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