164 1165 case Bytecodes::_jsr: 1166 case Bytecodes::_jsr_w: 1167 inputs = 0; 1168 depth = 1; // S.B. depth=1, not zero 1169 break; 1170 1171 default: 1172 // bytecode produces a typed result 1173 inputs = rsize() - depth; 1174 assert(inputs >= 0, ""); 1175 break; 1176 } 1177 1178 #ifdef ASSERT 1179 // spot check 1180 int outputs = depth + inputs; 1181 assert(outputs >= 0, "sanity"); 1182 switch (code) { 1183 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1184 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1185 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1186 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1187 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1188 default: break; 1189 } 1190 #endif //ASSERT 1191 1192 return true; 1193 } 1194 1195 1196 1197 //------------------------------basic_plus_adr--------------------------------- 1198 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1199 // short-circuit a common case 1200 if (offset == intcon(0)) return ptr; 1201 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1202 } 1203 1204 Node* GraphKit::ConvI2L(Node* offset) { 1205 // short-circuit a common case 1206 jint offset_con = find_int_con(offset, Type::OffsetBot); 1207 if (offset_con != Type::OffsetBot) { 1208 return longcon((jlong) offset_con); 1209 } 1210 return _gvn.transform( new ConvI2LNode(offset)); 1211 } 1212 1213 Node* GraphKit::ConvI2UL(Node* offset) { 1214 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1215 if (offset_con != (juint) Type::OffsetBot) { 1216 return longcon((julong) offset_con); 1217 } 1218 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1219 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1220 return _gvn.transform( new AndLNode(conv, mask) ); 1221 } 1222 1223 Node* GraphKit::ConvL2I(Node* offset) { 1224 // short-circuit a common case 1225 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1226 if (offset_con != (jlong)Type::OffsetBot) { 1227 return intcon((int) offset_con); 1228 } 1229 return _gvn.transform( new ConvL2INode(offset)); 1230 } 1231 1232 //-------------------------load_object_klass----------------------------------- 1233 Node* GraphKit::load_object_klass(Node* obj) { 1234 // Special-case a fresh allocation to avoid building nodes: 1235 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1236 if (akls != nullptr) return akls; 1237 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1238 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 1239 } 1240 1241 //-------------------------load_array_length----------------------------------- 1242 Node* GraphKit::load_array_length(Node* array) { 1243 // Special-case a fresh allocation to avoid building nodes: 1244 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array); 1245 Node *alen; 1246 if (alloc == nullptr) { 1247 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1248 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS)); 1249 } else { 1250 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false); 1251 } 1252 return alen; 1253 } 1254 1255 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc, 1256 const TypeOopPtr* oop_type, 1257 bool replace_length_in_map) { 1258 Node* length = alloc->Ideal_length(); 1259 if (replace_length_in_map == false || map()->find_edge(length) >= 0) { 1260 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn); 1261 if (ccast != length) { 1262 // do not transform ccast here, it might convert to top node for 1263 // negative array length and break assumptions in parsing stage. 1264 _gvn.set_type_bottom(ccast); 1265 record_for_igvn(ccast); 1266 if (replace_length_in_map) { 1267 replace_in_map(length, ccast); 1268 } 1269 return ccast; 1270 } 1271 } 1272 return length; 1273 } 1274 1275 //------------------------------do_null_check---------------------------------- 1276 // Helper function to do a null pointer check. Returned value is 1277 // the incoming address with null casted away. You are allowed to use the 1278 // not-null value only if you are control dependent on the test. 1279 #ifndef PRODUCT 1280 extern uint explicit_null_checks_inserted, 1281 explicit_null_checks_elided; 1282 #endif 1283 Node* GraphKit::null_check_common(Node* value, BasicType type, 1284 // optional arguments for variations: 1285 bool assert_null, 1286 Node* *null_control, 1287 bool speculative, 1288 bool null_marker_check) { 1289 assert(!assert_null || null_control == nullptr, "not both at once"); 1290 if (stopped()) return top(); 1291 NOT_PRODUCT(explicit_null_checks_inserted++); 1292 1293 if (value->is_InlineType()) { 1294 // Null checking a scalarized but nullable inline type. Check the null marker 1295 // input instead of the oop input to avoid keeping buffer allocations alive. 1296 InlineTypeNode* vtptr = value->as_InlineType(); 1297 while (vtptr->get_oop()->is_InlineType()) { 1298 vtptr = vtptr->get_oop()->as_InlineType(); 1299 } 1300 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true); 1301 if (stopped()) { 1302 return top(); 1303 } 1304 if (assert_null) { 1305 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts) 1306 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass()); 1307 // replace_in_map(value, vtptr); 1308 // return vtptr; 1309 replace_in_map(value, null()); 1310 return null(); 1311 } 1312 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top()); 1313 return cast_not_null(value, do_replace_in_map); 1314 } 1315 1316 // Construct null check 1317 Node *chk = nullptr; 1318 switch(type) { 1319 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1320 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1321 case T_ARRAY : // fall through 1322 type = T_OBJECT; // simplify further tests 1323 case T_OBJECT : { 1324 const Type *t = _gvn.type( value ); 1325 1326 const TypeOopPtr* tp = t->isa_oopptr(); 1327 if (tp != nullptr && !tp->is_loaded() 1328 // Only for do_null_check, not any of its siblings: 1329 && !assert_null && null_control == nullptr) { 1330 // Usually, any field access or invocation on an unloaded oop type 1331 // will simply fail to link, since the statically linked class is 1332 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1333 // the static class is loaded but the sharper oop type is not. 1334 // Rather than checking for this obscure case in lots of places, 1335 // we simply observe that a null check on an unloaded class 1336 // will always be followed by a nonsense operation, so we 1337 // can just issue the uncommon trap here. 1338 // Our access to the unloaded class will only be correct 1339 // after it has been loaded and initialized, which requires 1340 // a trip through the interpreter. 1341 ciKlass* klass = tp->unloaded_klass(); 1342 #ifndef PRODUCT 1343 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); } 1344 #endif 1345 uncommon_trap(Deoptimization::Reason_unloaded, 1346 Deoptimization::Action_reinterpret, 1347 klass, "!loaded"); 1348 return top(); 1349 } 1350 1351 if (assert_null) { 1352 // See if the type is contained in NULL_PTR. 1353 // If so, then the value is already null. 1354 if (t->higher_equal(TypePtr::NULL_PTR)) { 1355 NOT_PRODUCT(explicit_null_checks_elided++); 1356 return value; // Elided null assert quickly! 1357 } 1358 } else { 1359 // See if mixing in the null pointer changes type. 1360 // If so, then the null pointer was not allowed in the original 1361 // type. In other words, "value" was not-null. 1362 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1363 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1364 NOT_PRODUCT(explicit_null_checks_elided++); 1365 return value; // Elided null check quickly! 1366 } 1367 } 1368 chk = new CmpPNode( value, null() ); 1369 break; 1370 } 1371 1372 default: 1373 fatal("unexpected type: %s", type2name(type)); 1374 } 1375 assert(chk != nullptr, "sanity check"); 1376 chk = _gvn.transform(chk); 1377 1378 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1379 BoolNode *btst = new BoolNode( chk, btest); 1380 Node *tst = _gvn.transform( btst ); 1381 1382 //----------- 1383 // if peephole optimizations occurred, a prior test existed. 1384 // If a prior test existed, maybe it dominates as we can avoid this test. 1385 if (tst != btst && type == T_OBJECT) { 1386 // At this point we want to scan up the CFG to see if we can 1387 // find an identical test (and so avoid this test altogether). 1388 Node *cfg = control(); 1389 int depth = 0; 1390 while( depth < 16 ) { // Limit search depth for speed 1391 if( cfg->Opcode() == Op_IfTrue && 1392 cfg->in(0)->in(1) == tst ) { 1393 // Found prior test. Use "cast_not_null" to construct an identical 1394 // CastPP (and hence hash to) as already exists for the prior test. 1395 // Return that casted value. 1396 if (assert_null) { 1397 replace_in_map(value, null()); 1398 return null(); // do not issue the redundant test 1399 } 1400 Node *oldcontrol = control(); 1401 set_control(cfg); 1402 Node *res = cast_not_null(value); 1403 set_control(oldcontrol); 1404 NOT_PRODUCT(explicit_null_checks_elided++); 1405 return res; 1406 } 1407 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1408 if (cfg == nullptr) break; // Quit at region nodes 1409 depth++; 1410 } 1411 } 1412 1413 //----------- 1414 // Branch to failure if null 1415 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1416 Deoptimization::DeoptReason reason; 1417 if (assert_null) { 1418 reason = Deoptimization::reason_null_assert(speculative); 1419 } else if (type == T_OBJECT || null_marker_check) { 1420 reason = Deoptimization::reason_null_check(speculative); 1421 } else { 1422 reason = Deoptimization::Reason_div0_check; 1423 } 1424 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1425 // ciMethodData::has_trap_at will return a conservative -1 if any 1426 // must-be-null assertion has failed. This could cause performance 1427 // problems for a method after its first do_null_assert failure. 1428 // Consider using 'Reason_class_check' instead? 1429 1430 // To cause an implicit null check, we set the not-null probability 1431 // to the maximum (PROB_MAX). For an explicit check the probability 1432 // is set to a smaller value. 1433 if (null_control != nullptr || too_many_traps(reason)) { 1434 // probability is less likely 1435 ok_prob = PROB_LIKELY_MAG(3); 1436 } else if (!assert_null && 1437 (ImplicitNullCheckThreshold > 0) && 1438 method() != nullptr && 1439 (method()->method_data()->trap_count(reason) 1440 >= (uint)ImplicitNullCheckThreshold)) { 1441 ok_prob = PROB_LIKELY_MAG(3); 1442 } 1443 1444 if (null_control != nullptr) { 1445 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1446 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1447 set_control( _gvn.transform( new IfTrueNode(iff))); 1448 #ifndef PRODUCT 1449 if (null_true == top()) { 1450 explicit_null_checks_elided++; 1451 } 1452 #endif 1453 (*null_control) = null_true; 1454 } else { 1455 BuildCutout unless(this, tst, ok_prob); 1456 // Check for optimizer eliding test at parse time 1457 if (stopped()) { 1458 // Failure not possible; do not bother making uncommon trap. 1459 NOT_PRODUCT(explicit_null_checks_elided++); 1460 } else if (assert_null) { 1461 uncommon_trap(reason, 1462 Deoptimization::Action_make_not_entrant, 1463 nullptr, "assert_null"); 1464 } else { 1465 replace_in_map(value, zerocon(type)); 1466 builtin_throw(reason); 1467 } 1468 } 1469 1470 // Must throw exception, fall-thru not possible? 1471 if (stopped()) { 1472 return top(); // No result 1473 } 1474 1475 if (assert_null) { 1476 // Cast obj to null on this path. 1477 replace_in_map(value, zerocon(type)); 1478 return zerocon(type); 1479 } 1480 1481 // Cast obj to not-null on this path, if there is no null_control. 1482 // (If there is a null_control, a non-null value may come back to haunt us.) 1483 if (type == T_OBJECT) { 1484 Node* cast = cast_not_null(value, false); 1485 if (null_control == nullptr || (*null_control) == top()) 1486 replace_in_map(value, cast); 1487 value = cast; 1488 } 1489 1490 return value; 1491 } 1492 1493 //------------------------------cast_not_null---------------------------------- 1494 // Cast obj to not-null on this path 1495 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1496 if (obj->is_InlineType()) { 1497 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map); 1498 vt->as_InlineType()->set_null_marker(_gvn); 1499 vt = _gvn.transform(vt); 1500 if (do_replace_in_map) { 1501 replace_in_map(obj, vt); 1502 } 1503 return vt; 1504 } 1505 const Type *t = _gvn.type(obj); 1506 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1507 // Object is already not-null? 1508 if( t == t_not_null ) return obj; 1509 1510 Node* cast = new CastPPNode(control(), obj,t_not_null); 1511 cast = _gvn.transform( cast ); 1512 1513 // Scan for instances of 'obj' in the current JVM mapping. 1514 // These instances are known to be not-null after the test. 1515 if (do_replace_in_map) 1516 replace_in_map(obj, cast); 1517 1518 return cast; // Return casted value 1519 } 1520 1521 Node* GraphKit::cast_to_non_larval(Node* obj) { 1522 const Type* obj_type = gvn().type(obj); 1523 if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) { 1524 return obj; 1525 } 1526 1527 Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass()); 1528 replace_in_map(obj, new_obj); 1529 return new_obj; 1530 } 1531 1532 // Sometimes in intrinsics, we implicitly know an object is not null 1533 // (there's no actual null check) so we can cast it to not null. In 1534 // the course of optimizations, the input to the cast can become null. 1535 // In that case that data path will die and we need the control path 1536 // to become dead as well to keep the graph consistent. So we have to 1537 // add a check for null for which one branch can't be taken. It uses 1538 // an OpaqueNotNull node that will cause the check to be removed after loop 1539 // opts so the test goes away and the compiled code doesn't execute a 1540 // useless check. 1541 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1542 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1543 return value; 1544 } 1545 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1546 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1547 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst)); 1548 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1549 _gvn.set_type(iff, iff->Value(&_gvn)); 1550 if (!tst->is_Con()) { 1551 record_for_igvn(iff); 1552 } 1553 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1554 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1555 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1556 C->root()->add_req(halt); 1557 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1558 set_control(if_t); 1559 return cast_not_null(value, do_replace_in_map); 1560 } 1561 1562 1563 //--------------------------replace_in_map------------------------------------- 1564 void GraphKit::replace_in_map(Node* old, Node* neww) { 1565 if (old == neww) { 1566 return; 1567 } 1568 1569 map()->replace_edge(old, neww); 1570 1571 // Note: This operation potentially replaces any edge 1572 // on the map. This includes locals, stack, and monitors 1573 // of the current (innermost) JVM state. 1574 1575 // don't let inconsistent types from profiling escape this 1576 // method 1577 1578 const Type* told = _gvn.type(old); 1579 const Type* tnew = _gvn.type(neww); 1580 1581 if (!tnew->higher_equal(told)) { 1582 return; 1583 } 1584 1585 map()->record_replaced_node(old, neww); 1586 } 1587 1588 1589 //============================================================================= 1590 //--------------------------------memory--------------------------------------- 1591 Node* GraphKit::memory(uint alias_idx) { 1592 MergeMemNode* mem = merged_memory(); 1593 Node* p = mem->memory_at(alias_idx); 1594 assert(p != mem->empty_memory(), "empty"); 1595 _gvn.set_type(p, Type::MEMORY); // must be mapped 1596 return p; 1597 } 1598 1599 //-----------------------------reset_memory------------------------------------ 1600 Node* GraphKit::reset_memory() { 1601 Node* mem = map()->memory(); 1602 // do not use this node for any more parsing! 1603 DEBUG_ONLY( map()->set_memory((Node*)nullptr) ); 1604 return _gvn.transform( mem ); 1605 } 1606 1607 //------------------------------set_all_memory--------------------------------- 1608 void GraphKit::set_all_memory(Node* newmem) { 1609 Node* mergemem = MergeMemNode::make(newmem); 1610 gvn().set_type_bottom(mergemem); 1611 map()->set_memory(mergemem); 1612 } 1613 1614 //------------------------------set_all_memory_call---------------------------- 1615 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1616 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1617 set_all_memory(newmem); 1618 } 1619 1620 //============================================================================= 1621 // 1622 // parser factory methods for MemNodes 1623 // 1624 // These are layered on top of the factory methods in LoadNode and StoreNode, 1625 // and integrate with the parser's memory state and _gvn engine. 1626 // 1627 1628 // factory methods in "int adr_idx" 1629 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1630 MemNode::MemOrd mo, 1631 LoadNode::ControlDependency control_dependency, 1632 bool require_atomic_access, 1633 bool unaligned, 1634 bool mismatched, 1635 bool unsafe, 1636 uint8_t barrier_data) { 1637 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr()); 1638 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1639 const TypePtr* adr_type = nullptr; // debug-mode-only argument 1640 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx)); 1641 Node* mem = memory(adr_idx); 1642 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data); 1643 ld = _gvn.transform(ld); 1644 1645 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1646 // Improve graph before escape analysis and boxing elimination. 1647 record_for_igvn(ld); 1648 if (ld->is_DecodeN()) { 1649 // Also record the actual load (LoadN) in case ld is DecodeN. In some 1650 // rare corner cases, ld->in(1) can be something other than LoadN (e.g., 1651 // a Phi). Recording such cases is still perfectly sound, but may be 1652 // unnecessary and result in some minor IGVN overhead. 1653 record_for_igvn(ld->in(1)); 1654 } 1655 } 1656 return ld; 1657 } 1658 1659 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1660 MemNode::MemOrd mo, 1661 bool require_atomic_access, 1662 bool unaligned, 1663 bool mismatched, 1664 bool unsafe, 1665 int barrier_data) { 1666 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr()); 1667 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1668 const TypePtr* adr_type = nullptr; 1669 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx)); 1670 Node *mem = memory(adr_idx); 1671 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access); 1672 if (unaligned) { 1673 st->as_Store()->set_unaligned_access(); 1674 } 1675 if (mismatched) { 1676 st->as_Store()->set_mismatched_access(); 1677 } 1678 if (unsafe) { 1679 st->as_Store()->set_unsafe_access(); 1680 } 1681 st->as_Store()->set_barrier_data(barrier_data); 1682 st = _gvn.transform(st); 1683 set_memory(st, adr_idx); 1684 // Back-to-back stores can only remove intermediate store with DU info 1685 // so push on worklist for optimizer. 1686 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1687 record_for_igvn(st); 1688 1689 return st; 1690 } 1691 1692 Node* GraphKit::access_store_at(Node* obj, 1693 Node* adr, 1694 const TypePtr* adr_type, 1695 Node* val, 1696 const Type* val_type, 1697 BasicType bt, 1698 DecoratorSet decorators, 1699 bool safe_for_replace, 1700 const InlineTypeNode* vt) { 1701 // Transformation of a value which could be null pointer (CastPP #null) 1702 // could be delayed during Parse (for example, in adjust_map_after_if()). 1703 // Execute transformation here to avoid barrier generation in such case. 1704 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1705 val = _gvn.makecon(TypePtr::NULL_PTR); 1706 } 1707 1708 if (stopped()) { 1709 return top(); // Dead path ? 1710 } 1711 1712 assert(val != nullptr, "not dead path"); 1713 if (val->is_InlineType()) { 1714 // Store to non-flat field. Buffer the inline type and make sure 1715 // the store is re-executed if the allocation triggers deoptimization. 1716 PreserveReexecuteState preexecs(this); 1717 jvms()->set_should_reexecute(true); 1718 val = val->as_InlineType()->buffer(this, safe_for_replace); 1719 } 1720 1721 C2AccessValuePtr addr(adr, adr_type); 1722 C2AccessValue value(val, val_type); 1723 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt); 1724 if (access.is_raw()) { 1725 return _barrier_set->BarrierSetC2::store_at(access, value); 1726 } else { 1727 return _barrier_set->store_at(access, value); 1728 } 1729 } 1730 1731 Node* GraphKit::access_load_at(Node* obj, // containing obj 1732 Node* adr, // actual address to store val at 1733 const TypePtr* adr_type, 1734 const Type* val_type, 1735 BasicType bt, 1736 DecoratorSet decorators, 1737 Node* ctl) { 1738 if (stopped()) { 1739 return top(); // Dead path ? 1740 } 1741 1742 C2AccessValuePtr addr(adr, adr_type); 1743 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl); 1744 if (access.is_raw()) { 1745 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1746 } else { 1747 return _barrier_set->load_at(access, val_type); 1748 } 1749 } 1750 1751 Node* GraphKit::access_load(Node* adr, // actual address to load val at 1752 const Type* val_type, 1753 BasicType bt, 1754 DecoratorSet decorators) { 1755 if (stopped()) { 1756 return top(); // Dead path ? 1757 } 1758 1759 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr()); 1760 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr); 1761 if (access.is_raw()) { 1762 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1763 } else { 1764 return _barrier_set->load_at(access, val_type); 1765 } 1766 } 1767 1768 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1769 Node* adr, 1770 const TypePtr* adr_type, 1771 int alias_idx, 1772 Node* expected_val, 1773 Node* new_val, 1774 const Type* value_type, 1775 BasicType bt, 1776 DecoratorSet decorators) { 1777 C2AccessValuePtr addr(adr, adr_type); 1778 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1779 bt, obj, addr, alias_idx); 1780 if (access.is_raw()) { 1781 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1782 } else { 1783 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1784 } 1785 } 1786 1787 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1788 Node* adr, 1789 const TypePtr* adr_type, 1790 int alias_idx, 1791 Node* expected_val, 1792 Node* new_val, 1793 const Type* value_type, 1794 BasicType bt, 1795 DecoratorSet decorators) { 1796 C2AccessValuePtr addr(adr, adr_type); 1797 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1798 bt, obj, addr, alias_idx); 1799 if (access.is_raw()) { 1800 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1801 } else { 1802 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1803 } 1804 } 1805 1806 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1807 Node* adr, 1808 const TypePtr* adr_type, 1809 int alias_idx, 1810 Node* new_val, 1811 const Type* value_type, 1812 BasicType bt, 1813 DecoratorSet decorators) { 1814 C2AccessValuePtr addr(adr, adr_type); 1815 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1816 bt, obj, addr, alias_idx); 1817 if (access.is_raw()) { 1818 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1819 } else { 1820 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1821 } 1822 } 1823 1824 Node* GraphKit::access_atomic_add_at(Node* obj, 1825 Node* adr, 1826 const TypePtr* adr_type, 1827 int alias_idx, 1828 Node* new_val, 1829 const Type* value_type, 1830 BasicType bt, 1831 DecoratorSet decorators) { 1832 C2AccessValuePtr addr(adr, adr_type); 1833 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1834 if (access.is_raw()) { 1835 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1836 } else { 1837 return _barrier_set->atomic_add_at(access, new_val, value_type); 1838 } 1839 } 1840 1841 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) { 1842 return _barrier_set->clone(this, src, dst, size, is_array); 1843 } 1844 1845 //-------------------------array_element_address------------------------- 1846 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1847 const TypeInt* sizetype, Node* ctrl) { 1848 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 1849 uint shift; 1850 uint header; 1851 if (arytype->is_flat() && arytype->klass_is_exact()) { 1852 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different 1853 // value classes at runtime with a potentially different layout. The caller needs to fall back to call 1854 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that 1855 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even 1856 // though we don't need the address node in this case and throw it away again. 1857 shift = arytype->flat_log_elem_size(); 1858 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT); 1859 } else { 1860 shift = exact_log2(type2aelembytes(elembt)); 1861 header = arrayOopDesc::base_offset_in_bytes(elembt); 1862 } 1863 1864 // short-circuit a common case (saves lots of confusing waste motion) 1865 jint idx_con = find_int_con(idx, -1); 1866 if (idx_con >= 0) { 1867 intptr_t offset = header + ((intptr_t)idx_con << shift); 1868 return basic_plus_adr(ary, offset); 1869 } 1870 1871 // must be correct type for alignment purposes 1872 Node* base = basic_plus_adr(ary, header); 1873 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1874 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1875 return basic_plus_adr(ary, base, scale); 1876 } 1877 1878 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* vk, bool is_null_free, bool is_not_null_free, bool is_atomic) { 1879 assert(vk->maybe_flat_in_array(), "element of type %s cannot be flat in array", vk->name()->as_utf8()); 1880 if (!vk->has_nullable_atomic_layout()) { 1881 // Element does not have a nullable flat layout, cannot be nullable 1882 is_null_free = true; 1883 } 1884 if (!vk->has_atomic_layout() && !vk->has_non_atomic_layout()) { 1885 // Element does not have a null-free flat layout, cannot be null-free 1886 is_not_null_free = true; 1887 } 1888 if (is_null_free) { 1889 // TODO 8350865 Impossible type 1890 is_not_null_free = false; 1891 } 1892 1893 bool is_exact = is_null_free || is_not_null_free; 1894 ciArrayKlass* array_klass = ciArrayKlass::make(vk, is_null_free, is_atomic, true); 1895 assert(array_klass->is_elem_null_free() == is_null_free, "inconsistency"); 1896 assert(array_klass->is_elem_atomic() == is_atomic, "inconsistency"); 1897 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr(); 1898 arytype = arytype->cast_to_exactness(is_exact); 1899 arytype = arytype->cast_to_not_null_free(is_not_null_free); 1900 assert(arytype->is_null_free() == is_null_free, "inconsistency"); 1901 assert(arytype->is_not_null_free() == is_not_null_free, "inconsistency"); 1902 assert(arytype->is_atomic() == is_atomic, "inconsistency"); 1903 return _gvn.transform(new CastPPNode(control(), array, arytype, ConstraintCastNode::StrongDependency)); 1904 } 1905 1906 //-------------------------load_array_element------------------------- 1907 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) { 1908 const Type* elemtype = arytype->elem(); 1909 BasicType elembt = elemtype->array_element_basic_type(); 1910 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1911 if (elembt == T_NARROWOOP) { 1912 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1913 } 1914 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt, 1915 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0)); 1916 return ld; 1917 } 1918 1919 //-------------------------set_arguments_for_java_call------------------------- 1920 // Arguments (pre-popped from the stack) are taken from the JVMS. 1921 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) { 1922 PreserveReexecuteState preexecs(this); 1923 if (EnableValhalla) { 1924 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization. 1925 // At this point, the call hasn't been executed yet, so we will only ever execute the call once. 1926 jvms()->set_should_reexecute(true); 1927 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc()); 1928 inc_sp(arg_size); 1929 } 1930 // Add the call arguments 1931 const TypeTuple* domain = call->tf()->domain_sig(); 1932 uint nargs = domain->cnt(); 1933 int arg_num = 0; 1934 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1935 Node* arg = argument(i-TypeFunc::Parms); 1936 const Type* t = domain->field_at(i); 1937 // TODO 8284443 A static call to a mismatched method should still be scalarized 1938 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) { 1939 // We don't pass inline type arguments by reference but instead pass each field of the inline type 1940 if (!arg->is_InlineType()) { 1941 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type"); 1942 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass()); 1943 } 1944 InlineTypeNode* vt = arg->as_InlineType(); 1945 vt->pass_fields(this, call, idx, true, !t->maybe_null()); 1946 // If an inline type argument is passed as fields, attach the Method* to the call site 1947 // to be able to access the extended signature later via attached_method_before_pc(). 1948 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1949 call->set_override_symbolic_info(true); 1950 // Register an evol dependency on the callee method to make sure that this method is deoptimized and 1951 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched. 1952 C->dependencies()->assert_evol_method(call->method()); 1953 arg_num++; 1954 continue; 1955 } else if (arg->is_InlineType()) { 1956 // Pass inline type argument via oop to callee 1957 arg = arg->as_InlineType()->buffer(this, true); 1958 } 1959 if (t != Type::HALF) { 1960 arg_num++; 1961 } 1962 call->init_req(idx++, arg); 1963 } 1964 } 1965 1966 //---------------------------set_edges_for_java_call--------------------------- 1967 // Connect a newly created call into the current JVMS. 1968 // A return value node (if any) is returned from set_edges_for_java_call. 1969 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1970 1971 // Add the predefined inputs: 1972 call->init_req( TypeFunc::Control, control() ); 1973 call->init_req( TypeFunc::I_O , i_o() ); 1974 call->init_req( TypeFunc::Memory , reset_memory() ); 1975 call->init_req( TypeFunc::FramePtr, frameptr() ); 1976 call->init_req( TypeFunc::ReturnAdr, top() ); 1977 1978 add_safepoint_edges(call, must_throw); 1979 1980 Node* xcall = _gvn.transform(call); 1981 1982 if (xcall == top()) { 1983 set_control(top()); 1984 return; 1985 } 1986 assert(xcall == call, "call identity is stable"); 1987 1988 // Re-use the current map to produce the result. 1989 1990 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1991 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1992 set_all_memory_call(xcall, separate_io_proj); 1993 1994 //return xcall; // no need, caller already has it 1995 } 1996 1997 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1998 if (stopped()) return top(); // maybe the call folded up? 1999 2000 // Note: Since any out-of-line call can produce an exception, 2001 // we always insert an I_O projection from the call into the result. 2002 2003 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 2004 2005 if (separate_io_proj) { 2006 // The caller requested separate projections be used by the fall 2007 // through and exceptional paths, so replace the projections for 2008 // the fall through path. 2009 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 2010 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 2011 } 2012 2013 // Capture the return value, if any. 2014 Node* ret; 2015 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) { 2016 ret = top(); 2017 } else if (call->tf()->returns_inline_type_as_fields()) { 2018 // Return of multiple values (inline type fields): we create a 2019 // InlineType node, each field is a projection from the call. 2020 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass(); 2021 uint base_input = TypeFunc::Parms; 2022 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false); 2023 } else { 2024 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 2025 ciType* t = call->method()->return_type(); 2026 if (!t->is_loaded() && InlineTypeReturnedAsFields) { 2027 // The return type is unloaded but the callee might later be C2 compiled and then return 2028 // in scalarized form when the return type is loaded. Handle this similar to what we do in 2029 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer. 2030 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop. 2031 IdealKit ideal(this); 2032 IdealVariable res(ideal); 2033 ideal.declarations_done(); 2034 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); { 2035 // Return value is null 2036 ideal.set(res, ret); 2037 } ideal.else_(); { 2038 // Return value is non-null 2039 sync_kit(ideal); 2040 2041 // Change return type of call to scalarized return 2042 const TypeFunc* tf = call->_tf; 2043 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc(); 2044 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2045 call->_tf = new_tf; 2046 _gvn.set_type(call, call->Value(&_gvn)); 2047 _gvn.set_type(ret, ret->Value(&_gvn)); 2048 2049 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO, 2050 OptoRuntime::store_inline_type_fields_Type(), 2051 StubRoutines::store_inline_type_fields_to_buf(), 2052 nullptr, TypePtr::BOTTOM, ret); 2053 2054 // We don't know how many values are returned. This assumes the 2055 // worst case, that all available registers are used. 2056 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2057 if (domain->field_at(i) == Type::HALF) { 2058 store_to_buf_call->init_req(i, top()); 2059 continue; 2060 } 2061 Node* proj =_gvn.transform(new ProjNode(call, i)); 2062 store_to_buf_call->init_req(i, proj); 2063 } 2064 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false); 2065 2066 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms)); 2067 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL); 2068 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type)); 2069 2070 ideal.set(res, buf); 2071 ideal.sync_kit(this); 2072 } ideal.end_if(); 2073 sync_kit(ideal); 2074 ret = _gvn.transform(ideal.value(res)); 2075 } 2076 if (t->is_klass()) { 2077 const Type* type = TypeOopPtr::make_from_klass(t->as_klass()); 2078 if (type->is_inlinetypeptr()) { 2079 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass()); 2080 } 2081 } 2082 } 2083 2084 return ret; 2085 } 2086 2087 //--------------------set_predefined_input_for_runtime_call-------------------- 2088 // Reading and setting the memory state is way conservative here. 2089 // The real problem is that I am not doing real Type analysis on memory, 2090 // so I cannot distinguish card mark stores from other stores. Across a GC 2091 // point the Store Barrier and the card mark memory has to agree. I cannot 2092 // have a card mark store and its barrier split across the GC point from 2093 // either above or below. Here I get that to happen by reading ALL of memory. 2094 // A better answer would be to separate out card marks from other memory. 2095 // For now, return the input memory state, so that it can be reused 2096 // after the call, if this call has restricted memory effects. 2097 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 2098 // Set fixed predefined input arguments 2099 call->init_req(TypeFunc::Control, control()); 2100 call->init_req(TypeFunc::I_O, top()); // does no i/o 2101 call->init_req(TypeFunc::ReturnAdr, top()); 2102 if (call->is_CallLeafPure()) { 2103 call->init_req(TypeFunc::Memory, top()); 2104 call->init_req(TypeFunc::FramePtr, top()); 2105 return nullptr; 2106 } else { 2107 Node* memory = reset_memory(); 2108 Node* m = narrow_mem == nullptr ? memory : narrow_mem; 2109 call->init_req(TypeFunc::Memory, m); // may gc ptrs 2110 call->init_req(TypeFunc::FramePtr, frameptr()); 2111 return memory; 2112 } 2113 } 2114 2115 //-------------------set_predefined_output_for_runtime_call-------------------- 2116 // Set control and memory (not i_o) from the call. 2117 // If keep_mem is not null, use it for the output state, 2118 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 2119 // If hook_mem is null, this call produces no memory effects at all. 2120 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 2121 // then only that memory slice is taken from the call. 2122 // In the last case, we must put an appropriate memory barrier before 2123 // the call, so as to create the correct anti-dependencies on loads 2124 // preceding the call. 2125 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 2126 Node* keep_mem, 2127 const TypePtr* hook_mem) { 2128 // no i/o 2129 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 2130 if (call->is_CallLeafPure()) { 2131 // Pure function have only control (for now) and data output, in particular 2132 // they don't touch the memory, so we don't want a memory proj that is set after. 2133 return; 2134 } 2135 if (keep_mem) { 2136 // First clone the existing memory state 2137 set_all_memory(keep_mem); 2138 if (hook_mem != nullptr) { 2139 // Make memory for the call 2140 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 2141 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 2142 // We also use hook_mem to extract specific effects from arraycopy stubs. 2143 set_memory(mem, hook_mem); 2144 } 2145 // ...else the call has NO memory effects. 2146 2147 // Make sure the call advertises its memory effects precisely. 2148 // This lets us build accurate anti-dependences in gcm.cpp. 2149 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 2150 "call node must be constructed correctly"); 2151 } else { 2152 assert(hook_mem == nullptr, ""); 2153 // This is not a "slow path" call; all memory comes from the call. 2154 set_all_memory_call(call); 2155 } 2156 } 2157 2158 // Keep track of MergeMems feeding into other MergeMems 2159 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 2160 if (!mem->is_MergeMem()) { 2161 return; 2162 } 2163 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 2164 Node* use = i.get(); 2165 if (use->is_MergeMem()) { 2166 wl.push(use); 2167 } 2168 } 2169 } 2170 2171 // Replace the call with the current state of the kit. 2172 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) { 2173 JVMState* ejvms = nullptr; 2174 if (has_exceptions()) { 2175 ejvms = transfer_exceptions_into_jvms(); 2176 } 2177 2178 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 2179 ReplacedNodes replaced_nodes_exception; 2180 Node* ex_ctl = top(); 2181 2182 SafePointNode* final_state = stop(); 2183 2184 // Find all the needed outputs of this call 2185 CallProjections* callprojs = call->extract_projections(true, do_asserts); 2186 2187 Unique_Node_List wl; 2188 Node* init_mem = call->in(TypeFunc::Memory); 2189 Node* final_mem = final_state->in(TypeFunc::Memory); 2190 Node* final_ctl = final_state->in(TypeFunc::Control); 2191 Node* final_io = final_state->in(TypeFunc::I_O); 2192 2193 // Replace all the old call edges with the edges from the inlining result 2194 if (callprojs->fallthrough_catchproj != nullptr) { 2195 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl); 2196 } 2197 if (callprojs->fallthrough_memproj != nullptr) { 2198 if (final_mem->is_MergeMem()) { 2199 // Parser's exits MergeMem was not transformed but may be optimized 2200 final_mem = _gvn.transform(final_mem); 2201 } 2202 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem); 2203 add_mergemem_users_to_worklist(wl, final_mem); 2204 } 2205 if (callprojs->fallthrough_ioproj != nullptr) { 2206 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io); 2207 } 2208 2209 // Replace the result with the new result if it exists and is used 2210 if (callprojs->resproj[0] != nullptr && result != nullptr) { 2211 // If the inlined code is dead, the result projections for an inline type returned as 2212 // fields have not been replaced. They will go away once the call is replaced by TOP below. 2213 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()), 2214 "unexpected number of results"); 2215 C->gvn_replace_by(callprojs->resproj[0], result); 2216 } 2217 2218 if (ejvms == nullptr) { 2219 // No exception edges to simply kill off those paths 2220 if (callprojs->catchall_catchproj != nullptr) { 2221 C->gvn_replace_by(callprojs->catchall_catchproj, C->top()); 2222 } 2223 if (callprojs->catchall_memproj != nullptr) { 2224 C->gvn_replace_by(callprojs->catchall_memproj, C->top()); 2225 } 2226 if (callprojs->catchall_ioproj != nullptr) { 2227 C->gvn_replace_by(callprojs->catchall_ioproj, C->top()); 2228 } 2229 // Replace the old exception object with top 2230 if (callprojs->exobj != nullptr) { 2231 C->gvn_replace_by(callprojs->exobj, C->top()); 2232 } 2233 } else { 2234 GraphKit ekit(ejvms); 2235 2236 // Load my combined exception state into the kit, with all phis transformed: 2237 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 2238 replaced_nodes_exception = ex_map->replaced_nodes(); 2239 2240 Node* ex_oop = ekit.use_exception_state(ex_map); 2241 2242 if (callprojs->catchall_catchproj != nullptr) { 2243 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control()); 2244 ex_ctl = ekit.control(); 2245 } 2246 if (callprojs->catchall_memproj != nullptr) { 2247 Node* ex_mem = ekit.reset_memory(); 2248 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem); 2249 add_mergemem_users_to_worklist(wl, ex_mem); 2250 } 2251 if (callprojs->catchall_ioproj != nullptr) { 2252 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o()); 2253 } 2254 2255 // Replace the old exception object with the newly created one 2256 if (callprojs->exobj != nullptr) { 2257 C->gvn_replace_by(callprojs->exobj, ex_oop); 2258 } 2259 } 2260 2261 // Disconnect the call from the graph 2262 call->disconnect_inputs(C); 2263 C->gvn_replace_by(call, C->top()); 2264 2265 // Clean up any MergeMems that feed other MergeMems since the 2266 // optimizer doesn't like that. 2267 while (wl.size() > 0) { 2268 _gvn.transform(wl.pop()); 2269 } 2270 2271 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) { 2272 replaced_nodes.apply(C, final_ctl); 2273 } 2274 if (!ex_ctl->is_top() && do_replaced_nodes) { 2275 replaced_nodes_exception.apply(C, ex_ctl); 2276 } 2277 } 2278 2279 2280 //------------------------------increment_counter------------------------------ 2281 // for statistics: increment a VM counter by 1 2282 2283 void GraphKit::increment_counter(address counter_addr) { 2284 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2285 increment_counter(adr1); 2286 } 2287 2288 void GraphKit::increment_counter(Node* counter_addr) { 2289 Node* ctrl = control(); 2290 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered); 2291 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1))); 2292 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered); 2293 } 2294 2295 2296 //------------------------------uncommon_trap---------------------------------- 2297 // Bail out to the interpreter in mid-method. Implemented by calling the 2298 // uncommon_trap blob. This helper function inserts a runtime call with the 2299 // right debug info. 2300 Node* GraphKit::uncommon_trap(int trap_request, 2301 ciKlass* klass, const char* comment, 2302 bool must_throw, 2303 bool keep_exact_action) { 2304 if (failing_internal()) { 2305 stop(); 2306 } 2307 if (stopped()) return nullptr; // trap reachable? 2308 2309 // Note: If ProfileTraps is true, and if a deopt. actually 2310 // occurs here, the runtime will make sure an MDO exists. There is 2311 // no need to call method()->ensure_method_data() at this point. 2312 2313 // Set the stack pointer to the right value for reexecution: 2314 set_sp(reexecute_sp()); 2315 2316 #ifdef ASSERT 2317 if (!must_throw) { 2318 // Make sure the stack has at least enough depth to execute 2319 // the current bytecode. 2320 int inputs, ignored_depth; 2321 if (compute_stack_effects(inputs, ignored_depth)) { 2322 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2323 Bytecodes::name(java_bc()), sp(), inputs); 2324 } 2325 } 2326 #endif 2327 2328 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2329 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2330 2331 switch (action) { 2332 case Deoptimization::Action_maybe_recompile: 2333 case Deoptimization::Action_reinterpret: 2334 // Temporary fix for 6529811 to allow virtual calls to be sure they 2335 // get the chance to go from mono->bi->mega 2336 if (!keep_exact_action && 2337 Deoptimization::trap_request_index(trap_request) < 0 && 2338 too_many_recompiles(reason)) { 2339 // This BCI is causing too many recompilations. 2340 if (C->log() != nullptr) { 2341 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2342 Deoptimization::trap_reason_name(reason), 2343 Deoptimization::trap_action_name(action)); 2344 } 2345 action = Deoptimization::Action_none; 2346 trap_request = Deoptimization::make_trap_request(reason, action); 2347 } else { 2348 C->set_trap_can_recompile(true); 2349 } 2350 break; 2351 case Deoptimization::Action_make_not_entrant: 2352 C->set_trap_can_recompile(true); 2353 break; 2354 case Deoptimization::Action_none: 2355 case Deoptimization::Action_make_not_compilable: 2356 break; 2357 default: 2358 #ifdef ASSERT 2359 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2360 #endif 2361 break; 2362 } 2363 2364 if (TraceOptoParse) { 2365 char buf[100]; 2366 tty->print_cr("Uncommon trap %s at bci:%d", 2367 Deoptimization::format_trap_request(buf, sizeof(buf), 2368 trap_request), bci()); 2369 } 2370 2371 CompileLog* log = C->log(); 2372 if (log != nullptr) { 2373 int kid = (klass == nullptr)? -1: log->identify(klass); 2374 log->begin_elem("uncommon_trap bci='%d'", bci()); 2375 char buf[100]; 2376 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2377 trap_request)); 2378 if (kid >= 0) log->print(" klass='%d'", kid); 2379 if (comment != nullptr) log->print(" comment='%s'", comment); 2380 log->end_elem(); 2381 } 2382 2383 // Make sure any guarding test views this path as very unlikely 2384 Node *i0 = control()->in(0); 2385 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test? 2386 IfNode *iff = i0->as_If(); 2387 float f = iff->_prob; // Get prob 2388 if (control()->Opcode() == Op_IfTrue) { 2389 if (f > PROB_UNLIKELY_MAG(4)) 2390 iff->_prob = PROB_MIN; 2391 } else { 2392 if (f < PROB_LIKELY_MAG(4)) 2393 iff->_prob = PROB_MAX; 2394 } 2395 } 2396 2397 // Clear out dead values from the debug info. 2398 kill_dead_locals(); 2399 2400 // Now insert the uncommon trap subroutine call 2401 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point(); 2402 const TypePtr* no_memory_effects = nullptr; 2403 // Pass the index of the class to be loaded 2404 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2405 (must_throw ? RC_MUST_THROW : 0), 2406 OptoRuntime::uncommon_trap_Type(), 2407 call_addr, "uncommon_trap", no_memory_effects, 2408 intcon(trap_request)); 2409 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2410 "must extract request correctly from the graph"); 2411 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2412 2413 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2414 // The debug info is the only real input to this call. 2415 2416 // Halt-and-catch fire here. The above call should never return! 2417 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2418 PRODUCT_ONLY(COMMA /*reachable*/false)); 2419 _gvn.set_type_bottom(halt); 2420 root()->add_req(halt); 2421 2422 stop_and_kill_map(); 2423 return call; 2424 } 2425 2426 2427 //--------------------------just_allocated_object------------------------------ 2428 // Report the object that was just allocated. 2429 // It must be the case that there are no intervening safepoints. 2430 // We use this to determine if an object is so "fresh" that 2431 // it does not require card marks. 2432 Node* GraphKit::just_allocated_object(Node* current_control) { 2433 Node* ctrl = current_control; 2434 // Object::<init> is invoked after allocation, most of invoke nodes 2435 // will be reduced, but a region node is kept in parse time, we check 2436 // the pattern and skip the region node if it degraded to a copy. 2437 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 && 2438 ctrl->as_Region()->is_copy()) { 2439 ctrl = ctrl->as_Region()->is_copy(); 2440 } 2441 if (C->recent_alloc_ctl() == ctrl) { 2442 return C->recent_alloc_obj(); 2443 } 2444 return nullptr; 2445 } 2446 2447 2448 /** 2449 * Record profiling data exact_kls for Node n with the type system so 2450 * that it can propagate it (speculation) 2451 * 2452 * @param n node that the type applies to 2453 * @param exact_kls type from profiling 2454 * @param maybe_null did profiling see null? 2455 * 2456 * @return node with improved type 2457 */ 2458 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2459 const Type* current_type = _gvn.type(n); 2460 assert(UseTypeSpeculation, "type speculation must be on"); 2461 2462 const TypePtr* speculative = current_type->speculative(); 2463 2464 // Should the klass from the profile be recorded in the speculative type? 2465 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2466 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces); 2467 const TypeOopPtr* xtype = tklass->as_instance_type(); 2468 assert(xtype->klass_is_exact(), "Should be exact"); 2469 // Any reason to believe n is not null (from this profiling or a previous one)? 2470 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2471 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2472 // record the new speculative type's depth 2473 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2474 speculative = speculative->with_inline_depth(jvms()->depth()); 2475 } else if (current_type->would_improve_ptr(ptr_kind)) { 2476 // Profiling report that null was never seen so we can change the 2477 // speculative type to non null ptr. 2478 if (ptr_kind == ProfileAlwaysNull) { 2479 speculative = TypePtr::NULL_PTR; 2480 } else { 2481 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2482 const TypePtr* ptr = TypePtr::NOTNULL; 2483 if (speculative != nullptr) { 2484 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2485 } else { 2486 speculative = ptr; 2487 } 2488 } 2489 } 2490 2491 if (speculative != current_type->speculative()) { 2492 // Build a type with a speculative type (what we think we know 2493 // about the type but will need a guard when we use it) 2494 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2495 // We're changing the type, we need a new CheckCast node to carry 2496 // the new type. The new type depends on the control: what 2497 // profiling tells us is only valid from here as far as we can 2498 // tell. 2499 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2500 cast = _gvn.transform(cast); 2501 replace_in_map(n, cast); 2502 n = cast; 2503 } 2504 2505 return n; 2506 } 2507 2508 /** 2509 * Record profiling data from receiver profiling at an invoke with the 2510 * type system so that it can propagate it (speculation) 2511 * 2512 * @param n receiver node 2513 * 2514 * @return node with improved type 2515 */ 2516 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2517 if (!UseTypeSpeculation) { 2518 return n; 2519 } 2520 ciKlass* exact_kls = profile_has_unique_klass(); 2521 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2522 if ((java_bc() == Bytecodes::_checkcast || 2523 java_bc() == Bytecodes::_instanceof || 2524 java_bc() == Bytecodes::_aastore) && 2525 method()->method_data()->is_mature()) { 2526 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2527 if (data != nullptr) { 2528 if (java_bc() == Bytecodes::_aastore) { 2529 ciKlass* array_type = nullptr; 2530 ciKlass* element_type = nullptr; 2531 ProfilePtrKind element_ptr = ProfileMaybeNull; 2532 bool flat_array = true; 2533 bool null_free_array = true; 2534 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 2535 exact_kls = element_type; 2536 ptr_kind = element_ptr; 2537 } else { 2538 if (!data->as_BitData()->null_seen()) { 2539 ptr_kind = ProfileNeverNull; 2540 } else { 2541 if (TypeProfileCasts) { 2542 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2543 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2544 uint i = 0; 2545 for (; i < call->row_limit(); i++) { 2546 ciKlass* receiver = call->receiver(i); 2547 if (receiver != nullptr) { 2548 break; 2549 } 2550 } 2551 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2552 } 2553 } 2554 } 2555 } 2556 } 2557 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2558 } 2559 2560 /** 2561 * Record profiling data from argument profiling at an invoke with the 2562 * type system so that it can propagate it (speculation) 2563 * 2564 * @param dest_method target method for the call 2565 * @param bc what invoke bytecode is this? 2566 */ 2567 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2568 if (!UseTypeSpeculation) { 2569 return; 2570 } 2571 const TypeFunc* tf = TypeFunc::make(dest_method); 2572 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2573 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2574 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2575 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2576 if (is_reference_type(targ->basic_type())) { 2577 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2578 ciKlass* better_type = nullptr; 2579 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2580 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2581 } 2582 i++; 2583 } 2584 } 2585 } 2586 2587 /** 2588 * Record profiling data from parameter profiling at an invoke with 2589 * the type system so that it can propagate it (speculation) 2590 */ 2591 void GraphKit::record_profiled_parameters_for_speculation() { 2592 if (!UseTypeSpeculation) { 2593 return; 2594 } 2595 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2596 if (_gvn.type(local(i))->isa_oopptr()) { 2597 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2598 ciKlass* better_type = nullptr; 2599 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2600 record_profile_for_speculation(local(i), better_type, ptr_kind); 2601 } 2602 j++; 2603 } 2604 } 2605 } 2606 2607 /** 2608 * Record profiling data from return value profiling at an invoke with 2609 * the type system so that it can propagate it (speculation) 2610 */ 2611 void GraphKit::record_profiled_return_for_speculation() { 2612 if (!UseTypeSpeculation) { 2613 return; 2614 } 2615 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2616 ciKlass* better_type = nullptr; 2617 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2618 // If profiling reports a single type for the return value, 2619 // feed it to the type system so it can propagate it as a 2620 // speculative type 2621 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2622 } 2623 } 2624 2625 2626 //============================================================================= 2627 // Generate a fast path/slow path idiom. Graph looks like: 2628 // [foo] indicates that 'foo' is a parameter 2629 // 2630 // [in] null 2631 // \ / 2632 // CmpP 2633 // Bool ne 2634 // If 2635 // / \ 2636 // True False-<2> 2637 // / | 2638 // / cast_not_null 2639 // Load | | ^ 2640 // [fast_test] | | 2641 // gvn to opt_test | | 2642 // / \ | <1> 2643 // True False | 2644 // | \\ | 2645 // [slow_call] \[fast_result] 2646 // Ctl Val \ \ 2647 // | \ \ 2648 // Catch <1> \ \ 2649 // / \ ^ \ \ 2650 // Ex No_Ex | \ \ 2651 // | \ \ | \ <2> \ 2652 // ... \ [slow_res] | | \ [null_result] 2653 // \ \--+--+--- | | 2654 // \ | / \ | / 2655 // --------Region Phi 2656 // 2657 //============================================================================= 2658 // Code is structured as a series of driver functions all called 'do_XXX' that 2659 // call a set of helper functions. Helper functions first, then drivers. 2660 2661 //------------------------------null_check_oop--------------------------------- 2662 // Null check oop. Set null-path control into Region in slot 3. 2663 // Make a cast-not-nullness use the other not-null control. Return cast. 2664 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2665 bool never_see_null, 2666 bool safe_for_replace, 2667 bool speculative) { 2668 // Initial null check taken path 2669 (*null_control) = top(); 2670 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2671 2672 // Generate uncommon_trap: 2673 if (never_see_null && (*null_control) != top()) { 2674 // If we see an unexpected null at a check-cast we record it and force a 2675 // recompile; the offending check-cast will be compiled to handle nulls. 2676 // If we see more than one offending BCI, then all checkcasts in the 2677 // method will be compiled to handle nulls. 2678 PreserveJVMState pjvms(this); 2679 set_control(*null_control); 2680 replace_in_map(value, null()); 2681 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2682 uncommon_trap(reason, 2683 Deoptimization::Action_make_not_entrant); 2684 (*null_control) = top(); // null path is dead 2685 } 2686 if ((*null_control) == top() && safe_for_replace) { 2687 replace_in_map(value, cast); 2688 } 2689 2690 // Cast away null-ness on the result 2691 return cast; 2692 } 2693 2694 //------------------------------opt_iff---------------------------------------- 2695 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2696 // Return slow-path control. 2697 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2698 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2699 2700 // Fast path taken; set region slot 2 2701 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2702 region->init_req(2,fast_taken); // Capture fast-control 2703 2704 // Fast path not-taken, i.e. slow path 2705 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2706 return slow_taken; 2707 } 2708 2709 //-----------------------------make_runtime_call------------------------------- 2710 Node* GraphKit::make_runtime_call(int flags, 2711 const TypeFunc* call_type, address call_addr, 2712 const char* call_name, 2713 const TypePtr* adr_type, 2714 // The following parms are all optional. 2715 // The first null ends the list. 2716 Node* parm0, Node* parm1, 2717 Node* parm2, Node* parm3, 2718 Node* parm4, Node* parm5, 2719 Node* parm6, Node* parm7) { 2720 assert(call_addr != nullptr, "must not call null targets"); 2721 2722 // Slow-path call 2723 bool is_leaf = !(flags & RC_NO_LEAF); 2724 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2725 if (call_name == nullptr) { 2726 assert(!is_leaf, "must supply name for leaf"); 2727 call_name = OptoRuntime::stub_name(call_addr); 2728 } 2729 CallNode* call; 2730 if (!is_leaf) { 2731 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type); 2732 } else if (flags & RC_NO_FP) { 2733 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2734 } else if (flags & RC_VECTOR){ 2735 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte; 2736 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits); 2737 } else if (flags & RC_PURE) { 2738 call = new CallLeafPureNode(call_type, call_addr, call_name, adr_type); 2739 } else { 2740 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2741 } 2742 2743 // The following is similar to set_edges_for_java_call, 2744 // except that the memory effects of the call are restricted to AliasIdxRaw. 2745 2746 // Slow path call has no side-effects, uses few values 2747 bool wide_in = !(flags & RC_NARROW_MEM); 2748 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2749 2750 Node* prev_mem = nullptr; 2751 if (wide_in) { 2752 prev_mem = set_predefined_input_for_runtime_call(call); 2753 } else { 2754 assert(!wide_out, "narrow in => narrow out"); 2755 Node* narrow_mem = memory(adr_type); 2756 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2757 } 2758 2759 // Hook each parm in order. Stop looking at the first null. 2760 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0); 2761 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1); 2762 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2); 2763 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3); 2764 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4); 2765 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5); 2766 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6); 2767 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7); 2768 /* close each nested if ===> */ } } } } } } } } 2769 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms"); 2770 2771 if (!is_leaf) { 2772 // Non-leaves can block and take safepoints: 2773 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2774 } 2775 // Non-leaves can throw exceptions: 2776 if (has_io) { 2777 call->set_req(TypeFunc::I_O, i_o()); 2778 } 2779 2780 if (flags & RC_UNCOMMON) { 2781 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2782 // (An "if" probability corresponds roughly to an unconditional count. 2783 // Sort of.) 2784 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2785 } 2786 2787 Node* c = _gvn.transform(call); 2788 assert(c == call, "cannot disappear"); 2789 2790 if (wide_out) { 2791 // Slow path call has full side-effects. 2792 set_predefined_output_for_runtime_call(call); 2793 } else { 2794 // Slow path call has few side-effects, and/or sets few values. 2795 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2796 } 2797 2798 if (has_io) { 2799 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2800 } 2801 return call; 2802 2803 } 2804 2805 // i2b 2806 Node* GraphKit::sign_extend_byte(Node* in) { 2807 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24))); 2808 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24))); 2809 } 2810 2811 // i2s 2812 Node* GraphKit::sign_extend_short(Node* in) { 2813 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16))); 2814 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16))); 2815 } 2816 2817 2818 //------------------------------merge_memory----------------------------------- 2819 // Merge memory from one path into the current memory state. 2820 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2821 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2822 Node* old_slice = mms.force_memory(); 2823 Node* new_slice = mms.memory2(); 2824 if (old_slice != new_slice) { 2825 PhiNode* phi; 2826 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2827 if (mms.is_empty()) { 2828 // clone base memory Phi's inputs for this memory slice 2829 assert(old_slice == mms.base_memory(), "sanity"); 2830 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C)); 2831 _gvn.set_type(phi, Type::MEMORY); 2832 for (uint i = 1; i < phi->req(); i++) { 2833 phi->init_req(i, old_slice->in(i)); 2834 } 2835 } else { 2836 phi = old_slice->as_Phi(); // Phi was generated already 2837 } 2838 } else { 2839 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2840 _gvn.set_type(phi, Type::MEMORY); 2841 } 2842 phi->set_req(new_path, new_slice); 2843 mms.set_memory(phi); 2844 } 2845 } 2846 } 2847 2848 //------------------------------make_slow_call_ex------------------------------ 2849 // Make the exception handler hookups for the slow call 2850 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2851 if (stopped()) return; 2852 2853 // Make a catch node with just two handlers: fall-through and catch-all 2854 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2855 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2856 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2857 _gvn.set_type_bottom(norm); 2858 C->record_for_igvn(norm); 2859 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2860 2861 { PreserveJVMState pjvms(this); 2862 set_control(excp); 2863 set_i_o(i_o); 2864 2865 if (excp != top()) { 2866 if (deoptimize) { 2867 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2868 uncommon_trap(Deoptimization::Reason_unhandled, 2869 Deoptimization::Action_none); 2870 } else { 2871 // Create an exception state also. 2872 // Use an exact type if the caller has a specific exception. 2873 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2874 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2875 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2876 } 2877 } 2878 } 2879 2880 // Get the no-exception control from the CatchNode. 2881 set_control(norm); 2882 } 2883 2884 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) { 2885 Node* cmp = nullptr; 2886 switch(bt) { 2887 case T_INT: cmp = new CmpINode(in1, in2); break; 2888 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2889 default: fatal("unexpected comparison type %s", type2name(bt)); 2890 } 2891 cmp = gvn.transform(cmp); 2892 Node* bol = gvn.transform(new BoolNode(cmp, test)); 2893 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2894 gvn.transform(iff); 2895 if (!bol->is_Con()) gvn.record_for_igvn(iff); 2896 return iff; 2897 } 2898 2899 //-------------------------------gen_subtype_check----------------------------- 2900 // Generate a subtyping check. Takes as input the subtype and supertype. 2901 // Returns 2 values: sets the default control() to the true path and returns 2902 // the false path. Only reads invariant memory; sets no (visible) memory. 2903 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2904 // but that's not exposed to the optimizer. This call also doesn't take in an 2905 // Object; if you wish to check an Object you need to load the Object's class 2906 // prior to coming here. 2907 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn, 2908 ciMethod* method, int bci) { 2909 Compile* C = gvn.C; 2910 if ((*ctrl)->is_top()) { 2911 return C->top(); 2912 } 2913 2914 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr(); 2915 const TypeAryKlassPtr* ary_klass_t = klass_ptr_type->isa_aryklassptr(); 2916 Node* vm_superklass = superklass; 2917 // TODO 8366668 Compute the VM type here for when we do a direct pointer comparison 2918 if (ary_klass_t && ary_klass_t->klass_is_exact() && ary_klass_t->exact_klass()->is_obj_array_klass()) { 2919 ary_klass_t = ary_klass_t->get_vm_type(); 2920 vm_superklass = gvn.makecon(ary_klass_t); 2921 } 2922 2923 // Fast check for identical types, perhaps identical constants. 2924 // The types can even be identical non-constants, in cases 2925 // involving Array.newInstance, Object.clone, etc. 2926 if (subklass == superklass) 2927 return C->top(); // false path is dead; no test needed. 2928 2929 if (gvn.type(superklass)->singleton()) { 2930 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2931 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr(); 2932 2933 // In the common case of an exact superklass, try to fold up the 2934 // test before generating code. You may ask, why not just generate 2935 // the code and then let it fold up? The answer is that the generated 2936 // code will necessarily include null checks, which do not always 2937 // completely fold away. If they are also needless, then they turn 2938 // into a performance loss. Example: 2939 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2940 // Here, the type of 'fa' is often exact, so the store check 2941 // of fa[1]=x will fold up, without testing the nullness of x. 2942 // 2943 // At macro expansion, we would have already folded the SubTypeCheckNode 2944 // being expanded here because we always perform the static sub type 2945 // check in SubTypeCheckNode::sub() regardless of whether 2946 // StressReflectiveCode is set or not. We can therefore skip this 2947 // static check when StressReflectiveCode is on. 2948 switch (C->static_subtype_check(superk, subk)) { 2949 case Compile::SSC_always_false: 2950 { 2951 Node* always_fail = *ctrl; 2952 *ctrl = gvn.C->top(); 2953 return always_fail; 2954 } 2955 case Compile::SSC_always_true: 2956 return C->top(); 2957 case Compile::SSC_easy_test: 2958 { 2959 // Just do a direct pointer compare and be done. 2960 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2961 *ctrl = gvn.transform(new IfTrueNode(iff)); 2962 return gvn.transform(new IfFalseNode(iff)); 2963 } 2964 case Compile::SSC_full_test: 2965 break; 2966 default: 2967 ShouldNotReachHere(); 2968 } 2969 } 2970 2971 // %%% Possible further optimization: Even if the superklass is not exact, 2972 // if the subklass is the unique subtype of the superklass, the check 2973 // will always succeed. We could leave a dependency behind to ensure this. 2974 2975 // First load the super-klass's check-offset 2976 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2977 Node* m = C->immutable_memory(); 2978 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2979 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2980 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int(); 2981 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con; 2982 // TODO 8366668 Re-enable. This breaks test/hotspot/jtreg/compiler/c2/irTests/ProfileAtTypeCheck.java 2983 bool might_be_cache = true;//(chk_off_con == cacheoff_con); 2984 2985 // Load from the sub-klass's super-class display list, or a 1-word cache of 2986 // the secondary superclass list, or a failing value with a sentinel offset 2987 // if the super-klass is an interface or exceptionally deep in the Java 2988 // hierarchy and we have to scan the secondary superclass list the hard way. 2989 // Worst-case type is a little odd: null is allowed as a result (usually 2990 // klass loads can never produce a null). 2991 Node *chk_off_X = chk_off; 2992 #ifdef _LP64 2993 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X)); 2994 #endif 2995 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X)); 2996 // For some types like interfaces the following loadKlass is from a 1-word 2997 // cache which is mutable so can't use immutable memory. Other 2998 // types load from the super-class display table which is immutable. 2999 Node *kmem = C->immutable_memory(); 3000 // secondary_super_cache is not immutable but can be treated as such because: 3001 // - no ideal node writes to it in a way that could cause an 3002 // incorrect/missed optimization of the following Load. 3003 // - it's a cache so, worse case, not reading the latest value 3004 // wouldn't cause incorrect execution 3005 if (might_be_cache && mem != nullptr) { 3006 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem; 3007 } 3008 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL)); 3009 3010 // Compile speed common case: ARE a subtype and we canNOT fail 3011 if (superklass == nkls) { 3012 return C->top(); // false path is dead; no test needed. 3013 } 3014 3015 // Gather the various success & failures here 3016 RegionNode* r_not_subtype = new RegionNode(3); 3017 gvn.record_for_igvn(r_not_subtype); 3018 RegionNode* r_ok_subtype = new RegionNode(4); 3019 gvn.record_for_igvn(r_ok_subtype); 3020 3021 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the 3022 // SubTypeCheck node 3023 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) { 3024 ciCallProfile profile = method->call_profile_at_bci(bci); 3025 float total_prob = 0; 3026 for (int i = 0; profile.has_receiver(i); ++i) { 3027 float prob = profile.receiver_prob(i); 3028 total_prob += prob; 3029 } 3030 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) { 3031 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 3032 for (int i = 0; profile.has_receiver(i); ++i) { 3033 ciKlass* klass = profile.receiver(i); 3034 // TODO 8366668 Do we need adjustments here?? 3035 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass); 3036 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t); 3037 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) { 3038 continue; 3039 } 3040 float prob = profile.receiver_prob(i); 3041 ConNode* klass_node = gvn.makecon(klass_t); 3042 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS); 3043 Node* iftrue = gvn.transform(new IfTrueNode(iff)); 3044 3045 if (result == Compile::SSC_always_true) { 3046 r_ok_subtype->add_req(iftrue); 3047 } else { 3048 assert(result == Compile::SSC_always_false, ""); 3049 r_not_subtype->add_req(iftrue); 3050 } 3051 *ctrl = gvn.transform(new IfFalseNode(iff)); 3052 } 3053 } 3054 } 3055 3056 // See if we get an immediate positive hit. Happens roughly 83% of the 3057 // time. Test to see if the value loaded just previously from the subklass 3058 // is exactly the superklass. 3059 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 3060 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1)); 3061 *ctrl = gvn.transform(new IfFalseNode(iff1)); 3062 3063 // Compile speed common case: Check for being deterministic right now. If 3064 // chk_off is a constant and not equal to cacheoff then we are NOT a 3065 // subklass. In this case we need exactly the 1 test above and we can 3066 // return those results immediately. 3067 if (!might_be_cache) { 3068 Node* not_subtype_ctrl = *ctrl; 3069 *ctrl = iftrue1; // We need exactly the 1 test above 3070 PhaseIterGVN* igvn = gvn.is_IterGVN(); 3071 if (igvn != nullptr) { 3072 igvn->remove_globally_dead_node(r_ok_subtype); 3073 igvn->remove_globally_dead_node(r_not_subtype); 3074 } 3075 return not_subtype_ctrl; 3076 } 3077 3078 r_ok_subtype->init_req(1, iftrue1); 3079 3080 // Check for immediate negative hit. Happens roughly 11% of the time (which 3081 // is roughly 63% of the remaining cases). Test to see if the loaded 3082 // check-offset points into the subklass display list or the 1-element 3083 // cache. If it points to the display (and NOT the cache) and the display 3084 // missed then it's not a subtype. 3085 // TODO 8366668 Re-enable 3086 /* 3087 Node *cacheoff = gvn.intcon(cacheoff_con); 3088 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 3089 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2))); 3090 *ctrl = gvn.transform(new IfFalseNode(iff2)); 3091 */ 3092 // Check for self. Very rare to get here, but it is taken 1/3 the time. 3093 // No performance impact (too rare) but allows sharing of secondary arrays 3094 // which has some footprint reduction. 3095 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 3096 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3))); 3097 *ctrl = gvn.transform(new IfFalseNode(iff3)); 3098 3099 // -- Roads not taken here: -- 3100 // We could also have chosen to perform the self-check at the beginning 3101 // of this code sequence, as the assembler does. This would not pay off 3102 // the same way, since the optimizer, unlike the assembler, can perform 3103 // static type analysis to fold away many successful self-checks. 3104 // Non-foldable self checks work better here in second position, because 3105 // the initial primary superclass check subsumes a self-check for most 3106 // types. An exception would be a secondary type like array-of-interface, 3107 // which does not appear in its own primary supertype display. 3108 // Finally, we could have chosen to move the self-check into the 3109 // PartialSubtypeCheckNode, and from there out-of-line in a platform 3110 // dependent manner. But it is worthwhile to have the check here, 3111 // where it can be perhaps be optimized. The cost in code space is 3112 // small (register compare, branch). 3113 3114 // Now do a linear scan of the secondary super-klass array. Again, no real 3115 // performance impact (too rare) but it's gotta be done. 3116 // Since the code is rarely used, there is no penalty for moving it 3117 // out of line, and it can only improve I-cache density. 3118 // The decision to inline or out-of-line this final check is platform 3119 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 3120 Node* psc = gvn.transform( 3121 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 3122 3123 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 3124 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4))); 3125 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4))); 3126 3127 // Return false path; set default control to true path. 3128 *ctrl = gvn.transform(r_ok_subtype); 3129 return gvn.transform(r_not_subtype); 3130 } 3131 3132 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) { 3133 const Type* sub_t = _gvn.type(obj_or_subklass); 3134 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) { 3135 sub_t = TypeKlassPtr::make(sub_t->inline_klass()); 3136 obj_or_subklass = makecon(sub_t); 3137 } 3138 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over 3139 if (expand_subtype_check) { 3140 MergeMemNode* mem = merged_memory(); 3141 Node* ctrl = control(); 3142 Node* subklass = obj_or_subklass; 3143 if (!sub_t->isa_klassptr()) { 3144 subklass = load_object_klass(obj_or_subklass); 3145 } 3146 3147 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci()); 3148 set_control(ctrl); 3149 return n; 3150 } 3151 3152 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci())); 3153 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq)); 3154 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3155 set_control(_gvn.transform(new IfTrueNode(iff))); 3156 return _gvn.transform(new IfFalseNode(iff)); 3157 } 3158 3159 // Profile-driven exact type check: 3160 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 3161 float prob, Node* *casted_receiver) { 3162 assert(!klass->is_interface(), "no exact type check on interfaces"); 3163 Node* fail = top(); 3164 const Type* rec_t = _gvn.type(receiver); 3165 if (rec_t->is_inlinetypeptr()) { 3166 if (klass->equals(rec_t->inline_klass())) { 3167 (*casted_receiver) = receiver; // Always passes 3168 } else { 3169 (*casted_receiver) = top(); // Always fails 3170 fail = control(); 3171 set_control(top()); 3172 } 3173 return fail; 3174 } 3175 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces); 3176 const TypeAryKlassPtr* ary_klass_t = tklass->isa_aryklassptr(); 3177 // TODO 8366668 Compute the VM type 3178 if (ary_klass_t && ary_klass_t->klass_is_exact() && ary_klass_t->exact_klass()->is_obj_array_klass()) { 3179 tklass = ary_klass_t->get_vm_type(); 3180 } 3181 Node* recv_klass = load_object_klass(receiver); 3182 fail = type_check(recv_klass, tklass, prob); 3183 3184 if (!stopped()) { 3185 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3186 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 3187 assert(recv_xtype->klass_is_exact(), ""); 3188 3189 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts 3190 // Subsume downstream occurrences of receiver with a cast to 3191 // recv_xtype, since now we know what the type will be. 3192 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 3193 Node* res = _gvn.transform(cast); 3194 if (recv_xtype->is_inlinetypeptr()) { 3195 assert(!gvn().type(res)->maybe_null(), "receiver should never be null"); 3196 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass()); 3197 } 3198 (*casted_receiver) = res; 3199 assert(!(*casted_receiver)->is_top(), "that path should be unreachable"); 3200 // (User must make the replace_in_map call.) 3201 } 3202 } 3203 3204 return fail; 3205 } 3206 3207 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 3208 float prob) { 3209 Node* want_klass = makecon(tklass); 3210 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass)); 3211 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3212 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 3213 set_control(_gvn.transform(new IfTrueNode (iff))); 3214 Node* fail = _gvn.transform(new IfFalseNode(iff)); 3215 return fail; 3216 } 3217 3218 //------------------------------subtype_check_receiver------------------------- 3219 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 3220 Node** casted_receiver) { 3221 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve(); 3222 Node* want_klass = makecon(tklass); 3223 3224 Node* slow_ctl = gen_subtype_check(receiver, want_klass); 3225 3226 // Ignore interface type information until interface types are properly tracked. 3227 if (!stopped() && !klass->is_interface()) { 3228 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3229 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 3230 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts 3231 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type)); 3232 if (recv_type->is_inlinetypeptr()) { 3233 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass()); 3234 } 3235 (*casted_receiver) = cast; 3236 } 3237 } 3238 3239 return slow_ctl; 3240 } 3241 3242 //------------------------------seems_never_null------------------------------- 3243 // Use null_seen information if it is available from the profile. 3244 // If we see an unexpected null at a type check we record it and force a 3245 // recompile; the offending check will be recompiled to handle nulls. 3246 // If we see several offending BCIs, then all checks in the 3247 // method will be recompiled. 3248 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 3249 speculating = !_gvn.type(obj)->speculative_maybe_null(); 3250 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 3251 if (UncommonNullCast // Cutout for this technique 3252 && obj != null() // And not the -Xcomp stupid case? 3253 && !too_many_traps(reason) 3254 ) { 3255 if (speculating) { 3256 return true; 3257 } 3258 if (data == nullptr) 3259 // Edge case: no mature data. Be optimistic here. 3260 return true; 3261 // If the profile has not seen a null, assume it won't happen. 3262 assert(java_bc() == Bytecodes::_checkcast || 3263 java_bc() == Bytecodes::_instanceof || 3264 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 3265 return !data->as_BitData()->null_seen(); 3266 } 3267 speculating = false; 3268 return false; 3269 } 3270 3271 void GraphKit::guard_klass_being_initialized(Node* klass) { 3272 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 3273 Node* adr = basic_plus_adr(top(), klass, init_state_off); 3274 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3275 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 3276 T_BYTE, MemNode::acquire); 3277 init_state = _gvn.transform(init_state); 3278 3279 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 3280 3281 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 3282 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3283 3284 { BuildCutout unless(this, tst, PROB_MAX); 3285 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 3286 } 3287 } 3288 3289 void GraphKit::guard_init_thread(Node* klass) { 3290 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 3291 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 3292 3293 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr, 3294 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 3295 T_ADDRESS, MemNode::unordered); 3296 init_thread = _gvn.transform(init_thread); 3297 3298 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 3299 3300 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 3301 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3302 3303 { BuildCutout unless(this, tst, PROB_MAX); 3304 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 3305 } 3306 } 3307 3308 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3309 if (ik->is_being_initialized()) { 3310 if (C->needs_clinit_barrier(ik, context)) { 3311 Node* klass = makecon(TypeKlassPtr::make(ik)); 3312 guard_klass_being_initialized(klass); 3313 guard_init_thread(klass); 3314 insert_mem_bar(Op_MemBarCPUOrder); 3315 } 3316 } else if (ik->is_initialized()) { 3317 return; // no barrier needed 3318 } else { 3319 uncommon_trap(Deoptimization::Reason_uninitialized, 3320 Deoptimization::Action_reinterpret, 3321 nullptr); 3322 } 3323 } 3324 3325 //------------------------maybe_cast_profiled_receiver------------------------- 3326 // If the profile has seen exactly one type, narrow to exactly that type. 3327 // Subsequent type checks will always fold up. 3328 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3329 const TypeKlassPtr* require_klass, 3330 ciKlass* spec_klass, 3331 bool safe_for_replace) { 3332 if (!UseTypeProfile || !TypeProfileCasts) return nullptr; 3333 3334 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr); 3335 3336 // Make sure we haven't already deoptimized from this tactic. 3337 if (too_many_traps_or_recompiles(reason)) 3338 return nullptr; 3339 3340 // (No, this isn't a call, but it's enough like a virtual call 3341 // to use the same ciMethod accessor to get the profile info...) 3342 // If we have a speculative type use it instead of profiling (which 3343 // may not help us) 3344 ciKlass* exact_kls = spec_klass; 3345 if (exact_kls == nullptr) { 3346 if (java_bc() == Bytecodes::_aastore) { 3347 ciKlass* array_type = nullptr; 3348 ciKlass* element_type = nullptr; 3349 ProfilePtrKind element_ptr = ProfileMaybeNull; 3350 bool flat_array = true; 3351 bool null_free_array = true; 3352 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 3353 exact_kls = element_type; 3354 } else { 3355 exact_kls = profile_has_unique_klass(); 3356 } 3357 } 3358 if (exact_kls != nullptr) {// no cast failures here 3359 if (require_klass == nullptr || 3360 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) { 3361 // If we narrow the type to match what the type profile sees or 3362 // the speculative type, we can then remove the rest of the 3363 // cast. 3364 // This is a win, even if the exact_kls is very specific, 3365 // because downstream operations, such as method calls, 3366 // will often benefit from the sharper type. 3367 Node* exact_obj = not_null_obj; // will get updated in place... 3368 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3369 &exact_obj); 3370 { PreserveJVMState pjvms(this); 3371 set_control(slow_ctl); 3372 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3373 } 3374 if (safe_for_replace) { 3375 replace_in_map(not_null_obj, exact_obj); 3376 } 3377 return exact_obj; 3378 } 3379 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3380 } 3381 3382 return nullptr; 3383 } 3384 3385 /** 3386 * Cast obj to type and emit guard unless we had too many traps here 3387 * already 3388 * 3389 * @param obj node being casted 3390 * @param type type to cast the node to 3391 * @param not_null true if we know node cannot be null 3392 */ 3393 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3394 ciKlass* type, 3395 bool not_null) { 3396 if (stopped()) { 3397 return obj; 3398 } 3399 3400 // type is null if profiling tells us this object is always null 3401 if (type != nullptr) { 3402 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3403 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3404 3405 if (!too_many_traps_or_recompiles(null_reason) && 3406 !too_many_traps_or_recompiles(class_reason)) { 3407 Node* not_null_obj = nullptr; 3408 // not_null is true if we know the object is not null and 3409 // there's no need for a null check 3410 if (!not_null) { 3411 Node* null_ctl = top(); 3412 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3413 assert(null_ctl->is_top(), "no null control here"); 3414 } else { 3415 not_null_obj = obj; 3416 } 3417 3418 Node* exact_obj = not_null_obj; 3419 ciKlass* exact_kls = type; 3420 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3421 &exact_obj); 3422 { 3423 PreserveJVMState pjvms(this); 3424 set_control(slow_ctl); 3425 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3426 } 3427 replace_in_map(not_null_obj, exact_obj); 3428 obj = exact_obj; 3429 } 3430 } else { 3431 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3432 Node* exact_obj = null_assert(obj); 3433 replace_in_map(obj, exact_obj); 3434 obj = exact_obj; 3435 } 3436 } 3437 return obj; 3438 } 3439 3440 //-------------------------------gen_instanceof-------------------------------- 3441 // Generate an instance-of idiom. Used by both the instance-of bytecode 3442 // and the reflective instance-of call. 3443 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3444 kill_dead_locals(); // Benefit all the uncommon traps 3445 assert( !stopped(), "dead parse path should be checked in callers" ); 3446 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3447 "must check for not-null not-dead klass in callers"); 3448 3449 // Make the merge point 3450 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3451 RegionNode* region = new RegionNode(PATH_LIMIT); 3452 Node* phi = new PhiNode(region, TypeInt::BOOL); 3453 C->set_has_split_ifs(true); // Has chance for split-if optimization 3454 3455 ciProfileData* data = nullptr; 3456 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3457 data = method()->method_data()->bci_to_data(bci()); 3458 } 3459 bool speculative_not_null = false; 3460 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3461 && seems_never_null(obj, data, speculative_not_null)); 3462 3463 // Null check; get casted pointer; set region slot 3 3464 Node* null_ctl = top(); 3465 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3466 3467 // If not_null_obj is dead, only null-path is taken 3468 if (stopped()) { // Doing instance-of on a null? 3469 set_control(null_ctl); 3470 return intcon(0); 3471 } 3472 region->init_req(_null_path, null_ctl); 3473 phi ->init_req(_null_path, intcon(0)); // Set null path value 3474 if (null_ctl == top()) { 3475 // Do this eagerly, so that pattern matches like is_diamond_phi 3476 // will work even during parsing. 3477 assert(_null_path == PATH_LIMIT-1, "delete last"); 3478 region->del_req(_null_path); 3479 phi ->del_req(_null_path); 3480 } 3481 3482 // Do we know the type check always succeed? 3483 bool known_statically = false; 3484 if (_gvn.type(superklass)->singleton()) { 3485 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr(); 3486 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type(); 3487 if (subk != nullptr && subk->is_loaded()) { 3488 int static_res = C->static_subtype_check(superk, subk); 3489 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3490 } 3491 } 3492 3493 if (!known_statically) { 3494 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3495 // We may not have profiling here or it may not help us. If we 3496 // have a speculative type use it to perform an exact cast. 3497 ciKlass* spec_obj_type = obj_type->speculative_type(); 3498 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) { 3499 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace); 3500 if (stopped()) { // Profile disagrees with this path. 3501 set_control(null_ctl); // Null is the only remaining possibility. 3502 return intcon(0); 3503 } 3504 if (cast_obj != nullptr) { 3505 not_null_obj = cast_obj; 3506 } 3507 } 3508 } 3509 3510 // Generate the subtype check 3511 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3512 3513 // Plug in the success path to the general merge in slot 1. 3514 region->init_req(_obj_path, control()); 3515 phi ->init_req(_obj_path, intcon(1)); 3516 3517 // Plug in the failing path to the general merge in slot 2. 3518 region->init_req(_fail_path, not_subtype_ctrl); 3519 phi ->init_req(_fail_path, intcon(0)); 3520 3521 // Return final merged results 3522 set_control( _gvn.transform(region) ); 3523 record_for_igvn(region); 3524 3525 // If we know the type check always succeeds then we don't use the 3526 // profiling data at this bytecode. Don't lose it, feed it to the 3527 // type system as a speculative type. 3528 if (safe_for_replace) { 3529 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3530 replace_in_map(obj, casted_obj); 3531 } 3532 3533 return _gvn.transform(phi); 3534 } 3535 3536 //-------------------------------gen_checkcast--------------------------------- 3537 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3538 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3539 // uncommon-trap paths work. Adjust stack after this call. 3540 // If failure_control is supplied and not null, it is filled in with 3541 // the control edge for the cast failure. Otherwise, an appropriate 3542 // uncommon trap or exception is thrown. 3543 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) { 3544 kill_dead_locals(); // Benefit all the uncommon traps 3545 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr(); 3546 const Type* obj_type = _gvn.type(obj); 3547 if (obj_type->is_inlinetypeptr() && !obj_type->maybe_null() && klass_ptr_type->klass_is_exact() && obj_type->inline_klass() == klass_ptr_type->exact_klass(true)) { 3548 // Special case: larval inline objects must not be scalarized. They are also generally not 3549 // allowed to participate in most operations except as the first operand of putfield, or as an 3550 // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being 3551 // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize 3552 // value objects in all other places. This special case comes from the limitation of the Java 3553 // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete 3554 // value type. We must do this first because C->static_subtype_check may do nothing when 3555 // StressReflectiveCode is set. 3556 return obj; 3557 } 3558 3559 // Else it must be a non-larval object 3560 obj = cast_to_non_larval(obj); 3561 3562 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve(); 3563 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type(); 3564 bool safe_for_replace = (failure_control == nullptr); 3565 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer"); 3566 3567 // Fast cutout: Check the case that the cast is vacuously true. 3568 // This detects the common cases where the test will short-circuit 3569 // away completely. We do this before we perform the null check, 3570 // because if the test is going to turn into zero code, we don't 3571 // want a residual null check left around. (Causes a slowdown, 3572 // for example, in some objArray manipulations, such as a[i]=a[j].) 3573 if (improved_klass_ptr_type->singleton()) { 3574 const TypeKlassPtr* kptr = nullptr; 3575 if (obj_type->isa_oop_ptr()) { 3576 kptr = obj_type->is_oopptr()->as_klass_type(); 3577 } else if (obj->is_InlineType()) { 3578 ciInlineKlass* vk = obj_type->inline_klass(); 3579 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0)); 3580 } 3581 3582 if (kptr != nullptr) { 3583 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) { 3584 case Compile::SSC_always_true: 3585 // If we know the type check always succeed then we don't use 3586 // the profiling data at this bytecode. Don't lose it, feed it 3587 // to the type system as a speculative type. 3588 obj = record_profiled_receiver_for_speculation(obj); 3589 if (null_free) { 3590 assert(safe_for_replace, "must be"); 3591 obj = null_check(obj); 3592 } 3593 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized"); 3594 return obj; 3595 case Compile::SSC_always_false: 3596 if (null_free) { 3597 assert(safe_for_replace, "must be"); 3598 obj = null_check(obj); 3599 } 3600 // It needs a null check because a null will *pass* the cast check. 3601 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) { 3602 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3603 Deoptimization::DeoptReason reason = is_aastore ? 3604 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3605 builtin_throw(reason); 3606 return top(); 3607 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3608 return null_assert(obj); 3609 } 3610 break; // Fall through to full check 3611 default: 3612 break; 3613 } 3614 } 3615 } 3616 3617 ciProfileData* data = nullptr; 3618 if (failure_control == nullptr) { // use MDO in regular case only 3619 assert(java_bc() == Bytecodes::_aastore || 3620 java_bc() == Bytecodes::_checkcast, 3621 "interpreter profiles type checks only for these BCs"); 3622 if (method()->method_data()->is_mature()) { 3623 data = method()->method_data()->bci_to_data(bci()); 3624 } 3625 } 3626 3627 // Make the merge point 3628 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3629 RegionNode* region = new RegionNode(PATH_LIMIT); 3630 Node* phi = new PhiNode(region, toop); 3631 _gvn.set_type(region, Type::CONTROL); 3632 _gvn.set_type(phi, toop); 3633 3634 C->set_has_split_ifs(true); // Has chance for split-if optimization 3635 3636 // Use null-cast information if it is available 3637 bool speculative_not_null = false; 3638 bool never_see_null = ((failure_control == nullptr) // regular case only 3639 && seems_never_null(obj, data, speculative_not_null)); 3640 3641 if (obj->is_InlineType()) { 3642 // Re-execute if buffering during triggers deoptimization 3643 PreserveReexecuteState preexecs(this); 3644 jvms()->set_should_reexecute(true); 3645 obj = obj->as_InlineType()->buffer(this, safe_for_replace); 3646 } 3647 3648 // Null check; get casted pointer; set region slot 3 3649 Node* null_ctl = top(); 3650 Node* not_null_obj = nullptr; 3651 if (null_free) { 3652 assert(safe_for_replace, "must be"); 3653 not_null_obj = null_check(obj); 3654 } else { 3655 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3656 } 3657 3658 // If not_null_obj is dead, only null-path is taken 3659 if (stopped()) { // Doing instance-of on a null? 3660 set_control(null_ctl); 3661 if (toop->is_inlinetypeptr()) { 3662 return InlineTypeNode::make_null(_gvn, toop->inline_klass()); 3663 } 3664 return null(); 3665 } 3666 region->init_req(_null_path, null_ctl); 3667 phi ->init_req(_null_path, null()); // Set null path value 3668 if (null_ctl == top()) { 3669 // Do this eagerly, so that pattern matches like is_diamond_phi 3670 // will work even during parsing. 3671 assert(_null_path == PATH_LIMIT-1, "delete last"); 3672 region->del_req(_null_path); 3673 phi ->del_req(_null_path); 3674 } 3675 3676 Node* cast_obj = nullptr; 3677 if (improved_klass_ptr_type->klass_is_exact()) { 3678 // The following optimization tries to statically cast the speculative type of the object 3679 // (for example obtained during profiling) to the type of the superklass and then do a 3680 // dynamic check that the type of the object is what we expect. To work correctly 3681 // for checkcast and aastore the type of superklass should be exact. 3682 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3683 // We may not have profiling here or it may not help us. If we have 3684 // a speculative type use it to perform an exact cast. 3685 ciKlass* spec_obj_type = obj_type->speculative_type(); 3686 if (spec_obj_type != nullptr || data != nullptr) { 3687 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace); 3688 if (cast_obj != nullptr) { 3689 if (failure_control != nullptr) // failure is now impossible 3690 (*failure_control) = top(); 3691 // adjust the type of the phi to the exact klass: 3692 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3693 } 3694 } 3695 } 3696 3697 if (cast_obj == nullptr) { 3698 // Generate the subtype check 3699 Node* improved_superklass = superklass; 3700 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) { 3701 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up. 3702 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top. 3703 // Additionally, the benefit would only be minor in non-constant cases. 3704 improved_superklass = makecon(improved_klass_ptr_type); 3705 } 3706 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass); 3707 // Plug in success path into the merge 3708 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3709 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3710 if (failure_control == nullptr) { 3711 if (not_subtype_ctrl != top()) { // If failure is possible 3712 PreserveJVMState pjvms(this); 3713 set_control(not_subtype_ctrl); 3714 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3715 Deoptimization::DeoptReason reason = is_aastore ? 3716 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3717 builtin_throw(reason); 3718 } 3719 } else { 3720 (*failure_control) = not_subtype_ctrl; 3721 } 3722 } 3723 3724 region->init_req(_obj_path, control()); 3725 phi ->init_req(_obj_path, cast_obj); 3726 3727 // A merge of null or Casted-NotNull obj 3728 Node* res = _gvn.transform(phi); 3729 3730 // Note I do NOT always 'replace_in_map(obj,result)' here. 3731 // if( tk->klass()->can_be_primary_super() ) 3732 // This means that if I successfully store an Object into an array-of-String 3733 // I 'forget' that the Object is really now known to be a String. I have to 3734 // do this because we don't have true union types for interfaces - if I store 3735 // a Baz into an array-of-Interface and then tell the optimizer it's an 3736 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3737 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3738 // replace_in_map( obj, res ); 3739 3740 // Return final merged results 3741 set_control( _gvn.transform(region) ); 3742 record_for_igvn(region); 3743 3744 bool not_inline = !toop->can_be_inline_type(); 3745 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array()); 3746 if (EnableValhalla && (not_inline || not_flat_in_array)) { 3747 // Check if obj has been loaded from an array 3748 obj = obj->isa_DecodeN() ? obj->in(1) : obj; 3749 Node* array = nullptr; 3750 if (obj->isa_Load()) { 3751 Node* address = obj->in(MemNode::Address); 3752 if (address->isa_AddP()) { 3753 array = address->as_AddP()->in(AddPNode::Base); 3754 } 3755 } else if (obj->is_Phi()) { 3756 Node* region = obj->in(0); 3757 // TODO make this more robust (see JDK-8231346) 3758 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) { 3759 IfNode* iff = region->in(2)->in(0)->isa_If(); 3760 if (iff != nullptr) { 3761 iff->is_flat_array_check(&_gvn, &array); 3762 } 3763 } 3764 } 3765 if (array != nullptr) { 3766 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr(); 3767 if (ary_t != nullptr) { 3768 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) { 3769 // Casting array element to a non-inline-type, mark array as not null-free. 3770 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free())); 3771 replace_in_map(array, cast); 3772 array = cast; 3773 } 3774 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) { 3775 // Casting array element to a non-flat-in-array type, mark array as not flat. 3776 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat())); 3777 replace_in_map(array, cast); 3778 array = cast; 3779 } 3780 } 3781 } 3782 } 3783 3784 if (!stopped() && !res->is_InlineType()) { 3785 res = record_profiled_receiver_for_speculation(res); 3786 if (toop->is_inlinetypeptr() && !maybe_larval) { 3787 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass()); 3788 res = vt; 3789 if (safe_for_replace) { 3790 replace_in_map(obj, vt); 3791 replace_in_map(not_null_obj, vt); 3792 replace_in_map(res, vt); 3793 } 3794 } 3795 } 3796 return res; 3797 } 3798 3799 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) { 3800 // Load markword 3801 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3802 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3803 if (check_lock && !UseCompactObjectHeaders) { 3804 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword. 3805 // Check if obj is locked 3806 Node* locked_bit = MakeConX(markWord::unlocked_value); 3807 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark)); 3808 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0))); 3809 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3810 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN); 3811 _gvn.transform(iff); 3812 Node* locked_region = new RegionNode(3); 3813 Node* mark_phi = new PhiNode(locked_region, TypeX_X); 3814 3815 // Unlocked: Use bits from mark word 3816 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff))); 3817 mark_phi->init_req(1, mark); 3818 3819 // Locked: Load prototype header from klass 3820 set_control(_gvn.transform(new IfFalseNode(iff))); 3821 // Make loads control dependent to make sure they are only executed if array is locked 3822 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 3823 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 3824 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset())); 3825 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered)); 3826 3827 locked_region->init_req(2, control()); 3828 mark_phi->init_req(2, proto); 3829 set_control(_gvn.transform(locked_region)); 3830 record_for_igvn(locked_region); 3831 3832 mark = mark_phi; 3833 } 3834 3835 // Now check if mark word bits are set 3836 Node* mask = MakeConX(mask_val); 3837 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask)); 3838 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN 3839 Node* cmp = _gvn.transform(new CmpXNode(masked, mask)); 3840 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne)); 3841 } 3842 3843 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) { 3844 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false); 3845 } 3846 3847 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) { 3848 // We can't use immutable memory here because the mark word is mutable. 3849 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the 3850 // check is moved out of loops (mainly to enable loop unswitching). 3851 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass)); 3852 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN 3853 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne)); 3854 } 3855 3856 Node* GraphKit::null_free_array_test(Node* array, bool null_free) { 3857 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free); 3858 } 3859 3860 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) { 3861 assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat"); 3862 3863 // TODO 8350865 Add a stress flag to always access atomic if layout exists? 3864 if (!vk->has_non_atomic_layout()) { 3865 return intcon(1); // Always atomic 3866 } else if (!vk->has_atomic_layout()) { 3867 return intcon(0); // Never atomic 3868 } 3869 3870 Node* array_klass = load_object_klass(array); 3871 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset()); 3872 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset); 3873 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered); 3874 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::ATOMIC_FLAT))); 3875 return _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3876 } 3877 3878 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null 3879 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) { 3880 RegionNode* region = new RegionNode(3); 3881 Node* null_ctl = top(); 3882 null_check_oop(val, &null_ctl); 3883 if (null_ctl != top()) { 3884 PreserveJVMState pjvms(this); 3885 set_control(null_ctl); 3886 { 3887 // Deoptimize if null-free array 3888 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX); 3889 inc_sp(nargs); 3890 uncommon_trap(Deoptimization::Reason_null_check, 3891 Deoptimization::Action_none); 3892 } 3893 region->init_req(1, control()); 3894 } 3895 region->init_req(2, control()); 3896 set_control(_gvn.transform(region)); 3897 record_for_igvn(region); 3898 if (_gvn.type(val) == TypePtr::NULL_PTR) { 3899 // Since we were just successfully storing null, the array can't be null free. 3900 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 3901 ary_t = ary_t->cast_to_not_null_free(); 3902 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t)); 3903 if (safe_for_replace) { 3904 replace_in_map(ary, cast); 3905 } 3906 ary = cast; 3907 } 3908 return ary; 3909 } 3910 3911 //------------------------------next_monitor----------------------------------- 3912 // What number should be given to the next monitor? 3913 int GraphKit::next_monitor() { 3914 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3915 int next = current + C->sync_stack_slots(); 3916 // Keep the toplevel high water mark current: 3917 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3918 return current; 3919 } 3920 3921 //------------------------------insert_mem_bar--------------------------------- 3922 // Memory barrier to avoid floating things around 3923 // The membar serves as a pinch point between both control and all memory slices. 3924 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3925 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3926 mb->init_req(TypeFunc::Control, control()); 3927 mb->init_req(TypeFunc::Memory, reset_memory()); 3928 Node* membar = _gvn.transform(mb); 3929 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3930 set_all_memory_call(membar); 3931 return membar; 3932 } 3933 3934 //-------------------------insert_mem_bar_volatile---------------------------- 3935 // Memory barrier to avoid floating things around 3936 // The membar serves as a pinch point between both control and memory(alias_idx). 3937 // If you want to make a pinch point on all memory slices, do not use this 3938 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3939 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3940 // When Parse::do_put_xxx updates a volatile field, it appends a series 3941 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3942 // The first membar is on the same memory slice as the field store opcode. 3943 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3944 // All the other membars (for other volatile slices, including AliasIdxBot, 3945 // which stands for all unknown volatile slices) are control-dependent 3946 // on the first membar. This prevents later volatile loads or stores 3947 // from sliding up past the just-emitted store. 3948 3949 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3950 mb->set_req(TypeFunc::Control,control()); 3951 if (alias_idx == Compile::AliasIdxBot) { 3952 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3953 } else { 3954 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3955 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3956 } 3957 Node* membar = _gvn.transform(mb); 3958 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3959 if (alias_idx == Compile::AliasIdxBot) { 3960 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3961 } else { 3962 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3963 } 3964 return membar; 3965 } 3966 3967 //------------------------------shared_lock------------------------------------ 3968 // Emit locking code. 3969 FastLockNode* GraphKit::shared_lock(Node* obj) { 3970 // bci is either a monitorenter bc or InvocationEntryBci 3971 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3972 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3973 3974 if (stopped()) // Dead monitor? 3975 return nullptr; 3976 3977 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3978 3979 // Box the stack location 3980 Node* box = new BoxLockNode(next_monitor()); 3981 // Check for bailout after new BoxLockNode 3982 if (failing()) { return nullptr; } 3983 box = _gvn.transform(box); 3984 Node* mem = reset_memory(); 3985 3986 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock(); 3987 3988 // Add monitor to debug info for the slow path. If we block inside the 3989 // slow path and de-opt, we need the monitor hanging around 3990 map()->push_monitor( flock ); 3991 3992 const TypeFunc *tf = LockNode::lock_type(); 3993 LockNode *lock = new LockNode(C, tf); 3994 3995 lock->init_req( TypeFunc::Control, control() ); 3996 lock->init_req( TypeFunc::Memory , mem ); 3997 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3998 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3999 lock->init_req( TypeFunc::ReturnAdr, top() ); 4000 4001 lock->init_req(TypeFunc::Parms + 0, obj); 4002 lock->init_req(TypeFunc::Parms + 1, box); 4003 lock->init_req(TypeFunc::Parms + 2, flock); 4004 add_safepoint_edges(lock); 4005 4006 lock = _gvn.transform( lock )->as_Lock(); 4007 4008 // lock has no side-effects, sets few values 4009 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 4010 4011 insert_mem_bar(Op_MemBarAcquireLock); 4012 4013 // Add this to the worklist so that the lock can be eliminated 4014 record_for_igvn(lock); 4015 4016 #ifndef PRODUCT 4017 if (PrintLockStatistics) { 4018 // Update the counter for this lock. Don't bother using an atomic 4019 // operation since we don't require absolute accuracy. 4020 lock->create_lock_counter(map()->jvms()); 4021 increment_counter(lock->counter()->addr()); 4022 } 4023 #endif 4024 4025 return flock; 4026 } 4027 4028 4029 //------------------------------shared_unlock---------------------------------- 4030 // Emit unlocking code. 4031 void GraphKit::shared_unlock(Node* box, Node* obj) { 4032 // bci is either a monitorenter bc or InvocationEntryBci 4033 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 4034 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 4035 4036 if (stopped()) { // Dead monitor? 4037 map()->pop_monitor(); // Kill monitor from debug info 4038 return; 4039 } 4040 assert(!obj->is_InlineType(), "should not unlock on inline type"); 4041 4042 // Memory barrier to avoid floating things down past the locked region 4043 insert_mem_bar(Op_MemBarReleaseLock); 4044 4045 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 4046 UnlockNode *unlock = new UnlockNode(C, tf); 4047 #ifdef ASSERT 4048 unlock->set_dbg_jvms(sync_jvms()); 4049 #endif 4050 uint raw_idx = Compile::AliasIdxRaw; 4051 unlock->init_req( TypeFunc::Control, control() ); 4052 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 4053 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 4054 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 4055 unlock->init_req( TypeFunc::ReturnAdr, top() ); 4056 4057 unlock->init_req(TypeFunc::Parms + 0, obj); 4058 unlock->init_req(TypeFunc::Parms + 1, box); 4059 unlock = _gvn.transform(unlock)->as_Unlock(); 4060 4061 Node* mem = reset_memory(); 4062 4063 // unlock has no side-effects, sets few values 4064 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 4065 4066 // Kill monitor from debug info 4067 map()->pop_monitor( ); 4068 } 4069 4070 //-------------------------------get_layout_helper----------------------------- 4071 // If the given klass is a constant or known to be an array, 4072 // fetch the constant layout helper value into constant_value 4073 // and return null. Otherwise, load the non-constant 4074 // layout helper value, and return the node which represents it. 4075 // This two-faced routine is useful because allocation sites 4076 // almost always feature constant types. 4077 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 4078 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr(); 4079 if (!StressReflectiveCode && klass_t != nullptr) { 4080 bool xklass = klass_t->klass_is_exact(); 4081 bool can_be_flat = false; 4082 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr(); 4083 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) { 4084 // Don't constant fold if the runtime type might be a flat array but the static type is not. 4085 const TypeOopPtr* elem = ary_type->elem()->make_oopptr(); 4086 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array()); 4087 } 4088 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) { 4089 jint lhelper; 4090 if (klass_t->is_flat()) { 4091 lhelper = ary_type->flat_layout_helper(); 4092 } else if (klass_t->isa_aryklassptr()) { 4093 BasicType elem = ary_type->elem()->array_element_basic_type(); 4094 if (is_reference_type(elem, true)) { 4095 elem = T_OBJECT; 4096 } 4097 lhelper = Klass::array_layout_helper(elem); 4098 } else { 4099 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper(); 4100 } 4101 if (lhelper != Klass::_lh_neutral_value) { 4102 constant_value = lhelper; 4103 return (Node*) nullptr; 4104 } 4105 } 4106 } 4107 constant_value = Klass::_lh_neutral_value; // put in a known value 4108 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 4109 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered); 4110 } 4111 4112 // We just put in an allocate/initialize with a big raw-memory effect. 4113 // Hook selected additional alias categories on the initialization. 4114 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 4115 MergeMemNode* init_in_merge, 4116 Node* init_out_raw) { 4117 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 4118 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 4119 4120 Node* prevmem = kit.memory(alias_idx); 4121 init_in_merge->set_memory_at(alias_idx, prevmem); 4122 if (init_out_raw != nullptr) { 4123 kit.set_memory(init_out_raw, alias_idx); 4124 } 4125 } 4126 4127 //---------------------------set_output_for_allocation------------------------- 4128 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 4129 const TypeOopPtr* oop_type, 4130 bool deoptimize_on_exception) { 4131 int rawidx = Compile::AliasIdxRaw; 4132 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 4133 add_safepoint_edges(alloc); 4134 Node* allocx = _gvn.transform(alloc); 4135 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 4136 // create memory projection for i_o 4137 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 4138 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 4139 4140 // create a memory projection as for the normal control path 4141 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 4142 set_memory(malloc, rawidx); 4143 4144 // a normal slow-call doesn't change i_o, but an allocation does 4145 // we create a separate i_o projection for the normal control path 4146 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 4147 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 4148 4149 // put in an initialization barrier 4150 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 4151 rawoop)->as_Initialize(); 4152 assert(alloc->initialization() == init, "2-way macro link must work"); 4153 assert(init ->allocation() == alloc, "2-way macro link must work"); 4154 { 4155 // Extract memory strands which may participate in the new object's 4156 // initialization, and source them from the new InitializeNode. 4157 // This will allow us to observe initializations when they occur, 4158 // and link them properly (as a group) to the InitializeNode. 4159 assert(init->in(InitializeNode::Memory) == malloc, ""); 4160 MergeMemNode* minit_in = MergeMemNode::make(malloc); 4161 init->set_req(InitializeNode::Memory, minit_in); 4162 record_for_igvn(minit_in); // fold it up later, if possible 4163 _gvn.set_type(minit_in, Type::MEMORY); 4164 Node* minit_out = memory(rawidx); 4165 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 4166 // Add an edge in the MergeMem for the header fields so an access 4167 // to one of those has correct memory state 4168 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 4169 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 4170 if (oop_type->isa_aryptr()) { 4171 const TypeAryPtr* arytype = oop_type->is_aryptr(); 4172 if (arytype->is_flat()) { 4173 // Initially all flat array accesses share a single slice 4174 // but that changes after parsing. Prepare the memory graph so 4175 // it can optimize flat array accesses properly once they 4176 // don't share a single slice. 4177 assert(C->flat_accesses_share_alias(), "should be set at parse time"); 4178 C->set_flat_accesses_share_alias(false); 4179 ciInlineKlass* vk = arytype->elem()->inline_klass(); 4180 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) { 4181 ciField* field = vk->nonstatic_field_at(i); 4182 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize) 4183 continue; // do not bother to track really large numbers of fields 4184 int off_in_vt = field->offset_in_bytes() - vk->payload_offset(); 4185 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot); 4186 int fieldidx = C->get_alias_index(adr_type, true); 4187 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node 4188 // can result in per flat array field Phis to be created which confuses the logic of 4189 // Compile::adjust_flat_array_access_aliases(). 4190 hook_memory_on_init(*this, fieldidx, minit_in, nullptr); 4191 } 4192 C->set_flat_accesses_share_alias(true); 4193 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out); 4194 } else { 4195 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 4196 int elemidx = C->get_alias_index(telemref); 4197 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 4198 } 4199 } else if (oop_type->isa_instptr()) { 4200 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word 4201 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass(); 4202 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 4203 ciField* field = ik->nonstatic_field_at(i); 4204 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize) 4205 continue; // do not bother to track really large numbers of fields 4206 // Find (or create) the alias category for this field: 4207 int fieldidx = C->alias_type(field)->index(); 4208 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 4209 } 4210 } 4211 } 4212 4213 // Cast raw oop to the real thing... 4214 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 4215 javaoop = _gvn.transform(javaoop); 4216 C->set_recent_alloc(control(), javaoop); 4217 assert(just_allocated_object(control()) == javaoop, "just allocated"); 4218 4219 #ifdef ASSERT 4220 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 4221 assert(AllocateNode::Ideal_allocation(rawoop) == alloc, 4222 "Ideal_allocation works"); 4223 assert(AllocateNode::Ideal_allocation(javaoop) == alloc, 4224 "Ideal_allocation works"); 4225 if (alloc->is_AllocateArray()) { 4226 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(), 4227 "Ideal_allocation works"); 4228 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(), 4229 "Ideal_allocation works"); 4230 } else { 4231 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 4232 } 4233 } 4234 #endif //ASSERT 4235 4236 return javaoop; 4237 } 4238 4239 //---------------------------new_instance-------------------------------------- 4240 // This routine takes a klass_node which may be constant (for a static type) 4241 // or may be non-constant (for reflective code). It will work equally well 4242 // for either, and the graph will fold nicely if the optimizer later reduces 4243 // the type to a constant. 4244 // The optional arguments are for specialized use by intrinsics: 4245 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 4246 // - If 'return_size_val', report the total object size to the caller. 4247 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 4248 Node* GraphKit::new_instance(Node* klass_node, 4249 Node* extra_slow_test, 4250 Node* *return_size_val, 4251 bool deoptimize_on_exception, 4252 InlineTypeNode* inline_type_node) { 4253 // Compute size in doublewords 4254 // The size is always an integral number of doublewords, represented 4255 // as a positive bytewise size stored in the klass's layout_helper. 4256 // The layout_helper also encodes (in a low bit) the need for a slow path. 4257 jint layout_con = Klass::_lh_neutral_value; 4258 Node* layout_val = get_layout_helper(klass_node, layout_con); 4259 bool layout_is_con = (layout_val == nullptr); 4260 4261 if (extra_slow_test == nullptr) extra_slow_test = intcon(0); 4262 // Generate the initial go-slow test. It's either ALWAYS (return a 4263 // Node for 1) or NEVER (return a null) or perhaps (in the reflective 4264 // case) a computed value derived from the layout_helper. 4265 Node* initial_slow_test = nullptr; 4266 if (layout_is_con) { 4267 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4268 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 4269 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 4270 } else { // reflective case 4271 // This reflective path is used by Unsafe.allocateInstance. 4272 // (It may be stress-tested by specifying StressReflectiveCode.) 4273 // Basically, we want to get into the VM is there's an illegal argument. 4274 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 4275 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 4276 if (extra_slow_test != intcon(0)) { 4277 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 4278 } 4279 // (Macro-expander will further convert this to a Bool, if necessary.) 4280 } 4281 4282 // Find the size in bytes. This is easy; it's the layout_helper. 4283 // The size value must be valid even if the slow path is taken. 4284 Node* size = nullptr; 4285 if (layout_is_con) { 4286 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 4287 } else { // reflective case 4288 // This reflective path is used by clone and Unsafe.allocateInstance. 4289 size = ConvI2X(layout_val); 4290 4291 // Clear the low bits to extract layout_helper_size_in_bytes: 4292 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 4293 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 4294 size = _gvn.transform( new AndXNode(size, mask) ); 4295 } 4296 if (return_size_val != nullptr) { 4297 (*return_size_val) = size; 4298 } 4299 4300 // This is a precise notnull oop of the klass. 4301 // (Actually, it need not be precise if this is a reflective allocation.) 4302 // It's what we cast the result to. 4303 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 4304 if (!tklass) tklass = TypeInstKlassPtr::OBJECT; 4305 const TypeOopPtr* oop_type = tklass->as_instance_type(); 4306 4307 // Now generate allocation code 4308 4309 // The entire memory state is needed for slow path of the allocation 4310 // since GC and deoptimization can happen. 4311 Node *mem = reset_memory(); 4312 set_all_memory(mem); // Create new memory state 4313 4314 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 4315 control(), mem, i_o(), 4316 size, klass_node, 4317 initial_slow_test, inline_type_node); 4318 4319 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 4320 } 4321 4322 //-------------------------------new_array------------------------------------- 4323 // helper for newarray and anewarray 4324 // The 'length' parameter is (obviously) the length of the array. 4325 // The optional arguments are for specialized use by intrinsics: 4326 // - If 'return_size_val', report the non-padded array size (sum of header size 4327 // and array body) to the caller. 4328 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 4329 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 4330 Node* length, // number of array elements 4331 int nargs, // number of arguments to push back for uncommon trap 4332 Node* *return_size_val, 4333 bool deoptimize_on_exception, 4334 Node* init_val) { 4335 jint layout_con = Klass::_lh_neutral_value; 4336 Node* layout_val = get_layout_helper(klass_node, layout_con); 4337 bool layout_is_con = (layout_val == nullptr); 4338 4339 if (!layout_is_con && !StressReflectiveCode && 4340 !too_many_traps(Deoptimization::Reason_class_check)) { 4341 // This is a reflective array creation site. 4342 // Optimistically assume that it is a subtype of Object[], 4343 // so that we can fold up all the address arithmetic. 4344 layout_con = Klass::array_layout_helper(T_OBJECT); 4345 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 4346 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 4347 { BuildCutout unless(this, bol_lh, PROB_MAX); 4348 inc_sp(nargs); 4349 uncommon_trap(Deoptimization::Reason_class_check, 4350 Deoptimization::Action_maybe_recompile); 4351 } 4352 layout_val = nullptr; 4353 layout_is_con = true; 4354 } 4355 4356 // Generate the initial go-slow test. Make sure we do not overflow 4357 // if length is huge (near 2Gig) or negative! We do not need 4358 // exact double-words here, just a close approximation of needed 4359 // double-words. We can't add any offset or rounding bits, lest we 4360 // take a size -1 of bytes and make it positive. Use an unsigned 4361 // compare, so negative sizes look hugely positive. 4362 int fast_size_limit = FastAllocateSizeLimit; 4363 if (layout_is_con) { 4364 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4365 // Increase the size limit if we have exact knowledge of array type. 4366 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 4367 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0); 4368 } 4369 4370 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 4371 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 4372 4373 // --- Size Computation --- 4374 // array_size = round_to_heap(array_header + (length << elem_shift)); 4375 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 4376 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 4377 // The rounding mask is strength-reduced, if possible. 4378 int round_mask = MinObjAlignmentInBytes - 1; 4379 Node* header_size = nullptr; 4380 // (T_BYTE has the weakest alignment and size restrictions...) 4381 if (layout_is_con) { 4382 int hsize = Klass::layout_helper_header_size(layout_con); 4383 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4384 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con); 4385 if ((round_mask & ~right_n_bits(eshift)) == 0) 4386 round_mask = 0; // strength-reduce it if it goes away completely 4387 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 4388 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 4389 assert(header_size_min <= hsize, "generic minimum is smallest"); 4390 header_size = intcon(hsize); 4391 } else { 4392 Node* hss = intcon(Klass::_lh_header_size_shift); 4393 Node* hsm = intcon(Klass::_lh_header_size_mask); 4394 header_size = _gvn.transform(new URShiftINode(layout_val, hss)); 4395 header_size = _gvn.transform(new AndINode(header_size, hsm)); 4396 } 4397 4398 Node* elem_shift = nullptr; 4399 if (layout_is_con) { 4400 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4401 if (eshift != 0) 4402 elem_shift = intcon(eshift); 4403 } else { 4404 // There is no need to mask or shift this value. 4405 // The semantics of LShiftINode include an implicit mask to 0x1F. 4406 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 4407 elem_shift = layout_val; 4408 } 4409 4410 // Transition to native address size for all offset calculations: 4411 Node* lengthx = ConvI2X(length); 4412 Node* headerx = ConvI2X(header_size); 4413 #ifdef _LP64 4414 { const TypeInt* tilen = _gvn.find_int_type(length); 4415 if (tilen != nullptr && tilen->_lo < 0) { 4416 // Add a manual constraint to a positive range. Cf. array_element_address. 4417 jint size_max = fast_size_limit; 4418 if (size_max > tilen->_hi && tilen->_hi >= 0) { 4419 size_max = tilen->_hi; 4420 } 4421 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 4422 4423 // Only do a narrow I2L conversion if the range check passed. 4424 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 4425 _gvn.transform(iff); 4426 RegionNode* region = new RegionNode(3); 4427 _gvn.set_type(region, Type::CONTROL); 4428 lengthx = new PhiNode(region, TypeLong::LONG); 4429 _gvn.set_type(lengthx, TypeLong::LONG); 4430 4431 // Range check passed. Use ConvI2L node with narrow type. 4432 Node* passed = IfFalse(iff); 4433 region->init_req(1, passed); 4434 // Make I2L conversion control dependent to prevent it from 4435 // floating above the range check during loop optimizations. 4436 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 4437 4438 // Range check failed. Use ConvI2L with wide type because length may be invalid. 4439 region->init_req(2, IfTrue(iff)); 4440 lengthx->init_req(2, ConvI2X(length)); 4441 4442 set_control(region); 4443 record_for_igvn(region); 4444 record_for_igvn(lengthx); 4445 } 4446 } 4447 #endif 4448 4449 // Combine header size and body size for the array copy part, then align (if 4450 // necessary) for the allocation part. This computation cannot overflow, 4451 // because it is used only in two places, one where the length is sharply 4452 // limited, and the other after a successful allocation. 4453 Node* abody = lengthx; 4454 if (elem_shift != nullptr) { 4455 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift)); 4456 } 4457 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody)); 4458 4459 if (return_size_val != nullptr) { 4460 // This is the size 4461 (*return_size_val) = non_rounded_size; 4462 } 4463 4464 Node* size = non_rounded_size; 4465 if (round_mask != 0) { 4466 Node* mask1 = MakeConX(round_mask); 4467 size = _gvn.transform(new AddXNode(size, mask1)); 4468 Node* mask2 = MakeConX(~round_mask); 4469 size = _gvn.transform(new AndXNode(size, mask2)); 4470 } 4471 // else if round_mask == 0, the size computation is self-rounding 4472 4473 // Now generate allocation code 4474 4475 // The entire memory state is needed for slow path of the allocation 4476 // since GC and deoptimization can happen. 4477 Node *mem = reset_memory(); 4478 set_all_memory(mem); // Create new memory state 4479 4480 if (initial_slow_test->is_Bool()) { 4481 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4482 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4483 } 4484 4485 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr(); 4486 const TypeOopPtr* ary_type = ary_klass->as_instance_type(); 4487 4488 Node* raw_init_value = nullptr; 4489 if (init_val != nullptr) { 4490 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays 4491 if (ary_type->is_flat()) { 4492 initial_slow_test = intcon(1); 4493 } 4494 4495 if (UseCompressedOops) { 4496 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops 4497 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop())); 4498 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val)); 4499 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32))); 4500 raw_init_value = _gvn.transform(new OrLNode(lower, upper)); 4501 } else { 4502 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val)); 4503 } 4504 } 4505 4506 Node* valid_length_test = _gvn.intcon(1); 4507 if (ary_type->isa_aryptr()) { 4508 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type(); 4509 jint max = TypeAryPtr::max_array_length(bt); 4510 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max))); 4511 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le)); 4512 } 4513 4514 // Create the AllocateArrayNode and its result projections 4515 AllocateArrayNode* alloc 4516 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4517 control(), mem, i_o(), 4518 size, klass_node, 4519 initial_slow_test, 4520 length, valid_length_test, 4521 init_val, raw_init_value); 4522 // Cast to correct type. Note that the klass_node may be constant or not, 4523 // and in the latter case the actual array type will be inexact also. 4524 // (This happens via a non-constant argument to inline_native_newArray.) 4525 // In any case, the value of klass_node provides the desired array type. 4526 const TypeInt* length_type = _gvn.find_int_type(length); 4527 if (ary_type->isa_aryptr() && length_type != nullptr) { 4528 // Try to get a better type than POS for the size 4529 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4530 } 4531 4532 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4533 4534 array_ideal_length(alloc, ary_type, true); 4535 return javaoop; 4536 } 4537 4538 // The following "Ideal_foo" functions are placed here because they recognize 4539 // the graph shapes created by the functions immediately above. 4540 4541 //---------------------------Ideal_allocation---------------------------------- 4542 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4543 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) { 4544 if (ptr == nullptr) { // reduce dumb test in callers 4545 return nullptr; 4546 } 4547 4548 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4549 ptr = bs->step_over_gc_barrier(ptr); 4550 4551 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4552 ptr = ptr->in(1); 4553 if (ptr == nullptr) return nullptr; 4554 } 4555 // Return null for allocations with several casts: 4556 // j.l.reflect.Array.newInstance(jobject, jint) 4557 // Object.clone() 4558 // to keep more precise type from last cast. 4559 if (ptr->is_Proj()) { 4560 Node* allo = ptr->in(0); 4561 if (allo != nullptr && allo->is_Allocate()) { 4562 return allo->as_Allocate(); 4563 } 4564 } 4565 // Report failure to match. 4566 return nullptr; 4567 } 4568 4569 // Fancy version which also strips off an offset (and reports it to caller). 4570 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase, 4571 intptr_t& offset) { 4572 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4573 if (base == nullptr) return nullptr; 4574 return Ideal_allocation(base); 4575 } 4576 4577 // Trace Initialize <- Proj[Parm] <- Allocate 4578 AllocateNode* InitializeNode::allocation() { 4579 Node* rawoop = in(InitializeNode::RawAddress); 4580 if (rawoop->is_Proj()) { 4581 Node* alloc = rawoop->in(0); 4582 if (alloc->is_Allocate()) { 4583 return alloc->as_Allocate(); 4584 } 4585 } 4586 return nullptr; 4587 } 4588 4589 // Trace Allocate -> Proj[Parm] -> Initialize 4590 InitializeNode* AllocateNode::initialization() { 4591 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4592 if (rawoop == nullptr) return nullptr; 4593 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4594 Node* init = rawoop->fast_out(i); 4595 if (init->is_Initialize()) { 4596 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4597 return init->as_Initialize(); 4598 } 4599 } 4600 return nullptr; 4601 } 4602 4603 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path. 4604 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) { 4605 // Too many traps seen? 4606 if (too_many_traps(reason)) { 4607 #ifdef ASSERT 4608 if (TraceLoopPredicate) { 4609 int tc = C->trap_count(reason); 4610 tty->print("too many traps=%s tcount=%d in ", 4611 Deoptimization::trap_reason_name(reason), tc); 4612 method()->print(); // which method has too many predicate traps 4613 tty->cr(); 4614 } 4615 #endif 4616 // We cannot afford to take more traps here, 4617 // do not generate Parse Predicate. 4618 return; 4619 } 4620 4621 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn); 4622 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn)); 4623 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate)); 4624 { 4625 PreserveJVMState pjvms(this); 4626 set_control(if_false); 4627 inc_sp(nargs); 4628 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4629 } 4630 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate)); 4631 set_control(if_true); 4632 } 4633 4634 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All 4635 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate. 4636 void GraphKit::add_parse_predicates(int nargs) { 4637 if (ShortRunningLongLoop) { 4638 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when 4639 // walking up from the loop. 4640 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs); 4641 } 4642 if (UseLoopPredicate) { 4643 add_parse_predicate(Deoptimization::Reason_predicate, nargs); 4644 if (UseProfiledLoopPredicate) { 4645 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs); 4646 } 4647 } 4648 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs); 4649 // Loop Limit Check Predicate should be near the loop. 4650 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs); 4651 } 4652 4653 void GraphKit::sync_kit(IdealKit& ideal) { 4654 set_all_memory(ideal.merged_memory()); 4655 set_i_o(ideal.i_o()); 4656 set_control(ideal.ctrl()); 4657 } 4658 4659 void GraphKit::final_sync(IdealKit& ideal) { 4660 // Final sync IdealKit and graphKit. 4661 sync_kit(ideal); 4662 } 4663 4664 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4665 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4666 Node* coder = load_String_coder(str, set_ctrl); 4667 // Divide length by 2 if coder is UTF16 4668 return _gvn.transform(new RShiftINode(len, coder)); 4669 } 4670 4671 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4672 int value_offset = java_lang_String::value_offset(); 4673 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4674 false, nullptr, Type::Offset(0)); 4675 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4676 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4677 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true), 4678 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4679 Node* p = basic_plus_adr(str, str, value_offset); 4680 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4681 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4682 return load; 4683 } 4684 4685 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4686 if (!CompactStrings) { 4687 return intcon(java_lang_String::CODER_UTF16); 4688 } 4689 int coder_offset = java_lang_String::coder_offset(); 4690 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4691 false, nullptr, Type::Offset(0)); 4692 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4693 4694 Node* p = basic_plus_adr(str, str, coder_offset); 4695 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4696 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4697 return load; 4698 } 4699 4700 void GraphKit::store_String_value(Node* str, Node* value) { 4701 int value_offset = java_lang_String::value_offset(); 4702 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4703 false, nullptr, Type::Offset(0)); 4704 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4705 4706 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4707 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4708 } 4709 4710 void GraphKit::store_String_coder(Node* str, Node* value) { 4711 int coder_offset = java_lang_String::coder_offset(); 4712 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4713 false, nullptr, Type::Offset(0)); 4714 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4715 4716 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4717 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4718 } 4719 4720 // Capture src and dst memory state with a MergeMemNode 4721 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4722 if (src_type == dst_type) { 4723 // Types are equal, we don't need a MergeMemNode 4724 return memory(src_type); 4725 } 4726 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4727 record_for_igvn(merge); // fold it up later, if possible 4728 int src_idx = C->get_alias_index(src_type); 4729 int dst_idx = C->get_alias_index(dst_type); 4730 merge->set_memory_at(src_idx, memory(src_idx)); 4731 merge->set_memory_at(dst_idx, memory(dst_idx)); 4732 return merge; 4733 } 4734 4735 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4736 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4737 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4738 // If input and output memory types differ, capture both states to preserve 4739 // the dependency between preceding and subsequent loads/stores. 4740 // For example, the following program: 4741 // StoreB 4742 // compress_string 4743 // LoadB 4744 // has this memory graph (use->def): 4745 // LoadB -> compress_string -> CharMem 4746 // ... -> StoreB -> ByteMem 4747 // The intrinsic hides the dependency between LoadB and StoreB, causing 4748 // the load to read from memory not containing the result of the StoreB. 4749 // The correct memory graph should look like this: 4750 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4751 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4752 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4753 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str))); 4754 set_memory(res_mem, TypeAryPtr::BYTES); 4755 return str; 4756 } 4757 4758 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4759 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4760 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4761 // Capture src and dst memory (see comment in 'compress_string'). 4762 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4763 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4764 set_memory(_gvn.transform(str), dst_type); 4765 } 4766 4767 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4768 /** 4769 * int i_char = start; 4770 * for (int i_byte = 0; i_byte < count; i_byte++) { 4771 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4772 * } 4773 */ 4774 add_parse_predicates(); 4775 C->set_has_loops(true); 4776 4777 RegionNode* head = new RegionNode(3); 4778 head->init_req(1, control()); 4779 gvn().set_type(head, Type::CONTROL); 4780 record_for_igvn(head); 4781 4782 Node* i_byte = new PhiNode(head, TypeInt::INT); 4783 i_byte->init_req(1, intcon(0)); 4784 gvn().set_type(i_byte, TypeInt::INT); 4785 record_for_igvn(i_byte); 4786 4787 Node* i_char = new PhiNode(head, TypeInt::INT); 4788 i_char->init_req(1, start); 4789 gvn().set_type(i_char, TypeInt::INT); 4790 record_for_igvn(i_char); 4791 4792 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4793 gvn().set_type(mem, Type::MEMORY); 4794 record_for_igvn(mem); 4795 set_control(head); 4796 set_memory(mem, TypeAryPtr::BYTES); 4797 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true); 4798 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4799 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false, 4800 false, true /* mismatched */); 4801 4802 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4803 head->init_req(2, IfTrue(iff)); 4804 mem->init_req(2, st); 4805 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4806 i_char->init_req(2, AddI(i_char, intcon(2))); 4807 4808 set_control(IfFalse(iff)); 4809 set_memory(st, TypeAryPtr::BYTES); 4810 } 4811 4812 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4813 if (!field->is_constant()) { 4814 return nullptr; // Field not marked as constant. 4815 } 4816 ciInstance* holder = nullptr; 4817 if (!field->is_static()) { 4818 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4819 if (const_oop != nullptr && const_oop->is_instance()) { 4820 holder = const_oop->as_instance(); 4821 } 4822 } 4823 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4824 /*is_unsigned_load=*/false); 4825 if (con_type != nullptr) { 4826 Node* con = makecon(con_type); 4827 if (field->type()->is_inlinetype()) { 4828 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass()); 4829 } else if (con_type->is_inlinetypeptr()) { 4830 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass()); 4831 } 4832 return con; 4833 } 4834 return nullptr; 4835 } 4836 4837 //---------------------------load_mirror_from_klass---------------------------- 4838 // Given a klass oop, load its java mirror (a java.lang.Class oop). 4839 Node* GraphKit::load_mirror_from_klass(Node* klass) { 4840 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 4841 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4842 // mirror = ((OopHandle)mirror)->resolve(); 4843 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 4844 } 4845 4846 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) { 4847 const Type* obj_type = obj->bottom_type(); 4848 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type); 4849 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) { 4850 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part 4851 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type)); 4852 obj = casted_obj; 4853 } 4854 if (sig_type->is_inlinetypeptr()) { 4855 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass()); 4856 } 4857 return obj; 4858 }