1 /* 2 * Copyright (c) 2001, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciFlatArrayKlass.hpp" 27 #include "ci/ciInlineKlass.hpp" 28 #include "ci/ciUtilities.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "ci/ciObjArray.hpp" 31 #include "asm/register.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "gc/shared/barrierSet.hpp" 34 #include "gc/shared/c2/barrierSetC2.hpp" 35 #include "interpreter/interpreter.hpp" 36 #include "memory/resourceArea.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/castnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/graphKit.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/inlinetypenode.hpp" 43 #include "opto/intrinsicnode.hpp" 44 #include "opto/locknode.hpp" 45 #include "opto/machnode.hpp" 46 #include "opto/narrowptrnode.hpp" 47 #include "opto/opaquenode.hpp" 48 #include "opto/parse.hpp" 49 #include "opto/rootnode.hpp" 50 #include "opto/runtime.hpp" 51 #include "opto/subtypenode.hpp" 52 #include "runtime/deoptimization.hpp" 53 #include "runtime/sharedRuntime.hpp" 54 #include "utilities/bitMap.inline.hpp" 55 #include "utilities/powerOfTwo.hpp" 56 #include "utilities/growableArray.hpp" 57 58 //----------------------------GraphKit----------------------------------------- 59 // Main utility constructor. 60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 61 : Phase(Phase::Parser), 62 _env(C->env()), 63 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()), 64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 65 { 66 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled"); 67 _exceptions = jvms->map()->next_exception(); 68 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 69 set_jvms(jvms); 70 #ifdef ASSERT 71 if (_gvn.is_IterGVN() != NULL) { 72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used"); 73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 74 _worklist_size = _gvn.C->for_igvn()->size(); 75 } 76 #endif 77 } 78 79 // Private constructor for parser. 80 GraphKit::GraphKit() 81 : Phase(Phase::Parser), 82 _env(C->env()), 83 _gvn(*C->initial_gvn()), 84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 85 { 86 _exceptions = NULL; 87 set_map(NULL); 88 debug_only(_sp = -99); 89 debug_only(set_bci(-99)); 90 } 91 92 93 94 //---------------------------clean_stack--------------------------------------- 95 // Clear away rubbish from the stack area of the JVM state. 96 // This destroys any arguments that may be waiting on the stack. 97 void GraphKit::clean_stack(int from_sp) { 98 SafePointNode* map = this->map(); 99 JVMState* jvms = this->jvms(); 100 int stk_size = jvms->stk_size(); 101 int stkoff = jvms->stkoff(); 102 Node* top = this->top(); 103 for (int i = from_sp; i < stk_size; i++) { 104 if (map->in(stkoff + i) != top) { 105 map->set_req(stkoff + i, top); 106 } 107 } 108 } 109 110 111 //--------------------------------sync_jvms----------------------------------- 112 // Make sure our current jvms agrees with our parse state. 113 JVMState* GraphKit::sync_jvms() const { 114 JVMState* jvms = this->jvms(); 115 jvms->set_bci(bci()); // Record the new bci in the JVMState 116 jvms->set_sp(sp()); // Record the new sp in the JVMState 117 assert(jvms_in_sync(), "jvms is now in sync"); 118 return jvms; 119 } 120 121 //--------------------------------sync_jvms_for_reexecute--------------------- 122 // Make sure our current jvms agrees with our parse state. This version 123 // uses the reexecute_sp for reexecuting bytecodes. 124 JVMState* GraphKit::sync_jvms_for_reexecute() { 125 JVMState* jvms = this->jvms(); 126 jvms->set_bci(bci()); // Record the new bci in the JVMState 127 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 128 return jvms; 129 } 130 131 #ifdef ASSERT 132 bool GraphKit::jvms_in_sync() const { 133 Parse* parse = is_Parse(); 134 if (parse == NULL) { 135 if (bci() != jvms()->bci()) return false; 136 if (sp() != (int)jvms()->sp()) return false; 137 return true; 138 } 139 if (jvms()->method() != parse->method()) return false; 140 if (jvms()->bci() != parse->bci()) return false; 141 int jvms_sp = jvms()->sp(); 142 if (jvms_sp != parse->sp()) return false; 143 int jvms_depth = jvms()->depth(); 144 if (jvms_depth != parse->depth()) return false; 145 return true; 146 } 147 148 // Local helper checks for special internal merge points 149 // used to accumulate and merge exception states. 150 // They are marked by the region's in(0) edge being the map itself. 151 // Such merge points must never "escape" into the parser at large, 152 // until they have been handed to gvn.transform. 153 static bool is_hidden_merge(Node* reg) { 154 if (reg == NULL) return false; 155 if (reg->is_Phi()) { 156 reg = reg->in(0); 157 if (reg == NULL) return false; 158 } 159 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 160 } 161 162 void GraphKit::verify_map() const { 163 if (map() == NULL) return; // null map is OK 164 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 165 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 166 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 167 } 168 169 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 170 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 171 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 172 } 173 #endif 174 175 //---------------------------stop_and_kill_map--------------------------------- 176 // Set _map to NULL, signalling a stop to further bytecode execution. 177 // First smash the current map's control to a constant, to mark it dead. 178 void GraphKit::stop_and_kill_map() { 179 SafePointNode* dead_map = stop(); 180 if (dead_map != NULL) { 181 dead_map->disconnect_inputs(C); // Mark the map as killed. 182 assert(dead_map->is_killed(), "must be so marked"); 183 } 184 } 185 186 187 //--------------------------------stopped-------------------------------------- 188 // Tell if _map is NULL, or control is top. 189 bool GraphKit::stopped() { 190 if (map() == NULL) return true; 191 else if (control() == top()) return true; 192 else return false; 193 } 194 195 196 //-----------------------------has_ex_handler---------------------------------- 197 // Tell if this method or any caller method has exception handlers. 198 bool GraphKit::has_ex_handler() { 199 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 200 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 201 return true; 202 } 203 } 204 return false; 205 } 206 207 //------------------------------save_ex_oop------------------------------------ 208 // Save an exception without blowing stack contents or other JVM state. 209 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 210 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 211 ex_map->add_req(ex_oop); 212 debug_only(verify_exception_state(ex_map)); 213 } 214 215 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 216 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 217 Node* ex_oop = ex_map->in(ex_map->req()-1); 218 if (clear_it) ex_map->del_req(ex_map->req()-1); 219 return ex_oop; 220 } 221 222 //-----------------------------saved_ex_oop------------------------------------ 223 // Recover a saved exception from its map. 224 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 225 return common_saved_ex_oop(ex_map, false); 226 } 227 228 //--------------------------clear_saved_ex_oop--------------------------------- 229 // Erase a previously saved exception from its map. 230 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 231 return common_saved_ex_oop(ex_map, true); 232 } 233 234 #ifdef ASSERT 235 //---------------------------has_saved_ex_oop---------------------------------- 236 // Erase a previously saved exception from its map. 237 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 238 return ex_map->req() == ex_map->jvms()->endoff()+1; 239 } 240 #endif 241 242 //-------------------------make_exception_state-------------------------------- 243 // Turn the current JVM state into an exception state, appending the ex_oop. 244 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 245 sync_jvms(); 246 SafePointNode* ex_map = stop(); // do not manipulate this map any more 247 set_saved_ex_oop(ex_map, ex_oop); 248 return ex_map; 249 } 250 251 252 //--------------------------add_exception_state-------------------------------- 253 // Add an exception to my list of exceptions. 254 void GraphKit::add_exception_state(SafePointNode* ex_map) { 255 if (ex_map == NULL || ex_map->control() == top()) { 256 return; 257 } 258 #ifdef ASSERT 259 verify_exception_state(ex_map); 260 if (has_exceptions()) { 261 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 262 } 263 #endif 264 265 // If there is already an exception of exactly this type, merge with it. 266 // In particular, null-checks and other low-level exceptions common up here. 267 Node* ex_oop = saved_ex_oop(ex_map); 268 const Type* ex_type = _gvn.type(ex_oop); 269 if (ex_oop == top()) { 270 // No action needed. 271 return; 272 } 273 assert(ex_type->isa_instptr(), "exception must be an instance"); 274 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 275 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 276 // We check sp also because call bytecodes can generate exceptions 277 // both before and after arguments are popped! 278 if (ex_type2 == ex_type 279 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 280 combine_exception_states(ex_map, e2); 281 return; 282 } 283 } 284 285 // No pre-existing exception of the same type. Chain it on the list. 286 push_exception_state(ex_map); 287 } 288 289 //-----------------------add_exception_states_from----------------------------- 290 void GraphKit::add_exception_states_from(JVMState* jvms) { 291 SafePointNode* ex_map = jvms->map()->next_exception(); 292 if (ex_map != NULL) { 293 jvms->map()->set_next_exception(NULL); 294 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 295 next_map = ex_map->next_exception(); 296 ex_map->set_next_exception(NULL); 297 add_exception_state(ex_map); 298 } 299 } 300 } 301 302 //-----------------------transfer_exceptions_into_jvms------------------------- 303 JVMState* GraphKit::transfer_exceptions_into_jvms() { 304 if (map() == NULL) { 305 // We need a JVMS to carry the exceptions, but the map has gone away. 306 // Create a scratch JVMS, cloned from any of the exception states... 307 if (has_exceptions()) { 308 _map = _exceptions; 309 _map = clone_map(); 310 _map->set_next_exception(NULL); 311 clear_saved_ex_oop(_map); 312 debug_only(verify_map()); 313 } else { 314 // ...or created from scratch 315 JVMState* jvms = new (C) JVMState(_method, NULL); 316 jvms->set_bci(_bci); 317 jvms->set_sp(_sp); 318 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 319 set_jvms(jvms); 320 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 321 set_all_memory(top()); 322 while (map()->req() < jvms->endoff()) map()->add_req(top()); 323 } 324 // (This is a kludge, in case you didn't notice.) 325 set_control(top()); 326 } 327 JVMState* jvms = sync_jvms(); 328 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 329 jvms->map()->set_next_exception(_exceptions); 330 _exceptions = NULL; // done with this set of exceptions 331 return jvms; 332 } 333 334 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 335 assert(is_hidden_merge(dstphi), "must be a special merge node"); 336 assert(is_hidden_merge(srcphi), "must be a special merge node"); 337 uint limit = srcphi->req(); 338 for (uint i = PhiNode::Input; i < limit; i++) { 339 dstphi->add_req(srcphi->in(i)); 340 } 341 } 342 static inline void add_one_req(Node* dstphi, Node* src) { 343 assert(is_hidden_merge(dstphi), "must be a special merge node"); 344 assert(!is_hidden_merge(src), "must not be a special merge node"); 345 dstphi->add_req(src); 346 } 347 348 //-----------------------combine_exception_states------------------------------ 349 // This helper function combines exception states by building phis on a 350 // specially marked state-merging region. These regions and phis are 351 // untransformed, and can build up gradually. The region is marked by 352 // having a control input of its exception map, rather than NULL. Such 353 // regions do not appear except in this function, and in use_exception_state. 354 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 355 if (failing()) return; // dying anyway... 356 JVMState* ex_jvms = ex_map->_jvms; 357 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 358 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 359 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 360 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 361 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 362 assert(ex_map->req() == phi_map->req(), "matching maps"); 363 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 364 Node* hidden_merge_mark = root(); 365 Node* region = phi_map->control(); 366 MergeMemNode* phi_mem = phi_map->merged_memory(); 367 MergeMemNode* ex_mem = ex_map->merged_memory(); 368 if (region->in(0) != hidden_merge_mark) { 369 // The control input is not (yet) a specially-marked region in phi_map. 370 // Make it so, and build some phis. 371 region = new RegionNode(2); 372 _gvn.set_type(region, Type::CONTROL); 373 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 374 region->init_req(1, phi_map->control()); 375 phi_map->set_control(region); 376 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 377 record_for_igvn(io_phi); 378 _gvn.set_type(io_phi, Type::ABIO); 379 phi_map->set_i_o(io_phi); 380 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 381 Node* m = mms.memory(); 382 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 383 record_for_igvn(m_phi); 384 _gvn.set_type(m_phi, Type::MEMORY); 385 mms.set_memory(m_phi); 386 } 387 } 388 389 // Either or both of phi_map and ex_map might already be converted into phis. 390 Node* ex_control = ex_map->control(); 391 // if there is special marking on ex_map also, we add multiple edges from src 392 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 393 // how wide was the destination phi_map, originally? 394 uint orig_width = region->req(); 395 396 if (add_multiple) { 397 add_n_reqs(region, ex_control); 398 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 399 } else { 400 // ex_map has no merges, so we just add single edges everywhere 401 add_one_req(region, ex_control); 402 add_one_req(phi_map->i_o(), ex_map->i_o()); 403 } 404 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 405 if (mms.is_empty()) { 406 // get a copy of the base memory, and patch some inputs into it 407 const TypePtr* adr_type = mms.adr_type(C); 408 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 409 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 410 mms.set_memory(phi); 411 // Prepare to append interesting stuff onto the newly sliced phi: 412 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 413 } 414 // Append stuff from ex_map: 415 if (add_multiple) { 416 add_n_reqs(mms.memory(), mms.memory2()); 417 } else { 418 add_one_req(mms.memory(), mms.memory2()); 419 } 420 } 421 uint limit = ex_map->req(); 422 for (uint i = TypeFunc::Parms; i < limit; i++) { 423 // Skip everything in the JVMS after tos. (The ex_oop follows.) 424 if (i == tos) i = ex_jvms->monoff(); 425 Node* src = ex_map->in(i); 426 Node* dst = phi_map->in(i); 427 if (src != dst) { 428 PhiNode* phi; 429 if (dst->in(0) != region) { 430 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 431 record_for_igvn(phi); 432 _gvn.set_type(phi, phi->type()); 433 phi_map->set_req(i, dst); 434 // Prepare to append interesting stuff onto the new phi: 435 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 436 } else { 437 assert(dst->is_Phi(), "nobody else uses a hidden region"); 438 phi = dst->as_Phi(); 439 } 440 if (add_multiple && src->in(0) == ex_control) { 441 // Both are phis. 442 add_n_reqs(dst, src); 443 } else { 444 while (dst->req() < region->req()) add_one_req(dst, src); 445 } 446 const Type* srctype = _gvn.type(src); 447 if (phi->type() != srctype) { 448 const Type* dsttype = phi->type()->meet_speculative(srctype); 449 if (phi->type() != dsttype) { 450 phi->set_type(dsttype); 451 _gvn.set_type(phi, dsttype); 452 } 453 } 454 } 455 } 456 phi_map->merge_replaced_nodes_with(ex_map); 457 } 458 459 //--------------------------use_exception_state-------------------------------- 460 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 461 if (failing()) { stop(); return top(); } 462 Node* region = phi_map->control(); 463 Node* hidden_merge_mark = root(); 464 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 465 Node* ex_oop = clear_saved_ex_oop(phi_map); 466 if (region->in(0) == hidden_merge_mark) { 467 // Special marking for internal ex-states. Process the phis now. 468 region->set_req(0, region); // now it's an ordinary region 469 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 470 // Note: Setting the jvms also sets the bci and sp. 471 set_control(_gvn.transform(region)); 472 uint tos = jvms()->stkoff() + sp(); 473 for (uint i = 1; i < tos; i++) { 474 Node* x = phi_map->in(i); 475 if (x->in(0) == region) { 476 assert(x->is_Phi(), "expected a special phi"); 477 phi_map->set_req(i, _gvn.transform(x)); 478 } 479 } 480 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 481 Node* x = mms.memory(); 482 if (x->in(0) == region) { 483 assert(x->is_Phi(), "nobody else uses a hidden region"); 484 mms.set_memory(_gvn.transform(x)); 485 } 486 } 487 if (ex_oop->in(0) == region) { 488 assert(ex_oop->is_Phi(), "expected a special phi"); 489 ex_oop = _gvn.transform(ex_oop); 490 } 491 } else { 492 set_jvms(phi_map->jvms()); 493 } 494 495 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 496 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 497 return ex_oop; 498 } 499 500 //---------------------------------java_bc------------------------------------- 501 Bytecodes::Code GraphKit::java_bc() const { 502 ciMethod* method = this->method(); 503 int bci = this->bci(); 504 if (method != NULL && bci != InvocationEntryBci) 505 return method->java_code_at_bci(bci); 506 else 507 return Bytecodes::_illegal; 508 } 509 510 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 511 bool must_throw) { 512 // if the exception capability is set, then we will generate code 513 // to check the JavaThread.should_post_on_exceptions flag to see 514 // if we actually need to report exception events (for this 515 // thread). If we don't need to report exception events, we will 516 // take the normal fast path provided by add_exception_events. If 517 // exception event reporting is enabled for this thread, we will 518 // take the uncommon_trap in the BuildCutout below. 519 520 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 521 Node* jthread = _gvn.transform(new ThreadLocalNode()); 522 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 523 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 524 525 // Test the should_post_on_exceptions_flag vs. 0 526 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 527 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 528 529 // Branch to slow_path if should_post_on_exceptions_flag was true 530 { BuildCutout unless(this, tst, PROB_MAX); 531 // Do not try anything fancy if we're notifying the VM on every throw. 532 // Cf. case Bytecodes::_athrow in parse2.cpp. 533 uncommon_trap(reason, Deoptimization::Action_none, 534 (ciKlass*)NULL, (char*)NULL, must_throw); 535 } 536 537 } 538 539 //------------------------------builtin_throw---------------------------------- 540 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) { 541 bool must_throw = true; 542 543 // If this particular condition has not yet happened at this 544 // bytecode, then use the uncommon trap mechanism, and allow for 545 // a future recompilation if several traps occur here. 546 // If the throw is hot, try to use a more complicated inline mechanism 547 // which keeps execution inside the compiled code. 548 bool treat_throw_as_hot = false; 549 ciMethodData* md = method()->method_data(); 550 551 if (ProfileTraps) { 552 if (too_many_traps(reason)) { 553 treat_throw_as_hot = true; 554 } 555 // (If there is no MDO at all, assume it is early in 556 // execution, and that any deopts are part of the 557 // startup transient, and don't need to be remembered.) 558 559 // Also, if there is a local exception handler, treat all throws 560 // as hot if there has been at least one in this method. 561 if (C->trap_count(reason) != 0 562 && method()->method_data()->trap_count(reason) != 0 563 && has_ex_handler()) { 564 treat_throw_as_hot = true; 565 } 566 } 567 568 // If this throw happens frequently, an uncommon trap might cause 569 // a performance pothole. If there is a local exception handler, 570 // and if this particular bytecode appears to be deoptimizing often, 571 // let us handle the throw inline, with a preconstructed instance. 572 // Note: If the deopt count has blown up, the uncommon trap 573 // runtime is going to flush this nmethod, not matter what. 574 if (treat_throw_as_hot && method()->can_omit_stack_trace()) { 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 ex_obj = env()->ClassCastException_instance(); 594 break; 595 case Deoptimization::Reason_array_check: 596 ex_obj = env()->ArrayStoreException_instance(); 597 break; 598 default: 599 break; 600 } 601 if (failing()) { stop(); return; } // exception allocation might fail 602 if (ex_obj != NULL) { 603 if (env()->jvmti_can_post_on_exceptions()) { 604 // check if we must post exception events, take uncommon trap if so 605 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 606 // here if should_post_on_exceptions is false 607 // continue on with the normal codegen 608 } 609 610 // Cheat with a preallocated exception object. 611 if (C->log() != NULL) 612 C->log()->elem("hot_throw preallocated='1' reason='%s'", 613 Deoptimization::trap_reason_name(reason)); 614 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 615 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 616 617 // Clear the detail message of the preallocated exception object. 618 // Weblogic sometimes mutates the detail message of exceptions 619 // using reflection. 620 int offset = java_lang_Throwable::get_detailMessage_offset(); 621 const TypePtr* adr_typ = ex_con->add_offset(offset); 622 623 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 624 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 625 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 626 627 if (!method()->has_exception_handlers()) { 628 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway. 629 // This prevents issues with exception handling and late inlining. 630 set_sp(0); 631 clean_stack(0); 632 } 633 634 add_exception_state(make_exception_state(ex_node)); 635 return; 636 } 637 } 638 639 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 640 // It won't be much cheaper than bailing to the interp., since we'll 641 // have to pass up all the debug-info, and the runtime will have to 642 // create the stack trace. 643 644 // Usual case: Bail to interpreter. 645 // Reserve the right to recompile if we haven't seen anything yet. 646 647 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 648 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 649 if (treat_throw_as_hot 650 && (method()->method_data()->trap_recompiled_at(bci(), m) 651 || C->too_many_traps(reason))) { 652 // We cannot afford to take more traps here. Suffer in the interpreter. 653 if (C->log() != NULL) 654 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 655 Deoptimization::trap_reason_name(reason), 656 C->trap_count(reason)); 657 action = Deoptimization::Action_none; 658 } 659 660 // "must_throw" prunes the JVM state to include only the stack, if there 661 // are no local exception handlers. This should cut down on register 662 // allocation time and code size, by drastically reducing the number 663 // of in-edges on the call to the uncommon trap. 664 665 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 666 } 667 668 669 //----------------------------PreserveJVMState--------------------------------- 670 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 671 debug_only(kit->verify_map()); 672 _kit = kit; 673 _map = kit->map(); // preserve the map 674 _sp = kit->sp(); 675 kit->set_map(clone_map ? kit->clone_map() : NULL); 676 #ifdef ASSERT 677 _bci = kit->bci(); 678 Parse* parser = kit->is_Parse(); 679 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 680 _block = block; 681 #endif 682 } 683 PreserveJVMState::~PreserveJVMState() { 684 GraphKit* kit = _kit; 685 #ifdef ASSERT 686 assert(kit->bci() == _bci, "bci must not shift"); 687 Parse* parser = kit->is_Parse(); 688 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 689 assert(block == _block, "block must not shift"); 690 #endif 691 kit->set_map(_map); 692 kit->set_sp(_sp); 693 } 694 695 696 //-----------------------------BuildCutout------------------------------------- 697 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 698 : PreserveJVMState(kit) 699 { 700 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 701 SafePointNode* outer_map = _map; // preserved map is caller's 702 SafePointNode* inner_map = kit->map(); 703 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 704 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 705 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 706 } 707 BuildCutout::~BuildCutout() { 708 GraphKit* kit = _kit; 709 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 710 } 711 712 //---------------------------PreserveReexecuteState---------------------------- 713 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 714 assert(!kit->stopped(), "must call stopped() before"); 715 _kit = kit; 716 _sp = kit->sp(); 717 _reexecute = kit->jvms()->_reexecute; 718 } 719 PreserveReexecuteState::~PreserveReexecuteState() { 720 if (_kit->stopped()) return; 721 _kit->jvms()->_reexecute = _reexecute; 722 _kit->set_sp(_sp); 723 } 724 725 //------------------------------clone_map-------------------------------------- 726 // Implementation of PreserveJVMState 727 // 728 // Only clone_map(...) here. If this function is only used in the 729 // PreserveJVMState class we may want to get rid of this extra 730 // function eventually and do it all there. 731 732 SafePointNode* GraphKit::clone_map() { 733 if (map() == NULL) return NULL; 734 735 // Clone the memory edge first 736 Node* mem = MergeMemNode::make(map()->memory()); 737 gvn().set_type_bottom(mem); 738 739 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 740 JVMState* jvms = this->jvms(); 741 JVMState* clonejvms = jvms->clone_shallow(C); 742 clonemap->set_memory(mem); 743 clonemap->set_jvms(clonejvms); 744 clonejvms->set_map(clonemap); 745 record_for_igvn(clonemap); 746 gvn().set_type_bottom(clonemap); 747 return clonemap; 748 } 749 750 751 //-----------------------------set_map_clone----------------------------------- 752 void GraphKit::set_map_clone(SafePointNode* m) { 753 _map = m; 754 _map = clone_map(); 755 _map->set_next_exception(NULL); 756 debug_only(verify_map()); 757 } 758 759 760 //----------------------------kill_dead_locals--------------------------------- 761 // Detect any locals which are known to be dead, and force them to top. 762 void GraphKit::kill_dead_locals() { 763 // Consult the liveness information for the locals. If any 764 // of them are unused, then they can be replaced by top(). This 765 // should help register allocation time and cut down on the size 766 // of the deoptimization information. 767 768 // This call is made from many of the bytecode handling 769 // subroutines called from the Big Switch in do_one_bytecode. 770 // Every bytecode which might include a slow path is responsible 771 // for killing its dead locals. The more consistent we 772 // are about killing deads, the fewer useless phis will be 773 // constructed for them at various merge points. 774 775 // bci can be -1 (InvocationEntryBci). We return the entry 776 // liveness for the method. 777 778 if (method() == NULL || method()->code_size() == 0) { 779 // We are building a graph for a call to a native method. 780 // All locals are live. 781 return; 782 } 783 784 ResourceMark rm; 785 786 // Consult the liveness information for the locals. If any 787 // of them are unused, then they can be replaced by top(). This 788 // should help register allocation time and cut down on the size 789 // of the deoptimization information. 790 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 791 792 int len = (int)live_locals.size(); 793 assert(len <= jvms()->loc_size(), "too many live locals"); 794 for (int local = 0; local < len; local++) { 795 if (!live_locals.at(local)) { 796 set_local(local, top()); 797 } 798 } 799 } 800 801 #ifdef ASSERT 802 //-------------------------dead_locals_are_killed------------------------------ 803 // Return true if all dead locals are set to top in the map. 804 // Used to assert "clean" debug info at various points. 805 bool GraphKit::dead_locals_are_killed() { 806 if (method() == NULL || method()->code_size() == 0) { 807 // No locals need to be dead, so all is as it should be. 808 return true; 809 } 810 811 // Make sure somebody called kill_dead_locals upstream. 812 ResourceMark rm; 813 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 814 if (jvms->loc_size() == 0) continue; // no locals to consult 815 SafePointNode* map = jvms->map(); 816 ciMethod* method = jvms->method(); 817 int bci = jvms->bci(); 818 if (jvms == this->jvms()) { 819 bci = this->bci(); // it might not yet be synched 820 } 821 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 822 int len = (int)live_locals.size(); 823 if (!live_locals.is_valid() || len == 0) 824 // This method is trivial, or is poisoned by a breakpoint. 825 return true; 826 assert(len == jvms->loc_size(), "live map consistent with locals map"); 827 for (int local = 0; local < len; local++) { 828 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 829 if (PrintMiscellaneous && (Verbose || WizardMode)) { 830 tty->print_cr("Zombie local %d: ", local); 831 jvms->dump(); 832 } 833 return false; 834 } 835 } 836 } 837 return true; 838 } 839 840 #endif //ASSERT 841 842 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens. 843 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 844 ciMethod* cur_method = jvms->method(); 845 int cur_bci = jvms->bci(); 846 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 847 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 848 return Interpreter::bytecode_should_reexecute(code) || 849 (is_anewarray && (code == Bytecodes::_multianewarray)); 850 // Reexecute _multianewarray bytecode which was replaced with 851 // sequence of [a]newarray. See Parse::do_multianewarray(). 852 // 853 // Note: interpreter should not have it set since this optimization 854 // is limited by dimensions and guarded by flag so in some cases 855 // multianewarray() runtime calls will be generated and 856 // the bytecode should not be reexecutes (stack will not be reset). 857 } else { 858 return false; 859 } 860 } 861 862 // Helper function for adding JVMState and debug information to node 863 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 864 // Add the safepoint edges to the call (or other safepoint). 865 866 // Make sure dead locals are set to top. This 867 // should help register allocation time and cut down on the size 868 // of the deoptimization information. 869 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 870 871 // Walk the inline list to fill in the correct set of JVMState's 872 // Also fill in the associated edges for each JVMState. 873 874 // If the bytecode needs to be reexecuted we need to put 875 // the arguments back on the stack. 876 const bool should_reexecute = jvms()->should_reexecute(); 877 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 878 879 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 880 // undefined if the bci is different. This is normal for Parse but it 881 // should not happen for LibraryCallKit because only one bci is processed. 882 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 883 "in LibraryCallKit the reexecute bit should not change"); 884 885 // If we are guaranteed to throw, we can prune everything but the 886 // input to the current bytecode. 887 bool can_prune_locals = false; 888 uint stack_slots_not_pruned = 0; 889 int inputs = 0, depth = 0; 890 if (must_throw) { 891 assert(method() == youngest_jvms->method(), "sanity"); 892 if (compute_stack_effects(inputs, depth)) { 893 can_prune_locals = true; 894 stack_slots_not_pruned = inputs; 895 } 896 } 897 898 if (env()->should_retain_local_variables()) { 899 // At any safepoint, this method can get breakpointed, which would 900 // then require an immediate deoptimization. 901 can_prune_locals = false; // do not prune locals 902 stack_slots_not_pruned = 0; 903 } 904 905 // do not scribble on the input jvms 906 JVMState* out_jvms = youngest_jvms->clone_deep(C); 907 call->set_jvms(out_jvms); // Start jvms list for call node 908 909 // For a known set of bytecodes, the interpreter should reexecute them if 910 // deoptimization happens. We set the reexecute state for them here 911 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 912 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 913 #ifdef ASSERT 914 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects() 915 assert(method() == youngest_jvms->method(), "sanity"); 916 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc())); 917 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution"); 918 #endif // ASSERT 919 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 920 } 921 922 // Presize the call: 923 DEBUG_ONLY(uint non_debug_edges = call->req()); 924 call->add_req_batch(top(), youngest_jvms->debug_depth()); 925 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 926 927 // Set up edges so that the call looks like this: 928 // Call [state:] ctl io mem fptr retadr 929 // [parms:] parm0 ... parmN 930 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 931 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 932 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 933 // Note that caller debug info precedes callee debug info. 934 935 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 936 uint debug_ptr = call->req(); 937 938 // Loop over the map input edges associated with jvms, add them 939 // to the call node, & reset all offsets to match call node array. 940 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 941 uint debug_end = debug_ptr; 942 uint debug_start = debug_ptr - in_jvms->debug_size(); 943 debug_ptr = debug_start; // back up the ptr 944 945 uint p = debug_start; // walks forward in [debug_start, debug_end) 946 uint j, k, l; 947 SafePointNode* in_map = in_jvms->map(); 948 out_jvms->set_map(call); 949 950 if (can_prune_locals) { 951 assert(in_jvms->method() == out_jvms->method(), "sanity"); 952 // If the current throw can reach an exception handler in this JVMS, 953 // then we must keep everything live that can reach that handler. 954 // As a quick and dirty approximation, we look for any handlers at all. 955 if (in_jvms->method()->has_exception_handlers()) { 956 can_prune_locals = false; 957 } 958 } 959 960 // Add the Locals 961 k = in_jvms->locoff(); 962 l = in_jvms->loc_size(); 963 out_jvms->set_locoff(p); 964 if (!can_prune_locals) { 965 for (j = 0; j < l; j++) 966 call->set_req(p++, in_map->in(k+j)); 967 } else { 968 p += l; // already set to top above by add_req_batch 969 } 970 971 // Add the Expression Stack 972 k = in_jvms->stkoff(); 973 l = in_jvms->sp(); 974 out_jvms->set_stkoff(p); 975 if (!can_prune_locals) { 976 for (j = 0; j < l; j++) 977 call->set_req(p++, in_map->in(k+j)); 978 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 979 // Divide stack into {S0,...,S1}, where S0 is set to top. 980 uint s1 = stack_slots_not_pruned; 981 stack_slots_not_pruned = 0; // for next iteration 982 if (s1 > l) s1 = l; 983 uint s0 = l - s1; 984 p += s0; // skip the tops preinstalled by add_req_batch 985 for (j = s0; j < l; j++) 986 call->set_req(p++, in_map->in(k+j)); 987 } else { 988 p += l; // already set to top above by add_req_batch 989 } 990 991 // Add the Monitors 992 k = in_jvms->monoff(); 993 l = in_jvms->mon_size(); 994 out_jvms->set_monoff(p); 995 for (j = 0; j < l; j++) 996 call->set_req(p++, in_map->in(k+j)); 997 998 // Copy any scalar object fields. 999 k = in_jvms->scloff(); 1000 l = in_jvms->scl_size(); 1001 out_jvms->set_scloff(p); 1002 for (j = 0; j < l; j++) 1003 call->set_req(p++, in_map->in(k+j)); 1004 1005 // Finish the new jvms. 1006 out_jvms->set_endoff(p); 1007 1008 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 1009 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 1010 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 1011 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 1012 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 1013 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 1014 1015 // Update the two tail pointers in parallel. 1016 out_jvms = out_jvms->caller(); 1017 in_jvms = in_jvms->caller(); 1018 } 1019 1020 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1021 1022 // Test the correctness of JVMState::debug_xxx accessors: 1023 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1024 assert(call->jvms()->debug_end() == call->req(), ""); 1025 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1026 } 1027 1028 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1029 Bytecodes::Code code = java_bc(); 1030 if (code == Bytecodes::_wide) { 1031 code = method()->java_code_at_bci(bci() + 1); 1032 } 1033 1034 if (code != Bytecodes::_illegal) { 1035 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1036 } 1037 1038 auto rsize = [&]() { 1039 assert(code != Bytecodes::_illegal, "code is illegal!"); 1040 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1041 return (rtype < T_CONFLICT) ? type2size[rtype] : 0; 1042 }; 1043 1044 switch (code) { 1045 case Bytecodes::_illegal: 1046 return false; 1047 1048 case Bytecodes::_ldc: 1049 case Bytecodes::_ldc_w: 1050 case Bytecodes::_ldc2_w: 1051 inputs = 0; 1052 break; 1053 1054 case Bytecodes::_dup: inputs = 1; break; 1055 case Bytecodes::_dup_x1: inputs = 2; break; 1056 case Bytecodes::_dup_x2: inputs = 3; break; 1057 case Bytecodes::_dup2: inputs = 2; break; 1058 case Bytecodes::_dup2_x1: inputs = 3; break; 1059 case Bytecodes::_dup2_x2: inputs = 4; break; 1060 case Bytecodes::_swap: inputs = 2; break; 1061 case Bytecodes::_arraylength: inputs = 1; break; 1062 1063 case Bytecodes::_getstatic: 1064 case Bytecodes::_putstatic: 1065 case Bytecodes::_getfield: 1066 case Bytecodes::_putfield: 1067 { 1068 bool ignored_will_link; 1069 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1070 int size = field->type()->size(); 1071 bool is_get = (depth >= 0), is_static = (depth & 1); 1072 inputs = (is_static ? 0 : 1); 1073 if (is_get) { 1074 depth = size - inputs; 1075 } else { 1076 inputs += size; // putxxx pops the value from the stack 1077 depth = - inputs; 1078 } 1079 } 1080 break; 1081 1082 case Bytecodes::_invokevirtual: 1083 case Bytecodes::_invokespecial: 1084 case Bytecodes::_invokestatic: 1085 case Bytecodes::_invokedynamic: 1086 case Bytecodes::_invokeinterface: 1087 { 1088 bool ignored_will_link; 1089 ciSignature* declared_signature = NULL; 1090 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1091 assert(declared_signature != NULL, "cannot be null"); 1092 inputs = declared_signature->arg_size_for_bc(code); 1093 int size = declared_signature->return_type()->size(); 1094 depth = size - inputs; 1095 } 1096 break; 1097 1098 case Bytecodes::_multianewarray: 1099 { 1100 ciBytecodeStream iter(method()); 1101 iter.reset_to_bci(bci()); 1102 iter.next(); 1103 inputs = iter.get_dimensions(); 1104 assert(rsize() == 1, ""); 1105 depth = 1 - inputs; 1106 } 1107 break; 1108 1109 case Bytecodes::_withfield: { 1110 bool ignored_will_link; 1111 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1112 int size = field->type()->size(); 1113 inputs = size+1; 1114 depth = rsize() - inputs; 1115 break; 1116 } 1117 1118 case Bytecodes::_ireturn: 1119 case Bytecodes::_lreturn: 1120 case Bytecodes::_freturn: 1121 case Bytecodes::_dreturn: 1122 case Bytecodes::_areturn: 1123 assert(rsize() == -depth, ""); 1124 inputs = -depth; 1125 break; 1126 1127 case Bytecodes::_jsr: 1128 case Bytecodes::_jsr_w: 1129 inputs = 0; 1130 depth = 1; // S.B. depth=1, not zero 1131 break; 1132 1133 default: 1134 // bytecode produces a typed result 1135 inputs = rsize() - depth; 1136 assert(inputs >= 0, ""); 1137 break; 1138 } 1139 1140 #ifdef ASSERT 1141 // spot check 1142 int outputs = depth + inputs; 1143 assert(outputs >= 0, "sanity"); 1144 switch (code) { 1145 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1146 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1147 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1148 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1149 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1150 default: break; 1151 } 1152 #endif //ASSERT 1153 1154 return true; 1155 } 1156 1157 1158 1159 //------------------------------basic_plus_adr--------------------------------- 1160 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1161 // short-circuit a common case 1162 if (offset == intcon(0)) return ptr; 1163 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1164 } 1165 1166 Node* GraphKit::ConvI2L(Node* offset) { 1167 // short-circuit a common case 1168 jint offset_con = find_int_con(offset, Type::OffsetBot); 1169 if (offset_con != Type::OffsetBot) { 1170 return longcon((jlong) offset_con); 1171 } 1172 return _gvn.transform( new ConvI2LNode(offset)); 1173 } 1174 1175 Node* GraphKit::ConvI2UL(Node* offset) { 1176 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1177 if (offset_con != (juint) Type::OffsetBot) { 1178 return longcon((julong) offset_con); 1179 } 1180 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1181 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1182 return _gvn.transform( new AndLNode(conv, mask) ); 1183 } 1184 1185 Node* GraphKit::ConvL2I(Node* offset) { 1186 // short-circuit a common case 1187 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1188 if (offset_con != (jlong)Type::OffsetBot) { 1189 return intcon((int) offset_con); 1190 } 1191 return _gvn.transform( new ConvL2INode(offset)); 1192 } 1193 1194 //-------------------------load_object_klass----------------------------------- 1195 Node* GraphKit::load_object_klass(Node* obj) { 1196 // Special-case a fresh allocation to avoid building nodes: 1197 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1198 if (akls != NULL) return akls; 1199 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1200 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 1201 } 1202 1203 //-------------------------load_array_length----------------------------------- 1204 Node* GraphKit::load_array_length(Node* array) { 1205 // Special-case a fresh allocation to avoid building nodes: 1206 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1207 Node *alen; 1208 if (alloc == NULL) { 1209 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1210 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1211 } else { 1212 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false); 1213 } 1214 return alen; 1215 } 1216 1217 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc, 1218 const TypeOopPtr* oop_type, 1219 bool replace_length_in_map) { 1220 Node* length = alloc->Ideal_length(); 1221 if (replace_length_in_map == false || map()->find_edge(length) >= 0) { 1222 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn); 1223 if (ccast != length) { 1224 // do not transform ccast here, it might convert to top node for 1225 // negative array length and break assumptions in parsing stage. 1226 _gvn.set_type_bottom(ccast); 1227 record_for_igvn(ccast); 1228 if (replace_length_in_map) { 1229 replace_in_map(length, ccast); 1230 } 1231 return ccast; 1232 } 1233 } 1234 return length; 1235 } 1236 1237 //------------------------------do_null_check---------------------------------- 1238 // Helper function to do a NULL pointer check. Returned value is 1239 // the incoming address with NULL casted away. You are allowed to use the 1240 // not-null value only if you are control dependent on the test. 1241 #ifndef PRODUCT 1242 extern int explicit_null_checks_inserted, 1243 explicit_null_checks_elided; 1244 #endif 1245 Node* GraphKit::null_check_common(Node* value, BasicType type, 1246 // optional arguments for variations: 1247 bool assert_null, 1248 Node* *null_control, 1249 bool speculative, 1250 bool is_init_check) { 1251 assert(!assert_null || null_control == NULL, "not both at once"); 1252 if (stopped()) return top(); 1253 NOT_PRODUCT(explicit_null_checks_inserted++); 1254 1255 if (value->is_InlineType()) { 1256 // Null checking a scalarized but nullable inline type. Check the IsInit 1257 // input instead of the oop input to avoid keeping buffer allocations alive. 1258 InlineTypeNode* vtptr = value->as_InlineType(); 1259 while (vtptr->get_oop()->is_InlineType()) { 1260 vtptr = vtptr->get_oop()->as_InlineType(); 1261 } 1262 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true); 1263 if (stopped()) { 1264 return top(); 1265 } 1266 if (assert_null) { 1267 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts) 1268 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass()); 1269 // replace_in_map(value, vtptr); 1270 // return vtptr; 1271 return null(); 1272 } 1273 bool do_replace_in_map = (null_control == NULL || (*null_control) == top()); 1274 return cast_not_null(value, do_replace_in_map); 1275 } 1276 1277 // Construct NULL check 1278 Node *chk = NULL; 1279 switch(type) { 1280 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1281 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1282 case T_PRIMITIVE_OBJECT : // fall through 1283 case T_ARRAY : // fall through 1284 type = T_OBJECT; // simplify further tests 1285 case T_OBJECT : { 1286 const Type *t = _gvn.type( value ); 1287 1288 const TypeOopPtr* tp = t->isa_oopptr(); 1289 if (tp != NULL && !tp->is_loaded() 1290 // Only for do_null_check, not any of its siblings: 1291 && !assert_null && null_control == NULL) { 1292 // Usually, any field access or invocation on an unloaded oop type 1293 // will simply fail to link, since the statically linked class is 1294 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1295 // the static class is loaded but the sharper oop type is not. 1296 // Rather than checking for this obscure case in lots of places, 1297 // we simply observe that a null check on an unloaded class 1298 // will always be followed by a nonsense operation, so we 1299 // can just issue the uncommon trap here. 1300 // Our access to the unloaded class will only be correct 1301 // after it has been loaded and initialized, which requires 1302 // a trip through the interpreter. 1303 ciKlass* klass = tp->unloaded_klass(); 1304 #ifndef PRODUCT 1305 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); } 1306 #endif 1307 uncommon_trap(Deoptimization::Reason_unloaded, 1308 Deoptimization::Action_reinterpret, 1309 klass, "!loaded"); 1310 return top(); 1311 } 1312 1313 if (assert_null) { 1314 // See if the type is contained in NULL_PTR. 1315 // If so, then the value is already null. 1316 if (t->higher_equal(TypePtr::NULL_PTR)) { 1317 NOT_PRODUCT(explicit_null_checks_elided++); 1318 return value; // Elided null assert quickly! 1319 } 1320 } else { 1321 // See if mixing in the NULL pointer changes type. 1322 // If so, then the NULL pointer was not allowed in the original 1323 // type. In other words, "value" was not-null. 1324 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1325 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1326 NOT_PRODUCT(explicit_null_checks_elided++); 1327 return value; // Elided null check quickly! 1328 } 1329 } 1330 chk = new CmpPNode( value, null() ); 1331 break; 1332 } 1333 1334 default: 1335 fatal("unexpected type: %s", type2name(type)); 1336 } 1337 assert(chk != NULL, "sanity check"); 1338 chk = _gvn.transform(chk); 1339 1340 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1341 BoolNode *btst = new BoolNode( chk, btest); 1342 Node *tst = _gvn.transform( btst ); 1343 1344 //----------- 1345 // if peephole optimizations occurred, a prior test existed. 1346 // If a prior test existed, maybe it dominates as we can avoid this test. 1347 if (tst != btst && type == T_OBJECT) { 1348 // At this point we want to scan up the CFG to see if we can 1349 // find an identical test (and so avoid this test altogether). 1350 Node *cfg = control(); 1351 int depth = 0; 1352 while( depth < 16 ) { // Limit search depth for speed 1353 if( cfg->Opcode() == Op_IfTrue && 1354 cfg->in(0)->in(1) == tst ) { 1355 // Found prior test. Use "cast_not_null" to construct an identical 1356 // CastPP (and hence hash to) as already exists for the prior test. 1357 // Return that casted value. 1358 if (assert_null) { 1359 replace_in_map(value, null()); 1360 return null(); // do not issue the redundant test 1361 } 1362 Node *oldcontrol = control(); 1363 set_control(cfg); 1364 Node *res = cast_not_null(value); 1365 set_control(oldcontrol); 1366 NOT_PRODUCT(explicit_null_checks_elided++); 1367 return res; 1368 } 1369 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1370 if (cfg == NULL) break; // Quit at region nodes 1371 depth++; 1372 } 1373 } 1374 1375 //----------- 1376 // Branch to failure if null 1377 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1378 Deoptimization::DeoptReason reason; 1379 if (assert_null) { 1380 reason = Deoptimization::reason_null_assert(speculative); 1381 } else if (type == T_OBJECT || is_init_check) { 1382 reason = Deoptimization::reason_null_check(speculative); 1383 } else { 1384 reason = Deoptimization::Reason_div0_check; 1385 } 1386 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1387 // ciMethodData::has_trap_at will return a conservative -1 if any 1388 // must-be-null assertion has failed. This could cause performance 1389 // problems for a method after its first do_null_assert failure. 1390 // Consider using 'Reason_class_check' instead? 1391 1392 // To cause an implicit null check, we set the not-null probability 1393 // to the maximum (PROB_MAX). For an explicit check the probability 1394 // is set to a smaller value. 1395 if (null_control != NULL || too_many_traps(reason)) { 1396 // probability is less likely 1397 ok_prob = PROB_LIKELY_MAG(3); 1398 } else if (!assert_null && 1399 (ImplicitNullCheckThreshold > 0) && 1400 method() != NULL && 1401 (method()->method_data()->trap_count(reason) 1402 >= (uint)ImplicitNullCheckThreshold)) { 1403 ok_prob = PROB_LIKELY_MAG(3); 1404 } 1405 1406 if (null_control != NULL) { 1407 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1408 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1409 set_control( _gvn.transform( new IfTrueNode(iff))); 1410 #ifndef PRODUCT 1411 if (null_true == top()) { 1412 explicit_null_checks_elided++; 1413 } 1414 #endif 1415 (*null_control) = null_true; 1416 } else { 1417 BuildCutout unless(this, tst, ok_prob); 1418 // Check for optimizer eliding test at parse time 1419 if (stopped()) { 1420 // Failure not possible; do not bother making uncommon trap. 1421 NOT_PRODUCT(explicit_null_checks_elided++); 1422 } else if (assert_null) { 1423 uncommon_trap(reason, 1424 Deoptimization::Action_make_not_entrant, 1425 NULL, "assert_null"); 1426 } else { 1427 replace_in_map(value, zerocon(type)); 1428 builtin_throw(reason); 1429 } 1430 } 1431 1432 // Must throw exception, fall-thru not possible? 1433 if (stopped()) { 1434 return top(); // No result 1435 } 1436 1437 if (assert_null) { 1438 // Cast obj to null on this path. 1439 replace_in_map(value, zerocon(type)); 1440 return zerocon(type); 1441 } 1442 1443 // Cast obj to not-null on this path, if there is no null_control. 1444 // (If there is a null_control, a non-null value may come back to haunt us.) 1445 if (type == T_OBJECT) { 1446 Node* cast = cast_not_null(value, false); 1447 if (null_control == NULL || (*null_control) == top()) 1448 replace_in_map(value, cast); 1449 value = cast; 1450 } 1451 1452 return value; 1453 } 1454 1455 //------------------------------cast_not_null---------------------------------- 1456 // Cast obj to not-null on this path 1457 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1458 if (obj->is_InlineType()) { 1459 Node* vt = obj->clone(); 1460 vt->as_InlineType()->set_is_init(_gvn); 1461 vt = _gvn.transform(vt); 1462 if (do_replace_in_map) { 1463 replace_in_map(obj, vt); 1464 } 1465 return vt; 1466 } 1467 const Type *t = _gvn.type(obj); 1468 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1469 // Object is already not-null? 1470 if( t == t_not_null ) return obj; 1471 1472 Node *cast = new CastPPNode(obj,t_not_null); 1473 cast->init_req(0, control()); 1474 cast = _gvn.transform( cast ); 1475 1476 // Scan for instances of 'obj' in the current JVM mapping. 1477 // These instances are known to be not-null after the test. 1478 if (do_replace_in_map) 1479 replace_in_map(obj, cast); 1480 1481 return cast; // Return casted value 1482 } 1483 1484 // Sometimes in intrinsics, we implicitly know an object is not null 1485 // (there's no actual null check) so we can cast it to not null. In 1486 // the course of optimizations, the input to the cast can become null. 1487 // In that case that data path will die and we need the control path 1488 // to become dead as well to keep the graph consistent. So we have to 1489 // add a check for null for which one branch can't be taken. It uses 1490 // an Opaque4 node that will cause the check to be removed after loop 1491 // opts so the test goes away and the compiled code doesn't execute a 1492 // useless check. 1493 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1494 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1495 return value; 1496 } 1497 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1498 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1499 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1500 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1501 _gvn.set_type(iff, iff->Value(&_gvn)); 1502 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1503 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1504 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1505 C->root()->add_req(halt); 1506 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1507 set_control(if_t); 1508 return cast_not_null(value, do_replace_in_map); 1509 } 1510 1511 1512 //--------------------------replace_in_map------------------------------------- 1513 void GraphKit::replace_in_map(Node* old, Node* neww) { 1514 if (old == neww) { 1515 return; 1516 } 1517 1518 map()->replace_edge(old, neww); 1519 1520 // Note: This operation potentially replaces any edge 1521 // on the map. This includes locals, stack, and monitors 1522 // of the current (innermost) JVM state. 1523 1524 // don't let inconsistent types from profiling escape this 1525 // method 1526 1527 const Type* told = _gvn.type(old); 1528 const Type* tnew = _gvn.type(neww); 1529 1530 if (!tnew->higher_equal(told)) { 1531 return; 1532 } 1533 1534 map()->record_replaced_node(old, neww); 1535 } 1536 1537 1538 //============================================================================= 1539 //--------------------------------memory--------------------------------------- 1540 Node* GraphKit::memory(uint alias_idx) { 1541 MergeMemNode* mem = merged_memory(); 1542 Node* p = mem->memory_at(alias_idx); 1543 assert(p != mem->empty_memory(), "empty"); 1544 _gvn.set_type(p, Type::MEMORY); // must be mapped 1545 return p; 1546 } 1547 1548 //-----------------------------reset_memory------------------------------------ 1549 Node* GraphKit::reset_memory() { 1550 Node* mem = map()->memory(); 1551 // do not use this node for any more parsing! 1552 debug_only( map()->set_memory((Node*)NULL) ); 1553 return _gvn.transform( mem ); 1554 } 1555 1556 //------------------------------set_all_memory--------------------------------- 1557 void GraphKit::set_all_memory(Node* newmem) { 1558 Node* mergemem = MergeMemNode::make(newmem); 1559 gvn().set_type_bottom(mergemem); 1560 map()->set_memory(mergemem); 1561 } 1562 1563 //------------------------------set_all_memory_call---------------------------- 1564 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1565 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1566 set_all_memory(newmem); 1567 } 1568 1569 //============================================================================= 1570 // 1571 // parser factory methods for MemNodes 1572 // 1573 // These are layered on top of the factory methods in LoadNode and StoreNode, 1574 // and integrate with the parser's memory state and _gvn engine. 1575 // 1576 1577 // factory methods in "int adr_idx" 1578 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1579 int adr_idx, 1580 MemNode::MemOrd mo, 1581 LoadNode::ControlDependency control_dependency, 1582 bool require_atomic_access, 1583 bool unaligned, 1584 bool mismatched, 1585 bool unsafe, 1586 uint8_t barrier_data) { 1587 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1588 const TypePtr* adr_type = NULL; // debug-mode-only argument 1589 debug_only(adr_type = C->get_adr_type(adr_idx)); 1590 Node* mem = memory(adr_idx); 1591 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data); 1592 ld = _gvn.transform(ld); 1593 1594 if (((bt == T_OBJECT || bt == T_PRIMITIVE_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1595 // Improve graph before escape analysis and boxing elimination. 1596 record_for_igvn(ld); 1597 } 1598 return ld; 1599 } 1600 1601 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1602 int adr_idx, 1603 MemNode::MemOrd mo, 1604 bool require_atomic_access, 1605 bool unaligned, 1606 bool mismatched, 1607 bool unsafe, 1608 int barrier_data) { 1609 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1610 const TypePtr* adr_type = NULL; 1611 debug_only(adr_type = C->get_adr_type(adr_idx)); 1612 Node *mem = memory(adr_idx); 1613 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access); 1614 if (unaligned) { 1615 st->as_Store()->set_unaligned_access(); 1616 } 1617 if (mismatched) { 1618 st->as_Store()->set_mismatched_access(); 1619 } 1620 if (unsafe) { 1621 st->as_Store()->set_unsafe_access(); 1622 } 1623 st->as_Store()->set_barrier_data(barrier_data); 1624 st = _gvn.transform(st); 1625 set_memory(st, adr_idx); 1626 // Back-to-back stores can only remove intermediate store with DU info 1627 // so push on worklist for optimizer. 1628 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1629 record_for_igvn(st); 1630 1631 return st; 1632 } 1633 1634 Node* GraphKit::access_store_at(Node* obj, 1635 Node* adr, 1636 const TypePtr* adr_type, 1637 Node* val, 1638 const Type* val_type, 1639 BasicType bt, 1640 DecoratorSet decorators, 1641 bool safe_for_replace) { 1642 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1643 // could be delayed during Parse (for example, in adjust_map_after_if()). 1644 // Execute transformation here to avoid barrier generation in such case. 1645 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1646 val = _gvn.makecon(TypePtr::NULL_PTR); 1647 } 1648 1649 if (stopped()) { 1650 return top(); // Dead path ? 1651 } 1652 1653 assert(val != NULL, "not dead path"); 1654 if (val->is_InlineType()) { 1655 // Store to non-flattened field. Buffer the inline type and make sure 1656 // the store is re-executed if the allocation triggers deoptimization. 1657 PreserveReexecuteState preexecs(this); 1658 jvms()->set_should_reexecute(true); 1659 val = val->as_InlineType()->buffer(this, safe_for_replace); 1660 } 1661 1662 C2AccessValuePtr addr(adr, adr_type); 1663 C2AccessValue value(val, val_type); 1664 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1665 if (access.is_raw()) { 1666 return _barrier_set->BarrierSetC2::store_at(access, value); 1667 } else { 1668 return _barrier_set->store_at(access, value); 1669 } 1670 } 1671 1672 Node* GraphKit::access_load_at(Node* obj, // containing obj 1673 Node* adr, // actual address to store val at 1674 const TypePtr* adr_type, 1675 const Type* val_type, 1676 BasicType bt, 1677 DecoratorSet decorators, 1678 Node* ctl) { 1679 if (stopped()) { 1680 return top(); // Dead path ? 1681 } 1682 1683 C2AccessValuePtr addr(adr, adr_type); 1684 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl); 1685 if (access.is_raw()) { 1686 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1687 } else { 1688 return _barrier_set->load_at(access, val_type); 1689 } 1690 } 1691 1692 Node* GraphKit::access_load(Node* adr, // actual address to load val at 1693 const Type* val_type, 1694 BasicType bt, 1695 DecoratorSet decorators) { 1696 if (stopped()) { 1697 return top(); // Dead path ? 1698 } 1699 1700 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr()); 1701 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr); 1702 if (access.is_raw()) { 1703 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1704 } else { 1705 return _barrier_set->load_at(access, val_type); 1706 } 1707 } 1708 1709 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1710 Node* adr, 1711 const TypePtr* adr_type, 1712 int alias_idx, 1713 Node* expected_val, 1714 Node* new_val, 1715 const Type* value_type, 1716 BasicType bt, 1717 DecoratorSet decorators) { 1718 C2AccessValuePtr addr(adr, adr_type); 1719 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1720 bt, obj, addr, alias_idx); 1721 if (access.is_raw()) { 1722 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1723 } else { 1724 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1725 } 1726 } 1727 1728 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1729 Node* adr, 1730 const TypePtr* adr_type, 1731 int alias_idx, 1732 Node* expected_val, 1733 Node* new_val, 1734 const Type* value_type, 1735 BasicType bt, 1736 DecoratorSet decorators) { 1737 C2AccessValuePtr addr(adr, adr_type); 1738 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1739 bt, obj, addr, alias_idx); 1740 if (access.is_raw()) { 1741 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1742 } else { 1743 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1744 } 1745 } 1746 1747 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1748 Node* adr, 1749 const TypePtr* adr_type, 1750 int alias_idx, 1751 Node* new_val, 1752 const Type* value_type, 1753 BasicType bt, 1754 DecoratorSet decorators) { 1755 C2AccessValuePtr addr(adr, adr_type); 1756 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1757 bt, obj, addr, alias_idx); 1758 if (access.is_raw()) { 1759 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1760 } else { 1761 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1762 } 1763 } 1764 1765 Node* GraphKit::access_atomic_add_at(Node* obj, 1766 Node* adr, 1767 const TypePtr* adr_type, 1768 int alias_idx, 1769 Node* new_val, 1770 const Type* value_type, 1771 BasicType bt, 1772 DecoratorSet decorators) { 1773 C2AccessValuePtr addr(adr, adr_type); 1774 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1775 if (access.is_raw()) { 1776 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1777 } else { 1778 return _barrier_set->atomic_add_at(access, new_val, value_type); 1779 } 1780 } 1781 1782 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) { 1783 return _barrier_set->clone(this, src, dst, size, is_array); 1784 } 1785 1786 //-------------------------array_element_address------------------------- 1787 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1788 const TypeInt* sizetype, Node* ctrl) { 1789 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 1790 uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt)); 1791 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1792 1793 // short-circuit a common case (saves lots of confusing waste motion) 1794 jint idx_con = find_int_con(idx, -1); 1795 if (idx_con >= 0) { 1796 intptr_t offset = header + ((intptr_t)idx_con << shift); 1797 return basic_plus_adr(ary, offset); 1798 } 1799 1800 // must be correct type for alignment purposes 1801 Node* base = basic_plus_adr(ary, header); 1802 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1803 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1804 return basic_plus_adr(ary, base, scale); 1805 } 1806 1807 //-------------------------load_array_element------------------------- 1808 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) { 1809 const Type* elemtype = arytype->elem(); 1810 BasicType elembt = elemtype->array_element_basic_type(); 1811 assert(elembt != T_PRIMITIVE_OBJECT, "inline types are not supported by this method"); 1812 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1813 if (elembt == T_NARROWOOP) { 1814 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1815 } 1816 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt, 1817 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0)); 1818 return ld; 1819 } 1820 1821 //-------------------------set_arguments_for_java_call------------------------- 1822 // Arguments (pre-popped from the stack) are taken from the JVMS. 1823 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) { 1824 PreserveReexecuteState preexecs(this); 1825 if (EnableValhalla) { 1826 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization. 1827 // At this point, the call hasn't been executed yet, so we will only ever execute the call once. 1828 jvms()->set_should_reexecute(true); 1829 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc()); 1830 inc_sp(arg_size); 1831 } 1832 // Add the call arguments 1833 const TypeTuple* domain = call->tf()->domain_sig(); 1834 uint nargs = domain->cnt(); 1835 int arg_num = 0; 1836 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1837 Node* arg = argument(i-TypeFunc::Parms); 1838 const Type* t = domain->field_at(i); 1839 if (t->is_inlinetypeptr() && call->method()->is_scalarized_arg(arg_num)) { 1840 // We don't pass inline type arguments by reference but instead pass each field of the inline type 1841 if (!arg->is_InlineType()) { 1842 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type"); 1843 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free()); 1844 } 1845 InlineTypeNode* vt = arg->as_InlineType(); 1846 vt->pass_fields(this, call, idx, true, !t->maybe_null()); 1847 // If an inline type argument is passed as fields, attach the Method* to the call site 1848 // to be able to access the extended signature later via attached_method_before_pc(). 1849 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1850 call->set_override_symbolic_info(true); 1851 arg_num++; 1852 continue; 1853 } else if (arg->is_InlineType()) { 1854 // Pass inline type argument via oop to callee 1855 arg = arg->as_InlineType()->buffer(this); 1856 if (!is_late_inline) { 1857 arg = arg->as_InlineType()->get_oop(); 1858 } 1859 } 1860 if (t != Type::HALF) { 1861 arg_num++; 1862 } 1863 call->init_req(idx++, arg); 1864 } 1865 } 1866 1867 //---------------------------set_edges_for_java_call--------------------------- 1868 // Connect a newly created call into the current JVMS. 1869 // A return value node (if any) is returned from set_edges_for_java_call. 1870 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1871 1872 // Add the predefined inputs: 1873 call->init_req( TypeFunc::Control, control() ); 1874 call->init_req( TypeFunc::I_O , i_o() ); 1875 call->init_req( TypeFunc::Memory , reset_memory() ); 1876 call->init_req( TypeFunc::FramePtr, frameptr() ); 1877 call->init_req( TypeFunc::ReturnAdr, top() ); 1878 1879 add_safepoint_edges(call, must_throw); 1880 1881 Node* xcall = _gvn.transform(call); 1882 1883 if (xcall == top()) { 1884 set_control(top()); 1885 return; 1886 } 1887 assert(xcall == call, "call identity is stable"); 1888 1889 // Re-use the current map to produce the result. 1890 1891 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1892 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1893 set_all_memory_call(xcall, separate_io_proj); 1894 1895 //return xcall; // no need, caller already has it 1896 } 1897 1898 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1899 if (stopped()) return top(); // maybe the call folded up? 1900 1901 // Note: Since any out-of-line call can produce an exception, 1902 // we always insert an I_O projection from the call into the result. 1903 1904 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1905 1906 if (separate_io_proj) { 1907 // The caller requested separate projections be used by the fall 1908 // through and exceptional paths, so replace the projections for 1909 // the fall through path. 1910 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1911 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1912 } 1913 1914 // Capture the return value, if any. 1915 Node* ret; 1916 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) { 1917 ret = top(); 1918 } else if (call->tf()->returns_inline_type_as_fields()) { 1919 // Return of multiple values (inline type fields): we create a 1920 // InlineType node, each field is a projection from the call. 1921 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass(); 1922 uint base_input = TypeFunc::Parms; 1923 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, call->method()->signature()->returns_null_free_inline_type()); 1924 } else { 1925 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1926 } 1927 1928 return ret; 1929 } 1930 1931 //--------------------set_predefined_input_for_runtime_call-------------------- 1932 // Reading and setting the memory state is way conservative here. 1933 // The real problem is that I am not doing real Type analysis on memory, 1934 // so I cannot distinguish card mark stores from other stores. Across a GC 1935 // point the Store Barrier and the card mark memory has to agree. I cannot 1936 // have a card mark store and its barrier split across the GC point from 1937 // either above or below. Here I get that to happen by reading ALL of memory. 1938 // A better answer would be to separate out card marks from other memory. 1939 // For now, return the input memory state, so that it can be reused 1940 // after the call, if this call has restricted memory effects. 1941 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1942 // Set fixed predefined input arguments 1943 Node* memory = reset_memory(); 1944 Node* m = narrow_mem == NULL ? memory : narrow_mem; 1945 call->init_req( TypeFunc::Control, control() ); 1946 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1947 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1948 call->init_req( TypeFunc::FramePtr, frameptr() ); 1949 call->init_req( TypeFunc::ReturnAdr, top() ); 1950 return memory; 1951 } 1952 1953 //-------------------set_predefined_output_for_runtime_call-------------------- 1954 // Set control and memory (not i_o) from the call. 1955 // If keep_mem is not NULL, use it for the output state, 1956 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1957 // If hook_mem is NULL, this call produces no memory effects at all. 1958 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1959 // then only that memory slice is taken from the call. 1960 // In the last case, we must put an appropriate memory barrier before 1961 // the call, so as to create the correct anti-dependencies on loads 1962 // preceding the call. 1963 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1964 Node* keep_mem, 1965 const TypePtr* hook_mem) { 1966 // no i/o 1967 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1968 if (keep_mem) { 1969 // First clone the existing memory state 1970 set_all_memory(keep_mem); 1971 if (hook_mem != NULL) { 1972 // Make memory for the call 1973 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1974 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1975 // We also use hook_mem to extract specific effects from arraycopy stubs. 1976 set_memory(mem, hook_mem); 1977 } 1978 // ...else the call has NO memory effects. 1979 1980 // Make sure the call advertises its memory effects precisely. 1981 // This lets us build accurate anti-dependences in gcm.cpp. 1982 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1983 "call node must be constructed correctly"); 1984 } else { 1985 assert(hook_mem == NULL, ""); 1986 // This is not a "slow path" call; all memory comes from the call. 1987 set_all_memory_call(call); 1988 } 1989 } 1990 1991 // Keep track of MergeMems feeding into other MergeMems 1992 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 1993 if (!mem->is_MergeMem()) { 1994 return; 1995 } 1996 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 1997 Node* use = i.get(); 1998 if (use->is_MergeMem()) { 1999 wl.push(use); 2000 } 2001 } 2002 } 2003 2004 // Replace the call with the current state of the kit. 2005 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 2006 JVMState* ejvms = NULL; 2007 if (has_exceptions()) { 2008 ejvms = transfer_exceptions_into_jvms(); 2009 } 2010 2011 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 2012 ReplacedNodes replaced_nodes_exception; 2013 Node* ex_ctl = top(); 2014 2015 SafePointNode* final_state = stop(); 2016 2017 // Find all the needed outputs of this call 2018 CallProjections* callprojs = call->extract_projections(true); 2019 2020 Unique_Node_List wl; 2021 Node* init_mem = call->in(TypeFunc::Memory); 2022 Node* final_mem = final_state->in(TypeFunc::Memory); 2023 Node* final_ctl = final_state->in(TypeFunc::Control); 2024 Node* final_io = final_state->in(TypeFunc::I_O); 2025 2026 // Replace all the old call edges with the edges from the inlining result 2027 if (callprojs->fallthrough_catchproj != NULL) { 2028 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl); 2029 } 2030 if (callprojs->fallthrough_memproj != NULL) { 2031 if (final_mem->is_MergeMem()) { 2032 // Parser's exits MergeMem was not transformed but may be optimized 2033 final_mem = _gvn.transform(final_mem); 2034 } 2035 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem); 2036 add_mergemem_users_to_worklist(wl, final_mem); 2037 } 2038 if (callprojs->fallthrough_ioproj != NULL) { 2039 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io); 2040 } 2041 2042 // Replace the result with the new result if it exists and is used 2043 if (callprojs->resproj[0] != NULL && result != NULL) { 2044 // If the inlined code is dead, the result projections for an inline type returned as 2045 // fields have not been replaced. They will go away once the call is replaced by TOP below. 2046 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()), 2047 "unexpected number of results"); 2048 C->gvn_replace_by(callprojs->resproj[0], result); 2049 } 2050 2051 if (ejvms == NULL) { 2052 // No exception edges to simply kill off those paths 2053 if (callprojs->catchall_catchproj != NULL) { 2054 C->gvn_replace_by(callprojs->catchall_catchproj, C->top()); 2055 } 2056 if (callprojs->catchall_memproj != NULL) { 2057 C->gvn_replace_by(callprojs->catchall_memproj, C->top()); 2058 } 2059 if (callprojs->catchall_ioproj != NULL) { 2060 C->gvn_replace_by(callprojs->catchall_ioproj, C->top()); 2061 } 2062 // Replace the old exception object with top 2063 if (callprojs->exobj != NULL) { 2064 C->gvn_replace_by(callprojs->exobj, C->top()); 2065 } 2066 } else { 2067 GraphKit ekit(ejvms); 2068 2069 // Load my combined exception state into the kit, with all phis transformed: 2070 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 2071 replaced_nodes_exception = ex_map->replaced_nodes(); 2072 2073 Node* ex_oop = ekit.use_exception_state(ex_map); 2074 2075 if (callprojs->catchall_catchproj != NULL) { 2076 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control()); 2077 ex_ctl = ekit.control(); 2078 } 2079 if (callprojs->catchall_memproj != NULL) { 2080 Node* ex_mem = ekit.reset_memory(); 2081 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem); 2082 add_mergemem_users_to_worklist(wl, ex_mem); 2083 } 2084 if (callprojs->catchall_ioproj != NULL) { 2085 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o()); 2086 } 2087 2088 // Replace the old exception object with the newly created one 2089 if (callprojs->exobj != NULL) { 2090 C->gvn_replace_by(callprojs->exobj, ex_oop); 2091 } 2092 } 2093 2094 // Disconnect the call from the graph 2095 call->disconnect_inputs(C); 2096 C->gvn_replace_by(call, C->top()); 2097 2098 // Clean up any MergeMems that feed other MergeMems since the 2099 // optimizer doesn't like that. 2100 while (wl.size() > 0) { 2101 _gvn.transform(wl.pop()); 2102 } 2103 2104 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 2105 replaced_nodes.apply(C, final_ctl); 2106 } 2107 if (!ex_ctl->is_top() && do_replaced_nodes) { 2108 replaced_nodes_exception.apply(C, ex_ctl); 2109 } 2110 } 2111 2112 2113 //------------------------------increment_counter------------------------------ 2114 // for statistics: increment a VM counter by 1 2115 2116 void GraphKit::increment_counter(address counter_addr) { 2117 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2118 increment_counter(adr1); 2119 } 2120 2121 void GraphKit::increment_counter(Node* counter_addr) { 2122 int adr_type = Compile::AliasIdxRaw; 2123 Node* ctrl = control(); 2124 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered); 2125 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1))); 2126 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered); 2127 } 2128 2129 2130 //------------------------------uncommon_trap---------------------------------- 2131 // Bail out to the interpreter in mid-method. Implemented by calling the 2132 // uncommon_trap blob. This helper function inserts a runtime call with the 2133 // right debug info. 2134 Node* GraphKit::uncommon_trap(int trap_request, 2135 ciKlass* klass, const char* comment, 2136 bool must_throw, 2137 bool keep_exact_action) { 2138 if (failing()) stop(); 2139 if (stopped()) return NULL; // trap reachable? 2140 2141 // Note: If ProfileTraps is true, and if a deopt. actually 2142 // occurs here, the runtime will make sure an MDO exists. There is 2143 // no need to call method()->ensure_method_data() at this point. 2144 2145 // Set the stack pointer to the right value for reexecution: 2146 set_sp(reexecute_sp()); 2147 2148 #ifdef ASSERT 2149 if (!must_throw) { 2150 // Make sure the stack has at least enough depth to execute 2151 // the current bytecode. 2152 int inputs, ignored_depth; 2153 if (compute_stack_effects(inputs, ignored_depth)) { 2154 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2155 Bytecodes::name(java_bc()), sp(), inputs); 2156 } 2157 } 2158 #endif 2159 2160 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2161 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2162 2163 switch (action) { 2164 case Deoptimization::Action_maybe_recompile: 2165 case Deoptimization::Action_reinterpret: 2166 // Temporary fix for 6529811 to allow virtual calls to be sure they 2167 // get the chance to go from mono->bi->mega 2168 if (!keep_exact_action && 2169 Deoptimization::trap_request_index(trap_request) < 0 && 2170 too_many_recompiles(reason)) { 2171 // This BCI is causing too many recompilations. 2172 if (C->log() != NULL) { 2173 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2174 Deoptimization::trap_reason_name(reason), 2175 Deoptimization::trap_action_name(action)); 2176 } 2177 action = Deoptimization::Action_none; 2178 trap_request = Deoptimization::make_trap_request(reason, action); 2179 } else { 2180 C->set_trap_can_recompile(true); 2181 } 2182 break; 2183 case Deoptimization::Action_make_not_entrant: 2184 C->set_trap_can_recompile(true); 2185 break; 2186 case Deoptimization::Action_none: 2187 case Deoptimization::Action_make_not_compilable: 2188 break; 2189 default: 2190 #ifdef ASSERT 2191 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2192 #endif 2193 break; 2194 } 2195 2196 if (TraceOptoParse) { 2197 char buf[100]; 2198 tty->print_cr("Uncommon trap %s at bci:%d", 2199 Deoptimization::format_trap_request(buf, sizeof(buf), 2200 trap_request), bci()); 2201 } 2202 2203 CompileLog* log = C->log(); 2204 if (log != NULL) { 2205 int kid = (klass == NULL)? -1: log->identify(klass); 2206 log->begin_elem("uncommon_trap bci='%d'", bci()); 2207 char buf[100]; 2208 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2209 trap_request)); 2210 if (kid >= 0) log->print(" klass='%d'", kid); 2211 if (comment != NULL) log->print(" comment='%s'", comment); 2212 log->end_elem(); 2213 } 2214 2215 // Make sure any guarding test views this path as very unlikely 2216 Node *i0 = control()->in(0); 2217 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2218 IfNode *iff = i0->as_If(); 2219 float f = iff->_prob; // Get prob 2220 if (control()->Opcode() == Op_IfTrue) { 2221 if (f > PROB_UNLIKELY_MAG(4)) 2222 iff->_prob = PROB_MIN; 2223 } else { 2224 if (f < PROB_LIKELY_MAG(4)) 2225 iff->_prob = PROB_MAX; 2226 } 2227 } 2228 2229 // Clear out dead values from the debug info. 2230 kill_dead_locals(); 2231 2232 // Now insert the uncommon trap subroutine call 2233 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2234 const TypePtr* no_memory_effects = NULL; 2235 // Pass the index of the class to be loaded 2236 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2237 (must_throw ? RC_MUST_THROW : 0), 2238 OptoRuntime::uncommon_trap_Type(), 2239 call_addr, "uncommon_trap", no_memory_effects, 2240 intcon(trap_request)); 2241 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2242 "must extract request correctly from the graph"); 2243 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2244 2245 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2246 // The debug info is the only real input to this call. 2247 2248 // Halt-and-catch fire here. The above call should never return! 2249 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2250 PRODUCT_ONLY(COMMA /*reachable*/false)); 2251 _gvn.set_type_bottom(halt); 2252 root()->add_req(halt); 2253 2254 stop_and_kill_map(); 2255 return call; 2256 } 2257 2258 2259 //--------------------------just_allocated_object------------------------------ 2260 // Report the object that was just allocated. 2261 // It must be the case that there are no intervening safepoints. 2262 // We use this to determine if an object is so "fresh" that 2263 // it does not require card marks. 2264 Node* GraphKit::just_allocated_object(Node* current_control) { 2265 Node* ctrl = current_control; 2266 // Object::<init> is invoked after allocation, most of invoke nodes 2267 // will be reduced, but a region node is kept in parse time, we check 2268 // the pattern and skip the region node if it degraded to a copy. 2269 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 && 2270 ctrl->as_Region()->is_copy()) { 2271 ctrl = ctrl->as_Region()->is_copy(); 2272 } 2273 if (C->recent_alloc_ctl() == ctrl) { 2274 return C->recent_alloc_obj(); 2275 } 2276 return NULL; 2277 } 2278 2279 2280 /** 2281 * Record profiling data exact_kls for Node n with the type system so 2282 * that it can propagate it (speculation) 2283 * 2284 * @param n node that the type applies to 2285 * @param exact_kls type from profiling 2286 * @param maybe_null did profiling see null? 2287 * 2288 * @return node with improved type 2289 */ 2290 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2291 const Type* current_type = _gvn.type(n); 2292 assert(UseTypeSpeculation, "type speculation must be on"); 2293 2294 const TypePtr* speculative = current_type->speculative(); 2295 2296 // Should the klass from the profile be recorded in the speculative type? 2297 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2298 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces); 2299 const TypeOopPtr* xtype = tklass->as_instance_type(); 2300 assert(xtype->klass_is_exact(), "Should be exact"); 2301 // Any reason to believe n is not null (from this profiling or a previous one)? 2302 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2303 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2304 // record the new speculative type's depth 2305 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2306 speculative = speculative->with_inline_depth(jvms()->depth()); 2307 } else if (current_type->would_improve_ptr(ptr_kind)) { 2308 // Profiling report that null was never seen so we can change the 2309 // speculative type to non null ptr. 2310 if (ptr_kind == ProfileAlwaysNull) { 2311 speculative = TypePtr::NULL_PTR; 2312 } else { 2313 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2314 const TypePtr* ptr = TypePtr::NOTNULL; 2315 if (speculative != NULL) { 2316 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2317 } else { 2318 speculative = ptr; 2319 } 2320 } 2321 } 2322 2323 if (speculative != current_type->speculative()) { 2324 // Build a type with a speculative type (what we think we know 2325 // about the type but will need a guard when we use it) 2326 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2327 // We're changing the type, we need a new CheckCast node to carry 2328 // the new type. The new type depends on the control: what 2329 // profiling tells us is only valid from here as far as we can 2330 // tell. 2331 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2332 cast = _gvn.transform(cast); 2333 replace_in_map(n, cast); 2334 n = cast; 2335 } 2336 2337 return n; 2338 } 2339 2340 /** 2341 * Record profiling data from receiver profiling at an invoke with the 2342 * type system so that it can propagate it (speculation) 2343 * 2344 * @param n receiver node 2345 * 2346 * @return node with improved type 2347 */ 2348 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2349 if (!UseTypeSpeculation) { 2350 return n; 2351 } 2352 ciKlass* exact_kls = profile_has_unique_klass(); 2353 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2354 if ((java_bc() == Bytecodes::_checkcast || 2355 java_bc() == Bytecodes::_instanceof || 2356 java_bc() == Bytecodes::_aastore) && 2357 method()->method_data()->is_mature()) { 2358 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2359 if (data != NULL) { 2360 if (java_bc() == Bytecodes::_aastore) { 2361 ciKlass* array_type = NULL; 2362 ciKlass* element_type = NULL; 2363 ProfilePtrKind element_ptr = ProfileMaybeNull; 2364 bool flat_array = true; 2365 bool null_free_array = true; 2366 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 2367 exact_kls = element_type; 2368 ptr_kind = element_ptr; 2369 } else { 2370 if (!data->as_BitData()->null_seen()) { 2371 ptr_kind = ProfileNeverNull; 2372 } else { 2373 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2374 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2375 uint i = 0; 2376 for (; i < call->row_limit(); i++) { 2377 ciKlass* receiver = call->receiver(i); 2378 if (receiver != NULL) { 2379 break; 2380 } 2381 } 2382 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2383 } 2384 } 2385 } 2386 } 2387 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2388 } 2389 2390 /** 2391 * Record profiling data from argument profiling at an invoke with the 2392 * type system so that it can propagate it (speculation) 2393 * 2394 * @param dest_method target method for the call 2395 * @param bc what invoke bytecode is this? 2396 */ 2397 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2398 if (!UseTypeSpeculation) { 2399 return; 2400 } 2401 const TypeFunc* tf = TypeFunc::make(dest_method); 2402 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2403 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2404 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2405 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2406 if (is_reference_type(targ->basic_type())) { 2407 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2408 ciKlass* better_type = NULL; 2409 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2410 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2411 } 2412 i++; 2413 } 2414 } 2415 } 2416 2417 /** 2418 * Record profiling data from parameter profiling at an invoke with 2419 * the type system so that it can propagate it (speculation) 2420 */ 2421 void GraphKit::record_profiled_parameters_for_speculation() { 2422 if (!UseTypeSpeculation) { 2423 return; 2424 } 2425 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2426 if (_gvn.type(local(i))->isa_oopptr()) { 2427 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2428 ciKlass* better_type = NULL; 2429 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2430 record_profile_for_speculation(local(i), better_type, ptr_kind); 2431 } 2432 j++; 2433 } 2434 } 2435 } 2436 2437 /** 2438 * Record profiling data from return value profiling at an invoke with 2439 * the type system so that it can propagate it (speculation) 2440 */ 2441 void GraphKit::record_profiled_return_for_speculation() { 2442 if (!UseTypeSpeculation) { 2443 return; 2444 } 2445 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2446 ciKlass* better_type = NULL; 2447 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2448 // If profiling reports a single type for the return value, 2449 // feed it to the type system so it can propagate it as a 2450 // speculative type 2451 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2452 } 2453 } 2454 2455 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2456 if (Matcher::strict_fp_requires_explicit_rounding) { 2457 // (Note: TypeFunc::make has a cache that makes this fast.) 2458 const TypeFunc* tf = TypeFunc::make(dest_method); 2459 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2460 for (int j = 0; j < nargs; j++) { 2461 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2462 if (targ->basic_type() == T_DOUBLE) { 2463 // If any parameters are doubles, they must be rounded before 2464 // the call, dprecision_rounding does gvn.transform 2465 Node *arg = argument(j); 2466 arg = dprecision_rounding(arg); 2467 set_argument(j, arg); 2468 } 2469 } 2470 } 2471 } 2472 2473 // rounding for strict float precision conformance 2474 Node* GraphKit::precision_rounding(Node* n) { 2475 if (Matcher::strict_fp_requires_explicit_rounding) { 2476 #ifdef IA32 2477 if (UseSSE == 0) { 2478 return _gvn.transform(new RoundFloatNode(0, n)); 2479 } 2480 #else 2481 Unimplemented(); 2482 #endif // IA32 2483 } 2484 return n; 2485 } 2486 2487 // rounding for strict double precision conformance 2488 Node* GraphKit::dprecision_rounding(Node *n) { 2489 if (Matcher::strict_fp_requires_explicit_rounding) { 2490 #ifdef IA32 2491 if (UseSSE < 2) { 2492 return _gvn.transform(new RoundDoubleNode(0, n)); 2493 } 2494 #else 2495 Unimplemented(); 2496 #endif // IA32 2497 } 2498 return n; 2499 } 2500 2501 //============================================================================= 2502 // Generate a fast path/slow path idiom. Graph looks like: 2503 // [foo] indicates that 'foo' is a parameter 2504 // 2505 // [in] NULL 2506 // \ / 2507 // CmpP 2508 // Bool ne 2509 // If 2510 // / \ 2511 // True False-<2> 2512 // / | 2513 // / cast_not_null 2514 // Load | | ^ 2515 // [fast_test] | | 2516 // gvn to opt_test | | 2517 // / \ | <1> 2518 // True False | 2519 // | \\ | 2520 // [slow_call] \[fast_result] 2521 // Ctl Val \ \ 2522 // | \ \ 2523 // Catch <1> \ \ 2524 // / \ ^ \ \ 2525 // Ex No_Ex | \ \ 2526 // | \ \ | \ <2> \ 2527 // ... \ [slow_res] | | \ [null_result] 2528 // \ \--+--+--- | | 2529 // \ | / \ | / 2530 // --------Region Phi 2531 // 2532 //============================================================================= 2533 // Code is structured as a series of driver functions all called 'do_XXX' that 2534 // call a set of helper functions. Helper functions first, then drivers. 2535 2536 //------------------------------null_check_oop--------------------------------- 2537 // Null check oop. Set null-path control into Region in slot 3. 2538 // Make a cast-not-nullness use the other not-null control. Return cast. 2539 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2540 bool never_see_null, 2541 bool safe_for_replace, 2542 bool speculative) { 2543 // Initial NULL check taken path 2544 (*null_control) = top(); 2545 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2546 2547 // Generate uncommon_trap: 2548 if (never_see_null && (*null_control) != top()) { 2549 // If we see an unexpected null at a check-cast we record it and force a 2550 // recompile; the offending check-cast will be compiled to handle NULLs. 2551 // If we see more than one offending BCI, then all checkcasts in the 2552 // method will be compiled to handle NULLs. 2553 PreserveJVMState pjvms(this); 2554 set_control(*null_control); 2555 replace_in_map(value, null()); 2556 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2557 uncommon_trap(reason, 2558 Deoptimization::Action_make_not_entrant); 2559 (*null_control) = top(); // NULL path is dead 2560 } 2561 if ((*null_control) == top() && safe_for_replace) { 2562 replace_in_map(value, cast); 2563 } 2564 2565 // Cast away null-ness on the result 2566 return cast; 2567 } 2568 2569 //------------------------------opt_iff---------------------------------------- 2570 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2571 // Return slow-path control. 2572 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2573 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2574 2575 // Fast path taken; set region slot 2 2576 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2577 region->init_req(2,fast_taken); // Capture fast-control 2578 2579 // Fast path not-taken, i.e. slow path 2580 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2581 return slow_taken; 2582 } 2583 2584 //-----------------------------make_runtime_call------------------------------- 2585 Node* GraphKit::make_runtime_call(int flags, 2586 const TypeFunc* call_type, address call_addr, 2587 const char* call_name, 2588 const TypePtr* adr_type, 2589 // The following parms are all optional. 2590 // The first NULL ends the list. 2591 Node* parm0, Node* parm1, 2592 Node* parm2, Node* parm3, 2593 Node* parm4, Node* parm5, 2594 Node* parm6, Node* parm7) { 2595 assert(call_addr != NULL, "must not call NULL targets"); 2596 2597 // Slow-path call 2598 bool is_leaf = !(flags & RC_NO_LEAF); 2599 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2600 if (call_name == NULL) { 2601 assert(!is_leaf, "must supply name for leaf"); 2602 call_name = OptoRuntime::stub_name(call_addr); 2603 } 2604 CallNode* call; 2605 if (!is_leaf) { 2606 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type); 2607 } else if (flags & RC_NO_FP) { 2608 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2609 } else if (flags & RC_VECTOR){ 2610 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte; 2611 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits); 2612 } else { 2613 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2614 } 2615 2616 // The following is similar to set_edges_for_java_call, 2617 // except that the memory effects of the call are restricted to AliasIdxRaw. 2618 2619 // Slow path call has no side-effects, uses few values 2620 bool wide_in = !(flags & RC_NARROW_MEM); 2621 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2622 2623 Node* prev_mem = NULL; 2624 if (wide_in) { 2625 prev_mem = set_predefined_input_for_runtime_call(call); 2626 } else { 2627 assert(!wide_out, "narrow in => narrow out"); 2628 Node* narrow_mem = memory(adr_type); 2629 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2630 } 2631 2632 // Hook each parm in order. Stop looking at the first NULL. 2633 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2634 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2635 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2636 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2637 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2638 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2639 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2640 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2641 /* close each nested if ===> */ } } } } } } } } 2642 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2643 2644 if (!is_leaf) { 2645 // Non-leaves can block and take safepoints: 2646 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2647 } 2648 // Non-leaves can throw exceptions: 2649 if (has_io) { 2650 call->set_req(TypeFunc::I_O, i_o()); 2651 } 2652 2653 if (flags & RC_UNCOMMON) { 2654 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2655 // (An "if" probability corresponds roughly to an unconditional count. 2656 // Sort of.) 2657 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2658 } 2659 2660 Node* c = _gvn.transform(call); 2661 assert(c == call, "cannot disappear"); 2662 2663 if (wide_out) { 2664 // Slow path call has full side-effects. 2665 set_predefined_output_for_runtime_call(call); 2666 } else { 2667 // Slow path call has few side-effects, and/or sets few values. 2668 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2669 } 2670 2671 if (has_io) { 2672 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2673 } 2674 return call; 2675 2676 } 2677 2678 // i2b 2679 Node* GraphKit::sign_extend_byte(Node* in) { 2680 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24))); 2681 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24))); 2682 } 2683 2684 // i2s 2685 Node* GraphKit::sign_extend_short(Node* in) { 2686 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16))); 2687 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16))); 2688 } 2689 2690 2691 //------------------------------merge_memory----------------------------------- 2692 // Merge memory from one path into the current memory state. 2693 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2694 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2695 Node* old_slice = mms.force_memory(); 2696 Node* new_slice = mms.memory2(); 2697 if (old_slice != new_slice) { 2698 PhiNode* phi; 2699 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2700 if (mms.is_empty()) { 2701 // clone base memory Phi's inputs for this memory slice 2702 assert(old_slice == mms.base_memory(), "sanity"); 2703 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2704 _gvn.set_type(phi, Type::MEMORY); 2705 for (uint i = 1; i < phi->req(); i++) { 2706 phi->init_req(i, old_slice->in(i)); 2707 } 2708 } else { 2709 phi = old_slice->as_Phi(); // Phi was generated already 2710 } 2711 } else { 2712 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2713 _gvn.set_type(phi, Type::MEMORY); 2714 } 2715 phi->set_req(new_path, new_slice); 2716 mms.set_memory(phi); 2717 } 2718 } 2719 } 2720 2721 //------------------------------make_slow_call_ex------------------------------ 2722 // Make the exception handler hookups for the slow call 2723 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2724 if (stopped()) return; 2725 2726 // Make a catch node with just two handlers: fall-through and catch-all 2727 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2728 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2729 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2730 _gvn.set_type_bottom(norm); 2731 C->record_for_igvn(norm); 2732 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2733 2734 { PreserveJVMState pjvms(this); 2735 set_control(excp); 2736 set_i_o(i_o); 2737 2738 if (excp != top()) { 2739 if (deoptimize) { 2740 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2741 uncommon_trap(Deoptimization::Reason_unhandled, 2742 Deoptimization::Action_none); 2743 } else { 2744 // Create an exception state also. 2745 // Use an exact type if the caller has a specific exception. 2746 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2747 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2748 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2749 } 2750 } 2751 } 2752 2753 // Get the no-exception control from the CatchNode. 2754 set_control(norm); 2755 } 2756 2757 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) { 2758 Node* cmp = NULL; 2759 switch(bt) { 2760 case T_INT: cmp = new CmpINode(in1, in2); break; 2761 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2762 default: fatal("unexpected comparison type %s", type2name(bt)); 2763 } 2764 cmp = gvn.transform(cmp); 2765 Node* bol = gvn.transform(new BoolNode(cmp, test)); 2766 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2767 gvn.transform(iff); 2768 if (!bol->is_Con()) gvn.record_for_igvn(iff); 2769 return iff; 2770 } 2771 2772 //-------------------------------gen_subtype_check----------------------------- 2773 // Generate a subtyping check. Takes as input the subtype and supertype. 2774 // Returns 2 values: sets the default control() to the true path and returns 2775 // the false path. Only reads invariant memory; sets no (visible) memory. 2776 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2777 // but that's not exposed to the optimizer. This call also doesn't take in an 2778 // Object; if you wish to check an Object you need to load the Object's class 2779 // prior to coming here. 2780 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) { 2781 Compile* C = gvn.C; 2782 if ((*ctrl)->is_top()) { 2783 return C->top(); 2784 } 2785 2786 // Fast check for identical types, perhaps identical constants. 2787 // The types can even be identical non-constants, in cases 2788 // involving Array.newInstance, Object.clone, etc. 2789 if (subklass == superklass) 2790 return C->top(); // false path is dead; no test needed. 2791 2792 if (gvn.type(superklass)->singleton()) { 2793 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr(); 2794 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr(); 2795 2796 // In the common case of an exact superklass, try to fold up the 2797 // test before generating code. You may ask, why not just generate 2798 // the code and then let it fold up? The answer is that the generated 2799 // code will necessarily include null checks, which do not always 2800 // completely fold away. If they are also needless, then they turn 2801 // into a performance loss. Example: 2802 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2803 // Here, the type of 'fa' is often exact, so the store check 2804 // of fa[1]=x will fold up, without testing the nullness of x. 2805 switch (C->static_subtype_check(superk, subk)) { 2806 case Compile::SSC_always_false: 2807 { 2808 Node* always_fail = *ctrl; 2809 *ctrl = gvn.C->top(); 2810 return always_fail; 2811 } 2812 case Compile::SSC_always_true: 2813 return C->top(); 2814 case Compile::SSC_easy_test: 2815 { 2816 // Just do a direct pointer compare and be done. 2817 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2818 *ctrl = gvn.transform(new IfTrueNode(iff)); 2819 return gvn.transform(new IfFalseNode(iff)); 2820 } 2821 case Compile::SSC_full_test: 2822 break; 2823 default: 2824 ShouldNotReachHere(); 2825 } 2826 } 2827 2828 // %%% Possible further optimization: Even if the superklass is not exact, 2829 // if the subklass is the unique subtype of the superklass, the check 2830 // will always succeed. We could leave a dependency behind to ensure this. 2831 2832 // First load the super-klass's check-offset 2833 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2834 Node* m = C->immutable_memory(); 2835 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2836 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2837 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2838 2839 // Load from the sub-klass's super-class display list, or a 1-word cache of 2840 // the secondary superclass list, or a failing value with a sentinel offset 2841 // if the super-klass is an interface or exceptionally deep in the Java 2842 // hierarchy and we have to scan the secondary superclass list the hard way. 2843 // Worst-case type is a little odd: NULL is allowed as a result (usually 2844 // klass loads can never produce a NULL). 2845 Node *chk_off_X = chk_off; 2846 #ifdef _LP64 2847 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X)); 2848 #endif 2849 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X)); 2850 // For some types like interfaces the following loadKlass is from a 1-word 2851 // cache which is mutable so can't use immutable memory. Other 2852 // types load from the super-class display table which is immutable. 2853 Node *kmem = C->immutable_memory(); 2854 // secondary_super_cache is not immutable but can be treated as such because: 2855 // - no ideal node writes to it in a way that could cause an 2856 // incorrect/missed optimization of the following Load. 2857 // - it's a cache so, worse case, not reading the latest value 2858 // wouldn't cause incorrect execution 2859 if (might_be_cache && mem != NULL) { 2860 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem; 2861 } 2862 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL)); 2863 2864 // Compile speed common case: ARE a subtype and we canNOT fail 2865 if( superklass == nkls ) 2866 return C->top(); // false path is dead; no test needed. 2867 2868 // See if we get an immediate positive hit. Happens roughly 83% of the 2869 // time. Test to see if the value loaded just previously from the subklass 2870 // is exactly the superklass. 2871 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2872 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1)); 2873 *ctrl = gvn.transform(new IfFalseNode(iff1)); 2874 2875 // Compile speed common case: Check for being deterministic right now. If 2876 // chk_off is a constant and not equal to cacheoff then we are NOT a 2877 // subklass. In this case we need exactly the 1 test above and we can 2878 // return those results immediately. 2879 if (!might_be_cache) { 2880 Node* not_subtype_ctrl = *ctrl; 2881 *ctrl = iftrue1; // We need exactly the 1 test above 2882 return not_subtype_ctrl; 2883 } 2884 2885 // Gather the various success & failures here 2886 RegionNode *r_ok_subtype = new RegionNode(4); 2887 gvn.record_for_igvn(r_ok_subtype); 2888 RegionNode *r_not_subtype = new RegionNode(3); 2889 gvn.record_for_igvn(r_not_subtype); 2890 2891 r_ok_subtype->init_req(1, iftrue1); 2892 2893 // Check for immediate negative hit. Happens roughly 11% of the time (which 2894 // is roughly 63% of the remaining cases). Test to see if the loaded 2895 // check-offset points into the subklass display list or the 1-element 2896 // cache. If it points to the display (and NOT the cache) and the display 2897 // missed then it's not a subtype. 2898 Node *cacheoff = gvn.intcon(cacheoff_con); 2899 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2900 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2))); 2901 *ctrl = gvn.transform(new IfFalseNode(iff2)); 2902 2903 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2904 // No performance impact (too rare) but allows sharing of secondary arrays 2905 // which has some footprint reduction. 2906 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2907 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3))); 2908 *ctrl = gvn.transform(new IfFalseNode(iff3)); 2909 2910 // -- Roads not taken here: -- 2911 // We could also have chosen to perform the self-check at the beginning 2912 // of this code sequence, as the assembler does. This would not pay off 2913 // the same way, since the optimizer, unlike the assembler, can perform 2914 // static type analysis to fold away many successful self-checks. 2915 // Non-foldable self checks work better here in second position, because 2916 // the initial primary superclass check subsumes a self-check for most 2917 // types. An exception would be a secondary type like array-of-interface, 2918 // which does not appear in its own primary supertype display. 2919 // Finally, we could have chosen to move the self-check into the 2920 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2921 // dependent manner. But it is worthwhile to have the check here, 2922 // where it can be perhaps be optimized. The cost in code space is 2923 // small (register compare, branch). 2924 2925 // Now do a linear scan of the secondary super-klass array. Again, no real 2926 // performance impact (too rare) but it's gotta be done. 2927 // Since the code is rarely used, there is no penalty for moving it 2928 // out of line, and it can only improve I-cache density. 2929 // The decision to inline or out-of-line this final check is platform 2930 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2931 Node* psc = gvn.transform( 2932 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2933 2934 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2935 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4))); 2936 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4))); 2937 2938 // Return false path; set default control to true path. 2939 *ctrl = gvn.transform(r_ok_subtype); 2940 return gvn.transform(r_not_subtype); 2941 } 2942 2943 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) { 2944 const Type* sub_t = _gvn.type(obj_or_subklass); 2945 if (sub_t->make_oopptr() != NULL && sub_t->make_oopptr()->is_inlinetypeptr()) { 2946 sub_t = TypeKlassPtr::make(sub_t->inline_klass()); 2947 obj_or_subklass = makecon(sub_t); 2948 } 2949 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over 2950 ExpandSubTypeCheckAtParseTime; // forced expansion 2951 if (expand_subtype_check) { 2952 MergeMemNode* mem = merged_memory(); 2953 Node* ctrl = control(); 2954 Node* subklass = obj_or_subklass; 2955 if (!sub_t->isa_klassptr()) { 2956 subklass = load_object_klass(obj_or_subklass); 2957 } 2958 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn); 2959 set_control(ctrl); 2960 return n; 2961 } 2962 2963 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass)); 2964 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq)); 2965 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2966 set_control(_gvn.transform(new IfTrueNode(iff))); 2967 return _gvn.transform(new IfFalseNode(iff)); 2968 } 2969 2970 // Profile-driven exact type check: 2971 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2972 float prob, Node* *casted_receiver) { 2973 assert(!klass->is_interface(), "no exact type check on interfaces"); 2974 Node* fail = top(); 2975 const Type* rec_t = _gvn.type(receiver); 2976 if (rec_t->is_inlinetypeptr()) { 2977 if (klass->equals(rec_t->inline_klass())) { 2978 (*casted_receiver) = receiver; // Always passes 2979 } else { 2980 (*casted_receiver) = top(); // Always fails 2981 fail = control(); 2982 set_control(top()); 2983 } 2984 return fail; 2985 } 2986 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces); 2987 Node* recv_klass = load_object_klass(receiver); 2988 fail = type_check(recv_klass, tklass, prob); 2989 2990 if (!stopped()) { 2991 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 2992 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2993 assert(recv_xtype->klass_is_exact(), ""); 2994 2995 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts 2996 // Subsume downstream occurrences of receiver with a cast to 2997 // recv_xtype, since now we know what the type will be. 2998 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2999 Node* res = _gvn.transform(cast); 3000 if (recv_xtype->is_inlinetypeptr()) { 3001 assert(!gvn().type(res)->maybe_null(), "receiver should never be null"); 3002 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass()); 3003 } 3004 (*casted_receiver) = res; 3005 // (User must make the replace_in_map call.) 3006 } 3007 } 3008 3009 return fail; 3010 } 3011 3012 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 3013 float prob) { 3014 Node* want_klass = makecon(tklass); 3015 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass)); 3016 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3017 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 3018 set_control(_gvn.transform(new IfTrueNode (iff))); 3019 Node* fail = _gvn.transform(new IfFalseNode(iff)); 3020 return fail; 3021 } 3022 3023 //------------------------------subtype_check_receiver------------------------- 3024 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 3025 Node** casted_receiver) { 3026 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve(); 3027 Node* want_klass = makecon(tklass); 3028 3029 Node* slow_ctl = gen_subtype_check(receiver, want_klass); 3030 3031 // Ignore interface type information until interface types are properly tracked. 3032 if (!stopped() && !klass->is_interface()) { 3033 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3034 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 3035 if (receiver_type != NULL && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts 3036 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type)); 3037 if (recv_type->is_inlinetypeptr()) { 3038 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass()); 3039 } 3040 (*casted_receiver) = cast; 3041 } 3042 } 3043 3044 return slow_ctl; 3045 } 3046 3047 //------------------------------seems_never_null------------------------------- 3048 // Use null_seen information if it is available from the profile. 3049 // If we see an unexpected null at a type check we record it and force a 3050 // recompile; the offending check will be recompiled to handle NULLs. 3051 // If we see several offending BCIs, then all checks in the 3052 // method will be recompiled. 3053 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 3054 speculating = !_gvn.type(obj)->speculative_maybe_null(); 3055 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 3056 if (UncommonNullCast // Cutout for this technique 3057 && obj != null() // And not the -Xcomp stupid case? 3058 && !too_many_traps(reason) 3059 ) { 3060 if (speculating) { 3061 return true; 3062 } 3063 if (data == NULL) 3064 // Edge case: no mature data. Be optimistic here. 3065 return true; 3066 // If the profile has not seen a null, assume it won't happen. 3067 assert(java_bc() == Bytecodes::_checkcast || 3068 java_bc() == Bytecodes::_instanceof || 3069 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 3070 if (java_bc() == Bytecodes::_aastore) { 3071 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull; 3072 } 3073 return !data->as_BitData()->null_seen(); 3074 } 3075 speculating = false; 3076 return false; 3077 } 3078 3079 void GraphKit::guard_klass_being_initialized(Node* klass) { 3080 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 3081 Node* adr = basic_plus_adr(top(), klass, init_state_off); 3082 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr, 3083 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 3084 T_BYTE, MemNode::unordered); 3085 init_state = _gvn.transform(init_state); 3086 3087 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 3088 3089 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 3090 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3091 3092 { BuildCutout unless(this, tst, PROB_MAX); 3093 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 3094 } 3095 } 3096 3097 void GraphKit::guard_init_thread(Node* klass) { 3098 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 3099 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 3100 3101 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr, 3102 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 3103 T_ADDRESS, MemNode::unordered); 3104 init_thread = _gvn.transform(init_thread); 3105 3106 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 3107 3108 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 3109 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3110 3111 { BuildCutout unless(this, tst, PROB_MAX); 3112 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 3113 } 3114 } 3115 3116 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3117 if (ik->is_being_initialized()) { 3118 if (C->needs_clinit_barrier(ik, context)) { 3119 Node* klass = makecon(TypeKlassPtr::make(ik)); 3120 guard_klass_being_initialized(klass); 3121 guard_init_thread(klass); 3122 insert_mem_bar(Op_MemBarCPUOrder); 3123 } 3124 } else if (ik->is_initialized()) { 3125 return; // no barrier needed 3126 } else { 3127 uncommon_trap(Deoptimization::Reason_uninitialized, 3128 Deoptimization::Action_reinterpret, 3129 NULL); 3130 } 3131 } 3132 3133 //------------------------maybe_cast_profiled_receiver------------------------- 3134 // If the profile has seen exactly one type, narrow to exactly that type. 3135 // Subsequent type checks will always fold up. 3136 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3137 const TypeKlassPtr* require_klass, 3138 ciKlass* spec_klass, 3139 bool safe_for_replace) { 3140 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 3141 3142 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 3143 3144 // Make sure we haven't already deoptimized from this tactic. 3145 if (too_many_traps_or_recompiles(reason)) 3146 return NULL; 3147 3148 // (No, this isn't a call, but it's enough like a virtual call 3149 // to use the same ciMethod accessor to get the profile info...) 3150 // If we have a speculative type use it instead of profiling (which 3151 // may not help us) 3152 ciKlass* exact_kls = spec_klass; 3153 if (exact_kls == NULL) { 3154 if (java_bc() == Bytecodes::_aastore) { 3155 ciKlass* array_type = NULL; 3156 ciKlass* element_type = NULL; 3157 ProfilePtrKind element_ptr = ProfileMaybeNull; 3158 bool flat_array = true; 3159 bool null_free_array = true; 3160 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 3161 exact_kls = element_type; 3162 } else { 3163 exact_kls = profile_has_unique_klass(); 3164 } 3165 } 3166 if (exact_kls != NULL) {// no cast failures here 3167 if (require_klass == NULL || 3168 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) { 3169 // If we narrow the type to match what the type profile sees or 3170 // the speculative type, we can then remove the rest of the 3171 // cast. 3172 // This is a win, even if the exact_kls is very specific, 3173 // because downstream operations, such as method calls, 3174 // will often benefit from the sharper type. 3175 Node* exact_obj = not_null_obj; // will get updated in place... 3176 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3177 &exact_obj); 3178 { PreserveJVMState pjvms(this); 3179 set_control(slow_ctl); 3180 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3181 } 3182 if (safe_for_replace) { 3183 replace_in_map(not_null_obj, exact_obj); 3184 } 3185 return exact_obj; 3186 } 3187 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3188 } 3189 3190 return NULL; 3191 } 3192 3193 /** 3194 * Cast obj to type and emit guard unless we had too many traps here 3195 * already 3196 * 3197 * @param obj node being casted 3198 * @param type type to cast the node to 3199 * @param not_null true if we know node cannot be null 3200 */ 3201 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3202 ciKlass* type, 3203 bool not_null) { 3204 if (stopped()) { 3205 return obj; 3206 } 3207 3208 // type == NULL if profiling tells us this object is always null 3209 if (type != NULL) { 3210 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3211 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3212 3213 if (!too_many_traps_or_recompiles(null_reason) && 3214 !too_many_traps_or_recompiles(class_reason)) { 3215 Node* not_null_obj = NULL; 3216 // not_null is true if we know the object is not null and 3217 // there's no need for a null check 3218 if (!not_null) { 3219 Node* null_ctl = top(); 3220 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3221 assert(null_ctl->is_top(), "no null control here"); 3222 } else { 3223 not_null_obj = obj; 3224 } 3225 3226 Node* exact_obj = not_null_obj; 3227 ciKlass* exact_kls = type; 3228 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3229 &exact_obj); 3230 { 3231 PreserveJVMState pjvms(this); 3232 set_control(slow_ctl); 3233 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3234 } 3235 replace_in_map(not_null_obj, exact_obj); 3236 obj = exact_obj; 3237 } 3238 } else { 3239 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3240 Node* exact_obj = null_assert(obj); 3241 replace_in_map(obj, exact_obj); 3242 obj = exact_obj; 3243 } 3244 } 3245 return obj; 3246 } 3247 3248 //-------------------------------gen_instanceof-------------------------------- 3249 // Generate an instance-of idiom. Used by both the instance-of bytecode 3250 // and the reflective instance-of call. 3251 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3252 kill_dead_locals(); // Benefit all the uncommon traps 3253 assert( !stopped(), "dead parse path should be checked in callers" ); 3254 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3255 "must check for not-null not-dead klass in callers"); 3256 3257 // Make the merge point 3258 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3259 RegionNode* region = new RegionNode(PATH_LIMIT); 3260 Node* phi = new PhiNode(region, TypeInt::BOOL); 3261 C->set_has_split_ifs(true); // Has chance for split-if optimization 3262 3263 ciProfileData* data = NULL; 3264 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3265 data = method()->method_data()->bci_to_data(bci()); 3266 } 3267 bool speculative_not_null = false; 3268 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3269 && seems_never_null(obj, data, speculative_not_null)); 3270 3271 // Null check; get casted pointer; set region slot 3 3272 Node* null_ctl = top(); 3273 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3274 3275 // If not_null_obj is dead, only null-path is taken 3276 if (stopped()) { // Doing instance-of on a NULL? 3277 set_control(null_ctl); 3278 return intcon(0); 3279 } 3280 region->init_req(_null_path, null_ctl); 3281 phi ->init_req(_null_path, intcon(0)); // Set null path value 3282 if (null_ctl == top()) { 3283 // Do this eagerly, so that pattern matches like is_diamond_phi 3284 // will work even during parsing. 3285 assert(_null_path == PATH_LIMIT-1, "delete last"); 3286 region->del_req(_null_path); 3287 phi ->del_req(_null_path); 3288 } 3289 3290 // Do we know the type check always succeed? 3291 bool known_statically = false; 3292 if (_gvn.type(superklass)->singleton()) { 3293 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr(); 3294 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type(); 3295 if (subk != NULL && subk->is_loaded()) { 3296 int static_res = C->static_subtype_check(superk, subk); 3297 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3298 } 3299 } 3300 3301 if (!known_statically) { 3302 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3303 // We may not have profiling here or it may not help us. If we 3304 // have a speculative type use it to perform an exact cast. 3305 ciKlass* spec_obj_type = obj_type->speculative_type(); 3306 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 3307 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 3308 if (stopped()) { // Profile disagrees with this path. 3309 set_control(null_ctl); // Null is the only remaining possibility. 3310 return intcon(0); 3311 } 3312 if (cast_obj != NULL) { 3313 not_null_obj = cast_obj; 3314 } 3315 } 3316 } 3317 3318 // Generate the subtype check 3319 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3320 3321 // Plug in the success path to the general merge in slot 1. 3322 region->init_req(_obj_path, control()); 3323 phi ->init_req(_obj_path, intcon(1)); 3324 3325 // Plug in the failing path to the general merge in slot 2. 3326 region->init_req(_fail_path, not_subtype_ctrl); 3327 phi ->init_req(_fail_path, intcon(0)); 3328 3329 // Return final merged results 3330 set_control( _gvn.transform(region) ); 3331 record_for_igvn(region); 3332 3333 // If we know the type check always succeeds then we don't use the 3334 // profiling data at this bytecode. Don't lose it, feed it to the 3335 // type system as a speculative type. 3336 if (safe_for_replace) { 3337 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3338 replace_in_map(obj, casted_obj); 3339 } 3340 3341 return _gvn.transform(phi); 3342 } 3343 3344 //-------------------------------gen_checkcast--------------------------------- 3345 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3346 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3347 // uncommon-trap paths work. Adjust stack after this call. 3348 // If failure_control is supplied and not null, it is filled in with 3349 // the control edge for the cast failure. Otherwise, an appropriate 3350 // uncommon trap or exception is thrown. 3351 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) { 3352 kill_dead_locals(); // Benefit all the uncommon traps 3353 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr()->try_improve(); 3354 const TypeOopPtr *toop = tk->cast_to_exactness(false)->as_instance_type(); 3355 bool safe_for_replace = (failure_control == NULL); 3356 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer"); 3357 3358 // Fast cutout: Check the case that the cast is vacuously true. 3359 // This detects the common cases where the test will short-circuit 3360 // away completely. We do this before we perform the null check, 3361 // because if the test is going to turn into zero code, we don't 3362 // want a residual null check left around. (Causes a slowdown, 3363 // for example, in some objArray manipulations, such as a[i]=a[j].) 3364 if (tk->singleton()) { 3365 const TypeKlassPtr* kptr = NULL; 3366 const Type* t = _gvn.type(obj); 3367 if (t->isa_oop_ptr()) { 3368 kptr = t->is_oopptr()->as_klass_type(); 3369 } else if (obj->is_InlineType()) { 3370 ciInlineKlass* vk = t->inline_klass(); 3371 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0)); 3372 } 3373 if (kptr != NULL) { 3374 switch (C->static_subtype_check(tk, kptr)) { 3375 case Compile::SSC_always_true: 3376 // If we know the type check always succeed then we don't use 3377 // the profiling data at this bytecode. Don't lose it, feed it 3378 // to the type system as a speculative type. 3379 obj = record_profiled_receiver_for_speculation(obj); 3380 if (null_free) { 3381 assert(safe_for_replace, "must be"); 3382 obj = null_check(obj); 3383 } 3384 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized"); 3385 return obj; 3386 case Compile::SSC_always_false: 3387 if (null_free) { 3388 assert(safe_for_replace, "must be"); 3389 obj = null_check(obj); 3390 } 3391 // It needs a null check because a null will *pass* the cast check. 3392 if (t->isa_oopptr() != NULL && !t->is_oopptr()->maybe_null()) { 3393 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3394 Deoptimization::DeoptReason reason = is_aastore ? 3395 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3396 builtin_throw(reason); 3397 return top(); 3398 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3399 return null_assert(obj); 3400 } 3401 break; // Fall through to full check 3402 default: 3403 break; 3404 } 3405 } 3406 } 3407 3408 ciProfileData* data = NULL; 3409 if (failure_control == NULL) { // use MDO in regular case only 3410 assert(java_bc() == Bytecodes::_aastore || 3411 java_bc() == Bytecodes::_checkcast, 3412 "interpreter profiles type checks only for these BCs"); 3413 if (method()->method_data()->is_mature()) { 3414 data = method()->method_data()->bci_to_data(bci()); 3415 } 3416 } 3417 3418 // Make the merge point 3419 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3420 RegionNode* region = new RegionNode(PATH_LIMIT); 3421 Node* phi = new PhiNode(region, toop); 3422 _gvn.set_type(region, Type::CONTROL); 3423 _gvn.set_type(phi, toop); 3424 3425 C->set_has_split_ifs(true); // Has chance for split-if optimization 3426 3427 // Use null-cast information if it is available 3428 bool speculative_not_null = false; 3429 bool never_see_null = ((failure_control == NULL) // regular case only 3430 && seems_never_null(obj, data, speculative_not_null)); 3431 3432 if (obj->is_InlineType()) { 3433 // Re-execute if buffering during triggers deoptimization 3434 PreserveReexecuteState preexecs(this); 3435 jvms()->set_should_reexecute(true); 3436 obj = obj->as_InlineType()->buffer(this, safe_for_replace); 3437 } 3438 3439 // Null check; get casted pointer; set region slot 3 3440 Node* null_ctl = top(); 3441 Node* not_null_obj = NULL; 3442 if (null_free) { 3443 assert(safe_for_replace, "must be"); 3444 not_null_obj = null_check(obj); 3445 } else { 3446 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3447 } 3448 3449 // If not_null_obj is dead, only null-path is taken 3450 if (stopped()) { // Doing instance-of on a NULL? 3451 set_control(null_ctl); 3452 if (toop->is_inlinetypeptr()) { 3453 return InlineTypeNode::make_null(_gvn, toop->inline_klass()); 3454 } 3455 return null(); 3456 } 3457 region->init_req(_null_path, null_ctl); 3458 phi ->init_req(_null_path, null()); // Set null path value 3459 if (null_ctl == top()) { 3460 // Do this eagerly, so that pattern matches like is_diamond_phi 3461 // will work even during parsing. 3462 assert(_null_path == PATH_LIMIT-1, "delete last"); 3463 region->del_req(_null_path); 3464 phi ->del_req(_null_path); 3465 } 3466 3467 Node* cast_obj = NULL; 3468 if (tk->klass_is_exact()) { 3469 // The following optimization tries to statically cast the speculative type of the object 3470 // (for example obtained during profiling) to the type of the superklass and then do a 3471 // dynamic check that the type of the object is what we expect. To work correctly 3472 // for checkcast and aastore the type of superklass should be exact. 3473 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3474 // We may not have profiling here or it may not help us. If we have 3475 // a speculative type use it to perform an exact cast. 3476 ciKlass* spec_obj_type = obj_type->speculative_type(); 3477 if (spec_obj_type != NULL || data != NULL) { 3478 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk, spec_obj_type, safe_for_replace); 3479 if (cast_obj != NULL) { 3480 if (failure_control != NULL) // failure is now impossible 3481 (*failure_control) = top(); 3482 // adjust the type of the phi to the exact klass: 3483 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3484 } 3485 } 3486 } 3487 3488 if (cast_obj == NULL) { 3489 // Generate the subtype check 3490 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3491 3492 // Plug in success path into the merge 3493 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3494 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3495 if (failure_control == NULL) { 3496 if (not_subtype_ctrl != top()) { // If failure is possible 3497 PreserveJVMState pjvms(this); 3498 set_control(not_subtype_ctrl); 3499 Node* obj_klass = NULL; 3500 if (not_null_obj->is_InlineType()) { 3501 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass())); 3502 } else { 3503 obj_klass = load_object_klass(not_null_obj); 3504 } 3505 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3506 Deoptimization::DeoptReason reason = is_aastore ? 3507 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3508 builtin_throw(reason); 3509 } 3510 } else { 3511 (*failure_control) = not_subtype_ctrl; 3512 } 3513 } 3514 3515 region->init_req(_obj_path, control()); 3516 phi ->init_req(_obj_path, cast_obj); 3517 3518 // A merge of NULL or Casted-NotNull obj 3519 Node* res = _gvn.transform(phi); 3520 3521 // Note I do NOT always 'replace_in_map(obj,result)' here. 3522 // if( tk->klass()->can_be_primary_super() ) 3523 // This means that if I successfully store an Object into an array-of-String 3524 // I 'forget' that the Object is really now known to be a String. I have to 3525 // do this because we don't have true union types for interfaces - if I store 3526 // a Baz into an array-of-Interface and then tell the optimizer it's an 3527 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3528 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3529 // replace_in_map( obj, res ); 3530 3531 // Return final merged results 3532 set_control( _gvn.transform(region) ); 3533 record_for_igvn(region); 3534 3535 bool not_inline = !toop->can_be_inline_type(); 3536 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array()); 3537 if (EnableValhalla && not_flattened) { 3538 // Check if obj has been loaded from an array 3539 obj = obj->isa_DecodeN() ? obj->in(1) : obj; 3540 Node* array = NULL; 3541 if (obj->isa_Load()) { 3542 Node* address = obj->in(MemNode::Address); 3543 if (address->isa_AddP()) { 3544 array = address->as_AddP()->in(AddPNode::Base); 3545 } 3546 } else if (obj->is_Phi()) { 3547 Node* region = obj->in(0); 3548 // TODO make this more robust (see JDK-8231346) 3549 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) { 3550 IfNode* iff = region->in(2)->in(0)->isa_If(); 3551 if (iff != NULL) { 3552 iff->is_flat_array_check(&_gvn, &array); 3553 } 3554 } 3555 } 3556 if (array != NULL) { 3557 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr(); 3558 if (ary_t != NULL) { 3559 if (!ary_t->is_not_null_free() && not_inline) { 3560 // Casting array element to a non-inline-type, mark array as not null-free. 3561 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free())); 3562 replace_in_map(array, cast); 3563 } else if (!ary_t->is_not_flat()) { 3564 // Casting array element to a non-flattened type, mark array as not flat. 3565 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat())); 3566 replace_in_map(array, cast); 3567 } 3568 } 3569 } 3570 } 3571 3572 if (!stopped() && !res->is_InlineType()) { 3573 res = record_profiled_receiver_for_speculation(res); 3574 if (toop->is_inlinetypeptr()) { 3575 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null()); 3576 res = vt; 3577 if (safe_for_replace) { 3578 replace_in_map(obj, vt); 3579 replace_in_map(not_null_obj, vt); 3580 replace_in_map(res, vt); 3581 } 3582 } 3583 } 3584 return res; 3585 } 3586 3587 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) { 3588 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3589 Node* mark = make_load(NULL, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3590 Node* mask = MakeConX(markWord::inline_type_pattern); 3591 Node* masked = _gvn.transform(new AndXNode(mark, mask)); 3592 Node* cmp = _gvn.transform(new CmpXNode(masked, mask)); 3593 return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne)); 3594 } 3595 3596 Node* GraphKit::is_val_mirror(Node* mirror) { 3597 Node* p = basic_plus_adr(mirror, java_lang_Class::secondary_mirror_offset()); 3598 Node* secondary_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP); 3599 Node* cmp = _gvn.transform(new CmpPNode(mirror, secondary_mirror)); 3600 return _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3601 } 3602 3603 Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) { 3604 Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset())); 3605 // Make sure to use immutable memory here to enable hoisting the check out of loops 3606 Node* lh_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered)); 3607 Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask))); 3608 Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val))); 3609 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne)); 3610 } 3611 3612 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) { 3613 // We can't use immutable memory here because the mark word is mutable. 3614 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the 3615 // check is moved out of loops (mainly to enable loop unswitching). 3616 Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory(); 3617 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, array_or_klass)); 3618 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN 3619 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne)); 3620 } 3621 3622 Node* GraphKit::null_free_array_test(Node* klass, bool null_free) { 3623 return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free); 3624 } 3625 3626 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null 3627 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) { 3628 RegionNode* region = new RegionNode(3); 3629 Node* null_ctl = top(); 3630 null_check_oop(val, &null_ctl); 3631 if (null_ctl != top()) { 3632 PreserveJVMState pjvms(this); 3633 set_control(null_ctl); 3634 { 3635 // Deoptimize if null-free array 3636 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX); 3637 inc_sp(nargs); 3638 uncommon_trap(Deoptimization::Reason_null_check, 3639 Deoptimization::Action_none); 3640 } 3641 region->init_req(1, control()); 3642 } 3643 region->init_req(2, control()); 3644 set_control(_gvn.transform(region)); 3645 record_for_igvn(region); 3646 if (_gvn.type(val) == TypePtr::NULL_PTR) { 3647 // Since we were just successfully storing null, the array can't be null free. 3648 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 3649 ary_t = ary_t->cast_to_not_null_free(); 3650 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t)); 3651 if (safe_for_replace) { 3652 replace_in_map(ary, cast); 3653 } 3654 ary = cast; 3655 } 3656 return ary; 3657 } 3658 3659 //------------------------------next_monitor----------------------------------- 3660 // What number should be given to the next monitor? 3661 int GraphKit::next_monitor() { 3662 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3663 int next = current + C->sync_stack_slots(); 3664 // Keep the toplevel high water mark current: 3665 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3666 return current; 3667 } 3668 3669 //------------------------------insert_mem_bar--------------------------------- 3670 // Memory barrier to avoid floating things around 3671 // The membar serves as a pinch point between both control and all memory slices. 3672 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3673 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3674 mb->init_req(TypeFunc::Control, control()); 3675 mb->init_req(TypeFunc::Memory, reset_memory()); 3676 Node* membar = _gvn.transform(mb); 3677 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3678 set_all_memory_call(membar); 3679 return membar; 3680 } 3681 3682 //-------------------------insert_mem_bar_volatile---------------------------- 3683 // Memory barrier to avoid floating things around 3684 // The membar serves as a pinch point between both control and memory(alias_idx). 3685 // If you want to make a pinch point on all memory slices, do not use this 3686 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3687 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3688 // When Parse::do_put_xxx updates a volatile field, it appends a series 3689 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3690 // The first membar is on the same memory slice as the field store opcode. 3691 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3692 // All the other membars (for other volatile slices, including AliasIdxBot, 3693 // which stands for all unknown volatile slices) are control-dependent 3694 // on the first membar. This prevents later volatile loads or stores 3695 // from sliding up past the just-emitted store. 3696 3697 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3698 mb->set_req(TypeFunc::Control,control()); 3699 if (alias_idx == Compile::AliasIdxBot) { 3700 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3701 } else { 3702 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3703 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3704 } 3705 Node* membar = _gvn.transform(mb); 3706 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3707 if (alias_idx == Compile::AliasIdxBot) { 3708 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3709 } else { 3710 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3711 } 3712 return membar; 3713 } 3714 3715 //------------------------------shared_lock------------------------------------ 3716 // Emit locking code. 3717 FastLockNode* GraphKit::shared_lock(Node* obj) { 3718 // bci is either a monitorenter bc or InvocationEntryBci 3719 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3720 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3721 3722 if( !GenerateSynchronizationCode ) 3723 return NULL; // Not locking things? 3724 3725 if (stopped()) // Dead monitor? 3726 return NULL; 3727 3728 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3729 3730 // Box the stack location 3731 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3732 Node* mem = reset_memory(); 3733 3734 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3735 3736 // Create the rtm counters for this fast lock if needed. 3737 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3738 3739 // Add monitor to debug info for the slow path. If we block inside the 3740 // slow path and de-opt, we need the monitor hanging around 3741 map()->push_monitor( flock ); 3742 3743 const TypeFunc *tf = LockNode::lock_type(); 3744 LockNode *lock = new LockNode(C, tf); 3745 3746 lock->init_req( TypeFunc::Control, control() ); 3747 lock->init_req( TypeFunc::Memory , mem ); 3748 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3749 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3750 lock->init_req( TypeFunc::ReturnAdr, top() ); 3751 3752 lock->init_req(TypeFunc::Parms + 0, obj); 3753 lock->init_req(TypeFunc::Parms + 1, box); 3754 lock->init_req(TypeFunc::Parms + 2, flock); 3755 add_safepoint_edges(lock); 3756 3757 lock = _gvn.transform( lock )->as_Lock(); 3758 3759 // lock has no side-effects, sets few values 3760 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3761 3762 insert_mem_bar(Op_MemBarAcquireLock); 3763 3764 // Add this to the worklist so that the lock can be eliminated 3765 record_for_igvn(lock); 3766 3767 #ifndef PRODUCT 3768 if (PrintLockStatistics) { 3769 // Update the counter for this lock. Don't bother using an atomic 3770 // operation since we don't require absolute accuracy. 3771 lock->create_lock_counter(map()->jvms()); 3772 increment_counter(lock->counter()->addr()); 3773 } 3774 #endif 3775 3776 return flock; 3777 } 3778 3779 3780 //------------------------------shared_unlock---------------------------------- 3781 // Emit unlocking code. 3782 void GraphKit::shared_unlock(Node* box, Node* obj) { 3783 // bci is either a monitorenter bc or InvocationEntryBci 3784 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3785 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3786 3787 if( !GenerateSynchronizationCode ) 3788 return; 3789 if (stopped()) { // Dead monitor? 3790 map()->pop_monitor(); // Kill monitor from debug info 3791 return; 3792 } 3793 assert(!obj->is_InlineType(), "should not unlock on inline type"); 3794 3795 // Memory barrier to avoid floating things down past the locked region 3796 insert_mem_bar(Op_MemBarReleaseLock); 3797 3798 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3799 UnlockNode *unlock = new UnlockNode(C, tf); 3800 #ifdef ASSERT 3801 unlock->set_dbg_jvms(sync_jvms()); 3802 #endif 3803 uint raw_idx = Compile::AliasIdxRaw; 3804 unlock->init_req( TypeFunc::Control, control() ); 3805 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3806 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3807 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3808 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3809 3810 unlock->init_req(TypeFunc::Parms + 0, obj); 3811 unlock->init_req(TypeFunc::Parms + 1, box); 3812 unlock = _gvn.transform(unlock)->as_Unlock(); 3813 3814 Node* mem = reset_memory(); 3815 3816 // unlock has no side-effects, sets few values 3817 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3818 3819 // Kill monitor from debug info 3820 map()->pop_monitor( ); 3821 } 3822 3823 //-------------------------------get_layout_helper----------------------------- 3824 // If the given klass is a constant or known to be an array, 3825 // fetch the constant layout helper value into constant_value 3826 // and return (Node*)NULL. Otherwise, load the non-constant 3827 // layout helper value, and return the node which represents it. 3828 // This two-faced routine is useful because allocation sites 3829 // almost always feature constant types. 3830 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3831 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3832 if (!StressReflectiveCode && inst_klass != NULL) { 3833 bool xklass = inst_klass->klass_is_exact(); 3834 bool can_be_flattened = false; 3835 const TypeAryPtr* ary_type = inst_klass->as_instance_type()->isa_aryptr(); 3836 if (UseFlatArray && !xklass && ary_type != NULL && !ary_type->is_null_free()) { 3837 // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue. Don't constant fold. 3838 const TypeOopPtr* elem = ary_type->elem()->make_oopptr(); 3839 can_be_flattened = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flatten_array()); 3840 } 3841 if (!can_be_flattened && (xklass || inst_klass->isa_aryklassptr())) { 3842 jint lhelper; 3843 if (inst_klass->is_flat()) { 3844 lhelper = ary_type->flat_layout_helper(); 3845 } else if (inst_klass->isa_aryklassptr()) { 3846 BasicType elem = ary_type->elem()->array_element_basic_type(); 3847 if (is_reference_type(elem, true)) { 3848 elem = T_OBJECT; 3849 } 3850 lhelper = Klass::array_layout_helper(elem); 3851 } else { 3852 lhelper = inst_klass->is_instklassptr()->exact_klass()->layout_helper(); 3853 } 3854 if (lhelper != Klass::_lh_neutral_value) { 3855 constant_value = lhelper; 3856 return (Node*) NULL; 3857 } 3858 } 3859 } 3860 constant_value = Klass::_lh_neutral_value; // put in a known value 3861 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3862 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3863 } 3864 3865 // We just put in an allocate/initialize with a big raw-memory effect. 3866 // Hook selected additional alias categories on the initialization. 3867 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3868 MergeMemNode* init_in_merge, 3869 Node* init_out_raw) { 3870 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3871 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3872 3873 Node* prevmem = kit.memory(alias_idx); 3874 init_in_merge->set_memory_at(alias_idx, prevmem); 3875 if (init_out_raw != NULL) { 3876 kit.set_memory(init_out_raw, alias_idx); 3877 } 3878 } 3879 3880 //---------------------------set_output_for_allocation------------------------- 3881 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3882 const TypeOopPtr* oop_type, 3883 bool deoptimize_on_exception) { 3884 int rawidx = Compile::AliasIdxRaw; 3885 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3886 add_safepoint_edges(alloc); 3887 Node* allocx = _gvn.transform(alloc); 3888 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3889 // create memory projection for i_o 3890 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3891 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3892 3893 // create a memory projection as for the normal control path 3894 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3895 set_memory(malloc, rawidx); 3896 3897 // a normal slow-call doesn't change i_o, but an allocation does 3898 // we create a separate i_o projection for the normal control path 3899 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3900 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3901 3902 // put in an initialization barrier 3903 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3904 rawoop)->as_Initialize(); 3905 assert(alloc->initialization() == init, "2-way macro link must work"); 3906 assert(init ->allocation() == alloc, "2-way macro link must work"); 3907 { 3908 // Extract memory strands which may participate in the new object's 3909 // initialization, and source them from the new InitializeNode. 3910 // This will allow us to observe initializations when they occur, 3911 // and link them properly (as a group) to the InitializeNode. 3912 assert(init->in(InitializeNode::Memory) == malloc, ""); 3913 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3914 init->set_req(InitializeNode::Memory, minit_in); 3915 record_for_igvn(minit_in); // fold it up later, if possible 3916 _gvn.set_type(minit_in, Type::MEMORY); 3917 Node* minit_out = memory(rawidx); 3918 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3919 // Add an edge in the MergeMem for the header fields so an access 3920 // to one of those has correct memory state 3921 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 3922 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 3923 if (oop_type->isa_aryptr()) { 3924 const TypeAryPtr* arytype = oop_type->is_aryptr(); 3925 if (arytype->is_flat()) { 3926 // Initially all flattened array accesses share a single slice 3927 // but that changes after parsing. Prepare the memory graph so 3928 // it can optimize flattened array accesses properly once they 3929 // don't share a single slice. 3930 assert(C->flattened_accesses_share_alias(), "should be set at parse time"); 3931 C->set_flattened_accesses_share_alias(false); 3932 ciInlineKlass* vk = arytype->elem()->inline_klass(); 3933 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) { 3934 ciField* field = vk->nonstatic_field_at(i); 3935 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3936 continue; // do not bother to track really large numbers of fields 3937 int off_in_vt = field->offset() - vk->first_field_offset(); 3938 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot); 3939 int fieldidx = C->get_alias_index(adr_type, true); 3940 // Pass NULL for init_out. Having per flat array element field memory edges as uses of the Initialize node 3941 // can result in per flat array field Phis to be created which confuses the logic of 3942 // Compile::adjust_flattened_array_access_aliases(). 3943 hook_memory_on_init(*this, fieldidx, minit_in, NULL); 3944 } 3945 C->set_flattened_accesses_share_alias(true); 3946 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out); 3947 } else { 3948 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3949 int elemidx = C->get_alias_index(telemref); 3950 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3951 } 3952 } else if (oop_type->isa_instptr()) { 3953 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word 3954 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass(); 3955 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3956 ciField* field = ik->nonstatic_field_at(i); 3957 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3958 continue; // do not bother to track really large numbers of fields 3959 // Find (or create) the alias category for this field: 3960 int fieldidx = C->alias_type(field)->index(); 3961 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3962 } 3963 } 3964 } 3965 3966 // Cast raw oop to the real thing... 3967 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3968 javaoop = _gvn.transform(javaoop); 3969 C->set_recent_alloc(control(), javaoop); 3970 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3971 3972 #ifdef ASSERT 3973 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3974 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3975 "Ideal_allocation works"); 3976 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3977 "Ideal_allocation works"); 3978 if (alloc->is_AllocateArray()) { 3979 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3980 "Ideal_allocation works"); 3981 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3982 "Ideal_allocation works"); 3983 } else { 3984 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3985 } 3986 } 3987 #endif //ASSERT 3988 3989 return javaoop; 3990 } 3991 3992 //---------------------------new_instance-------------------------------------- 3993 // This routine takes a klass_node which may be constant (for a static type) 3994 // or may be non-constant (for reflective code). It will work equally well 3995 // for either, and the graph will fold nicely if the optimizer later reduces 3996 // the type to a constant. 3997 // The optional arguments are for specialized use by intrinsics: 3998 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3999 // - If 'return_size_val', report the total object size to the caller. 4000 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 4001 Node* GraphKit::new_instance(Node* klass_node, 4002 Node* extra_slow_test, 4003 Node* *return_size_val, 4004 bool deoptimize_on_exception, 4005 InlineTypeNode* inline_type_node) { 4006 // Compute size in doublewords 4007 // The size is always an integral number of doublewords, represented 4008 // as a positive bytewise size stored in the klass's layout_helper. 4009 // The layout_helper also encodes (in a low bit) the need for a slow path. 4010 jint layout_con = Klass::_lh_neutral_value; 4011 Node* layout_val = get_layout_helper(klass_node, layout_con); 4012 bool layout_is_con = (layout_val == NULL); 4013 4014 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 4015 // Generate the initial go-slow test. It's either ALWAYS (return a 4016 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 4017 // case) a computed value derived from the layout_helper. 4018 Node* initial_slow_test = NULL; 4019 if (layout_is_con) { 4020 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4021 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 4022 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 4023 } else { // reflective case 4024 // This reflective path is used by Unsafe.allocateInstance. 4025 // (It may be stress-tested by specifying StressReflectiveCode.) 4026 // Basically, we want to get into the VM is there's an illegal argument. 4027 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 4028 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 4029 if (extra_slow_test != intcon(0)) { 4030 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 4031 } 4032 // (Macro-expander will further convert this to a Bool, if necessary.) 4033 } 4034 4035 // Find the size in bytes. This is easy; it's the layout_helper. 4036 // The size value must be valid even if the slow path is taken. 4037 Node* size = NULL; 4038 if (layout_is_con) { 4039 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 4040 } else { // reflective case 4041 // This reflective path is used by clone and Unsafe.allocateInstance. 4042 size = ConvI2X(layout_val); 4043 4044 // Clear the low bits to extract layout_helper_size_in_bytes: 4045 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 4046 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 4047 size = _gvn.transform( new AndXNode(size, mask) ); 4048 } 4049 if (return_size_val != NULL) { 4050 (*return_size_val) = size; 4051 } 4052 4053 // This is a precise notnull oop of the klass. 4054 // (Actually, it need not be precise if this is a reflective allocation.) 4055 // It's what we cast the result to. 4056 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 4057 if (!tklass) tklass = TypeInstKlassPtr::OBJECT; 4058 const TypeOopPtr* oop_type = tklass->as_instance_type(); 4059 4060 // Now generate allocation code 4061 4062 // The entire memory state is needed for slow path of the allocation 4063 // since GC and deoptimization can happen. 4064 Node *mem = reset_memory(); 4065 set_all_memory(mem); // Create new memory state 4066 4067 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 4068 control(), mem, i_o(), 4069 size, klass_node, 4070 initial_slow_test, inline_type_node); 4071 4072 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 4073 } 4074 4075 //-------------------------------new_array------------------------------------- 4076 // helper for newarray and anewarray 4077 // The 'length' parameter is (obviously) the length of the array. 4078 // See comments on new_instance for the meaning of the other arguments. 4079 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 4080 Node* length, // number of array elements 4081 int nargs, // number of arguments to push back for uncommon trap 4082 Node* *return_size_val, 4083 bool deoptimize_on_exception) { 4084 jint layout_con = Klass::_lh_neutral_value; 4085 Node* layout_val = get_layout_helper(klass_node, layout_con); 4086 bool layout_is_con = (layout_val == NULL); 4087 4088 if (!layout_is_con && !StressReflectiveCode && 4089 !too_many_traps(Deoptimization::Reason_class_check)) { 4090 // This is a reflective array creation site. 4091 // Optimistically assume that it is a subtype of Object[], 4092 // so that we can fold up all the address arithmetic. 4093 layout_con = Klass::array_layout_helper(T_OBJECT); 4094 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 4095 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 4096 { BuildCutout unless(this, bol_lh, PROB_MAX); 4097 inc_sp(nargs); 4098 uncommon_trap(Deoptimization::Reason_class_check, 4099 Deoptimization::Action_maybe_recompile); 4100 } 4101 layout_val = NULL; 4102 layout_is_con = true; 4103 } 4104 4105 // Generate the initial go-slow test. Make sure we do not overflow 4106 // if length is huge (near 2Gig) or negative! We do not need 4107 // exact double-words here, just a close approximation of needed 4108 // double-words. We can't add any offset or rounding bits, lest we 4109 // take a size -1 of bytes and make it positive. Use an unsigned 4110 // compare, so negative sizes look hugely positive. 4111 int fast_size_limit = FastAllocateSizeLimit; 4112 if (layout_is_con) { 4113 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4114 // Increase the size limit if we have exact knowledge of array type. 4115 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 4116 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0); 4117 } 4118 4119 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 4120 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 4121 4122 // --- Size Computation --- 4123 // array_size = round_to_heap(array_header + (length << elem_shift)); 4124 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 4125 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 4126 // The rounding mask is strength-reduced, if possible. 4127 int round_mask = MinObjAlignmentInBytes - 1; 4128 Node* header_size = NULL; 4129 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 4130 // (T_BYTE has the weakest alignment and size restrictions...) 4131 if (layout_is_con) { 4132 int hsize = Klass::layout_helper_header_size(layout_con); 4133 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4134 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con); 4135 if ((round_mask & ~right_n_bits(eshift)) == 0) 4136 round_mask = 0; // strength-reduce it if it goes away completely 4137 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 4138 assert(header_size_min <= hsize, "generic minimum is smallest"); 4139 header_size_min = hsize; 4140 header_size = intcon(hsize + round_mask); 4141 } else { 4142 Node* hss = intcon(Klass::_lh_header_size_shift); 4143 Node* hsm = intcon(Klass::_lh_header_size_mask); 4144 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 4145 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 4146 Node* mask = intcon(round_mask); 4147 header_size = _gvn.transform( new AddINode(hsize, mask) ); 4148 } 4149 4150 Node* elem_shift = NULL; 4151 if (layout_is_con) { 4152 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4153 if (eshift != 0) 4154 elem_shift = intcon(eshift); 4155 } else { 4156 // There is no need to mask or shift this value. 4157 // The semantics of LShiftINode include an implicit mask to 0x1F. 4158 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 4159 elem_shift = layout_val; 4160 } 4161 4162 // Transition to native address size for all offset calculations: 4163 Node* lengthx = ConvI2X(length); 4164 Node* headerx = ConvI2X(header_size); 4165 #ifdef _LP64 4166 { const TypeInt* tilen = _gvn.find_int_type(length); 4167 if (tilen != NULL && tilen->_lo < 0) { 4168 // Add a manual constraint to a positive range. Cf. array_element_address. 4169 jint size_max = fast_size_limit; 4170 if (size_max > tilen->_hi) size_max = tilen->_hi; 4171 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 4172 4173 // Only do a narrow I2L conversion if the range check passed. 4174 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 4175 _gvn.transform(iff); 4176 RegionNode* region = new RegionNode(3); 4177 _gvn.set_type(region, Type::CONTROL); 4178 lengthx = new PhiNode(region, TypeLong::LONG); 4179 _gvn.set_type(lengthx, TypeLong::LONG); 4180 4181 // Range check passed. Use ConvI2L node with narrow type. 4182 Node* passed = IfFalse(iff); 4183 region->init_req(1, passed); 4184 // Make I2L conversion control dependent to prevent it from 4185 // floating above the range check during loop optimizations. 4186 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 4187 4188 // Range check failed. Use ConvI2L with wide type because length may be invalid. 4189 region->init_req(2, IfTrue(iff)); 4190 lengthx->init_req(2, ConvI2X(length)); 4191 4192 set_control(region); 4193 record_for_igvn(region); 4194 record_for_igvn(lengthx); 4195 } 4196 } 4197 #endif 4198 4199 // Combine header size (plus rounding) and body size. Then round down. 4200 // This computation cannot overflow, because it is used only in two 4201 // places, one where the length is sharply limited, and the other 4202 // after a successful allocation. 4203 Node* abody = lengthx; 4204 if (elem_shift != NULL) 4205 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 4206 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 4207 if (round_mask != 0) { 4208 Node* mask = MakeConX(~round_mask); 4209 size = _gvn.transform( new AndXNode(size, mask) ); 4210 } 4211 // else if round_mask == 0, the size computation is self-rounding 4212 4213 if (return_size_val != NULL) { 4214 // This is the size 4215 (*return_size_val) = size; 4216 } 4217 4218 // Now generate allocation code 4219 4220 // The entire memory state is needed for slow path of the allocation 4221 // since GC and deoptimization can happen. 4222 Node *mem = reset_memory(); 4223 set_all_memory(mem); // Create new memory state 4224 4225 if (initial_slow_test->is_Bool()) { 4226 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4227 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4228 } 4229 4230 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr(); 4231 const TypeOopPtr* ary_type = ary_klass->as_instance_type(); 4232 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 4233 4234 // Inline type array variants: 4235 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue") 4236 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue") 4237 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue") 4238 // Check if array is a null-free, non-flattened inline type array 4239 // that needs to be initialized with the default inline type. 4240 Node* default_value = NULL; 4241 Node* raw_default_value = NULL; 4242 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) { 4243 // Array type is known 4244 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) { 4245 ciInlineKlass* vk = ary_ptr->elem()->inline_klass(); 4246 default_value = InlineTypeNode::default_oop(gvn(), vk); 4247 } 4248 } else if (ary_type->can_be_inline_array()) { 4249 // Array type is not known, add runtime checks 4250 assert(!ary_klass->klass_is_exact(), "unexpected exact type"); 4251 Node* r = new RegionNode(3); 4252 default_value = new PhiNode(r, TypeInstPtr::BOTTOM); 4253 4254 Node* bol = array_lh_test(klass_node, Klass::_lh_array_tag_flat_value_bit_inplace | Klass::_lh_null_free_array_bit_inplace, Klass::_lh_null_free_array_bit_inplace); 4255 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4256 4257 // Null-free, non-flattened inline type array, initialize with the default value 4258 set_control(_gvn.transform(new IfTrueNode(iff))); 4259 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset())); 4260 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS)); 4261 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset())); 4262 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4263 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset())); 4264 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered); 4265 Node* elem_mirror = load_mirror_from_klass(eklass); 4266 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset)); 4267 Node* val = access_load_at(elem_mirror, default_value_addr, TypeInstPtr::MIRROR, TypeInstPtr::NOTNULL, T_OBJECT, IN_HEAP); 4268 r->init_req(1, control()); 4269 default_value->init_req(1, val); 4270 4271 // Otherwise initialize with all zero 4272 r->init_req(2, _gvn.transform(new IfFalseNode(iff))); 4273 default_value->init_req(2, null()); 4274 4275 set_control(_gvn.transform(r)); 4276 default_value = _gvn.transform(default_value); 4277 } 4278 if (default_value != NULL) { 4279 if (UseCompressedOops) { 4280 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops 4281 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop())); 4282 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value)); 4283 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32))); 4284 raw_default_value = _gvn.transform(new OrLNode(lower, upper)); 4285 } else { 4286 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value)); 4287 } 4288 } 4289 4290 Node* valid_length_test = _gvn.intcon(1); 4291 if (ary_type->isa_aryptr()) { 4292 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type(); 4293 jint max = TypeAryPtr::max_array_length(bt); 4294 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max))); 4295 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le)); 4296 } 4297 4298 // Create the AllocateArrayNode and its result projections 4299 AllocateArrayNode* alloc 4300 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4301 control(), mem, i_o(), 4302 size, klass_node, 4303 initial_slow_test, 4304 length, valid_length_test, 4305 default_value, raw_default_value); 4306 // Cast to correct type. Note that the klass_node may be constant or not, 4307 // and in the latter case the actual array type will be inexact also. 4308 // (This happens via a non-constant argument to inline_native_newArray.) 4309 // In any case, the value of klass_node provides the desired array type. 4310 const TypeInt* length_type = _gvn.find_int_type(length); 4311 if (ary_type->isa_aryptr() && length_type != NULL) { 4312 // Try to get a better type than POS for the size 4313 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4314 } 4315 4316 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4317 4318 array_ideal_length(alloc, ary_type, true); 4319 return javaoop; 4320 } 4321 4322 // The following "Ideal_foo" functions are placed here because they recognize 4323 // the graph shapes created by the functions immediately above. 4324 4325 //---------------------------Ideal_allocation---------------------------------- 4326 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4327 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 4328 if (ptr == NULL) { // reduce dumb test in callers 4329 return NULL; 4330 } 4331 4332 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4333 ptr = bs->step_over_gc_barrier(ptr); 4334 4335 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4336 ptr = ptr->in(1); 4337 if (ptr == NULL) return NULL; 4338 } 4339 // Return NULL for allocations with several casts: 4340 // j.l.reflect.Array.newInstance(jobject, jint) 4341 // Object.clone() 4342 // to keep more precise type from last cast. 4343 if (ptr->is_Proj()) { 4344 Node* allo = ptr->in(0); 4345 if (allo != NULL && allo->is_Allocate()) { 4346 return allo->as_Allocate(); 4347 } 4348 } 4349 // Report failure to match. 4350 return NULL; 4351 } 4352 4353 // Fancy version which also strips off an offset (and reports it to caller). 4354 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 4355 intptr_t& offset) { 4356 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4357 if (base == NULL) return NULL; 4358 return Ideal_allocation(base, phase); 4359 } 4360 4361 // Trace Initialize <- Proj[Parm] <- Allocate 4362 AllocateNode* InitializeNode::allocation() { 4363 Node* rawoop = in(InitializeNode::RawAddress); 4364 if (rawoop->is_Proj()) { 4365 Node* alloc = rawoop->in(0); 4366 if (alloc->is_Allocate()) { 4367 return alloc->as_Allocate(); 4368 } 4369 } 4370 return NULL; 4371 } 4372 4373 // Trace Allocate -> Proj[Parm] -> Initialize 4374 InitializeNode* AllocateNode::initialization() { 4375 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4376 if (rawoop == NULL) return NULL; 4377 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4378 Node* init = rawoop->fast_out(i); 4379 if (init->is_Initialize()) { 4380 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4381 return init->as_Initialize(); 4382 } 4383 } 4384 return NULL; 4385 } 4386 4387 //----------------------------- loop predicates --------------------------- 4388 4389 //------------------------------add_predicate_impl---------------------------- 4390 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 4391 // Too many traps seen? 4392 if (too_many_traps(reason)) { 4393 #ifdef ASSERT 4394 if (TraceLoopPredicate) { 4395 int tc = C->trap_count(reason); 4396 tty->print("too many traps=%s tcount=%d in ", 4397 Deoptimization::trap_reason_name(reason), tc); 4398 method()->print(); // which method has too many predicate traps 4399 tty->cr(); 4400 } 4401 #endif 4402 // We cannot afford to take more traps here, 4403 // do not generate predicate. 4404 return; 4405 } 4406 4407 Node *cont = _gvn.intcon(1); 4408 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 4409 Node *bol = _gvn.transform(new Conv2BNode(opq)); 4410 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4411 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4412 C->add_predicate_opaq(opq); 4413 { 4414 PreserveJVMState pjvms(this); 4415 set_control(iffalse); 4416 inc_sp(nargs); 4417 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4418 } 4419 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4420 set_control(iftrue); 4421 } 4422 4423 //------------------------------add_predicate--------------------------------- 4424 void GraphKit::add_empty_predicates(int nargs) { 4425 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding 4426 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of 4427 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate. 4428 if (UseLoopPredicate) { 4429 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs); 4430 } 4431 if (UseProfiledLoopPredicate) { 4432 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs); 4433 } 4434 // loop's limit check predicate should be near the loop. 4435 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 4436 } 4437 4438 void GraphKit::sync_kit(IdealKit& ideal) { 4439 set_all_memory(ideal.merged_memory()); 4440 set_i_o(ideal.i_o()); 4441 set_control(ideal.ctrl()); 4442 } 4443 4444 void GraphKit::final_sync(IdealKit& ideal) { 4445 // Final sync IdealKit and graphKit. 4446 sync_kit(ideal); 4447 } 4448 4449 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4450 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4451 Node* coder = load_String_coder(str, set_ctrl); 4452 // Divide length by 2 if coder is UTF16 4453 return _gvn.transform(new RShiftINode(len, coder)); 4454 } 4455 4456 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4457 int value_offset = java_lang_String::value_offset(); 4458 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4459 false, NULL, Type::Offset(0)); 4460 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4461 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4462 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true), 4463 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4464 Node* p = basic_plus_adr(str, str, value_offset); 4465 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4466 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4467 return load; 4468 } 4469 4470 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4471 if (!CompactStrings) { 4472 return intcon(java_lang_String::CODER_UTF16); 4473 } 4474 int coder_offset = java_lang_String::coder_offset(); 4475 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4476 false, NULL, Type::Offset(0)); 4477 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4478 4479 Node* p = basic_plus_adr(str, str, coder_offset); 4480 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4481 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4482 return load; 4483 } 4484 4485 void GraphKit::store_String_value(Node* str, Node* value) { 4486 int value_offset = java_lang_String::value_offset(); 4487 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4488 false, NULL, Type::Offset(0)); 4489 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4490 4491 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4492 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4493 } 4494 4495 void GraphKit::store_String_coder(Node* str, Node* value) { 4496 int coder_offset = java_lang_String::coder_offset(); 4497 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4498 false, NULL, Type::Offset(0)); 4499 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4500 4501 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4502 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4503 } 4504 4505 // Capture src and dst memory state with a MergeMemNode 4506 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4507 if (src_type == dst_type) { 4508 // Types are equal, we don't need a MergeMemNode 4509 return memory(src_type); 4510 } 4511 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4512 record_for_igvn(merge); // fold it up later, if possible 4513 int src_idx = C->get_alias_index(src_type); 4514 int dst_idx = C->get_alias_index(dst_type); 4515 merge->set_memory_at(src_idx, memory(src_idx)); 4516 merge->set_memory_at(dst_idx, memory(dst_idx)); 4517 return merge; 4518 } 4519 4520 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4521 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4522 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4523 // If input and output memory types differ, capture both states to preserve 4524 // the dependency between preceding and subsequent loads/stores. 4525 // For example, the following program: 4526 // StoreB 4527 // compress_string 4528 // LoadB 4529 // has this memory graph (use->def): 4530 // LoadB -> compress_string -> CharMem 4531 // ... -> StoreB -> ByteMem 4532 // The intrinsic hides the dependency between LoadB and StoreB, causing 4533 // the load to read from memory not containing the result of the StoreB. 4534 // The correct memory graph should look like this: 4535 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4536 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4537 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4538 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str))); 4539 set_memory(res_mem, TypeAryPtr::BYTES); 4540 return str; 4541 } 4542 4543 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4544 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4545 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4546 // Capture src and dst memory (see comment in 'compress_string'). 4547 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4548 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4549 set_memory(_gvn.transform(str), dst_type); 4550 } 4551 4552 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4553 /** 4554 * int i_char = start; 4555 * for (int i_byte = 0; i_byte < count; i_byte++) { 4556 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4557 * } 4558 */ 4559 add_empty_predicates(); 4560 C->set_has_loops(true); 4561 4562 RegionNode* head = new RegionNode(3); 4563 head->init_req(1, control()); 4564 gvn().set_type(head, Type::CONTROL); 4565 record_for_igvn(head); 4566 4567 Node* i_byte = new PhiNode(head, TypeInt::INT); 4568 i_byte->init_req(1, intcon(0)); 4569 gvn().set_type(i_byte, TypeInt::INT); 4570 record_for_igvn(i_byte); 4571 4572 Node* i_char = new PhiNode(head, TypeInt::INT); 4573 i_char->init_req(1, start); 4574 gvn().set_type(i_char, TypeInt::INT); 4575 record_for_igvn(i_char); 4576 4577 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4578 gvn().set_type(mem, Type::MEMORY); 4579 record_for_igvn(mem); 4580 set_control(head); 4581 set_memory(mem, TypeAryPtr::BYTES); 4582 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true); 4583 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4584 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4585 false, false, true /* mismatched */); 4586 4587 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4588 head->init_req(2, IfTrue(iff)); 4589 mem->init_req(2, st); 4590 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4591 i_char->init_req(2, AddI(i_char, intcon(2))); 4592 4593 set_control(IfFalse(iff)); 4594 set_memory(st, TypeAryPtr::BYTES); 4595 } 4596 4597 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4598 if (!field->is_constant()) { 4599 return NULL; // Field not marked as constant. 4600 } 4601 ciInstance* holder = NULL; 4602 if (!field->is_static()) { 4603 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4604 if (const_oop != NULL && const_oop->is_instance()) { 4605 holder = const_oop->as_instance(); 4606 } 4607 } 4608 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4609 /*is_unsigned_load=*/false); 4610 if (con_type != NULL) { 4611 Node* con = makecon(con_type); 4612 if (field->type()->is_inlinetype()) { 4613 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free()); 4614 } else if (con_type->is_inlinetypeptr()) { 4615 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free()); 4616 } 4617 return con; 4618 } 4619 return NULL; 4620 } 4621 4622 //---------------------------load_mirror_from_klass---------------------------- 4623 // Given a klass oop, load its java mirror (a java.lang.Class oop). 4624 Node* GraphKit::load_mirror_from_klass(Node* klass) { 4625 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 4626 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4627 // mirror = ((OopHandle)mirror)->resolve(); 4628 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 4629 }