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(control(), 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::filter_null(Node* value, bool null2default, int trap_bci) { 1398 Node* null_ctl = top(); 1399 value = null_check_oop(value, &null_ctl); 1400 if (!null_ctl->is_top()) { 1401 if (null2default) { 1402 // Return default value if oop is null 1403 Node* region = new RegionNode(3); 1404 region->init_req(1, control()); 1405 region->init_req(2, null_ctl); 1406 const Type* type = gvn().type(value); 1407 value = PhiNode::make(region, value, type); 1408 value->set_req(2, ValueTypeNode::default_oop(gvn(), type->value_klass())); 1409 set_control(gvn().transform(region)); 1410 value = gvn().transform(value); 1411 } else { 1412 // Deoptimize if oop is null 1413 PreserveJVMState pjvms(this); 1414 set_control(null_ctl); 1415 int cur_bci = bci(); 1416 if (trap_bci != -1) { 1417 // Put trap at different bytecode 1418 push(null()); 1419 set_bci(trap_bci); 1420 } 1421 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none); 1422 if (trap_bci != -1) { 1423 set_bci(cur_bci); 1424 } 1425 } 1426 } 1427 return value; 1428 } 1429 1430 //------------------------------cast_not_null---------------------------------- 1431 // Cast obj to not-null on this path 1432 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1433 if (obj->is_ValueType()) { 1434 return obj; 1435 } 1436 const Type *t = _gvn.type(obj); 1437 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1438 // Object is already not-null? 1439 if( t == t_not_null ) return obj; 1440 1441 Node *cast = new CastPPNode(obj,t_not_null); 1442 cast->init_req(0, control()); 1443 cast = _gvn.transform( cast ); 1444 1445 // Scan for instances of 'obj' in the current JVM mapping. 1446 // These instances are known to be not-null after the test. 1447 if (do_replace_in_map) 1448 replace_in_map(obj, cast); 1449 1450 return cast; // Return casted value 1451 } 1452 1453 // Sometimes in intrinsics, we implicitly know an object is not null 1454 // (there's no actual null check) so we can cast it to not null. In 1455 // the course of optimizations, the input to the cast can become null. 1456 // In that case that data path will die and we need the control path 1457 // to become dead as well to keep the graph consistent. So we have to 1458 // add a check for null for which one branch can't be taken. It uses 1459 // an Opaque4 node that will cause the check to be removed after loop 1460 // opts so the test goes away and the compiled code doesn't execute a 1461 // useless check. 1462 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1463 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1464 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1465 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1466 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1467 _gvn.set_type(iff, iff->Value(&_gvn)); 1468 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1469 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1470 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1471 C->root()->add_req(halt); 1472 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1473 set_control(if_t); 1474 return cast_not_null(value, do_replace_in_map); 1475 } 1476 1477 1478 //--------------------------replace_in_map------------------------------------- 1479 void GraphKit::replace_in_map(Node* old, Node* neww) { 1480 if (old == neww) { 1481 return; 1482 } 1483 1484 map()->replace_edge(old, neww); 1485 1486 // Note: This operation potentially replaces any edge 1487 // on the map. This includes locals, stack, and monitors 1488 // of the current (innermost) JVM state. 1489 1490 // don't let inconsistent types from profiling escape this 1491 // method 1492 1493 const Type* told = _gvn.type(old); 1494 const Type* tnew = _gvn.type(neww); 1495 1496 if (!tnew->higher_equal(told)) { 1497 return; 1498 } 1499 1500 map()->record_replaced_node(old, neww); 1501 } 1502 1503 1504 //============================================================================= 1505 //--------------------------------memory--------------------------------------- 1506 Node* GraphKit::memory(uint alias_idx) { 1507 MergeMemNode* mem = merged_memory(); 1508 Node* p = mem->memory_at(alias_idx); 1509 _gvn.set_type(p, Type::MEMORY); // must be mapped 1510 return p; 1511 } 1512 1513 //-----------------------------reset_memory------------------------------------ 1514 Node* GraphKit::reset_memory() { 1515 Node* mem = map()->memory(); 1516 // do not use this node for any more parsing! 1517 debug_only( map()->set_memory((Node*)NULL) ); 1518 return _gvn.transform( mem ); 1519 } 1520 1521 //------------------------------set_all_memory--------------------------------- 1522 void GraphKit::set_all_memory(Node* newmem) { 1523 Node* mergemem = MergeMemNode::make(newmem); 1524 gvn().set_type_bottom(mergemem); 1525 map()->set_memory(mergemem); 1526 } 1527 1528 //------------------------------set_all_memory_call---------------------------- 1529 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1530 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1531 set_all_memory(newmem); 1532 } 1533 1534 //============================================================================= 1535 // 1536 // parser factory methods for MemNodes 1537 // 1538 // These are layered on top of the factory methods in LoadNode and StoreNode, 1539 // and integrate with the parser's memory state and _gvn engine. 1540 // 1541 1542 // factory methods in "int adr_idx" 1543 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1544 int adr_idx, 1545 MemNode::MemOrd mo, 1546 LoadNode::ControlDependency control_dependency, 1547 bool require_atomic_access, 1548 bool unaligned, 1549 bool mismatched) { 1550 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1551 const TypePtr* adr_type = NULL; // debug-mode-only argument 1552 debug_only(adr_type = C->get_adr_type(adr_idx)); 1553 Node* mem = memory(adr_idx); 1554 Node* ld; 1555 if (require_atomic_access && bt == T_LONG) { 1556 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1557 } else if (require_atomic_access && bt == T_DOUBLE) { 1558 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1559 } else { 1560 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1561 } 1562 ld = _gvn.transform(ld); 1563 1564 if (((bt == T_OBJECT || bt == T_VALUETYPE || bt == T_VALUETYPEPTR) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1565 // Improve graph before escape analysis and boxing elimination. 1566 record_for_igvn(ld); 1567 } 1568 return ld; 1569 } 1570 1571 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1572 int adr_idx, 1573 MemNode::MemOrd mo, 1574 bool require_atomic_access, 1575 bool unaligned, 1576 bool mismatched) { 1577 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1578 const TypePtr* adr_type = NULL; 1579 debug_only(adr_type = C->get_adr_type(adr_idx)); 1580 Node *mem = memory(adr_idx); 1581 Node* st; 1582 if (require_atomic_access && bt == T_LONG) { 1583 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1584 } else if (require_atomic_access && bt == T_DOUBLE) { 1585 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1586 } else { 1587 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1588 } 1589 if (unaligned) { 1590 st->as_Store()->set_unaligned_access(); 1591 } 1592 if (mismatched) { 1593 st->as_Store()->set_mismatched_access(); 1594 } 1595 st = _gvn.transform(st); 1596 set_memory(st, adr_idx); 1597 // Back-to-back stores can only remove intermediate store with DU info 1598 // so push on worklist for optimizer. 1599 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1600 record_for_igvn(st); 1601 1602 return st; 1603 } 1604 1605 Node* GraphKit::access_store_at(Node* ctl, 1606 Node* obj, 1607 Node* adr, 1608 const TypePtr* adr_type, 1609 Node* val, 1610 const Type* val_type, 1611 BasicType bt, 1612 DecoratorSet decorators, 1613 bool deoptimize_on_exception) { 1614 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1615 // could be delayed during Parse (for example, in adjust_map_after_if()). 1616 // Execute transformation here to avoid barrier generation in such case. 1617 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1618 val = _gvn.makecon(TypePtr::NULL_PTR); 1619 } 1620 1621 set_control(ctl); 1622 if (stopped()) { 1623 return top(); // Dead path ? 1624 } 1625 1626 assert(val != NULL, "not dead path"); 1627 if (val->is_ValueType()) { 1628 // Allocate value type and get oop 1629 val = val->as_ValueType()->allocate(this, deoptimize_on_exception)->get_oop(); 1630 } 1631 1632 C2AccessValuePtr addr(adr, adr_type); 1633 C2AccessValue value(val, val_type); 1634 C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1635 if (access.is_raw()) { 1636 return _barrier_set->BarrierSetC2::store_at(access, value); 1637 } else { 1638 return _barrier_set->store_at(access, value); 1639 } 1640 } 1641 1642 Node* GraphKit::access_load_at(Node* obj, // containing obj 1643 Node* adr, // actual adress to store val at 1644 const TypePtr* adr_type, 1645 const Type* val_type, 1646 BasicType bt, 1647 DecoratorSet decorators) { 1648 if (stopped()) { 1649 return top(); // Dead path ? 1650 } 1651 1652 C2AccessValuePtr addr(adr, adr_type); 1653 C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr); 1654 if (access.is_raw()) { 1655 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1656 } else { 1657 return _barrier_set->load_at(access, val_type); 1658 } 1659 } 1660 1661 Node* GraphKit::access_load(Node* adr, // actual adress to load val at 1662 const Type* val_type, 1663 BasicType bt, 1664 DecoratorSet decorators) { 1665 if (stopped()) { 1666 return top(); // Dead path ? 1667 } 1668 1669 C2AccessValuePtr addr(adr, NULL); 1670 C2Access access(this, decorators | C2_READ_ACCESS, bt, NULL, addr); 1671 if (access.is_raw()) { 1672 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1673 } else { 1674 return _barrier_set->load_at(access, val_type); 1675 } 1676 } 1677 1678 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl, 1679 Node* obj, 1680 Node* adr, 1681 const TypePtr* adr_type, 1682 int alias_idx, 1683 Node* expected_val, 1684 Node* new_val, 1685 const Type* value_type, 1686 BasicType bt, 1687 DecoratorSet decorators) { 1688 set_control(ctl); 1689 C2AccessValuePtr addr(adr, adr_type); 1690 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1691 bt, obj, addr, alias_idx); 1692 if (access.is_raw()) { 1693 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1694 } else { 1695 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1696 } 1697 } 1698 1699 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl, 1700 Node* obj, 1701 Node* adr, 1702 const TypePtr* adr_type, 1703 int alias_idx, 1704 Node* expected_val, 1705 Node* new_val, 1706 const Type* value_type, 1707 BasicType bt, 1708 DecoratorSet decorators) { 1709 set_control(ctl); 1710 C2AccessValuePtr addr(adr, adr_type); 1711 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1712 bt, obj, addr, alias_idx); 1713 if (access.is_raw()) { 1714 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1715 } else { 1716 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1717 } 1718 } 1719 1720 Node* GraphKit::access_atomic_xchg_at(Node* ctl, 1721 Node* obj, 1722 Node* adr, 1723 const TypePtr* adr_type, 1724 int alias_idx, 1725 Node* new_val, 1726 const Type* value_type, 1727 BasicType bt, 1728 DecoratorSet decorators) { 1729 set_control(ctl); 1730 C2AccessValuePtr addr(adr, adr_type); 1731 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1732 bt, obj, addr, alias_idx); 1733 if (access.is_raw()) { 1734 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1735 } else { 1736 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1737 } 1738 } 1739 1740 Node* GraphKit::access_atomic_add_at(Node* ctl, 1741 Node* obj, 1742 Node* adr, 1743 const TypePtr* adr_type, 1744 int alias_idx, 1745 Node* new_val, 1746 const Type* value_type, 1747 BasicType bt, 1748 DecoratorSet decorators) { 1749 set_control(ctl); 1750 C2AccessValuePtr addr(adr, adr_type); 1751 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1752 if (access.is_raw()) { 1753 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1754 } else { 1755 return _barrier_set->atomic_add_at(access, new_val, value_type); 1756 } 1757 } 1758 1759 void GraphKit::access_clone(Node* ctl, Node* src_base, Node* dst_base, Node* countx, bool is_array) { 1760 set_control(ctl); 1761 return _barrier_set->clone(this, src_base, dst_base, countx, is_array); 1762 } 1763 1764 //-------------------------array_element_address------------------------- 1765 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1766 const TypeInt* sizetype, Node* ctrl) { 1767 uint shift = exact_log2(type2aelembytes(elembt)); 1768 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1769 if (arytype_klass->is_value_array_klass()) { 1770 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass(); 1771 shift = vak->log2_element_size(); 1772 } 1773 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1774 1775 // short-circuit a common case (saves lots of confusing waste motion) 1776 jint idx_con = find_int_con(idx, -1); 1777 if (idx_con >= 0) { 1778 intptr_t offset = header + ((intptr_t)idx_con << shift); 1779 return basic_plus_adr(ary, offset); 1780 } 1781 1782 // must be correct type for alignment purposes 1783 Node* base = basic_plus_adr(ary, header); 1784 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1785 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1786 return basic_plus_adr(ary, base, scale); 1787 } 1788 1789 //-------------------------load_array_element------------------------- 1790 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1791 const Type* elemtype = arytype->elem(); 1792 BasicType elembt = elemtype->array_element_basic_type(); 1793 assert(elembt != T_VALUETYPE, "value types are not supported by this method"); 1794 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1795 if (elembt == T_NARROWOOP) { 1796 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1797 } 1798 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1799 return ld; 1800 } 1801 1802 //-------------------------set_arguments_for_java_call------------------------- 1803 // Arguments (pre-popped from the stack) are taken from the JVMS. 1804 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1805 // Add the call arguments: 1806 const TypeTuple* domain = call->tf()->domain_sig(); 1807 uint nargs = domain->cnt(); 1808 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1809 Node* arg = argument(i-TypeFunc::Parms); 1810 const Type* t = domain->field_at(i); 1811 if (arg->is_ValueType()) { 1812 assert(t->is_oopptr()->can_be_value_type(), "wrong argument type"); 1813 ValueTypeNode* vt = arg->as_ValueType(); 1814 if (ValueTypePassFieldsAsArgs) { 1815 // We don't pass value type arguments by reference but instead 1816 // pass each field of the value type 1817 idx += vt->pass_fields(call, idx, *this); 1818 // If a value type argument is passed as fields, attach the Method* to the call site 1819 // to be able to access the extended signature later via attached_method_before_pc(). 1820 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1821 call->set_override_symbolic_info(true); 1822 continue; 1823 } else { 1824 // Pass value type argument via oop to callee 1825 arg = vt->allocate(this)->get_oop(); 1826 } 1827 } else if (t->is_valuetypeptr() && !arg->is_ValueTypePtr() && gvn().type(arg)->maybe_null()) { 1828 // Constant null passed for a value type argument 1829 assert(arg->bottom_type()->remove_speculative() == TypePtr::NULL_PTR, "Anything other than null?"); 1830 ciMethod* declared_method = method()->get_method_at_bci(bci()); 1831 int arg_size = declared_method->signature()->arg_size_for_bc(java_bc()); 1832 inc_sp(arg_size); // restore arguments 1833 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none); 1834 return; 1835 } 1836 call->init_req(idx, arg); 1837 idx++; 1838 } 1839 } 1840 1841 //---------------------------set_edges_for_java_call--------------------------- 1842 // Connect a newly created call into the current JVMS. 1843 // A return value node (if any) is returned from set_edges_for_java_call. 1844 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1845 1846 // Add the predefined inputs: 1847 call->init_req( TypeFunc::Control, control() ); 1848 call->init_req( TypeFunc::I_O , i_o() ); 1849 call->init_req( TypeFunc::Memory , reset_memory() ); 1850 call->init_req( TypeFunc::FramePtr, frameptr() ); 1851 call->init_req( TypeFunc::ReturnAdr, top() ); 1852 1853 add_safepoint_edges(call, must_throw); 1854 1855 Node* xcall = _gvn.transform(call); 1856 1857 if (xcall == top()) { 1858 set_control(top()); 1859 return; 1860 } 1861 assert(xcall == call, "call identity is stable"); 1862 1863 // Re-use the current map to produce the result. 1864 1865 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1866 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1867 set_all_memory_call(xcall, separate_io_proj); 1868 1869 //return xcall; // no need, caller already has it 1870 } 1871 1872 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1873 if (stopped()) return top(); // maybe the call folded up? 1874 1875 // Note: Since any out-of-line call can produce an exception, 1876 // we always insert an I_O projection from the call into the result. 1877 1878 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1879 1880 if (separate_io_proj) { 1881 // The caller requested separate projections be used by the fall 1882 // through and exceptional paths, so replace the projections for 1883 // the fall through path. 1884 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1885 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1886 } 1887 1888 // Capture the return value, if any. 1889 Node* ret; 1890 if (call->method() == NULL || 1891 call->method()->return_type()->basic_type() == T_VOID) { 1892 ret = top(); 1893 } else { 1894 if (!call->tf()->returns_value_type_as_fields()) { 1895 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1896 } else { 1897 // Return of multiple values (value type fields): we create a 1898 // ValueType node, each field is a projection from the call. 1899 const TypeTuple* range_sig = call->tf()->range_sig(); 1900 const Type* t = range_sig->field_at(TypeFunc::Parms); 1901 assert(t->is_valuetypeptr(), "only value types for multiple return values"); 1902 ciValueKlass* vk = t->value_klass(); 1903 ret = ValueTypeNode::make_from_multi(this, call, vk, TypeFunc::Parms+1, false); 1904 } 1905 } 1906 1907 return ret; 1908 } 1909 1910 //--------------------set_predefined_input_for_runtime_call-------------------- 1911 // Reading and setting the memory state is way conservative here. 1912 // The real problem is that I am not doing real Type analysis on memory, 1913 // so I cannot distinguish card mark stores from other stores. Across a GC 1914 // point the Store Barrier and the card mark memory has to agree. I cannot 1915 // have a card mark store and its barrier split across the GC point from 1916 // either above or below. Here I get that to happen by reading ALL of memory. 1917 // A better answer would be to separate out card marks from other memory. 1918 // For now, return the input memory state, so that it can be reused 1919 // after the call, if this call has restricted memory effects. 1920 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1921 // Set fixed predefined input arguments 1922 Node* memory = reset_memory(); 1923 call->init_req( TypeFunc::Control, control() ); 1924 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1925 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1926 call->init_req( TypeFunc::FramePtr, frameptr() ); 1927 call->init_req( TypeFunc::ReturnAdr, top() ); 1928 return memory; 1929 } 1930 1931 //-------------------set_predefined_output_for_runtime_call-------------------- 1932 // Set control and memory (not i_o) from the call. 1933 // If keep_mem is not NULL, use it for the output state, 1934 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1935 // If hook_mem is NULL, this call produces no memory effects at all. 1936 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1937 // then only that memory slice is taken from the call. 1938 // In the last case, we must put an appropriate memory barrier before 1939 // the call, so as to create the correct anti-dependencies on loads 1940 // preceding the call. 1941 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1942 Node* keep_mem, 1943 const TypePtr* hook_mem) { 1944 // no i/o 1945 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1946 if (keep_mem) { 1947 // First clone the existing memory state 1948 set_all_memory(keep_mem); 1949 if (hook_mem != NULL) { 1950 // Make memory for the call 1951 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1952 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1953 // We also use hook_mem to extract specific effects from arraycopy stubs. 1954 set_memory(mem, hook_mem); 1955 } 1956 // ...else the call has NO memory effects. 1957 1958 // Make sure the call advertises its memory effects precisely. 1959 // This lets us build accurate anti-dependences in gcm.cpp. 1960 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1961 "call node must be constructed correctly"); 1962 } else { 1963 assert(hook_mem == NULL, ""); 1964 // This is not a "slow path" call; all memory comes from the call. 1965 set_all_memory_call(call); 1966 } 1967 } 1968 1969 1970 // Replace the call with the current state of the kit. 1971 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1972 JVMState* ejvms = NULL; 1973 if (has_exceptions()) { 1974 ejvms = transfer_exceptions_into_jvms(); 1975 } 1976 1977 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1978 ReplacedNodes replaced_nodes_exception; 1979 Node* ex_ctl = top(); 1980 1981 SafePointNode* final_state = stop(); 1982 1983 // Find all the needed outputs of this call 1984 CallProjections* callprojs = call->extract_projections(true); 1985 1986 Node* init_mem = call->in(TypeFunc::Memory); 1987 Node* final_mem = final_state->in(TypeFunc::Memory); 1988 Node* final_ctl = final_state->in(TypeFunc::Control); 1989 Node* final_io = final_state->in(TypeFunc::I_O); 1990 1991 // Replace all the old call edges with the edges from the inlining result 1992 if (callprojs->fallthrough_catchproj != NULL) { 1993 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl); 1994 } 1995 if (callprojs->fallthrough_memproj != NULL) { 1996 if (final_mem->is_MergeMem()) { 1997 // Parser's exits MergeMem was not transformed but may be optimized 1998 final_mem = _gvn.transform(final_mem); 1999 } 2000 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem); 2001 } 2002 if (callprojs->fallthrough_ioproj != NULL) { 2003 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io); 2004 } 2005 2006 // Replace the result with the new result if it exists and is used 2007 if (callprojs->resproj[0] != NULL && result != NULL) { 2008 assert(callprojs->nb_resproj == 1, "unexpected number of results"); 2009 C->gvn_replace_by(callprojs->resproj[0], result); 2010 } 2011 2012 if (ejvms == NULL) { 2013 // No exception edges to simply kill off those paths 2014 if (callprojs->catchall_catchproj != NULL) { 2015 C->gvn_replace_by(callprojs->catchall_catchproj, C->top()); 2016 } 2017 if (callprojs->catchall_memproj != NULL) { 2018 C->gvn_replace_by(callprojs->catchall_memproj, C->top()); 2019 } 2020 if (callprojs->catchall_ioproj != NULL) { 2021 C->gvn_replace_by(callprojs->catchall_ioproj, C->top()); 2022 } 2023 // Replace the old exception object with top 2024 if (callprojs->exobj != NULL) { 2025 C->gvn_replace_by(callprojs->exobj, C->top()); 2026 } 2027 } else { 2028 GraphKit ekit(ejvms); 2029 2030 // Load my combined exception state into the kit, with all phis transformed: 2031 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 2032 replaced_nodes_exception = ex_map->replaced_nodes(); 2033 2034 Node* ex_oop = ekit.use_exception_state(ex_map); 2035 2036 if (callprojs->catchall_catchproj != NULL) { 2037 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control()); 2038 ex_ctl = ekit.control(); 2039 } 2040 if (callprojs->catchall_memproj != NULL) { 2041 C->gvn_replace_by(callprojs->catchall_memproj, ekit.reset_memory()); 2042 } 2043 if (callprojs->catchall_ioproj != NULL) { 2044 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o()); 2045 } 2046 2047 // Replace the old exception object with the newly created one 2048 if (callprojs->exobj != NULL) { 2049 C->gvn_replace_by(callprojs->exobj, ex_oop); 2050 } 2051 } 2052 2053 // Disconnect the call from the graph 2054 call->disconnect_inputs(NULL, C); 2055 C->gvn_replace_by(call, C->top()); 2056 2057 // Clean up any MergeMems that feed other MergeMems since the 2058 // optimizer doesn't like that. 2059 if (final_mem->is_MergeMem()) { 2060 Node_List wl; 2061 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 2062 Node* m = i.get(); 2063 if (m->is_MergeMem() && !wl.contains(m)) { 2064 wl.push(m); 2065 } 2066 } 2067 while (wl.size() > 0) { 2068 _gvn.transform(wl.pop()); 2069 } 2070 } 2071 2072 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 2073 replaced_nodes.apply(C, final_ctl); 2074 } 2075 if (!ex_ctl->is_top() && do_replaced_nodes) { 2076 replaced_nodes_exception.apply(C, ex_ctl); 2077 } 2078 } 2079 2080 2081 //------------------------------increment_counter------------------------------ 2082 // for statistics: increment a VM counter by 1 2083 2084 void GraphKit::increment_counter(address counter_addr) { 2085 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2086 increment_counter(adr1); 2087 } 2088 2089 void GraphKit::increment_counter(Node* counter_addr) { 2090 int adr_type = Compile::AliasIdxRaw; 2091 Node* ctrl = control(); 2092 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 2093 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 2094 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 2095 } 2096 2097 2098 //------------------------------uncommon_trap---------------------------------- 2099 // Bail out to the interpreter in mid-method. Implemented by calling the 2100 // uncommon_trap blob. This helper function inserts a runtime call with the 2101 // right debug info. 2102 void GraphKit::uncommon_trap(int trap_request, 2103 ciKlass* klass, const char* comment, 2104 bool must_throw, 2105 bool keep_exact_action) { 2106 if (failing()) stop(); 2107 if (stopped()) return; // trap reachable? 2108 2109 // Note: If ProfileTraps is true, and if a deopt. actually 2110 // occurs here, the runtime will make sure an MDO exists. There is 2111 // no need to call method()->ensure_method_data() at this point. 2112 2113 // Set the stack pointer to the right value for reexecution: 2114 set_sp(reexecute_sp()); 2115 2116 #ifdef ASSERT 2117 if (!must_throw) { 2118 // Make sure the stack has at least enough depth to execute 2119 // the current bytecode. 2120 int inputs, ignored_depth; 2121 if (compute_stack_effects(inputs, ignored_depth)) { 2122 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2123 Bytecodes::name(java_bc()), sp(), inputs); 2124 } 2125 } 2126 #endif 2127 2128 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2129 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2130 2131 switch (action) { 2132 case Deoptimization::Action_maybe_recompile: 2133 case Deoptimization::Action_reinterpret: 2134 // Temporary fix for 6529811 to allow virtual calls to be sure they 2135 // get the chance to go from mono->bi->mega 2136 if (!keep_exact_action && 2137 Deoptimization::trap_request_index(trap_request) < 0 && 2138 too_many_recompiles(reason)) { 2139 // This BCI is causing too many recompilations. 2140 if (C->log() != NULL) { 2141 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2142 Deoptimization::trap_reason_name(reason), 2143 Deoptimization::trap_action_name(action)); 2144 } 2145 action = Deoptimization::Action_none; 2146 trap_request = Deoptimization::make_trap_request(reason, action); 2147 } else { 2148 C->set_trap_can_recompile(true); 2149 } 2150 break; 2151 case Deoptimization::Action_make_not_entrant: 2152 C->set_trap_can_recompile(true); 2153 break; 2154 case Deoptimization::Action_none: 2155 case Deoptimization::Action_make_not_compilable: 2156 break; 2157 default: 2158 #ifdef ASSERT 2159 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2160 #endif 2161 break; 2162 } 2163 2164 if (TraceOptoParse) { 2165 char buf[100]; 2166 tty->print_cr("Uncommon trap %s at bci:%d", 2167 Deoptimization::format_trap_request(buf, sizeof(buf), 2168 trap_request), bci()); 2169 } 2170 2171 CompileLog* log = C->log(); 2172 if (log != NULL) { 2173 int kid = (klass == NULL)? -1: log->identify(klass); 2174 log->begin_elem("uncommon_trap bci='%d'", bci()); 2175 char buf[100]; 2176 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2177 trap_request)); 2178 if (kid >= 0) log->print(" klass='%d'", kid); 2179 if (comment != NULL) log->print(" comment='%s'", comment); 2180 log->end_elem(); 2181 } 2182 2183 // Make sure any guarding test views this path as very unlikely 2184 Node *i0 = control()->in(0); 2185 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2186 IfNode *iff = i0->as_If(); 2187 float f = iff->_prob; // Get prob 2188 if (control()->Opcode() == Op_IfTrue) { 2189 if (f > PROB_UNLIKELY_MAG(4)) 2190 iff->_prob = PROB_MIN; 2191 } else { 2192 if (f < PROB_LIKELY_MAG(4)) 2193 iff->_prob = PROB_MAX; 2194 } 2195 } 2196 2197 // Clear out dead values from the debug info. 2198 kill_dead_locals(); 2199 2200 // Now insert the uncommon trap subroutine call 2201 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2202 const TypePtr* no_memory_effects = NULL; 2203 // Pass the index of the class to be loaded 2204 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2205 (must_throw ? RC_MUST_THROW : 0), 2206 OptoRuntime::uncommon_trap_Type(), 2207 call_addr, "uncommon_trap", no_memory_effects, 2208 intcon(trap_request)); 2209 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2210 "must extract request correctly from the graph"); 2211 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2212 2213 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2214 // The debug info is the only real input to this call. 2215 2216 // Halt-and-catch fire here. The above call should never return! 2217 HaltNode* halt = new HaltNode(control(), frameptr()); 2218 _gvn.set_type_bottom(halt); 2219 root()->add_req(halt); 2220 2221 stop_and_kill_map(); 2222 } 2223 2224 2225 //--------------------------just_allocated_object------------------------------ 2226 // Report the object that was just allocated. 2227 // It must be the case that there are no intervening safepoints. 2228 // We use this to determine if an object is so "fresh" that 2229 // it does not require card marks. 2230 Node* GraphKit::just_allocated_object(Node* current_control) { 2231 if (C->recent_alloc_ctl() == current_control) 2232 return C->recent_alloc_obj(); 2233 return NULL; 2234 } 2235 2236 2237 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2238 // (Note: TypeFunc::make has a cache that makes this fast.) 2239 const TypeFunc* tf = TypeFunc::make(dest_method); 2240 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2241 for (int j = 0; j < nargs; j++) { 2242 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2243 if( targ->basic_type() == T_DOUBLE ) { 2244 // If any parameters are doubles, they must be rounded before 2245 // the call, dstore_rounding does gvn.transform 2246 Node *arg = argument(j); 2247 arg = dstore_rounding(arg); 2248 set_argument(j, arg); 2249 } 2250 } 2251 } 2252 2253 /** 2254 * Record profiling data exact_kls for Node n with the type system so 2255 * that it can propagate it (speculation) 2256 * 2257 * @param n node that the type applies to 2258 * @param exact_kls type from profiling 2259 * @param maybe_null did profiling see null? 2260 * 2261 * @return node with improved type 2262 */ 2263 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2264 const Type* current_type = _gvn.type(n); 2265 assert(UseTypeSpeculation, "type speculation must be on"); 2266 2267 const TypePtr* speculative = current_type->speculative(); 2268 2269 // Should the klass from the profile be recorded in the speculative type? 2270 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2271 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2272 const TypeOopPtr* xtype = tklass->as_instance_type(); 2273 assert(xtype->klass_is_exact(), "Should be exact"); 2274 // Any reason to believe n is not null (from this profiling or a previous one)? 2275 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2276 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2277 // record the new speculative type's depth 2278 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2279 speculative = speculative->with_inline_depth(jvms()->depth()); 2280 } else if (current_type->would_improve_ptr(ptr_kind)) { 2281 // Profiling report that null was never seen so we can change the 2282 // speculative type to non null ptr. 2283 if (ptr_kind == ProfileAlwaysNull) { 2284 speculative = TypePtr::NULL_PTR; 2285 } else { 2286 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2287 const TypePtr* ptr = TypePtr::NOTNULL; 2288 if (speculative != NULL) { 2289 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2290 } else { 2291 speculative = ptr; 2292 } 2293 } 2294 } 2295 2296 if (speculative != current_type->speculative()) { 2297 // Build a type with a speculative type (what we think we know 2298 // about the type but will need a guard when we use it) 2299 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2300 // We're changing the type, we need a new CheckCast node to carry 2301 // the new type. The new type depends on the control: what 2302 // profiling tells us is only valid from here as far as we can 2303 // tell. 2304 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2305 cast = _gvn.transform(cast); 2306 replace_in_map(n, cast); 2307 n = cast; 2308 } 2309 2310 return n; 2311 } 2312 2313 /** 2314 * Record profiling data from receiver profiling at an invoke with the 2315 * type system so that it can propagate it (speculation) 2316 * 2317 * @param n receiver node 2318 * 2319 * @return node with improved type 2320 */ 2321 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2322 if (!UseTypeSpeculation) { 2323 return n; 2324 } 2325 ciKlass* exact_kls = profile_has_unique_klass(); 2326 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2327 if ((java_bc() == Bytecodes::_checkcast || 2328 java_bc() == Bytecodes::_instanceof || 2329 java_bc() == Bytecodes::_aastore) && 2330 method()->method_data()->is_mature()) { 2331 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2332 if (data != NULL) { 2333 if (!data->as_BitData()->null_seen()) { 2334 ptr_kind = ProfileNeverNull; 2335 } else { 2336 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2337 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2338 uint i = 0; 2339 for (; i < call->row_limit(); i++) { 2340 ciKlass* receiver = call->receiver(i); 2341 if (receiver != NULL) { 2342 break; 2343 } 2344 } 2345 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2346 } 2347 } 2348 } 2349 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2350 } 2351 2352 /** 2353 * Record profiling data from argument profiling at an invoke with the 2354 * type system so that it can propagate it (speculation) 2355 * 2356 * @param dest_method target method for the call 2357 * @param bc what invoke bytecode is this? 2358 */ 2359 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2360 if (!UseTypeSpeculation) { 2361 return; 2362 } 2363 const TypeFunc* tf = TypeFunc::make(dest_method); 2364 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2365 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2366 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2367 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2368 if (targ->isa_oopptr()) { 2369 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2370 ciKlass* better_type = NULL; 2371 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2372 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2373 } 2374 i++; 2375 } 2376 } 2377 } 2378 2379 /** 2380 * Record profiling data from parameter profiling at an invoke with 2381 * the type system so that it can propagate it (speculation) 2382 */ 2383 void GraphKit::record_profiled_parameters_for_speculation() { 2384 if (!UseTypeSpeculation) { 2385 return; 2386 } 2387 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2388 if (_gvn.type(local(i))->isa_oopptr()) { 2389 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2390 ciKlass* better_type = NULL; 2391 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2392 record_profile_for_speculation(local(i), better_type, ptr_kind); 2393 } 2394 j++; 2395 } 2396 } 2397 } 2398 2399 /** 2400 * Record profiling data from return value profiling at an invoke with 2401 * the type system so that it can propagate it (speculation) 2402 */ 2403 void GraphKit::record_profiled_return_for_speculation() { 2404 if (!UseTypeSpeculation) { 2405 return; 2406 } 2407 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2408 ciKlass* better_type = NULL; 2409 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2410 // If profiling reports a single type for the return value, 2411 // feed it to the type system so it can propagate it as a 2412 // speculative type 2413 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2414 } 2415 } 2416 2417 void GraphKit::round_double_result(ciMethod* dest_method) { 2418 // A non-strict method may return a double value which has an extended 2419 // exponent, but this must not be visible in a caller which is 'strict' 2420 // If a strict caller invokes a non-strict callee, round a double result 2421 2422 BasicType result_type = dest_method->return_type()->basic_type(); 2423 assert( method() != NULL, "must have caller context"); 2424 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2425 // Destination method's return value is on top of stack 2426 // dstore_rounding() does gvn.transform 2427 Node *result = pop_pair(); 2428 result = dstore_rounding(result); 2429 push_pair(result); 2430 } 2431 } 2432 2433 // rounding for strict float precision conformance 2434 Node* GraphKit::precision_rounding(Node* n) { 2435 return UseStrictFP && _method->flags().is_strict() 2436 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2437 ? _gvn.transform( new RoundFloatNode(0, n) ) 2438 : n; 2439 } 2440 2441 // rounding for strict double precision conformance 2442 Node* GraphKit::dprecision_rounding(Node *n) { 2443 return UseStrictFP && _method->flags().is_strict() 2444 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2445 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2446 : n; 2447 } 2448 2449 // rounding for non-strict double stores 2450 Node* GraphKit::dstore_rounding(Node* n) { 2451 return Matcher::strict_fp_requires_explicit_rounding 2452 && UseSSE <= 1 2453 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2454 : n; 2455 } 2456 2457 //============================================================================= 2458 // Generate a fast path/slow path idiom. Graph looks like: 2459 // [foo] indicates that 'foo' is a parameter 2460 // 2461 // [in] NULL 2462 // \ / 2463 // CmpP 2464 // Bool ne 2465 // If 2466 // / \ 2467 // True False-<2> 2468 // / | 2469 // / cast_not_null 2470 // Load | | ^ 2471 // [fast_test] | | 2472 // gvn to opt_test | | 2473 // / \ | <1> 2474 // True False | 2475 // | \\ | 2476 // [slow_call] \[fast_result] 2477 // Ctl Val \ \ 2478 // | \ \ 2479 // Catch <1> \ \ 2480 // / \ ^ \ \ 2481 // Ex No_Ex | \ \ 2482 // | \ \ | \ <2> \ 2483 // ... \ [slow_res] | | \ [null_result] 2484 // \ \--+--+--- | | 2485 // \ | / \ | / 2486 // --------Region Phi 2487 // 2488 //============================================================================= 2489 // Code is structured as a series of driver functions all called 'do_XXX' that 2490 // call a set of helper functions. Helper functions first, then drivers. 2491 2492 //------------------------------null_check_oop--------------------------------- 2493 // Null check oop. Set null-path control into Region in slot 3. 2494 // Make a cast-not-nullness use the other not-null control. Return cast. 2495 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2496 bool never_see_null, 2497 bool safe_for_replace, 2498 bool speculative) { 2499 // Initial NULL check taken path 2500 (*null_control) = top(); 2501 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2502 2503 // Generate uncommon_trap: 2504 if (never_see_null && (*null_control) != top()) { 2505 // If we see an unexpected null at a check-cast we record it and force a 2506 // recompile; the offending check-cast will be compiled to handle NULLs. 2507 // If we see more than one offending BCI, then all checkcasts in the 2508 // method will be compiled to handle NULLs. 2509 PreserveJVMState pjvms(this); 2510 set_control(*null_control); 2511 replace_in_map(value, null()); 2512 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2513 uncommon_trap(reason, 2514 Deoptimization::Action_make_not_entrant); 2515 (*null_control) = top(); // NULL path is dead 2516 } 2517 if ((*null_control) == top() && safe_for_replace) { 2518 replace_in_map(value, cast); 2519 } 2520 2521 // Cast away null-ness on the result 2522 return cast; 2523 } 2524 2525 //------------------------------opt_iff---------------------------------------- 2526 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2527 // Return slow-path control. 2528 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2529 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2530 2531 // Fast path taken; set region slot 2 2532 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2533 region->init_req(2,fast_taken); // Capture fast-control 2534 2535 // Fast path not-taken, i.e. slow path 2536 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2537 return slow_taken; 2538 } 2539 2540 //-----------------------------make_runtime_call------------------------------- 2541 Node* GraphKit::make_runtime_call(int flags, 2542 const TypeFunc* call_type, address call_addr, 2543 const char* call_name, 2544 const TypePtr* adr_type, 2545 // The following parms are all optional. 2546 // The first NULL ends the list. 2547 Node* parm0, Node* parm1, 2548 Node* parm2, Node* parm3, 2549 Node* parm4, Node* parm5, 2550 Node* parm6, Node* parm7) { 2551 // Slow-path call 2552 bool is_leaf = !(flags & RC_NO_LEAF); 2553 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2554 if (call_name == NULL) { 2555 assert(!is_leaf, "must supply name for leaf"); 2556 call_name = OptoRuntime::stub_name(call_addr); 2557 } 2558 CallNode* call; 2559 if (!is_leaf) { 2560 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2561 bci(), adr_type); 2562 } else if (flags & RC_NO_FP) { 2563 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2564 } else { 2565 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2566 } 2567 2568 // The following is similar to set_edges_for_java_call, 2569 // except that the memory effects of the call are restricted to AliasIdxRaw. 2570 2571 // Slow path call has no side-effects, uses few values 2572 bool wide_in = !(flags & RC_NARROW_MEM); 2573 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2574 2575 Node* prev_mem = NULL; 2576 if (wide_in) { 2577 prev_mem = set_predefined_input_for_runtime_call(call); 2578 } else { 2579 assert(!wide_out, "narrow in => narrow out"); 2580 Node* narrow_mem = memory(adr_type); 2581 prev_mem = reset_memory(); 2582 map()->set_memory(narrow_mem); 2583 set_predefined_input_for_runtime_call(call); 2584 } 2585 2586 // Hook each parm in order. Stop looking at the first NULL. 2587 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2588 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2589 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2590 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2591 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2592 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2593 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2594 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2595 /* close each nested if ===> */ } } } } } } } } 2596 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2597 2598 if (!is_leaf) { 2599 // Non-leaves can block and take safepoints: 2600 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2601 } 2602 // Non-leaves can throw exceptions: 2603 if (has_io) { 2604 call->set_req(TypeFunc::I_O, i_o()); 2605 } 2606 2607 if (flags & RC_UNCOMMON) { 2608 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2609 // (An "if" probability corresponds roughly to an unconditional count. 2610 // Sort of.) 2611 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2612 } 2613 2614 Node* c = _gvn.transform(call); 2615 assert(c == call, "cannot disappear"); 2616 2617 if (wide_out) { 2618 // Slow path call has full side-effects. 2619 set_predefined_output_for_runtime_call(call); 2620 } else { 2621 // Slow path call has few side-effects, and/or sets few values. 2622 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2623 } 2624 2625 if (has_io) { 2626 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2627 } 2628 return call; 2629 2630 } 2631 2632 //------------------------------merge_memory----------------------------------- 2633 // Merge memory from one path into the current memory state. 2634 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2635 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2636 Node* old_slice = mms.force_memory(); 2637 Node* new_slice = mms.memory2(); 2638 if (old_slice != new_slice) { 2639 PhiNode* phi; 2640 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2641 if (mms.is_empty()) { 2642 // clone base memory Phi's inputs for this memory slice 2643 assert(old_slice == mms.base_memory(), "sanity"); 2644 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2645 _gvn.set_type(phi, Type::MEMORY); 2646 for (uint i = 1; i < phi->req(); i++) { 2647 phi->init_req(i, old_slice->in(i)); 2648 } 2649 } else { 2650 phi = old_slice->as_Phi(); // Phi was generated already 2651 } 2652 } else { 2653 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2654 _gvn.set_type(phi, Type::MEMORY); 2655 } 2656 phi->set_req(new_path, new_slice); 2657 mms.set_memory(phi); 2658 } 2659 } 2660 } 2661 2662 //------------------------------make_slow_call_ex------------------------------ 2663 // Make the exception handler hookups for the slow call 2664 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2665 if (stopped()) return; 2666 2667 // Make a catch node with just two handlers: fall-through and catch-all 2668 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2669 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2670 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2671 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2672 2673 { PreserveJVMState pjvms(this); 2674 set_control(excp); 2675 set_i_o(i_o); 2676 2677 if (excp != top()) { 2678 if (deoptimize) { 2679 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2680 uncommon_trap(Deoptimization::Reason_unhandled, 2681 Deoptimization::Action_none); 2682 } else { 2683 // Create an exception state also. 2684 // Use an exact type if the caller has a specific exception. 2685 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2686 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2687 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2688 } 2689 } 2690 } 2691 2692 // Get the no-exception control from the CatchNode. 2693 set_control(norm); 2694 } 2695 2696 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2697 Node* cmp = NULL; 2698 switch(bt) { 2699 case T_INT: cmp = new CmpINode(in1, in2); break; 2700 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2701 default: fatal("unexpected comparison type %s", type2name(bt)); 2702 } 2703 gvn->transform(cmp); 2704 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2705 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2706 gvn->transform(iff); 2707 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2708 return iff; 2709 } 2710 2711 2712 //-------------------------------gen_subtype_check----------------------------- 2713 // Generate a subtyping check. Takes as input the subtype and supertype. 2714 // Returns 2 values: sets the default control() to the true path and returns 2715 // the false path. Only reads invariant memory; sets no (visible) memory. 2716 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2717 // but that's not exposed to the optimizer. This call also doesn't take in an 2718 // Object; if you wish to check an Object you need to load the Object's class 2719 // prior to coming here. 2720 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2721 Compile* C = gvn->C; 2722 2723 if ((*ctrl)->is_top()) { 2724 return C->top(); 2725 } 2726 2727 // Fast check for identical types, perhaps identical constants. 2728 // The types can even be identical non-constants, in cases 2729 // involving Array.newInstance, Object.clone, etc. 2730 if (subklass == superklass) 2731 return C->top(); // false path is dead; no test needed. 2732 2733 if (gvn->type(superklass)->singleton()) { 2734 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2735 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2736 2737 // In the common case of an exact superklass, try to fold up the 2738 // test before generating code. You may ask, why not just generate 2739 // the code and then let it fold up? The answer is that the generated 2740 // code will necessarily include null checks, which do not always 2741 // completely fold away. If they are also needless, then they turn 2742 // into a performance loss. Example: 2743 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2744 // Here, the type of 'fa' is often exact, so the store check 2745 // of fa[1]=x will fold up, without testing the nullness of x. 2746 switch (C->static_subtype_check(superk, subk)) { 2747 case Compile::SSC_always_false: 2748 { 2749 Node* always_fail = *ctrl; 2750 *ctrl = gvn->C->top(); 2751 return always_fail; 2752 } 2753 case Compile::SSC_always_true: 2754 return C->top(); 2755 case Compile::SSC_easy_test: 2756 { 2757 // Just do a direct pointer compare and be done. 2758 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2759 *ctrl = gvn->transform(new IfTrueNode(iff)); 2760 return gvn->transform(new IfFalseNode(iff)); 2761 } 2762 case Compile::SSC_full_test: 2763 break; 2764 default: 2765 ShouldNotReachHere(); 2766 } 2767 } 2768 2769 // %%% Possible further optimization: Even if the superklass is not exact, 2770 // if the subklass is the unique subtype of the superklass, the check 2771 // will always succeed. We could leave a dependency behind to ensure this. 2772 2773 // First load the super-klass's check-offset 2774 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2775 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2776 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2777 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2778 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2779 2780 // Load from the sub-klass's super-class display list, or a 1-word cache of 2781 // the secondary superclass list, or a failing value with a sentinel offset 2782 // if the super-klass is an interface or exceptionally deep in the Java 2783 // hierarchy and we have to scan the secondary superclass list the hard way. 2784 // Worst-case type is a little odd: NULL is allowed as a result (usually 2785 // klass loads can never produce a NULL). 2786 Node *chk_off_X = chk_off; 2787 #ifdef _LP64 2788 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2789 #endif 2790 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2791 // For some types like interfaces the following loadKlass is from a 1-word 2792 // cache which is mutable so can't use immutable memory. Other 2793 // types load from the super-class display table which is immutable. 2794 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2795 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2796 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM)); 2797 2798 // Compile speed common case: ARE a subtype and we canNOT fail 2799 if( superklass == nkls ) 2800 return C->top(); // false path is dead; no test needed. 2801 2802 // See if we get an immediate positive hit. Happens roughly 83% of the 2803 // time. Test to see if the value loaded just previously from the subklass 2804 // is exactly the superklass. 2805 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2806 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2807 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2808 2809 // Compile speed common case: Check for being deterministic right now. If 2810 // chk_off is a constant and not equal to cacheoff then we are NOT a 2811 // subklass. In this case we need exactly the 1 test above and we can 2812 // return those results immediately. 2813 if (!might_be_cache) { 2814 Node* not_subtype_ctrl = *ctrl; 2815 *ctrl = iftrue1; // We need exactly the 1 test above 2816 return not_subtype_ctrl; 2817 } 2818 2819 // Gather the various success & failures here 2820 RegionNode *r_ok_subtype = new RegionNode(4); 2821 gvn->record_for_igvn(r_ok_subtype); 2822 RegionNode *r_not_subtype = new RegionNode(3); 2823 gvn->record_for_igvn(r_not_subtype); 2824 2825 r_ok_subtype->init_req(1, iftrue1); 2826 2827 // Check for immediate negative hit. Happens roughly 11% of the time (which 2828 // is roughly 63% of the remaining cases). Test to see if the loaded 2829 // check-offset points into the subklass display list or the 1-element 2830 // cache. If it points to the display (and NOT the cache) and the display 2831 // missed then it's not a subtype. 2832 Node *cacheoff = gvn->intcon(cacheoff_con); 2833 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2834 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2835 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2836 2837 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2838 // No performance impact (too rare) but allows sharing of secondary arrays 2839 // which has some footprint reduction. 2840 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2841 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2842 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2843 2844 // -- Roads not taken here: -- 2845 // We could also have chosen to perform the self-check at the beginning 2846 // of this code sequence, as the assembler does. This would not pay off 2847 // the same way, since the optimizer, unlike the assembler, can perform 2848 // static type analysis to fold away many successful self-checks. 2849 // Non-foldable self checks work better here in second position, because 2850 // the initial primary superclass check subsumes a self-check for most 2851 // types. An exception would be a secondary type like array-of-interface, 2852 // which does not appear in its own primary supertype display. 2853 // Finally, we could have chosen to move the self-check into the 2854 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2855 // dependent manner. But it is worthwhile to have the check here, 2856 // where it can be perhaps be optimized. The cost in code space is 2857 // small (register compare, branch). 2858 2859 // Now do a linear scan of the secondary super-klass array. Again, no real 2860 // performance impact (too rare) but it's gotta be done. 2861 // Since the code is rarely used, there is no penalty for moving it 2862 // out of line, and it can only improve I-cache density. 2863 // The decision to inline or out-of-line this final check is platform 2864 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2865 Node* psc = gvn->transform( 2866 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2867 2868 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2869 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2870 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2871 2872 // Return false path; set default control to true path. 2873 *ctrl = gvn->transform(r_ok_subtype); 2874 return gvn->transform(r_not_subtype); 2875 } 2876 2877 // Profile-driven exact type check: 2878 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2879 float prob, 2880 Node* *casted_receiver) { 2881 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2882 Node* recv_klass = load_object_klass(receiver); 2883 Node* fail = type_check(recv_klass, tklass, prob); 2884 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2885 assert(recv_xtype->klass_is_exact(), ""); 2886 2887 // Subsume downstream occurrences of receiver with a cast to 2888 // recv_xtype, since now we know what the type will be. 2889 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2890 Node* res = _gvn.transform(cast); 2891 if (recv_xtype->is_valuetypeptr()) { 2892 assert(!gvn().type(res)->maybe_null(), "should never be null"); 2893 if (recv_xtype->value_klass()->is_scalarizable()) { 2894 res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass()); 2895 } 2896 } 2897 2898 (*casted_receiver) = res; 2899 // (User must make the replace_in_map call.) 2900 2901 return fail; 2902 } 2903 2904 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 2905 float prob) { 2906 Node* want_klass = makecon(tklass); 2907 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass)); 2908 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2909 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2910 set_control( _gvn.transform( new IfTrueNode (iff))); 2911 Node* fail = _gvn.transform( new IfFalseNode(iff)); 2912 return fail; 2913 } 2914 2915 2916 //------------------------------seems_never_null------------------------------- 2917 // Use null_seen information if it is available from the profile. 2918 // If we see an unexpected null at a type check we record it and force a 2919 // recompile; the offending check will be recompiled to handle NULLs. 2920 // If we see several offending BCIs, then all checks in the 2921 // method will be recompiled. 2922 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2923 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2924 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2925 if (UncommonNullCast // Cutout for this technique 2926 && obj != null() // And not the -Xcomp stupid case? 2927 && !too_many_traps(reason) 2928 ) { 2929 if (speculating) { 2930 return true; 2931 } 2932 if (data == NULL) 2933 // Edge case: no mature data. Be optimistic here. 2934 return true; 2935 // If the profile has not seen a null, assume it won't happen. 2936 assert(java_bc() == Bytecodes::_checkcast || 2937 java_bc() == Bytecodes::_instanceof || 2938 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2939 return !data->as_BitData()->null_seen(); 2940 } 2941 speculating = false; 2942 return false; 2943 } 2944 2945 //------------------------maybe_cast_profiled_receiver------------------------- 2946 // If the profile has seen exactly one type, narrow to exactly that type. 2947 // Subsequent type checks will always fold up. 2948 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2949 ciKlass* require_klass, 2950 ciKlass* spec_klass, 2951 bool safe_for_replace) { 2952 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2953 2954 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2955 2956 // Make sure we haven't already deoptimized from this tactic. 2957 if (too_many_traps(reason) || too_many_recompiles(reason)) 2958 return NULL; 2959 2960 // (No, this isn't a call, but it's enough like a virtual call 2961 // to use the same ciMethod accessor to get the profile info...) 2962 // If we have a speculative type use it instead of profiling (which 2963 // may not help us) 2964 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2965 if (exact_kls != NULL) {// no cast failures here 2966 if (require_klass == NULL || 2967 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2968 // If we narrow the type to match what the type profile sees or 2969 // the speculative type, we can then remove the rest of the 2970 // cast. 2971 // This is a win, even if the exact_kls is very specific, 2972 // because downstream operations, such as method calls, 2973 // will often benefit from the sharper type. 2974 Node* exact_obj = not_null_obj; // will get updated in place... 2975 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2976 &exact_obj); 2977 { PreserveJVMState pjvms(this); 2978 set_control(slow_ctl); 2979 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2980 } 2981 if (safe_for_replace) { 2982 replace_in_map(not_null_obj, exact_obj); 2983 } 2984 return exact_obj; 2985 } 2986 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2987 } 2988 2989 return NULL; 2990 } 2991 2992 /** 2993 * Cast obj to type and emit guard unless we had too many traps here 2994 * already 2995 * 2996 * @param obj node being casted 2997 * @param type type to cast the node to 2998 * @param not_null true if we know node cannot be null 2999 */ 3000 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3001 ciKlass* type, 3002 bool not_null) { 3003 if (stopped()) { 3004 return obj; 3005 } 3006 3007 // type == NULL if profiling tells us this object is always null 3008 if (type != NULL) { 3009 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3010 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3011 3012 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 3013 !too_many_traps(class_reason) && 3014 !too_many_recompiles(class_reason)) { 3015 Node* not_null_obj = NULL; 3016 // not_null is true if we know the object is not null and 3017 // there's no need for a null check 3018 if (!not_null) { 3019 Node* null_ctl = top(); 3020 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3021 assert(null_ctl->is_top(), "no null control here"); 3022 } else { 3023 not_null_obj = obj; 3024 } 3025 3026 Node* exact_obj = not_null_obj; 3027 ciKlass* exact_kls = type; 3028 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3029 &exact_obj); 3030 { 3031 PreserveJVMState pjvms(this); 3032 set_control(slow_ctl); 3033 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3034 } 3035 replace_in_map(not_null_obj, exact_obj); 3036 obj = exact_obj; 3037 } 3038 } else { 3039 if (!too_many_traps(Deoptimization::Reason_null_assert) && 3040 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 3041 Node* exact_obj = null_assert(obj); 3042 replace_in_map(obj, exact_obj); 3043 obj = exact_obj; 3044 } 3045 } 3046 return obj; 3047 } 3048 3049 //-------------------------------gen_instanceof-------------------------------- 3050 // Generate an instance-of idiom. Used by both the instance-of bytecode 3051 // and the reflective instance-of call. 3052 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3053 kill_dead_locals(); // Benefit all the uncommon traps 3054 assert( !stopped(), "dead parse path should be checked in callers" ); 3055 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3056 "must check for not-null not-dead klass in callers"); 3057 3058 // Make the merge point 3059 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3060 RegionNode* region = new RegionNode(PATH_LIMIT); 3061 Node* phi = new PhiNode(region, TypeInt::BOOL); 3062 C->set_has_split_ifs(true); // Has chance for split-if optimization 3063 3064 ciProfileData* data = NULL; 3065 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3066 data = method()->method_data()->bci_to_data(bci()); 3067 } 3068 bool speculative_not_null = false; 3069 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3070 && seems_never_null(obj, data, speculative_not_null)); 3071 bool is_value = obj->is_ValueType(); 3072 3073 // Null check; get casted pointer; set region slot 3 3074 Node* null_ctl = top(); 3075 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3076 3077 // If not_null_obj is dead, only null-path is taken 3078 if (stopped()) { // Doing instance-of on a NULL? 3079 set_control(null_ctl); 3080 return intcon(0); 3081 } 3082 region->init_req(_null_path, null_ctl); 3083 phi ->init_req(_null_path, intcon(0)); // Set null path value 3084 if (null_ctl == top()) { 3085 // Do this eagerly, so that pattern matches like is_diamond_phi 3086 // will work even during parsing. 3087 assert(_null_path == PATH_LIMIT-1, "delete last"); 3088 region->del_req(_null_path); 3089 phi ->del_req(_null_path); 3090 } 3091 3092 // Do we know the type check always succeed? 3093 if (!is_value) { 3094 bool known_statically = false; 3095 if (_gvn.type(superklass)->singleton()) { 3096 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 3097 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 3098 if (subk != NULL && subk->is_loaded()) { 3099 int static_res = C->static_subtype_check(superk, subk); 3100 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3101 } 3102 } 3103 3104 if (!known_statically) { 3105 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3106 // We may not have profiling here or it may not help us. If we 3107 // have a speculative type use it to perform an exact cast. 3108 ciKlass* spec_obj_type = obj_type->speculative_type(); 3109 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 3110 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 3111 if (stopped()) { // Profile disagrees with this path. 3112 set_control(null_ctl); // Null is the only remaining possibility. 3113 return intcon(0); 3114 } 3115 if (cast_obj != NULL && 3116 // A value that's sometimes null is not something we can optimize well 3117 !(cast_obj->is_ValueType() && null_ctl != top())) { 3118 not_null_obj = cast_obj; 3119 is_value = not_null_obj->is_ValueType(); 3120 } 3121 } 3122 } 3123 } 3124 3125 // Load the object's klass 3126 Node* obj_klass = NULL; 3127 if (is_value) { 3128 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass())); 3129 } else { 3130 obj_klass = load_object_klass(not_null_obj); 3131 } 3132 3133 // Generate the subtype check 3134 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 3135 3136 // Plug in the success path to the general merge in slot 1. 3137 region->init_req(_obj_path, control()); 3138 phi ->init_req(_obj_path, intcon(1)); 3139 3140 // Plug in the failing path to the general merge in slot 2. 3141 region->init_req(_fail_path, not_subtype_ctrl); 3142 phi ->init_req(_fail_path, intcon(0)); 3143 3144 // Return final merged results 3145 set_control( _gvn.transform(region) ); 3146 record_for_igvn(region); 3147 3148 // If we know the type check always succeeds then we don't use the 3149 // profiling data at this bytecode. Don't lose it, feed it to the 3150 // type system as a speculative type. 3151 if (safe_for_replace && !is_value) { 3152 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3153 replace_in_map(obj, casted_obj); 3154 } 3155 3156 return _gvn.transform(phi); 3157 } 3158 3159 //-------------------------------gen_checkcast--------------------------------- 3160 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3161 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3162 // uncommon-trap paths work. Adjust stack after this call. 3163 // If failure_control is supplied and not null, it is filled in with 3164 // the control edge for the cast failure. Otherwise, an appropriate 3165 // uncommon trap or exception is thrown. 3166 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3167 Node* *failure_control) { 3168 kill_dead_locals(); // Benefit all the uncommon traps 3169 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr(); 3170 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass()); 3171 3172 bool is_value = obj->is_ValueType(); 3173 3174 // Fast cutout: Check the case that the cast is vacuously true. 3175 // This detects the common cases where the test will short-circuit 3176 // away completely. We do this before we perform the null check, 3177 // because if the test is going to turn into zero code, we don't 3178 // want a residual null check left around. (Causes a slowdown, 3179 // for example, in some objArray manipulations, such as a[i]=a[j].) 3180 if (tk->singleton()) { 3181 ciKlass* klass = NULL; 3182 if (is_value) { 3183 klass = _gvn.type(obj)->is_valuetype()->value_klass(); 3184 } else { 3185 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3186 if (objtp != NULL) { 3187 klass = objtp->klass(); 3188 } 3189 } 3190 if (klass != NULL) { 3191 switch (C->static_subtype_check(tk->klass(), klass)) { 3192 case Compile::SSC_always_true: 3193 // If we know the type check always succeed then we don't use 3194 // the profiling data at this bytecode. Don't lose it, feed it 3195 // to the type system as a speculative type. 3196 if (!is_value) { 3197 obj = record_profiled_receiver_for_speculation(obj); 3198 if (toop->is_valuetypeptr()) { 3199 if (toop->value_klass()->is_scalarizable()) { 3200 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass(), /* buffer_check */ false, /* null2default */ false); 3201 } else { 3202 obj = filter_null(obj); 3203 } 3204 } 3205 } 3206 return obj; 3207 case Compile::SSC_always_false: 3208 // It needs a null check because a null will *pass* the cast check. 3209 if (is_value || toop->is_valuetypeptr()) { 3210 // Value types are never null - always throw an exception 3211 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass))); 3212 return top(); 3213 } else { 3214 return null_assert(obj); 3215 } 3216 } 3217 } 3218 } 3219 3220 ciProfileData* data = NULL; 3221 bool safe_for_replace = false; 3222 if (failure_control == NULL) { // use MDO in regular case only 3223 assert(java_bc() == Bytecodes::_aastore || 3224 java_bc() == Bytecodes::_checkcast, 3225 "interpreter profiles type checks only for these BCs"); 3226 data = method()->method_data()->bci_to_data(bci()); 3227 safe_for_replace = true; 3228 } 3229 3230 // Make the merge point 3231 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3232 RegionNode* region = new RegionNode(PATH_LIMIT); 3233 Node* phi = new PhiNode(region, toop); 3234 C->set_has_split_ifs(true); // Has chance for split-if optimization 3235 3236 // Use null-cast information if it is available 3237 bool speculative_not_null = false; 3238 bool never_see_null = ((failure_control == NULL) // regular case only 3239 && seems_never_null(obj, data, speculative_not_null)); 3240 3241 // Null check; get casted pointer; set region slot 3 3242 Node* null_ctl = top(); 3243 Node* not_null_obj = NULL; 3244 if (is_value) { 3245 not_null_obj = obj; 3246 } else if (toop->is_valuetypeptr()) { 3247 not_null_obj = filter_null(obj); 3248 } else { 3249 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3250 } 3251 3252 // If not_null_obj is dead, only null-path is taken 3253 if (stopped()) { // Doing instance-of on a NULL? 3254 set_control(null_ctl); 3255 return null(); 3256 } 3257 region->init_req(_null_path, null_ctl); 3258 phi ->init_req(_null_path, null()); // Set null path value 3259 if (null_ctl == top()) { 3260 // Do this eagerly, so that pattern matches like is_diamond_phi 3261 // will work even during parsing. 3262 assert(_null_path == PATH_LIMIT-1, "delete last"); 3263 region->del_req(_null_path); 3264 phi ->del_req(_null_path); 3265 } 3266 3267 Node* cast_obj = NULL; 3268 if (!is_value && tk->klass_is_exact()) { 3269 // The following optimization tries to statically cast the speculative type of the object 3270 // (for example obtained during profiling) to the type of the superklass and then do a 3271 // dynamic check that the type of the object is what we expect. To work correctly 3272 // for checkcast and aastore the type of superklass should be exact. 3273 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3274 // We may not have profiling here or it may not help us. If we have 3275 // a speculative type use it to perform an exact cast. 3276 ciKlass* spec_obj_type = obj_type->speculative_type(); 3277 if (spec_obj_type != NULL || data != NULL) { 3278 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3279 if (cast_obj != NULL && cast_obj->is_ValueType()) { 3280 if (null_ctl != top()) { 3281 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well 3282 } else { 3283 return cast_obj; 3284 } 3285 } 3286 if (cast_obj != NULL) { 3287 if (failure_control != NULL) // failure is now impossible 3288 (*failure_control) = top(); 3289 // adjust the type of the phi to the exact klass: 3290 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3291 } 3292 } 3293 } 3294 3295 if (cast_obj == NULL) { 3296 // Load the object's klass 3297 Node* obj_klass = NULL; 3298 if (is_value) { 3299 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->is_valuetype()->value_klass())); 3300 } else { 3301 obj_klass = load_object_klass(not_null_obj); 3302 } 3303 3304 // Generate the subtype check 3305 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3306 3307 // Plug in success path into the merge 3308 cast_obj = is_value ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3309 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3310 if (failure_control == NULL) { 3311 if (not_subtype_ctrl != top()) { // If failure is possible 3312 PreserveJVMState pjvms(this); 3313 set_control(not_subtype_ctrl); 3314 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3315 } 3316 } else { 3317 (*failure_control) = not_subtype_ctrl; 3318 } 3319 } 3320 3321 region->init_req(_obj_path, control()); 3322 phi ->init_req(_obj_path, cast_obj); 3323 3324 // A merge of NULL or Casted-NotNull obj 3325 Node* res = _gvn.transform(phi); 3326 3327 // Note I do NOT always 'replace_in_map(obj,result)' here. 3328 // if( tk->klass()->can_be_primary_super() ) 3329 // This means that if I successfully store an Object into an array-of-String 3330 // I 'forget' that the Object is really now known to be a String. I have to 3331 // do this because we don't have true union types for interfaces - if I store 3332 // a Baz into an array-of-Interface and then tell the optimizer it's an 3333 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3334 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3335 // replace_in_map( obj, res ); 3336 3337 // Return final merged results 3338 set_control( _gvn.transform(region) ); 3339 record_for_igvn(region); 3340 3341 if (!is_value) { 3342 res = record_profiled_receiver_for_speculation(res); 3343 if (toop->is_valuetypeptr()) { 3344 assert(!gvn().type(res)->maybe_null(), "should never be null"); 3345 if (toop->value_klass()->is_scalarizable()) { 3346 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass()); 3347 } 3348 } 3349 } 3350 return res; 3351 } 3352 3353 Node* GraphKit::is_always_locked(Node* obj) { 3354 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3355 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3356 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern); 3357 return _gvn.transform(new AndXNode(mark, value_mask)); 3358 } 3359 3360 Node* GraphKit::gen_value_type_test(Node* kls) { 3361 Node* flags_addr = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3362 Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered); 3363 Node* is_value = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE))); 3364 Node* cmp = _gvn.transform(new CmpINode(is_value, intcon(0))); 3365 return cmp; 3366 } 3367 3368 // Deoptimize if 'obj' is a value type 3369 void GraphKit::gen_value_type_guard(Node* obj, int nargs) { 3370 assert(EnableValhalla, "should only be used if value types are enabled"); 3371 Node* bol = NULL; 3372 if (obj->is_ValueTypeBase()) { 3373 bol = intcon(0); 3374 } else { 3375 Node* is_value = is_always_locked(obj); 3376 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern); 3377 Node* cmp = _gvn.transform(new CmpXNode(is_value, value_mask)); 3378 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3379 } 3380 { BuildCutout unless(this, bol, PROB_MAX); 3381 inc_sp(nargs); 3382 uncommon_trap(Deoptimization::Reason_class_check, 3383 Deoptimization::Action_none); 3384 } 3385 } 3386 3387 // Deoptimize if 'ary' is flattened or if 'obj' is null and 'ary' is a value type array 3388 void GraphKit::gen_value_type_array_guard(Node* ary, Node* obj, int nargs) { 3389 assert(EnableValhalla, "should only be used if value types are enabled"); 3390 // Load array element klass 3391 Node* kls = load_object_klass(ary); 3392 Node* k_adr = basic_plus_adr(kls, in_bytes(ArrayKlass::element_klass_offset())); 3393 Node* elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 3394 // Check if element is a value type 3395 Node* flags_addr = basic_plus_adr(elem_klass, in_bytes(Klass::access_flags_offset())); 3396 Node* flags = make_load(NULL, flags_addr, TypeInt::INT, T_INT, MemNode::unordered); 3397 Node* is_value_elem = _gvn.transform(new AndINode(flags, intcon(JVM_ACC_VALUE))); 3398 3399 const Type* objtype = _gvn.type(obj); 3400 if (objtype == TypePtr::NULL_PTR) { 3401 // Object is always null, check if array is a value type array 3402 Node* cmp = _gvn.transform(new CmpINode(is_value_elem, intcon(0))); 3403 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3404 { BuildCutout unless(this, bol, PROB_MAX); 3405 // TODO just deoptimize for now if we store null to a value type array 3406 inc_sp(nargs); 3407 uncommon_trap(Deoptimization::Reason_array_check, 3408 Deoptimization::Action_none); 3409 } 3410 } else { 3411 // Check if (is_value_elem && obj_is_null) <=> (!is_value_elem | !obj_is_null == 0) 3412 // TODO what if we later figure out that obj is never null? 3413 Node* not_value = _gvn.transform(new XorINode(is_value_elem, intcon(JVM_ACC_VALUE))); 3414 not_value = _gvn.transform(new ConvI2LNode(not_value)); 3415 Node* not_null = _gvn.transform(new CastP2XNode(NULL, obj)); 3416 Node* both = _gvn.transform(new OrLNode(not_null, not_value)); 3417 Node* cmp = _gvn.transform(new CmpLNode(both, longcon(0))); 3418 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3419 { BuildCutout unless(this, bol, PROB_MAX); 3420 // TODO just deoptimize for now if we store null to a value type array 3421 inc_sp(nargs); 3422 uncommon_trap(Deoptimization::Reason_array_check, 3423 Deoptimization::Action_none); 3424 } 3425 } 3426 } 3427 3428 Node* GraphKit::load_lh_array_tag(Node* kls) { 3429 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset())); 3430 Node* layout_val = make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3431 return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift))); 3432 } 3433 3434 3435 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) { 3436 Node* layout_val = load_lh_array_tag(kls); 3437 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value))); 3438 return cmp; 3439 } 3440 3441 3442 //------------------------------next_monitor----------------------------------- 3443 // What number should be given to the next monitor? 3444 int GraphKit::next_monitor() { 3445 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3446 int next = current + C->sync_stack_slots(); 3447 // Keep the toplevel high water mark current: 3448 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3449 return current; 3450 } 3451 3452 //------------------------------insert_mem_bar--------------------------------- 3453 // Memory barrier to avoid floating things around 3454 // The membar serves as a pinch point between both control and all memory slices. 3455 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3456 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3457 mb->init_req(TypeFunc::Control, control()); 3458 mb->init_req(TypeFunc::Memory, reset_memory()); 3459 Node* membar = _gvn.transform(mb); 3460 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3461 set_all_memory_call(membar); 3462 return membar; 3463 } 3464 3465 //-------------------------insert_mem_bar_volatile---------------------------- 3466 // Memory barrier to avoid floating things around 3467 // The membar serves as a pinch point between both control and memory(alias_idx). 3468 // If you want to make a pinch point on all memory slices, do not use this 3469 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3470 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3471 // When Parse::do_put_xxx updates a volatile field, it appends a series 3472 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3473 // The first membar is on the same memory slice as the field store opcode. 3474 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3475 // All the other membars (for other volatile slices, including AliasIdxBot, 3476 // which stands for all unknown volatile slices) are control-dependent 3477 // on the first membar. This prevents later volatile loads or stores 3478 // from sliding up past the just-emitted store. 3479 3480 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3481 mb->set_req(TypeFunc::Control,control()); 3482 if (alias_idx == Compile::AliasIdxBot) { 3483 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3484 } else { 3485 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3486 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3487 } 3488 Node* membar = _gvn.transform(mb); 3489 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3490 if (alias_idx == Compile::AliasIdxBot) { 3491 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3492 } else { 3493 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3494 } 3495 return membar; 3496 } 3497 3498 //------------------------------shared_lock------------------------------------ 3499 // Emit locking code. 3500 FastLockNode* GraphKit::shared_lock(Node* obj) { 3501 // bci is either a monitorenter bc or InvocationEntryBci 3502 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3503 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3504 3505 if( !GenerateSynchronizationCode ) 3506 return NULL; // Not locking things? 3507 3508 // We cannot lock on a value type 3509 const TypeOopPtr* objptr = _gvn.type(obj)->make_oopptr(); 3510 if (objptr->can_be_value_type()) { 3511 gen_value_type_guard(obj, 1); 3512 } 3513 3514 if (stopped()) // Dead monitor? 3515 return NULL; 3516 3517 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3518 3519 // Box the stack location 3520 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3521 Node* mem = reset_memory(); 3522 3523 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3524 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3525 // Create the counters for this fast lock. 3526 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3527 } 3528 3529 // Create the rtm counters for this fast lock if needed. 3530 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3531 3532 // Add monitor to debug info for the slow path. If we block inside the 3533 // slow path and de-opt, we need the monitor hanging around 3534 map()->push_monitor( flock ); 3535 3536 const TypeFunc *tf = LockNode::lock_type(); 3537 LockNode *lock = new LockNode(C, tf); 3538 3539 lock->init_req( TypeFunc::Control, control() ); 3540 lock->init_req( TypeFunc::Memory , mem ); 3541 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3542 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3543 lock->init_req( TypeFunc::ReturnAdr, top() ); 3544 3545 lock->init_req(TypeFunc::Parms + 0, obj); 3546 lock->init_req(TypeFunc::Parms + 1, box); 3547 lock->init_req(TypeFunc::Parms + 2, flock); 3548 add_safepoint_edges(lock); 3549 3550 lock = _gvn.transform( lock )->as_Lock(); 3551 3552 // lock has no side-effects, sets few values 3553 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3554 3555 insert_mem_bar(Op_MemBarAcquireLock); 3556 3557 // Add this to the worklist so that the lock can be eliminated 3558 record_for_igvn(lock); 3559 3560 #ifndef PRODUCT 3561 if (PrintLockStatistics) { 3562 // Update the counter for this lock. Don't bother using an atomic 3563 // operation since we don't require absolute accuracy. 3564 lock->create_lock_counter(map()->jvms()); 3565 increment_counter(lock->counter()->addr()); 3566 } 3567 #endif 3568 3569 return flock; 3570 } 3571 3572 3573 //------------------------------shared_unlock---------------------------------- 3574 // Emit unlocking code. 3575 void GraphKit::shared_unlock(Node* box, Node* obj) { 3576 // bci is either a monitorenter bc or InvocationEntryBci 3577 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3578 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3579 3580 if( !GenerateSynchronizationCode ) 3581 return; 3582 if (stopped()) { // Dead monitor? 3583 map()->pop_monitor(); // Kill monitor from debug info 3584 return; 3585 } 3586 assert(!obj->is_ValueTypeBase(), "should not unlock on value type"); 3587 3588 // Memory barrier to avoid floating things down past the locked region 3589 insert_mem_bar(Op_MemBarReleaseLock); 3590 3591 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3592 UnlockNode *unlock = new UnlockNode(C, tf); 3593 #ifdef ASSERT 3594 unlock->set_dbg_jvms(sync_jvms()); 3595 #endif 3596 uint raw_idx = Compile::AliasIdxRaw; 3597 unlock->init_req( TypeFunc::Control, control() ); 3598 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3599 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3600 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3601 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3602 3603 unlock->init_req(TypeFunc::Parms + 0, obj); 3604 unlock->init_req(TypeFunc::Parms + 1, box); 3605 unlock = _gvn.transform(unlock)->as_Unlock(); 3606 3607 Node* mem = reset_memory(); 3608 3609 // unlock has no side-effects, sets few values 3610 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3611 3612 // Kill monitor from debug info 3613 map()->pop_monitor( ); 3614 } 3615 3616 //-------------------------------get_layout_helper----------------------------- 3617 // If the given klass is a constant or known to be an array, 3618 // fetch the constant layout helper value into constant_value 3619 // and return (Node*)NULL. Otherwise, load the non-constant 3620 // layout helper value, and return the node which represents it. 3621 // This two-faced routine is useful because allocation sites 3622 // almost always feature constant types. 3623 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3624 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3625 if (!StressReflectiveCode && inst_klass != NULL) { 3626 ciKlass* klass = inst_klass->klass(); 3627 assert(klass != NULL, "klass should not be NULL"); 3628 bool xklass = inst_klass->klass_is_exact(); 3629 bool can_be_value_array = false; 3630 if (klass->is_array_klass() && EnableValhalla && ValueArrayFlatten) { 3631 ciKlass* elem = klass->as_array_klass()->element_klass(); 3632 can_be_value_array = elem != NULL && (elem->is_java_lang_Object() || elem->is_interface()); 3633 } 3634 if (xklass || (klass->is_array_klass() && !can_be_value_array)) { 3635 jint lhelper = klass->layout_helper(); 3636 if (lhelper != Klass::_lh_neutral_value) { 3637 constant_value = lhelper; 3638 return (Node*) NULL; 3639 } 3640 } 3641 } 3642 constant_value = Klass::_lh_neutral_value; // put in a known value 3643 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3644 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3645 } 3646 3647 // We just put in an allocate/initialize with a big raw-memory effect. 3648 // Hook selected additional alias categories on the initialization. 3649 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3650 MergeMemNode* init_in_merge, 3651 Node* init_out_raw) { 3652 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3653 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3654 3655 Node* prevmem = kit.memory(alias_idx); 3656 init_in_merge->set_memory_at(alias_idx, prevmem); 3657 kit.set_memory(init_out_raw, alias_idx); 3658 } 3659 3660 //---------------------------set_output_for_allocation------------------------- 3661 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3662 const TypeOopPtr* oop_type, 3663 bool deoptimize_on_exception) { 3664 int rawidx = Compile::AliasIdxRaw; 3665 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3666 add_safepoint_edges(alloc); 3667 Node* allocx = _gvn.transform(alloc); 3668 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3669 // create memory projection for i_o 3670 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3671 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3672 3673 // create a memory projection as for the normal control path 3674 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3675 set_memory(malloc, rawidx); 3676 3677 // a normal slow-call doesn't change i_o, but an allocation does 3678 // we create a separate i_o projection for the normal control path 3679 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3680 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3681 3682 // put in an initialization barrier 3683 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3684 rawoop)->as_Initialize(); 3685 assert(alloc->initialization() == init, "2-way macro link must work"); 3686 assert(init ->allocation() == alloc, "2-way macro link must work"); 3687 { 3688 // Extract memory strands which may participate in the new object's 3689 // initialization, and source them from the new InitializeNode. 3690 // This will allow us to observe initializations when they occur, 3691 // and link them properly (as a group) to the InitializeNode. 3692 assert(init->in(InitializeNode::Memory) == malloc, ""); 3693 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3694 init->set_req(InitializeNode::Memory, minit_in); 3695 record_for_igvn(minit_in); // fold it up later, if possible 3696 Node* minit_out = memory(rawidx); 3697 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3698 if (oop_type->isa_aryptr()) { 3699 const TypeAryPtr* arytype = oop_type->is_aryptr(); 3700 if (arytype->klass()->is_value_array_klass()) { 3701 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass(); 3702 ciValueKlass* vk = vak->element_klass()->as_value_klass(); 3703 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) { 3704 ciField* field = vk->nonstatic_field_at(i); 3705 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3706 continue; // do not bother to track really large numbers of fields 3707 int off_in_vt = field->offset() - vk->first_field_offset(); 3708 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot); 3709 int fieldidx = C->get_alias_index(adr_type); 3710 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3711 } 3712 } else { 3713 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3714 int elemidx = C->get_alias_index(telemref); 3715 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3716 } 3717 } else if (oop_type->isa_instptr()) { 3718 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3719 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3720 ciField* field = ik->nonstatic_field_at(i); 3721 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3722 continue; // do not bother to track really large numbers of fields 3723 // Find (or create) the alias category for this field: 3724 int fieldidx = C->alias_type(field)->index(); 3725 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3726 } 3727 } 3728 } 3729 3730 // Cast raw oop to the real thing... 3731 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3732 javaoop = _gvn.transform(javaoop); 3733 C->set_recent_alloc(control(), javaoop); 3734 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3735 3736 #ifdef ASSERT 3737 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3738 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3739 "Ideal_allocation works"); 3740 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3741 "Ideal_allocation works"); 3742 if (alloc->is_AllocateArray()) { 3743 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3744 "Ideal_allocation works"); 3745 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3746 "Ideal_allocation works"); 3747 } else { 3748 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3749 } 3750 } 3751 #endif //ASSERT 3752 3753 return javaoop; 3754 } 3755 3756 //---------------------------new_instance-------------------------------------- 3757 // This routine takes a klass_node which may be constant (for a static type) 3758 // or may be non-constant (for reflective code). It will work equally well 3759 // for either, and the graph will fold nicely if the optimizer later reduces 3760 // the type to a constant. 3761 // The optional arguments are for specialized use by intrinsics: 3762 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3763 // - If 'return_size_val', report the the total object size to the caller. 3764 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3765 Node* GraphKit::new_instance(Node* klass_node, 3766 Node* extra_slow_test, 3767 Node* *return_size_val, 3768 bool deoptimize_on_exception, 3769 ValueTypeBaseNode* value_node) { 3770 // Compute size in doublewords 3771 // The size is always an integral number of doublewords, represented 3772 // as a positive bytewise size stored in the klass's layout_helper. 3773 // The layout_helper also encodes (in a low bit) the need for a slow path. 3774 jint layout_con = Klass::_lh_neutral_value; 3775 Node* layout_val = get_layout_helper(klass_node, layout_con); 3776 bool layout_is_con = (layout_val == NULL); 3777 3778 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3779 // Generate the initial go-slow test. It's either ALWAYS (return a 3780 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3781 // case) a computed value derived from the layout_helper. 3782 Node* initial_slow_test = NULL; 3783 if (layout_is_con) { 3784 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3785 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3786 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3787 } else { // reflective case 3788 // This reflective path is used by Unsafe.allocateInstance. 3789 // (It may be stress-tested by specifying StressReflectiveCode.) 3790 // Basically, we want to get into the VM is there's an illegal argument. 3791 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3792 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3793 if (extra_slow_test != intcon(0)) { 3794 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3795 } 3796 // (Macro-expander will further convert this to a Bool, if necessary.) 3797 } 3798 3799 // Find the size in bytes. This is easy; it's the layout_helper. 3800 // The size value must be valid even if the slow path is taken. 3801 Node* size = NULL; 3802 if (layout_is_con) { 3803 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3804 } else { // reflective case 3805 // This reflective path is used by clone and Unsafe.allocateInstance. 3806 size = ConvI2X(layout_val); 3807 3808 // Clear the low bits to extract layout_helper_size_in_bytes: 3809 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3810 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3811 size = _gvn.transform( new AndXNode(size, mask) ); 3812 } 3813 if (return_size_val != NULL) { 3814 (*return_size_val) = size; 3815 } 3816 3817 // This is a precise notnull oop of the klass. 3818 // (Actually, it need not be precise if this is a reflective allocation.) 3819 // It's what we cast the result to. 3820 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3821 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3822 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3823 3824 // Now generate allocation code 3825 3826 // The entire memory state is needed for slow path of the allocation 3827 // since GC and deoptimization can happen. 3828 Node *mem = reset_memory(); 3829 set_all_memory(mem); // Create new memory state 3830 3831 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3832 control(), mem, i_o(), 3833 size, klass_node, 3834 initial_slow_test, value_node); 3835 3836 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3837 } 3838 3839 // With compressed oops, the 64 bit init value for non flattened value 3840 // arrays is built from 2 32 bit compressed oops 3841 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) { 3842 Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value)); 3843 Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32))); 3844 return kit.gvn().transform(new OrLNode(lower, upper)); 3845 } 3846 3847 //-------------------------------new_array------------------------------------- 3848 // helper for newarray and anewarray 3849 // The 'length' parameter is (obviously) the length of the array. 3850 // See comments on new_instance for the meaning of the other arguments. 3851 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3852 Node* length, // number of array elements 3853 int nargs, // number of arguments to push back for uncommon trap 3854 Node* *return_size_val, 3855 bool deoptimize_on_exception) { 3856 jint layout_con = Klass::_lh_neutral_value; 3857 Node* layout_val = get_layout_helper(klass_node, layout_con); 3858 bool layout_is_con = (layout_val == NULL); 3859 3860 if (!layout_is_con && !StressReflectiveCode && 3861 !too_many_traps(Deoptimization::Reason_class_check)) { 3862 // This is a reflective array creation site. 3863 // Optimistically assume that it is a subtype of Object[], 3864 // so that we can fold up all the address arithmetic. 3865 layout_con = Klass::array_layout_helper(T_OBJECT); 3866 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3867 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3868 { BuildCutout unless(this, bol_lh, PROB_MAX); 3869 inc_sp(nargs); 3870 uncommon_trap(Deoptimization::Reason_class_check, 3871 Deoptimization::Action_maybe_recompile); 3872 } 3873 layout_val = NULL; 3874 layout_is_con = true; 3875 } 3876 3877 // Generate the initial go-slow test. Make sure we do not overflow 3878 // if length is huge (near 2Gig) or negative! We do not need 3879 // exact double-words here, just a close approximation of needed 3880 // double-words. We can't add any offset or rounding bits, lest we 3881 // take a size -1 of bytes and make it positive. Use an unsigned 3882 // compare, so negative sizes look hugely positive. 3883 int fast_size_limit = FastAllocateSizeLimit; 3884 if (layout_is_con) { 3885 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3886 // Increase the size limit if we have exact knowledge of array type. 3887 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3888 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0); 3889 } 3890 3891 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3892 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3893 3894 // --- Size Computation --- 3895 // array_size = round_to_heap(array_header + (length << elem_shift)); 3896 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3897 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3898 // The rounding mask is strength-reduced, if possible. 3899 int round_mask = MinObjAlignmentInBytes - 1; 3900 Node* header_size = NULL; 3901 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3902 // (T_BYTE has the weakest alignment and size restrictions...) 3903 if (layout_is_con) { 3904 int hsize = Klass::layout_helper_header_size(layout_con); 3905 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3906 BasicType etype = Klass::layout_helper_element_type(layout_con); 3907 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con); 3908 if ((round_mask & ~right_n_bits(eshift)) == 0) 3909 round_mask = 0; // strength-reduce it if it goes away completely 3910 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3911 assert(header_size_min <= hsize, "generic minimum is smallest"); 3912 header_size_min = hsize; 3913 header_size = intcon(hsize + round_mask); 3914 } else { 3915 Node* hss = intcon(Klass::_lh_header_size_shift); 3916 Node* hsm = intcon(Klass::_lh_header_size_mask); 3917 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3918 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3919 Node* mask = intcon(round_mask); 3920 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3921 } 3922 3923 Node* elem_shift = NULL; 3924 if (layout_is_con) { 3925 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3926 if (eshift != 0) 3927 elem_shift = intcon(eshift); 3928 } else { 3929 // There is no need to mask or shift this value. 3930 // The semantics of LShiftINode include an implicit mask to 0x1F. 3931 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3932 elem_shift = layout_val; 3933 } 3934 3935 // Transition to native address size for all offset calculations: 3936 Node* lengthx = ConvI2X(length); 3937 Node* headerx = ConvI2X(header_size); 3938 #ifdef _LP64 3939 { const TypeInt* tilen = _gvn.find_int_type(length); 3940 if (tilen != NULL && tilen->_lo < 0) { 3941 // Add a manual constraint to a positive range. Cf. array_element_address. 3942 jint size_max = fast_size_limit; 3943 if (size_max > tilen->_hi) size_max = tilen->_hi; 3944 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3945 3946 // Only do a narrow I2L conversion if the range check passed. 3947 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3948 _gvn.transform(iff); 3949 RegionNode* region = new RegionNode(3); 3950 _gvn.set_type(region, Type::CONTROL); 3951 lengthx = new PhiNode(region, TypeLong::LONG); 3952 _gvn.set_type(lengthx, TypeLong::LONG); 3953 3954 // Range check passed. Use ConvI2L node with narrow type. 3955 Node* passed = IfFalse(iff); 3956 region->init_req(1, passed); 3957 // Make I2L conversion control dependent to prevent it from 3958 // floating above the range check during loop optimizations. 3959 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3960 3961 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3962 region->init_req(2, IfTrue(iff)); 3963 lengthx->init_req(2, ConvI2X(length)); 3964 3965 set_control(region); 3966 record_for_igvn(region); 3967 record_for_igvn(lengthx); 3968 } 3969 } 3970 #endif 3971 3972 // Combine header size (plus rounding) and body size. Then round down. 3973 // This computation cannot overflow, because it is used only in two 3974 // places, one where the length is sharply limited, and the other 3975 // after a successful allocation. 3976 Node* abody = lengthx; 3977 if (elem_shift != NULL) 3978 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3979 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3980 if (round_mask != 0) { 3981 Node* mask = MakeConX(~round_mask); 3982 size = _gvn.transform( new AndXNode(size, mask) ); 3983 } 3984 // else if round_mask == 0, the size computation is self-rounding 3985 3986 if (return_size_val != NULL) { 3987 // This is the size 3988 (*return_size_val) = size; 3989 } 3990 3991 // Now generate allocation code 3992 3993 // The entire memory state is needed for slow path of the allocation 3994 // since GC and deoptimization can happen. 3995 Node *mem = reset_memory(); 3996 set_all_memory(mem); // Create new memory state 3997 3998 if (initial_slow_test->is_Bool()) { 3999 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4000 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4001 } 4002 4003 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 4004 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 4005 const Type* elem = NULL; 4006 ciKlass* elem_klass = NULL; 4007 if (ary_ptr != NULL) { 4008 elem = ary_ptr->elem(); 4009 elem_klass = ary_ptr->klass()->as_array_klass()->element_klass(); 4010 } 4011 Node* default_value = NULL; 4012 Node* raw_default_value = NULL; 4013 if (elem != NULL && elem->make_ptr()) { 4014 if (elem_klass != NULL && elem_klass->is_valuetype()) { 4015 ciValueKlass* vk = elem_klass->as_value_klass(); 4016 if (!vk->flatten_array()) { 4017 default_value = ValueTypeNode::default_oop(gvn(), vk); 4018 if (elem->isa_narrowoop()) { 4019 default_value = _gvn.transform(new EncodePNode(default_value, elem)); 4020 raw_default_value = raw_default_for_coops(default_value, *this); 4021 } else { 4022 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value)); 4023 } 4024 } 4025 } 4026 } 4027 4028 if (EnableValhalla && (elem == NULL || (elem_klass != NULL && elem_klass->is_java_lang_Object() && !ary_type->klass_is_exact()))) { 4029 assert(raw_default_value == NULL, "shouldn't be set yet"); 4030 4031 // unkown array type, could be a non flattened value array that's 4032 // initialize to a non zero default value 4033 4034 Node* r = new RegionNode(4); 4035 Node* phi = new PhiNode(r, TypeX_X); 4036 4037 Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_obj_value); 4038 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 4039 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4040 r->init_req(1, _gvn.transform(new IfFalseNode(iff))); 4041 phi->init_req(1, MakeConX(0)); 4042 set_control(_gvn.transform(new IfTrueNode(iff))); 4043 Node* k_adr = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset())); 4044 Node* elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), k_adr, TypeInstPtr::KLASS)); 4045 cmp = gen_value_type_test(elem_klass); 4046 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 4047 iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4048 r->init_req(2, _gvn.transform(new IfTrueNode(iff))); 4049 phi->init_req(2, MakeConX(0)); 4050 set_control(_gvn.transform(new IfFalseNode(iff))); 4051 4052 Node* adr_fixed_block_addr = basic_plus_adr(elem_klass, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())); 4053 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4054 4055 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(ValueKlass::default_value_offset_offset())); 4056 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered); 4057 4058 Node* elem_mirror = load_mirror_from_klass(elem_klass); 4059 4060 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset)); 4061 const TypePtr* adr_type = _gvn.type(default_value_addr)->is_ptr(); 4062 Node* val = access_load_at(elem_mirror, default_value_addr, adr_type, TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP); 4063 4064 if (UseCompressedOops) { 4065 val = _gvn.transform(new EncodePNode(val, elem)); 4066 val = raw_default_for_coops(val, *this); 4067 } else { 4068 val = _gvn.transform(new CastP2XNode(control(), val)); 4069 } 4070 r->init_req(3, control()); 4071 phi->init_req(3, val); 4072 set_control(_gvn.transform(r)); 4073 raw_default_value = _gvn.transform(phi); 4074 } 4075 4076 // Create the AllocateArrayNode and its result projections 4077 AllocateArrayNode* alloc 4078 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4079 control(), mem, i_o(), 4080 size, klass_node, 4081 initial_slow_test, 4082 length, default_value, 4083 raw_default_value); 4084 4085 // Cast to correct type. Note that the klass_node may be constant or not, 4086 // and in the latter case the actual array type will be inexact also. 4087 // (This happens via a non-constant argument to inline_native_newArray.) 4088 // In any case, the value of klass_node provides the desired array type. 4089 const TypeInt* length_type = _gvn.find_int_type(length); 4090 if (ary_type->isa_aryptr() && length_type != NULL) { 4091 // Try to get a better type than POS for the size 4092 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4093 } 4094 4095 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4096 4097 // Cast length on remaining path to be as narrow as possible 4098 if (map()->find_edge(length) >= 0) { 4099 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 4100 if (ccast != length) { 4101 _gvn.set_type_bottom(ccast); 4102 record_for_igvn(ccast); 4103 replace_in_map(length, ccast); 4104 } 4105 } 4106 4107 return javaoop; 4108 } 4109 4110 // The following "Ideal_foo" functions are placed here because they recognize 4111 // the graph shapes created by the functions immediately above. 4112 4113 //---------------------------Ideal_allocation---------------------------------- 4114 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4115 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 4116 if (ptr == NULL) { // reduce dumb test in callers 4117 return NULL; 4118 } 4119 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4120 ptr = ptr->in(1); 4121 if (ptr == NULL) return NULL; 4122 } 4123 // Return NULL for allocations with several casts: 4124 // j.l.reflect.Array.newInstance(jobject, jint) 4125 // Object.clone() 4126 // to keep more precise type from last cast. 4127 if (ptr->is_Proj()) { 4128 Node* allo = ptr->in(0); 4129 if (allo != NULL && allo->is_Allocate()) { 4130 return allo->as_Allocate(); 4131 } 4132 } 4133 // Report failure to match. 4134 return NULL; 4135 } 4136 4137 // Fancy version which also strips off an offset (and reports it to caller). 4138 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 4139 intptr_t& offset) { 4140 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4141 if (base == NULL) return NULL; 4142 return Ideal_allocation(base, phase); 4143 } 4144 4145 // Trace Initialize <- Proj[Parm] <- Allocate 4146 AllocateNode* InitializeNode::allocation() { 4147 Node* rawoop = in(InitializeNode::RawAddress); 4148 if (rawoop->is_Proj()) { 4149 Node* alloc = rawoop->in(0); 4150 if (alloc->is_Allocate()) { 4151 return alloc->as_Allocate(); 4152 } 4153 } 4154 return NULL; 4155 } 4156 4157 // Trace Allocate -> Proj[Parm] -> Initialize 4158 InitializeNode* AllocateNode::initialization() { 4159 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4160 if (rawoop == NULL) return NULL; 4161 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4162 Node* init = rawoop->fast_out(i); 4163 if (init->is_Initialize()) { 4164 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4165 return init->as_Initialize(); 4166 } 4167 } 4168 return NULL; 4169 } 4170 4171 //----------------------------- loop predicates --------------------------- 4172 4173 //------------------------------add_predicate_impl---------------------------- 4174 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 4175 // Too many traps seen? 4176 if (too_many_traps(reason)) { 4177 #ifdef ASSERT 4178 if (TraceLoopPredicate) { 4179 int tc = C->trap_count(reason); 4180 tty->print("too many traps=%s tcount=%d in ", 4181 Deoptimization::trap_reason_name(reason), tc); 4182 method()->print(); // which method has too many predicate traps 4183 tty->cr(); 4184 } 4185 #endif 4186 // We cannot afford to take more traps here, 4187 // do not generate predicate. 4188 return; 4189 } 4190 4191 Node *cont = _gvn.intcon(1); 4192 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 4193 Node *bol = _gvn.transform(new Conv2BNode(opq)); 4194 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4195 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4196 C->add_predicate_opaq(opq); 4197 { 4198 PreserveJVMState pjvms(this); 4199 set_control(iffalse); 4200 inc_sp(nargs); 4201 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4202 } 4203 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4204 set_control(iftrue); 4205 } 4206 4207 //------------------------------add_predicate--------------------------------- 4208 void GraphKit::add_predicate(int nargs) { 4209 if (UseLoopPredicate) { 4210 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 4211 } 4212 if (UseProfiledLoopPredicate) { 4213 add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs); 4214 } 4215 // loop's limit check predicate should be near the loop. 4216 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 4217 } 4218 4219 void GraphKit::sync_kit(IdealKit& ideal) { 4220 set_all_memory(ideal.merged_memory()); 4221 set_i_o(ideal.i_o()); 4222 set_control(ideal.ctrl()); 4223 } 4224 4225 void GraphKit::final_sync(IdealKit& ideal) { 4226 // Final sync IdealKit and graphKit. 4227 sync_kit(ideal); 4228 } 4229 4230 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4231 Node* len = load_array_length(load_String_value(ctrl, str)); 4232 Node* coder = load_String_coder(ctrl, str); 4233 // Divide length by 2 if coder is UTF16 4234 return _gvn.transform(new RShiftINode(len, coder)); 4235 } 4236 4237 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4238 int value_offset = java_lang_String::value_offset_in_bytes(); 4239 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4240 false, NULL, Type::Offset(0)); 4241 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4242 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4243 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4244 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4245 Node* p = basic_plus_adr(str, str, value_offset); 4246 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4247 IN_HEAP | C2_CONTROL_DEPENDENT_LOAD); 4248 // String.value field is known to be @Stable. 4249 if (UseImplicitStableValues) { 4250 load = cast_array_to_stable(load, value_type); 4251 } 4252 return load; 4253 } 4254 4255 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4256 if (!CompactStrings) { 4257 return intcon(java_lang_String::CODER_UTF16); 4258 } 4259 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4260 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4261 false, NULL, Type::Offset(0)); 4262 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4263 int coder_field_idx = C->get_alias_index(coder_field_type); 4264 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4265 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4266 } 4267 4268 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4269 int value_offset = java_lang_String::value_offset_in_bytes(); 4270 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4271 false, NULL, Type::Offset(0)); 4272 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4273 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4274 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP); 4275 } 4276 4277 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4278 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4279 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4280 false, NULL, Type::Offset(0)); 4281 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4282 int coder_field_idx = C->get_alias_index(coder_field_type); 4283 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4284 value, T_BYTE, coder_field_idx, MemNode::unordered); 4285 } 4286 4287 // Capture src and dst memory state with a MergeMemNode 4288 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4289 if (src_type == dst_type) { 4290 // Types are equal, we don't need a MergeMemNode 4291 return memory(src_type); 4292 } 4293 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4294 record_for_igvn(merge); // fold it up later, if possible 4295 int src_idx = C->get_alias_index(src_type); 4296 int dst_idx = C->get_alias_index(dst_type); 4297 merge->set_memory_at(src_idx, memory(src_idx)); 4298 merge->set_memory_at(dst_idx, memory(dst_idx)); 4299 return merge; 4300 } 4301 4302 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4303 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4304 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4305 // If input and output memory types differ, capture both states to preserve 4306 // the dependency between preceding and subsequent loads/stores. 4307 // For example, the following program: 4308 // StoreB 4309 // compress_string 4310 // LoadB 4311 // has this memory graph (use->def): 4312 // LoadB -> compress_string -> CharMem 4313 // ... -> StoreB -> ByteMem 4314 // The intrinsic hides the dependency between LoadB and StoreB, causing 4315 // the load to read from memory not containing the result of the StoreB. 4316 // The correct memory graph should look like this: 4317 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4318 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4319 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4320 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4321 set_memory(res_mem, TypeAryPtr::BYTES); 4322 return str; 4323 } 4324 4325 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4326 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4327 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4328 // Capture src and dst memory (see comment in 'compress_string'). 4329 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4330 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4331 set_memory(_gvn.transform(str), dst_type); 4332 } 4333 4334 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4335 /** 4336 * int i_char = start; 4337 * for (int i_byte = 0; i_byte < count; i_byte++) { 4338 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4339 * } 4340 */ 4341 add_predicate(); 4342 RegionNode* head = new RegionNode(3); 4343 head->init_req(1, control()); 4344 gvn().set_type(head, Type::CONTROL); 4345 record_for_igvn(head); 4346 4347 Node* i_byte = new PhiNode(head, TypeInt::INT); 4348 i_byte->init_req(1, intcon(0)); 4349 gvn().set_type(i_byte, TypeInt::INT); 4350 record_for_igvn(i_byte); 4351 4352 Node* i_char = new PhiNode(head, TypeInt::INT); 4353 i_char->init_req(1, start); 4354 gvn().set_type(i_char, TypeInt::INT); 4355 record_for_igvn(i_char); 4356 4357 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4358 gvn().set_type(mem, Type::MEMORY); 4359 record_for_igvn(mem); 4360 set_control(head); 4361 set_memory(mem, TypeAryPtr::BYTES); 4362 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4363 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4364 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4365 false, false, true /* mismatched */); 4366 4367 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4368 head->init_req(2, IfTrue(iff)); 4369 mem->init_req(2, st); 4370 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4371 i_char->init_req(2, AddI(i_char, intcon(2))); 4372 4373 set_control(IfFalse(iff)); 4374 set_memory(st, TypeAryPtr::BYTES); 4375 } 4376 4377 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4378 if (!field->is_constant()) { 4379 return NULL; // Field not marked as constant. 4380 } 4381 ciInstance* holder = NULL; 4382 if (!field->is_static()) { 4383 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4384 if (const_oop != NULL && const_oop->is_instance()) { 4385 holder = const_oop->as_instance(); 4386 } 4387 } 4388 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4389 /*is_unsigned_load=*/false); 4390 if (con_type != NULL) { 4391 Node* con = makecon(con_type); 4392 if (field->layout_type() == T_VALUETYPE) { 4393 // Load value type from constant oop 4394 assert(!con_type->maybe_null(), "should never be null"); 4395 if (field->type()->as_value_klass()->is_scalarizable()) { 4396 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass()); 4397 } 4398 } 4399 return con; 4400 } 4401 return NULL; 4402 } 4403 4404 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4405 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4406 // assumption of CCP analysis. 4407 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4408 } 4409 4410 //---------------------------load_mirror_from_klass---------------------------- 4411 // Given a klass oop, load its java mirror (a java.lang.Class oop). 4412 Node* GraphKit::load_mirror_from_klass(Node* klass) { 4413 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 4414 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4415 // mirror = ((OopHandle)mirror)->resolve(); 4416 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 4417 }