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