1 /* 2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciUtilities.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "ci/ciValueKlass.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/c2/barrierSetC2.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/graphKit.hpp" 37 #include "opto/idealKit.hpp" 38 #include "opto/intrinsicnode.hpp" 39 #include "opto/locknode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/narrowptrnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/valuetypenode.hpp" 47 #include "runtime/deoptimization.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 50 //----------------------------GraphKit----------------------------------------- 51 // Main utility constructor. 52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 53 : Phase(Phase::Parser), 54 _env(C->env()), 55 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()), 56 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 57 { 58 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled"); 59 _exceptions = jvms->map()->next_exception(); 60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 61 set_jvms(jvms); 62 #ifdef ASSERT 63 if (_gvn.is_IterGVN() != NULL) { 64 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used"); 65 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 66 _worklist_size = _gvn.C->for_igvn()->size(); 67 } 68 #endif 69 } 70 71 // Private constructor for parser. 72 GraphKit::GraphKit() 73 : Phase(Phase::Parser), 74 _env(C->env()), 75 _gvn(*C->initial_gvn()), 76 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 77 { 78 _exceptions = NULL; 79 set_map(NULL); 80 debug_only(_sp = -99); 81 debug_only(set_bci(-99)); 82 } 83 84 85 86 //---------------------------clean_stack--------------------------------------- 87 // Clear away rubbish from the stack area of the JVM state. 88 // This destroys any arguments that may be waiting on the stack. 89 void GraphKit::clean_stack(int from_sp) { 90 SafePointNode* map = this->map(); 91 JVMState* jvms = this->jvms(); 92 int stk_size = jvms->stk_size(); 93 int stkoff = jvms->stkoff(); 94 Node* top = this->top(); 95 for (int i = from_sp; i < stk_size; i++) { 96 if (map->in(stkoff + i) != top) { 97 map->set_req(stkoff + i, top); 98 } 99 } 100 } 101 102 103 //--------------------------------sync_jvms----------------------------------- 104 // Make sure our current jvms agrees with our parse state. 105 JVMState* GraphKit::sync_jvms() const { 106 JVMState* jvms = this->jvms(); 107 jvms->set_bci(bci()); // Record the new bci in the JVMState 108 jvms->set_sp(sp()); // Record the new sp in the JVMState 109 assert(jvms_in_sync(), "jvms is now in sync"); 110 return jvms; 111 } 112 113 //--------------------------------sync_jvms_for_reexecute--------------------- 114 // Make sure our current jvms agrees with our parse state. This version 115 // uses the reexecute_sp for reexecuting bytecodes. 116 JVMState* GraphKit::sync_jvms_for_reexecute() { 117 JVMState* jvms = this->jvms(); 118 jvms->set_bci(bci()); // Record the new bci in the JVMState 119 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 120 return jvms; 121 } 122 123 #ifdef ASSERT 124 bool GraphKit::jvms_in_sync() const { 125 Parse* parse = is_Parse(); 126 if (parse == NULL) { 127 if (bci() != jvms()->bci()) return false; 128 if (sp() != (int)jvms()->sp()) return false; 129 return true; 130 } 131 if (jvms()->method() != parse->method()) return false; 132 if (jvms()->bci() != parse->bci()) return false; 133 int jvms_sp = jvms()->sp(); 134 if (jvms_sp != parse->sp()) return false; 135 int jvms_depth = jvms()->depth(); 136 if (jvms_depth != parse->depth()) return false; 137 return true; 138 } 139 140 // Local helper checks for special internal merge points 141 // used to accumulate and merge exception states. 142 // They are marked by the region's in(0) edge being the map itself. 143 // Such merge points must never "escape" into the parser at large, 144 // until they have been handed to gvn.transform. 145 static bool is_hidden_merge(Node* reg) { 146 if (reg == NULL) return false; 147 if (reg->is_Phi()) { 148 reg = reg->in(0); 149 if (reg == NULL) return false; 150 } 151 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 152 } 153 154 void GraphKit::verify_map() const { 155 if (map() == NULL) return; // null map is OK 156 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 157 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 158 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 159 } 160 161 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 162 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 163 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 164 } 165 #endif 166 167 //---------------------------stop_and_kill_map--------------------------------- 168 // Set _map to NULL, signalling a stop to further bytecode execution. 169 // First smash the current map's control to a constant, to mark it dead. 170 void GraphKit::stop_and_kill_map() { 171 SafePointNode* dead_map = stop(); 172 if (dead_map != NULL) { 173 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 174 assert(dead_map->is_killed(), "must be so marked"); 175 } 176 } 177 178 179 //--------------------------------stopped-------------------------------------- 180 // Tell if _map is NULL, or control is top. 181 bool GraphKit::stopped() { 182 if (map() == NULL) return true; 183 else if (control() == top()) return true; 184 else return false; 185 } 186 187 188 //-----------------------------has_ex_handler---------------------------------- 189 // Tell if this method or any caller method has exception handlers. 190 bool GraphKit::has_ex_handler() { 191 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 192 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 193 return true; 194 } 195 } 196 return false; 197 } 198 199 //------------------------------save_ex_oop------------------------------------ 200 // Save an exception without blowing stack contents or other JVM state. 201 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 202 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 203 ex_map->add_req(ex_oop); 204 debug_only(verify_exception_state(ex_map)); 205 } 206 207 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 208 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 209 Node* ex_oop = ex_map->in(ex_map->req()-1); 210 if (clear_it) ex_map->del_req(ex_map->req()-1); 211 return ex_oop; 212 } 213 214 //-----------------------------saved_ex_oop------------------------------------ 215 // Recover a saved exception from its map. 216 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 217 return common_saved_ex_oop(ex_map, false); 218 } 219 220 //--------------------------clear_saved_ex_oop--------------------------------- 221 // Erase a previously saved exception from its map. 222 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 223 return common_saved_ex_oop(ex_map, true); 224 } 225 226 #ifdef ASSERT 227 //---------------------------has_saved_ex_oop---------------------------------- 228 // Erase a previously saved exception from its map. 229 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 230 return ex_map->req() == ex_map->jvms()->endoff()+1; 231 } 232 #endif 233 234 //-------------------------make_exception_state-------------------------------- 235 // Turn the current JVM state into an exception state, appending the ex_oop. 236 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 237 sync_jvms(); 238 SafePointNode* ex_map = stop(); // do not manipulate this map any more 239 set_saved_ex_oop(ex_map, ex_oop); 240 return ex_map; 241 } 242 243 244 //--------------------------add_exception_state-------------------------------- 245 // Add an exception to my list of exceptions. 246 void GraphKit::add_exception_state(SafePointNode* ex_map) { 247 if (ex_map == NULL || ex_map->control() == top()) { 248 return; 249 } 250 #ifdef ASSERT 251 verify_exception_state(ex_map); 252 if (has_exceptions()) { 253 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 254 } 255 #endif 256 257 // If there is already an exception of exactly this type, merge with it. 258 // In particular, null-checks and other low-level exceptions common up here. 259 Node* ex_oop = saved_ex_oop(ex_map); 260 const Type* ex_type = _gvn.type(ex_oop); 261 if (ex_oop == top()) { 262 // No action needed. 263 return; 264 } 265 assert(ex_type->isa_instptr(), "exception must be an instance"); 266 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 267 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 268 // We check sp also because call bytecodes can generate exceptions 269 // both before and after arguments are popped! 270 if (ex_type2 == ex_type 271 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 272 combine_exception_states(ex_map, e2); 273 return; 274 } 275 } 276 277 // No pre-existing exception of the same type. Chain it on the list. 278 push_exception_state(ex_map); 279 } 280 281 //-----------------------add_exception_states_from----------------------------- 282 void GraphKit::add_exception_states_from(JVMState* jvms) { 283 SafePointNode* ex_map = jvms->map()->next_exception(); 284 if (ex_map != NULL) { 285 jvms->map()->set_next_exception(NULL); 286 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 287 next_map = ex_map->next_exception(); 288 ex_map->set_next_exception(NULL); 289 add_exception_state(ex_map); 290 } 291 } 292 } 293 294 //-----------------------transfer_exceptions_into_jvms------------------------- 295 JVMState* GraphKit::transfer_exceptions_into_jvms() { 296 if (map() == NULL) { 297 // We need a JVMS to carry the exceptions, but the map has gone away. 298 // Create a scratch JVMS, cloned from any of the exception states... 299 if (has_exceptions()) { 300 _map = _exceptions; 301 _map = clone_map(); 302 _map->set_next_exception(NULL); 303 clear_saved_ex_oop(_map); 304 debug_only(verify_map()); 305 } else { 306 // ...or created from scratch 307 JVMState* jvms = new (C) JVMState(_method, NULL); 308 jvms->set_bci(_bci); 309 jvms->set_sp(_sp); 310 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 311 set_jvms(jvms); 312 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 313 set_all_memory(top()); 314 while (map()->req() < jvms->endoff()) map()->add_req(top()); 315 } 316 // (This is a kludge, in case you didn't notice.) 317 set_control(top()); 318 } 319 JVMState* jvms = sync_jvms(); 320 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 321 jvms->map()->set_next_exception(_exceptions); 322 _exceptions = NULL; // done with this set of exceptions 323 return jvms; 324 } 325 326 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 327 assert(is_hidden_merge(dstphi), "must be a special merge node"); 328 assert(is_hidden_merge(srcphi), "must be a special merge node"); 329 uint limit = srcphi->req(); 330 for (uint i = PhiNode::Input; i < limit; i++) { 331 dstphi->add_req(srcphi->in(i)); 332 } 333 } 334 static inline void add_one_req(Node* dstphi, Node* src) { 335 assert(is_hidden_merge(dstphi), "must be a special merge node"); 336 assert(!is_hidden_merge(src), "must not be a special merge node"); 337 dstphi->add_req(src); 338 } 339 340 //-----------------------combine_exception_states------------------------------ 341 // This helper function combines exception states by building phis on a 342 // specially marked state-merging region. These regions and phis are 343 // untransformed, and can build up gradually. The region is marked by 344 // having a control input of its exception map, rather than NULL. Such 345 // regions do not appear except in this function, and in use_exception_state. 346 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 347 if (failing()) return; // dying anyway... 348 JVMState* ex_jvms = ex_map->_jvms; 349 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 350 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 351 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 352 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 353 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 354 assert(ex_map->req() == phi_map->req(), "matching maps"); 355 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 356 Node* hidden_merge_mark = root(); 357 Node* region = phi_map->control(); 358 MergeMemNode* phi_mem = phi_map->merged_memory(); 359 MergeMemNode* ex_mem = ex_map->merged_memory(); 360 if (region->in(0) != hidden_merge_mark) { 361 // The control input is not (yet) a specially-marked region in phi_map. 362 // Make it so, and build some phis. 363 region = new RegionNode(2); 364 _gvn.set_type(region, Type::CONTROL); 365 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 366 region->init_req(1, phi_map->control()); 367 phi_map->set_control(region); 368 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 369 record_for_igvn(io_phi); 370 _gvn.set_type(io_phi, Type::ABIO); 371 phi_map->set_i_o(io_phi); 372 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 373 Node* m = mms.memory(); 374 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 375 record_for_igvn(m_phi); 376 _gvn.set_type(m_phi, Type::MEMORY); 377 mms.set_memory(m_phi); 378 } 379 } 380 381 // Either or both of phi_map and ex_map might already be converted into phis. 382 Node* ex_control = ex_map->control(); 383 // if there is special marking on ex_map also, we add multiple edges from src 384 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 385 // how wide was the destination phi_map, originally? 386 uint orig_width = region->req(); 387 388 if (add_multiple) { 389 add_n_reqs(region, ex_control); 390 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 391 } else { 392 // ex_map has no merges, so we just add single edges everywhere 393 add_one_req(region, ex_control); 394 add_one_req(phi_map->i_o(), ex_map->i_o()); 395 } 396 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 397 if (mms.is_empty()) { 398 // get a copy of the base memory, and patch some inputs into it 399 const TypePtr* adr_type = mms.adr_type(C); 400 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 401 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 402 mms.set_memory(phi); 403 // Prepare to append interesting stuff onto the newly sliced phi: 404 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 405 } 406 // Append stuff from ex_map: 407 if (add_multiple) { 408 add_n_reqs(mms.memory(), mms.memory2()); 409 } else { 410 add_one_req(mms.memory(), mms.memory2()); 411 } 412 } 413 uint limit = ex_map->req(); 414 for (uint i = TypeFunc::Parms; i < limit; i++) { 415 // Skip everything in the JVMS after tos. (The ex_oop follows.) 416 if (i == tos) i = ex_jvms->monoff(); 417 Node* src = ex_map->in(i); 418 Node* dst = phi_map->in(i); 419 if (src != dst) { 420 PhiNode* phi; 421 if (dst->in(0) != region) { 422 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 423 record_for_igvn(phi); 424 _gvn.set_type(phi, phi->type()); 425 phi_map->set_req(i, dst); 426 // Prepare to append interesting stuff onto the new phi: 427 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 428 } else { 429 assert(dst->is_Phi(), "nobody else uses a hidden region"); 430 phi = dst->as_Phi(); 431 } 432 if (add_multiple && src->in(0) == ex_control) { 433 // Both are phis. 434 add_n_reqs(dst, src); 435 } else { 436 while (dst->req() < region->req()) add_one_req(dst, src); 437 } 438 const Type* srctype = _gvn.type(src); 439 if (phi->type() != srctype) { 440 const Type* dsttype = phi->type()->meet_speculative(srctype); 441 if (phi->type() != dsttype) { 442 phi->set_type(dsttype); 443 _gvn.set_type(phi, dsttype); 444 } 445 } 446 } 447 } 448 phi_map->merge_replaced_nodes_with(ex_map); 449 } 450 451 //--------------------------use_exception_state-------------------------------- 452 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 453 if (failing()) { stop(); return top(); } 454 Node* region = phi_map->control(); 455 Node* hidden_merge_mark = root(); 456 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 457 Node* ex_oop = clear_saved_ex_oop(phi_map); 458 if (region->in(0) == hidden_merge_mark) { 459 // Special marking for internal ex-states. Process the phis now. 460 region->set_req(0, region); // now it's an ordinary region 461 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 462 // Note: Setting the jvms also sets the bci and sp. 463 set_control(_gvn.transform(region)); 464 uint tos = jvms()->stkoff() + sp(); 465 for (uint i = 1; i < tos; i++) { 466 Node* x = phi_map->in(i); 467 if (x->in(0) == region) { 468 assert(x->is_Phi(), "expected a special phi"); 469 phi_map->set_req(i, _gvn.transform(x)); 470 } 471 } 472 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 473 Node* x = mms.memory(); 474 if (x->in(0) == region) { 475 assert(x->is_Phi(), "nobody else uses a hidden region"); 476 mms.set_memory(_gvn.transform(x)); 477 } 478 } 479 if (ex_oop->in(0) == region) { 480 assert(ex_oop->is_Phi(), "expected a special phi"); 481 ex_oop = _gvn.transform(ex_oop); 482 } 483 } else { 484 set_jvms(phi_map->jvms()); 485 } 486 487 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 488 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 489 return ex_oop; 490 } 491 492 //---------------------------------java_bc------------------------------------- 493 Bytecodes::Code GraphKit::java_bc() const { 494 ciMethod* method = this->method(); 495 int bci = this->bci(); 496 if (method != NULL && bci != InvocationEntryBci) 497 return method->java_code_at_bci(bci); 498 else 499 return Bytecodes::_illegal; 500 } 501 502 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 503 bool must_throw) { 504 // if the exception capability is set, then we will generate code 505 // to check the JavaThread.should_post_on_exceptions flag to see 506 // if we actually need to report exception events (for this 507 // thread). If we don't need to report exception events, we will 508 // take the normal fast path provided by add_exception_events. If 509 // exception event reporting is enabled for this thread, we will 510 // take the uncommon_trap in the BuildCutout below. 511 512 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 513 Node* jthread = _gvn.transform(new ThreadLocalNode()); 514 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 515 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 516 517 // Test the should_post_on_exceptions_flag vs. 0 518 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 519 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 520 521 // Branch to slow_path if should_post_on_exceptions_flag was true 522 { BuildCutout unless(this, tst, PROB_MAX); 523 // Do not try anything fancy if we're notifying the VM on every throw. 524 // Cf. case Bytecodes::_athrow in parse2.cpp. 525 uncommon_trap(reason, Deoptimization::Action_none, 526 (ciKlass*)NULL, (char*)NULL, must_throw); 527 } 528 529 } 530 531 //------------------------------builtin_throw---------------------------------- 532 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 533 bool must_throw = true; 534 535 if (env()->jvmti_can_post_on_exceptions()) { 536 // check if we must post exception events, take uncommon trap if so 537 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 538 // here if should_post_on_exceptions is false 539 // continue on with the normal codegen 540 } 541 542 // If this particular condition has not yet happened at this 543 // bytecode, then use the uncommon trap mechanism, and allow for 544 // a future recompilation if several traps occur here. 545 // If the throw is hot, try to use a more complicated inline mechanism 546 // which keeps execution inside the compiled code. 547 bool treat_throw_as_hot = false; 548 ciMethodData* md = method()->method_data(); 549 550 if (ProfileTraps) { 551 if (too_many_traps(reason)) { 552 treat_throw_as_hot = true; 553 } 554 // (If there is no MDO at all, assume it is early in 555 // execution, and that any deopts are part of the 556 // startup transient, and don't need to be remembered.) 557 558 // Also, if there is a local exception handler, treat all throws 559 // as hot if there has been at least one in this method. 560 if (C->trap_count(reason) != 0 561 && method()->method_data()->trap_count(reason) != 0 562 && has_ex_handler()) { 563 treat_throw_as_hot = true; 564 } 565 } 566 567 // If this throw happens frequently, an uncommon trap might cause 568 // a performance pothole. If there is a local exception handler, 569 // and if this particular bytecode appears to be deoptimizing often, 570 // let us handle the throw inline, with a preconstructed instance. 571 // Note: If the deopt count has blown up, the uncommon trap 572 // runtime is going to flush this nmethod, not matter what. 573 if (treat_throw_as_hot 574 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 575 // If the throw is local, we use a pre-existing instance and 576 // punt on the backtrace. This would lead to a missing backtrace 577 // (a repeat of 4292742) if the backtrace object is ever asked 578 // for its backtrace. 579 // Fixing this remaining case of 4292742 requires some flavor of 580 // escape analysis. Leave that for the future. 581 ciInstance* ex_obj = NULL; 582 switch (reason) { 583 case Deoptimization::Reason_null_check: 584 ex_obj = env()->NullPointerException_instance(); 585 break; 586 case Deoptimization::Reason_div0_check: 587 ex_obj = env()->ArithmeticException_instance(); 588 break; 589 case Deoptimization::Reason_range_check: 590 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 591 break; 592 case Deoptimization::Reason_class_check: 593 if (java_bc() == Bytecodes::_aastore) { 594 ex_obj = env()->ArrayStoreException_instance(); 595 } else { 596 ex_obj = env()->ClassCastException_instance(); 597 } 598 break; 599 default: 600 break; 601 } 602 if (failing()) { stop(); return; } // exception allocation might fail 603 if (ex_obj != NULL) { 604 // Cheat with a preallocated exception object. 605 if (C->log() != NULL) 606 C->log()->elem("hot_throw preallocated='1' reason='%s'", 607 Deoptimization::trap_reason_name(reason)); 608 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 609 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 610 611 // Clear the detail message of the preallocated exception object. 612 // Weblogic sometimes mutates the detail message of exceptions 613 // using reflection. 614 int offset = java_lang_Throwable::get_detailMessage_offset(); 615 const TypePtr* adr_typ = ex_con->add_offset(offset); 616 617 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 618 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 619 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 620 621 add_exception_state(make_exception_state(ex_node)); 622 return; 623 } 624 } 625 626 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 627 // It won't be much cheaper than bailing to the interp., since we'll 628 // have to pass up all the debug-info, and the runtime will have to 629 // create the stack trace. 630 631 // Usual case: Bail to interpreter. 632 // Reserve the right to recompile if we haven't seen anything yet. 633 634 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 635 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 636 if (treat_throw_as_hot 637 && (method()->method_data()->trap_recompiled_at(bci(), m) 638 || C->too_many_traps(reason))) { 639 // We cannot afford to take more traps here. Suffer in the interpreter. 640 if (C->log() != NULL) 641 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 642 Deoptimization::trap_reason_name(reason), 643 C->trap_count(reason)); 644 action = Deoptimization::Action_none; 645 } 646 647 // "must_throw" prunes the JVM state to include only the stack, if there 648 // are no local exception handlers. This should cut down on register 649 // allocation time and code size, by drastically reducing the number 650 // of in-edges on the call to the uncommon trap. 651 652 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 653 } 654 655 656 //----------------------------PreserveJVMState--------------------------------- 657 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 658 debug_only(kit->verify_map()); 659 _kit = kit; 660 _map = kit->map(); // preserve the map 661 _sp = kit->sp(); 662 kit->set_map(clone_map ? kit->clone_map() : NULL); 663 #ifdef ASSERT 664 _bci = kit->bci(); 665 Parse* parser = kit->is_Parse(); 666 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 667 _block = block; 668 #endif 669 } 670 PreserveJVMState::~PreserveJVMState() { 671 GraphKit* kit = _kit; 672 #ifdef ASSERT 673 assert(kit->bci() == _bci, "bci must not shift"); 674 Parse* parser = kit->is_Parse(); 675 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 676 assert(block == _block, "block must not shift"); 677 #endif 678 kit->set_map(_map); 679 kit->set_sp(_sp); 680 } 681 682 683 //-----------------------------BuildCutout------------------------------------- 684 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 685 : PreserveJVMState(kit) 686 { 687 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 688 SafePointNode* outer_map = _map; // preserved map is caller's 689 SafePointNode* inner_map = kit->map(); 690 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 691 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 692 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 693 } 694 BuildCutout::~BuildCutout() { 695 GraphKit* kit = _kit; 696 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 697 } 698 699 //---------------------------PreserveReexecuteState---------------------------- 700 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 701 assert(!kit->stopped(), "must call stopped() before"); 702 _kit = kit; 703 _sp = kit->sp(); 704 _reexecute = kit->jvms()->_reexecute; 705 } 706 PreserveReexecuteState::~PreserveReexecuteState() { 707 if (_kit->stopped()) return; 708 _kit->jvms()->_reexecute = _reexecute; 709 _kit->set_sp(_sp); 710 } 711 712 //------------------------------clone_map-------------------------------------- 713 // Implementation of PreserveJVMState 714 // 715 // Only clone_map(...) here. If this function is only used in the 716 // PreserveJVMState class we may want to get rid of this extra 717 // function eventually and do it all there. 718 719 SafePointNode* GraphKit::clone_map() { 720 if (map() == NULL) return NULL; 721 722 // Clone the memory edge first 723 Node* mem = MergeMemNode::make(map()->memory()); 724 gvn().set_type_bottom(mem); 725 726 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 727 JVMState* jvms = this->jvms(); 728 JVMState* clonejvms = jvms->clone_shallow(C); 729 clonemap->set_memory(mem); 730 clonemap->set_jvms(clonejvms); 731 clonejvms->set_map(clonemap); 732 record_for_igvn(clonemap); 733 gvn().set_type_bottom(clonemap); 734 return clonemap; 735 } 736 737 738 //-----------------------------set_map_clone----------------------------------- 739 void GraphKit::set_map_clone(SafePointNode* m) { 740 _map = m; 741 _map = clone_map(); 742 _map->set_next_exception(NULL); 743 debug_only(verify_map()); 744 } 745 746 747 //----------------------------kill_dead_locals--------------------------------- 748 // Detect any locals which are known to be dead, and force them to top. 749 void GraphKit::kill_dead_locals() { 750 // Consult the liveness information for the locals. If any 751 // of them are unused, then they can be replaced by top(). This 752 // should help register allocation time and cut down on the size 753 // of the deoptimization information. 754 755 // This call is made from many of the bytecode handling 756 // subroutines called from the Big Switch in do_one_bytecode. 757 // Every bytecode which might include a slow path is responsible 758 // for killing its dead locals. The more consistent we 759 // are about killing deads, the fewer useless phis will be 760 // constructed for them at various merge points. 761 762 // bci can be -1 (InvocationEntryBci). We return the entry 763 // liveness for the method. 764 765 if (method() == NULL || method()->code_size() == 0) { 766 // We are building a graph for a call to a native method. 767 // All locals are live. 768 return; 769 } 770 771 ResourceMark rm; 772 773 // Consult the liveness information for the locals. If any 774 // of them are unused, then they can be replaced by top(). This 775 // should help register allocation time and cut down on the size 776 // of the deoptimization information. 777 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 778 779 int len = (int)live_locals.size(); 780 assert(len <= jvms()->loc_size(), "too many live locals"); 781 for (int local = 0; local < len; local++) { 782 if (!live_locals.at(local)) { 783 set_local(local, top()); 784 } 785 } 786 } 787 788 #ifdef ASSERT 789 //-------------------------dead_locals_are_killed------------------------------ 790 // Return true if all dead locals are set to top in the map. 791 // Used to assert "clean" debug info at various points. 792 bool GraphKit::dead_locals_are_killed() { 793 if (method() == NULL || method()->code_size() == 0) { 794 // No locals need to be dead, so all is as it should be. 795 return true; 796 } 797 798 // Make sure somebody called kill_dead_locals upstream. 799 ResourceMark rm; 800 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 801 if (jvms->loc_size() == 0) continue; // no locals to consult 802 SafePointNode* map = jvms->map(); 803 ciMethod* method = jvms->method(); 804 int bci = jvms->bci(); 805 if (jvms == this->jvms()) { 806 bci = this->bci(); // it might not yet be synched 807 } 808 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 809 int len = (int)live_locals.size(); 810 if (!live_locals.is_valid() || len == 0) 811 // This method is trivial, or is poisoned by a breakpoint. 812 return true; 813 assert(len == jvms->loc_size(), "live map consistent with locals map"); 814 for (int local = 0; local < len; local++) { 815 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 816 if (PrintMiscellaneous && (Verbose || WizardMode)) { 817 tty->print_cr("Zombie local %d: ", local); 818 jvms->dump(); 819 } 820 return false; 821 } 822 } 823 } 824 return true; 825 } 826 827 #endif //ASSERT 828 829 // Helper function for enforcing certain bytecodes to reexecute if 830 // deoptimization happens 831 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 832 ciMethod* cur_method = jvms->method(); 833 int cur_bci = jvms->bci(); 834 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 835 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 836 return Interpreter::bytecode_should_reexecute(code) || 837 (is_anewarray && (code == Bytecodes::_multianewarray)); 838 // Reexecute _multianewarray bytecode which was replaced with 839 // sequence of [a]newarray. See Parse::do_multianewarray(). 840 // 841 // Note: interpreter should not have it set since this optimization 842 // is limited by dimensions and guarded by flag so in some cases 843 // multianewarray() runtime calls will be generated and 844 // the bytecode should not be reexecutes (stack will not be reset). 845 } else { 846 return false; 847 } 848 } 849 850 // Helper function for adding JVMState and debug information to node 851 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 852 // Add the safepoint edges to the call (or other safepoint). 853 854 // Make sure dead locals are set to top. This 855 // should help register allocation time and cut down on the size 856 // of the deoptimization information. 857 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 858 859 // Walk the inline list to fill in the correct set of JVMState's 860 // Also fill in the associated edges for each JVMState. 861 862 // If the bytecode needs to be reexecuted we need to put 863 // the arguments back on the stack. 864 const bool should_reexecute = jvms()->should_reexecute(); 865 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 866 867 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 868 // undefined if the bci is different. This is normal for Parse but it 869 // should not happen for LibraryCallKit because only one bci is processed. 870 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 871 "in LibraryCallKit the reexecute bit should not change"); 872 873 // If we are guaranteed to throw, we can prune everything but the 874 // input to the current bytecode. 875 bool can_prune_locals = false; 876 uint stack_slots_not_pruned = 0; 877 int inputs = 0, depth = 0; 878 if (must_throw) { 879 assert(method() == youngest_jvms->method(), "sanity"); 880 if (compute_stack_effects(inputs, depth)) { 881 can_prune_locals = true; 882 stack_slots_not_pruned = inputs; 883 } 884 } 885 886 if (env()->should_retain_local_variables()) { 887 // At any safepoint, this method can get breakpointed, which would 888 // then require an immediate deoptimization. 889 can_prune_locals = false; // do not prune locals 890 stack_slots_not_pruned = 0; 891 } 892 893 // do not scribble on the input jvms 894 JVMState* out_jvms = youngest_jvms->clone_deep(C); 895 call->set_jvms(out_jvms); // Start jvms list for call node 896 897 // For a known set of bytecodes, the interpreter should reexecute them if 898 // deoptimization happens. We set the reexecute state for them here 899 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 900 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 901 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 902 } 903 904 // Presize the call: 905 DEBUG_ONLY(uint non_debug_edges = call->req()); 906 call->add_req_batch(top(), youngest_jvms->debug_depth()); 907 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 908 909 // Set up edges so that the call looks like this: 910 // Call [state:] ctl io mem fptr retadr 911 // [parms:] parm0 ... parmN 912 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 913 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 914 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 915 // Note that caller debug info precedes callee debug info. 916 917 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 918 uint debug_ptr = call->req(); 919 920 // Loop over the map input edges associated with jvms, add them 921 // to the call node, & reset all offsets to match call node array. 922 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 923 uint debug_end = debug_ptr; 924 uint debug_start = debug_ptr - in_jvms->debug_size(); 925 debug_ptr = debug_start; // back up the ptr 926 927 uint p = debug_start; // walks forward in [debug_start, debug_end) 928 uint j, k, l; 929 SafePointNode* in_map = in_jvms->map(); 930 out_jvms->set_map(call); 931 932 if (can_prune_locals) { 933 assert(in_jvms->method() == out_jvms->method(), "sanity"); 934 // If the current throw can reach an exception handler in this JVMS, 935 // then we must keep everything live that can reach that handler. 936 // As a quick and dirty approximation, we look for any handlers at all. 937 if (in_jvms->method()->has_exception_handlers()) { 938 can_prune_locals = false; 939 } 940 } 941 942 // Add the Locals 943 k = in_jvms->locoff(); 944 l = in_jvms->loc_size(); 945 out_jvms->set_locoff(p); 946 if (!can_prune_locals) { 947 for (j = 0; j < l; j++) 948 call->set_req(p++, in_map->in(k+j)); 949 } else { 950 p += l; // already set to top above by add_req_batch 951 } 952 953 // Add the Expression Stack 954 k = in_jvms->stkoff(); 955 l = in_jvms->sp(); 956 out_jvms->set_stkoff(p); 957 if (!can_prune_locals) { 958 for (j = 0; j < l; j++) 959 call->set_req(p++, in_map->in(k+j)); 960 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 961 // Divide stack into {S0,...,S1}, where S0 is set to top. 962 uint s1 = stack_slots_not_pruned; 963 stack_slots_not_pruned = 0; // for next iteration 964 if (s1 > l) s1 = l; 965 uint s0 = l - s1; 966 p += s0; // skip the tops preinstalled by add_req_batch 967 for (j = s0; j < l; j++) 968 call->set_req(p++, in_map->in(k+j)); 969 } else { 970 p += l; // already set to top above by add_req_batch 971 } 972 973 // Add the Monitors 974 k = in_jvms->monoff(); 975 l = in_jvms->mon_size(); 976 out_jvms->set_monoff(p); 977 for (j = 0; j < l; j++) 978 call->set_req(p++, in_map->in(k+j)); 979 980 // Copy any scalar object fields. 981 k = in_jvms->scloff(); 982 l = in_jvms->scl_size(); 983 out_jvms->set_scloff(p); 984 for (j = 0; j < l; j++) 985 call->set_req(p++, in_map->in(k+j)); 986 987 // Finish the new jvms. 988 out_jvms->set_endoff(p); 989 990 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 991 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 992 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 993 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 994 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 995 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 996 997 // Update the two tail pointers in parallel. 998 out_jvms = out_jvms->caller(); 999 in_jvms = in_jvms->caller(); 1000 } 1001 1002 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1003 1004 // Test the correctness of JVMState::debug_xxx accessors: 1005 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1006 assert(call->jvms()->debug_end() == call->req(), ""); 1007 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1008 } 1009 1010 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1011 Bytecodes::Code code = java_bc(); 1012 if (code == Bytecodes::_wide) { 1013 code = method()->java_code_at_bci(bci() + 1); 1014 } 1015 1016 BasicType rtype = T_ILLEGAL; 1017 int rsize = 0; 1018 1019 if (code != Bytecodes::_illegal) { 1020 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1021 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1022 if (rtype < T_CONFLICT) 1023 rsize = type2size[rtype]; 1024 } 1025 1026 switch (code) { 1027 case Bytecodes::_illegal: 1028 return false; 1029 1030 case Bytecodes::_ldc: 1031 case Bytecodes::_ldc_w: 1032 case Bytecodes::_ldc2_w: 1033 inputs = 0; 1034 break; 1035 1036 case Bytecodes::_dup: inputs = 1; break; 1037 case Bytecodes::_dup_x1: inputs = 2; break; 1038 case Bytecodes::_dup_x2: inputs = 3; break; 1039 case Bytecodes::_dup2: inputs = 2; break; 1040 case Bytecodes::_dup2_x1: inputs = 3; break; 1041 case Bytecodes::_dup2_x2: inputs = 4; break; 1042 case Bytecodes::_swap: inputs = 2; break; 1043 case Bytecodes::_arraylength: inputs = 1; break; 1044 1045 case Bytecodes::_getstatic: 1046 case Bytecodes::_putstatic: 1047 case Bytecodes::_getfield: 1048 case Bytecodes::_putfield: 1049 { 1050 bool ignored_will_link; 1051 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1052 int size = field->type()->size(); 1053 bool is_get = (depth >= 0), is_static = (depth & 1); 1054 inputs = (is_static ? 0 : 1); 1055 if (is_get) { 1056 depth = size - inputs; 1057 } else { 1058 inputs += size; // putxxx pops the value from the stack 1059 depth = - inputs; 1060 } 1061 } 1062 break; 1063 1064 case Bytecodes::_invokevirtual: 1065 case Bytecodes::_invokespecial: 1066 case Bytecodes::_invokestatic: 1067 case Bytecodes::_invokedynamic: 1068 case Bytecodes::_invokeinterface: 1069 { 1070 bool ignored_will_link; 1071 ciSignature* declared_signature = NULL; 1072 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1073 assert(declared_signature != NULL, "cannot be null"); 1074 inputs = declared_signature->arg_size_for_bc(code); 1075 int size = declared_signature->return_type()->size(); 1076 depth = size - inputs; 1077 } 1078 break; 1079 1080 case Bytecodes::_multianewarray: 1081 { 1082 ciBytecodeStream iter(method()); 1083 iter.reset_to_bci(bci()); 1084 iter.next(); 1085 inputs = iter.get_dimensions(); 1086 assert(rsize == 1, ""); 1087 depth = rsize - inputs; 1088 } 1089 break; 1090 1091 case Bytecodes::_withfield: { 1092 bool ignored_will_link; 1093 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1094 int size = field->type()->size(); 1095 inputs = size+1; 1096 depth = rsize - inputs; 1097 break; 1098 } 1099 1100 case Bytecodes::_ireturn: 1101 case Bytecodes::_lreturn: 1102 case Bytecodes::_freturn: 1103 case Bytecodes::_dreturn: 1104 case Bytecodes::_areturn: 1105 assert(rsize == -depth, ""); 1106 inputs = rsize; 1107 break; 1108 1109 case Bytecodes::_jsr: 1110 case Bytecodes::_jsr_w: 1111 inputs = 0; 1112 depth = 1; // S.B. depth=1, not zero 1113 break; 1114 1115 default: 1116 // bytecode produces a typed result 1117 inputs = rsize - depth; 1118 assert(inputs >= 0, ""); 1119 break; 1120 } 1121 1122 #ifdef ASSERT 1123 // spot check 1124 int outputs = depth + inputs; 1125 assert(outputs >= 0, "sanity"); 1126 switch (code) { 1127 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1128 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1129 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1130 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1131 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1132 default: break; 1133 } 1134 #endif //ASSERT 1135 1136 return true; 1137 } 1138 1139 1140 1141 //------------------------------basic_plus_adr--------------------------------- 1142 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1143 // short-circuit a common case 1144 if (offset == intcon(0)) return ptr; 1145 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1146 } 1147 1148 Node* GraphKit::ConvI2L(Node* offset) { 1149 // short-circuit a common case 1150 jint offset_con = find_int_con(offset, Type::OffsetBot); 1151 if (offset_con != Type::OffsetBot) { 1152 return longcon((jlong) offset_con); 1153 } 1154 return _gvn.transform( new ConvI2LNode(offset)); 1155 } 1156 1157 Node* GraphKit::ConvI2UL(Node* offset) { 1158 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1159 if (offset_con != (juint) Type::OffsetBot) { 1160 return longcon((julong) offset_con); 1161 } 1162 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1163 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1164 return _gvn.transform( new AndLNode(conv, mask) ); 1165 } 1166 1167 Node* GraphKit::ConvL2I(Node* offset) { 1168 // short-circuit a common case 1169 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1170 if (offset_con != (jlong)Type::OffsetBot) { 1171 return intcon((int) offset_con); 1172 } 1173 return _gvn.transform( new ConvL2INode(offset)); 1174 } 1175 1176 //-------------------------load_object_klass----------------------------------- 1177 Node* GraphKit::load_object_klass(Node* obj) { 1178 // Special-case a fresh allocation to avoid building nodes: 1179 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1180 if (akls != NULL) return akls; 1181 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1182 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1183 } 1184 1185 //-------------------------load_array_length----------------------------------- 1186 Node* GraphKit::load_array_length(Node* array) { 1187 // Special-case a fresh allocation to avoid building nodes: 1188 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1189 Node *alen; 1190 if (alloc == NULL) { 1191 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1192 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1193 } else { 1194 alen = alloc->Ideal_length(); 1195 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1196 if (ccast != alen) { 1197 alen = _gvn.transform(ccast); 1198 } 1199 } 1200 return alen; 1201 } 1202 1203 //------------------------------do_null_check---------------------------------- 1204 // Helper function to do a NULL pointer check. Returned value is 1205 // the incoming address with NULL casted away. You are allowed to use the 1206 // not-null value only if you are control dependent on the test. 1207 #ifndef PRODUCT 1208 extern int explicit_null_checks_inserted, 1209 explicit_null_checks_elided; 1210 #endif 1211 Node* GraphKit::null_check_common(Node* value, BasicType type, 1212 // optional arguments for variations: 1213 bool assert_null, 1214 Node* *null_control, 1215 bool speculative) { 1216 assert(!assert_null || null_control == NULL, "not both at once"); 1217 if (stopped()) return top(); 1218 NOT_PRODUCT(explicit_null_checks_inserted++); 1219 1220 // Construct NULL check 1221 Node *chk = NULL; 1222 switch(type) { 1223 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1224 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1225 case T_VALUETYPE : // fall through 1226 case T_ARRAY : // fall through 1227 type = T_OBJECT; // simplify further tests 1228 case T_OBJECT : { 1229 const Type *t = _gvn.type( value ); 1230 1231 const TypeOopPtr* tp = t->isa_oopptr(); 1232 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1233 // Only for do_null_check, not any of its siblings: 1234 && !assert_null && null_control == NULL) { 1235 // Usually, any field access or invocation on an unloaded oop type 1236 // will simply fail to link, since the statically linked class is 1237 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1238 // the static class is loaded but the sharper oop type is not. 1239 // Rather than checking for this obscure case in lots of places, 1240 // we simply observe that a null check on an unloaded class 1241 // will always be followed by a nonsense operation, so we 1242 // can just issue the uncommon trap here. 1243 // Our access to the unloaded class will only be correct 1244 // after it has been loaded and initialized, which requires 1245 // a trip through the interpreter. 1246 #ifndef PRODUCT 1247 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1248 #endif 1249 uncommon_trap(Deoptimization::Reason_unloaded, 1250 Deoptimization::Action_reinterpret, 1251 tp->klass(), "!loaded"); 1252 return top(); 1253 } 1254 1255 if (assert_null) { 1256 // See if the type is contained in NULL_PTR. 1257 // If so, then the value is already null. 1258 if (t->higher_equal(TypePtr::NULL_PTR)) { 1259 NOT_PRODUCT(explicit_null_checks_elided++); 1260 return value; // Elided null assert quickly! 1261 } 1262 } else { 1263 // See if mixing in the NULL pointer changes type. 1264 // If so, then the NULL pointer was not allowed in the original 1265 // type. In other words, "value" was not-null. 1266 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1267 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1268 NOT_PRODUCT(explicit_null_checks_elided++); 1269 return value; // Elided null check quickly! 1270 } 1271 } 1272 chk = new CmpPNode( value, null() ); 1273 break; 1274 } 1275 1276 default: 1277 fatal("unexpected type: %s", type2name(type)); 1278 } 1279 assert(chk != NULL, "sanity check"); 1280 chk = _gvn.transform(chk); 1281 1282 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1283 BoolNode *btst = new BoolNode( chk, btest); 1284 Node *tst = _gvn.transform( btst ); 1285 1286 //----------- 1287 // if peephole optimizations occurred, a prior test existed. 1288 // If a prior test existed, maybe it dominates as we can avoid this test. 1289 if (tst != btst && type == T_OBJECT) { 1290 // At this point we want to scan up the CFG to see if we can 1291 // find an identical test (and so avoid this test altogether). 1292 Node *cfg = control(); 1293 int depth = 0; 1294 while( depth < 16 ) { // Limit search depth for speed 1295 if( cfg->Opcode() == Op_IfTrue && 1296 cfg->in(0)->in(1) == tst ) { 1297 // Found prior test. Use "cast_not_null" to construct an identical 1298 // CastPP (and hence hash to) as already exists for the prior test. 1299 // Return that casted value. 1300 if (assert_null) { 1301 replace_in_map(value, null()); 1302 return null(); // do not issue the redundant test 1303 } 1304 Node *oldcontrol = control(); 1305 set_control(cfg); 1306 Node *res = cast_not_null(value); 1307 set_control(oldcontrol); 1308 NOT_PRODUCT(explicit_null_checks_elided++); 1309 return res; 1310 } 1311 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1312 if (cfg == NULL) break; // Quit at region nodes 1313 depth++; 1314 } 1315 } 1316 1317 //----------- 1318 // Branch to failure if null 1319 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1320 Deoptimization::DeoptReason reason; 1321 if (assert_null) { 1322 reason = Deoptimization::reason_null_assert(speculative); 1323 } else if (type == T_OBJECT) { 1324 reason = Deoptimization::reason_null_check(speculative); 1325 } else { 1326 reason = Deoptimization::Reason_div0_check; 1327 } 1328 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1329 // ciMethodData::has_trap_at will return a conservative -1 if any 1330 // must-be-null assertion has failed. This could cause performance 1331 // problems for a method after its first do_null_assert failure. 1332 // Consider using 'Reason_class_check' instead? 1333 1334 // To cause an implicit null check, we set the not-null probability 1335 // to the maximum (PROB_MAX). For an explicit check the probability 1336 // is set to a smaller value. 1337 if (null_control != NULL || too_many_traps(reason)) { 1338 // probability is less likely 1339 ok_prob = PROB_LIKELY_MAG(3); 1340 } else if (!assert_null && 1341 (ImplicitNullCheckThreshold > 0) && 1342 method() != NULL && 1343 (method()->method_data()->trap_count(reason) 1344 >= (uint)ImplicitNullCheckThreshold)) { 1345 ok_prob = PROB_LIKELY_MAG(3); 1346 } 1347 1348 if (null_control != NULL) { 1349 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1350 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1351 set_control( _gvn.transform( new IfTrueNode(iff))); 1352 #ifndef PRODUCT 1353 if (null_true == top()) { 1354 explicit_null_checks_elided++; 1355 } 1356 #endif 1357 (*null_control) = null_true; 1358 } else { 1359 BuildCutout unless(this, tst, ok_prob); 1360 // Check for optimizer eliding test at parse time 1361 if (stopped()) { 1362 // Failure not possible; do not bother making uncommon trap. 1363 NOT_PRODUCT(explicit_null_checks_elided++); 1364 } else if (assert_null) { 1365 uncommon_trap(reason, 1366 Deoptimization::Action_make_not_entrant, 1367 NULL, "assert_null"); 1368 } else { 1369 replace_in_map(value, zerocon(type)); 1370 builtin_throw(reason); 1371 } 1372 } 1373 1374 // Must throw exception, fall-thru not possible? 1375 if (stopped()) { 1376 return top(); // No result 1377 } 1378 1379 if (assert_null) { 1380 // Cast obj to null on this path. 1381 replace_in_map(value, zerocon(type)); 1382 return zerocon(type); 1383 } 1384 1385 // Cast obj to not-null on this path, if there is no null_control. 1386 // (If there is a null_control, a non-null value may come back to haunt us.) 1387 return cast_not_null(value, (null_control == NULL || (*null_control) == top())); 1388 } 1389 1390 Node* GraphKit::null2default(Node* value, ciValueKlass* vk) { 1391 Node* null_ctl = top(); 1392 value = null_check_oop(value, &null_ctl); 1393 if (!null_ctl->is_top()) { 1394 // Return default value if oop is null 1395 Node* region = new RegionNode(3); 1396 region->init_req(1, control()); 1397 region->init_req(2, null_ctl); 1398 value = PhiNode::make(region, value, TypeInstPtr::make(TypePtr::BotPTR, vk)); 1399 value->set_req(2, ValueTypeNode::default_oop(gvn(), vk)); 1400 set_control(gvn().transform(region)); 1401 value = gvn().transform(value); 1402 } 1403 return value; 1404 } 1405 1406 //------------------------------cast_not_null---------------------------------- 1407 // Cast obj to not-null on this path 1408 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1409 if (obj->is_ValueType()) { 1410 return obj; 1411 } 1412 Node* cast = NULL; 1413 const Type* t = _gvn.type(obj); 1414 if (t->make_ptr() != NULL) { 1415 const Type* t_not_null = t->join_speculative(TypePtr::NOTNULL); 1416 // Object is already not-null? 1417 if (t == t_not_null) { 1418 return obj; 1419 } 1420 cast = ConstraintCastNode::make_cast(Op_CastPP, control(), obj, t_not_null, false); 1421 } else if (t->isa_int() != NULL) { 1422 cast = ConstraintCastNode::make_cast(Op_CastII, control(), obj, TypeInt::INT, true); 1423 } else if (t->isa_long() != NULL) { 1424 cast = ConstraintCastNode::make_cast(Op_CastLL, control(), obj, TypeLong::LONG, true); 1425 } else { 1426 fatal("unexpected type: %s", type2name(t->basic_type())); 1427 } 1428 cast = _gvn.transform(cast); 1429 1430 // Scan for instances of 'obj' in the current JVM mapping. 1431 // These instances are known to be not-null after the test. 1432 if (do_replace_in_map) { 1433 replace_in_map(obj, cast); 1434 } 1435 return cast; 1436 } 1437 1438 // Sometimes in intrinsics, we implicitly know an object is not null 1439 // (there's no actual null check) so we can cast it to not null. In 1440 // the course of optimizations, the input to the cast can become null. 1441 // In that case that data path will die and we need the control path 1442 // to become dead as well to keep the graph consistent. So we have to 1443 // add a check for null for which one branch can't be taken. It uses 1444 // an Opaque4 node that will cause the check to be removed after loop 1445 // opts so the test goes away and the compiled code doesn't execute a 1446 // useless check. 1447 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1448 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1449 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1450 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1451 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1452 _gvn.set_type(iff, iff->Value(&_gvn)); 1453 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1454 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1455 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1456 C->root()->add_req(halt); 1457 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1458 set_control(if_t); 1459 return cast_not_null(value, do_replace_in_map); 1460 } 1461 1462 1463 //--------------------------replace_in_map------------------------------------- 1464 void GraphKit::replace_in_map(Node* old, Node* neww) { 1465 if (old == neww) { 1466 return; 1467 } 1468 1469 map()->replace_edge(old, neww); 1470 1471 // Note: This operation potentially replaces any edge 1472 // on the map. This includes locals, stack, and monitors 1473 // of the current (innermost) JVM state. 1474 1475 // don't let inconsistent types from profiling escape this 1476 // method 1477 1478 const Type* told = _gvn.type(old); 1479 const Type* tnew = _gvn.type(neww); 1480 1481 if (!tnew->higher_equal(told)) { 1482 return; 1483 } 1484 1485 map()->record_replaced_node(old, neww); 1486 } 1487 1488 1489 //============================================================================= 1490 //--------------------------------memory--------------------------------------- 1491 Node* GraphKit::memory(uint alias_idx) { 1492 MergeMemNode* mem = merged_memory(); 1493 Node* p = mem->memory_at(alias_idx); 1494 _gvn.set_type(p, Type::MEMORY); // must be mapped 1495 return p; 1496 } 1497 1498 //-----------------------------reset_memory------------------------------------ 1499 Node* GraphKit::reset_memory() { 1500 Node* mem = map()->memory(); 1501 // do not use this node for any more parsing! 1502 debug_only( map()->set_memory((Node*)NULL) ); 1503 return _gvn.transform( mem ); 1504 } 1505 1506 //------------------------------set_all_memory--------------------------------- 1507 void GraphKit::set_all_memory(Node* newmem) { 1508 Node* mergemem = MergeMemNode::make(newmem); 1509 gvn().set_type_bottom(mergemem); 1510 map()->set_memory(mergemem); 1511 } 1512 1513 //------------------------------set_all_memory_call---------------------------- 1514 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1515 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1516 set_all_memory(newmem); 1517 } 1518 1519 //============================================================================= 1520 // 1521 // parser factory methods for MemNodes 1522 // 1523 // These are layered on top of the factory methods in LoadNode and StoreNode, 1524 // and integrate with the parser's memory state and _gvn engine. 1525 // 1526 1527 // factory methods in "int adr_idx" 1528 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1529 int adr_idx, 1530 MemNode::MemOrd mo, 1531 LoadNode::ControlDependency control_dependency, 1532 bool require_atomic_access, 1533 bool unaligned, 1534 bool mismatched, 1535 bool unsafe) { 1536 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1537 const TypePtr* adr_type = NULL; // debug-mode-only argument 1538 debug_only(adr_type = C->get_adr_type(adr_idx)); 1539 Node* mem = memory(adr_idx); 1540 Node* ld; 1541 if (require_atomic_access && bt == T_LONG) { 1542 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe); 1543 } else if (require_atomic_access && bt == T_DOUBLE) { 1544 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe); 1545 } else { 1546 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe); 1547 } 1548 ld = _gvn.transform(ld); 1549 1550 if (((bt == T_OBJECT || bt == T_VALUETYPE) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1551 // Improve graph before escape analysis and boxing elimination. 1552 record_for_igvn(ld); 1553 } 1554 return ld; 1555 } 1556 1557 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1558 int adr_idx, 1559 MemNode::MemOrd mo, 1560 bool require_atomic_access, 1561 bool unaligned, 1562 bool mismatched, 1563 bool unsafe) { 1564 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1565 const TypePtr* adr_type = NULL; 1566 debug_only(adr_type = C->get_adr_type(adr_idx)); 1567 Node *mem = memory(adr_idx); 1568 Node* st; 1569 if (require_atomic_access && bt == T_LONG) { 1570 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1571 } else if (require_atomic_access && bt == T_DOUBLE) { 1572 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1573 } else { 1574 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1575 } 1576 if (unaligned) { 1577 st->as_Store()->set_unaligned_access(); 1578 } 1579 if (mismatched) { 1580 st->as_Store()->set_mismatched_access(); 1581 } 1582 if (unsafe) { 1583 st->as_Store()->set_unsafe_access(); 1584 } 1585 st = _gvn.transform(st); 1586 set_memory(st, adr_idx); 1587 // Back-to-back stores can only remove intermediate store with DU info 1588 // so push on worklist for optimizer. 1589 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1590 record_for_igvn(st); 1591 1592 return st; 1593 } 1594 1595 Node* GraphKit::access_store_at(Node* obj, 1596 Node* adr, 1597 const TypePtr* adr_type, 1598 Node* val, 1599 const Type* val_type, 1600 BasicType bt, 1601 DecoratorSet decorators, 1602 bool deoptimize_on_exception, 1603 bool safe_for_replace) { 1604 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1605 // could be delayed during Parse (for example, in adjust_map_after_if()). 1606 // Execute transformation here to avoid barrier generation in such case. 1607 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1608 val = _gvn.makecon(TypePtr::NULL_PTR); 1609 } 1610 1611 if (stopped()) { 1612 return top(); // Dead path ? 1613 } 1614 1615 assert(val != NULL, "not dead path"); 1616 if (val->is_ValueType()) { 1617 // Allocate value type and get oop 1618 val = val->as_ValueType()->allocate(this, deoptimize_on_exception, safe_for_replace)->get_oop(); 1619 } 1620 1621 C2AccessValuePtr addr(adr, adr_type); 1622 C2AccessValue value(val, val_type); 1623 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1624 if (access.is_raw()) { 1625 return _barrier_set->BarrierSetC2::store_at(access, value); 1626 } else { 1627 return _barrier_set->store_at(access, value); 1628 } 1629 } 1630 1631 Node* GraphKit::access_load_at(Node* obj, // containing obj 1632 Node* adr, // actual adress to store val at 1633 const TypePtr* adr_type, 1634 const Type* val_type, 1635 BasicType bt, 1636 DecoratorSet decorators, 1637 Node* ctl) { 1638 if (stopped()) { 1639 return top(); // Dead path ? 1640 } 1641 1642 C2AccessValuePtr addr(adr, adr_type); 1643 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl); 1644 if (access.is_raw()) { 1645 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1646 } else { 1647 return _barrier_set->load_at(access, val_type); 1648 } 1649 } 1650 1651 Node* GraphKit::access_load(Node* adr, // actual adress to load val at 1652 const Type* val_type, 1653 BasicType bt, 1654 DecoratorSet decorators) { 1655 if (stopped()) { 1656 return top(); // Dead path ? 1657 } 1658 1659 C2AccessValuePtr addr(adr, NULL); 1660 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr); 1661 if (access.is_raw()) { 1662 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1663 } else { 1664 return _barrier_set->load_at(access, val_type); 1665 } 1666 } 1667 1668 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1669 Node* adr, 1670 const TypePtr* adr_type, 1671 int alias_idx, 1672 Node* expected_val, 1673 Node* new_val, 1674 const Type* value_type, 1675 BasicType bt, 1676 DecoratorSet decorators) { 1677 C2AccessValuePtr addr(adr, adr_type); 1678 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1679 bt, obj, addr, alias_idx); 1680 if (access.is_raw()) { 1681 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1682 } else { 1683 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1684 } 1685 } 1686 1687 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1688 Node* adr, 1689 const TypePtr* adr_type, 1690 int alias_idx, 1691 Node* expected_val, 1692 Node* new_val, 1693 const Type* value_type, 1694 BasicType bt, 1695 DecoratorSet decorators) { 1696 C2AccessValuePtr addr(adr, adr_type); 1697 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1698 bt, obj, addr, alias_idx); 1699 if (access.is_raw()) { 1700 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1701 } else { 1702 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1703 } 1704 } 1705 1706 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1707 Node* adr, 1708 const TypePtr* adr_type, 1709 int alias_idx, 1710 Node* new_val, 1711 const Type* value_type, 1712 BasicType bt, 1713 DecoratorSet decorators) { 1714 C2AccessValuePtr addr(adr, adr_type); 1715 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1716 bt, obj, addr, alias_idx); 1717 if (access.is_raw()) { 1718 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1719 } else { 1720 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1721 } 1722 } 1723 1724 Node* GraphKit::access_atomic_add_at(Node* obj, 1725 Node* adr, 1726 const TypePtr* adr_type, 1727 int alias_idx, 1728 Node* new_val, 1729 const Type* value_type, 1730 BasicType bt, 1731 DecoratorSet decorators) { 1732 C2AccessValuePtr addr(adr, adr_type); 1733 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1734 if (access.is_raw()) { 1735 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1736 } else { 1737 return _barrier_set->atomic_add_at(access, new_val, value_type); 1738 } 1739 } 1740 1741 void GraphKit::access_clone(Node* src_base, Node* dst_base, Node* countx, bool is_array) { 1742 return _barrier_set->clone(this, src_base, dst_base, countx, is_array); 1743 } 1744 1745 Node* GraphKit::access_resolve(Node* n, DecoratorSet decorators) { 1746 // Use stronger ACCESS_WRITE|ACCESS_READ by default. 1747 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) { 1748 decorators |= ACCESS_READ | ACCESS_WRITE; 1749 } 1750 return _barrier_set->resolve(this, n, decorators); 1751 } 1752 1753 //-------------------------array_element_address------------------------- 1754 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1755 const TypeInt* sizetype, Node* ctrl) { 1756 uint shift = exact_log2(type2aelembytes(elembt)); 1757 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1758 if (arytype_klass != NULL && arytype_klass->is_value_array_klass()) { 1759 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass(); 1760 shift = vak->log2_element_size(); 1761 } 1762 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1763 1764 // short-circuit a common case (saves lots of confusing waste motion) 1765 jint idx_con = find_int_con(idx, -1); 1766 if (idx_con >= 0) { 1767 intptr_t offset = header + ((intptr_t)idx_con << shift); 1768 return basic_plus_adr(ary, offset); 1769 } 1770 1771 // must be correct type for alignment purposes 1772 Node* base = basic_plus_adr(ary, header); 1773 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1774 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1775 return basic_plus_adr(ary, base, scale); 1776 } 1777 1778 //-------------------------load_array_element------------------------- 1779 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1780 const Type* elemtype = arytype->elem(); 1781 BasicType elembt = elemtype->array_element_basic_type(); 1782 assert(elembt != T_VALUETYPE, "value types are not supported by this method"); 1783 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1784 if (elembt == T_NARROWOOP) { 1785 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1786 } 1787 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1788 return ld; 1789 } 1790 1791 //-------------------------set_arguments_for_java_call------------------------- 1792 // Arguments (pre-popped from the stack) are taken from the JVMS. 1793 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool incremental_inlining) { 1794 // Add the call arguments: 1795 const TypeTuple* domain = call->tf()->domain_sig(); 1796 ExtendedSignature sig_cc = ExtendedSignature(call->method()->get_sig_cc(), SigEntryFilter()); 1797 uint nargs = domain->cnt(); 1798 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1799 Node* arg = argument(i-TypeFunc::Parms); 1800 const Type* t = domain->field_at(i); 1801 if (call->method()->has_scalarized_args() && t->is_valuetypeptr() && !t->maybe_null()) { 1802 // We don't pass value type arguments by reference but instead pass each field of the value type 1803 ValueTypeNode* vt = arg->as_ValueType(); 1804 vt->pass_fields(this, call, sig_cc, idx); 1805 // If a value type argument is passed as fields, attach the Method* to the call site 1806 // to be able to access the extended signature later via attached_method_before_pc(). 1807 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1808 call->set_override_symbolic_info(true); 1809 continue; 1810 } else if (arg->is_ValueType()) { 1811 // Pass value type argument via oop to callee 1812 if (!incremental_inlining) { 1813 arg = arg->as_ValueType()->allocate(this)->get_oop(); 1814 } else { 1815 arg = ValueTypePtrNode::make_from_value_type(this, arg->as_ValueType()); 1816 } 1817 } 1818 call->init_req(idx++, arg); 1819 // Skip reserved arguments 1820 BasicType bt = t->basic_type(); 1821 while (SigEntry::next_is_reserved(sig_cc, bt, true)) { 1822 call->init_req(idx++, top()); 1823 if (type2size[bt] == 2) { 1824 call->init_req(idx++, top()); 1825 } 1826 } 1827 } 1828 } 1829 1830 //---------------------------set_edges_for_java_call--------------------------- 1831 // Connect a newly created call into the current JVMS. 1832 // A return value node (if any) is returned from set_edges_for_java_call. 1833 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1834 1835 // Add the predefined inputs: 1836 call->init_req( TypeFunc::Control, control() ); 1837 call->init_req( TypeFunc::I_O , i_o() ); 1838 call->init_req( TypeFunc::Memory , reset_memory() ); 1839 call->init_req( TypeFunc::FramePtr, frameptr() ); 1840 call->init_req( TypeFunc::ReturnAdr, top() ); 1841 1842 add_safepoint_edges(call, must_throw); 1843 1844 Node* xcall = _gvn.transform(call); 1845 1846 if (xcall == top()) { 1847 set_control(top()); 1848 return; 1849 } 1850 assert(xcall == call, "call identity is stable"); 1851 1852 // Re-use the current map to produce the result. 1853 1854 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1855 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1856 set_all_memory_call(xcall, separate_io_proj); 1857 1858 //return xcall; // no need, caller already has it 1859 } 1860 1861 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1862 if (stopped()) return top(); // maybe the call folded up? 1863 1864 // Note: Since any out-of-line call can produce an exception, 1865 // we always insert an I_O projection from the call into the result. 1866 1867 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1868 1869 if (separate_io_proj) { 1870 // The caller requested separate projections be used by the fall 1871 // through and exceptional paths, so replace the projections for 1872 // the fall through path. 1873 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1874 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1875 } 1876 1877 // Capture the return value, if any. 1878 Node* ret; 1879 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) { 1880 ret = top(); 1881 } else if (call->tf()->returns_value_type_as_fields()) { 1882 // Return of multiple values (value type fields): we create a 1883 // ValueType node, each field is a projection from the call. 1884 ciValueKlass* vk = call->method()->return_type()->as_value_klass(); 1885 const Array<SigEntry>* sig_array = vk->extended_sig(); 1886 GrowableArray<SigEntry> sig = GrowableArray<SigEntry>(sig_array->length()); 1887 sig.appendAll(sig_array); 1888 ExtendedSignature sig_cc = ExtendedSignature(&sig, SigEntryFilter()); 1889 uint base_input = TypeFunc::Parms + 1; 1890 ret = ValueTypeNode::make_from_multi(this, call, sig_cc, vk, base_input, false); 1891 } else { 1892 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1893 } 1894 1895 return ret; 1896 } 1897 1898 //--------------------set_predefined_input_for_runtime_call-------------------- 1899 // Reading and setting the memory state is way conservative here. 1900 // The real problem is that I am not doing real Type analysis on memory, 1901 // so I cannot distinguish card mark stores from other stores. Across a GC 1902 // point the Store Barrier and the card mark memory has to agree. I cannot 1903 // have a card mark store and its barrier split across the GC point from 1904 // either above or below. Here I get that to happen by reading ALL of memory. 1905 // A better answer would be to separate out card marks from other memory. 1906 // For now, return the input memory state, so that it can be reused 1907 // after the call, if this call has restricted memory effects. 1908 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1909 // Set fixed predefined input arguments 1910 Node* memory = reset_memory(); 1911 Node* m = narrow_mem == NULL ? memory : narrow_mem; 1912 call->init_req( TypeFunc::Control, control() ); 1913 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1914 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1915 call->init_req( TypeFunc::FramePtr, frameptr() ); 1916 call->init_req( TypeFunc::ReturnAdr, top() ); 1917 return memory; 1918 } 1919 1920 //-------------------set_predefined_output_for_runtime_call-------------------- 1921 // Set control and memory (not i_o) from the call. 1922 // If keep_mem is not NULL, use it for the output state, 1923 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1924 // If hook_mem is NULL, this call produces no memory effects at all. 1925 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1926 // then only that memory slice is taken from the call. 1927 // In the last case, we must put an appropriate memory barrier before 1928 // the call, so as to create the correct anti-dependencies on loads 1929 // preceding the call. 1930 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1931 Node* keep_mem, 1932 const TypePtr* hook_mem) { 1933 // no i/o 1934 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1935 if (keep_mem) { 1936 // First clone the existing memory state 1937 set_all_memory(keep_mem); 1938 if (hook_mem != NULL) { 1939 // Make memory for the call 1940 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1941 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1942 // We also use hook_mem to extract specific effects from arraycopy stubs. 1943 set_memory(mem, hook_mem); 1944 } 1945 // ...else the call has NO memory effects. 1946 1947 // Make sure the call advertises its memory effects precisely. 1948 // This lets us build accurate anti-dependences in gcm.cpp. 1949 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1950 "call node must be constructed correctly"); 1951 } else { 1952 assert(hook_mem == NULL, ""); 1953 // This is not a "slow path" call; all memory comes from the call. 1954 set_all_memory_call(call); 1955 } 1956 } 1957 1958 // Keep track of MergeMems feeding into other MergeMems 1959 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 1960 if (!mem->is_MergeMem()) { 1961 return; 1962 } 1963 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 1964 Node* use = i.get(); 1965 if (use->is_MergeMem()) { 1966 wl.push(use); 1967 } 1968 } 1969 } 1970 1971 // Replace the call with the current state of the kit. 1972 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1973 JVMState* ejvms = NULL; 1974 if (has_exceptions()) { 1975 ejvms = transfer_exceptions_into_jvms(); 1976 } 1977 1978 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1979 ReplacedNodes replaced_nodes_exception; 1980 Node* ex_ctl = top(); 1981 1982 SafePointNode* final_state = stop(); 1983 1984 // Find all the needed outputs of this call 1985 CallProjections* callprojs = call->extract_projections(true); 1986 1987 Unique_Node_List wl; 1988 Node* init_mem = call->in(TypeFunc::Memory); 1989 Node* final_mem = final_state->in(TypeFunc::Memory); 1990 Node* final_ctl = final_state->in(TypeFunc::Control); 1991 Node* final_io = final_state->in(TypeFunc::I_O); 1992 1993 // Replace all the old call edges with the edges from the inlining result 1994 if (callprojs->fallthrough_catchproj != NULL) { 1995 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl); 1996 } 1997 if (callprojs->fallthrough_memproj != NULL) { 1998 if (final_mem->is_MergeMem()) { 1999 // Parser's exits MergeMem was not transformed but may be optimized 2000 final_mem = _gvn.transform(final_mem); 2001 } 2002 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem); 2003 add_mergemem_users_to_worklist(wl, final_mem); 2004 } 2005 if (callprojs->fallthrough_ioproj != NULL) { 2006 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io); 2007 } 2008 2009 // Replace the result with the new result if it exists and is used 2010 if (callprojs->resproj[0] != NULL && result != NULL) { 2011 assert(callprojs->nb_resproj == 1, "unexpected number of results"); 2012 C->gvn_replace_by(callprojs->resproj[0], result); 2013 } 2014 2015 if (ejvms == NULL) { 2016 // No exception edges to simply kill off those paths 2017 if (callprojs->catchall_catchproj != NULL) { 2018 C->gvn_replace_by(callprojs->catchall_catchproj, C->top()); 2019 } 2020 if (callprojs->catchall_memproj != NULL) { 2021 C->gvn_replace_by(callprojs->catchall_memproj, C->top()); 2022 } 2023 if (callprojs->catchall_ioproj != NULL) { 2024 C->gvn_replace_by(callprojs->catchall_ioproj, C->top()); 2025 } 2026 // Replace the old exception object with top 2027 if (callprojs->exobj != NULL) { 2028 C->gvn_replace_by(callprojs->exobj, C->top()); 2029 } 2030 } else { 2031 GraphKit ekit(ejvms); 2032 2033 // Load my combined exception state into the kit, with all phis transformed: 2034 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 2035 replaced_nodes_exception = ex_map->replaced_nodes(); 2036 2037 Node* ex_oop = ekit.use_exception_state(ex_map); 2038 2039 if (callprojs->catchall_catchproj != NULL) { 2040 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control()); 2041 ex_ctl = ekit.control(); 2042 } 2043 if (callprojs->catchall_memproj != NULL) { 2044 Node* ex_mem = ekit.reset_memory(); 2045 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem); 2046 add_mergemem_users_to_worklist(wl, ex_mem); 2047 } 2048 if (callprojs->catchall_ioproj != NULL) { 2049 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o()); 2050 } 2051 2052 // Replace the old exception object with the newly created one 2053 if (callprojs->exobj != NULL) { 2054 C->gvn_replace_by(callprojs->exobj, ex_oop); 2055 } 2056 } 2057 2058 // Disconnect the call from the graph 2059 call->disconnect_inputs(NULL, C); 2060 C->gvn_replace_by(call, C->top()); 2061 2062 // Clean up any MergeMems that feed other MergeMems since the 2063 // optimizer doesn't like that. 2064 while (wl.size() > 0) { 2065 _gvn.transform(wl.pop()); 2066 } 2067 2068 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 2069 replaced_nodes.apply(C, final_ctl); 2070 } 2071 if (!ex_ctl->is_top() && do_replaced_nodes) { 2072 replaced_nodes_exception.apply(C, ex_ctl); 2073 } 2074 } 2075 2076 2077 //------------------------------increment_counter------------------------------ 2078 // for statistics: increment a VM counter by 1 2079 2080 void GraphKit::increment_counter(address counter_addr) { 2081 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2082 increment_counter(adr1); 2083 } 2084 2085 void GraphKit::increment_counter(Node* counter_addr) { 2086 int adr_type = Compile::AliasIdxRaw; 2087 Node* ctrl = control(); 2088 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 2089 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 2090 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 2091 } 2092 2093 2094 //------------------------------uncommon_trap---------------------------------- 2095 // Bail out to the interpreter in mid-method. Implemented by calling the 2096 // uncommon_trap blob. This helper function inserts a runtime call with the 2097 // right debug info. 2098 void GraphKit::uncommon_trap(int trap_request, 2099 ciKlass* klass, const char* comment, 2100 bool must_throw, 2101 bool keep_exact_action) { 2102 if (failing()) stop(); 2103 if (stopped()) return; // trap reachable? 2104 2105 // Note: If ProfileTraps is true, and if a deopt. actually 2106 // occurs here, the runtime will make sure an MDO exists. There is 2107 // no need to call method()->ensure_method_data() at this point. 2108 2109 // Set the stack pointer to the right value for reexecution: 2110 set_sp(reexecute_sp()); 2111 2112 #ifdef ASSERT 2113 if (!must_throw) { 2114 // Make sure the stack has at least enough depth to execute 2115 // the current bytecode. 2116 int inputs, ignored_depth; 2117 if (compute_stack_effects(inputs, ignored_depth)) { 2118 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2119 Bytecodes::name(java_bc()), sp(), inputs); 2120 } 2121 } 2122 #endif 2123 2124 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2125 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2126 2127 switch (action) { 2128 case Deoptimization::Action_maybe_recompile: 2129 case Deoptimization::Action_reinterpret: 2130 // Temporary fix for 6529811 to allow virtual calls to be sure they 2131 // get the chance to go from mono->bi->mega 2132 if (!keep_exact_action && 2133 Deoptimization::trap_request_index(trap_request) < 0 && 2134 too_many_recompiles(reason)) { 2135 // This BCI is causing too many recompilations. 2136 if (C->log() != NULL) { 2137 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2138 Deoptimization::trap_reason_name(reason), 2139 Deoptimization::trap_action_name(action)); 2140 } 2141 action = Deoptimization::Action_none; 2142 trap_request = Deoptimization::make_trap_request(reason, action); 2143 } else { 2144 C->set_trap_can_recompile(true); 2145 } 2146 break; 2147 case Deoptimization::Action_make_not_entrant: 2148 C->set_trap_can_recompile(true); 2149 break; 2150 case Deoptimization::Action_none: 2151 case Deoptimization::Action_make_not_compilable: 2152 break; 2153 default: 2154 #ifdef ASSERT 2155 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2156 #endif 2157 break; 2158 } 2159 2160 if (TraceOptoParse) { 2161 char buf[100]; 2162 tty->print_cr("Uncommon trap %s at bci:%d", 2163 Deoptimization::format_trap_request(buf, sizeof(buf), 2164 trap_request), bci()); 2165 } 2166 2167 CompileLog* log = C->log(); 2168 if (log != NULL) { 2169 int kid = (klass == NULL)? -1: log->identify(klass); 2170 log->begin_elem("uncommon_trap bci='%d'", bci()); 2171 char buf[100]; 2172 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2173 trap_request)); 2174 if (kid >= 0) log->print(" klass='%d'", kid); 2175 if (comment != NULL) log->print(" comment='%s'", comment); 2176 log->end_elem(); 2177 } 2178 2179 // Make sure any guarding test views this path as very unlikely 2180 Node *i0 = control()->in(0); 2181 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2182 IfNode *iff = i0->as_If(); 2183 float f = iff->_prob; // Get prob 2184 if (control()->Opcode() == Op_IfTrue) { 2185 if (f > PROB_UNLIKELY_MAG(4)) 2186 iff->_prob = PROB_MIN; 2187 } else { 2188 if (f < PROB_LIKELY_MAG(4)) 2189 iff->_prob = PROB_MAX; 2190 } 2191 } 2192 2193 // Clear out dead values from the debug info. 2194 kill_dead_locals(); 2195 2196 // Now insert the uncommon trap subroutine call 2197 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2198 const TypePtr* no_memory_effects = NULL; 2199 // Pass the index of the class to be loaded 2200 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2201 (must_throw ? RC_MUST_THROW : 0), 2202 OptoRuntime::uncommon_trap_Type(), 2203 call_addr, "uncommon_trap", no_memory_effects, 2204 intcon(trap_request)); 2205 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2206 "must extract request correctly from the graph"); 2207 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2208 2209 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2210 // The debug info is the only real input to this call. 2211 2212 // Halt-and-catch fire here. The above call should never return! 2213 HaltNode* halt = new HaltNode(control(), frameptr()); 2214 _gvn.set_type_bottom(halt); 2215 root()->add_req(halt); 2216 2217 stop_and_kill_map(); 2218 } 2219 2220 2221 //--------------------------just_allocated_object------------------------------ 2222 // Report the object that was just allocated. 2223 // It must be the case that there are no intervening safepoints. 2224 // We use this to determine if an object is so "fresh" that 2225 // it does not require card marks. 2226 Node* GraphKit::just_allocated_object(Node* current_control) { 2227 Node* ctrl = current_control; 2228 // Object::<init> is invoked after allocation, most of invoke nodes 2229 // will be reduced, but a region node is kept in parse time, we check 2230 // the pattern and skip the region node if it degraded to a copy. 2231 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 && 2232 ctrl->as_Region()->is_copy()) { 2233 ctrl = ctrl->as_Region()->is_copy(); 2234 } 2235 if (C->recent_alloc_ctl() == ctrl) { 2236 return C->recent_alloc_obj(); 2237 } 2238 return NULL; 2239 } 2240 2241 2242 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2243 // (Note: TypeFunc::make has a cache that makes this fast.) 2244 const TypeFunc* tf = TypeFunc::make(dest_method); 2245 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2246 for (int j = 0; j < nargs; j++) { 2247 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2248 if( targ->basic_type() == T_DOUBLE ) { 2249 // If any parameters are doubles, they must be rounded before 2250 // the call, dstore_rounding does gvn.transform 2251 Node *arg = argument(j); 2252 arg = dstore_rounding(arg); 2253 set_argument(j, arg); 2254 } 2255 } 2256 } 2257 2258 /** 2259 * Record profiling data exact_kls for Node n with the type system so 2260 * that it can propagate it (speculation) 2261 * 2262 * @param n node that the type applies to 2263 * @param exact_kls type from profiling 2264 * @param maybe_null did profiling see null? 2265 * 2266 * @return node with improved type 2267 */ 2268 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2269 const Type* current_type = _gvn.type(n); 2270 assert(UseTypeSpeculation, "type speculation must be on"); 2271 2272 const TypePtr* speculative = current_type->speculative(); 2273 2274 // Should the klass from the profile be recorded in the speculative type? 2275 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2276 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2277 const TypeOopPtr* xtype = tklass->as_instance_type(); 2278 assert(xtype->klass_is_exact(), "Should be exact"); 2279 // Any reason to believe n is not null (from this profiling or a previous one)? 2280 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2281 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2282 // record the new speculative type's depth 2283 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2284 speculative = speculative->with_inline_depth(jvms()->depth()); 2285 } else if (current_type->would_improve_ptr(ptr_kind)) { 2286 // Profiling report that null was never seen so we can change the 2287 // speculative type to non null ptr. 2288 if (ptr_kind == ProfileAlwaysNull) { 2289 speculative = TypePtr::NULL_PTR; 2290 } else { 2291 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2292 const TypePtr* ptr = TypePtr::NOTNULL; 2293 if (speculative != NULL) { 2294 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2295 } else { 2296 speculative = ptr; 2297 } 2298 } 2299 } 2300 2301 if (speculative != current_type->speculative()) { 2302 // Build a type with a speculative type (what we think we know 2303 // about the type but will need a guard when we use it) 2304 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2305 // We're changing the type, we need a new CheckCast node to carry 2306 // the new type. The new type depends on the control: what 2307 // profiling tells us is only valid from here as far as we can 2308 // tell. 2309 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2310 cast = _gvn.transform(cast); 2311 replace_in_map(n, cast); 2312 n = cast; 2313 } 2314 2315 return n; 2316 } 2317 2318 /** 2319 * Record profiling data from receiver profiling at an invoke with the 2320 * type system so that it can propagate it (speculation) 2321 * 2322 * @param n receiver node 2323 * 2324 * @return node with improved type 2325 */ 2326 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2327 if (!UseTypeSpeculation) { 2328 return n; 2329 } 2330 ciKlass* exact_kls = profile_has_unique_klass(); 2331 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2332 if ((java_bc() == Bytecodes::_checkcast || 2333 java_bc() == Bytecodes::_instanceof || 2334 java_bc() == Bytecodes::_aastore) && 2335 method()->method_data()->is_mature()) { 2336 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2337 if (data != NULL) { 2338 if (!data->as_BitData()->null_seen()) { 2339 ptr_kind = ProfileNeverNull; 2340 } else { 2341 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2342 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2343 uint i = 0; 2344 for (; i < call->row_limit(); i++) { 2345 ciKlass* receiver = call->receiver(i); 2346 if (receiver != NULL) { 2347 break; 2348 } 2349 } 2350 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2351 } 2352 } 2353 } 2354 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2355 } 2356 2357 /** 2358 * Record profiling data from argument profiling at an invoke with the 2359 * type system so that it can propagate it (speculation) 2360 * 2361 * @param dest_method target method for the call 2362 * @param bc what invoke bytecode is this? 2363 */ 2364 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2365 if (!UseTypeSpeculation) { 2366 return; 2367 } 2368 const TypeFunc* tf = TypeFunc::make(dest_method); 2369 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2370 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2371 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2372 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2373 if (targ->isa_oopptr()) { 2374 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2375 ciKlass* better_type = NULL; 2376 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2377 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2378 } 2379 i++; 2380 } 2381 } 2382 } 2383 2384 /** 2385 * Record profiling data from parameter profiling at an invoke with 2386 * the type system so that it can propagate it (speculation) 2387 */ 2388 void GraphKit::record_profiled_parameters_for_speculation() { 2389 if (!UseTypeSpeculation) { 2390 return; 2391 } 2392 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2393 if (_gvn.type(local(i))->isa_oopptr()) { 2394 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2395 ciKlass* better_type = NULL; 2396 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2397 record_profile_for_speculation(local(i), better_type, ptr_kind); 2398 } 2399 j++; 2400 } 2401 } 2402 } 2403 2404 /** 2405 * Record profiling data from return value profiling at an invoke with 2406 * the type system so that it can propagate it (speculation) 2407 */ 2408 void GraphKit::record_profiled_return_for_speculation() { 2409 if (!UseTypeSpeculation) { 2410 return; 2411 } 2412 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2413 ciKlass* better_type = NULL; 2414 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2415 // If profiling reports a single type for the return value, 2416 // feed it to the type system so it can propagate it as a 2417 // speculative type 2418 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2419 } 2420 } 2421 2422 void GraphKit::round_double_result(ciMethod* dest_method) { 2423 // A non-strict method may return a double value which has an extended 2424 // exponent, but this must not be visible in a caller which is 'strict' 2425 // If a strict caller invokes a non-strict callee, round a double result 2426 2427 BasicType result_type = dest_method->return_type()->basic_type(); 2428 assert( method() != NULL, "must have caller context"); 2429 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2430 // Destination method's return value is on top of stack 2431 // dstore_rounding() does gvn.transform 2432 Node *result = pop_pair(); 2433 result = dstore_rounding(result); 2434 push_pair(result); 2435 } 2436 } 2437 2438 // rounding for strict float precision conformance 2439 Node* GraphKit::precision_rounding(Node* n) { 2440 return UseStrictFP && _method->flags().is_strict() 2441 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2442 ? _gvn.transform( new RoundFloatNode(0, n) ) 2443 : n; 2444 } 2445 2446 // rounding for strict double precision conformance 2447 Node* GraphKit::dprecision_rounding(Node *n) { 2448 return UseStrictFP && _method->flags().is_strict() 2449 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2450 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2451 : n; 2452 } 2453 2454 // rounding for non-strict double stores 2455 Node* GraphKit::dstore_rounding(Node* n) { 2456 return Matcher::strict_fp_requires_explicit_rounding 2457 && UseSSE <= 1 2458 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2459 : n; 2460 } 2461 2462 //============================================================================= 2463 // Generate a fast path/slow path idiom. Graph looks like: 2464 // [foo] indicates that 'foo' is a parameter 2465 // 2466 // [in] NULL 2467 // \ / 2468 // CmpP 2469 // Bool ne 2470 // If 2471 // / \ 2472 // True False-<2> 2473 // / | 2474 // / cast_not_null 2475 // Load | | ^ 2476 // [fast_test] | | 2477 // gvn to opt_test | | 2478 // / \ | <1> 2479 // True False | 2480 // | \\ | 2481 // [slow_call] \[fast_result] 2482 // Ctl Val \ \ 2483 // | \ \ 2484 // Catch <1> \ \ 2485 // / \ ^ \ \ 2486 // Ex No_Ex | \ \ 2487 // | \ \ | \ <2> \ 2488 // ... \ [slow_res] | | \ [null_result] 2489 // \ \--+--+--- | | 2490 // \ | / \ | / 2491 // --------Region Phi 2492 // 2493 //============================================================================= 2494 // Code is structured as a series of driver functions all called 'do_XXX' that 2495 // call a set of helper functions. Helper functions first, then drivers. 2496 2497 //------------------------------null_check_oop--------------------------------- 2498 // Null check oop. Set null-path control into Region in slot 3. 2499 // Make a cast-not-nullness use the other not-null control. Return cast. 2500 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2501 bool never_see_null, 2502 bool safe_for_replace, 2503 bool speculative) { 2504 // Initial NULL check taken path 2505 (*null_control) = top(); 2506 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2507 2508 // Generate uncommon_trap: 2509 if (never_see_null && (*null_control) != top()) { 2510 // If we see an unexpected null at a check-cast we record it and force a 2511 // recompile; the offending check-cast will be compiled to handle NULLs. 2512 // If we see more than one offending BCI, then all checkcasts in the 2513 // method will be compiled to handle NULLs. 2514 PreserveJVMState pjvms(this); 2515 set_control(*null_control); 2516 replace_in_map(value, null()); 2517 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2518 uncommon_trap(reason, 2519 Deoptimization::Action_make_not_entrant); 2520 (*null_control) = top(); // NULL path is dead 2521 } 2522 if ((*null_control) == top() && safe_for_replace) { 2523 replace_in_map(value, cast); 2524 } 2525 2526 // Cast away null-ness on the result 2527 return cast; 2528 } 2529 2530 //------------------------------opt_iff---------------------------------------- 2531 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2532 // Return slow-path control. 2533 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2534 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2535 2536 // Fast path taken; set region slot 2 2537 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2538 region->init_req(2,fast_taken); // Capture fast-control 2539 2540 // Fast path not-taken, i.e. slow path 2541 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2542 return slow_taken; 2543 } 2544 2545 //-----------------------------make_runtime_call------------------------------- 2546 Node* GraphKit::make_runtime_call(int flags, 2547 const TypeFunc* call_type, address call_addr, 2548 const char* call_name, 2549 const TypePtr* adr_type, 2550 // The following parms are all optional. 2551 // The first NULL ends the list. 2552 Node* parm0, Node* parm1, 2553 Node* parm2, Node* parm3, 2554 Node* parm4, Node* parm5, 2555 Node* parm6, Node* parm7) { 2556 assert(call_addr != NULL, "must not call NULL targets"); 2557 2558 // Slow-path call 2559 bool is_leaf = !(flags & RC_NO_LEAF); 2560 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2561 if (call_name == NULL) { 2562 assert(!is_leaf, "must supply name for leaf"); 2563 call_name = OptoRuntime::stub_name(call_addr); 2564 } 2565 CallNode* call; 2566 if (!is_leaf) { 2567 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2568 bci(), adr_type); 2569 } else if (flags & RC_NO_FP) { 2570 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2571 } else { 2572 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2573 } 2574 2575 // The following is similar to set_edges_for_java_call, 2576 // except that the memory effects of the call are restricted to AliasIdxRaw. 2577 2578 // Slow path call has no side-effects, uses few values 2579 bool wide_in = !(flags & RC_NARROW_MEM); 2580 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2581 2582 Node* prev_mem = NULL; 2583 if (wide_in) { 2584 prev_mem = set_predefined_input_for_runtime_call(call); 2585 } else { 2586 assert(!wide_out, "narrow in => narrow out"); 2587 Node* narrow_mem = memory(adr_type); 2588 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2589 } 2590 2591 // Hook each parm in order. Stop looking at the first NULL. 2592 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2593 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2594 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2595 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2596 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2597 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2598 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2599 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2600 /* close each nested if ===> */ } } } } } } } } 2601 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2602 2603 if (!is_leaf) { 2604 // Non-leaves can block and take safepoints: 2605 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2606 } 2607 // Non-leaves can throw exceptions: 2608 if (has_io) { 2609 call->set_req(TypeFunc::I_O, i_o()); 2610 } 2611 2612 if (flags & RC_UNCOMMON) { 2613 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2614 // (An "if" probability corresponds roughly to an unconditional count. 2615 // Sort of.) 2616 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2617 } 2618 2619 Node* c = _gvn.transform(call); 2620 assert(c == call, "cannot disappear"); 2621 2622 if (wide_out) { 2623 // Slow path call has full side-effects. 2624 set_predefined_output_for_runtime_call(call); 2625 } else { 2626 // Slow path call has few side-effects, and/or sets few values. 2627 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2628 } 2629 2630 if (has_io) { 2631 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2632 } 2633 return call; 2634 2635 } 2636 2637 //------------------------------merge_memory----------------------------------- 2638 // Merge memory from one path into the current memory state. 2639 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2640 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2641 Node* old_slice = mms.force_memory(); 2642 Node* new_slice = mms.memory2(); 2643 if (old_slice != new_slice) { 2644 PhiNode* phi; 2645 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2646 if (mms.is_empty()) { 2647 // clone base memory Phi's inputs for this memory slice 2648 assert(old_slice == mms.base_memory(), "sanity"); 2649 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2650 _gvn.set_type(phi, Type::MEMORY); 2651 for (uint i = 1; i < phi->req(); i++) { 2652 phi->init_req(i, old_slice->in(i)); 2653 } 2654 } else { 2655 phi = old_slice->as_Phi(); // Phi was generated already 2656 } 2657 } else { 2658 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2659 _gvn.set_type(phi, Type::MEMORY); 2660 } 2661 phi->set_req(new_path, new_slice); 2662 mms.set_memory(phi); 2663 } 2664 } 2665 } 2666 2667 //------------------------------make_slow_call_ex------------------------------ 2668 // Make the exception handler hookups for the slow call 2669 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2670 if (stopped()) return; 2671 2672 // Make a catch node with just two handlers: fall-through and catch-all 2673 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2674 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2675 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2676 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2677 2678 { PreserveJVMState pjvms(this); 2679 set_control(excp); 2680 set_i_o(i_o); 2681 2682 if (excp != top()) { 2683 if (deoptimize) { 2684 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2685 uncommon_trap(Deoptimization::Reason_unhandled, 2686 Deoptimization::Action_none); 2687 } else { 2688 // Create an exception state also. 2689 // Use an exact type if the caller has a specific exception. 2690 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2691 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2692 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2693 } 2694 } 2695 } 2696 2697 // Get the no-exception control from the CatchNode. 2698 set_control(norm); 2699 } 2700 2701 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2702 Node* cmp = NULL; 2703 switch(bt) { 2704 case T_INT: cmp = new CmpINode(in1, in2); break; 2705 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2706 default: fatal("unexpected comparison type %s", type2name(bt)); 2707 } 2708 gvn->transform(cmp); 2709 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2710 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2711 gvn->transform(iff); 2712 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2713 return iff; 2714 } 2715 2716 2717 //-------------------------------gen_subtype_check----------------------------- 2718 // Generate a subtyping check. Takes as input the subtype and supertype. 2719 // Returns 2 values: sets the default control() to the true path and returns 2720 // the false path. Only reads invariant memory; sets no (visible) memory. 2721 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2722 // but that's not exposed to the optimizer. This call also doesn't take in an 2723 // Object; if you wish to check an Object you need to load the Object's class 2724 // prior to coming here. 2725 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2726 Compile* C = gvn->C; 2727 2728 if ((*ctrl)->is_top()) { 2729 return C->top(); 2730 } 2731 2732 // Fast check for identical types, perhaps identical constants. 2733 // The types can even be identical non-constants, in cases 2734 // involving Array.newInstance, Object.clone, etc. 2735 if (subklass == superklass) 2736 return C->top(); // false path is dead; no test needed. 2737 2738 if (gvn->type(superklass)->singleton()) { 2739 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2740 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2741 2742 // In the common case of an exact superklass, try to fold up the 2743 // test before generating code. You may ask, why not just generate 2744 // the code and then let it fold up? The answer is that the generated 2745 // code will necessarily include null checks, which do not always 2746 // completely fold away. If they are also needless, then they turn 2747 // into a performance loss. Example: 2748 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2749 // Here, the type of 'fa' is often exact, so the store check 2750 // of fa[1]=x will fold up, without testing the nullness of x. 2751 switch (C->static_subtype_check(superk, subk)) { 2752 case Compile::SSC_always_false: 2753 { 2754 Node* always_fail = *ctrl; 2755 *ctrl = gvn->C->top(); 2756 return always_fail; 2757 } 2758 case Compile::SSC_always_true: 2759 return C->top(); 2760 case Compile::SSC_easy_test: 2761 { 2762 // Just do a direct pointer compare and be done. 2763 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2764 *ctrl = gvn->transform(new IfTrueNode(iff)); 2765 return gvn->transform(new IfFalseNode(iff)); 2766 } 2767 case Compile::SSC_full_test: 2768 break; 2769 default: 2770 ShouldNotReachHere(); 2771 } 2772 } 2773 2774 // %%% Possible further optimization: Even if the superklass is not exact, 2775 // if the subklass is the unique subtype of the superklass, the check 2776 // will always succeed. We could leave a dependency behind to ensure this. 2777 2778 // First load the super-klass's check-offset 2779 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2780 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2781 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2782 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2783 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2784 2785 // Load from the sub-klass's super-class display list, or a 1-word cache of 2786 // the secondary superclass list, or a failing value with a sentinel offset 2787 // if the super-klass is an interface or exceptionally deep in the Java 2788 // hierarchy and we have to scan the secondary superclass list the hard way. 2789 // Worst-case type is a little odd: NULL is allowed as a result (usually 2790 // klass loads can never produce a NULL). 2791 Node *chk_off_X = chk_off; 2792 #ifdef _LP64 2793 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2794 #endif 2795 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2796 // For some types like interfaces the following loadKlass is from a 1-word 2797 // cache which is mutable so can't use immutable memory. Other 2798 // types load from the super-class display table which is immutable. 2799 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2800 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2801 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2802 2803 // Compile speed common case: ARE a subtype and we canNOT fail 2804 if( superklass == nkls ) 2805 return C->top(); // false path is dead; no test needed. 2806 2807 // See if we get an immediate positive hit. Happens roughly 83% of the 2808 // time. Test to see if the value loaded just previously from the subklass 2809 // is exactly the superklass. 2810 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2811 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2812 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2813 2814 // Compile speed common case: Check for being deterministic right now. If 2815 // chk_off is a constant and not equal to cacheoff then we are NOT a 2816 // subklass. In this case we need exactly the 1 test above and we can 2817 // return those results immediately. 2818 if (!might_be_cache) { 2819 Node* not_subtype_ctrl = *ctrl; 2820 *ctrl = iftrue1; // We need exactly the 1 test above 2821 return not_subtype_ctrl; 2822 } 2823 2824 // Gather the various success & failures here 2825 RegionNode *r_ok_subtype = new RegionNode(4); 2826 gvn->record_for_igvn(r_ok_subtype); 2827 RegionNode *r_not_subtype = new RegionNode(3); 2828 gvn->record_for_igvn(r_not_subtype); 2829 2830 r_ok_subtype->init_req(1, iftrue1); 2831 2832 // Check for immediate negative hit. Happens roughly 11% of the time (which 2833 // is roughly 63% of the remaining cases). Test to see if the loaded 2834 // check-offset points into the subklass display list or the 1-element 2835 // cache. If it points to the display (and NOT the cache) and the display 2836 // missed then it's not a subtype. 2837 Node *cacheoff = gvn->intcon(cacheoff_con); 2838 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2839 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2840 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2841 2842 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2843 // No performance impact (too rare) but allows sharing of secondary arrays 2844 // which has some footprint reduction. 2845 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2846 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2847 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2848 2849 // -- Roads not taken here: -- 2850 // We could also have chosen to perform the self-check at the beginning 2851 // of this code sequence, as the assembler does. This would not pay off 2852 // the same way, since the optimizer, unlike the assembler, can perform 2853 // static type analysis to fold away many successful self-checks. 2854 // Non-foldable self checks work better here in second position, because 2855 // the initial primary superclass check subsumes a self-check for most 2856 // types. An exception would be a secondary type like array-of-interface, 2857 // which does not appear in its own primary supertype display. 2858 // Finally, we could have chosen to move the self-check into the 2859 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2860 // dependent manner. But it is worthwhile to have the check here, 2861 // where it can be perhaps be optimized. The cost in code space is 2862 // small (register compare, branch). 2863 2864 // Now do a linear scan of the secondary super-klass array. Again, no real 2865 // performance impact (too rare) but it's gotta be done. 2866 // Since the code is rarely used, there is no penalty for moving it 2867 // out of line, and it can only improve I-cache density. 2868 // The decision to inline or out-of-line this final check is platform 2869 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2870 Node* psc = gvn->transform( 2871 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2872 2873 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2874 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2875 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2876 2877 // Return false path; set default control to true path. 2878 *ctrl = gvn->transform(r_ok_subtype); 2879 return gvn->transform(r_not_subtype); 2880 } 2881 2882 // Profile-driven exact type check: 2883 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2884 float prob, 2885 Node* *casted_receiver) { 2886 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2887 Node* recv_klass = load_object_klass(receiver); 2888 Node* fail = type_check(recv_klass, tklass, prob); 2889 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2890 assert(recv_xtype->klass_is_exact(), ""); 2891 2892 // Subsume downstream occurrences of receiver with a cast to 2893 // recv_xtype, since now we know what the type will be. 2894 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2895 Node* res = _gvn.transform(cast); 2896 if (recv_xtype->is_valuetypeptr() && recv_xtype->value_klass()->is_scalarizable()) { 2897 assert(!gvn().type(res)->maybe_null(), "receiver should never be null"); 2898 res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass()); 2899 } 2900 2901 (*casted_receiver) = res; 2902 // (User must make the replace_in_map call.) 2903 2904 return fail; 2905 } 2906 2907 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 2908 float prob) { 2909 Node* want_klass = makecon(tklass); 2910 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass)); 2911 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2912 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2913 set_control( _gvn.transform( new IfTrueNode (iff))); 2914 Node* fail = _gvn.transform( new IfFalseNode(iff)); 2915 return fail; 2916 } 2917 2918 //------------------------------subtype_check_receiver------------------------- 2919 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 2920 Node** casted_receiver) { 2921 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2922 Node* recv_klass = load_object_klass(receiver); 2923 Node* want_klass = makecon(tklass); 2924 2925 Node* slow_ctl = gen_subtype_check(recv_klass, want_klass); 2926 2927 // Cast receiver after successful check 2928 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 2929 Node* cast = new CheckCastPPNode(control(), receiver, recv_type); 2930 (*casted_receiver) = _gvn.transform(cast); 2931 2932 return slow_ctl; 2933 } 2934 2935 //------------------------------seems_never_null------------------------------- 2936 // Use null_seen information if it is available from the profile. 2937 // If we see an unexpected null at a type check we record it and force a 2938 // recompile; the offending check will be recompiled to handle NULLs. 2939 // If we see several offending BCIs, then all checks in the 2940 // method will be recompiled. 2941 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2942 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2943 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2944 if (UncommonNullCast // Cutout for this technique 2945 && obj != null() // And not the -Xcomp stupid case? 2946 && !too_many_traps(reason) 2947 ) { 2948 if (speculating) { 2949 return true; 2950 } 2951 if (data == NULL) 2952 // Edge case: no mature data. Be optimistic here. 2953 return true; 2954 // If the profile has not seen a null, assume it won't happen. 2955 assert(java_bc() == Bytecodes::_checkcast || 2956 java_bc() == Bytecodes::_instanceof || 2957 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2958 return !data->as_BitData()->null_seen(); 2959 } 2960 speculating = false; 2961 return false; 2962 } 2963 2964 void GraphKit::guard_klass_being_initialized(Node* klass) { 2965 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 2966 Node* adr = basic_plus_adr(top(), klass, init_state_off); 2967 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr, 2968 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 2969 T_BYTE, MemNode::unordered); 2970 init_state = _gvn.transform(init_state); 2971 2972 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 2973 2974 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 2975 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 2976 2977 { BuildCutout unless(this, tst, PROB_MAX); 2978 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 2979 } 2980 } 2981 2982 void GraphKit::guard_init_thread(Node* klass) { 2983 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 2984 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 2985 2986 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr, 2987 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 2988 T_ADDRESS, MemNode::unordered); 2989 init_thread = _gvn.transform(init_thread); 2990 2991 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 2992 2993 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 2994 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 2995 2996 { BuildCutout unless(this, tst, PROB_MAX); 2997 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 2998 } 2999 } 3000 3001 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3002 if (ik->is_being_initialized()) { 3003 if (C->needs_clinit_barrier(ik, context)) { 3004 Node* klass = makecon(TypeKlassPtr::make(ik)); 3005 guard_klass_being_initialized(klass); 3006 guard_init_thread(klass); 3007 insert_mem_bar(Op_MemBarCPUOrder); 3008 } 3009 } else if (ik->is_initialized()) { 3010 return; // no barrier needed 3011 } else { 3012 uncommon_trap(Deoptimization::Reason_uninitialized, 3013 Deoptimization::Action_reinterpret, 3014 NULL); 3015 } 3016 } 3017 3018 //------------------------maybe_cast_profiled_receiver------------------------- 3019 // If the profile has seen exactly one type, narrow to exactly that type. 3020 // Subsequent type checks will always fold up. 3021 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3022 ciKlass* require_klass, 3023 ciKlass* spec_klass, 3024 bool safe_for_replace) { 3025 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 3026 3027 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 3028 3029 // Make sure we haven't already deoptimized from this tactic. 3030 if (too_many_traps_or_recompiles(reason)) 3031 return NULL; 3032 3033 // (No, this isn't a call, but it's enough like a virtual call 3034 // to use the same ciMethod accessor to get the profile info...) 3035 // If we have a speculative type use it instead of profiling (which 3036 // may not help us) 3037 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 3038 if (exact_kls != NULL) {// no cast failures here 3039 if (require_klass == NULL || 3040 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 3041 // If we narrow the type to match what the type profile sees or 3042 // the speculative type, we can then remove the rest of the 3043 // cast. 3044 // This is a win, even if the exact_kls is very specific, 3045 // because downstream operations, such as method calls, 3046 // will often benefit from the sharper type. 3047 Node* exact_obj = not_null_obj; // will get updated in place... 3048 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3049 &exact_obj); 3050 { PreserveJVMState pjvms(this); 3051 set_control(slow_ctl); 3052 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3053 } 3054 if (safe_for_replace) { 3055 replace_in_map(not_null_obj, exact_obj); 3056 } 3057 return exact_obj; 3058 } 3059 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3060 } 3061 3062 return NULL; 3063 } 3064 3065 /** 3066 * Cast obj to type and emit guard unless we had too many traps here 3067 * already 3068 * 3069 * @param obj node being casted 3070 * @param type type to cast the node to 3071 * @param not_null true if we know node cannot be null 3072 */ 3073 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3074 ciKlass* type, 3075 bool not_null) { 3076 if (stopped()) { 3077 return obj; 3078 } 3079 3080 // type == NULL if profiling tells us this object is always null 3081 if (type != NULL) { 3082 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3083 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3084 3085 if (!too_many_traps_or_recompiles(null_reason) && 3086 !too_many_traps_or_recompiles(class_reason)) { 3087 Node* not_null_obj = NULL; 3088 // not_null is true if we know the object is not null and 3089 // there's no need for a null check 3090 if (!not_null) { 3091 Node* null_ctl = top(); 3092 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3093 assert(null_ctl->is_top(), "no null control here"); 3094 } else { 3095 not_null_obj = obj; 3096 } 3097 3098 Node* exact_obj = not_null_obj; 3099 ciKlass* exact_kls = type; 3100 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3101 &exact_obj); 3102 { 3103 PreserveJVMState pjvms(this); 3104 set_control(slow_ctl); 3105 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3106 } 3107 replace_in_map(not_null_obj, exact_obj); 3108 obj = exact_obj; 3109 } 3110 } else { 3111 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3112 Node* exact_obj = null_assert(obj); 3113 replace_in_map(obj, exact_obj); 3114 obj = exact_obj; 3115 } 3116 } 3117 return obj; 3118 } 3119 3120 //-------------------------------gen_instanceof-------------------------------- 3121 // Generate an instance-of idiom. Used by both the instance-of bytecode 3122 // and the reflective instance-of call. 3123 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3124 kill_dead_locals(); // Benefit all the uncommon traps 3125 assert( !stopped(), "dead parse path should be checked in callers" ); 3126 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3127 "must check for not-null not-dead klass in callers"); 3128 3129 // Make the merge point 3130 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3131 RegionNode* region = new RegionNode(PATH_LIMIT); 3132 Node* phi = new PhiNode(region, TypeInt::BOOL); 3133 C->set_has_split_ifs(true); // Has chance for split-if optimization 3134 3135 ciProfileData* data = NULL; 3136 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3137 data = method()->method_data()->bci_to_data(bci()); 3138 } 3139 bool speculative_not_null = false; 3140 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3141 && seems_never_null(obj, data, speculative_not_null)); 3142 bool is_value = obj->is_ValueType(); 3143 3144 // Null check; get casted pointer; set region slot 3 3145 Node* null_ctl = top(); 3146 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3147 3148 // If not_null_obj is dead, only null-path is taken 3149 if (stopped()) { // Doing instance-of on a NULL? 3150 set_control(null_ctl); 3151 return intcon(0); 3152 } 3153 region->init_req(_null_path, null_ctl); 3154 phi ->init_req(_null_path, intcon(0)); // Set null path value 3155 if (null_ctl == top()) { 3156 // Do this eagerly, so that pattern matches like is_diamond_phi 3157 // will work even during parsing. 3158 assert(_null_path == PATH_LIMIT-1, "delete last"); 3159 region->del_req(_null_path); 3160 phi ->del_req(_null_path); 3161 } 3162 3163 // Do we know the type check always succeed? 3164 if (!is_value) { 3165 bool known_statically = false; 3166 if (_gvn.type(superklass)->singleton()) { 3167 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 3168 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 3169 if (subk != NULL && subk->is_loaded()) { 3170 int static_res = C->static_subtype_check(superk, subk); 3171 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3172 } 3173 } 3174 3175 if (!known_statically) { 3176 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3177 // We may not have profiling here or it may not help us. If we 3178 // have a speculative type use it to perform an exact cast. 3179 ciKlass* spec_obj_type = obj_type->speculative_type(); 3180 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 3181 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 3182 if (stopped()) { // Profile disagrees with this path. 3183 set_control(null_ctl); // Null is the only remaining possibility. 3184 return intcon(0); 3185 } 3186 if (cast_obj != NULL && 3187 // A value that's sometimes null is not something we can optimize well 3188 !(cast_obj->is_ValueType() && null_ctl != top())) { 3189 not_null_obj = cast_obj; 3190 is_value = not_null_obj->is_ValueType(); 3191 } 3192 } 3193 } 3194 } 3195 3196 // Load the object's klass 3197 Node* obj_klass = NULL; 3198 if (is_value) { 3199 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->value_klass())); 3200 } else { 3201 obj_klass = load_object_klass(not_null_obj); 3202 } 3203 3204 // Generate the subtype check 3205 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 3206 3207 // Plug in the success path to the general merge in slot 1. 3208 region->init_req(_obj_path, control()); 3209 phi ->init_req(_obj_path, intcon(1)); 3210 3211 // Plug in the failing path to the general merge in slot 2. 3212 region->init_req(_fail_path, not_subtype_ctrl); 3213 phi ->init_req(_fail_path, intcon(0)); 3214 3215 // Return final merged results 3216 set_control( _gvn.transform(region) ); 3217 record_for_igvn(region); 3218 3219 // If we know the type check always succeeds then we don't use the 3220 // profiling data at this bytecode. Don't lose it, feed it to the 3221 // type system as a speculative type. 3222 if (safe_for_replace && !is_value) { 3223 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3224 replace_in_map(obj, casted_obj); 3225 } 3226 3227 return _gvn.transform(phi); 3228 } 3229 3230 //-------------------------------gen_checkcast--------------------------------- 3231 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3232 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3233 // uncommon-trap paths work. Adjust stack after this call. 3234 // If failure_control is supplied and not null, it is filled in with 3235 // the control edge for the cast failure. Otherwise, an appropriate 3236 // uncommon trap or exception is thrown. 3237 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool never_null) { 3238 kill_dead_locals(); // Benefit all the uncommon traps 3239 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr(); 3240 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass()); 3241 assert(!never_null || toop->is_valuetypeptr(), "must be a value type pointer"); 3242 bool is_value = obj->is_ValueType(); 3243 3244 // Fast cutout: Check the case that the cast is vacuously true. 3245 // This detects the common cases where the test will short-circuit 3246 // away completely. We do this before we perform the null check, 3247 // because if the test is going to turn into zero code, we don't 3248 // want a residual null check left around. (Causes a slowdown, 3249 // for example, in some objArray manipulations, such as a[i]=a[j].) 3250 if (tk->singleton()) { 3251 ciKlass* klass = NULL; 3252 if (is_value) { 3253 klass = _gvn.type(obj)->value_klass(); 3254 } else { 3255 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3256 if (objtp != NULL) { 3257 klass = objtp->klass(); 3258 } 3259 } 3260 if (klass != NULL) { 3261 switch (C->static_subtype_check(tk->klass(), klass)) { 3262 case Compile::SSC_always_true: 3263 // If we know the type check always succeed then we don't use 3264 // the profiling data at this bytecode. Don't lose it, feed it 3265 // to the type system as a speculative type. 3266 if (!is_value) { 3267 obj = record_profiled_receiver_for_speculation(obj); 3268 if (never_null) { 3269 obj = null_check(obj); 3270 } 3271 if (toop->is_valuetypeptr() && toop->value_klass()->is_scalarizable() && !gvn().type(obj)->maybe_null()) { 3272 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass()); 3273 } 3274 } 3275 return obj; 3276 case Compile::SSC_always_false: 3277 if (is_value || never_null) { 3278 if (!is_value) { 3279 null_check(obj); 3280 } 3281 // Value type is never null. Always throw an exception. 3282 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass))); 3283 return top(); 3284 } else { 3285 // It needs a null check because a null will *pass* the cast check. 3286 return null_assert(obj); 3287 } 3288 } 3289 } 3290 } 3291 3292 ciProfileData* data = NULL; 3293 bool safe_for_replace = false; 3294 if (failure_control == NULL) { // use MDO in regular case only 3295 assert(java_bc() == Bytecodes::_aastore || 3296 java_bc() == Bytecodes::_checkcast, 3297 "interpreter profiles type checks only for these BCs"); 3298 data = method()->method_data()->bci_to_data(bci()); 3299 safe_for_replace = true; 3300 } 3301 3302 // Make the merge point 3303 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3304 RegionNode* region = new RegionNode(PATH_LIMIT); 3305 Node* phi = new PhiNode(region, toop); 3306 _gvn.set_type(region, Type::CONTROL); 3307 _gvn.set_type(phi, toop); 3308 3309 C->set_has_split_ifs(true); // Has chance for split-if optimization 3310 3311 // Use null-cast information if it is available 3312 bool speculative_not_null = false; 3313 bool never_see_null = ((failure_control == NULL) // regular case only 3314 && seems_never_null(obj, data, speculative_not_null)); 3315 3316 // Null check; get casted pointer; set region slot 3 3317 Node* null_ctl = top(); 3318 Node* not_null_obj = NULL; 3319 if (is_value) { 3320 not_null_obj = obj; 3321 } else if (never_null) { 3322 not_null_obj = null_check(obj); 3323 } else { 3324 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3325 } 3326 3327 // If not_null_obj is dead, only null-path is taken 3328 if (stopped()) { // Doing instance-of on a NULL? 3329 set_control(null_ctl); 3330 return null(); 3331 } 3332 region->init_req(_null_path, null_ctl); 3333 phi ->init_req(_null_path, null()); // Set null path value 3334 if (null_ctl == top()) { 3335 // Do this eagerly, so that pattern matches like is_diamond_phi 3336 // will work even during parsing. 3337 assert(_null_path == PATH_LIMIT-1, "delete last"); 3338 region->del_req(_null_path); 3339 phi ->del_req(_null_path); 3340 } 3341 3342 Node* cast_obj = NULL; 3343 if (!is_value && tk->klass_is_exact()) { 3344 // The following optimization tries to statically cast the speculative type of the object 3345 // (for example obtained during profiling) to the type of the superklass and then do a 3346 // dynamic check that the type of the object is what we expect. To work correctly 3347 // for checkcast and aastore the type of superklass should be exact. 3348 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3349 // We may not have profiling here or it may not help us. If we have 3350 // a speculative type use it to perform an exact cast. 3351 ciKlass* spec_obj_type = obj_type->speculative_type(); 3352 if (spec_obj_type != NULL || data != NULL) { 3353 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3354 if (cast_obj != NULL && cast_obj->is_ValueType()) { 3355 if (null_ctl != top()) { 3356 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well 3357 } else { 3358 return cast_obj; 3359 } 3360 } 3361 if (cast_obj != NULL) { 3362 if (failure_control != NULL) // failure is now impossible 3363 (*failure_control) = top(); 3364 // adjust the type of the phi to the exact klass: 3365 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3366 } 3367 } 3368 } 3369 3370 if (cast_obj == NULL) { 3371 // Load the object's klass 3372 Node* obj_klass = NULL; 3373 if (is_value) { 3374 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->value_klass())); 3375 } else { 3376 obj_klass = load_object_klass(not_null_obj); 3377 } 3378 3379 // Generate the subtype check 3380 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3381 3382 // Plug in success path into the merge 3383 cast_obj = is_value ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3384 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3385 if (failure_control == NULL) { 3386 if (not_subtype_ctrl != top()) { // If failure is possible 3387 PreserveJVMState pjvms(this); 3388 set_control(not_subtype_ctrl); 3389 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3390 } 3391 } else { 3392 (*failure_control) = not_subtype_ctrl; 3393 } 3394 } 3395 3396 region->init_req(_obj_path, control()); 3397 phi ->init_req(_obj_path, cast_obj); 3398 3399 // A merge of NULL or Casted-NotNull obj 3400 Node* res = _gvn.transform(phi); 3401 3402 // Note I do NOT always 'replace_in_map(obj,result)' here. 3403 // if( tk->klass()->can_be_primary_super() ) 3404 // This means that if I successfully store an Object into an array-of-String 3405 // I 'forget' that the Object is really now known to be a String. I have to 3406 // do this because we don't have true union types for interfaces - if I store 3407 // a Baz into an array-of-Interface and then tell the optimizer it's an 3408 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3409 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3410 // replace_in_map( obj, res ); 3411 3412 // Return final merged results 3413 set_control( _gvn.transform(region) ); 3414 record_for_igvn(region); 3415 3416 bool not_null_free = !toop->can_be_value_type(); 3417 bool not_flattenable = !ValueArrayFlatten || not_null_free || (toop->is_valuetypeptr() && !toop->value_klass()->flatten_array()); 3418 if (EnableValhalla && not_flattenable) { 3419 // Check if obj has been loaded from an array 3420 obj = obj->isa_DecodeN() ? obj->in(1) : obj; 3421 Node* array = NULL; 3422 if (obj->isa_Load()) { 3423 Node* address = obj->in(MemNode::Address); 3424 if (address->isa_AddP()) { 3425 array = address->as_AddP()->in(AddPNode::Base); 3426 } 3427 } else if (obj->is_Phi()) { 3428 Node* region = obj->in(0); 3429 if (region->req() == 3 && region->in(1) != NULL) { 3430 IfNode* iff = region->in(1)->in(0)->isa_If(); 3431 if (iff != NULL) { 3432 iff->is_flattened_array_check(&_gvn, array); 3433 } 3434 } 3435 } 3436 if (array != NULL) { 3437 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr(); 3438 if (ary_t != NULL) { 3439 if (!ary_t->is_not_null_free() && not_null_free) { 3440 // Casting array element to a non-inline-type, mark array as not null-free. 3441 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free())); 3442 replace_in_map(array, cast); 3443 } else if (!ary_t->is_not_flat()) { 3444 // Casting array element to a non-flattenable type, mark array as not flat. 3445 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat())); 3446 replace_in_map(array, cast); 3447 } 3448 } 3449 } 3450 } 3451 3452 if (!is_value) { 3453 res = record_profiled_receiver_for_speculation(res); 3454 if (toop->is_valuetypeptr() && toop->value_klass()->is_scalarizable() && !gvn().type(res)->maybe_null()) { 3455 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass()); 3456 } 3457 } 3458 return res; 3459 } 3460 3461 Node* GraphKit::is_always_locked(Node* obj) { 3462 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3463 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3464 Node* value_mask = _gvn.MakeConX(markWord::always_locked_pattern); 3465 return _gvn.transform(new AndXNode(mark, value_mask)); 3466 } 3467 3468 Node* GraphKit::is_value_mirror(Node* mirror) { 3469 Node* p = basic_plus_adr(mirror, java_lang_Class::inline_mirror_offset_in_bytes()); 3470 Node* inline_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP); 3471 Node* cmp = _gvn.transform(new CmpPNode(mirror, inline_mirror)); 3472 return _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3473 } 3474 3475 // Deoptimize if 'obj' is a value type 3476 void GraphKit::gen_value_type_guard(Node* obj, int nargs) { 3477 assert(EnableValhalla, "should only be used if value types are enabled"); 3478 Node* bol = NULL; 3479 if (obj->is_ValueTypeBase()) { 3480 bol = intcon(0); 3481 } else { 3482 Node* is_value = is_always_locked(obj); 3483 Node* value_mask = _gvn.MakeConX(markWord::always_locked_pattern); 3484 Node* cmp = _gvn.transform(new CmpXNode(is_value, value_mask)); 3485 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3486 } 3487 { BuildCutout unless(this, bol, PROB_MAX); 3488 inc_sp(nargs); 3489 uncommon_trap(Deoptimization::Reason_class_check, 3490 Deoptimization::Action_none); 3491 } 3492 } 3493 3494 // Check if 'ary' is a null-free value type array 3495 Node* GraphKit::gen_null_free_array_check(Node* ary) { 3496 assert(EnableValhalla, "should only be used if value types are enabled"); 3497 // Extract null free property from klass pointer 3498 Node* k_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes()); 3499 const TypePtr* k_adr_type = k_adr->bottom_type()->isa_ptr(); 3500 Node* klass = NULL; 3501 if (k_adr_type->is_ptr_to_narrowklass()) { 3502 klass = _gvn.transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered)); 3503 } else { 3504 klass = _gvn.transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered)); 3505 } 3506 Node* null_free = _gvn.transform(new GetNullFreePropertyNode(klass)); 3507 Node* cmp = NULL; 3508 if (_gvn.type(klass)->isa_klassptr()) { 3509 cmp = _gvn.transform(new CmpLNode(null_free, zerocon(T_LONG))); 3510 } else { 3511 cmp = _gvn.transform(new CmpINode(null_free, zerocon(T_INT))); 3512 } 3513 return _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3514 } 3515 3516 Node* GraphKit::gen_flattened_array_test(Node* ary) { 3517 assert(EnableValhalla, "should only be used if value types are enabled"); 3518 // Extract flattened property from klass pointer 3519 Node* k_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes()); 3520 const TypePtr* k_adr_type = k_adr->bottom_type()->isa_ptr(); 3521 Node* klass = NULL; 3522 if (k_adr_type->is_ptr_to_narrowklass()) { 3523 klass = _gvn.transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered)); 3524 } else { 3525 klass = _gvn.transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered)); 3526 } 3527 return _gvn.transform(new GetFlattenedPropertyNode(klass)); 3528 } 3529 3530 // Deoptimize if 'ary' is a null-free value type array and 'val' is null 3531 Node* GraphKit::gen_value_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) { 3532 const Type* val_t = _gvn.type(val); 3533 if (val->is_ValueType() || !TypePtr::NULL_PTR->higher_equal(val_t)) { 3534 return ary; // Never null 3535 } 3536 RegionNode* region = new RegionNode(3); 3537 Node* null_ctl = top(); 3538 null_check_oop(val, &null_ctl); 3539 if (null_ctl != top()) { 3540 PreserveJVMState pjvms(this); 3541 set_control(null_ctl); 3542 // Deoptimize if null-free array 3543 Node* bol = gen_null_free_array_check(ary); 3544 { BuildCutout unless(this, bol, PROB_MAX); 3545 inc_sp(nargs); 3546 uncommon_trap(Deoptimization::Reason_null_check, 3547 Deoptimization::Action_none); 3548 } 3549 region->init_req(1, control()); 3550 } 3551 region->init_req(2, control()); 3552 set_control(_gvn.transform(region)); 3553 record_for_igvn(region); 3554 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 3555 if (val_t == TypePtr::NULL_PTR && !ary_t->is_not_null_free()) { 3556 // Since we were just successfully storing null, the array can't be null free. 3557 ary_t = ary_t->cast_to_not_null_free(); 3558 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t)); 3559 if (safe_for_replace) { 3560 replace_in_map(ary, cast); 3561 } 3562 ary = cast; 3563 } 3564 return ary; 3565 } 3566 3567 Node* GraphKit::load_lh_array_tag(Node* kls) { 3568 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset())); 3569 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered)); 3570 3571 return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3572 } 3573 3574 3575 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) { 3576 Node* layout_val = load_lh_array_tag(kls); 3577 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value))); 3578 return cmp; 3579 } 3580 3581 3582 //------------------------------next_monitor----------------------------------- 3583 // What number should be given to the next monitor? 3584 int GraphKit::next_monitor() { 3585 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3586 int next = current + C->sync_stack_slots(); 3587 // Keep the toplevel high water mark current: 3588 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3589 return current; 3590 } 3591 3592 //------------------------------insert_mem_bar--------------------------------- 3593 // Memory barrier to avoid floating things around 3594 // The membar serves as a pinch point between both control and all memory slices. 3595 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3596 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3597 mb->init_req(TypeFunc::Control, control()); 3598 mb->init_req(TypeFunc::Memory, reset_memory()); 3599 Node* membar = _gvn.transform(mb); 3600 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3601 set_all_memory_call(membar); 3602 return membar; 3603 } 3604 3605 //-------------------------insert_mem_bar_volatile---------------------------- 3606 // Memory barrier to avoid floating things around 3607 // The membar serves as a pinch point between both control and memory(alias_idx). 3608 // If you want to make a pinch point on all memory slices, do not use this 3609 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3610 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3611 // When Parse::do_put_xxx updates a volatile field, it appends a series 3612 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3613 // The first membar is on the same memory slice as the field store opcode. 3614 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3615 // All the other membars (for other volatile slices, including AliasIdxBot, 3616 // which stands for all unknown volatile slices) are control-dependent 3617 // on the first membar. This prevents later volatile loads or stores 3618 // from sliding up past the just-emitted store. 3619 3620 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3621 mb->set_req(TypeFunc::Control,control()); 3622 if (alias_idx == Compile::AliasIdxBot) { 3623 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3624 } else { 3625 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3626 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3627 } 3628 Node* membar = _gvn.transform(mb); 3629 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3630 if (alias_idx == Compile::AliasIdxBot) { 3631 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3632 } else { 3633 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3634 } 3635 return membar; 3636 } 3637 3638 //------------------------------shared_lock------------------------------------ 3639 // Emit locking code. 3640 FastLockNode* GraphKit::shared_lock(Node* obj) { 3641 // bci is either a monitorenter bc or InvocationEntryBci 3642 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3643 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3644 3645 if( !GenerateSynchronizationCode ) 3646 return NULL; // Not locking things? 3647 3648 if (stopped()) // Dead monitor? 3649 return NULL; 3650 3651 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3652 3653 obj = access_resolve(obj, ACCESS_READ | ACCESS_WRITE); 3654 3655 // Box the stack location 3656 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3657 Node* mem = reset_memory(); 3658 3659 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3660 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3661 // Create the counters for this fast lock. 3662 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3663 } 3664 3665 // Create the rtm counters for this fast lock if needed. 3666 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3667 3668 // Add monitor to debug info for the slow path. If we block inside the 3669 // slow path and de-opt, we need the monitor hanging around 3670 map()->push_monitor( flock ); 3671 3672 const TypeFunc *tf = LockNode::lock_type(); 3673 LockNode *lock = new LockNode(C, tf); 3674 3675 lock->init_req( TypeFunc::Control, control() ); 3676 lock->init_req( TypeFunc::Memory , mem ); 3677 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3678 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3679 lock->init_req( TypeFunc::ReturnAdr, top() ); 3680 3681 lock->init_req(TypeFunc::Parms + 0, obj); 3682 lock->init_req(TypeFunc::Parms + 1, box); 3683 lock->init_req(TypeFunc::Parms + 2, flock); 3684 add_safepoint_edges(lock); 3685 3686 lock = _gvn.transform( lock )->as_Lock(); 3687 3688 // lock has no side-effects, sets few values 3689 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3690 3691 insert_mem_bar(Op_MemBarAcquireLock); 3692 3693 // Add this to the worklist so that the lock can be eliminated 3694 record_for_igvn(lock); 3695 3696 #ifndef PRODUCT 3697 if (PrintLockStatistics) { 3698 // Update the counter for this lock. Don't bother using an atomic 3699 // operation since we don't require absolute accuracy. 3700 lock->create_lock_counter(map()->jvms()); 3701 increment_counter(lock->counter()->addr()); 3702 } 3703 #endif 3704 3705 return flock; 3706 } 3707 3708 3709 //------------------------------shared_unlock---------------------------------- 3710 // Emit unlocking code. 3711 void GraphKit::shared_unlock(Node* box, Node* obj) { 3712 // bci is either a monitorenter bc or InvocationEntryBci 3713 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3714 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3715 3716 if( !GenerateSynchronizationCode ) 3717 return; 3718 if (stopped()) { // Dead monitor? 3719 map()->pop_monitor(); // Kill monitor from debug info 3720 return; 3721 } 3722 assert(!obj->is_ValueTypeBase(), "should not unlock on value type"); 3723 3724 // Memory barrier to avoid floating things down past the locked region 3725 insert_mem_bar(Op_MemBarReleaseLock); 3726 3727 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3728 UnlockNode *unlock = new UnlockNode(C, tf); 3729 #ifdef ASSERT 3730 unlock->set_dbg_jvms(sync_jvms()); 3731 #endif 3732 uint raw_idx = Compile::AliasIdxRaw; 3733 unlock->init_req( TypeFunc::Control, control() ); 3734 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3735 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3736 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3737 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3738 3739 unlock->init_req(TypeFunc::Parms + 0, obj); 3740 unlock->init_req(TypeFunc::Parms + 1, box); 3741 unlock = _gvn.transform(unlock)->as_Unlock(); 3742 3743 Node* mem = reset_memory(); 3744 3745 // unlock has no side-effects, sets few values 3746 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3747 3748 // Kill monitor from debug info 3749 map()->pop_monitor( ); 3750 } 3751 3752 //-------------------------------get_layout_helper----------------------------- 3753 // If the given klass is a constant or known to be an array, 3754 // fetch the constant layout helper value into constant_value 3755 // and return (Node*)NULL. Otherwise, load the non-constant 3756 // layout helper value, and return the node which represents it. 3757 // This two-faced routine is useful because allocation sites 3758 // almost always feature constant types. 3759 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3760 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3761 if (!StressReflectiveCode && inst_klass != NULL) { 3762 ciKlass* klass = inst_klass->klass(); 3763 assert(klass != NULL, "klass should not be NULL"); 3764 bool xklass = inst_klass->klass_is_exact(); 3765 bool can_be_flattened = false; 3766 if (ValueArrayFlatten && klass->is_obj_array_klass()) { 3767 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 3768 can_be_flattened = elem->is_java_lang_Object() || elem->is_interface() || (elem->is_valuetype() && !klass->as_array_klass()->storage_properties().is_null_free()); 3769 } 3770 if (xklass || (klass->is_array_klass() && !can_be_flattened)) { 3771 jint lhelper = klass->layout_helper(); 3772 if (lhelper != Klass::_lh_neutral_value) { 3773 constant_value = lhelper; 3774 return (Node*) NULL; 3775 } 3776 } 3777 } 3778 constant_value = Klass::_lh_neutral_value; // put in a known value 3779 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3780 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3781 } 3782 3783 // We just put in an allocate/initialize with a big raw-memory effect. 3784 // Hook selected additional alias categories on the initialization. 3785 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3786 MergeMemNode* init_in_merge, 3787 Node* init_out_raw) { 3788 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3789 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3790 3791 Node* prevmem = kit.memory(alias_idx); 3792 init_in_merge->set_memory_at(alias_idx, prevmem); 3793 kit.set_memory(init_out_raw, alias_idx); 3794 } 3795 3796 //---------------------------set_output_for_allocation------------------------- 3797 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3798 const TypeOopPtr* oop_type, 3799 bool deoptimize_on_exception) { 3800 int rawidx = Compile::AliasIdxRaw; 3801 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3802 add_safepoint_edges(alloc); 3803 Node* allocx = _gvn.transform(alloc); 3804 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3805 // create memory projection for i_o 3806 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3807 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3808 3809 // create a memory projection as for the normal control path 3810 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3811 set_memory(malloc, rawidx); 3812 3813 // a normal slow-call doesn't change i_o, but an allocation does 3814 // we create a separate i_o projection for the normal control path 3815 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3816 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3817 3818 // put in an initialization barrier 3819 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3820 rawoop)->as_Initialize(); 3821 assert(alloc->initialization() == init, "2-way macro link must work"); 3822 assert(init ->allocation() == alloc, "2-way macro link must work"); 3823 { 3824 // Extract memory strands which may participate in the new object's 3825 // initialization, and source them from the new InitializeNode. 3826 // This will allow us to observe initializations when they occur, 3827 // and link them properly (as a group) to the InitializeNode. 3828 assert(init->in(InitializeNode::Memory) == malloc, ""); 3829 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3830 init->set_req(InitializeNode::Memory, minit_in); 3831 record_for_igvn(minit_in); // fold it up later, if possible 3832 _gvn.set_type(minit_in, Type::MEMORY); 3833 Node* minit_out = memory(rawidx); 3834 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3835 // Add an edge in the MergeMem for the header fields so an access 3836 // to one of those has correct memory state 3837 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 3838 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 3839 if (oop_type->isa_aryptr()) { 3840 const TypeAryPtr* arytype = oop_type->is_aryptr(); 3841 if (arytype->klass()->is_value_array_klass()) { 3842 // Initially all flattened array accesses share a single slice 3843 // but that changes after parsing. Prepare the memory graph so 3844 // it can optimize flattened array accesses properly once they 3845 // don't share a single slice. 3846 assert(C->flattened_accesses_share_alias(), "should be set at parse time"); 3847 C->set_flattened_accesses_share_alias(false); 3848 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass(); 3849 ciValueKlass* vk = vak->element_klass()->as_value_klass(); 3850 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) { 3851 ciField* field = vk->nonstatic_field_at(i); 3852 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3853 continue; // do not bother to track really large numbers of fields 3854 int off_in_vt = field->offset() - vk->first_field_offset(); 3855 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot); 3856 int fieldidx = C->get_alias_index(adr_type, true); 3857 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3858 } 3859 C->set_flattened_accesses_share_alias(true); 3860 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::VALUES), minit_in, minit_out); 3861 } else { 3862 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3863 int elemidx = C->get_alias_index(telemref); 3864 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3865 } 3866 } else if (oop_type->isa_instptr()) { 3867 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word 3868 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3869 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3870 ciField* field = ik->nonstatic_field_at(i); 3871 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3872 continue; // do not bother to track really large numbers of fields 3873 // Find (or create) the alias category for this field: 3874 int fieldidx = C->alias_type(field)->index(); 3875 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3876 } 3877 } 3878 } 3879 3880 // Cast raw oop to the real thing... 3881 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3882 javaoop = _gvn.transform(javaoop); 3883 C->set_recent_alloc(control(), javaoop); 3884 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3885 3886 #ifdef ASSERT 3887 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3888 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3889 "Ideal_allocation works"); 3890 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3891 "Ideal_allocation works"); 3892 if (alloc->is_AllocateArray()) { 3893 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3894 "Ideal_allocation works"); 3895 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3896 "Ideal_allocation works"); 3897 } else { 3898 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3899 } 3900 } 3901 #endif //ASSERT 3902 3903 return javaoop; 3904 } 3905 3906 //---------------------------new_instance-------------------------------------- 3907 // This routine takes a klass_node which may be constant (for a static type) 3908 // or may be non-constant (for reflective code). It will work equally well 3909 // for either, and the graph will fold nicely if the optimizer later reduces 3910 // the type to a constant. 3911 // The optional arguments are for specialized use by intrinsics: 3912 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3913 // - If 'return_size_val', report the the total object size to the caller. 3914 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3915 Node* GraphKit::new_instance(Node* klass_node, 3916 Node* extra_slow_test, 3917 Node* *return_size_val, 3918 bool deoptimize_on_exception, 3919 ValueTypeBaseNode* value_node) { 3920 // Compute size in doublewords 3921 // The size is always an integral number of doublewords, represented 3922 // as a positive bytewise size stored in the klass's layout_helper. 3923 // The layout_helper also encodes (in a low bit) the need for a slow path. 3924 jint layout_con = Klass::_lh_neutral_value; 3925 Node* layout_val = get_layout_helper(klass_node, layout_con); 3926 bool layout_is_con = (layout_val == NULL); 3927 3928 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3929 // Generate the initial go-slow test. It's either ALWAYS (return a 3930 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3931 // case) a computed value derived from the layout_helper. 3932 Node* initial_slow_test = NULL; 3933 if (layout_is_con) { 3934 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3935 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3936 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3937 } else { // reflective case 3938 // This reflective path is used by Unsafe.allocateInstance. 3939 // (It may be stress-tested by specifying StressReflectiveCode.) 3940 // Basically, we want to get into the VM is there's an illegal argument. 3941 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3942 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3943 if (extra_slow_test != intcon(0)) { 3944 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3945 } 3946 // (Macro-expander will further convert this to a Bool, if necessary.) 3947 } 3948 3949 // Find the size in bytes. This is easy; it's the layout_helper. 3950 // The size value must be valid even if the slow path is taken. 3951 Node* size = NULL; 3952 if (layout_is_con) { 3953 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3954 } else { // reflective case 3955 // This reflective path is used by clone and Unsafe.allocateInstance. 3956 size = ConvI2X(layout_val); 3957 3958 // Clear the low bits to extract layout_helper_size_in_bytes: 3959 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3960 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3961 size = _gvn.transform( new AndXNode(size, mask) ); 3962 } 3963 if (return_size_val != NULL) { 3964 (*return_size_val) = size; 3965 } 3966 3967 // This is a precise notnull oop of the klass. 3968 // (Actually, it need not be precise if this is a reflective allocation.) 3969 // It's what we cast the result to. 3970 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3971 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3972 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3973 3974 // Now generate allocation code 3975 3976 // The entire memory state is needed for slow path of the allocation 3977 // since GC and deoptimization can happen. 3978 Node *mem = reset_memory(); 3979 set_all_memory(mem); // Create new memory state 3980 3981 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3982 control(), mem, i_o(), 3983 size, klass_node, 3984 initial_slow_test, value_node); 3985 3986 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3987 } 3988 3989 // With compressed oops, the 64 bit init value for non flattened value 3990 // arrays is built from 2 32 bit compressed oops 3991 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) { 3992 Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value)); 3993 Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32))); 3994 return kit.gvn().transform(new OrLNode(lower, upper)); 3995 } 3996 3997 //-------------------------------new_array------------------------------------- 3998 // helper for newarray and anewarray 3999 // The 'length' parameter is (obviously) the length of the array. 4000 // See comments on new_instance for the meaning of the other arguments. 4001 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 4002 Node* length, // number of array elements 4003 int nargs, // number of arguments to push back for uncommon trap 4004 Node* *return_size_val, 4005 bool deoptimize_on_exception, 4006 Node* elem_mirror) { 4007 jint layout_con = Klass::_lh_neutral_value; 4008 Node* layout_val = get_layout_helper(klass_node, layout_con); 4009 bool layout_is_con = (layout_val == NULL); 4010 4011 if (!layout_is_con && !StressReflectiveCode && 4012 !too_many_traps(Deoptimization::Reason_class_check)) { 4013 // This is a reflective array creation site. 4014 // Optimistically assume that it is a subtype of Object[], 4015 // so that we can fold up all the address arithmetic. 4016 layout_con = Klass::array_layout_helper(T_OBJECT); 4017 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 4018 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 4019 { BuildCutout unless(this, bol_lh, PROB_MAX); 4020 inc_sp(nargs); 4021 uncommon_trap(Deoptimization::Reason_class_check, 4022 Deoptimization::Action_maybe_recompile); 4023 } 4024 layout_val = NULL; 4025 layout_is_con = true; 4026 } 4027 4028 // Generate the initial go-slow test. Make sure we do not overflow 4029 // if length is huge (near 2Gig) or negative! We do not need 4030 // exact double-words here, just a close approximation of needed 4031 // double-words. We can't add any offset or rounding bits, lest we 4032 // take a size -1 of bytes and make it positive. Use an unsigned 4033 // compare, so negative sizes look hugely positive. 4034 int fast_size_limit = FastAllocateSizeLimit; 4035 if (layout_is_con) { 4036 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4037 // Increase the size limit if we have exact knowledge of array type. 4038 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 4039 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0); 4040 } 4041 4042 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 4043 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 4044 4045 // --- Size Computation --- 4046 // array_size = round_to_heap(array_header + (length << elem_shift)); 4047 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 4048 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 4049 // The rounding mask is strength-reduced, if possible. 4050 int round_mask = MinObjAlignmentInBytes - 1; 4051 Node* header_size = NULL; 4052 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 4053 // (T_BYTE has the weakest alignment and size restrictions...) 4054 if (layout_is_con) { 4055 int hsize = Klass::layout_helper_header_size(layout_con); 4056 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4057 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con); 4058 if ((round_mask & ~right_n_bits(eshift)) == 0) 4059 round_mask = 0; // strength-reduce it if it goes away completely 4060 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 4061 assert(header_size_min <= hsize, "generic minimum is smallest"); 4062 header_size_min = hsize; 4063 header_size = intcon(hsize + round_mask); 4064 } else { 4065 Node* hss = intcon(Klass::_lh_header_size_shift); 4066 Node* hsm = intcon(Klass::_lh_header_size_mask); 4067 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 4068 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 4069 Node* mask = intcon(round_mask); 4070 header_size = _gvn.transform( new AddINode(hsize, mask) ); 4071 } 4072 4073 Node* elem_shift = NULL; 4074 if (layout_is_con) { 4075 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4076 if (eshift != 0) 4077 elem_shift = intcon(eshift); 4078 } else { 4079 // There is no need to mask or shift this value. 4080 // The semantics of LShiftINode include an implicit mask to 0x1F. 4081 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 4082 elem_shift = layout_val; 4083 } 4084 4085 // Transition to native address size for all offset calculations: 4086 Node* lengthx = ConvI2X(length); 4087 Node* headerx = ConvI2X(header_size); 4088 #ifdef _LP64 4089 { const TypeInt* tilen = _gvn.find_int_type(length); 4090 if (tilen != NULL && tilen->_lo < 0) { 4091 // Add a manual constraint to a positive range. Cf. array_element_address. 4092 jint size_max = fast_size_limit; 4093 if (size_max > tilen->_hi) size_max = tilen->_hi; 4094 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 4095 4096 // Only do a narrow I2L conversion if the range check passed. 4097 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 4098 _gvn.transform(iff); 4099 RegionNode* region = new RegionNode(3); 4100 _gvn.set_type(region, Type::CONTROL); 4101 lengthx = new PhiNode(region, TypeLong::LONG); 4102 _gvn.set_type(lengthx, TypeLong::LONG); 4103 4104 // Range check passed. Use ConvI2L node with narrow type. 4105 Node* passed = IfFalse(iff); 4106 region->init_req(1, passed); 4107 // Make I2L conversion control dependent to prevent it from 4108 // floating above the range check during loop optimizations. 4109 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 4110 4111 // Range check failed. Use ConvI2L with wide type because length may be invalid. 4112 region->init_req(2, IfTrue(iff)); 4113 lengthx->init_req(2, ConvI2X(length)); 4114 4115 set_control(region); 4116 record_for_igvn(region); 4117 record_for_igvn(lengthx); 4118 } 4119 } 4120 #endif 4121 4122 // Combine header size (plus rounding) and body size. Then round down. 4123 // This computation cannot overflow, because it is used only in two 4124 // places, one where the length is sharply limited, and the other 4125 // after a successful allocation. 4126 Node* abody = lengthx; 4127 if (elem_shift != NULL) 4128 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 4129 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 4130 if (round_mask != 0) { 4131 Node* mask = MakeConX(~round_mask); 4132 size = _gvn.transform( new AndXNode(size, mask) ); 4133 } 4134 // else if round_mask == 0, the size computation is self-rounding 4135 4136 if (return_size_val != NULL) { 4137 // This is the size 4138 (*return_size_val) = size; 4139 } 4140 4141 // Now generate allocation code 4142 4143 // The entire memory state is needed for slow path of the allocation 4144 // since GC and deoptimization can happen. 4145 Node *mem = reset_memory(); 4146 set_all_memory(mem); // Create new memory state 4147 4148 if (initial_slow_test->is_Bool()) { 4149 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4150 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4151 } 4152 4153 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 4154 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 4155 const Type* elem = NULL; 4156 ciKlass* elem_klass = NULL; 4157 4158 // Compute default value and storage properties for value type arrays: 4159 // - null-ok: MyValue.box[] (ciObjArrayKlass "[LMyValue") 4160 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue") 4161 // - null-free, flattened: MyValue.val[] (ciValueArrayKlass "[QMyValue") 4162 Node* storage_properties = NULL; 4163 Node* default_value = NULL; 4164 Node* raw_default_value = NULL; 4165 int props_shift = UseCompressedClassPointers ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift; 4166 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) { 4167 // Array type is known 4168 elem = ary_ptr->elem(); 4169 ciArrayKlass* ary_klass = ary_ptr->klass()->as_array_klass(); 4170 elem_klass = ary_klass->element_klass(); 4171 4172 ArrayStorageProperties props = ary_klass->storage_properties(); 4173 if (!props.is_empty() && elem_klass->is_valuetype()) { 4174 if (props.is_null_free() && !props.is_flattened()) { 4175 default_value = ValueTypeNode::default_oop(gvn(), elem_klass->as_value_klass()); 4176 if (elem->isa_narrowoop()) { 4177 default_value = _gvn.transform(new EncodePNode(default_value, elem)); 4178 raw_default_value = raw_default_for_coops(default_value, *this); 4179 } else { 4180 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value)); 4181 } 4182 } 4183 storage_properties = MakeConX(props.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift)); 4184 } 4185 } 4186 4187 if (EnableValhalla && (elem == NULL || (elem_klass != NULL && (elem_klass->is_java_lang_Object() || elem_klass->is_valuetype()) && 4188 !ary_type->klass_is_exact()))) { 4189 // Array type is not known, compute default value and storage properties for initialization. 4190 assert(default_value == NULL && raw_default_value == NULL && storage_properties == NULL, "shouldn't be set yet"); 4191 assert(elem_mirror != NULL, "should not be null"); 4192 4193 Node* r = new RegionNode(4); 4194 default_value = new PhiNode(r, TypeInstPtr::BOTTOM); 4195 storage_properties = new PhiNode(r, TypeX_X); 4196 4197 Node* empty = MakeConX(ArrayStorageProperties::empty.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift)); 4198 Node* null_free = MakeConX(ArrayStorageProperties::null_free.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift)); 4199 Node* flat = MakeConX(ArrayStorageProperties::flattened_and_null_free.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift)); 4200 4201 // Check if element mirror is a value mirror 4202 IfNode* iff = create_and_map_if(control(), is_value_mirror(elem_mirror), PROB_FAIR, COUNT_UNKNOWN); 4203 4204 // Not a value mirror but a box mirror or not a value type array, initialize with all zero 4205 r->init_req(1, _gvn.transform(new IfFalseNode(iff))); 4206 default_value->init_req(1, null()); 4207 storage_properties->init_req(1, empty); 4208 4209 // Value mirror (= null-free), check if flattened 4210 set_control(_gvn.transform(new IfTrueNode(iff))); 4211 Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_vt_value); 4212 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 4213 iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4214 4215 // Flattened, initialize with all zero 4216 r->init_req(2, _gvn.transform(new IfTrueNode(iff))); 4217 default_value->init_req(2, null()); 4218 storage_properties->init_req(2, flat); 4219 4220 // Non-flattened, initialize with the default value 4221 set_control(_gvn.transform(new IfFalseNode(iff))); 4222 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset())); 4223 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS)); 4224 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())); 4225 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4226 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(ValueKlass::default_value_offset_offset())); 4227 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered); 4228 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset)); 4229 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP); 4230 r->init_req(3, control()); 4231 default_value->init_req(3, val); 4232 storage_properties->init_req(3, null_free); 4233 4234 set_control(_gvn.transform(r)); 4235 default_value = _gvn.transform(default_value); 4236 storage_properties = _gvn.transform(storage_properties); 4237 if (UseCompressedOops) { 4238 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop())); 4239 raw_default_value = raw_default_for_coops(default_value, *this); 4240 } else { 4241 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value)); 4242 } 4243 } 4244 4245 // Create the AllocateArrayNode and its result projections 4246 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4247 control(), mem, i_o(), 4248 size, klass_node, 4249 initial_slow_test, 4250 length, default_value, 4251 raw_default_value, 4252 storage_properties); 4253 4254 // Cast to correct type. Note that the klass_node may be constant or not, 4255 // and in the latter case the actual array type will be inexact also. 4256 // (This happens via a non-constant argument to inline_native_newArray.) 4257 // In any case, the value of klass_node provides the desired array type. 4258 const TypeInt* length_type = _gvn.find_int_type(length); 4259 if (ary_type->isa_aryptr() && length_type != NULL) { 4260 // Try to get a better type than POS for the size 4261 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4262 } 4263 4264 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4265 4266 // Cast length on remaining path to be as narrow as possible 4267 if (map()->find_edge(length) >= 0) { 4268 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 4269 if (ccast != length) { 4270 _gvn.set_type_bottom(ccast); 4271 record_for_igvn(ccast); 4272 replace_in_map(length, ccast); 4273 } 4274 } 4275 4276 return javaoop; 4277 } 4278 4279 // The following "Ideal_foo" functions are placed here because they recognize 4280 // the graph shapes created by the functions immediately above. 4281 4282 //---------------------------Ideal_allocation---------------------------------- 4283 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4284 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 4285 if (ptr == NULL) { // reduce dumb test in callers 4286 return NULL; 4287 } 4288 4289 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4290 ptr = bs->step_over_gc_barrier(ptr); 4291 4292 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4293 ptr = ptr->in(1); 4294 if (ptr == NULL) return NULL; 4295 } 4296 // Return NULL for allocations with several casts: 4297 // j.l.reflect.Array.newInstance(jobject, jint) 4298 // Object.clone() 4299 // to keep more precise type from last cast. 4300 if (ptr->is_Proj()) { 4301 Node* allo = ptr->in(0); 4302 if (allo != NULL && allo->is_Allocate()) { 4303 return allo->as_Allocate(); 4304 } 4305 } 4306 // Report failure to match. 4307 return NULL; 4308 } 4309 4310 // Fancy version which also strips off an offset (and reports it to caller). 4311 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 4312 intptr_t& offset) { 4313 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4314 if (base == NULL) return NULL; 4315 return Ideal_allocation(base, phase); 4316 } 4317 4318 // Trace Initialize <- Proj[Parm] <- Allocate 4319 AllocateNode* InitializeNode::allocation() { 4320 Node* rawoop = in(InitializeNode::RawAddress); 4321 if (rawoop->is_Proj()) { 4322 Node* alloc = rawoop->in(0); 4323 if (alloc->is_Allocate()) { 4324 return alloc->as_Allocate(); 4325 } 4326 } 4327 return NULL; 4328 } 4329 4330 // Trace Allocate -> Proj[Parm] -> Initialize 4331 InitializeNode* AllocateNode::initialization() { 4332 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4333 if (rawoop == NULL) return NULL; 4334 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4335 Node* init = rawoop->fast_out(i); 4336 if (init->is_Initialize()) { 4337 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4338 return init->as_Initialize(); 4339 } 4340 } 4341 return NULL; 4342 } 4343 4344 //----------------------------- loop predicates --------------------------- 4345 4346 //------------------------------add_predicate_impl---------------------------- 4347 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 4348 // Too many traps seen? 4349 if (too_many_traps(reason)) { 4350 #ifdef ASSERT 4351 if (TraceLoopPredicate) { 4352 int tc = C->trap_count(reason); 4353 tty->print("too many traps=%s tcount=%d in ", 4354 Deoptimization::trap_reason_name(reason), tc); 4355 method()->print(); // which method has too many predicate traps 4356 tty->cr(); 4357 } 4358 #endif 4359 // We cannot afford to take more traps here, 4360 // do not generate predicate. 4361 return; 4362 } 4363 4364 Node *cont = _gvn.intcon(1); 4365 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 4366 Node *bol = _gvn.transform(new Conv2BNode(opq)); 4367 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4368 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4369 C->add_predicate_opaq(opq); 4370 { 4371 PreserveJVMState pjvms(this); 4372 set_control(iffalse); 4373 inc_sp(nargs); 4374 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4375 } 4376 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4377 set_control(iftrue); 4378 } 4379 4380 //------------------------------add_predicate--------------------------------- 4381 void GraphKit::add_predicate(int nargs) { 4382 if (UseLoopPredicate) { 4383 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 4384 } 4385 if (UseProfiledLoopPredicate) { 4386 add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs); 4387 } 4388 // loop's limit check predicate should be near the loop. 4389 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 4390 } 4391 4392 void GraphKit::sync_kit(IdealKit& ideal) { 4393 set_all_memory(ideal.merged_memory()); 4394 set_i_o(ideal.i_o()); 4395 set_control(ideal.ctrl()); 4396 } 4397 4398 void GraphKit::final_sync(IdealKit& ideal) { 4399 // Final sync IdealKit and graphKit. 4400 sync_kit(ideal); 4401 } 4402 4403 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4404 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4405 Node* coder = load_String_coder(str, set_ctrl); 4406 // Divide length by 2 if coder is UTF16 4407 return _gvn.transform(new RShiftINode(len, coder)); 4408 } 4409 4410 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4411 int value_offset = java_lang_String::value_offset_in_bytes(); 4412 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4413 false, NULL, Type::Offset(0), false); 4414 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4415 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4416 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true), 4417 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4418 Node* p = basic_plus_adr(str, str, value_offset); 4419 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4420 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4421 return load; 4422 } 4423 4424 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4425 if (!CompactStrings) { 4426 return intcon(java_lang_String::CODER_UTF16); 4427 } 4428 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4429 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4430 false, NULL, Type::Offset(0), false); 4431 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4432 4433 Node* p = basic_plus_adr(str, str, coder_offset); 4434 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4435 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4436 return load; 4437 } 4438 4439 void GraphKit::store_String_value(Node* str, Node* value) { 4440 int value_offset = java_lang_String::value_offset_in_bytes(); 4441 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4442 false, NULL, Type::Offset(0), false); 4443 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4444 4445 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4446 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4447 } 4448 4449 void GraphKit::store_String_coder(Node* str, Node* value) { 4450 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4451 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4452 false, NULL, Type::Offset(0), false); 4453 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4454 4455 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4456 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4457 } 4458 4459 // Capture src and dst memory state with a MergeMemNode 4460 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4461 if (src_type == dst_type) { 4462 // Types are equal, we don't need a MergeMemNode 4463 return memory(src_type); 4464 } 4465 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4466 record_for_igvn(merge); // fold it up later, if possible 4467 int src_idx = C->get_alias_index(src_type); 4468 int dst_idx = C->get_alias_index(dst_type); 4469 merge->set_memory_at(src_idx, memory(src_idx)); 4470 merge->set_memory_at(dst_idx, memory(dst_idx)); 4471 return merge; 4472 } 4473 4474 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4475 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4476 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4477 // If input and output memory types differ, capture both states to preserve 4478 // the dependency between preceding and subsequent loads/stores. 4479 // For example, the following program: 4480 // StoreB 4481 // compress_string 4482 // LoadB 4483 // has this memory graph (use->def): 4484 // LoadB -> compress_string -> CharMem 4485 // ... -> StoreB -> ByteMem 4486 // The intrinsic hides the dependency between LoadB and StoreB, causing 4487 // the load to read from memory not containing the result of the StoreB. 4488 // The correct memory graph should look like this: 4489 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4490 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4491 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4492 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4493 set_memory(res_mem, TypeAryPtr::BYTES); 4494 return str; 4495 } 4496 4497 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4498 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4499 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4500 // Capture src and dst memory (see comment in 'compress_string'). 4501 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4502 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4503 set_memory(_gvn.transform(str), dst_type); 4504 } 4505 4506 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4507 /** 4508 * int i_char = start; 4509 * for (int i_byte = 0; i_byte < count; i_byte++) { 4510 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4511 * } 4512 */ 4513 src = access_resolve(src, ACCESS_READ); 4514 dst = access_resolve(dst, ACCESS_WRITE); 4515 add_predicate(); 4516 RegionNode* head = new RegionNode(3); 4517 head->init_req(1, control()); 4518 gvn().set_type(head, Type::CONTROL); 4519 record_for_igvn(head); 4520 4521 Node* i_byte = new PhiNode(head, TypeInt::INT); 4522 i_byte->init_req(1, intcon(0)); 4523 gvn().set_type(i_byte, TypeInt::INT); 4524 record_for_igvn(i_byte); 4525 4526 Node* i_char = new PhiNode(head, TypeInt::INT); 4527 i_char->init_req(1, start); 4528 gvn().set_type(i_char, TypeInt::INT); 4529 record_for_igvn(i_char); 4530 4531 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4532 gvn().set_type(mem, Type::MEMORY); 4533 record_for_igvn(mem); 4534 set_control(head); 4535 set_memory(mem, TypeAryPtr::BYTES); 4536 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4537 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4538 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4539 false, false, true /* mismatched */); 4540 4541 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4542 head->init_req(2, IfTrue(iff)); 4543 mem->init_req(2, st); 4544 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4545 i_char->init_req(2, AddI(i_char, intcon(2))); 4546 4547 set_control(IfFalse(iff)); 4548 set_memory(st, TypeAryPtr::BYTES); 4549 } 4550 4551 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4552 if (!field->is_constant()) { 4553 return NULL; // Field not marked as constant. 4554 } 4555 ciInstance* holder = NULL; 4556 if (!field->is_static()) { 4557 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4558 if (const_oop != NULL && const_oop->is_instance()) { 4559 holder = const_oop->as_instance(); 4560 } 4561 } 4562 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4563 /*is_unsigned_load=*/false); 4564 if (con_type != NULL) { 4565 Node* con = makecon(con_type); 4566 if (field->layout_type() == T_VALUETYPE && field->type()->as_value_klass()->is_scalarizable() && !con_type->maybe_null()) { 4567 // Load value type from constant oop 4568 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass()); 4569 } 4570 return con; 4571 } 4572 return NULL; 4573 } 4574 4575 //---------------------------load_mirror_from_klass---------------------------- 4576 // Given a klass oop, load its java mirror (a java.lang.Class oop). 4577 Node* GraphKit::load_mirror_from_klass(Node* klass) { 4578 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 4579 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4580 // mirror = ((OopHandle)mirror)->resolve(); 4581 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 4582 }