1 /* 2 * Copyright (c) 2001, 2016, 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 "compiler/compileLog.hpp" 27 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc/g1/heapRegion.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/cardTableModRefBS.hpp" 31 #include "gc/shared/collectedHeap.hpp" 32 #include "gc/shenandoah/brooksPointer.hpp" 33 #include "gc/shenandoah/shenandoahConnectionMatrix.hpp" 34 #include "gc/shenandoah/shenandoahHeap.hpp" 35 #include "gc/shenandoah/shenandoahHeapRegion.hpp" 36 #include "memory/resourceArea.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/castnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/graphKit.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/intrinsicnode.hpp" 43 #include "opto/locknode.hpp" 44 #include "opto/machnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/rootnode.hpp" 48 #include "opto/runtime.hpp" 49 #include "opto/shenandoahSupport.hpp" 50 #include "runtime/deoptimization.hpp" 51 #include "runtime/sharedRuntime.hpp" 52 53 //----------------------------GraphKit----------------------------------------- 54 // Main utility constructor. 55 GraphKit::GraphKit(JVMState* jvms) 56 : Phase(Phase::Parser), 57 _env(C->env()), 58 _gvn(*C->initial_gvn()) 59 { 60 _exceptions = jvms->map()->next_exception(); 61 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 62 set_jvms(jvms); 63 } 64 65 // Private constructor for parser. 66 GraphKit::GraphKit() 67 : Phase(Phase::Parser), 68 _env(C->env()), 69 _gvn(*C->initial_gvn()) 70 { 71 _exceptions = NULL; 72 set_map(NULL); 73 debug_only(_sp = -99); 74 debug_only(set_bci(-99)); 75 } 76 77 78 79 //---------------------------clean_stack--------------------------------------- 80 // Clear away rubbish from the stack area of the JVM state. 81 // This destroys any arguments that may be waiting on the stack. 82 void GraphKit::clean_stack(int from_sp) { 83 SafePointNode* map = this->map(); 84 JVMState* jvms = this->jvms(); 85 int stk_size = jvms->stk_size(); 86 int stkoff = jvms->stkoff(); 87 Node* top = this->top(); 88 for (int i = from_sp; i < stk_size; i++) { 89 if (map->in(stkoff + i) != top) { 90 map->set_req(stkoff + i, top); 91 } 92 } 93 } 94 95 96 //--------------------------------sync_jvms----------------------------------- 97 // Make sure our current jvms agrees with our parse state. 98 JVMState* GraphKit::sync_jvms() const { 99 JVMState* jvms = this->jvms(); 100 jvms->set_bci(bci()); // Record the new bci in the JVMState 101 jvms->set_sp(sp()); // Record the new sp in the JVMState 102 assert(jvms_in_sync(), "jvms is now in sync"); 103 return jvms; 104 } 105 106 //--------------------------------sync_jvms_for_reexecute--------------------- 107 // Make sure our current jvms agrees with our parse state. This version 108 // uses the reexecute_sp for reexecuting bytecodes. 109 JVMState* GraphKit::sync_jvms_for_reexecute() { 110 JVMState* jvms = this->jvms(); 111 jvms->set_bci(bci()); // Record the new bci in the JVMState 112 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 113 return jvms; 114 } 115 116 #ifdef ASSERT 117 bool GraphKit::jvms_in_sync() const { 118 Parse* parse = is_Parse(); 119 if (parse == NULL) { 120 if (bci() != jvms()->bci()) return false; 121 if (sp() != (int)jvms()->sp()) return false; 122 return true; 123 } 124 if (jvms()->method() != parse->method()) return false; 125 if (jvms()->bci() != parse->bci()) return false; 126 int jvms_sp = jvms()->sp(); 127 if (jvms_sp != parse->sp()) return false; 128 int jvms_depth = jvms()->depth(); 129 if (jvms_depth != parse->depth()) return false; 130 return true; 131 } 132 133 // Local helper checks for special internal merge points 134 // used to accumulate and merge exception states. 135 // They are marked by the region's in(0) edge being the map itself. 136 // Such merge points must never "escape" into the parser at large, 137 // until they have been handed to gvn.transform. 138 static bool is_hidden_merge(Node* reg) { 139 if (reg == NULL) return false; 140 if (reg->is_Phi()) { 141 reg = reg->in(0); 142 if (reg == NULL) return false; 143 } 144 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 145 } 146 147 void GraphKit::verify_map() const { 148 if (map() == NULL) return; // null map is OK 149 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 150 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 151 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 152 } 153 154 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 155 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 156 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 157 } 158 #endif 159 160 //---------------------------stop_and_kill_map--------------------------------- 161 // Set _map to NULL, signalling a stop to further bytecode execution. 162 // First smash the current map's control to a constant, to mark it dead. 163 void GraphKit::stop_and_kill_map() { 164 SafePointNode* dead_map = stop(); 165 if (dead_map != NULL) { 166 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 167 assert(dead_map->is_killed(), "must be so marked"); 168 } 169 } 170 171 172 //--------------------------------stopped-------------------------------------- 173 // Tell if _map is NULL, or control is top. 174 bool GraphKit::stopped() { 175 if (map() == NULL) return true; 176 else if (control() == top()) return true; 177 else return false; 178 } 179 180 181 //-----------------------------has_ex_handler---------------------------------- 182 // Tell if this method or any caller method has exception handlers. 183 bool GraphKit::has_ex_handler() { 184 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 185 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 186 return true; 187 } 188 } 189 return false; 190 } 191 192 //------------------------------save_ex_oop------------------------------------ 193 // Save an exception without blowing stack contents or other JVM state. 194 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 195 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 196 ex_map->add_req(ex_oop); 197 debug_only(verify_exception_state(ex_map)); 198 } 199 200 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 201 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 202 Node* ex_oop = ex_map->in(ex_map->req()-1); 203 if (clear_it) ex_map->del_req(ex_map->req()-1); 204 return ex_oop; 205 } 206 207 //-----------------------------saved_ex_oop------------------------------------ 208 // Recover a saved exception from its map. 209 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, false); 211 } 212 213 //--------------------------clear_saved_ex_oop--------------------------------- 214 // Erase a previously saved exception from its map. 215 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 216 return common_saved_ex_oop(ex_map, true); 217 } 218 219 #ifdef ASSERT 220 //---------------------------has_saved_ex_oop---------------------------------- 221 // Erase a previously saved exception from its map. 222 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 223 return ex_map->req() == ex_map->jvms()->endoff()+1; 224 } 225 #endif 226 227 //-------------------------make_exception_state-------------------------------- 228 // Turn the current JVM state into an exception state, appending the ex_oop. 229 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 230 sync_jvms(); 231 SafePointNode* ex_map = stop(); // do not manipulate this map any more 232 set_saved_ex_oop(ex_map, ex_oop); 233 return ex_map; 234 } 235 236 237 //--------------------------add_exception_state-------------------------------- 238 // Add an exception to my list of exceptions. 239 void GraphKit::add_exception_state(SafePointNode* ex_map) { 240 if (ex_map == NULL || ex_map->control() == top()) { 241 return; 242 } 243 #ifdef ASSERT 244 verify_exception_state(ex_map); 245 if (has_exceptions()) { 246 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 247 } 248 #endif 249 250 // If there is already an exception of exactly this type, merge with it. 251 // In particular, null-checks and other low-level exceptions common up here. 252 Node* ex_oop = saved_ex_oop(ex_map); 253 const Type* ex_type = _gvn.type(ex_oop); 254 if (ex_oop == top()) { 255 // No action needed. 256 return; 257 } 258 assert(ex_type->isa_instptr(), "exception must be an instance"); 259 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 260 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 261 // We check sp also because call bytecodes can generate exceptions 262 // both before and after arguments are popped! 263 if (ex_type2 == ex_type 264 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 265 combine_exception_states(ex_map, e2); 266 return; 267 } 268 } 269 270 // No pre-existing exception of the same type. Chain it on the list. 271 push_exception_state(ex_map); 272 } 273 274 //-----------------------add_exception_states_from----------------------------- 275 void GraphKit::add_exception_states_from(JVMState* jvms) { 276 SafePointNode* ex_map = jvms->map()->next_exception(); 277 if (ex_map != NULL) { 278 jvms->map()->set_next_exception(NULL); 279 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 280 next_map = ex_map->next_exception(); 281 ex_map->set_next_exception(NULL); 282 add_exception_state(ex_map); 283 } 284 } 285 } 286 287 //-----------------------transfer_exceptions_into_jvms------------------------- 288 JVMState* GraphKit::transfer_exceptions_into_jvms() { 289 if (map() == NULL) { 290 // We need a JVMS to carry the exceptions, but the map has gone away. 291 // Create a scratch JVMS, cloned from any of the exception states... 292 if (has_exceptions()) { 293 _map = _exceptions; 294 _map = clone_map(); 295 _map->set_next_exception(NULL); 296 clear_saved_ex_oop(_map); 297 debug_only(verify_map()); 298 } else { 299 // ...or created from scratch 300 JVMState* jvms = new (C) JVMState(_method, NULL); 301 jvms->set_bci(_bci); 302 jvms->set_sp(_sp); 303 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 304 set_jvms(jvms); 305 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 306 set_all_memory(top()); 307 while (map()->req() < jvms->endoff()) map()->add_req(top()); 308 } 309 // (This is a kludge, in case you didn't notice.) 310 set_control(top()); 311 } 312 JVMState* jvms = sync_jvms(); 313 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 314 jvms->map()->set_next_exception(_exceptions); 315 _exceptions = NULL; // done with this set of exceptions 316 return jvms; 317 } 318 319 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 320 assert(is_hidden_merge(dstphi), "must be a special merge node"); 321 assert(is_hidden_merge(srcphi), "must be a special merge node"); 322 uint limit = srcphi->req(); 323 for (uint i = PhiNode::Input; i < limit; i++) { 324 dstphi->add_req(srcphi->in(i)); 325 } 326 } 327 static inline void add_one_req(Node* dstphi, Node* src) { 328 assert(is_hidden_merge(dstphi), "must be a special merge node"); 329 assert(!is_hidden_merge(src), "must not be a special merge node"); 330 dstphi->add_req(src); 331 } 332 333 //-----------------------combine_exception_states------------------------------ 334 // This helper function combines exception states by building phis on a 335 // specially marked state-merging region. These regions and phis are 336 // untransformed, and can build up gradually. The region is marked by 337 // having a control input of its exception map, rather than NULL. Such 338 // regions do not appear except in this function, and in use_exception_state. 339 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 340 if (failing()) return; // dying anyway... 341 JVMState* ex_jvms = ex_map->_jvms; 342 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 343 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 344 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 345 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 346 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 347 assert(ex_map->req() == phi_map->req(), "matching maps"); 348 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 349 Node* hidden_merge_mark = root(); 350 Node* region = phi_map->control(); 351 MergeMemNode* phi_mem = phi_map->merged_memory(); 352 MergeMemNode* ex_mem = ex_map->merged_memory(); 353 if (region->in(0) != hidden_merge_mark) { 354 // The control input is not (yet) a specially-marked region in phi_map. 355 // Make it so, and build some phis. 356 region = new RegionNode(2); 357 _gvn.set_type(region, Type::CONTROL); 358 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 359 region->init_req(1, phi_map->control()); 360 phi_map->set_control(region); 361 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 362 record_for_igvn(io_phi); 363 _gvn.set_type(io_phi, Type::ABIO); 364 phi_map->set_i_o(io_phi); 365 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 366 Node* m = mms.memory(); 367 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 368 record_for_igvn(m_phi); 369 _gvn.set_type(m_phi, Type::MEMORY); 370 mms.set_memory(m_phi); 371 } 372 } 373 374 // Either or both of phi_map and ex_map might already be converted into phis. 375 Node* ex_control = ex_map->control(); 376 // if there is special marking on ex_map also, we add multiple edges from src 377 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 378 // how wide was the destination phi_map, originally? 379 uint orig_width = region->req(); 380 381 if (add_multiple) { 382 add_n_reqs(region, ex_control); 383 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 384 } else { 385 // ex_map has no merges, so we just add single edges everywhere 386 add_one_req(region, ex_control); 387 add_one_req(phi_map->i_o(), ex_map->i_o()); 388 } 389 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 390 if (mms.is_empty()) { 391 // get a copy of the base memory, and patch some inputs into it 392 const TypePtr* adr_type = mms.adr_type(C); 393 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 394 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 395 mms.set_memory(phi); 396 // Prepare to append interesting stuff onto the newly sliced phi: 397 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 398 } 399 // Append stuff from ex_map: 400 if (add_multiple) { 401 add_n_reqs(mms.memory(), mms.memory2()); 402 } else { 403 add_one_req(mms.memory(), mms.memory2()); 404 } 405 } 406 uint limit = ex_map->req(); 407 for (uint i = TypeFunc::Parms; i < limit; i++) { 408 // Skip everything in the JVMS after tos. (The ex_oop follows.) 409 if (i == tos) i = ex_jvms->monoff(); 410 Node* src = ex_map->in(i); 411 Node* dst = phi_map->in(i); 412 if (src != dst) { 413 PhiNode* phi; 414 if (dst->in(0) != region) { 415 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 416 record_for_igvn(phi); 417 _gvn.set_type(phi, phi->type()); 418 phi_map->set_req(i, dst); 419 // Prepare to append interesting stuff onto the new phi: 420 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 421 } else { 422 assert(dst->is_Phi(), "nobody else uses a hidden region"); 423 phi = dst->as_Phi(); 424 } 425 if (add_multiple && src->in(0) == ex_control) { 426 // Both are phis. 427 add_n_reqs(dst, src); 428 } else { 429 while (dst->req() < region->req()) add_one_req(dst, src); 430 } 431 const Type* srctype = _gvn.type(src); 432 if (phi->type() != srctype) { 433 const Type* dsttype = phi->type()->meet_speculative(srctype); 434 if (phi->type() != dsttype) { 435 phi->set_type(dsttype); 436 _gvn.set_type(phi, dsttype); 437 } 438 } 439 } 440 } 441 phi_map->merge_replaced_nodes_with(ex_map); 442 } 443 444 //--------------------------use_exception_state-------------------------------- 445 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 446 if (failing()) { stop(); return top(); } 447 Node* region = phi_map->control(); 448 Node* hidden_merge_mark = root(); 449 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 450 Node* ex_oop = clear_saved_ex_oop(phi_map); 451 if (region->in(0) == hidden_merge_mark) { 452 // Special marking for internal ex-states. Process the phis now. 453 region->set_req(0, region); // now it's an ordinary region 454 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 455 // Note: Setting the jvms also sets the bci and sp. 456 set_control(_gvn.transform(region)); 457 uint tos = jvms()->stkoff() + sp(); 458 for (uint i = 1; i < tos; i++) { 459 Node* x = phi_map->in(i); 460 if (x->in(0) == region) { 461 assert(x->is_Phi(), "expected a special phi"); 462 phi_map->set_req(i, _gvn.transform(x)); 463 } 464 } 465 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 466 Node* x = mms.memory(); 467 if (x->in(0) == region) { 468 assert(x->is_Phi(), "nobody else uses a hidden region"); 469 mms.set_memory(_gvn.transform(x)); 470 } 471 } 472 if (ex_oop->in(0) == region) { 473 assert(ex_oop->is_Phi(), "expected a special phi"); 474 ex_oop = _gvn.transform(ex_oop); 475 } 476 } else { 477 set_jvms(phi_map->jvms()); 478 } 479 480 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 481 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 482 return ex_oop; 483 } 484 485 //---------------------------------java_bc------------------------------------- 486 Bytecodes::Code GraphKit::java_bc() const { 487 ciMethod* method = this->method(); 488 int bci = this->bci(); 489 if (method != NULL && bci != InvocationEntryBci) 490 return method->java_code_at_bci(bci); 491 else 492 return Bytecodes::_illegal; 493 } 494 495 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 496 bool must_throw) { 497 // if the exception capability is set, then we will generate code 498 // to check the JavaThread.should_post_on_exceptions flag to see 499 // if we actually need to report exception events (for this 500 // thread). If we don't need to report exception events, we will 501 // take the normal fast path provided by add_exception_events. If 502 // exception event reporting is enabled for this thread, we will 503 // take the uncommon_trap in the BuildCutout below. 504 505 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 506 Node* jthread = _gvn.transform(new ThreadLocalNode()); 507 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 508 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 509 510 // Test the should_post_on_exceptions_flag vs. 0 511 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 512 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 513 514 // Branch to slow_path if should_post_on_exceptions_flag was true 515 { BuildCutout unless(this, tst, PROB_MAX); 516 // Do not try anything fancy if we're notifying the VM on every throw. 517 // Cf. case Bytecodes::_athrow in parse2.cpp. 518 uncommon_trap(reason, Deoptimization::Action_none, 519 (ciKlass*)NULL, (char*)NULL, must_throw); 520 } 521 522 } 523 524 //------------------------------builtin_throw---------------------------------- 525 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 526 bool must_throw = true; 527 528 if (env()->jvmti_can_post_on_exceptions()) { 529 // check if we must post exception events, take uncommon trap if so 530 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 531 // here if should_post_on_exceptions is false 532 // continue on with the normal codegen 533 } 534 535 // If this particular condition has not yet happened at this 536 // bytecode, then use the uncommon trap mechanism, and allow for 537 // a future recompilation if several traps occur here. 538 // If the throw is hot, try to use a more complicated inline mechanism 539 // which keeps execution inside the compiled code. 540 bool treat_throw_as_hot = false; 541 ciMethodData* md = method()->method_data(); 542 543 if (ProfileTraps) { 544 if (too_many_traps(reason)) { 545 treat_throw_as_hot = true; 546 } 547 // (If there is no MDO at all, assume it is early in 548 // execution, and that any deopts are part of the 549 // startup transient, and don't need to be remembered.) 550 551 // Also, if there is a local exception handler, treat all throws 552 // as hot if there has been at least one in this method. 553 if (C->trap_count(reason) != 0 554 && method()->method_data()->trap_count(reason) != 0 555 && has_ex_handler()) { 556 treat_throw_as_hot = true; 557 } 558 } 559 560 // If this throw happens frequently, an uncommon trap might cause 561 // a performance pothole. If there is a local exception handler, 562 // and if this particular bytecode appears to be deoptimizing often, 563 // let us handle the throw inline, with a preconstructed instance. 564 // Note: If the deopt count has blown up, the uncommon trap 565 // runtime is going to flush this nmethod, not matter what. 566 if (treat_throw_as_hot 567 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 568 // If the throw is local, we use a pre-existing instance and 569 // punt on the backtrace. This would lead to a missing backtrace 570 // (a repeat of 4292742) if the backtrace object is ever asked 571 // for its backtrace. 572 // Fixing this remaining case of 4292742 requires some flavor of 573 // escape analysis. Leave that for the future. 574 ciInstance* ex_obj = NULL; 575 switch (reason) { 576 case Deoptimization::Reason_null_check: 577 ex_obj = env()->NullPointerException_instance(); 578 break; 579 case Deoptimization::Reason_div0_check: 580 ex_obj = env()->ArithmeticException_instance(); 581 break; 582 case Deoptimization::Reason_range_check: 583 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 584 break; 585 case Deoptimization::Reason_class_check: 586 if (java_bc() == Bytecodes::_aastore) { 587 ex_obj = env()->ArrayStoreException_instance(); 588 } else { 589 ex_obj = env()->ClassCastException_instance(); 590 } 591 break; 592 } 593 if (failing()) { stop(); return; } // exception allocation might fail 594 if (ex_obj != NULL) { 595 // Cheat with a preallocated exception object. 596 if (C->log() != NULL) 597 C->log()->elem("hot_throw preallocated='1' reason='%s'", 598 Deoptimization::trap_reason_name(reason)); 599 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 600 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 601 602 // Clear the detail message of the preallocated exception object. 603 // Weblogic sometimes mutates the detail message of exceptions 604 // using reflection. 605 int offset = java_lang_Throwable::get_detailMessage_offset(); 606 const TypePtr* adr_typ = ex_con->add_offset(offset); 607 608 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 609 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 610 // Conservatively release stores of object references. 611 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 612 613 add_exception_state(make_exception_state(ex_node)); 614 return; 615 } 616 } 617 618 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 619 // It won't be much cheaper than bailing to the interp., since we'll 620 // have to pass up all the debug-info, and the runtime will have to 621 // create the stack trace. 622 623 // Usual case: Bail to interpreter. 624 // Reserve the right to recompile if we haven't seen anything yet. 625 626 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 627 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 628 if (treat_throw_as_hot 629 && (method()->method_data()->trap_recompiled_at(bci(), m) 630 || C->too_many_traps(reason))) { 631 // We cannot afford to take more traps here. Suffer in the interpreter. 632 if (C->log() != NULL) 633 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 634 Deoptimization::trap_reason_name(reason), 635 C->trap_count(reason)); 636 action = Deoptimization::Action_none; 637 } 638 639 // "must_throw" prunes the JVM state to include only the stack, if there 640 // are no local exception handlers. This should cut down on register 641 // allocation time and code size, by drastically reducing the number 642 // of in-edges on the call to the uncommon trap. 643 644 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 645 } 646 647 648 //----------------------------PreserveJVMState--------------------------------- 649 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 650 debug_only(kit->verify_map()); 651 _kit = kit; 652 _map = kit->map(); // preserve the map 653 _sp = kit->sp(); 654 kit->set_map(clone_map ? kit->clone_map() : NULL); 655 #ifdef ASSERT 656 _bci = kit->bci(); 657 Parse* parser = kit->is_Parse(); 658 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 659 _block = block; 660 #endif 661 } 662 PreserveJVMState::~PreserveJVMState() { 663 GraphKit* kit = _kit; 664 #ifdef ASSERT 665 assert(kit->bci() == _bci, "bci must not shift"); 666 Parse* parser = kit->is_Parse(); 667 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 668 assert(block == _block, "block must not shift"); 669 #endif 670 kit->set_map(_map); 671 kit->set_sp(_sp); 672 } 673 674 675 //-----------------------------BuildCutout------------------------------------- 676 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 677 : PreserveJVMState(kit) 678 { 679 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 680 SafePointNode* outer_map = _map; // preserved map is caller's 681 SafePointNode* inner_map = kit->map(); 682 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 683 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 684 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 685 } 686 BuildCutout::~BuildCutout() { 687 GraphKit* kit = _kit; 688 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 689 } 690 691 //---------------------------PreserveReexecuteState---------------------------- 692 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 693 assert(!kit->stopped(), "must call stopped() before"); 694 _kit = kit; 695 _sp = kit->sp(); 696 _reexecute = kit->jvms()->_reexecute; 697 } 698 PreserveReexecuteState::~PreserveReexecuteState() { 699 if (_kit->stopped()) return; 700 _kit->jvms()->_reexecute = _reexecute; 701 _kit->set_sp(_sp); 702 } 703 704 //------------------------------clone_map-------------------------------------- 705 // Implementation of PreserveJVMState 706 // 707 // Only clone_map(...) here. If this function is only used in the 708 // PreserveJVMState class we may want to get rid of this extra 709 // function eventually and do it all there. 710 711 SafePointNode* GraphKit::clone_map() { 712 if (map() == NULL) return NULL; 713 714 // Clone the memory edge first 715 Node* mem = MergeMemNode::make(map()->memory()); 716 gvn().set_type_bottom(mem); 717 718 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 719 JVMState* jvms = this->jvms(); 720 JVMState* clonejvms = jvms->clone_shallow(C); 721 clonemap->set_memory(mem); 722 clonemap->set_jvms(clonejvms); 723 clonejvms->set_map(clonemap); 724 record_for_igvn(clonemap); 725 gvn().set_type_bottom(clonemap); 726 return clonemap; 727 } 728 729 730 //-----------------------------set_map_clone----------------------------------- 731 void GraphKit::set_map_clone(SafePointNode* m) { 732 _map = m; 733 _map = clone_map(); 734 _map->set_next_exception(NULL); 735 debug_only(verify_map()); 736 } 737 738 739 //----------------------------kill_dead_locals--------------------------------- 740 // Detect any locals which are known to be dead, and force them to top. 741 void GraphKit::kill_dead_locals() { 742 // Consult the liveness information for the locals. If any 743 // of them are unused, then they can be replaced by top(). This 744 // should help register allocation time and cut down on the size 745 // of the deoptimization information. 746 747 // This call is made from many of the bytecode handling 748 // subroutines called from the Big Switch in do_one_bytecode. 749 // Every bytecode which might include a slow path is responsible 750 // for killing its dead locals. The more consistent we 751 // are about killing deads, the fewer useless phis will be 752 // constructed for them at various merge points. 753 754 // bci can be -1 (InvocationEntryBci). We return the entry 755 // liveness for the method. 756 757 if (method() == NULL || method()->code_size() == 0) { 758 // We are building a graph for a call to a native method. 759 // All locals are live. 760 return; 761 } 762 763 ResourceMark rm; 764 765 // Consult the liveness information for the locals. If any 766 // of them are unused, then they can be replaced by top(). This 767 // should help register allocation time and cut down on the size 768 // of the deoptimization information. 769 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 770 771 int len = (int)live_locals.size(); 772 assert(len <= jvms()->loc_size(), "too many live locals"); 773 for (int local = 0; local < len; local++) { 774 if (!live_locals.at(local)) { 775 set_local(local, top()); 776 } 777 } 778 } 779 780 #ifdef ASSERT 781 //-------------------------dead_locals_are_killed------------------------------ 782 // Return true if all dead locals are set to top in the map. 783 // Used to assert "clean" debug info at various points. 784 bool GraphKit::dead_locals_are_killed() { 785 if (method() == NULL || method()->code_size() == 0) { 786 // No locals need to be dead, so all is as it should be. 787 return true; 788 } 789 790 // Make sure somebody called kill_dead_locals upstream. 791 ResourceMark rm; 792 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 793 if (jvms->loc_size() == 0) continue; // no locals to consult 794 SafePointNode* map = jvms->map(); 795 ciMethod* method = jvms->method(); 796 int bci = jvms->bci(); 797 if (jvms == this->jvms()) { 798 bci = this->bci(); // it might not yet be synched 799 } 800 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 801 int len = (int)live_locals.size(); 802 if (!live_locals.is_valid() || len == 0) 803 // This method is trivial, or is poisoned by a breakpoint. 804 return true; 805 assert(len == jvms->loc_size(), "live map consistent with locals map"); 806 for (int local = 0; local < len; local++) { 807 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 808 if (PrintMiscellaneous && (Verbose || WizardMode)) { 809 tty->print_cr("Zombie local %d: ", local); 810 jvms->dump(); 811 } 812 return false; 813 } 814 } 815 } 816 return true; 817 } 818 819 #endif //ASSERT 820 821 // Helper function for enforcing certain bytecodes to reexecute if 822 // deoptimization happens 823 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 824 ciMethod* cur_method = jvms->method(); 825 int cur_bci = jvms->bci(); 826 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 827 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 828 return Interpreter::bytecode_should_reexecute(code) || 829 is_anewarray && code == Bytecodes::_multianewarray; 830 // Reexecute _multianewarray bytecode which was replaced with 831 // sequence of [a]newarray. See Parse::do_multianewarray(). 832 // 833 // Note: interpreter should not have it set since this optimization 834 // is limited by dimensions and guarded by flag so in some cases 835 // multianewarray() runtime calls will be generated and 836 // the bytecode should not be reexecutes (stack will not be reset). 837 } else 838 return false; 839 } 840 841 // Helper function for adding JVMState and debug information to node 842 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 843 // Add the safepoint edges to the call (or other safepoint). 844 845 // Make sure dead locals are set to top. This 846 // should help register allocation time and cut down on the size 847 // of the deoptimization information. 848 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 849 850 // Walk the inline list to fill in the correct set of JVMState's 851 // Also fill in the associated edges for each JVMState. 852 853 // If the bytecode needs to be reexecuted we need to put 854 // the arguments back on the stack. 855 const bool should_reexecute = jvms()->should_reexecute(); 856 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 857 858 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 859 // undefined if the bci is different. This is normal for Parse but it 860 // should not happen for LibraryCallKit because only one bci is processed. 861 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 862 "in LibraryCallKit the reexecute bit should not change"); 863 864 // If we are guaranteed to throw, we can prune everything but the 865 // input to the current bytecode. 866 bool can_prune_locals = false; 867 uint stack_slots_not_pruned = 0; 868 int inputs = 0, depth = 0; 869 if (must_throw) { 870 assert(method() == youngest_jvms->method(), "sanity"); 871 if (compute_stack_effects(inputs, depth)) { 872 can_prune_locals = true; 873 stack_slots_not_pruned = inputs; 874 } 875 } 876 877 if (env()->should_retain_local_variables()) { 878 // At any safepoint, this method can get breakpointed, which would 879 // then require an immediate deoptimization. 880 can_prune_locals = false; // do not prune locals 881 stack_slots_not_pruned = 0; 882 } 883 884 // do not scribble on the input jvms 885 JVMState* out_jvms = youngest_jvms->clone_deep(C); 886 call->set_jvms(out_jvms); // Start jvms list for call node 887 888 // For a known set of bytecodes, the interpreter should reexecute them if 889 // deoptimization happens. We set the reexecute state for them here 890 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 891 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 892 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 893 } 894 895 // Presize the call: 896 DEBUG_ONLY(uint non_debug_edges = call->req()); 897 call->add_req_batch(top(), youngest_jvms->debug_depth()); 898 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 899 900 // Set up edges so that the call looks like this: 901 // Call [state:] ctl io mem fptr retadr 902 // [parms:] parm0 ... parmN 903 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 904 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 905 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 906 // Note that caller debug info precedes callee debug info. 907 908 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 909 uint debug_ptr = call->req(); 910 911 // Loop over the map input edges associated with jvms, add them 912 // to the call node, & reset all offsets to match call node array. 913 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 914 uint debug_end = debug_ptr; 915 uint debug_start = debug_ptr - in_jvms->debug_size(); 916 debug_ptr = debug_start; // back up the ptr 917 918 uint p = debug_start; // walks forward in [debug_start, debug_end) 919 uint j, k, l; 920 SafePointNode* in_map = in_jvms->map(); 921 out_jvms->set_map(call); 922 923 if (can_prune_locals) { 924 assert(in_jvms->method() == out_jvms->method(), "sanity"); 925 // If the current throw can reach an exception handler in this JVMS, 926 // then we must keep everything live that can reach that handler. 927 // As a quick and dirty approximation, we look for any handlers at all. 928 if (in_jvms->method()->has_exception_handlers()) { 929 can_prune_locals = false; 930 } 931 } 932 933 // Add the Locals 934 k = in_jvms->locoff(); 935 l = in_jvms->loc_size(); 936 out_jvms->set_locoff(p); 937 if (!can_prune_locals) { 938 for (j = 0; j < l; j++) 939 call->set_req(p++, in_map->in(k+j)); 940 } else { 941 p += l; // already set to top above by add_req_batch 942 } 943 944 // Add the Expression Stack 945 k = in_jvms->stkoff(); 946 l = in_jvms->sp(); 947 out_jvms->set_stkoff(p); 948 if (!can_prune_locals) { 949 for (j = 0; j < l; j++) 950 call->set_req(p++, in_map->in(k+j)); 951 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 952 // Divide stack into {S0,...,S1}, where S0 is set to top. 953 uint s1 = stack_slots_not_pruned; 954 stack_slots_not_pruned = 0; // for next iteration 955 if (s1 > l) s1 = l; 956 uint s0 = l - s1; 957 p += s0; // skip the tops preinstalled by add_req_batch 958 for (j = s0; j < l; j++) 959 call->set_req(p++, in_map->in(k+j)); 960 } else { 961 p += l; // already set to top above by add_req_batch 962 } 963 964 // Add the Monitors 965 k = in_jvms->monoff(); 966 l = in_jvms->mon_size(); 967 out_jvms->set_monoff(p); 968 for (j = 0; j < l; j++) 969 call->set_req(p++, in_map->in(k+j)); 970 971 // Copy any scalar object fields. 972 k = in_jvms->scloff(); 973 l = in_jvms->scl_size(); 974 out_jvms->set_scloff(p); 975 for (j = 0; j < l; j++) 976 call->set_req(p++, in_map->in(k+j)); 977 978 // Finish the new jvms. 979 out_jvms->set_endoff(p); 980 981 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 982 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 983 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 984 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 985 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 986 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 987 988 // Update the two tail pointers in parallel. 989 out_jvms = out_jvms->caller(); 990 in_jvms = in_jvms->caller(); 991 } 992 993 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 994 995 // Test the correctness of JVMState::debug_xxx accessors: 996 assert(call->jvms()->debug_start() == non_debug_edges, ""); 997 assert(call->jvms()->debug_end() == call->req(), ""); 998 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 999 } 1000 1001 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1002 Bytecodes::Code code = java_bc(); 1003 if (code == Bytecodes::_wide) { 1004 code = method()->java_code_at_bci(bci() + 1); 1005 } 1006 1007 BasicType rtype = T_ILLEGAL; 1008 int rsize = 0; 1009 1010 if (code != Bytecodes::_illegal) { 1011 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1012 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1013 if (rtype < T_CONFLICT) 1014 rsize = type2size[rtype]; 1015 } 1016 1017 switch (code) { 1018 case Bytecodes::_illegal: 1019 return false; 1020 1021 case Bytecodes::_ldc: 1022 case Bytecodes::_ldc_w: 1023 case Bytecodes::_ldc2_w: 1024 inputs = 0; 1025 break; 1026 1027 case Bytecodes::_dup: inputs = 1; break; 1028 case Bytecodes::_dup_x1: inputs = 2; break; 1029 case Bytecodes::_dup_x2: inputs = 3; break; 1030 case Bytecodes::_dup2: inputs = 2; break; 1031 case Bytecodes::_dup2_x1: inputs = 3; break; 1032 case Bytecodes::_dup2_x2: inputs = 4; break; 1033 case Bytecodes::_swap: inputs = 2; break; 1034 case Bytecodes::_arraylength: inputs = 1; break; 1035 1036 case Bytecodes::_getstatic: 1037 case Bytecodes::_putstatic: 1038 case Bytecodes::_getfield: 1039 case Bytecodes::_putfield: 1040 { 1041 bool ignored_will_link; 1042 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1043 int size = field->type()->size(); 1044 bool is_get = (depth >= 0), is_static = (depth & 1); 1045 inputs = (is_static ? 0 : 1); 1046 if (is_get) { 1047 depth = size - inputs; 1048 } else { 1049 inputs += size; // putxxx pops the value from the stack 1050 depth = - inputs; 1051 } 1052 } 1053 break; 1054 1055 case Bytecodes::_invokevirtual: 1056 case Bytecodes::_invokespecial: 1057 case Bytecodes::_invokestatic: 1058 case Bytecodes::_invokedynamic: 1059 case Bytecodes::_invokeinterface: 1060 { 1061 bool ignored_will_link; 1062 ciSignature* declared_signature = NULL; 1063 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1064 assert(declared_signature != NULL, "cannot be null"); 1065 inputs = declared_signature->arg_size_for_bc(code); 1066 int size = declared_signature->return_type()->size(); 1067 depth = size - inputs; 1068 } 1069 break; 1070 1071 case Bytecodes::_multianewarray: 1072 { 1073 ciBytecodeStream iter(method()); 1074 iter.reset_to_bci(bci()); 1075 iter.next(); 1076 inputs = iter.get_dimensions(); 1077 assert(rsize == 1, ""); 1078 depth = rsize - inputs; 1079 } 1080 break; 1081 1082 case Bytecodes::_ireturn: 1083 case Bytecodes::_lreturn: 1084 case Bytecodes::_freturn: 1085 case Bytecodes::_dreturn: 1086 case Bytecodes::_areturn: 1087 assert(rsize == -depth, ""); 1088 inputs = rsize; 1089 break; 1090 1091 case Bytecodes::_jsr: 1092 case Bytecodes::_jsr_w: 1093 inputs = 0; 1094 depth = 1; // S.B. depth=1, not zero 1095 break; 1096 1097 default: 1098 // bytecode produces a typed result 1099 inputs = rsize - depth; 1100 assert(inputs >= 0, ""); 1101 break; 1102 } 1103 1104 #ifdef ASSERT 1105 // spot check 1106 int outputs = depth + inputs; 1107 assert(outputs >= 0, "sanity"); 1108 switch (code) { 1109 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1110 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1111 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1112 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1113 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1114 } 1115 #endif //ASSERT 1116 1117 return true; 1118 } 1119 1120 1121 1122 //------------------------------basic_plus_adr--------------------------------- 1123 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1124 // short-circuit a common case 1125 if (offset == intcon(0)) return ptr; 1126 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1127 } 1128 1129 Node* GraphKit::ConvI2L(Node* offset) { 1130 // short-circuit a common case 1131 jint offset_con = find_int_con(offset, Type::OffsetBot); 1132 if (offset_con != Type::OffsetBot) { 1133 return longcon((jlong) offset_con); 1134 } 1135 return _gvn.transform( new ConvI2LNode(offset)); 1136 } 1137 1138 Node* GraphKit::ConvI2UL(Node* offset) { 1139 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1140 if (offset_con != (juint) Type::OffsetBot) { 1141 return longcon((julong) offset_con); 1142 } 1143 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1144 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1145 return _gvn.transform( new AndLNode(conv, mask) ); 1146 } 1147 1148 Node* GraphKit::ConvL2I(Node* offset) { 1149 // short-circuit a common case 1150 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1151 if (offset_con != (jlong)Type::OffsetBot) { 1152 return intcon((int) offset_con); 1153 } 1154 return _gvn.transform( new ConvL2INode(offset)); 1155 } 1156 1157 //-------------------------load_object_klass----------------------------------- 1158 Node* GraphKit::load_object_klass(Node* obj) { 1159 // Special-case a fresh allocation to avoid building nodes: 1160 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1161 if (akls != NULL) return akls; 1162 if (ShenandoahVerifyReadsToFromSpace) { 1163 obj = shenandoah_read_barrier(obj); 1164 } 1165 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1166 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1167 } 1168 1169 //-------------------------load_array_length----------------------------------- 1170 Node* GraphKit::load_array_length(Node* array) { 1171 // Special-case a fresh allocation to avoid building nodes: 1172 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1173 Node *alen; 1174 if (alloc == NULL) { 1175 if (ShenandoahVerifyReadsToFromSpace) { 1176 array = shenandoah_read_barrier(array); 1177 } 1178 1179 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1180 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1181 } else { 1182 alen = alloc->Ideal_length(); 1183 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1184 if (ccast != alen) { 1185 alen = _gvn.transform(ccast); 1186 } 1187 } 1188 return alen; 1189 } 1190 1191 //------------------------------do_null_check---------------------------------- 1192 // Helper function to do a NULL pointer check. Returned value is 1193 // the incoming address with NULL casted away. You are allowed to use the 1194 // not-null value only if you are control dependent on the test. 1195 #ifndef PRODUCT 1196 extern int explicit_null_checks_inserted, 1197 explicit_null_checks_elided; 1198 #endif 1199 Node* GraphKit::null_check_common(Node* value, BasicType type, 1200 // optional arguments for variations: 1201 bool assert_null, 1202 Node* *null_control, 1203 bool speculative) { 1204 assert(!assert_null || null_control == NULL, "not both at once"); 1205 if (stopped()) return top(); 1206 NOT_PRODUCT(explicit_null_checks_inserted++); 1207 1208 // Construct NULL check 1209 Node *chk = NULL; 1210 switch(type) { 1211 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1212 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1213 case T_ARRAY : // fall through 1214 type = T_OBJECT; // simplify further tests 1215 case T_OBJECT : { 1216 const Type *t = _gvn.type( value ); 1217 1218 const TypeOopPtr* tp = t->isa_oopptr(); 1219 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1220 // Only for do_null_check, not any of its siblings: 1221 && !assert_null && null_control == NULL) { 1222 // Usually, any field access or invocation on an unloaded oop type 1223 // will simply fail to link, since the statically linked class is 1224 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1225 // the static class is loaded but the sharper oop type is not. 1226 // Rather than checking for this obscure case in lots of places, 1227 // we simply observe that a null check on an unloaded class 1228 // will always be followed by a nonsense operation, so we 1229 // can just issue the uncommon trap here. 1230 // Our access to the unloaded class will only be correct 1231 // after it has been loaded and initialized, which requires 1232 // a trip through the interpreter. 1233 #ifndef PRODUCT 1234 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1235 #endif 1236 uncommon_trap(Deoptimization::Reason_unloaded, 1237 Deoptimization::Action_reinterpret, 1238 tp->klass(), "!loaded"); 1239 return top(); 1240 } 1241 1242 if (assert_null) { 1243 // See if the type is contained in NULL_PTR. 1244 // If so, then the value is already null. 1245 if (t->higher_equal(TypePtr::NULL_PTR)) { 1246 NOT_PRODUCT(explicit_null_checks_elided++); 1247 return value; // Elided null assert quickly! 1248 } 1249 } else { 1250 // See if mixing in the NULL pointer changes type. 1251 // If so, then the NULL pointer was not allowed in the original 1252 // type. In other words, "value" was not-null. 1253 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1254 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1255 NOT_PRODUCT(explicit_null_checks_elided++); 1256 return value; // Elided null check quickly! 1257 } 1258 } 1259 chk = new CmpPNode( value, null() ); 1260 break; 1261 } 1262 1263 default: 1264 fatal("unexpected type: %s", type2name(type)); 1265 } 1266 assert(chk != NULL, "sanity check"); 1267 chk = _gvn.transform(chk); 1268 1269 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1270 BoolNode *btst = new BoolNode( chk, btest); 1271 Node *tst = _gvn.transform( btst ); 1272 1273 //----------- 1274 // if peephole optimizations occurred, a prior test existed. 1275 // If a prior test existed, maybe it dominates as we can avoid this test. 1276 if (tst != btst && type == T_OBJECT) { 1277 // At this point we want to scan up the CFG to see if we can 1278 // find an identical test (and so avoid this test altogether). 1279 Node *cfg = control(); 1280 int depth = 0; 1281 while( depth < 16 ) { // Limit search depth for speed 1282 if( cfg->Opcode() == Op_IfTrue && 1283 cfg->in(0)->in(1) == tst ) { 1284 // Found prior test. Use "cast_not_null" to construct an identical 1285 // CastPP (and hence hash to) as already exists for the prior test. 1286 // Return that casted value. 1287 if (assert_null) { 1288 replace_in_map(value, null()); 1289 return null(); // do not issue the redundant test 1290 } 1291 Node *oldcontrol = control(); 1292 set_control(cfg); 1293 Node *res = cast_not_null(value); 1294 set_control(oldcontrol); 1295 NOT_PRODUCT(explicit_null_checks_elided++); 1296 return res; 1297 } 1298 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1299 if (cfg == NULL) break; // Quit at region nodes 1300 depth++; 1301 } 1302 } 1303 1304 //----------- 1305 // Branch to failure if null 1306 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1307 Deoptimization::DeoptReason reason; 1308 if (assert_null) { 1309 reason = Deoptimization::Reason_null_assert; 1310 } else if (type == T_OBJECT) { 1311 reason = Deoptimization::reason_null_check(speculative); 1312 } else { 1313 reason = Deoptimization::Reason_div0_check; 1314 } 1315 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1316 // ciMethodData::has_trap_at will return a conservative -1 if any 1317 // must-be-null assertion has failed. This could cause performance 1318 // problems for a method after its first do_null_assert failure. 1319 // Consider using 'Reason_class_check' instead? 1320 1321 // To cause an implicit null check, we set the not-null probability 1322 // to the maximum (PROB_MAX). For an explicit check the probability 1323 // is set to a smaller value. 1324 if (null_control != NULL || too_many_traps(reason)) { 1325 // probability is less likely 1326 ok_prob = PROB_LIKELY_MAG(3); 1327 } else if (!assert_null && 1328 (ImplicitNullCheckThreshold > 0) && 1329 method() != NULL && 1330 (method()->method_data()->trap_count(reason) 1331 >= (uint)ImplicitNullCheckThreshold)) { 1332 ok_prob = PROB_LIKELY_MAG(3); 1333 } 1334 1335 if (null_control != NULL) { 1336 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1337 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1338 set_control( _gvn.transform( new IfTrueNode(iff))); 1339 #ifndef PRODUCT 1340 if (null_true == top()) { 1341 explicit_null_checks_elided++; 1342 } 1343 #endif 1344 (*null_control) = null_true; 1345 } else { 1346 BuildCutout unless(this, tst, ok_prob); 1347 // Check for optimizer eliding test at parse time 1348 if (stopped()) { 1349 // Failure not possible; do not bother making uncommon trap. 1350 NOT_PRODUCT(explicit_null_checks_elided++); 1351 } else if (assert_null) { 1352 uncommon_trap(reason, 1353 Deoptimization::Action_make_not_entrant, 1354 NULL, "assert_null"); 1355 } else { 1356 replace_in_map(value, zerocon(type)); 1357 builtin_throw(reason); 1358 } 1359 } 1360 1361 // Must throw exception, fall-thru not possible? 1362 if (stopped()) { 1363 return top(); // No result 1364 } 1365 1366 if (assert_null) { 1367 // Cast obj to null on this path. 1368 replace_in_map(value, zerocon(type)); 1369 return zerocon(type); 1370 } 1371 1372 // Cast obj to not-null on this path, if there is no null_control. 1373 // (If there is a null_control, a non-null value may come back to haunt us.) 1374 if (type == T_OBJECT) { 1375 Node* cast = cast_not_null(value, false); 1376 if (null_control == NULL || (*null_control) == top()) 1377 replace_in_map(value, cast); 1378 value = cast; 1379 } 1380 1381 return value; 1382 } 1383 1384 1385 //------------------------------cast_not_null---------------------------------- 1386 // Cast obj to not-null on this path 1387 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1388 const Type *t = _gvn.type(obj); 1389 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1390 // Object is already not-null? 1391 if( t == t_not_null ) return obj; 1392 1393 Node *cast = new CastPPNode(obj,t_not_null); 1394 cast->init_req(0, control()); 1395 cast = _gvn.transform( cast ); 1396 1397 // Scan for instances of 'obj' in the current JVM mapping. 1398 // These instances are known to be not-null after the test. 1399 if (do_replace_in_map) 1400 replace_in_map(obj, cast); 1401 1402 return cast; // Return casted value 1403 } 1404 1405 1406 //--------------------------replace_in_map------------------------------------- 1407 void GraphKit::replace_in_map(Node* old, Node* neww) { 1408 if (old == neww) { 1409 return; 1410 } 1411 1412 map()->replace_edge(old, neww); 1413 1414 // Note: This operation potentially replaces any edge 1415 // on the map. This includes locals, stack, and monitors 1416 // of the current (innermost) JVM state. 1417 1418 // don't let inconsistent types from profiling escape this 1419 // method 1420 1421 const Type* told = _gvn.type(old); 1422 const Type* tnew = _gvn.type(neww); 1423 1424 if (!tnew->higher_equal(told)) { 1425 return; 1426 } 1427 1428 map()->record_replaced_node(old, neww); 1429 } 1430 1431 1432 //============================================================================= 1433 //--------------------------------memory--------------------------------------- 1434 Node* GraphKit::memory(uint alias_idx) { 1435 MergeMemNode* mem = merged_memory(); 1436 Node* p = mem->memory_at(alias_idx); 1437 _gvn.set_type(p, Type::MEMORY); // must be mapped 1438 return p; 1439 } 1440 1441 //-----------------------------reset_memory------------------------------------ 1442 Node* GraphKit::reset_memory() { 1443 Node* mem = map()->memory(); 1444 // do not use this node for any more parsing! 1445 debug_only( map()->set_memory((Node*)NULL) ); 1446 return _gvn.transform( mem ); 1447 } 1448 1449 //------------------------------set_all_memory--------------------------------- 1450 void GraphKit::set_all_memory(Node* newmem) { 1451 Node* mergemem = MergeMemNode::make(newmem); 1452 gvn().set_type_bottom(mergemem); 1453 map()->set_memory(mergemem); 1454 } 1455 1456 //------------------------------set_all_memory_call---------------------------- 1457 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1458 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1459 set_all_memory(newmem); 1460 } 1461 1462 //============================================================================= 1463 // 1464 // parser factory methods for MemNodes 1465 // 1466 // These are layered on top of the factory methods in LoadNode and StoreNode, 1467 // and integrate with the parser's memory state and _gvn engine. 1468 // 1469 1470 // factory methods in "int adr_idx" 1471 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1472 int adr_idx, 1473 MemNode::MemOrd mo, 1474 LoadNode::ControlDependency control_dependency, 1475 bool require_atomic_access, 1476 bool unaligned, 1477 bool mismatched) { 1478 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1479 const TypePtr* adr_type = NULL; // debug-mode-only argument 1480 debug_only(adr_type = C->get_adr_type(adr_idx)); 1481 Node* mem = memory(adr_idx); 1482 Node* ld; 1483 if (require_atomic_access && bt == T_LONG) { 1484 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1485 } else if (require_atomic_access && bt == T_DOUBLE) { 1486 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1487 } else { 1488 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1489 } 1490 ld = _gvn.transform(ld); 1491 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1492 // Improve graph before escape analysis and boxing elimination. 1493 record_for_igvn(ld); 1494 } 1495 return ld; 1496 } 1497 1498 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1499 int adr_idx, 1500 MemNode::MemOrd mo, 1501 bool require_atomic_access, 1502 bool unaligned, 1503 bool mismatched) { 1504 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1505 const TypePtr* adr_type = NULL; 1506 debug_only(adr_type = C->get_adr_type(adr_idx)); 1507 Node *mem = memory(adr_idx); 1508 Node* st; 1509 if (require_atomic_access && bt == T_LONG) { 1510 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1511 } else if (require_atomic_access && bt == T_DOUBLE) { 1512 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1513 } else { 1514 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1515 } 1516 if (unaligned) { 1517 st->as_Store()->set_unaligned_access(); 1518 } 1519 if (mismatched) { 1520 st->as_Store()->set_mismatched_access(); 1521 } 1522 st = _gvn.transform(st); 1523 set_memory(st, adr_idx); 1524 // Back-to-back stores can only remove intermediate store with DU info 1525 // so push on worklist for optimizer. 1526 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1527 record_for_igvn(st); 1528 1529 return st; 1530 } 1531 1532 1533 Node* GraphKit::pre_barrier(bool do_load, 1534 Node* ctl, 1535 Node* obj, 1536 Node* adr, 1537 uint adr_idx, 1538 Node* val, 1539 const TypeOopPtr* val_type, 1540 Node* pre_val, 1541 BasicType bt) { 1542 1543 BarrierSet* bs = Universe::heap()->barrier_set(); 1544 set_control(ctl); 1545 switch (bs->kind()) { 1546 case BarrierSet::G1SATBCTLogging: 1547 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1548 return val; 1549 case BarrierSet::ShenandoahBarrierSet: 1550 return shenandoah_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1551 break; 1552 1553 case BarrierSet::CardTableForRS: 1554 case BarrierSet::CardTableExtension: 1555 case BarrierSet::ModRef: 1556 break; 1557 1558 default : 1559 ShouldNotReachHere(); 1560 1561 } 1562 return val; 1563 } 1564 1565 bool GraphKit::can_move_pre_barrier() const { 1566 BarrierSet* bs = Universe::heap()->barrier_set(); 1567 switch (bs->kind()) { 1568 case BarrierSet::G1SATBCTLogging: 1569 case BarrierSet::ShenandoahBarrierSet: 1570 return true; // Can move it if no safepoint 1571 1572 case BarrierSet::CardTableForRS: 1573 case BarrierSet::CardTableExtension: 1574 case BarrierSet::ModRef: 1575 return true; // There is no pre-barrier 1576 1577 default : 1578 ShouldNotReachHere(); 1579 } 1580 return false; 1581 } 1582 1583 void GraphKit::post_barrier(Node* ctl, 1584 Node* store, 1585 Node* obj, 1586 Node* adr, 1587 uint adr_idx, 1588 Node* val, 1589 BasicType bt, 1590 bool use_precise) { 1591 BarrierSet* bs = Universe::heap()->barrier_set(); 1592 set_control(ctl); 1593 switch (bs->kind()) { 1594 case BarrierSet::G1SATBCTLogging: 1595 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1596 break; 1597 1598 case BarrierSet::CardTableForRS: 1599 case BarrierSet::CardTableExtension: 1600 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1601 break; 1602 1603 case BarrierSet::ModRef: 1604 case BarrierSet::ShenandoahBarrierSet: 1605 break; 1606 1607 default : 1608 ShouldNotReachHere(); 1609 1610 } 1611 } 1612 1613 Node* GraphKit::store_oop(Node* ctl, 1614 Node* obj, 1615 Node* adr, 1616 const TypePtr* adr_type, 1617 Node* val, 1618 const TypeOopPtr* val_type, 1619 BasicType bt, 1620 bool use_precise, 1621 MemNode::MemOrd mo, 1622 bool mismatched) { 1623 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1624 // could be delayed during Parse (for example, in adjust_map_after_if()). 1625 // Execute transformation here to avoid barrier generation in such case. 1626 if (_gvn.type(val) == TypePtr::NULL_PTR) 1627 val = _gvn.makecon(TypePtr::NULL_PTR); 1628 1629 set_control(ctl); 1630 if (stopped()) return top(); // Dead path ? 1631 1632 assert(bt == T_OBJECT, "sanity"); 1633 assert(val != NULL, "not dead path"); 1634 uint adr_idx = C->get_alias_index(adr_type); 1635 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1636 1637 val = pre_barrier(true /* do_load */, 1638 control(), obj, adr, adr_idx, val, val_type, 1639 NULL /* pre_val */, 1640 bt); 1641 1642 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1643 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1644 return store; 1645 } 1646 1647 // Could be an array or object we don't know at compile time (unsafe ref.) 1648 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1649 Node* obj, // containing obj 1650 Node* adr, // actual adress to store val at 1651 const TypePtr* adr_type, 1652 Node* val, 1653 BasicType bt, 1654 MemNode::MemOrd mo, 1655 bool mismatched) { 1656 Compile::AliasType* at = C->alias_type(adr_type); 1657 const TypeOopPtr* val_type = NULL; 1658 if (adr_type->isa_instptr()) { 1659 if (at->field() != NULL) { 1660 // known field. This code is a copy of the do_put_xxx logic. 1661 ciField* field = at->field(); 1662 if (!field->type()->is_loaded()) { 1663 val_type = TypeInstPtr::BOTTOM; 1664 } else { 1665 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1666 } 1667 } 1668 } else if (adr_type->isa_aryptr()) { 1669 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1670 } 1671 if (val_type == NULL) { 1672 val_type = TypeInstPtr::BOTTOM; 1673 } 1674 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1675 } 1676 1677 1678 //-------------------------array_element_address------------------------- 1679 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1680 const TypeInt* sizetype, Node* ctrl) { 1681 uint shift = exact_log2(type2aelembytes(elembt)); 1682 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1683 1684 // short-circuit a common case (saves lots of confusing waste motion) 1685 jint idx_con = find_int_con(idx, -1); 1686 if (idx_con >= 0) { 1687 intptr_t offset = header + ((intptr_t)idx_con << shift); 1688 return basic_plus_adr(ary, offset); 1689 } 1690 1691 // must be correct type for alignment purposes 1692 Node* base = basic_plus_adr(ary, header); 1693 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1694 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1695 return basic_plus_adr(ary, base, scale); 1696 } 1697 1698 //-------------------------load_array_element------------------------- 1699 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1700 const Type* elemtype = arytype->elem(); 1701 BasicType elembt = elemtype->array_element_basic_type(); 1702 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1703 if (elembt == T_NARROWOOP) { 1704 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1705 } 1706 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1707 return ld; 1708 } 1709 1710 //-------------------------set_arguments_for_java_call------------------------- 1711 // Arguments (pre-popped from the stack) are taken from the JVMS. 1712 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1713 // Add the call arguments: 1714 uint nargs = call->method()->arg_size(); 1715 for (uint i = 0; i < nargs; i++) { 1716 Node* arg = argument(i); 1717 if (ShenandoahVerifyReadsToFromSpace && call->is_CallDynamicJava() && i == 0) { 1718 arg = shenandoah_read_barrier(arg); 1719 } 1720 call->init_req(i + TypeFunc::Parms, arg); 1721 } 1722 } 1723 1724 //---------------------------set_edges_for_java_call--------------------------- 1725 // Connect a newly created call into the current JVMS. 1726 // A return value node (if any) is returned from set_edges_for_java_call. 1727 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1728 1729 // Add the predefined inputs: 1730 call->init_req( TypeFunc::Control, control() ); 1731 call->init_req( TypeFunc::I_O , i_o() ); 1732 call->init_req( TypeFunc::Memory , reset_memory() ); 1733 call->init_req( TypeFunc::FramePtr, frameptr() ); 1734 call->init_req( TypeFunc::ReturnAdr, top() ); 1735 1736 add_safepoint_edges(call, must_throw); 1737 1738 Node* xcall = _gvn.transform(call); 1739 1740 if (xcall == top()) { 1741 set_control(top()); 1742 return; 1743 } 1744 assert(xcall == call, "call identity is stable"); 1745 1746 // Re-use the current map to produce the result. 1747 1748 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1749 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1750 set_all_memory_call(xcall, separate_io_proj); 1751 1752 //return xcall; // no need, caller already has it 1753 } 1754 1755 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1756 if (stopped()) return top(); // maybe the call folded up? 1757 1758 // Capture the return value, if any. 1759 Node* ret; 1760 if (call->method() == NULL || 1761 call->method()->return_type()->basic_type() == T_VOID) 1762 ret = top(); 1763 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1764 1765 // Note: Since any out-of-line call can produce an exception, 1766 // we always insert an I_O projection from the call into the result. 1767 1768 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1769 1770 if (separate_io_proj) { 1771 // The caller requested separate projections be used by the fall 1772 // through and exceptional paths, so replace the projections for 1773 // the fall through path. 1774 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1775 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1776 } 1777 return ret; 1778 } 1779 1780 //--------------------set_predefined_input_for_runtime_call-------------------- 1781 // Reading and setting the memory state is way conservative here. 1782 // The real problem is that I am not doing real Type analysis on memory, 1783 // so I cannot distinguish card mark stores from other stores. Across a GC 1784 // point the Store Barrier and the card mark memory has to agree. I cannot 1785 // have a card mark store and its barrier split across the GC point from 1786 // either above or below. Here I get that to happen by reading ALL of memory. 1787 // A better answer would be to separate out card marks from other memory. 1788 // For now, return the input memory state, so that it can be reused 1789 // after the call, if this call has restricted memory effects. 1790 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1791 // Set fixed predefined input arguments 1792 Node* memory = reset_memory(); 1793 call->init_req( TypeFunc::Control, control() ); 1794 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1795 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1796 call->init_req( TypeFunc::FramePtr, frameptr() ); 1797 call->init_req( TypeFunc::ReturnAdr, top() ); 1798 return memory; 1799 } 1800 1801 //-------------------set_predefined_output_for_runtime_call-------------------- 1802 // Set control and memory (not i_o) from the call. 1803 // If keep_mem is not NULL, use it for the output state, 1804 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1805 // If hook_mem is NULL, this call produces no memory effects at all. 1806 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1807 // then only that memory slice is taken from the call. 1808 // In the last case, we must put an appropriate memory barrier before 1809 // the call, so as to create the correct anti-dependencies on loads 1810 // preceding the call. 1811 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1812 Node* keep_mem, 1813 const TypePtr* hook_mem) { 1814 // no i/o 1815 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1816 if (keep_mem) { 1817 // First clone the existing memory state 1818 set_all_memory(keep_mem); 1819 if (hook_mem != NULL) { 1820 // Make memory for the call 1821 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1822 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1823 // We also use hook_mem to extract specific effects from arraycopy stubs. 1824 set_memory(mem, hook_mem); 1825 } 1826 // ...else the call has NO memory effects. 1827 1828 // Make sure the call advertises its memory effects precisely. 1829 // This lets us build accurate anti-dependences in gcm.cpp. 1830 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1831 "call node must be constructed correctly"); 1832 } else { 1833 assert(hook_mem == NULL, ""); 1834 // This is not a "slow path" call; all memory comes from the call. 1835 set_all_memory_call(call); 1836 } 1837 } 1838 1839 1840 // Replace the call with the current state of the kit. 1841 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1842 JVMState* ejvms = NULL; 1843 if (has_exceptions()) { 1844 ejvms = transfer_exceptions_into_jvms(); 1845 } 1846 1847 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1848 ReplacedNodes replaced_nodes_exception; 1849 Node* ex_ctl = top(); 1850 1851 SafePointNode* final_state = stop(); 1852 1853 // Find all the needed outputs of this call 1854 CallProjections callprojs; 1855 call->extract_projections(&callprojs, true); 1856 1857 Node* init_mem = call->in(TypeFunc::Memory); 1858 Node* final_mem = final_state->in(TypeFunc::Memory); 1859 Node* final_ctl = final_state->in(TypeFunc::Control); 1860 Node* final_io = final_state->in(TypeFunc::I_O); 1861 1862 // Replace all the old call edges with the edges from the inlining result 1863 if (callprojs.fallthrough_catchproj != NULL) { 1864 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1865 } 1866 if (callprojs.fallthrough_memproj != NULL) { 1867 if (final_mem->is_MergeMem()) { 1868 // Parser's exits MergeMem was not transformed but may be optimized 1869 final_mem = _gvn.transform(final_mem); 1870 } 1871 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1872 } 1873 if (callprojs.fallthrough_ioproj != NULL) { 1874 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1875 } 1876 1877 // Replace the result with the new result if it exists and is used 1878 if (callprojs.resproj != NULL && result != NULL) { 1879 C->gvn_replace_by(callprojs.resproj, result); 1880 } 1881 1882 if (ejvms == NULL) { 1883 // No exception edges to simply kill off those paths 1884 if (callprojs.catchall_catchproj != NULL) { 1885 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1886 } 1887 if (callprojs.catchall_memproj != NULL) { 1888 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1889 } 1890 if (callprojs.catchall_ioproj != NULL) { 1891 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1892 } 1893 // Replace the old exception object with top 1894 if (callprojs.exobj != NULL) { 1895 C->gvn_replace_by(callprojs.exobj, C->top()); 1896 } 1897 } else { 1898 GraphKit ekit(ejvms); 1899 1900 // Load my combined exception state into the kit, with all phis transformed: 1901 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1902 replaced_nodes_exception = ex_map->replaced_nodes(); 1903 1904 Node* ex_oop = ekit.use_exception_state(ex_map); 1905 1906 if (callprojs.catchall_catchproj != NULL) { 1907 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1908 ex_ctl = ekit.control(); 1909 } 1910 if (callprojs.catchall_memproj != NULL) { 1911 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1912 } 1913 if (callprojs.catchall_ioproj != NULL) { 1914 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1915 } 1916 1917 // Replace the old exception object with the newly created one 1918 if (callprojs.exobj != NULL) { 1919 C->gvn_replace_by(callprojs.exobj, ex_oop); 1920 } 1921 } 1922 1923 // Disconnect the call from the graph 1924 call->disconnect_inputs(NULL, C); 1925 C->gvn_replace_by(call, C->top()); 1926 1927 // Clean up any MergeMems that feed other MergeMems since the 1928 // optimizer doesn't like that. 1929 if (final_mem->is_MergeMem()) { 1930 Node_List wl; 1931 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1932 Node* m = i.get(); 1933 if (m->is_MergeMem() && !wl.contains(m)) { 1934 wl.push(m); 1935 } 1936 } 1937 while (wl.size() > 0) { 1938 _gvn.transform(wl.pop()); 1939 } 1940 } 1941 1942 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1943 replaced_nodes.apply(C, final_ctl); 1944 } 1945 if (!ex_ctl->is_top() && do_replaced_nodes) { 1946 replaced_nodes_exception.apply(C, ex_ctl); 1947 } 1948 } 1949 1950 1951 //------------------------------increment_counter------------------------------ 1952 // for statistics: increment a VM counter by 1 1953 1954 void GraphKit::increment_counter(address counter_addr) { 1955 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1956 increment_counter(adr1); 1957 } 1958 1959 void GraphKit::increment_counter(Node* counter_addr) { 1960 int adr_type = Compile::AliasIdxRaw; 1961 Node* ctrl = control(); 1962 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1963 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1964 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1965 } 1966 1967 1968 //------------------------------uncommon_trap---------------------------------- 1969 // Bail out to the interpreter in mid-method. Implemented by calling the 1970 // uncommon_trap blob. This helper function inserts a runtime call with the 1971 // right debug info. 1972 void GraphKit::uncommon_trap(int trap_request, 1973 ciKlass* klass, const char* comment, 1974 bool must_throw, 1975 bool keep_exact_action) { 1976 if (failing()) stop(); 1977 if (stopped()) return; // trap reachable? 1978 1979 // Note: If ProfileTraps is true, and if a deopt. actually 1980 // occurs here, the runtime will make sure an MDO exists. There is 1981 // no need to call method()->ensure_method_data() at this point. 1982 1983 // Set the stack pointer to the right value for reexecution: 1984 set_sp(reexecute_sp()); 1985 1986 #ifdef ASSERT 1987 if (!must_throw) { 1988 // Make sure the stack has at least enough depth to execute 1989 // the current bytecode. 1990 int inputs, ignored_depth; 1991 if (compute_stack_effects(inputs, ignored_depth)) { 1992 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1993 Bytecodes::name(java_bc()), sp(), inputs); 1994 } 1995 } 1996 #endif 1997 1998 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1999 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2000 2001 switch (action) { 2002 case Deoptimization::Action_maybe_recompile: 2003 case Deoptimization::Action_reinterpret: 2004 // Temporary fix for 6529811 to allow virtual calls to be sure they 2005 // get the chance to go from mono->bi->mega 2006 if (!keep_exact_action && 2007 Deoptimization::trap_request_index(trap_request) < 0 && 2008 too_many_recompiles(reason)) { 2009 // This BCI is causing too many recompilations. 2010 if (C->log() != NULL) { 2011 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2012 Deoptimization::trap_reason_name(reason), 2013 Deoptimization::trap_action_name(action)); 2014 } 2015 action = Deoptimization::Action_none; 2016 trap_request = Deoptimization::make_trap_request(reason, action); 2017 } else { 2018 C->set_trap_can_recompile(true); 2019 } 2020 break; 2021 case Deoptimization::Action_make_not_entrant: 2022 C->set_trap_can_recompile(true); 2023 break; 2024 #ifdef ASSERT 2025 case Deoptimization::Action_none: 2026 case Deoptimization::Action_make_not_compilable: 2027 break; 2028 default: 2029 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2030 break; 2031 #endif 2032 } 2033 2034 if (TraceOptoParse) { 2035 char buf[100]; 2036 tty->print_cr("Uncommon trap %s at bci:%d", 2037 Deoptimization::format_trap_request(buf, sizeof(buf), 2038 trap_request), bci()); 2039 } 2040 2041 CompileLog* log = C->log(); 2042 if (log != NULL) { 2043 int kid = (klass == NULL)? -1: log->identify(klass); 2044 log->begin_elem("uncommon_trap bci='%d'", bci()); 2045 char buf[100]; 2046 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2047 trap_request)); 2048 if (kid >= 0) log->print(" klass='%d'", kid); 2049 if (comment != NULL) log->print(" comment='%s'", comment); 2050 log->end_elem(); 2051 } 2052 2053 // Make sure any guarding test views this path as very unlikely 2054 Node *i0 = control()->in(0); 2055 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2056 IfNode *iff = i0->as_If(); 2057 float f = iff->_prob; // Get prob 2058 if (control()->Opcode() == Op_IfTrue) { 2059 if (f > PROB_UNLIKELY_MAG(4)) 2060 iff->_prob = PROB_MIN; 2061 } else { 2062 if (f < PROB_LIKELY_MAG(4)) 2063 iff->_prob = PROB_MAX; 2064 } 2065 } 2066 2067 // Clear out dead values from the debug info. 2068 kill_dead_locals(); 2069 2070 // Now insert the uncommon trap subroutine call 2071 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2072 const TypePtr* no_memory_effects = NULL; 2073 // Pass the index of the class to be loaded 2074 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2075 (must_throw ? RC_MUST_THROW : 0), 2076 OptoRuntime::uncommon_trap_Type(), 2077 call_addr, "uncommon_trap", no_memory_effects, 2078 intcon(trap_request)); 2079 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2080 "must extract request correctly from the graph"); 2081 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2082 2083 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2084 // The debug info is the only real input to this call. 2085 2086 // Halt-and-catch fire here. The above call should never return! 2087 HaltNode* halt = new HaltNode(control(), frameptr()); 2088 _gvn.set_type_bottom(halt); 2089 root()->add_req(halt); 2090 2091 stop_and_kill_map(); 2092 } 2093 2094 2095 //--------------------------just_allocated_object------------------------------ 2096 // Report the object that was just allocated. 2097 // It must be the case that there are no intervening safepoints. 2098 // We use this to determine if an object is so "fresh" that 2099 // it does not require card marks. 2100 Node* GraphKit::just_allocated_object(Node* current_control) { 2101 if (C->recent_alloc_ctl() == current_control) 2102 return C->recent_alloc_obj(); 2103 return NULL; 2104 } 2105 2106 2107 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2108 // (Note: TypeFunc::make has a cache that makes this fast.) 2109 const TypeFunc* tf = TypeFunc::make(dest_method); 2110 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2111 for (int j = 0; j < nargs; j++) { 2112 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2113 if( targ->basic_type() == T_DOUBLE ) { 2114 // If any parameters are doubles, they must be rounded before 2115 // the call, dstore_rounding does gvn.transform 2116 Node *arg = argument(j); 2117 arg = dstore_rounding(arg); 2118 set_argument(j, arg); 2119 } 2120 } 2121 } 2122 2123 /** 2124 * Record profiling data exact_kls for Node n with the type system so 2125 * that it can propagate it (speculation) 2126 * 2127 * @param n node that the type applies to 2128 * @param exact_kls type from profiling 2129 * @param maybe_null did profiling see null? 2130 * 2131 * @return node with improved type 2132 */ 2133 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2134 const Type* current_type = _gvn.type(n); 2135 assert(UseTypeSpeculation, "type speculation must be on"); 2136 2137 const TypePtr* speculative = current_type->speculative(); 2138 2139 // Should the klass from the profile be recorded in the speculative type? 2140 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2141 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2142 const TypeOopPtr* xtype = tklass->as_instance_type(); 2143 assert(xtype->klass_is_exact(), "Should be exact"); 2144 // Any reason to believe n is not null (from this profiling or a previous one)? 2145 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2146 // record the new speculative type's depth 2147 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2148 speculative = speculative->with_inline_depth(jvms()->depth()); 2149 } else if (current_type->would_improve_ptr(maybe_null)) { 2150 // Profiling report that null was never seen so we can change the 2151 // speculative type to non null ptr. 2152 assert(!maybe_null, "nothing to improve"); 2153 if (speculative == NULL) { 2154 speculative = TypePtr::NOTNULL; 2155 } else { 2156 const TypePtr* ptr = TypePtr::NOTNULL; 2157 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2158 } 2159 } 2160 2161 if (speculative != current_type->speculative()) { 2162 // Build a type with a speculative type (what we think we know 2163 // about the type but will need a guard when we use it) 2164 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2165 // We're changing the type, we need a new CheckCast node to carry 2166 // the new type. The new type depends on the control: what 2167 // profiling tells us is only valid from here as far as we can 2168 // tell. 2169 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2170 cast = _gvn.transform(cast); 2171 replace_in_map(n, cast); 2172 n = cast; 2173 } 2174 2175 return n; 2176 } 2177 2178 /** 2179 * Record profiling data from receiver profiling at an invoke with the 2180 * type system so that it can propagate it (speculation) 2181 * 2182 * @param n receiver node 2183 * 2184 * @return node with improved type 2185 */ 2186 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2187 if (!UseTypeSpeculation) { 2188 return n; 2189 } 2190 ciKlass* exact_kls = profile_has_unique_klass(); 2191 bool maybe_null = true; 2192 if (java_bc() == Bytecodes::_checkcast || 2193 java_bc() == Bytecodes::_instanceof || 2194 java_bc() == Bytecodes::_aastore) { 2195 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2196 maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2197 } 2198 return record_profile_for_speculation(n, exact_kls, maybe_null); 2199 } 2200 2201 /** 2202 * Record profiling data from argument profiling at an invoke with the 2203 * type system so that it can propagate it (speculation) 2204 * 2205 * @param dest_method target method for the call 2206 * @param bc what invoke bytecode is this? 2207 */ 2208 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2209 if (!UseTypeSpeculation) { 2210 return; 2211 } 2212 const TypeFunc* tf = TypeFunc::make(dest_method); 2213 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2214 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2215 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2216 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2217 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2218 bool maybe_null = true; 2219 ciKlass* better_type = NULL; 2220 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2221 record_profile_for_speculation(argument(j), better_type, maybe_null); 2222 } 2223 i++; 2224 } 2225 } 2226 } 2227 2228 /** 2229 * Record profiling data from parameter profiling at an invoke with 2230 * the type system so that it can propagate it (speculation) 2231 */ 2232 void GraphKit::record_profiled_parameters_for_speculation() { 2233 if (!UseTypeSpeculation) { 2234 return; 2235 } 2236 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2237 if (_gvn.type(local(i))->isa_oopptr()) { 2238 bool maybe_null = true; 2239 ciKlass* better_type = NULL; 2240 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2241 record_profile_for_speculation(local(i), better_type, maybe_null); 2242 } 2243 j++; 2244 } 2245 } 2246 } 2247 2248 /** 2249 * Record profiling data from return value profiling at an invoke with 2250 * the type system so that it can propagate it (speculation) 2251 */ 2252 void GraphKit::record_profiled_return_for_speculation() { 2253 if (!UseTypeSpeculation) { 2254 return; 2255 } 2256 bool maybe_null = true; 2257 ciKlass* better_type = NULL; 2258 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2259 // If profiling reports a single type for the return value, 2260 // feed it to the type system so it can propagate it as a 2261 // speculative type 2262 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2263 } 2264 } 2265 2266 void GraphKit::round_double_result(ciMethod* dest_method) { 2267 // A non-strict method may return a double value which has an extended 2268 // exponent, but this must not be visible in a caller which is 'strict' 2269 // If a strict caller invokes a non-strict callee, round a double result 2270 2271 BasicType result_type = dest_method->return_type()->basic_type(); 2272 assert( method() != NULL, "must have caller context"); 2273 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2274 // Destination method's return value is on top of stack 2275 // dstore_rounding() does gvn.transform 2276 Node *result = pop_pair(); 2277 result = dstore_rounding(result); 2278 push_pair(result); 2279 } 2280 } 2281 2282 // rounding for strict float precision conformance 2283 Node* GraphKit::precision_rounding(Node* n) { 2284 return UseStrictFP && _method->flags().is_strict() 2285 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2286 ? _gvn.transform( new RoundFloatNode(0, n) ) 2287 : n; 2288 } 2289 2290 // rounding for strict double precision conformance 2291 Node* GraphKit::dprecision_rounding(Node *n) { 2292 return UseStrictFP && _method->flags().is_strict() 2293 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2294 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2295 : n; 2296 } 2297 2298 // rounding for non-strict double stores 2299 Node* GraphKit::dstore_rounding(Node* n) { 2300 return Matcher::strict_fp_requires_explicit_rounding 2301 && UseSSE <= 1 2302 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2303 : n; 2304 } 2305 2306 //============================================================================= 2307 // Generate a fast path/slow path idiom. Graph looks like: 2308 // [foo] indicates that 'foo' is a parameter 2309 // 2310 // [in] NULL 2311 // \ / 2312 // CmpP 2313 // Bool ne 2314 // If 2315 // / \ 2316 // True False-<2> 2317 // / | 2318 // / cast_not_null 2319 // Load | | ^ 2320 // [fast_test] | | 2321 // gvn to opt_test | | 2322 // / \ | <1> 2323 // True False | 2324 // | \\ | 2325 // [slow_call] \[fast_result] 2326 // Ctl Val \ \ 2327 // | \ \ 2328 // Catch <1> \ \ 2329 // / \ ^ \ \ 2330 // Ex No_Ex | \ \ 2331 // | \ \ | \ <2> \ 2332 // ... \ [slow_res] | | \ [null_result] 2333 // \ \--+--+--- | | 2334 // \ | / \ | / 2335 // --------Region Phi 2336 // 2337 //============================================================================= 2338 // Code is structured as a series of driver functions all called 'do_XXX' that 2339 // call a set of helper functions. Helper functions first, then drivers. 2340 2341 //------------------------------null_check_oop--------------------------------- 2342 // Null check oop. Set null-path control into Region in slot 3. 2343 // Make a cast-not-nullness use the other not-null control. Return cast. 2344 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2345 bool never_see_null, 2346 bool safe_for_replace, 2347 bool speculative) { 2348 // Initial NULL check taken path 2349 (*null_control) = top(); 2350 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2351 2352 // Generate uncommon_trap: 2353 if (never_see_null && (*null_control) != top()) { 2354 // If we see an unexpected null at a check-cast we record it and force a 2355 // recompile; the offending check-cast will be compiled to handle NULLs. 2356 // If we see more than one offending BCI, then all checkcasts in the 2357 // method will be compiled to handle NULLs. 2358 PreserveJVMState pjvms(this); 2359 set_control(*null_control); 2360 replace_in_map(value, null()); 2361 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2362 uncommon_trap(reason, 2363 Deoptimization::Action_make_not_entrant); 2364 (*null_control) = top(); // NULL path is dead 2365 } 2366 if ((*null_control) == top() && safe_for_replace) { 2367 replace_in_map(value, cast); 2368 } 2369 2370 // Cast away null-ness on the result 2371 return cast; 2372 } 2373 2374 //------------------------------opt_iff---------------------------------------- 2375 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2376 // Return slow-path control. 2377 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2378 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2379 2380 // Fast path taken; set region slot 2 2381 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2382 region->init_req(2,fast_taken); // Capture fast-control 2383 2384 // Fast path not-taken, i.e. slow path 2385 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2386 return slow_taken; 2387 } 2388 2389 //-----------------------------make_runtime_call------------------------------- 2390 Node* GraphKit::make_runtime_call(int flags, 2391 const TypeFunc* call_type, address call_addr, 2392 const char* call_name, 2393 const TypePtr* adr_type, 2394 // The following parms are all optional. 2395 // The first NULL ends the list. 2396 Node* parm0, Node* parm1, 2397 Node* parm2, Node* parm3, 2398 Node* parm4, Node* parm5, 2399 Node* parm6, Node* parm7) { 2400 // Slow-path call 2401 bool is_leaf = !(flags & RC_NO_LEAF); 2402 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2403 if (call_name == NULL) { 2404 assert(!is_leaf, "must supply name for leaf"); 2405 call_name = OptoRuntime::stub_name(call_addr); 2406 } 2407 CallNode* call; 2408 if (!is_leaf) { 2409 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2410 bci(), adr_type); 2411 } else if (flags & RC_NO_FP) { 2412 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2413 } else { 2414 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2415 } 2416 2417 // The following is similar to set_edges_for_java_call, 2418 // except that the memory effects of the call are restricted to AliasIdxRaw. 2419 2420 // Slow path call has no side-effects, uses few values 2421 bool wide_in = !(flags & RC_NARROW_MEM); 2422 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2423 2424 Node* prev_mem = NULL; 2425 if (wide_in) { 2426 prev_mem = set_predefined_input_for_runtime_call(call); 2427 } else { 2428 assert(!wide_out, "narrow in => narrow out"); 2429 Node* narrow_mem = memory(adr_type); 2430 prev_mem = reset_memory(); 2431 map()->set_memory(narrow_mem); 2432 set_predefined_input_for_runtime_call(call); 2433 } 2434 2435 // Hook each parm in order. Stop looking at the first NULL. 2436 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2437 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2438 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2439 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2440 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2441 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2442 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2443 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2444 /* close each nested if ===> */ } } } } } } } } 2445 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2446 2447 if (!is_leaf) { 2448 // Non-leaves can block and take safepoints: 2449 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2450 } 2451 // Non-leaves can throw exceptions: 2452 if (has_io) { 2453 call->set_req(TypeFunc::I_O, i_o()); 2454 } 2455 2456 if (flags & RC_UNCOMMON) { 2457 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2458 // (An "if" probability corresponds roughly to an unconditional count. 2459 // Sort of.) 2460 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2461 } 2462 2463 Node* c = _gvn.transform(call); 2464 assert(c == call, "cannot disappear"); 2465 2466 if (wide_out) { 2467 // Slow path call has full side-effects. 2468 set_predefined_output_for_runtime_call(call); 2469 } else { 2470 // Slow path call has few side-effects, and/or sets few values. 2471 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2472 } 2473 2474 if (has_io) { 2475 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2476 } 2477 return call; 2478 2479 } 2480 2481 //------------------------------merge_memory----------------------------------- 2482 // Merge memory from one path into the current memory state. 2483 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2484 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2485 Node* old_slice = mms.force_memory(); 2486 Node* new_slice = mms.memory2(); 2487 if (old_slice != new_slice) { 2488 PhiNode* phi; 2489 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2490 if (mms.is_empty()) { 2491 // clone base memory Phi's inputs for this memory slice 2492 assert(old_slice == mms.base_memory(), "sanity"); 2493 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2494 _gvn.set_type(phi, Type::MEMORY); 2495 for (uint i = 1; i < phi->req(); i++) { 2496 phi->init_req(i, old_slice->in(i)); 2497 } 2498 } else { 2499 phi = old_slice->as_Phi(); // Phi was generated already 2500 } 2501 } else { 2502 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2503 _gvn.set_type(phi, Type::MEMORY); 2504 } 2505 phi->set_req(new_path, new_slice); 2506 mms.set_memory(phi); 2507 } 2508 } 2509 } 2510 2511 //------------------------------make_slow_call_ex------------------------------ 2512 // Make the exception handler hookups for the slow call 2513 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2514 if (stopped()) return; 2515 2516 // Make a catch node with just two handlers: fall-through and catch-all 2517 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2518 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2519 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2520 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2521 2522 { PreserveJVMState pjvms(this); 2523 set_control(excp); 2524 set_i_o(i_o); 2525 2526 if (excp != top()) { 2527 if (deoptimize) { 2528 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2529 uncommon_trap(Deoptimization::Reason_unhandled, 2530 Deoptimization::Action_none); 2531 } else { 2532 // Create an exception state also. 2533 // Use an exact type if the caller has specified a specific exception. 2534 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2535 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2536 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2537 } 2538 } 2539 } 2540 2541 // Get the no-exception control from the CatchNode. 2542 set_control(norm); 2543 } 2544 2545 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2546 Node* cmp = NULL; 2547 switch(bt) { 2548 case T_INT: cmp = new CmpINode(in1, in2); break; 2549 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2550 default: fatal("unexpected comparison type %s", type2name(bt)); 2551 } 2552 gvn->transform(cmp); 2553 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2554 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2555 gvn->transform(iff); 2556 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2557 return iff; 2558 } 2559 2560 2561 //-------------------------------gen_subtype_check----------------------------- 2562 // Generate a subtyping check. Takes as input the subtype and supertype. 2563 // Returns 2 values: sets the default control() to the true path and returns 2564 // the false path. Only reads invariant memory; sets no (visible) memory. 2565 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2566 // but that's not exposed to the optimizer. This call also doesn't take in an 2567 // Object; if you wish to check an Object you need to load the Object's class 2568 // prior to coming here. 2569 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2570 Compile* C = gvn->C; 2571 2572 if ((*ctrl)->is_top()) { 2573 return C->top(); 2574 } 2575 2576 // Fast check for identical types, perhaps identical constants. 2577 // The types can even be identical non-constants, in cases 2578 // involving Array.newInstance, Object.clone, etc. 2579 if (subklass == superklass) 2580 return C->top(); // false path is dead; no test needed. 2581 2582 if (gvn->type(superklass)->singleton()) { 2583 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2584 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2585 2586 // In the common case of an exact superklass, try to fold up the 2587 // test before generating code. You may ask, why not just generate 2588 // the code and then let it fold up? The answer is that the generated 2589 // code will necessarily include null checks, which do not always 2590 // completely fold away. If they are also needless, then they turn 2591 // into a performance loss. Example: 2592 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2593 // Here, the type of 'fa' is often exact, so the store check 2594 // of fa[1]=x will fold up, without testing the nullness of x. 2595 switch (C->static_subtype_check(superk, subk)) { 2596 case Compile::SSC_always_false: 2597 { 2598 Node* always_fail = *ctrl; 2599 *ctrl = gvn->C->top(); 2600 return always_fail; 2601 } 2602 case Compile::SSC_always_true: 2603 return C->top(); 2604 case Compile::SSC_easy_test: 2605 { 2606 // Just do a direct pointer compare and be done. 2607 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2608 *ctrl = gvn->transform(new IfTrueNode(iff)); 2609 return gvn->transform(new IfFalseNode(iff)); 2610 } 2611 case Compile::SSC_full_test: 2612 break; 2613 default: 2614 ShouldNotReachHere(); 2615 } 2616 } 2617 2618 // %%% Possible further optimization: Even if the superklass is not exact, 2619 // if the subklass is the unique subtype of the superklass, the check 2620 // will always succeed. We could leave a dependency behind to ensure this. 2621 2622 // First load the super-klass's check-offset 2623 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2624 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2625 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2626 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2627 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2628 2629 // Load from the sub-klass's super-class display list, or a 1-word cache of 2630 // the secondary superclass list, or a failing value with a sentinel offset 2631 // if the super-klass is an interface or exceptionally deep in the Java 2632 // hierarchy and we have to scan the secondary superclass list the hard way. 2633 // Worst-case type is a little odd: NULL is allowed as a result (usually 2634 // klass loads can never produce a NULL). 2635 Node *chk_off_X = chk_off; 2636 #ifdef _LP64 2637 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2638 #endif 2639 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2640 // For some types like interfaces the following loadKlass is from a 1-word 2641 // cache which is mutable so can't use immutable memory. Other 2642 // types load from the super-class display table which is immutable. 2643 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2644 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2645 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2646 2647 // Compile speed common case: ARE a subtype and we canNOT fail 2648 if( superklass == nkls ) 2649 return C->top(); // false path is dead; no test needed. 2650 2651 // See if we get an immediate positive hit. Happens roughly 83% of the 2652 // time. Test to see if the value loaded just previously from the subklass 2653 // is exactly the superklass. 2654 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2655 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2656 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2657 2658 // Compile speed common case: Check for being deterministic right now. If 2659 // chk_off is a constant and not equal to cacheoff then we are NOT a 2660 // subklass. In this case we need exactly the 1 test above and we can 2661 // return those results immediately. 2662 if (!might_be_cache) { 2663 Node* not_subtype_ctrl = *ctrl; 2664 *ctrl = iftrue1; // We need exactly the 1 test above 2665 return not_subtype_ctrl; 2666 } 2667 2668 // Gather the various success & failures here 2669 RegionNode *r_ok_subtype = new RegionNode(4); 2670 gvn->record_for_igvn(r_ok_subtype); 2671 RegionNode *r_not_subtype = new RegionNode(3); 2672 gvn->record_for_igvn(r_not_subtype); 2673 2674 r_ok_subtype->init_req(1, iftrue1); 2675 2676 // Check for immediate negative hit. Happens roughly 11% of the time (which 2677 // is roughly 63% of the remaining cases). Test to see if the loaded 2678 // check-offset points into the subklass display list or the 1-element 2679 // cache. If it points to the display (and NOT the cache) and the display 2680 // missed then it's not a subtype. 2681 Node *cacheoff = gvn->intcon(cacheoff_con); 2682 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2683 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2684 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2685 2686 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2687 // No performance impact (too rare) but allows sharing of secondary arrays 2688 // which has some footprint reduction. 2689 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2690 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2691 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2692 2693 // -- Roads not taken here: -- 2694 // We could also have chosen to perform the self-check at the beginning 2695 // of this code sequence, as the assembler does. This would not pay off 2696 // the same way, since the optimizer, unlike the assembler, can perform 2697 // static type analysis to fold away many successful self-checks. 2698 // Non-foldable self checks work better here in second position, because 2699 // the initial primary superclass check subsumes a self-check for most 2700 // types. An exception would be a secondary type like array-of-interface, 2701 // which does not appear in its own primary supertype display. 2702 // Finally, we could have chosen to move the self-check into the 2703 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2704 // dependent manner. But it is worthwhile to have the check here, 2705 // where it can be perhaps be optimized. The cost in code space is 2706 // small (register compare, branch). 2707 2708 // Now do a linear scan of the secondary super-klass array. Again, no real 2709 // performance impact (too rare) but it's gotta be done. 2710 // Since the code is rarely used, there is no penalty for moving it 2711 // out of line, and it can only improve I-cache density. 2712 // The decision to inline or out-of-line this final check is platform 2713 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2714 Node* psc = gvn->transform( 2715 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2716 2717 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2718 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2719 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2720 2721 // Return false path; set default control to true path. 2722 *ctrl = gvn->transform(r_ok_subtype); 2723 return gvn->transform(r_not_subtype); 2724 } 2725 2726 // Profile-driven exact type check: 2727 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2728 float prob, 2729 Node* *casted_receiver) { 2730 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2731 Node* recv_klass = load_object_klass(receiver); 2732 Node* want_klass = makecon(tklass); 2733 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2734 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2735 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2736 set_control( _gvn.transform( new IfTrueNode (iff) )); 2737 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2738 2739 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2740 assert(recv_xtype->klass_is_exact(), ""); 2741 2742 // Subsume downstream occurrences of receiver with a cast to 2743 // recv_xtype, since now we know what the type will be. 2744 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2745 (*casted_receiver) = _gvn.transform(cast); 2746 // (User must make the replace_in_map call.) 2747 2748 return fail; 2749 } 2750 2751 2752 //------------------------------seems_never_null------------------------------- 2753 // Use null_seen information if it is available from the profile. 2754 // If we see an unexpected null at a type check we record it and force a 2755 // recompile; the offending check will be recompiled to handle NULLs. 2756 // If we see several offending BCIs, then all checks in the 2757 // method will be recompiled. 2758 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2759 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2760 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2761 if (UncommonNullCast // Cutout for this technique 2762 && obj != null() // And not the -Xcomp stupid case? 2763 && !too_many_traps(reason) 2764 ) { 2765 if (speculating) { 2766 return true; 2767 } 2768 if (data == NULL) 2769 // Edge case: no mature data. Be optimistic here. 2770 return true; 2771 // If the profile has not seen a null, assume it won't happen. 2772 assert(java_bc() == Bytecodes::_checkcast || 2773 java_bc() == Bytecodes::_instanceof || 2774 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2775 return !data->as_BitData()->null_seen(); 2776 } 2777 speculating = false; 2778 return false; 2779 } 2780 2781 //------------------------maybe_cast_profiled_receiver------------------------- 2782 // If the profile has seen exactly one type, narrow to exactly that type. 2783 // Subsequent type checks will always fold up. 2784 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2785 ciKlass* require_klass, 2786 ciKlass* spec_klass, 2787 bool safe_for_replace) { 2788 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2789 2790 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2791 2792 // Make sure we haven't already deoptimized from this tactic. 2793 if (too_many_traps(reason) || too_many_recompiles(reason)) 2794 return NULL; 2795 2796 // (No, this isn't a call, but it's enough like a virtual call 2797 // to use the same ciMethod accessor to get the profile info...) 2798 // If we have a speculative type use it instead of profiling (which 2799 // may not help us) 2800 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2801 if (exact_kls != NULL) {// no cast failures here 2802 if (require_klass == NULL || 2803 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2804 // If we narrow the type to match what the type profile sees or 2805 // the speculative type, we can then remove the rest of the 2806 // cast. 2807 // This is a win, even if the exact_kls is very specific, 2808 // because downstream operations, such as method calls, 2809 // will often benefit from the sharper type. 2810 Node* exact_obj = not_null_obj; // will get updated in place... 2811 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2812 &exact_obj); 2813 { PreserveJVMState pjvms(this); 2814 set_control(slow_ctl); 2815 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2816 } 2817 if (safe_for_replace) { 2818 replace_in_map(not_null_obj, exact_obj); 2819 } 2820 return exact_obj; 2821 } 2822 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2823 } 2824 2825 return NULL; 2826 } 2827 2828 /** 2829 * Cast obj to type and emit guard unless we had too many traps here 2830 * already 2831 * 2832 * @param obj node being casted 2833 * @param type type to cast the node to 2834 * @param not_null true if we know node cannot be null 2835 */ 2836 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2837 ciKlass* type, 2838 bool not_null) { 2839 if (stopped()) { 2840 return obj; 2841 } 2842 2843 // type == NULL if profiling tells us this object is always null 2844 if (type != NULL) { 2845 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2846 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2847 2848 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2849 !too_many_traps(class_reason) && 2850 !too_many_recompiles(class_reason)) { 2851 Node* not_null_obj = NULL; 2852 // not_null is true if we know the object is not null and 2853 // there's no need for a null check 2854 if (!not_null) { 2855 Node* null_ctl = top(); 2856 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2857 assert(null_ctl->is_top(), "no null control here"); 2858 } else { 2859 not_null_obj = obj; 2860 } 2861 2862 Node* exact_obj = not_null_obj; 2863 ciKlass* exact_kls = type; 2864 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2865 &exact_obj); 2866 { 2867 PreserveJVMState pjvms(this); 2868 set_control(slow_ctl); 2869 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2870 } 2871 replace_in_map(not_null_obj, exact_obj); 2872 obj = exact_obj; 2873 } 2874 } else { 2875 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2876 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2877 Node* exact_obj = null_assert(obj); 2878 replace_in_map(obj, exact_obj); 2879 obj = exact_obj; 2880 } 2881 } 2882 return obj; 2883 } 2884 2885 //-------------------------------gen_instanceof-------------------------------- 2886 // Generate an instance-of idiom. Used by both the instance-of bytecode 2887 // and the reflective instance-of call. 2888 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2889 kill_dead_locals(); // Benefit all the uncommon traps 2890 assert( !stopped(), "dead parse path should be checked in callers" ); 2891 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2892 "must check for not-null not-dead klass in callers"); 2893 2894 // Make the merge point 2895 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2896 RegionNode* region = new RegionNode(PATH_LIMIT); 2897 Node* phi = new PhiNode(region, TypeInt::BOOL); 2898 C->set_has_split_ifs(true); // Has chance for split-if optimization 2899 2900 ciProfileData* data = NULL; 2901 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2902 data = method()->method_data()->bci_to_data(bci()); 2903 } 2904 bool speculative_not_null = false; 2905 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2906 && seems_never_null(obj, data, speculative_not_null)); 2907 2908 // Null check; get casted pointer; set region slot 3 2909 Node* null_ctl = top(); 2910 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2911 2912 // If not_null_obj is dead, only null-path is taken 2913 if (stopped()) { // Doing instance-of on a NULL? 2914 set_control(null_ctl); 2915 return intcon(0); 2916 } 2917 region->init_req(_null_path, null_ctl); 2918 phi ->init_req(_null_path, intcon(0)); // Set null path value 2919 if (null_ctl == top()) { 2920 // Do this eagerly, so that pattern matches like is_diamond_phi 2921 // will work even during parsing. 2922 assert(_null_path == PATH_LIMIT-1, "delete last"); 2923 region->del_req(_null_path); 2924 phi ->del_req(_null_path); 2925 } 2926 2927 // Do we know the type check always succeed? 2928 bool known_statically = false; 2929 if (_gvn.type(superklass)->singleton()) { 2930 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2931 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2932 if (subk != NULL && subk->is_loaded()) { 2933 int static_res = C->static_subtype_check(superk, subk); 2934 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2935 } 2936 } 2937 2938 if (known_statically && UseTypeSpeculation) { 2939 // If we know the type check always succeeds then we don't use the 2940 // profiling data at this bytecode. Don't lose it, feed it to the 2941 // type system as a speculative type. 2942 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2943 } else { 2944 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2945 // We may not have profiling here or it may not help us. If we 2946 // have a speculative type use it to perform an exact cast. 2947 ciKlass* spec_obj_type = obj_type->speculative_type(); 2948 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2949 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2950 if (stopped()) { // Profile disagrees with this path. 2951 set_control(null_ctl); // Null is the only remaining possibility. 2952 return intcon(0); 2953 } 2954 if (cast_obj != NULL) { 2955 not_null_obj = cast_obj; 2956 } 2957 } 2958 } 2959 2960 if (ShenandoahVerifyReadsToFromSpace) { 2961 not_null_obj = shenandoah_read_barrier(not_null_obj); 2962 } 2963 2964 // Load the object's klass 2965 Node* obj_klass = load_object_klass(not_null_obj); 2966 2967 // Generate the subtype check 2968 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2969 2970 // Plug in the success path to the general merge in slot 1. 2971 region->init_req(_obj_path, control()); 2972 phi ->init_req(_obj_path, intcon(1)); 2973 2974 // Plug in the failing path to the general merge in slot 2. 2975 region->init_req(_fail_path, not_subtype_ctrl); 2976 phi ->init_req(_fail_path, intcon(0)); 2977 2978 // Return final merged results 2979 set_control( _gvn.transform(region) ); 2980 record_for_igvn(region); 2981 return _gvn.transform(phi); 2982 } 2983 2984 //-------------------------------gen_checkcast--------------------------------- 2985 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2986 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2987 // uncommon-trap paths work. Adjust stack after this call. 2988 // If failure_control is supplied and not null, it is filled in with 2989 // the control edge for the cast failure. Otherwise, an appropriate 2990 // uncommon trap or exception is thrown. 2991 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2992 Node* *failure_control) { 2993 kill_dead_locals(); // Benefit all the uncommon traps 2994 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2995 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2996 2997 // Fast cutout: Check the case that the cast is vacuously true. 2998 // This detects the common cases where the test will short-circuit 2999 // away completely. We do this before we perform the null check, 3000 // because if the test is going to turn into zero code, we don't 3001 // want a residual null check left around. (Causes a slowdown, 3002 // for example, in some objArray manipulations, such as a[i]=a[j].) 3003 if (tk->singleton()) { 3004 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3005 if (objtp != NULL && objtp->klass() != NULL) { 3006 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3007 case Compile::SSC_always_true: 3008 // If we know the type check always succeed then we don't use 3009 // the profiling data at this bytecode. Don't lose it, feed it 3010 // to the type system as a speculative type. 3011 return record_profiled_receiver_for_speculation(obj); 3012 case Compile::SSC_always_false: 3013 // It needs a null check because a null will *pass* the cast check. 3014 // A non-null value will always produce an exception. 3015 return null_assert(obj); 3016 } 3017 } 3018 } 3019 3020 ciProfileData* data = NULL; 3021 bool safe_for_replace = false; 3022 if (failure_control == NULL) { // use MDO in regular case only 3023 assert(java_bc() == Bytecodes::_aastore || 3024 java_bc() == Bytecodes::_checkcast, 3025 "interpreter profiles type checks only for these BCs"); 3026 data = method()->method_data()->bci_to_data(bci()); 3027 safe_for_replace = true; 3028 } 3029 3030 // Make the merge point 3031 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3032 RegionNode* region = new RegionNode(PATH_LIMIT); 3033 Node* phi = new PhiNode(region, toop); 3034 C->set_has_split_ifs(true); // Has chance for split-if optimization 3035 3036 // Use null-cast information if it is available 3037 bool speculative_not_null = false; 3038 bool never_see_null = ((failure_control == NULL) // regular case only 3039 && seems_never_null(obj, data, speculative_not_null)); 3040 3041 // Null check; get casted pointer; set region slot 3 3042 Node* null_ctl = top(); 3043 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3044 3045 if (ShenandoahVerifyReadsToFromSpace) { 3046 not_null_obj = shenandoah_read_barrier(not_null_obj); 3047 } 3048 3049 // If not_null_obj is dead, only null-path is taken 3050 if (stopped()) { // Doing instance-of on a NULL? 3051 set_control(null_ctl); 3052 return null(); 3053 } 3054 region->init_req(_null_path, null_ctl); 3055 phi ->init_req(_null_path, null()); // Set null path value 3056 if (null_ctl == top()) { 3057 // Do this eagerly, so that pattern matches like is_diamond_phi 3058 // will work even during parsing. 3059 assert(_null_path == PATH_LIMIT-1, "delete last"); 3060 region->del_req(_null_path); 3061 phi ->del_req(_null_path); 3062 } 3063 3064 Node* cast_obj = NULL; 3065 if (tk->klass_is_exact()) { 3066 // The following optimization tries to statically cast the speculative type of the object 3067 // (for example obtained during profiling) to the type of the superklass and then do a 3068 // dynamic check that the type of the object is what we expect. To work correctly 3069 // for checkcast and aastore the type of superklass should be exact. 3070 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3071 // We may not have profiling here or it may not help us. If we have 3072 // a speculative type use it to perform an exact cast. 3073 ciKlass* spec_obj_type = obj_type->speculative_type(); 3074 if (spec_obj_type != NULL || data != NULL) { 3075 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3076 if (cast_obj != NULL) { 3077 if (failure_control != NULL) // failure is now impossible 3078 (*failure_control) = top(); 3079 // adjust the type of the phi to the exact klass: 3080 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3081 } 3082 } 3083 } 3084 3085 if (cast_obj == NULL) { 3086 // Load the object's klass 3087 Node* obj_klass = load_object_klass(not_null_obj); 3088 3089 // Generate the subtype check 3090 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3091 3092 // Plug in success path into the merge 3093 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3094 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3095 if (failure_control == NULL) { 3096 if (not_subtype_ctrl != top()) { // If failure is possible 3097 PreserveJVMState pjvms(this); 3098 set_control(not_subtype_ctrl); 3099 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3100 } 3101 } else { 3102 (*failure_control) = not_subtype_ctrl; 3103 } 3104 } 3105 3106 region->init_req(_obj_path, control()); 3107 phi ->init_req(_obj_path, cast_obj); 3108 3109 // A merge of NULL or Casted-NotNull obj 3110 Node* res = _gvn.transform(phi); 3111 3112 // Note I do NOT always 'replace_in_map(obj,result)' here. 3113 // if( tk->klass()->can_be_primary_super() ) 3114 // This means that if I successfully store an Object into an array-of-String 3115 // I 'forget' that the Object is really now known to be a String. I have to 3116 // do this because we don't have true union types for interfaces - if I store 3117 // a Baz into an array-of-Interface and then tell the optimizer it's an 3118 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3119 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3120 // replace_in_map( obj, res ); 3121 3122 // Return final merged results 3123 set_control( _gvn.transform(region) ); 3124 record_for_igvn(region); 3125 return res; 3126 } 3127 3128 //------------------------------next_monitor----------------------------------- 3129 // What number should be given to the next monitor? 3130 int GraphKit::next_monitor() { 3131 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3132 int next = current + C->sync_stack_slots(); 3133 // Keep the toplevel high water mark current: 3134 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3135 return current; 3136 } 3137 3138 //------------------------------insert_mem_bar--------------------------------- 3139 // Memory barrier to avoid floating things around 3140 // The membar serves as a pinch point between both control and all memory slices. 3141 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3142 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3143 mb->init_req(TypeFunc::Control, control()); 3144 mb->init_req(TypeFunc::Memory, reset_memory()); 3145 Node* membar = _gvn.transform(mb); 3146 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3147 set_all_memory_call(membar); 3148 return membar; 3149 } 3150 3151 //-------------------------insert_mem_bar_volatile---------------------------- 3152 // Memory barrier to avoid floating things around 3153 // The membar serves as a pinch point between both control and memory(alias_idx). 3154 // If you want to make a pinch point on all memory slices, do not use this 3155 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3156 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3157 // When Parse::do_put_xxx updates a volatile field, it appends a series 3158 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3159 // The first membar is on the same memory slice as the field store opcode. 3160 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3161 // All the other membars (for other volatile slices, including AliasIdxBot, 3162 // which stands for all unknown volatile slices) are control-dependent 3163 // on the first membar. This prevents later volatile loads or stores 3164 // from sliding up past the just-emitted store. 3165 3166 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3167 mb->set_req(TypeFunc::Control,control()); 3168 if (alias_idx == Compile::AliasIdxBot) { 3169 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3170 } else { 3171 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3172 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3173 } 3174 Node* membar = _gvn.transform(mb); 3175 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3176 if (alias_idx == Compile::AliasIdxBot) { 3177 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3178 } else { 3179 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3180 } 3181 return membar; 3182 } 3183 3184 //------------------------------shared_lock------------------------------------ 3185 // Emit locking code. 3186 FastLockNode* GraphKit::shared_lock(Node* obj) { 3187 // bci is either a monitorenter bc or InvocationEntryBci 3188 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3189 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3190 3191 if( !GenerateSynchronizationCode ) 3192 return NULL; // Not locking things? 3193 if (stopped()) // Dead monitor? 3194 return NULL; 3195 3196 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3197 3198 obj = shenandoah_write_barrier(obj); 3199 3200 // Box the stack location 3201 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3202 Node* mem = reset_memory(); 3203 3204 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3205 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3206 // Create the counters for this fast lock. 3207 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3208 } 3209 3210 // Create the rtm counters for this fast lock if needed. 3211 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3212 3213 // Add monitor to debug info for the slow path. If we block inside the 3214 // slow path and de-opt, we need the monitor hanging around 3215 map()->push_monitor( flock ); 3216 3217 const TypeFunc *tf = LockNode::lock_type(); 3218 LockNode *lock = new LockNode(C, tf); 3219 3220 lock->init_req( TypeFunc::Control, control() ); 3221 lock->init_req( TypeFunc::Memory , mem ); 3222 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3223 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3224 lock->init_req( TypeFunc::ReturnAdr, top() ); 3225 3226 lock->init_req(TypeFunc::Parms + 0, obj); 3227 lock->init_req(TypeFunc::Parms + 1, box); 3228 lock->init_req(TypeFunc::Parms + 2, flock); 3229 add_safepoint_edges(lock); 3230 3231 lock = _gvn.transform( lock )->as_Lock(); 3232 3233 // lock has no side-effects, sets few values 3234 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3235 3236 insert_mem_bar(Op_MemBarAcquireLock); 3237 3238 // Add this to the worklist so that the lock can be eliminated 3239 record_for_igvn(lock); 3240 3241 #ifndef PRODUCT 3242 if (PrintLockStatistics) { 3243 // Update the counter for this lock. Don't bother using an atomic 3244 // operation since we don't require absolute accuracy. 3245 lock->create_lock_counter(map()->jvms()); 3246 increment_counter(lock->counter()->addr()); 3247 } 3248 #endif 3249 3250 return flock; 3251 } 3252 3253 3254 //------------------------------shared_unlock---------------------------------- 3255 // Emit unlocking code. 3256 void GraphKit::shared_unlock(Node* box, Node* obj) { 3257 // bci is either a monitorenter bc or InvocationEntryBci 3258 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3259 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3260 3261 if( !GenerateSynchronizationCode ) 3262 return; 3263 if (stopped()) { // Dead monitor? 3264 map()->pop_monitor(); // Kill monitor from debug info 3265 return; 3266 } 3267 3268 // Memory barrier to avoid floating things down past the locked region 3269 insert_mem_bar(Op_MemBarReleaseLock); 3270 3271 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3272 UnlockNode *unlock = new UnlockNode(C, tf); 3273 #ifdef ASSERT 3274 unlock->set_dbg_jvms(sync_jvms()); 3275 #endif 3276 uint raw_idx = Compile::AliasIdxRaw; 3277 unlock->init_req( TypeFunc::Control, control() ); 3278 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3279 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3280 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3281 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3282 3283 unlock->init_req(TypeFunc::Parms + 0, obj); 3284 unlock->init_req(TypeFunc::Parms + 1, box); 3285 unlock = _gvn.transform(unlock)->as_Unlock(); 3286 3287 Node* mem = reset_memory(); 3288 3289 // unlock has no side-effects, sets few values 3290 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3291 3292 // Kill monitor from debug info 3293 map()->pop_monitor( ); 3294 } 3295 3296 //-------------------------------get_layout_helper----------------------------- 3297 // If the given klass is a constant or known to be an array, 3298 // fetch the constant layout helper value into constant_value 3299 // and return (Node*)NULL. Otherwise, load the non-constant 3300 // layout helper value, and return the node which represents it. 3301 // This two-faced routine is useful because allocation sites 3302 // almost always feature constant types. 3303 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3304 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3305 if (!StressReflectiveCode && inst_klass != NULL) { 3306 ciKlass* klass = inst_klass->klass(); 3307 bool xklass = inst_klass->klass_is_exact(); 3308 if (xklass || klass->is_array_klass()) { 3309 jint lhelper = klass->layout_helper(); 3310 if (lhelper != Klass::_lh_neutral_value) { 3311 constant_value = lhelper; 3312 return (Node*) NULL; 3313 } 3314 } 3315 } 3316 constant_value = Klass::_lh_neutral_value; // put in a known value 3317 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3318 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3319 } 3320 3321 // We just put in an allocate/initialize with a big raw-memory effect. 3322 // Hook selected additional alias categories on the initialization. 3323 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3324 MergeMemNode* init_in_merge, 3325 Node* init_out_raw) { 3326 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3327 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3328 3329 Node* prevmem = kit.memory(alias_idx); 3330 init_in_merge->set_memory_at(alias_idx, prevmem); 3331 kit.set_memory(init_out_raw, alias_idx); 3332 } 3333 3334 //---------------------------set_output_for_allocation------------------------- 3335 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3336 const TypeOopPtr* oop_type, 3337 bool deoptimize_on_exception) { 3338 int rawidx = Compile::AliasIdxRaw; 3339 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3340 add_safepoint_edges(alloc); 3341 Node* allocx = _gvn.transform(alloc); 3342 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3343 // create memory projection for i_o 3344 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3345 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3346 3347 // create a memory projection as for the normal control path 3348 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3349 set_memory(malloc, rawidx); 3350 3351 // a normal slow-call doesn't change i_o, but an allocation does 3352 // we create a separate i_o projection for the normal control path 3353 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3354 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3355 3356 // put in an initialization barrier 3357 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3358 rawoop)->as_Initialize(); 3359 assert(alloc->initialization() == init, "2-way macro link must work"); 3360 assert(init ->allocation() == alloc, "2-way macro link must work"); 3361 { 3362 // Extract memory strands which may participate in the new object's 3363 // initialization, and source them from the new InitializeNode. 3364 // This will allow us to observe initializations when they occur, 3365 // and link them properly (as a group) to the InitializeNode. 3366 assert(init->in(InitializeNode::Memory) == malloc, ""); 3367 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3368 init->set_req(InitializeNode::Memory, minit_in); 3369 record_for_igvn(minit_in); // fold it up later, if possible 3370 Node* minit_out = memory(rawidx); 3371 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3372 if (oop_type->isa_aryptr()) { 3373 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3374 int elemidx = C->get_alias_index(telemref); 3375 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3376 } else if (oop_type->isa_instptr()) { 3377 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3378 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3379 ciField* field = ik->nonstatic_field_at(i); 3380 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3381 continue; // do not bother to track really large numbers of fields 3382 // Find (or create) the alias category for this field: 3383 int fieldidx = C->alias_type(field)->index(); 3384 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3385 } 3386 } 3387 } 3388 3389 // Cast raw oop to the real thing... 3390 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3391 javaoop = _gvn.transform(javaoop); 3392 C->set_recent_alloc(control(), javaoop); 3393 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3394 3395 #ifdef ASSERT 3396 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3397 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3398 "Ideal_allocation works"); 3399 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3400 "Ideal_allocation works"); 3401 if (alloc->is_AllocateArray()) { 3402 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3403 "Ideal_allocation works"); 3404 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3405 "Ideal_allocation works"); 3406 } else { 3407 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3408 } 3409 } 3410 #endif //ASSERT 3411 3412 return javaoop; 3413 } 3414 3415 //---------------------------new_instance-------------------------------------- 3416 // This routine takes a klass_node which may be constant (for a static type) 3417 // or may be non-constant (for reflective code). It will work equally well 3418 // for either, and the graph will fold nicely if the optimizer later reduces 3419 // the type to a constant. 3420 // The optional arguments are for specialized use by intrinsics: 3421 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3422 // - If 'return_size_val', report the the total object size to the caller. 3423 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3424 Node* GraphKit::new_instance(Node* klass_node, 3425 Node* extra_slow_test, 3426 Node* *return_size_val, 3427 bool deoptimize_on_exception) { 3428 // Compute size in doublewords 3429 // The size is always an integral number of doublewords, represented 3430 // as a positive bytewise size stored in the klass's layout_helper. 3431 // The layout_helper also encodes (in a low bit) the need for a slow path. 3432 jint layout_con = Klass::_lh_neutral_value; 3433 Node* layout_val = get_layout_helper(klass_node, layout_con); 3434 int layout_is_con = (layout_val == NULL); 3435 3436 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3437 // Generate the initial go-slow test. It's either ALWAYS (return a 3438 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3439 // case) a computed value derived from the layout_helper. 3440 Node* initial_slow_test = NULL; 3441 if (layout_is_con) { 3442 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3443 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3444 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3445 } else { // reflective case 3446 // This reflective path is used by Unsafe.allocateInstance. 3447 // (It may be stress-tested by specifying StressReflectiveCode.) 3448 // Basically, we want to get into the VM is there's an illegal argument. 3449 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3450 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3451 if (extra_slow_test != intcon(0)) { 3452 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3453 } 3454 // (Macro-expander will further convert this to a Bool, if necessary.) 3455 } 3456 3457 // Find the size in bytes. This is easy; it's the layout_helper. 3458 // The size value must be valid even if the slow path is taken. 3459 Node* size = NULL; 3460 if (layout_is_con) { 3461 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3462 } else { // reflective case 3463 // This reflective path is used by clone and Unsafe.allocateInstance. 3464 size = ConvI2X(layout_val); 3465 3466 // Clear the low bits to extract layout_helper_size_in_bytes: 3467 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3468 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3469 size = _gvn.transform( new AndXNode(size, mask) ); 3470 } 3471 if (return_size_val != NULL) { 3472 (*return_size_val) = size; 3473 } 3474 3475 // This is a precise notnull oop of the klass. 3476 // (Actually, it need not be precise if this is a reflective allocation.) 3477 // It's what we cast the result to. 3478 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3479 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3480 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3481 3482 // Now generate allocation code 3483 3484 // The entire memory state is needed for slow path of the allocation 3485 // since GC and deoptimization can happened. 3486 Node *mem = reset_memory(); 3487 set_all_memory(mem); // Create new memory state 3488 3489 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3490 control(), mem, i_o(), 3491 size, klass_node, 3492 initial_slow_test); 3493 3494 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3495 } 3496 3497 //-------------------------------new_array------------------------------------- 3498 // helper for both newarray and anewarray 3499 // The 'length' parameter is (obviously) the length of the array. 3500 // See comments on new_instance for the meaning of the other arguments. 3501 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3502 Node* length, // number of array elements 3503 int nargs, // number of arguments to push back for uncommon trap 3504 Node* *return_size_val, 3505 bool deoptimize_on_exception) { 3506 jint layout_con = Klass::_lh_neutral_value; 3507 Node* layout_val = get_layout_helper(klass_node, layout_con); 3508 int layout_is_con = (layout_val == NULL); 3509 3510 if (!layout_is_con && !StressReflectiveCode && 3511 !too_many_traps(Deoptimization::Reason_class_check)) { 3512 // This is a reflective array creation site. 3513 // Optimistically assume that it is a subtype of Object[], 3514 // so that we can fold up all the address arithmetic. 3515 layout_con = Klass::array_layout_helper(T_OBJECT); 3516 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3517 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3518 { BuildCutout unless(this, bol_lh, PROB_MAX); 3519 inc_sp(nargs); 3520 uncommon_trap(Deoptimization::Reason_class_check, 3521 Deoptimization::Action_maybe_recompile); 3522 } 3523 layout_val = NULL; 3524 layout_is_con = true; 3525 } 3526 3527 // Generate the initial go-slow test. Make sure we do not overflow 3528 // if length is huge (near 2Gig) or negative! We do not need 3529 // exact double-words here, just a close approximation of needed 3530 // double-words. We can't add any offset or rounding bits, lest we 3531 // take a size -1 of bytes and make it positive. Use an unsigned 3532 // compare, so negative sizes look hugely positive. 3533 int fast_size_limit = FastAllocateSizeLimit; 3534 if (layout_is_con) { 3535 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3536 // Increase the size limit if we have exact knowledge of array type. 3537 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3538 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3539 } 3540 3541 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3542 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3543 3544 // --- Size Computation --- 3545 // array_size = round_to_heap(array_header + (length << elem_shift)); 3546 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3547 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3548 // The rounding mask is strength-reduced, if possible. 3549 int round_mask = MinObjAlignmentInBytes - 1; 3550 Node* header_size = NULL; 3551 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3552 // (T_BYTE has the weakest alignment and size restrictions...) 3553 if (layout_is_con) { 3554 int hsize = Klass::layout_helper_header_size(layout_con); 3555 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3556 BasicType etype = Klass::layout_helper_element_type(layout_con); 3557 if ((round_mask & ~right_n_bits(eshift)) == 0) 3558 round_mask = 0; // strength-reduce it if it goes away completely 3559 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3560 assert(header_size_min <= hsize, "generic minimum is smallest"); 3561 header_size_min = hsize; 3562 header_size = intcon(hsize + round_mask); 3563 } else { 3564 Node* hss = intcon(Klass::_lh_header_size_shift); 3565 Node* hsm = intcon(Klass::_lh_header_size_mask); 3566 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3567 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3568 Node* mask = intcon(round_mask); 3569 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3570 } 3571 3572 Node* elem_shift = NULL; 3573 if (layout_is_con) { 3574 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3575 if (eshift != 0) 3576 elem_shift = intcon(eshift); 3577 } else { 3578 // There is no need to mask or shift this value. 3579 // The semantics of LShiftINode include an implicit mask to 0x1F. 3580 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3581 elem_shift = layout_val; 3582 } 3583 3584 // Transition to native address size for all offset calculations: 3585 Node* lengthx = ConvI2X(length); 3586 Node* headerx = ConvI2X(header_size); 3587 #ifdef _LP64 3588 { const TypeInt* tilen = _gvn.find_int_type(length); 3589 if (tilen != NULL && tilen->_lo < 0) { 3590 // Add a manual constraint to a positive range. Cf. array_element_address. 3591 jint size_max = fast_size_limit; 3592 if (size_max > tilen->_hi) size_max = tilen->_hi; 3593 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3594 3595 // Only do a narrow I2L conversion if the range check passed. 3596 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3597 _gvn.transform(iff); 3598 RegionNode* region = new RegionNode(3); 3599 _gvn.set_type(region, Type::CONTROL); 3600 lengthx = new PhiNode(region, TypeLong::LONG); 3601 _gvn.set_type(lengthx, TypeLong::LONG); 3602 3603 // Range check passed. Use ConvI2L node with narrow type. 3604 Node* passed = IfFalse(iff); 3605 region->init_req(1, passed); 3606 // Make I2L conversion control dependent to prevent it from 3607 // floating above the range check during loop optimizations. 3608 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3609 3610 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3611 region->init_req(2, IfTrue(iff)); 3612 lengthx->init_req(2, ConvI2X(length)); 3613 3614 set_control(region); 3615 record_for_igvn(region); 3616 record_for_igvn(lengthx); 3617 } 3618 } 3619 #endif 3620 3621 // Combine header size (plus rounding) and body size. Then round down. 3622 // This computation cannot overflow, because it is used only in two 3623 // places, one where the length is sharply limited, and the other 3624 // after a successful allocation. 3625 Node* abody = lengthx; 3626 if (elem_shift != NULL) 3627 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3628 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3629 if (round_mask != 0) { 3630 Node* mask = MakeConX(~round_mask); 3631 size = _gvn.transform( new AndXNode(size, mask) ); 3632 } 3633 // else if round_mask == 0, the size computation is self-rounding 3634 3635 if (return_size_val != NULL) { 3636 // This is the size 3637 (*return_size_val) = size; 3638 } 3639 3640 // Now generate allocation code 3641 3642 // The entire memory state is needed for slow path of the allocation 3643 // since GC and deoptimization can happened. 3644 Node *mem = reset_memory(); 3645 set_all_memory(mem); // Create new memory state 3646 3647 if (initial_slow_test->is_Bool()) { 3648 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3649 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3650 } 3651 3652 // Create the AllocateArrayNode and its result projections 3653 AllocateArrayNode* alloc 3654 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3655 control(), mem, i_o(), 3656 size, klass_node, 3657 initial_slow_test, 3658 length); 3659 3660 // Cast to correct type. Note that the klass_node may be constant or not, 3661 // and in the latter case the actual array type will be inexact also. 3662 // (This happens via a non-constant argument to inline_native_newArray.) 3663 // In any case, the value of klass_node provides the desired array type. 3664 const TypeInt* length_type = _gvn.find_int_type(length); 3665 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3666 if (ary_type->isa_aryptr() && length_type != NULL) { 3667 // Try to get a better type than POS for the size 3668 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3669 } 3670 3671 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3672 3673 // Cast length on remaining path to be as narrow as possible 3674 if (map()->find_edge(length) >= 0) { 3675 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3676 if (ccast != length) { 3677 _gvn.set_type_bottom(ccast); 3678 record_for_igvn(ccast); 3679 replace_in_map(length, ccast); 3680 } 3681 } 3682 3683 return javaoop; 3684 } 3685 3686 // The following "Ideal_foo" functions are placed here because they recognize 3687 // the graph shapes created by the functions immediately above. 3688 3689 //---------------------------Ideal_allocation---------------------------------- 3690 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3691 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3692 if (ptr == NULL) { // reduce dumb test in callers 3693 return NULL; 3694 } 3695 3696 // Attempt to see through Shenandoah barriers. 3697 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr); 3698 3699 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3700 ptr = ptr->in(1); 3701 if (ptr == NULL) return NULL; 3702 } 3703 // Return NULL for allocations with several casts: 3704 // j.l.reflect.Array.newInstance(jobject, jint) 3705 // Object.clone() 3706 // to keep more precise type from last cast. 3707 if (ptr->is_Proj()) { 3708 Node* allo = ptr->in(0); 3709 if (allo != NULL && allo->is_Allocate()) { 3710 return allo->as_Allocate(); 3711 } 3712 } 3713 // Report failure to match. 3714 return NULL; 3715 } 3716 3717 // Fancy version which also strips off an offset (and reports it to caller). 3718 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3719 intptr_t& offset) { 3720 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3721 if (base == NULL) return NULL; 3722 return Ideal_allocation(base, phase); 3723 } 3724 3725 // Trace Initialize <- Proj[Parm] <- Allocate 3726 AllocateNode* InitializeNode::allocation() { 3727 Node* rawoop = in(InitializeNode::RawAddress); 3728 if (rawoop->is_Proj()) { 3729 Node* alloc = rawoop->in(0); 3730 if (alloc->is_Allocate()) { 3731 return alloc->as_Allocate(); 3732 } 3733 } 3734 return NULL; 3735 } 3736 3737 // Trace Allocate -> Proj[Parm] -> Initialize 3738 InitializeNode* AllocateNode::initialization() { 3739 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3740 if (rawoop == NULL) return NULL; 3741 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3742 Node* init = rawoop->fast_out(i); 3743 if (init->is_Initialize()) { 3744 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3745 return init->as_Initialize(); 3746 } 3747 } 3748 return NULL; 3749 } 3750 3751 //----------------------------- loop predicates --------------------------- 3752 3753 //------------------------------add_predicate_impl---------------------------- 3754 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3755 // Too many traps seen? 3756 if (too_many_traps(reason)) { 3757 #ifdef ASSERT 3758 if (TraceLoopPredicate) { 3759 int tc = C->trap_count(reason); 3760 tty->print("too many traps=%s tcount=%d in ", 3761 Deoptimization::trap_reason_name(reason), tc); 3762 method()->print(); // which method has too many predicate traps 3763 tty->cr(); 3764 } 3765 #endif 3766 // We cannot afford to take more traps here, 3767 // do not generate predicate. 3768 return; 3769 } 3770 3771 Node *cont = _gvn.intcon(1); 3772 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3773 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3774 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3775 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3776 C->add_predicate_opaq(opq); 3777 { 3778 PreserveJVMState pjvms(this); 3779 set_control(iffalse); 3780 inc_sp(nargs); 3781 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3782 } 3783 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3784 set_control(iftrue); 3785 } 3786 3787 //------------------------------add_predicate--------------------------------- 3788 void GraphKit::add_predicate(int nargs) { 3789 if (UseLoopPredicate) { 3790 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3791 } 3792 // loop's limit check predicate should be near the loop. 3793 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3794 } 3795 3796 //----------------------------- store barriers ---------------------------- 3797 #define __ ideal. 3798 3799 void GraphKit::sync_kit(IdealKit& ideal) { 3800 set_all_memory(__ merged_memory()); 3801 set_i_o(__ i_o()); 3802 set_control(__ ctrl()); 3803 } 3804 3805 void GraphKit::final_sync(IdealKit& ideal) { 3806 // Final sync IdealKit and graphKit. 3807 sync_kit(ideal); 3808 } 3809 3810 Node* GraphKit::byte_map_base_node() { 3811 // Get base of card map 3812 CardTableModRefBS* ct = 3813 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3814 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3815 if (ct->byte_map_base != NULL) { 3816 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3817 } else { 3818 return null(); 3819 } 3820 } 3821 3822 // vanilla/CMS post barrier 3823 // Insert a write-barrier store. This is to let generational GC work; we have 3824 // to flag all oop-stores before the next GC point. 3825 void GraphKit::write_barrier_post(Node* oop_store, 3826 Node* obj, 3827 Node* adr, 3828 uint adr_idx, 3829 Node* val, 3830 bool use_precise) { 3831 // No store check needed if we're storing a NULL or an old object 3832 // (latter case is probably a string constant). The concurrent 3833 // mark sweep garbage collector, however, needs to have all nonNull 3834 // oop updates flagged via card-marks. 3835 if (val != NULL && val->is_Con()) { 3836 // must be either an oop or NULL 3837 const Type* t = val->bottom_type(); 3838 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3839 // stores of null never (?) need barriers 3840 return; 3841 } 3842 3843 if (use_ReduceInitialCardMarks() 3844 && obj == just_allocated_object(control())) { 3845 // We can skip marks on a freshly-allocated object in Eden. 3846 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3847 // That routine informs GC to take appropriate compensating steps, 3848 // upon a slow-path allocation, so as to make this card-mark 3849 // elision safe. 3850 return; 3851 } 3852 3853 if (!use_precise) { 3854 // All card marks for a (non-array) instance are in one place: 3855 adr = obj; 3856 } 3857 // (Else it's an array (or unknown), and we want more precise card marks.) 3858 assert(adr != NULL, ""); 3859 3860 IdealKit ideal(this, true); 3861 3862 // Convert the pointer to an int prior to doing math on it 3863 Node* cast = __ CastPX(__ ctrl(), adr); 3864 3865 // Divide by card size 3866 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3867 "Only one we handle so far."); 3868 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3869 3870 // Combine card table base and card offset 3871 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3872 3873 // Get the alias_index for raw card-mark memory 3874 int adr_type = Compile::AliasIdxRaw; 3875 Node* zero = __ ConI(0); // Dirty card value 3876 BasicType bt = T_BYTE; 3877 3878 if (UseConcMarkSweepGC && UseCondCardMark) { 3879 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3880 __ sync_kit(this); 3881 } 3882 3883 if (UseCondCardMark) { 3884 // The classic GC reference write barrier is typically implemented 3885 // as a store into the global card mark table. Unfortunately 3886 // unconditional stores can result in false sharing and excessive 3887 // coherence traffic as well as false transactional aborts. 3888 // UseCondCardMark enables MP "polite" conditional card mark 3889 // stores. In theory we could relax the load from ctrl() to 3890 // no_ctrl, but that doesn't buy much latitude. 3891 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3892 __ if_then(card_val, BoolTest::ne, zero); 3893 } 3894 3895 // Smash zero into card 3896 if( !UseConcMarkSweepGC ) { 3897 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3898 } else { 3899 // Specialized path for CM store barrier 3900 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3901 } 3902 3903 if (UseCondCardMark) { 3904 __ end_if(); 3905 } 3906 3907 // Final sync IdealKit and GraphKit. 3908 final_sync(ideal); 3909 } 3910 /* 3911 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3912 * required by SATB to make sure all objects live at the start of the 3913 * marking are kept alive, all reference updates need to any previous 3914 * reference stored before writing. 3915 * 3916 * If the previous value is NULL there is no need to save the old value. 3917 * References that are NULL are filtered during runtime by the barrier 3918 * code to avoid unnecessary queuing. 3919 * 3920 * However in the case of newly allocated objects it might be possible to 3921 * prove that the reference about to be overwritten is NULL during compile 3922 * time and avoid adding the barrier code completely. 3923 * 3924 * The compiler needs to determine that the object in which a field is about 3925 * to be written is newly allocated, and that no prior store to the same field 3926 * has happened since the allocation. 3927 * 3928 * Returns true if the pre-barrier can be removed 3929 */ 3930 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3931 BasicType bt, uint adr_idx) { 3932 intptr_t offset = 0; 3933 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3934 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3935 3936 if (offset == Type::OffsetBot) { 3937 return false; // cannot unalias unless there are precise offsets 3938 } 3939 3940 if (alloc == NULL) { 3941 return false; // No allocation found 3942 } 3943 3944 intptr_t size_in_bytes = type2aelembytes(bt); 3945 3946 Node* mem = memory(adr_idx); // start searching here... 3947 3948 for (int cnt = 0; cnt < 50; cnt++) { 3949 3950 if (mem->is_Store()) { 3951 3952 Node* st_adr = mem->in(MemNode::Address); 3953 intptr_t st_offset = 0; 3954 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3955 3956 if (st_base == NULL) { 3957 break; // inscrutable pointer 3958 } 3959 3960 // Break we have found a store with same base and offset as ours so break 3961 if (st_base == base && st_offset == offset) { 3962 break; 3963 } 3964 3965 if (st_offset != offset && st_offset != Type::OffsetBot) { 3966 const int MAX_STORE = BytesPerLong; 3967 if (st_offset >= offset + size_in_bytes || 3968 st_offset <= offset - MAX_STORE || 3969 st_offset <= offset - mem->as_Store()->memory_size()) { 3970 // Success: The offsets are provably independent. 3971 // (You may ask, why not just test st_offset != offset and be done? 3972 // The answer is that stores of different sizes can co-exist 3973 // in the same sequence of RawMem effects. We sometimes initialize 3974 // a whole 'tile' of array elements with a single jint or jlong.) 3975 mem = mem->in(MemNode::Memory); 3976 continue; // advance through independent store memory 3977 } 3978 } 3979 3980 if (st_base != base 3981 && MemNode::detect_ptr_independence(base, alloc, st_base, 3982 AllocateNode::Ideal_allocation(st_base, phase), 3983 phase)) { 3984 // Success: The bases are provably independent. 3985 mem = mem->in(MemNode::Memory); 3986 continue; // advance through independent store memory 3987 } 3988 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3989 3990 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3991 AllocateNode* st_alloc = st_init->allocation(); 3992 3993 // Make sure that we are looking at the same allocation site. 3994 // The alloc variable is guaranteed to not be null here from earlier check. 3995 if (alloc == st_alloc) { 3996 // Check that the initialization is storing NULL so that no previous store 3997 // has been moved up and directly write a reference 3998 Node* captured_store = st_init->find_captured_store(offset, 3999 type2aelembytes(T_OBJECT), 4000 phase); 4001 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4002 return true; 4003 } 4004 } 4005 } 4006 4007 // Unless there is an explicit 'continue', we must bail out here, 4008 // because 'mem' is an inscrutable memory state (e.g., a call). 4009 break; 4010 } 4011 4012 return false; 4013 } 4014 4015 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) { 4016 if (UseShenandoahGC && ShenandoahWriteBarrier && adr != NULL) { 4017 Node* c = kit.control(); 4018 Node* call = c->in(1)->in(1)->in(1)->in(0); 4019 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected"); 4020 call->add_req(adr); 4021 } 4022 } 4023 4024 // G1 pre/post barriers 4025 void GraphKit::g1_write_barrier_pre(bool do_load, 4026 Node* obj, 4027 Node* adr, 4028 uint alias_idx, 4029 Node* val, 4030 const TypeOopPtr* val_type, 4031 Node* pre_val, 4032 BasicType bt) { 4033 4034 // Some sanity checks 4035 // Note: val is unused in this routine. 4036 4037 if (do_load) { 4038 // We need to generate the load of the previous value 4039 assert(obj != NULL, "must have a base"); 4040 assert(adr != NULL, "where are loading from?"); 4041 assert(pre_val == NULL, "loaded already?"); 4042 assert(val_type != NULL, "need a type"); 4043 4044 if (use_ReduceInitialCardMarks() 4045 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4046 return; 4047 } 4048 4049 } else { 4050 // In this case both val_type and alias_idx are unused. 4051 assert(pre_val != NULL, "must be loaded already"); 4052 // Nothing to be done if pre_val is null. 4053 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4054 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4055 } 4056 assert(bt == T_OBJECT, "or we shouldn't be here"); 4057 4058 IdealKit ideal(this, true); 4059 4060 Node* tls = __ thread(); // ThreadLocalStorage 4061 4062 Node* no_ctrl = NULL; 4063 Node* no_base = __ top(); 4064 Node* zero = __ ConI(0); 4065 Node* zeroX = __ ConX(0); 4066 4067 float likely = PROB_LIKELY(0.999); 4068 float unlikely = PROB_UNLIKELY(0.999); 4069 4070 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4071 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4072 4073 // Offsets into the thread 4074 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4075 SATBMarkQueue::byte_offset_of_active()); 4076 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4077 SATBMarkQueue::byte_offset_of_index()); 4078 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4079 SATBMarkQueue::byte_offset_of_buf()); 4080 4081 // Now the actual pointers into the thread 4082 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4083 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4084 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4085 4086 // Now some of the values 4087 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4088 4089 // if (!marking) 4090 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4091 BasicType index_bt = TypeX_X->basic_type(); 4092 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4093 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4094 4095 if (do_load) { 4096 // load original value 4097 // alias_idx correct?? 4098 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4099 } 4100 4101 // if (pre_val != NULL) 4102 __ if_then(pre_val, BoolTest::ne, null()); { 4103 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4104 4105 // is the queue for this thread full? 4106 __ if_then(index, BoolTest::ne, zeroX, likely); { 4107 4108 // decrement the index 4109 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4110 4111 // Now get the buffer location we will log the previous value into and store it 4112 Node *log_addr = __ AddP(no_base, buffer, next_index); 4113 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4114 // update the index 4115 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4116 4117 } __ else_(); { 4118 4119 // logging buffer is full, call the runtime 4120 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4121 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4122 } __ end_if(); // (!index) 4123 } __ end_if(); // (pre_val != NULL) 4124 } __ end_if(); // (!marking) 4125 4126 // Final sync IdealKit and GraphKit. 4127 final_sync(ideal); 4128 g1_write_barrier_pre_helper(*this, adr); 4129 } 4130 4131 Node* GraphKit::shenandoah_write_barrier_pre(bool do_load, 4132 Node* obj, 4133 Node* adr, 4134 uint alias_idx, 4135 Node* val, 4136 const TypeOopPtr* val_type, 4137 Node* pre_val, 4138 BasicType bt) { 4139 4140 // Some sanity checks 4141 // Note: val is unused in this routine. 4142 4143 if (val == NULL) { 4144 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 4145 return NULL; 4146 } 4147 4148 if (! ShenandoahReduceStoreValBarrier) { 4149 val = shenandoah_read_barrier_storeval(val); 4150 shenandoah_update_matrix(adr, val); 4151 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 4152 return val; 4153 } 4154 4155 if (do_load) { 4156 // We need to generate the load of the previous value 4157 assert(obj != NULL, "must have a base"); 4158 assert(adr != NULL, "where are loading from?"); 4159 assert(pre_val == NULL, "loaded already?"); 4160 assert(val_type != NULL, "need a type"); 4161 4162 if (use_ReduceInitialCardMarks() 4163 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4164 return shenandoah_read_barrier_storeval(val); 4165 } 4166 4167 } else { 4168 // In this case both val_type and alias_idx are unused. 4169 assert(pre_val != NULL, "must be loaded already"); 4170 // Nothing to be done if pre_val is null. 4171 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return val; 4172 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4173 } 4174 assert(bt == T_OBJECT, "or we shouldn't be here"); 4175 4176 IdealKit ideal(this, true, true); 4177 IdealVariable ival(ideal); 4178 __ declarations_done(); 4179 __ set(ival, val); 4180 Node* tls = __ thread(); // ThreadLocalStorage 4181 4182 Node* no_ctrl = NULL; 4183 Node* no_base = __ top(); 4184 Node* zero = __ ConI(0); 4185 Node* zeroX = __ ConX(0); 4186 4187 float likely = PROB_LIKELY(0.999); 4188 float unlikely = PROB_UNLIKELY(0.999); 4189 4190 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4191 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4192 4193 // Offsets into the thread 4194 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4195 SATBMarkQueue::byte_offset_of_active()); 4196 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4197 SATBMarkQueue::byte_offset_of_index()); 4198 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4199 SATBMarkQueue::byte_offset_of_buf()); 4200 4201 // Now the actual pointers into the thread 4202 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4203 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4204 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4205 4206 // Now some of the values 4207 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4208 4209 // if (!marking) 4210 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4211 4212 Node* storeval = ideal.value(ival); 4213 sync_kit(ideal); 4214 storeval = shenandoah_read_barrier_storeval(storeval); 4215 __ sync_kit(this); 4216 __ set(ival, storeval); 4217 4218 BasicType index_bt = TypeX_X->basic_type(); 4219 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4220 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4221 4222 if (do_load) { 4223 // load original value 4224 // alias_idx correct?? 4225 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4226 } 4227 4228 // if (pre_val != NULL) 4229 __ if_then(pre_val, BoolTest::ne, null()); { 4230 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4231 4232 // is the queue for this thread full? 4233 __ if_then(index, BoolTest::ne, zeroX, likely); { 4234 4235 // decrement the index 4236 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4237 4238 // Now get the buffer location we will log the previous value into and store it 4239 Node *log_addr = __ AddP(no_base, buffer, next_index); 4240 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4241 // update the index 4242 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4243 4244 } __ else_(); { 4245 4246 // logging buffer is full, call the runtime 4247 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4248 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4249 } __ end_if(); // (!index) 4250 } __ end_if(); // (pre_val != NULL) 4251 } __ end_if(); // (!marking) 4252 Node* new_val = __ value(ival); 4253 // IdealKit generates a Phi with very conservative type, and even 4254 // turns array types into TypeInstPtr (see type.cpp, _const_basic_type[T_ARRAY]). 4255 // We're forcing the result to be the original type. 4256 if (new_val != val) { 4257 const Type* t = _gvn.type(val); 4258 if (new_val->isa_Type()) { 4259 new_val->as_Type()->set_type(t); 4260 } 4261 _gvn.set_type(new_val, t); 4262 } 4263 val = new_val; 4264 __ dead(ival); 4265 // Final sync IdealKit and GraphKit. 4266 final_sync(ideal); 4267 g1_write_barrier_pre_helper(*this, adr); 4268 return val; 4269 } 4270 4271 void GraphKit::shenandoah_update_matrix(Node* adr, Node* val) { 4272 assert(val != NULL, "checked before"); 4273 if (adr == NULL) { 4274 return; // Nothing to do 4275 } 4276 assert(adr != NULL, "must not happen"); 4277 if (val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) { 4278 // Nothing to do. 4279 return; 4280 } 4281 4282 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix(); 4283 4284 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4285 RegionNode* region = new RegionNode(PATH_LIMIT); 4286 Node* prev_mem = memory(Compile::AliasIdxRaw); 4287 Node* memphi = PhiNode::make(region, prev_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 4288 Node* null_ctrl = top(); 4289 Node* not_null_val = null_check_oop(val, &null_ctrl); 4290 4291 // Null path: nothing to do. 4292 region->init_req(_null_path, null_ctrl); 4293 memphi->init_req(_null_path, prev_mem); 4294 4295 // Not null path. Update the matrix. 4296 Node* heapbase = MakeConX((intx) ShenandoahHeap::heap()->first_region_bottom()); 4297 Node* region_size_shift = intcon((jint) ShenandoahHeapRegion::RegionSizeShift); 4298 Node* stride = MakeConX((intx) matrix->stride()); 4299 Node* matrix_base = makecon(TypeRawPtr::make(matrix->matrix_addr())); 4300 // Compute region index for adr. 4301 Node* adr_idx = _gvn.transform(new CastP2XNode(control(), adr)); 4302 adr_idx = _gvn.transform(new SubXNode(adr_idx, heapbase)); 4303 adr_idx = _gvn.transform(new URShiftXNode(adr_idx, region_size_shift)); 4304 // Compute region index for val. 4305 Node* val_idx = _gvn.transform(new CastP2XNode(control(), not_null_val)); 4306 val_idx = _gvn.transform(new SubXNode(val_idx, heapbase)); 4307 val_idx = _gvn.transform(new URShiftXNode(val_idx, region_size_shift)); 4308 // Compute matrix index & address. 4309 Node* matrix_idx = _gvn.transform(new MulXNode(adr_idx, stride)); 4310 matrix_idx = _gvn.transform(new AddXNode(matrix_idx, val_idx)); 4311 Node* matrix_adr = basic_plus_adr(top(), matrix_base, matrix_idx); 4312 // Do the store. 4313 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 4314 Node* store = _gvn.transform(StoreNode::make(_gvn, control(), memory(Compile::AliasIdxRaw), matrix_adr, adr_type, intcon(1), T_BOOLEAN, MemNode::unordered)); 4315 4316 region->init_req(_not_null_path, control()); 4317 memphi->init_req(_not_null_path, store); 4318 4319 // Merge control flows and memory. 4320 set_control(_gvn.transform(region)); 4321 record_for_igvn(region); 4322 set_memory(_gvn.transform(memphi), Compile::AliasIdxRaw); 4323 } 4324 4325 /* 4326 * G1 similar to any GC with a Young Generation requires a way to keep track of 4327 * references from Old Generation to Young Generation to make sure all live 4328 * objects are found. G1 also requires to keep track of object references 4329 * between different regions to enable evacuation of old regions, which is done 4330 * as part of mixed collections. References are tracked in remembered sets and 4331 * is continuously updated as reference are written to with the help of the 4332 * post-barrier. 4333 * 4334 * To reduce the number of updates to the remembered set the post-barrier 4335 * filters updates to fields in objects located in the Young Generation, 4336 * the same region as the reference, when the NULL is being written or 4337 * if the card is already marked as dirty by an earlier write. 4338 * 4339 * Under certain circumstances it is possible to avoid generating the 4340 * post-barrier completely if it is possible during compile time to prove 4341 * the object is newly allocated and that no safepoint exists between the 4342 * allocation and the store. 4343 * 4344 * In the case of slow allocation the allocation code must handle the barrier 4345 * as part of the allocation in the case the allocated object is not located 4346 * in the nursery, this would happen for humongous objects. This is similar to 4347 * how CMS is required to handle this case, see the comments for the method 4348 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4349 * A deferred card mark is required for these objects and handled in the above 4350 * mentioned methods. 4351 * 4352 * Returns true if the post barrier can be removed 4353 */ 4354 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4355 Node* adr) { 4356 intptr_t offset = 0; 4357 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4358 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4359 4360 if (offset == Type::OffsetBot) { 4361 return false; // cannot unalias unless there are precise offsets 4362 } 4363 4364 if (alloc == NULL) { 4365 return false; // No allocation found 4366 } 4367 4368 // Start search from Store node 4369 Node* mem = store->in(MemNode::Control); 4370 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4371 4372 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4373 AllocateNode* st_alloc = st_init->allocation(); 4374 4375 // Make sure we are looking at the same allocation 4376 if (alloc == st_alloc) { 4377 return true; 4378 } 4379 } 4380 4381 return false; 4382 } 4383 4384 // 4385 // Update the card table and add card address to the queue 4386 // 4387 void GraphKit::g1_mark_card(IdealKit& ideal, 4388 Node* card_adr, 4389 Node* oop_store, 4390 uint oop_alias_idx, 4391 Node* index, 4392 Node* index_adr, 4393 Node* buffer, 4394 const TypeFunc* tf) { 4395 4396 Node* zero = __ ConI(0); 4397 Node* zeroX = __ ConX(0); 4398 Node* no_base = __ top(); 4399 BasicType card_bt = T_BYTE; 4400 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4401 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4402 4403 // Now do the queue work 4404 __ if_then(index, BoolTest::ne, zeroX); { 4405 4406 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4407 Node* log_addr = __ AddP(no_base, buffer, next_index); 4408 4409 // Order, see storeCM. 4410 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4411 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4412 4413 } __ else_(); { 4414 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4415 } __ end_if(); 4416 4417 } 4418 4419 void GraphKit::g1_write_barrier_post(Node* oop_store, 4420 Node* obj, 4421 Node* adr, 4422 uint alias_idx, 4423 Node* val, 4424 BasicType bt, 4425 bool use_precise) { 4426 // If we are writing a NULL then we need no post barrier 4427 4428 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4429 // Must be NULL 4430 const Type* t = val->bottom_type(); 4431 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4432 // No post barrier if writing NULLx 4433 return; 4434 } 4435 4436 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4437 // We can skip marks on a freshly-allocated object in Eden. 4438 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4439 // That routine informs GC to take appropriate compensating steps, 4440 // upon a slow-path allocation, so as to make this card-mark 4441 // elision safe. 4442 return; 4443 } 4444 4445 if (use_ReduceInitialCardMarks() 4446 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4447 return; 4448 } 4449 4450 if (!use_precise) { 4451 // All card marks for a (non-array) instance are in one place: 4452 adr = obj; 4453 } 4454 // (Else it's an array (or unknown), and we want more precise card marks.) 4455 assert(adr != NULL, ""); 4456 4457 IdealKit ideal(this, true); 4458 4459 Node* tls = __ thread(); // ThreadLocalStorage 4460 4461 Node* no_base = __ top(); 4462 float likely = PROB_LIKELY(0.999); 4463 float unlikely = PROB_UNLIKELY(0.999); 4464 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4465 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4466 Node* zeroX = __ ConX(0); 4467 4468 // Get the alias_index for raw card-mark memory 4469 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4470 4471 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4472 4473 // Offsets into the thread 4474 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4475 DirtyCardQueue::byte_offset_of_index()); 4476 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4477 DirtyCardQueue::byte_offset_of_buf()); 4478 4479 // Pointers into the thread 4480 4481 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4482 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4483 4484 // Now some values 4485 // Use ctrl to avoid hoisting these values past a safepoint, which could 4486 // potentially reset these fields in the JavaThread. 4487 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4488 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4489 4490 // Convert the store obj pointer to an int prior to doing math on it 4491 // Must use ctrl to prevent "integerized oop" existing across safepoint 4492 Node* cast = __ CastPX(__ ctrl(), adr); 4493 4494 // Divide pointer by card size 4495 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4496 4497 // Combine card table base and card offset 4498 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4499 4500 // If we know the value being stored does it cross regions? 4501 4502 if (val != NULL) { 4503 // Does the store cause us to cross regions? 4504 4505 // Should be able to do an unsigned compare of region_size instead of 4506 // and extra shift. Do we have an unsigned compare?? 4507 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4508 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4509 4510 // if (xor_res == 0) same region so skip 4511 __ if_then(xor_res, BoolTest::ne, zeroX); { 4512 4513 // No barrier if we are storing a NULL 4514 __ if_then(val, BoolTest::ne, null(), unlikely); { 4515 4516 // Ok must mark the card if not already dirty 4517 4518 // load the original value of the card 4519 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4520 4521 __ if_then(card_val, BoolTest::ne, young_card); { 4522 sync_kit(ideal); 4523 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4524 insert_mem_bar(Op_MemBarVolatile, oop_store); 4525 __ sync_kit(this); 4526 4527 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4528 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4529 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4530 } __ end_if(); 4531 } __ end_if(); 4532 } __ end_if(); 4533 } __ end_if(); 4534 } else { 4535 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4536 // We don't need a barrier here if the destination is a newly allocated object 4537 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4538 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4539 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4540 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4541 __ if_then(card_val, BoolTest::ne, young_card); { 4542 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4543 } __ end_if(); 4544 } 4545 4546 // Final sync IdealKit and GraphKit. 4547 final_sync(ideal); 4548 } 4549 #undef __ 4550 4551 4552 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4553 Node* len = load_array_length(load_String_value(ctrl, str)); 4554 Node* coder = load_String_coder(ctrl, str); 4555 // Divide length by 2 if coder is UTF16 4556 return _gvn.transform(new RShiftINode(len, coder)); 4557 } 4558 4559 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4560 int value_offset = java_lang_String::value_offset_in_bytes(); 4561 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4562 false, NULL, 0); 4563 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4564 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4565 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4566 ciTypeArrayKlass::make(T_BYTE), true, 0); 4567 int value_field_idx = C->get_alias_index(value_field_type); 4568 4569 if (! ShenandoahOptimizeFinals) { 4570 str = shenandoah_read_barrier(str); 4571 } 4572 4573 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4574 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4575 // String.value field is known to be @Stable. 4576 if (UseImplicitStableValues) { 4577 load = cast_array_to_stable(load, value_type); 4578 } 4579 return load; 4580 } 4581 4582 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4583 if (!CompactStrings) { 4584 return intcon(java_lang_String::CODER_UTF16); 4585 } 4586 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4587 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4588 false, NULL, 0); 4589 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4590 int coder_field_idx = C->get_alias_index(coder_field_type); 4591 4592 if (! ShenandoahOptimizeFinals) { 4593 str = shenandoah_read_barrier(str); 4594 } 4595 4596 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4597 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4598 } 4599 4600 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4601 int value_offset = java_lang_String::value_offset_in_bytes(); 4602 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4603 false, NULL, 0); 4604 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4605 4606 str = shenandoah_write_barrier(str); 4607 4608 store_oop_to_object(control(), str, basic_plus_adr(str, value_offset), value_field_type, 4609 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4610 } 4611 4612 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4613 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4614 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4615 false, NULL, 0); 4616 4617 str = shenandoah_write_barrier(str); 4618 4619 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4620 int coder_field_idx = C->get_alias_index(coder_field_type); 4621 store_to_memory(control(), basic_plus_adr(str, coder_offset), 4622 value, T_BYTE, coder_field_idx, MemNode::unordered); 4623 } 4624 4625 // Capture src and dst memory state with a MergeMemNode 4626 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4627 if (src_type == dst_type) { 4628 // Types are equal, we don't need a MergeMemNode 4629 return memory(src_type); 4630 } 4631 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4632 record_for_igvn(merge); // fold it up later, if possible 4633 int src_idx = C->get_alias_index(src_type); 4634 int dst_idx = C->get_alias_index(dst_type); 4635 merge->set_memory_at(src_idx, memory(src_idx)); 4636 merge->set_memory_at(dst_idx, memory(dst_idx)); 4637 return merge; 4638 } 4639 4640 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4641 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4642 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4643 // If input and output memory types differ, capture both states to preserve 4644 // the dependency between preceding and subsequent loads/stores. 4645 // For example, the following program: 4646 // StoreB 4647 // compress_string 4648 // LoadB 4649 // has this memory graph (use->def): 4650 // LoadB -> compress_string -> CharMem 4651 // ... -> StoreB -> ByteMem 4652 // The intrinsic hides the dependency between LoadB and StoreB, causing 4653 // the load to read from memory not containing the result of the StoreB. 4654 // The correct memory graph should look like this: 4655 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4656 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4657 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4658 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4659 set_memory(res_mem, TypeAryPtr::BYTES); 4660 return str; 4661 } 4662 4663 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4664 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4665 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4666 // Capture src and dst memory (see comment in 'compress_string'). 4667 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4668 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4669 set_memory(_gvn.transform(str), dst_type); 4670 } 4671 4672 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4673 4674 src = shenandoah_read_barrier(src); 4675 dst = shenandoah_write_barrier(dst); 4676 4677 /** 4678 * int i_char = start; 4679 * for (int i_byte = 0; i_byte < count; i_byte++) { 4680 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4681 * } 4682 */ 4683 add_predicate(); 4684 RegionNode* head = new RegionNode(3); 4685 head->init_req(1, control()); 4686 gvn().set_type(head, Type::CONTROL); 4687 record_for_igvn(head); 4688 4689 Node* i_byte = new PhiNode(head, TypeInt::INT); 4690 i_byte->init_req(1, intcon(0)); 4691 gvn().set_type(i_byte, TypeInt::INT); 4692 record_for_igvn(i_byte); 4693 4694 Node* i_char = new PhiNode(head, TypeInt::INT); 4695 i_char->init_req(1, start); 4696 gvn().set_type(i_char, TypeInt::INT); 4697 record_for_igvn(i_char); 4698 4699 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4700 gvn().set_type(mem, Type::MEMORY); 4701 record_for_igvn(mem); 4702 set_control(head); 4703 set_memory(mem, TypeAryPtr::BYTES); 4704 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4705 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4706 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4707 false, false, true /* mismatched */); 4708 4709 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4710 head->init_req(2, IfTrue(iff)); 4711 mem->init_req(2, st); 4712 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4713 i_char->init_req(2, AddI(i_char, intcon(2))); 4714 4715 set_control(IfFalse(iff)); 4716 set_memory(st, TypeAryPtr::BYTES); 4717 } 4718 4719 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4720 if (!field->is_constant()) { 4721 return NULL; // Field not marked as constant. 4722 } 4723 ciInstance* holder = NULL; 4724 if (!field->is_static()) { 4725 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4726 if (const_oop != NULL && const_oop->is_instance()) { 4727 holder = const_oop->as_instance(); 4728 } 4729 } 4730 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4731 /*is_unsigned_load=*/false); 4732 if (con_type != NULL) { 4733 return makecon(con_type); 4734 } 4735 return NULL; 4736 } 4737 4738 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4739 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4740 // assumption of CCP analysis. 4741 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4742 } 4743 4744 Node* GraphKit::shenandoah_read_barrier(Node* obj) { 4745 return shenandoah_read_barrier_impl(obj, false, true, true); 4746 } 4747 4748 Node* GraphKit::shenandoah_read_barrier_storeval(Node* obj) { 4749 return shenandoah_read_barrier_impl(obj, true, false, false); 4750 } 4751 4752 Node* GraphKit::shenandoah_read_barrier_impl(Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) { 4753 4754 if (UseShenandoahGC && ShenandoahReadBarrier) { 4755 const Type* obj_type = obj->bottom_type(); 4756 if (obj_type->higher_equal(TypePtr::NULL_PTR)) { 4757 return obj; 4758 } 4759 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 4760 Node* mem = use_mem ? memory(adr_type) : immutable_memory(); 4761 4762 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, mem, allow_fromspace)) { 4763 // We know it is null, no barrier needed. 4764 return obj; 4765 } 4766 4767 4768 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4769 4770 // We don't know if it's null or not. Need null-check. 4771 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4772 RegionNode* region = new RegionNode(PATH_LIMIT); 4773 Node* phi = new PhiNode(region, obj_type); 4774 Node* null_ctrl = top(); 4775 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4776 4777 region->init_req(_null_path, null_ctrl); 4778 phi ->init_req(_null_path, zerocon(T_OBJECT)); 4779 4780 Node* ctrl = use_ctrl ? control() : NULL; 4781 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace); 4782 Node* n = _gvn.transform(rb); 4783 4784 region->init_req(_not_null_path, control()); 4785 phi ->init_req(_not_null_path, n); 4786 4787 set_control(_gvn.transform(region)); 4788 record_for_igvn(region); 4789 return _gvn.transform(phi); 4790 4791 } else { 4792 // We know it is not null. Simple barrier is sufficient. 4793 Node* ctrl = use_ctrl ? control() : NULL; 4794 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace); 4795 Node* n = _gvn.transform(rb); 4796 record_for_igvn(n); 4797 return n; 4798 } 4799 4800 } else { 4801 return obj; 4802 } 4803 } 4804 4805 static Node* shenandoah_write_barrier_helper(GraphKit& kit, Node* obj, const TypePtr* adr_type) { 4806 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit.C, kit.control(), kit.memory(adr_type), obj); 4807 Node* n = kit.gvn().transform(wb); 4808 if (n == wb) { // New barrier needs memory projection. 4809 Node* proj = kit.gvn().transform(new ShenandoahWBMemProjNode(n)); 4810 kit.set_memory(proj, adr_type); 4811 } 4812 4813 return n; 4814 } 4815 4816 Node* GraphKit::shenandoah_write_barrier(Node* obj) { 4817 4818 if (UseShenandoahGC && ShenandoahWriteBarrier) { 4819 4820 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, NULL, true)) { 4821 return obj; 4822 } 4823 const Type* obj_type = obj->bottom_type(); 4824 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 4825 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4826 // We don't know if it's null or not. Need null-check. 4827 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4828 RegionNode* region = new RegionNode(PATH_LIMIT); 4829 Node* phi = new PhiNode(region, obj_type); 4830 Node* memphi = PhiNode::make(region, memory(adr_type), Type::MEMORY, C->alias_type(adr_type)->adr_type()); 4831 4832 Node* prev_mem = memory(adr_type); 4833 Node* null_ctrl = top(); 4834 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4835 4836 region->init_req(_null_path, null_ctrl); 4837 phi ->init_req(_null_path, zerocon(T_OBJECT)); 4838 memphi->init_req(_null_path, prev_mem); 4839 4840 Node* n = shenandoah_write_barrier_helper(*this, not_null_obj, adr_type); 4841 4842 region->init_req(_not_null_path, control()); 4843 phi ->init_req(_not_null_path, n); 4844 memphi->init_req(_not_null_path, memory(adr_type)); 4845 4846 set_control(_gvn.transform(region)); 4847 record_for_igvn(region); 4848 set_memory(_gvn.transform(memphi), adr_type); 4849 4850 Node* res_val = _gvn.transform(phi); 4851 // replace_in_map(obj, res_val); 4852 return res_val; 4853 } else { 4854 // We know it is not null. Simple barrier is sufficient. 4855 Node* n = shenandoah_write_barrier_helper(*this, obj, adr_type); 4856 // replace_in_map(obj, n); 4857 record_for_igvn(n); 4858 return n; 4859 } 4860 4861 } else { 4862 return obj; 4863 } 4864 } 4865 4866 /** 4867 * In Shenandoah, we need barriers on acmp (and similar instructions that compare two 4868 * oops) to avoid false negatives. If it compares a from-space and a to-space 4869 * copy of an object, a regular acmp would return false, even though both are 4870 * the same. The acmp barrier compares the two objects, and when they are 4871 * *not equal* it does a read-barrier on both, and compares them again. When it 4872 * failed because of different copies of the object, we know that the object 4873 * must already have been evacuated (and therefore doesn't require a write-barrier). 4874 */ 4875 Node* GraphKit::cmp_objects(Node* a, Node* b) { 4876 // TODO: Refactor into proper GC interface. 4877 if (UseShenandoahGC && ShenandoahAcmpBarrier) { 4878 const Type* a_type = a->bottom_type(); 4879 const Type* b_type = b->bottom_type(); 4880 if (a_type->higher_equal(TypePtr::NULL_PTR) || b_type->higher_equal(TypePtr::NULL_PTR)) { 4881 // We know one arg is gonna be null. No need for barriers. 4882 return _gvn.transform(new CmpPNode(b, a)); 4883 } 4884 4885 const TypePtr* a_adr_type = ShenandoahBarrierNode::brooks_pointer_type(a_type); 4886 const TypePtr* b_adr_type = ShenandoahBarrierNode::brooks_pointer_type(b_type); 4887 if ((! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, a, memory(a_adr_type), false)) && 4888 (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, b, memory(b_adr_type), false))) { 4889 // We know both args are in to-space already. No acmp barrier needed. 4890 return _gvn.transform(new CmpPNode(b, a)); 4891 } 4892 4893 C->set_has_split_ifs(true); 4894 4895 if (ShenandoahVerifyOptoBarriers) { 4896 a = shenandoah_write_barrier(a); 4897 b = shenandoah_write_barrier(b); 4898 return _gvn.transform(new CmpPNode(b, a)); 4899 } 4900 4901 enum { _equal = 1, _not_equal, PATH_LIMIT }; 4902 RegionNode* region = new RegionNode(PATH_LIMIT); 4903 PhiNode* phiA = PhiNode::make(region, a, _gvn.type(a)->is_oopptr()->cast_to_nonconst()); 4904 PhiNode* phiB = PhiNode::make(region, b, _gvn.type(b)->is_oopptr()->cast_to_nonconst()); 4905 4906 Node* cmp = _gvn.transform(new CmpPNode(b, a)); 4907 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 4908 4909 // TODO: Use profiling data. 4910 IfNode* iff = create_and_map_if(control(), tst, PROB_FAIR, COUNT_UNKNOWN); 4911 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4912 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4913 4914 // Equal path: Use original values. 4915 region->init_req(_equal, iftrue); 4916 phiA->init_req(_equal, a); 4917 phiB->init_req(_equal, b); 4918 4919 uint alias_a = C->get_alias_index(a_adr_type); 4920 uint alias_b = C->get_alias_index(b_adr_type); 4921 PhiNode* mem_phi = NULL; 4922 if (alias_a == alias_b) { 4923 mem_phi = PhiNode::make(region, memory(alias_a), Type::MEMORY, C->get_adr_type(alias_a)); 4924 } else { 4925 mem_phi = PhiNode::make(region, map()->memory(), Type::MEMORY, TypePtr::BOTTOM); 4926 } 4927 4928 // Unequal path: retry after read barriers. 4929 set_control(iffalse); 4930 if (!iffalse->is_top()) { 4931 Node* mb = NULL; 4932 if (alias_a == alias_b) { 4933 Node* mem = reset_memory(); 4934 mb = MemBarNode::make(C, Op_MemBarAcquire, alias_a); 4935 mb->init_req(TypeFunc::Control, control()); 4936 mb->init_req(TypeFunc::Memory, mem); 4937 Node* membar = _gvn.transform(mb); 4938 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 4939 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory)); 4940 set_all_memory(mem); 4941 set_memory(newmem, alias_a); 4942 } else { 4943 mb = insert_mem_bar(Op_MemBarAcquire); 4944 } 4945 } else { 4946 a = top(); 4947 b = top(); 4948 } 4949 4950 a = shenandoah_read_barrier_impl(a, true, true, false); 4951 b = shenandoah_read_barrier_impl(b, true, true, false); 4952 4953 region->init_req(_not_equal, control()); 4954 phiA->init_req(_not_equal, a); 4955 phiB->init_req(_not_equal, b); 4956 if (alias_a == alias_b) { 4957 mem_phi->init_req(_not_equal, memory(alias_a)); 4958 set_memory(mem_phi, alias_a); 4959 } else { 4960 mem_phi->init_req(_not_equal, reset_memory()); 4961 set_all_memory(mem_phi); 4962 } 4963 record_for_igvn(mem_phi); 4964 _gvn.set_type(mem_phi, Type::MEMORY); 4965 set_control(_gvn.transform(region)); 4966 record_for_igvn(region); 4967 4968 a = _gvn.transform(phiA); 4969 b = _gvn.transform(phiB); 4970 return _gvn.transform(new CmpPNode(b, a)); 4971 } else { 4972 return _gvn.transform(new CmpPNode(b, a)); 4973 } 4974 }