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 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1163 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1164 } 1165 1166 //-------------------------load_array_length----------------------------------- 1167 Node* GraphKit::load_array_length(Node* array) { 1168 // Special-case a fresh allocation to avoid building nodes: 1169 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1170 Node *alen; 1171 if (alloc == NULL) { 1172 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1173 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1174 } else { 1175 alen = alloc->Ideal_length(); 1176 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1177 if (ccast != alen) { 1178 alen = _gvn.transform(ccast); 1179 } 1180 } 1181 return alen; 1182 } 1183 1184 //------------------------------do_null_check---------------------------------- 1185 // Helper function to do a NULL pointer check. Returned value is 1186 // the incoming address with NULL casted away. You are allowed to use the 1187 // not-null value only if you are control dependent on the test. 1188 #ifndef PRODUCT 1189 extern int explicit_null_checks_inserted, 1190 explicit_null_checks_elided; 1191 #endif 1192 Node* GraphKit::null_check_common(Node* value, BasicType type, 1193 // optional arguments for variations: 1194 bool assert_null, 1195 Node* *null_control, 1196 bool speculative) { 1197 assert(!assert_null || null_control == NULL, "not both at once"); 1198 if (stopped()) return top(); 1199 NOT_PRODUCT(explicit_null_checks_inserted++); 1200 1201 // Construct NULL check 1202 Node *chk = NULL; 1203 switch(type) { 1204 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1205 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1206 case T_ARRAY : // fall through 1207 type = T_OBJECT; // simplify further tests 1208 case T_OBJECT : { 1209 const Type *t = _gvn.type( value ); 1210 1211 const TypeOopPtr* tp = t->isa_oopptr(); 1212 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1213 // Only for do_null_check, not any of its siblings: 1214 && !assert_null && null_control == NULL) { 1215 // Usually, any field access or invocation on an unloaded oop type 1216 // will simply fail to link, since the statically linked class is 1217 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1218 // the static class is loaded but the sharper oop type is not. 1219 // Rather than checking for this obscure case in lots of places, 1220 // we simply observe that a null check on an unloaded class 1221 // will always be followed by a nonsense operation, so we 1222 // can just issue the uncommon trap here. 1223 // Our access to the unloaded class will only be correct 1224 // after it has been loaded and initialized, which requires 1225 // a trip through the interpreter. 1226 #ifndef PRODUCT 1227 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1228 #endif 1229 uncommon_trap(Deoptimization::Reason_unloaded, 1230 Deoptimization::Action_reinterpret, 1231 tp->klass(), "!loaded"); 1232 return top(); 1233 } 1234 1235 if (assert_null) { 1236 // See if the type is contained in NULL_PTR. 1237 // If so, then the value is already null. 1238 if (t->higher_equal(TypePtr::NULL_PTR)) { 1239 NOT_PRODUCT(explicit_null_checks_elided++); 1240 return value; // Elided null assert quickly! 1241 } 1242 } else { 1243 // See if mixing in the NULL pointer changes type. 1244 // If so, then the NULL pointer was not allowed in the original 1245 // type. In other words, "value" was not-null. 1246 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1247 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1248 NOT_PRODUCT(explicit_null_checks_elided++); 1249 return value; // Elided null check quickly! 1250 } 1251 } 1252 chk = new CmpPNode( value, null() ); 1253 break; 1254 } 1255 1256 default: 1257 fatal("unexpected type: %s", type2name(type)); 1258 } 1259 assert(chk != NULL, "sanity check"); 1260 chk = _gvn.transform(chk); 1261 1262 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1263 BoolNode *btst = new BoolNode( chk, btest); 1264 Node *tst = _gvn.transform( btst ); 1265 1266 //----------- 1267 // if peephole optimizations occurred, a prior test existed. 1268 // If a prior test existed, maybe it dominates as we can avoid this test. 1269 if (tst != btst && type == T_OBJECT) { 1270 // At this point we want to scan up the CFG to see if we can 1271 // find an identical test (and so avoid this test altogether). 1272 Node *cfg = control(); 1273 int depth = 0; 1274 while( depth < 16 ) { // Limit search depth for speed 1275 if( cfg->Opcode() == Op_IfTrue && 1276 cfg->in(0)->in(1) == tst ) { 1277 // Found prior test. Use "cast_not_null" to construct an identical 1278 // CastPP (and hence hash to) as already exists for the prior test. 1279 // Return that casted value. 1280 if (assert_null) { 1281 replace_in_map(value, null()); 1282 return null(); // do not issue the redundant test 1283 } 1284 Node *oldcontrol = control(); 1285 set_control(cfg); 1286 Node *res = cast_not_null(value); 1287 set_control(oldcontrol); 1288 NOT_PRODUCT(explicit_null_checks_elided++); 1289 return res; 1290 } 1291 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1292 if (cfg == NULL) break; // Quit at region nodes 1293 depth++; 1294 } 1295 } 1296 1297 //----------- 1298 // Branch to failure if null 1299 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1300 Deoptimization::DeoptReason reason; 1301 if (assert_null) { 1302 reason = Deoptimization::Reason_null_assert; 1303 } else if (type == T_OBJECT) { 1304 reason = Deoptimization::reason_null_check(speculative); 1305 } else { 1306 reason = Deoptimization::Reason_div0_check; 1307 } 1308 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1309 // ciMethodData::has_trap_at will return a conservative -1 if any 1310 // must-be-null assertion has failed. This could cause performance 1311 // problems for a method after its first do_null_assert failure. 1312 // Consider using 'Reason_class_check' instead? 1313 1314 // To cause an implicit null check, we set the not-null probability 1315 // to the maximum (PROB_MAX). For an explicit check the probability 1316 // is set to a smaller value. 1317 if (null_control != NULL || too_many_traps(reason)) { 1318 // probability is less likely 1319 ok_prob = PROB_LIKELY_MAG(3); 1320 } else if (!assert_null && 1321 (ImplicitNullCheckThreshold > 0) && 1322 method() != NULL && 1323 (method()->method_data()->trap_count(reason) 1324 >= (uint)ImplicitNullCheckThreshold)) { 1325 ok_prob = PROB_LIKELY_MAG(3); 1326 } 1327 1328 if (null_control != NULL) { 1329 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1330 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1331 set_control( _gvn.transform( new IfTrueNode(iff))); 1332 #ifndef PRODUCT 1333 if (null_true == top()) { 1334 explicit_null_checks_elided++; 1335 } 1336 #endif 1337 (*null_control) = null_true; 1338 } else { 1339 BuildCutout unless(this, tst, ok_prob); 1340 // Check for optimizer eliding test at parse time 1341 if (stopped()) { 1342 // Failure not possible; do not bother making uncommon trap. 1343 NOT_PRODUCT(explicit_null_checks_elided++); 1344 } else if (assert_null) { 1345 uncommon_trap(reason, 1346 Deoptimization::Action_make_not_entrant, 1347 NULL, "assert_null"); 1348 } else { 1349 replace_in_map(value, zerocon(type)); 1350 builtin_throw(reason); 1351 } 1352 } 1353 1354 // Must throw exception, fall-thru not possible? 1355 if (stopped()) { 1356 return top(); // No result 1357 } 1358 1359 if (assert_null) { 1360 // Cast obj to null on this path. 1361 replace_in_map(value, zerocon(type)); 1362 return zerocon(type); 1363 } 1364 1365 // Cast obj to not-null on this path, if there is no null_control. 1366 // (If there is a null_control, a non-null value may come back to haunt us.) 1367 if (type == T_OBJECT) { 1368 Node* cast = cast_not_null(value, false); 1369 if (null_control == NULL || (*null_control) == top()) 1370 replace_in_map(value, cast); 1371 value = cast; 1372 } 1373 1374 return value; 1375 } 1376 1377 1378 //------------------------------cast_not_null---------------------------------- 1379 // Cast obj to not-null on this path 1380 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1381 const Type *t = _gvn.type(obj); 1382 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1383 // Object is already not-null? 1384 if( t == t_not_null ) return obj; 1385 1386 Node *cast = new CastPPNode(obj,t_not_null); 1387 cast->init_req(0, control()); 1388 cast = _gvn.transform( cast ); 1389 1390 // Scan for instances of 'obj' in the current JVM mapping. 1391 // These instances are known to be not-null after the test. 1392 if (do_replace_in_map) 1393 replace_in_map(obj, cast); 1394 1395 return cast; // Return casted value 1396 } 1397 1398 1399 //--------------------------replace_in_map------------------------------------- 1400 void GraphKit::replace_in_map(Node* old, Node* neww) { 1401 if (old == neww) { 1402 return; 1403 } 1404 1405 map()->replace_edge(old, neww); 1406 1407 // Note: This operation potentially replaces any edge 1408 // on the map. This includes locals, stack, and monitors 1409 // of the current (innermost) JVM state. 1410 1411 // don't let inconsistent types from profiling escape this 1412 // method 1413 1414 const Type* told = _gvn.type(old); 1415 const Type* tnew = _gvn.type(neww); 1416 1417 if (!tnew->higher_equal(told)) { 1418 return; 1419 } 1420 1421 map()->record_replaced_node(old, neww); 1422 } 1423 1424 1425 //============================================================================= 1426 //--------------------------------memory--------------------------------------- 1427 Node* GraphKit::memory(uint alias_idx) { 1428 MergeMemNode* mem = merged_memory(); 1429 Node* p = mem->memory_at(alias_idx); 1430 _gvn.set_type(p, Type::MEMORY); // must be mapped 1431 return p; 1432 } 1433 1434 //-----------------------------reset_memory------------------------------------ 1435 Node* GraphKit::reset_memory() { 1436 Node* mem = map()->memory(); 1437 // do not use this node for any more parsing! 1438 debug_only( map()->set_memory((Node*)NULL) ); 1439 return _gvn.transform( mem ); 1440 } 1441 1442 //------------------------------set_all_memory--------------------------------- 1443 void GraphKit::set_all_memory(Node* newmem) { 1444 Node* mergemem = MergeMemNode::make(newmem); 1445 gvn().set_type_bottom(mergemem); 1446 map()->set_memory(mergemem); 1447 } 1448 1449 //------------------------------set_all_memory_call---------------------------- 1450 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1451 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1452 set_all_memory(newmem); 1453 } 1454 1455 //============================================================================= 1456 // 1457 // parser factory methods for MemNodes 1458 // 1459 // These are layered on top of the factory methods in LoadNode and StoreNode, 1460 // and integrate with the parser's memory state and _gvn engine. 1461 // 1462 1463 // factory methods in "int adr_idx" 1464 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1465 int adr_idx, 1466 MemNode::MemOrd mo, 1467 LoadNode::ControlDependency control_dependency, 1468 bool require_atomic_access, 1469 bool unaligned, 1470 bool mismatched) { 1471 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1472 const TypePtr* adr_type = NULL; // debug-mode-only argument 1473 debug_only(adr_type = C->get_adr_type(adr_idx)); 1474 Node* mem = memory(adr_idx); 1475 Node* ld; 1476 if (require_atomic_access && bt == T_LONG) { 1477 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1478 } else if (require_atomic_access && bt == T_DOUBLE) { 1479 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1480 } else { 1481 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1482 } 1483 ld = _gvn.transform(ld); 1484 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1485 // Improve graph before escape analysis and boxing elimination. 1486 record_for_igvn(ld); 1487 } 1488 return ld; 1489 } 1490 1491 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1492 int adr_idx, 1493 MemNode::MemOrd mo, 1494 bool require_atomic_access, 1495 bool unaligned, 1496 bool mismatched) { 1497 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1498 const TypePtr* adr_type = NULL; 1499 debug_only(adr_type = C->get_adr_type(adr_idx)); 1500 Node *mem = memory(adr_idx); 1501 Node* st; 1502 if (require_atomic_access && bt == T_LONG) { 1503 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1504 } else if (require_atomic_access && bt == T_DOUBLE) { 1505 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1506 } else { 1507 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1508 } 1509 if (unaligned) { 1510 st->as_Store()->set_unaligned_access(); 1511 } 1512 if (mismatched) { 1513 st->as_Store()->set_mismatched_access(); 1514 } 1515 st = _gvn.transform(st); 1516 set_memory(st, adr_idx); 1517 // Back-to-back stores can only remove intermediate store with DU info 1518 // so push on worklist for optimizer. 1519 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1520 record_for_igvn(st); 1521 1522 return st; 1523 } 1524 1525 1526 Node* GraphKit::pre_barrier(bool do_load, 1527 Node* ctl, 1528 Node* obj, 1529 Node* adr, 1530 uint adr_idx, 1531 Node* val, 1532 const TypeOopPtr* val_type, 1533 Node* pre_val, 1534 BasicType bt) { 1535 1536 BarrierSet* bs = Universe::heap()->barrier_set(); 1537 set_control(ctl); 1538 switch (bs->kind()) { 1539 case BarrierSet::G1SATBCTLogging: 1540 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1541 return val; 1542 case BarrierSet::ShenandoahBarrierSet: 1543 return shenandoah_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1544 1545 case BarrierSet::CardTableForRS: 1546 case BarrierSet::CardTableExtension: 1547 case BarrierSet::ModRef: 1548 break; 1549 1550 default : 1551 ShouldNotReachHere(); 1552 1553 } 1554 return val; 1555 } 1556 1557 bool GraphKit::can_move_pre_barrier() const { 1558 BarrierSet* bs = Universe::heap()->barrier_set(); 1559 switch (bs->kind()) { 1560 case BarrierSet::G1SATBCTLogging: 1561 case BarrierSet::ShenandoahBarrierSet: 1562 return true; // Can move it if no safepoint 1563 1564 case BarrierSet::CardTableForRS: 1565 case BarrierSet::CardTableExtension: 1566 case BarrierSet::ModRef: 1567 return true; // There is no pre-barrier 1568 1569 default : 1570 ShouldNotReachHere(); 1571 } 1572 return false; 1573 } 1574 1575 void GraphKit::post_barrier(Node* ctl, 1576 Node* store, 1577 Node* obj, 1578 Node* adr, 1579 uint adr_idx, 1580 Node* val, 1581 BasicType bt, 1582 bool use_precise) { 1583 BarrierSet* bs = Universe::heap()->barrier_set(); 1584 set_control(ctl); 1585 switch (bs->kind()) { 1586 case BarrierSet::G1SATBCTLogging: 1587 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1588 break; 1589 1590 case BarrierSet::CardTableForRS: 1591 case BarrierSet::CardTableExtension: 1592 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1593 break; 1594 1595 case BarrierSet::ModRef: 1596 case BarrierSet::ShenandoahBarrierSet: 1597 break; 1598 1599 default : 1600 ShouldNotReachHere(); 1601 1602 } 1603 } 1604 1605 Node* GraphKit::store_oop(Node* ctl, 1606 Node* obj, 1607 Node* adr, 1608 const TypePtr* adr_type, 1609 Node* val, 1610 const TypeOopPtr* val_type, 1611 BasicType bt, 1612 bool use_precise, 1613 MemNode::MemOrd mo, 1614 bool mismatched) { 1615 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1616 // could be delayed during Parse (for example, in adjust_map_after_if()). 1617 // Execute transformation here to avoid barrier generation in such case. 1618 if (_gvn.type(val) == TypePtr::NULL_PTR) 1619 val = _gvn.makecon(TypePtr::NULL_PTR); 1620 1621 set_control(ctl); 1622 if (stopped()) return top(); // Dead path ? 1623 1624 assert(bt == T_OBJECT, "sanity"); 1625 assert(val != NULL, "not dead path"); 1626 uint adr_idx = C->get_alias_index(adr_type); 1627 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1628 1629 val = pre_barrier(true /* do_load */, 1630 control(), obj, adr, adr_idx, val, val_type, 1631 NULL /* pre_val */, 1632 bt); 1633 1634 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1635 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1636 return store; 1637 } 1638 1639 // Could be an array or object we don't know at compile time (unsafe ref.) 1640 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1641 Node* obj, // containing obj 1642 Node* adr, // actual adress to store val at 1643 const TypePtr* adr_type, 1644 Node* val, 1645 BasicType bt, 1646 MemNode::MemOrd mo, 1647 bool mismatched) { 1648 Compile::AliasType* at = C->alias_type(adr_type); 1649 const TypeOopPtr* val_type = NULL; 1650 if (adr_type->isa_instptr()) { 1651 if (at->field() != NULL) { 1652 // known field. This code is a copy of the do_put_xxx logic. 1653 ciField* field = at->field(); 1654 if (!field->type()->is_loaded()) { 1655 val_type = TypeInstPtr::BOTTOM; 1656 } else { 1657 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1658 } 1659 } 1660 } else if (adr_type->isa_aryptr()) { 1661 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1662 } 1663 if (val_type == NULL) { 1664 val_type = TypeInstPtr::BOTTOM; 1665 } 1666 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1667 } 1668 1669 1670 //-------------------------array_element_address------------------------- 1671 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1672 const TypeInt* sizetype, Node* ctrl) { 1673 uint shift = exact_log2(type2aelembytes(elembt)); 1674 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1675 1676 // short-circuit a common case (saves lots of confusing waste motion) 1677 jint idx_con = find_int_con(idx, -1); 1678 if (idx_con >= 0) { 1679 intptr_t offset = header + ((intptr_t)idx_con << shift); 1680 return basic_plus_adr(ary, offset); 1681 } 1682 1683 // must be correct type for alignment purposes 1684 Node* base = basic_plus_adr(ary, header); 1685 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1686 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1687 return basic_plus_adr(ary, base, scale); 1688 } 1689 1690 //-------------------------load_array_element------------------------- 1691 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1692 const Type* elemtype = arytype->elem(); 1693 BasicType elembt = elemtype->array_element_basic_type(); 1694 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1695 if (elembt == T_NARROWOOP) { 1696 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1697 } 1698 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1699 return ld; 1700 } 1701 1702 //-------------------------set_arguments_for_java_call------------------------- 1703 // Arguments (pre-popped from the stack) are taken from the JVMS. 1704 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1705 // Add the call arguments: 1706 uint nargs = call->method()->arg_size(); 1707 for (uint i = 0; i < nargs; i++) { 1708 Node* arg = argument(i); 1709 call->init_req(i + TypeFunc::Parms, arg); 1710 } 1711 } 1712 1713 //---------------------------set_edges_for_java_call--------------------------- 1714 // Connect a newly created call into the current JVMS. 1715 // A return value node (if any) is returned from set_edges_for_java_call. 1716 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1717 1718 // Add the predefined inputs: 1719 call->init_req( TypeFunc::Control, control() ); 1720 call->init_req( TypeFunc::I_O , i_o() ); 1721 call->init_req( TypeFunc::Memory , reset_memory() ); 1722 call->init_req( TypeFunc::FramePtr, frameptr() ); 1723 call->init_req( TypeFunc::ReturnAdr, top() ); 1724 1725 add_safepoint_edges(call, must_throw); 1726 1727 Node* xcall = _gvn.transform(call); 1728 1729 if (xcall == top()) { 1730 set_control(top()); 1731 return; 1732 } 1733 assert(xcall == call, "call identity is stable"); 1734 1735 // Re-use the current map to produce the result. 1736 1737 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1738 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1739 set_all_memory_call(xcall, separate_io_proj); 1740 1741 //return xcall; // no need, caller already has it 1742 } 1743 1744 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1745 if (stopped()) return top(); // maybe the call folded up? 1746 1747 // Capture the return value, if any. 1748 Node* ret; 1749 if (call->method() == NULL || 1750 call->method()->return_type()->basic_type() == T_VOID) 1751 ret = top(); 1752 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1753 1754 // Note: Since any out-of-line call can produce an exception, 1755 // we always insert an I_O projection from the call into the result. 1756 1757 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1758 1759 if (separate_io_proj) { 1760 // The caller requested separate projections be used by the fall 1761 // through and exceptional paths, so replace the projections for 1762 // the fall through path. 1763 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1764 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1765 } 1766 return ret; 1767 } 1768 1769 //--------------------set_predefined_input_for_runtime_call-------------------- 1770 // Reading and setting the memory state is way conservative here. 1771 // The real problem is that I am not doing real Type analysis on memory, 1772 // so I cannot distinguish card mark stores from other stores. Across a GC 1773 // point the Store Barrier and the card mark memory has to agree. I cannot 1774 // have a card mark store and its barrier split across the GC point from 1775 // either above or below. Here I get that to happen by reading ALL of memory. 1776 // A better answer would be to separate out card marks from other memory. 1777 // For now, return the input memory state, so that it can be reused 1778 // after the call, if this call has restricted memory effects. 1779 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1780 // Set fixed predefined input arguments 1781 Node* memory = reset_memory(); 1782 call->init_req( TypeFunc::Control, control() ); 1783 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1784 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1785 call->init_req( TypeFunc::FramePtr, frameptr() ); 1786 call->init_req( TypeFunc::ReturnAdr, top() ); 1787 return memory; 1788 } 1789 1790 //-------------------set_predefined_output_for_runtime_call-------------------- 1791 // Set control and memory (not i_o) from the call. 1792 // If keep_mem is not NULL, use it for the output state, 1793 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1794 // If hook_mem is NULL, this call produces no memory effects at all. 1795 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1796 // then only that memory slice is taken from the call. 1797 // In the last case, we must put an appropriate memory barrier before 1798 // the call, so as to create the correct anti-dependencies on loads 1799 // preceding the call. 1800 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1801 Node* keep_mem, 1802 const TypePtr* hook_mem) { 1803 // no i/o 1804 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1805 if (keep_mem) { 1806 // First clone the existing memory state 1807 set_all_memory(keep_mem); 1808 if (hook_mem != NULL) { 1809 // Make memory for the call 1810 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1811 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1812 // We also use hook_mem to extract specific effects from arraycopy stubs. 1813 set_memory(mem, hook_mem); 1814 } 1815 // ...else the call has NO memory effects. 1816 1817 // Make sure the call advertises its memory effects precisely. 1818 // This lets us build accurate anti-dependences in gcm.cpp. 1819 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1820 "call node must be constructed correctly"); 1821 } else { 1822 assert(hook_mem == NULL, ""); 1823 // This is not a "slow path" call; all memory comes from the call. 1824 set_all_memory_call(call); 1825 } 1826 } 1827 1828 1829 // Replace the call with the current state of the kit. 1830 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1831 JVMState* ejvms = NULL; 1832 if (has_exceptions()) { 1833 ejvms = transfer_exceptions_into_jvms(); 1834 } 1835 1836 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1837 ReplacedNodes replaced_nodes_exception; 1838 Node* ex_ctl = top(); 1839 1840 SafePointNode* final_state = stop(); 1841 1842 // Find all the needed outputs of this call 1843 CallProjections callprojs; 1844 call->extract_projections(&callprojs, true); 1845 1846 Node* init_mem = call->in(TypeFunc::Memory); 1847 Node* final_mem = final_state->in(TypeFunc::Memory); 1848 Node* final_ctl = final_state->in(TypeFunc::Control); 1849 Node* final_io = final_state->in(TypeFunc::I_O); 1850 1851 // Replace all the old call edges with the edges from the inlining result 1852 if (callprojs.fallthrough_catchproj != NULL) { 1853 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1854 } 1855 if (callprojs.fallthrough_memproj != NULL) { 1856 if (final_mem->is_MergeMem()) { 1857 // Parser's exits MergeMem was not transformed but may be optimized 1858 final_mem = _gvn.transform(final_mem); 1859 } 1860 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1861 } 1862 if (callprojs.fallthrough_ioproj != NULL) { 1863 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1864 } 1865 1866 // Replace the result with the new result if it exists and is used 1867 if (callprojs.resproj != NULL && result != NULL) { 1868 C->gvn_replace_by(callprojs.resproj, result); 1869 } 1870 1871 if (ejvms == NULL) { 1872 // No exception edges to simply kill off those paths 1873 if (callprojs.catchall_catchproj != NULL) { 1874 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1875 } 1876 if (callprojs.catchall_memproj != NULL) { 1877 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1878 } 1879 if (callprojs.catchall_ioproj != NULL) { 1880 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1881 } 1882 // Replace the old exception object with top 1883 if (callprojs.exobj != NULL) { 1884 C->gvn_replace_by(callprojs.exobj, C->top()); 1885 } 1886 } else { 1887 GraphKit ekit(ejvms); 1888 1889 // Load my combined exception state into the kit, with all phis transformed: 1890 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1891 replaced_nodes_exception = ex_map->replaced_nodes(); 1892 1893 Node* ex_oop = ekit.use_exception_state(ex_map); 1894 1895 if (callprojs.catchall_catchproj != NULL) { 1896 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1897 ex_ctl = ekit.control(); 1898 } 1899 if (callprojs.catchall_memproj != NULL) { 1900 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1901 } 1902 if (callprojs.catchall_ioproj != NULL) { 1903 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1904 } 1905 1906 // Replace the old exception object with the newly created one 1907 if (callprojs.exobj != NULL) { 1908 C->gvn_replace_by(callprojs.exobj, ex_oop); 1909 } 1910 } 1911 1912 // Disconnect the call from the graph 1913 call->disconnect_inputs(NULL, C); 1914 C->gvn_replace_by(call, C->top()); 1915 1916 // Clean up any MergeMems that feed other MergeMems since the 1917 // optimizer doesn't like that. 1918 if (final_mem->is_MergeMem()) { 1919 Node_List wl; 1920 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1921 Node* m = i.get(); 1922 if (m->is_MergeMem() && !wl.contains(m)) { 1923 wl.push(m); 1924 } 1925 } 1926 while (wl.size() > 0) { 1927 _gvn.transform(wl.pop()); 1928 } 1929 } 1930 1931 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1932 replaced_nodes.apply(C, final_ctl); 1933 } 1934 if (!ex_ctl->is_top() && do_replaced_nodes) { 1935 replaced_nodes_exception.apply(C, ex_ctl); 1936 } 1937 } 1938 1939 1940 //------------------------------increment_counter------------------------------ 1941 // for statistics: increment a VM counter by 1 1942 1943 void GraphKit::increment_counter(address counter_addr) { 1944 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1945 increment_counter(adr1); 1946 } 1947 1948 void GraphKit::increment_counter(Node* counter_addr) { 1949 int adr_type = Compile::AliasIdxRaw; 1950 Node* ctrl = control(); 1951 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1952 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1953 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1954 } 1955 1956 1957 //------------------------------uncommon_trap---------------------------------- 1958 // Bail out to the interpreter in mid-method. Implemented by calling the 1959 // uncommon_trap blob. This helper function inserts a runtime call with the 1960 // right debug info. 1961 void GraphKit::uncommon_trap(int trap_request, 1962 ciKlass* klass, const char* comment, 1963 bool must_throw, 1964 bool keep_exact_action) { 1965 if (failing()) stop(); 1966 if (stopped()) return; // trap reachable? 1967 1968 // Note: If ProfileTraps is true, and if a deopt. actually 1969 // occurs here, the runtime will make sure an MDO exists. There is 1970 // no need to call method()->ensure_method_data() at this point. 1971 1972 // Set the stack pointer to the right value for reexecution: 1973 set_sp(reexecute_sp()); 1974 1975 #ifdef ASSERT 1976 if (!must_throw) { 1977 // Make sure the stack has at least enough depth to execute 1978 // the current bytecode. 1979 int inputs, ignored_depth; 1980 if (compute_stack_effects(inputs, ignored_depth)) { 1981 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1982 Bytecodes::name(java_bc()), sp(), inputs); 1983 } 1984 } 1985 #endif 1986 1987 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1988 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1989 1990 switch (action) { 1991 case Deoptimization::Action_maybe_recompile: 1992 case Deoptimization::Action_reinterpret: 1993 // Temporary fix for 6529811 to allow virtual calls to be sure they 1994 // get the chance to go from mono->bi->mega 1995 if (!keep_exact_action && 1996 Deoptimization::trap_request_index(trap_request) < 0 && 1997 too_many_recompiles(reason)) { 1998 // This BCI is causing too many recompilations. 1999 if (C->log() != NULL) { 2000 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2001 Deoptimization::trap_reason_name(reason), 2002 Deoptimization::trap_action_name(action)); 2003 } 2004 action = Deoptimization::Action_none; 2005 trap_request = Deoptimization::make_trap_request(reason, action); 2006 } else { 2007 C->set_trap_can_recompile(true); 2008 } 2009 break; 2010 case Deoptimization::Action_make_not_entrant: 2011 C->set_trap_can_recompile(true); 2012 break; 2013 #ifdef ASSERT 2014 case Deoptimization::Action_none: 2015 case Deoptimization::Action_make_not_compilable: 2016 break; 2017 default: 2018 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2019 break; 2020 #endif 2021 } 2022 2023 if (TraceOptoParse) { 2024 char buf[100]; 2025 tty->print_cr("Uncommon trap %s at bci:%d", 2026 Deoptimization::format_trap_request(buf, sizeof(buf), 2027 trap_request), bci()); 2028 } 2029 2030 CompileLog* log = C->log(); 2031 if (log != NULL) { 2032 int kid = (klass == NULL)? -1: log->identify(klass); 2033 log->begin_elem("uncommon_trap bci='%d'", bci()); 2034 char buf[100]; 2035 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2036 trap_request)); 2037 if (kid >= 0) log->print(" klass='%d'", kid); 2038 if (comment != NULL) log->print(" comment='%s'", comment); 2039 log->end_elem(); 2040 } 2041 2042 // Make sure any guarding test views this path as very unlikely 2043 Node *i0 = control()->in(0); 2044 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2045 IfNode *iff = i0->as_If(); 2046 float f = iff->_prob; // Get prob 2047 if (control()->Opcode() == Op_IfTrue) { 2048 if (f > PROB_UNLIKELY_MAG(4)) 2049 iff->_prob = PROB_MIN; 2050 } else { 2051 if (f < PROB_LIKELY_MAG(4)) 2052 iff->_prob = PROB_MAX; 2053 } 2054 } 2055 2056 // Clear out dead values from the debug info. 2057 kill_dead_locals(); 2058 2059 // Now insert the uncommon trap subroutine call 2060 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2061 const TypePtr* no_memory_effects = NULL; 2062 // Pass the index of the class to be loaded 2063 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2064 (must_throw ? RC_MUST_THROW : 0), 2065 OptoRuntime::uncommon_trap_Type(), 2066 call_addr, "uncommon_trap", no_memory_effects, 2067 intcon(trap_request)); 2068 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2069 "must extract request correctly from the graph"); 2070 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2071 2072 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2073 // The debug info is the only real input to this call. 2074 2075 // Halt-and-catch fire here. The above call should never return! 2076 HaltNode* halt = new HaltNode(control(), frameptr()); 2077 _gvn.set_type_bottom(halt); 2078 root()->add_req(halt); 2079 2080 stop_and_kill_map(); 2081 } 2082 2083 2084 //--------------------------just_allocated_object------------------------------ 2085 // Report the object that was just allocated. 2086 // It must be the case that there are no intervening safepoints. 2087 // We use this to determine if an object is so "fresh" that 2088 // it does not require card marks. 2089 Node* GraphKit::just_allocated_object(Node* current_control) { 2090 if (C->recent_alloc_ctl() == current_control) 2091 return C->recent_alloc_obj(); 2092 return NULL; 2093 } 2094 2095 2096 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2097 // (Note: TypeFunc::make has a cache that makes this fast.) 2098 const TypeFunc* tf = TypeFunc::make(dest_method); 2099 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2100 for (int j = 0; j < nargs; j++) { 2101 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2102 if( targ->basic_type() == T_DOUBLE ) { 2103 // If any parameters are doubles, they must be rounded before 2104 // the call, dstore_rounding does gvn.transform 2105 Node *arg = argument(j); 2106 arg = dstore_rounding(arg); 2107 set_argument(j, arg); 2108 } 2109 } 2110 } 2111 2112 /** 2113 * Record profiling data exact_kls for Node n with the type system so 2114 * that it can propagate it (speculation) 2115 * 2116 * @param n node that the type applies to 2117 * @param exact_kls type from profiling 2118 * @param maybe_null did profiling see null? 2119 * 2120 * @return node with improved type 2121 */ 2122 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2123 const Type* current_type = _gvn.type(n); 2124 assert(UseTypeSpeculation, "type speculation must be on"); 2125 2126 const TypePtr* speculative = current_type->speculative(); 2127 2128 // Should the klass from the profile be recorded in the speculative type? 2129 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2130 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2131 const TypeOopPtr* xtype = tklass->as_instance_type(); 2132 assert(xtype->klass_is_exact(), "Should be exact"); 2133 // Any reason to believe n is not null (from this profiling or a previous one)? 2134 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2135 // record the new speculative type's depth 2136 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2137 speculative = speculative->with_inline_depth(jvms()->depth()); 2138 } else if (current_type->would_improve_ptr(maybe_null)) { 2139 // Profiling report that null was never seen so we can change the 2140 // speculative type to non null ptr. 2141 assert(!maybe_null, "nothing to improve"); 2142 if (speculative == NULL) { 2143 speculative = TypePtr::NOTNULL; 2144 } else { 2145 const TypePtr* ptr = TypePtr::NOTNULL; 2146 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2147 } 2148 } 2149 2150 if (speculative != current_type->speculative()) { 2151 // Build a type with a speculative type (what we think we know 2152 // about the type but will need a guard when we use it) 2153 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2154 // We're changing the type, we need a new CheckCast node to carry 2155 // the new type. The new type depends on the control: what 2156 // profiling tells us is only valid from here as far as we can 2157 // tell. 2158 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2159 cast = _gvn.transform(cast); 2160 replace_in_map(n, cast); 2161 n = cast; 2162 } 2163 2164 return n; 2165 } 2166 2167 /** 2168 * Record profiling data from receiver profiling at an invoke with the 2169 * type system so that it can propagate it (speculation) 2170 * 2171 * @param n receiver node 2172 * 2173 * @return node with improved type 2174 */ 2175 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2176 if (!UseTypeSpeculation) { 2177 return n; 2178 } 2179 ciKlass* exact_kls = profile_has_unique_klass(); 2180 bool maybe_null = true; 2181 if (java_bc() == Bytecodes::_checkcast || 2182 java_bc() == Bytecodes::_instanceof || 2183 java_bc() == Bytecodes::_aastore) { 2184 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2185 maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2186 } 2187 return record_profile_for_speculation(n, exact_kls, maybe_null); 2188 } 2189 2190 /** 2191 * Record profiling data from argument profiling at an invoke with the 2192 * type system so that it can propagate it (speculation) 2193 * 2194 * @param dest_method target method for the call 2195 * @param bc what invoke bytecode is this? 2196 */ 2197 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2198 if (!UseTypeSpeculation) { 2199 return; 2200 } 2201 const TypeFunc* tf = TypeFunc::make(dest_method); 2202 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2203 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2204 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2205 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2206 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2207 bool maybe_null = true; 2208 ciKlass* better_type = NULL; 2209 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2210 record_profile_for_speculation(argument(j), better_type, maybe_null); 2211 } 2212 i++; 2213 } 2214 } 2215 } 2216 2217 /** 2218 * Record profiling data from parameter profiling at an invoke with 2219 * the type system so that it can propagate it (speculation) 2220 */ 2221 void GraphKit::record_profiled_parameters_for_speculation() { 2222 if (!UseTypeSpeculation) { 2223 return; 2224 } 2225 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2226 if (_gvn.type(local(i))->isa_oopptr()) { 2227 bool maybe_null = true; 2228 ciKlass* better_type = NULL; 2229 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2230 record_profile_for_speculation(local(i), better_type, maybe_null); 2231 } 2232 j++; 2233 } 2234 } 2235 } 2236 2237 /** 2238 * Record profiling data from return value profiling at an invoke with 2239 * the type system so that it can propagate it (speculation) 2240 */ 2241 void GraphKit::record_profiled_return_for_speculation() { 2242 if (!UseTypeSpeculation) { 2243 return; 2244 } 2245 bool maybe_null = true; 2246 ciKlass* better_type = NULL; 2247 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2248 // If profiling reports a single type for the return value, 2249 // feed it to the type system so it can propagate it as a 2250 // speculative type 2251 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2252 } 2253 } 2254 2255 void GraphKit::round_double_result(ciMethod* dest_method) { 2256 // A non-strict method may return a double value which has an extended 2257 // exponent, but this must not be visible in a caller which is 'strict' 2258 // If a strict caller invokes a non-strict callee, round a double result 2259 2260 BasicType result_type = dest_method->return_type()->basic_type(); 2261 assert( method() != NULL, "must have caller context"); 2262 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2263 // Destination method's return value is on top of stack 2264 // dstore_rounding() does gvn.transform 2265 Node *result = pop_pair(); 2266 result = dstore_rounding(result); 2267 push_pair(result); 2268 } 2269 } 2270 2271 // rounding for strict float precision conformance 2272 Node* GraphKit::precision_rounding(Node* n) { 2273 return UseStrictFP && _method->flags().is_strict() 2274 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2275 ? _gvn.transform( new RoundFloatNode(0, n) ) 2276 : n; 2277 } 2278 2279 // rounding for strict double precision conformance 2280 Node* GraphKit::dprecision_rounding(Node *n) { 2281 return UseStrictFP && _method->flags().is_strict() 2282 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2283 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2284 : n; 2285 } 2286 2287 // rounding for non-strict double stores 2288 Node* GraphKit::dstore_rounding(Node* n) { 2289 return Matcher::strict_fp_requires_explicit_rounding 2290 && UseSSE <= 1 2291 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2292 : n; 2293 } 2294 2295 //============================================================================= 2296 // Generate a fast path/slow path idiom. Graph looks like: 2297 // [foo] indicates that 'foo' is a parameter 2298 // 2299 // [in] NULL 2300 // \ / 2301 // CmpP 2302 // Bool ne 2303 // If 2304 // / \ 2305 // True False-<2> 2306 // / | 2307 // / cast_not_null 2308 // Load | | ^ 2309 // [fast_test] | | 2310 // gvn to opt_test | | 2311 // / \ | <1> 2312 // True False | 2313 // | \\ | 2314 // [slow_call] \[fast_result] 2315 // Ctl Val \ \ 2316 // | \ \ 2317 // Catch <1> \ \ 2318 // / \ ^ \ \ 2319 // Ex No_Ex | \ \ 2320 // | \ \ | \ <2> \ 2321 // ... \ [slow_res] | | \ [null_result] 2322 // \ \--+--+--- | | 2323 // \ | / \ | / 2324 // --------Region Phi 2325 // 2326 //============================================================================= 2327 // Code is structured as a series of driver functions all called 'do_XXX' that 2328 // call a set of helper functions. Helper functions first, then drivers. 2329 2330 //------------------------------null_check_oop--------------------------------- 2331 // Null check oop. Set null-path control into Region in slot 3. 2332 // Make a cast-not-nullness use the other not-null control. Return cast. 2333 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2334 bool never_see_null, 2335 bool safe_for_replace, 2336 bool speculative) { 2337 // Initial NULL check taken path 2338 (*null_control) = top(); 2339 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2340 2341 // Generate uncommon_trap: 2342 if (never_see_null && (*null_control) != top()) { 2343 // If we see an unexpected null at a check-cast we record it and force a 2344 // recompile; the offending check-cast will be compiled to handle NULLs. 2345 // If we see more than one offending BCI, then all checkcasts in the 2346 // method will be compiled to handle NULLs. 2347 PreserveJVMState pjvms(this); 2348 set_control(*null_control); 2349 replace_in_map(value, null()); 2350 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2351 uncommon_trap(reason, 2352 Deoptimization::Action_make_not_entrant); 2353 (*null_control) = top(); // NULL path is dead 2354 } 2355 if ((*null_control) == top() && safe_for_replace) { 2356 replace_in_map(value, cast); 2357 } 2358 2359 // Cast away null-ness on the result 2360 return cast; 2361 } 2362 2363 //------------------------------opt_iff---------------------------------------- 2364 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2365 // Return slow-path control. 2366 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2367 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2368 2369 // Fast path taken; set region slot 2 2370 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2371 region->init_req(2,fast_taken); // Capture fast-control 2372 2373 // Fast path not-taken, i.e. slow path 2374 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2375 return slow_taken; 2376 } 2377 2378 //-----------------------------make_runtime_call------------------------------- 2379 Node* GraphKit::make_runtime_call(int flags, 2380 const TypeFunc* call_type, address call_addr, 2381 const char* call_name, 2382 const TypePtr* adr_type, 2383 // The following parms are all optional. 2384 // The first NULL ends the list. 2385 Node* parm0, Node* parm1, 2386 Node* parm2, Node* parm3, 2387 Node* parm4, Node* parm5, 2388 Node* parm6, Node* parm7) { 2389 // Slow-path call 2390 bool is_leaf = !(flags & RC_NO_LEAF); 2391 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2392 if (call_name == NULL) { 2393 assert(!is_leaf, "must supply name for leaf"); 2394 call_name = OptoRuntime::stub_name(call_addr); 2395 } 2396 CallNode* call; 2397 if (!is_leaf) { 2398 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2399 bci(), adr_type); 2400 } else if (flags & RC_NO_FP) { 2401 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2402 } else { 2403 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2404 } 2405 2406 // The following is similar to set_edges_for_java_call, 2407 // except that the memory effects of the call are restricted to AliasIdxRaw. 2408 2409 // Slow path call has no side-effects, uses few values 2410 bool wide_in = !(flags & RC_NARROW_MEM); 2411 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2412 2413 Node* prev_mem = NULL; 2414 if (wide_in) { 2415 prev_mem = set_predefined_input_for_runtime_call(call); 2416 } else { 2417 assert(!wide_out, "narrow in => narrow out"); 2418 Node* narrow_mem = memory(adr_type); 2419 prev_mem = reset_memory(); 2420 map()->set_memory(narrow_mem); 2421 set_predefined_input_for_runtime_call(call); 2422 } 2423 2424 // Hook each parm in order. Stop looking at the first NULL. 2425 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2426 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2427 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2428 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2429 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2430 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2431 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2432 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2433 /* close each nested if ===> */ } } } } } } } } 2434 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2435 2436 if (!is_leaf) { 2437 // Non-leaves can block and take safepoints: 2438 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2439 } 2440 // Non-leaves can throw exceptions: 2441 if (has_io) { 2442 call->set_req(TypeFunc::I_O, i_o()); 2443 } 2444 2445 if (flags & RC_UNCOMMON) { 2446 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2447 // (An "if" probability corresponds roughly to an unconditional count. 2448 // Sort of.) 2449 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2450 } 2451 2452 Node* c = _gvn.transform(call); 2453 assert(c == call, "cannot disappear"); 2454 2455 if (wide_out) { 2456 // Slow path call has full side-effects. 2457 set_predefined_output_for_runtime_call(call); 2458 } else { 2459 // Slow path call has few side-effects, and/or sets few values. 2460 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2461 } 2462 2463 if (has_io) { 2464 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2465 } 2466 return call; 2467 2468 } 2469 2470 //------------------------------merge_memory----------------------------------- 2471 // Merge memory from one path into the current memory state. 2472 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2473 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2474 Node* old_slice = mms.force_memory(); 2475 Node* new_slice = mms.memory2(); 2476 if (old_slice != new_slice) { 2477 PhiNode* phi; 2478 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2479 if (mms.is_empty()) { 2480 // clone base memory Phi's inputs for this memory slice 2481 assert(old_slice == mms.base_memory(), "sanity"); 2482 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2483 _gvn.set_type(phi, Type::MEMORY); 2484 for (uint i = 1; i < phi->req(); i++) { 2485 phi->init_req(i, old_slice->in(i)); 2486 } 2487 } else { 2488 phi = old_slice->as_Phi(); // Phi was generated already 2489 } 2490 } else { 2491 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2492 _gvn.set_type(phi, Type::MEMORY); 2493 } 2494 phi->set_req(new_path, new_slice); 2495 mms.set_memory(phi); 2496 } 2497 } 2498 } 2499 2500 //------------------------------make_slow_call_ex------------------------------ 2501 // Make the exception handler hookups for the slow call 2502 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2503 if (stopped()) return; 2504 2505 // Make a catch node with just two handlers: fall-through and catch-all 2506 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2507 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2508 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2509 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2510 2511 { PreserveJVMState pjvms(this); 2512 set_control(excp); 2513 set_i_o(i_o); 2514 2515 if (excp != top()) { 2516 if (deoptimize) { 2517 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2518 uncommon_trap(Deoptimization::Reason_unhandled, 2519 Deoptimization::Action_none); 2520 } else { 2521 // Create an exception state also. 2522 // Use an exact type if the caller has specified a specific exception. 2523 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2524 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2525 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2526 } 2527 } 2528 } 2529 2530 // Get the no-exception control from the CatchNode. 2531 set_control(norm); 2532 } 2533 2534 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2535 Node* cmp = NULL; 2536 switch(bt) { 2537 case T_INT: cmp = new CmpINode(in1, in2); break; 2538 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2539 default: fatal("unexpected comparison type %s", type2name(bt)); 2540 } 2541 gvn->transform(cmp); 2542 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2543 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2544 gvn->transform(iff); 2545 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2546 return iff; 2547 } 2548 2549 2550 //-------------------------------gen_subtype_check----------------------------- 2551 // Generate a subtyping check. Takes as input the subtype and supertype. 2552 // Returns 2 values: sets the default control() to the true path and returns 2553 // the false path. Only reads invariant memory; sets no (visible) memory. 2554 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2555 // but that's not exposed to the optimizer. This call also doesn't take in an 2556 // Object; if you wish to check an Object you need to load the Object's class 2557 // prior to coming here. 2558 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2559 Compile* C = gvn->C; 2560 2561 if ((*ctrl)->is_top()) { 2562 return C->top(); 2563 } 2564 2565 // Fast check for identical types, perhaps identical constants. 2566 // The types can even be identical non-constants, in cases 2567 // involving Array.newInstance, Object.clone, etc. 2568 if (subklass == superklass) 2569 return C->top(); // false path is dead; no test needed. 2570 2571 if (gvn->type(superklass)->singleton()) { 2572 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2573 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2574 2575 // In the common case of an exact superklass, try to fold up the 2576 // test before generating code. You may ask, why not just generate 2577 // the code and then let it fold up? The answer is that the generated 2578 // code will necessarily include null checks, which do not always 2579 // completely fold away. If they are also needless, then they turn 2580 // into a performance loss. Example: 2581 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2582 // Here, the type of 'fa' is often exact, so the store check 2583 // of fa[1]=x will fold up, without testing the nullness of x. 2584 switch (C->static_subtype_check(superk, subk)) { 2585 case Compile::SSC_always_false: 2586 { 2587 Node* always_fail = *ctrl; 2588 *ctrl = gvn->C->top(); 2589 return always_fail; 2590 } 2591 case Compile::SSC_always_true: 2592 return C->top(); 2593 case Compile::SSC_easy_test: 2594 { 2595 // Just do a direct pointer compare and be done. 2596 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2597 *ctrl = gvn->transform(new IfTrueNode(iff)); 2598 return gvn->transform(new IfFalseNode(iff)); 2599 } 2600 case Compile::SSC_full_test: 2601 break; 2602 default: 2603 ShouldNotReachHere(); 2604 } 2605 } 2606 2607 // %%% Possible further optimization: Even if the superklass is not exact, 2608 // if the subklass is the unique subtype of the superklass, the check 2609 // will always succeed. We could leave a dependency behind to ensure this. 2610 2611 // First load the super-klass's check-offset 2612 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2613 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2614 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2615 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2616 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2617 2618 // Load from the sub-klass's super-class display list, or a 1-word cache of 2619 // the secondary superclass list, or a failing value with a sentinel offset 2620 // if the super-klass is an interface or exceptionally deep in the Java 2621 // hierarchy and we have to scan the secondary superclass list the hard way. 2622 // Worst-case type is a little odd: NULL is allowed as a result (usually 2623 // klass loads can never produce a NULL). 2624 Node *chk_off_X = chk_off; 2625 #ifdef _LP64 2626 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2627 #endif 2628 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2629 // For some types like interfaces the following loadKlass is from a 1-word 2630 // cache which is mutable so can't use immutable memory. Other 2631 // types load from the super-class display table which is immutable. 2632 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2633 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2634 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2635 2636 // Compile speed common case: ARE a subtype and we canNOT fail 2637 if( superklass == nkls ) 2638 return C->top(); // false path is dead; no test needed. 2639 2640 // See if we get an immediate positive hit. Happens roughly 83% of the 2641 // time. Test to see if the value loaded just previously from the subklass 2642 // is exactly the superklass. 2643 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2644 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2645 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2646 2647 // Compile speed common case: Check for being deterministic right now. If 2648 // chk_off is a constant and not equal to cacheoff then we are NOT a 2649 // subklass. In this case we need exactly the 1 test above and we can 2650 // return those results immediately. 2651 if (!might_be_cache) { 2652 Node* not_subtype_ctrl = *ctrl; 2653 *ctrl = iftrue1; // We need exactly the 1 test above 2654 return not_subtype_ctrl; 2655 } 2656 2657 // Gather the various success & failures here 2658 RegionNode *r_ok_subtype = new RegionNode(4); 2659 gvn->record_for_igvn(r_ok_subtype); 2660 RegionNode *r_not_subtype = new RegionNode(3); 2661 gvn->record_for_igvn(r_not_subtype); 2662 2663 r_ok_subtype->init_req(1, iftrue1); 2664 2665 // Check for immediate negative hit. Happens roughly 11% of the time (which 2666 // is roughly 63% of the remaining cases). Test to see if the loaded 2667 // check-offset points into the subklass display list or the 1-element 2668 // cache. If it points to the display (and NOT the cache) and the display 2669 // missed then it's not a subtype. 2670 Node *cacheoff = gvn->intcon(cacheoff_con); 2671 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2672 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2673 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2674 2675 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2676 // No performance impact (too rare) but allows sharing of secondary arrays 2677 // which has some footprint reduction. 2678 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2679 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2680 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2681 2682 // -- Roads not taken here: -- 2683 // We could also have chosen to perform the self-check at the beginning 2684 // of this code sequence, as the assembler does. This would not pay off 2685 // the same way, since the optimizer, unlike the assembler, can perform 2686 // static type analysis to fold away many successful self-checks. 2687 // Non-foldable self checks work better here in second position, because 2688 // the initial primary superclass check subsumes a self-check for most 2689 // types. An exception would be a secondary type like array-of-interface, 2690 // which does not appear in its own primary supertype display. 2691 // Finally, we could have chosen to move the self-check into the 2692 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2693 // dependent manner. But it is worthwhile to have the check here, 2694 // where it can be perhaps be optimized. The cost in code space is 2695 // small (register compare, branch). 2696 2697 // Now do a linear scan of the secondary super-klass array. Again, no real 2698 // performance impact (too rare) but it's gotta be done. 2699 // Since the code is rarely used, there is no penalty for moving it 2700 // out of line, and it can only improve I-cache density. 2701 // The decision to inline or out-of-line this final check is platform 2702 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2703 Node* psc = gvn->transform( 2704 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2705 2706 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2707 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2708 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2709 2710 // Return false path; set default control to true path. 2711 *ctrl = gvn->transform(r_ok_subtype); 2712 return gvn->transform(r_not_subtype); 2713 } 2714 2715 // Profile-driven exact type check: 2716 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2717 float prob, 2718 Node* *casted_receiver) { 2719 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2720 Node* recv_klass = load_object_klass(receiver); 2721 Node* want_klass = makecon(tklass); 2722 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2723 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2724 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2725 set_control( _gvn.transform( new IfTrueNode (iff) )); 2726 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2727 2728 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2729 assert(recv_xtype->klass_is_exact(), ""); 2730 2731 // Subsume downstream occurrences of receiver with a cast to 2732 // recv_xtype, since now we know what the type will be. 2733 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2734 (*casted_receiver) = _gvn.transform(cast); 2735 // (User must make the replace_in_map call.) 2736 2737 return fail; 2738 } 2739 2740 2741 //------------------------------seems_never_null------------------------------- 2742 // Use null_seen information if it is available from the profile. 2743 // If we see an unexpected null at a type check we record it and force a 2744 // recompile; the offending check will be recompiled to handle NULLs. 2745 // If we see several offending BCIs, then all checks in the 2746 // method will be recompiled. 2747 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2748 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2749 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2750 if (UncommonNullCast // Cutout for this technique 2751 && obj != null() // And not the -Xcomp stupid case? 2752 && !too_many_traps(reason) 2753 ) { 2754 if (speculating) { 2755 return true; 2756 } 2757 if (data == NULL) 2758 // Edge case: no mature data. Be optimistic here. 2759 return true; 2760 // If the profile has not seen a null, assume it won't happen. 2761 assert(java_bc() == Bytecodes::_checkcast || 2762 java_bc() == Bytecodes::_instanceof || 2763 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2764 return !data->as_BitData()->null_seen(); 2765 } 2766 speculating = false; 2767 return false; 2768 } 2769 2770 //------------------------maybe_cast_profiled_receiver------------------------- 2771 // If the profile has seen exactly one type, narrow to exactly that type. 2772 // Subsequent type checks will always fold up. 2773 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2774 ciKlass* require_klass, 2775 ciKlass* spec_klass, 2776 bool safe_for_replace) { 2777 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2778 2779 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2780 2781 // Make sure we haven't already deoptimized from this tactic. 2782 if (too_many_traps(reason) || too_many_recompiles(reason)) 2783 return NULL; 2784 2785 // (No, this isn't a call, but it's enough like a virtual call 2786 // to use the same ciMethod accessor to get the profile info...) 2787 // If we have a speculative type use it instead of profiling (which 2788 // may not help us) 2789 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2790 if (exact_kls != NULL) {// no cast failures here 2791 if (require_klass == NULL || 2792 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2793 // If we narrow the type to match what the type profile sees or 2794 // the speculative type, we can then remove the rest of the 2795 // cast. 2796 // This is a win, even if the exact_kls is very specific, 2797 // because downstream operations, such as method calls, 2798 // will often benefit from the sharper type. 2799 Node* exact_obj = not_null_obj; // will get updated in place... 2800 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2801 &exact_obj); 2802 { PreserveJVMState pjvms(this); 2803 set_control(slow_ctl); 2804 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2805 } 2806 if (safe_for_replace) { 2807 replace_in_map(not_null_obj, exact_obj); 2808 } 2809 return exact_obj; 2810 } 2811 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2812 } 2813 2814 return NULL; 2815 } 2816 2817 /** 2818 * Cast obj to type and emit guard unless we had too many traps here 2819 * already 2820 * 2821 * @param obj node being casted 2822 * @param type type to cast the node to 2823 * @param not_null true if we know node cannot be null 2824 */ 2825 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2826 ciKlass* type, 2827 bool not_null) { 2828 if (stopped()) { 2829 return obj; 2830 } 2831 2832 // type == NULL if profiling tells us this object is always null 2833 if (type != NULL) { 2834 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2835 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2836 2837 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2838 !too_many_traps(class_reason) && 2839 !too_many_recompiles(class_reason)) { 2840 Node* not_null_obj = NULL; 2841 // not_null is true if we know the object is not null and 2842 // there's no need for a null check 2843 if (!not_null) { 2844 Node* null_ctl = top(); 2845 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2846 assert(null_ctl->is_top(), "no null control here"); 2847 } else { 2848 not_null_obj = obj; 2849 } 2850 2851 Node* exact_obj = not_null_obj; 2852 ciKlass* exact_kls = type; 2853 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2854 &exact_obj); 2855 { 2856 PreserveJVMState pjvms(this); 2857 set_control(slow_ctl); 2858 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2859 } 2860 replace_in_map(not_null_obj, exact_obj); 2861 obj = exact_obj; 2862 } 2863 } else { 2864 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2865 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2866 Node* exact_obj = null_assert(obj); 2867 replace_in_map(obj, exact_obj); 2868 obj = exact_obj; 2869 } 2870 } 2871 return obj; 2872 } 2873 2874 //-------------------------------gen_instanceof-------------------------------- 2875 // Generate an instance-of idiom. Used by both the instance-of bytecode 2876 // and the reflective instance-of call. 2877 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2878 kill_dead_locals(); // Benefit all the uncommon traps 2879 assert( !stopped(), "dead parse path should be checked in callers" ); 2880 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2881 "must check for not-null not-dead klass in callers"); 2882 2883 // Make the merge point 2884 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2885 RegionNode* region = new RegionNode(PATH_LIMIT); 2886 Node* phi = new PhiNode(region, TypeInt::BOOL); 2887 C->set_has_split_ifs(true); // Has chance for split-if optimization 2888 2889 ciProfileData* data = NULL; 2890 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2891 data = method()->method_data()->bci_to_data(bci()); 2892 } 2893 bool speculative_not_null = false; 2894 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2895 && seems_never_null(obj, data, speculative_not_null)); 2896 2897 // Null check; get casted pointer; set region slot 3 2898 Node* null_ctl = top(); 2899 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2900 2901 // If not_null_obj is dead, only null-path is taken 2902 if (stopped()) { // Doing instance-of on a NULL? 2903 set_control(null_ctl); 2904 return intcon(0); 2905 } 2906 region->init_req(_null_path, null_ctl); 2907 phi ->init_req(_null_path, intcon(0)); // Set null path value 2908 if (null_ctl == top()) { 2909 // Do this eagerly, so that pattern matches like is_diamond_phi 2910 // will work even during parsing. 2911 assert(_null_path == PATH_LIMIT-1, "delete last"); 2912 region->del_req(_null_path); 2913 phi ->del_req(_null_path); 2914 } 2915 2916 // Do we know the type check always succeed? 2917 bool known_statically = false; 2918 if (_gvn.type(superklass)->singleton()) { 2919 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2920 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2921 if (subk != NULL && subk->is_loaded()) { 2922 int static_res = C->static_subtype_check(superk, subk); 2923 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2924 } 2925 } 2926 2927 if (known_statically && UseTypeSpeculation) { 2928 // If we know the type check always succeeds then we don't use the 2929 // profiling data at this bytecode. Don't lose it, feed it to the 2930 // type system as a speculative type. 2931 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2932 } else { 2933 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2934 // We may not have profiling here or it may not help us. If we 2935 // have a speculative type use it to perform an exact cast. 2936 ciKlass* spec_obj_type = obj_type->speculative_type(); 2937 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2938 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2939 if (stopped()) { // Profile disagrees with this path. 2940 set_control(null_ctl); // Null is the only remaining possibility. 2941 return intcon(0); 2942 } 2943 if (cast_obj != NULL) { 2944 not_null_obj = cast_obj; 2945 } 2946 } 2947 } 2948 2949 // Load the object's klass 2950 Node* obj_klass = load_object_klass(not_null_obj); 2951 2952 // Generate the subtype check 2953 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2954 2955 // Plug in the success path to the general merge in slot 1. 2956 region->init_req(_obj_path, control()); 2957 phi ->init_req(_obj_path, intcon(1)); 2958 2959 // Plug in the failing path to the general merge in slot 2. 2960 region->init_req(_fail_path, not_subtype_ctrl); 2961 phi ->init_req(_fail_path, intcon(0)); 2962 2963 // Return final merged results 2964 set_control( _gvn.transform(region) ); 2965 record_for_igvn(region); 2966 return _gvn.transform(phi); 2967 } 2968 2969 //-------------------------------gen_checkcast--------------------------------- 2970 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2971 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2972 // uncommon-trap paths work. Adjust stack after this call. 2973 // If failure_control is supplied and not null, it is filled in with 2974 // the control edge for the cast failure. Otherwise, an appropriate 2975 // uncommon trap or exception is thrown. 2976 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2977 Node* *failure_control) { 2978 kill_dead_locals(); // Benefit all the uncommon traps 2979 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2980 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2981 2982 // Fast cutout: Check the case that the cast is vacuously true. 2983 // This detects the common cases where the test will short-circuit 2984 // away completely. We do this before we perform the null check, 2985 // because if the test is going to turn into zero code, we don't 2986 // want a residual null check left around. (Causes a slowdown, 2987 // for example, in some objArray manipulations, such as a[i]=a[j].) 2988 if (tk->singleton()) { 2989 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2990 if (objtp != NULL && objtp->klass() != NULL) { 2991 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2992 case Compile::SSC_always_true: 2993 // If we know the type check always succeed then we don't use 2994 // the profiling data at this bytecode. Don't lose it, feed it 2995 // to the type system as a speculative type. 2996 return record_profiled_receiver_for_speculation(obj); 2997 case Compile::SSC_always_false: 2998 // It needs a null check because a null will *pass* the cast check. 2999 // A non-null value will always produce an exception. 3000 return null_assert(obj); 3001 } 3002 } 3003 } 3004 3005 ciProfileData* data = NULL; 3006 bool safe_for_replace = false; 3007 if (failure_control == NULL) { // use MDO in regular case only 3008 assert(java_bc() == Bytecodes::_aastore || 3009 java_bc() == Bytecodes::_checkcast, 3010 "interpreter profiles type checks only for these BCs"); 3011 data = method()->method_data()->bci_to_data(bci()); 3012 safe_for_replace = true; 3013 } 3014 3015 // Make the merge point 3016 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3017 RegionNode* region = new RegionNode(PATH_LIMIT); 3018 Node* phi = new PhiNode(region, toop); 3019 C->set_has_split_ifs(true); // Has chance for split-if optimization 3020 3021 // Use null-cast information if it is available 3022 bool speculative_not_null = false; 3023 bool never_see_null = ((failure_control == NULL) // regular case only 3024 && seems_never_null(obj, data, speculative_not_null)); 3025 3026 // Null check; get casted pointer; set region slot 3 3027 Node* null_ctl = top(); 3028 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3029 3030 // If not_null_obj is dead, only null-path is taken 3031 if (stopped()) { // Doing instance-of on a NULL? 3032 set_control(null_ctl); 3033 return null(); 3034 } 3035 region->init_req(_null_path, null_ctl); 3036 phi ->init_req(_null_path, null()); // Set null path value 3037 if (null_ctl == top()) { 3038 // Do this eagerly, so that pattern matches like is_diamond_phi 3039 // will work even during parsing. 3040 assert(_null_path == PATH_LIMIT-1, "delete last"); 3041 region->del_req(_null_path); 3042 phi ->del_req(_null_path); 3043 } 3044 3045 Node* cast_obj = NULL; 3046 if (tk->klass_is_exact()) { 3047 // The following optimization tries to statically cast the speculative type of the object 3048 // (for example obtained during profiling) to the type of the superklass and then do a 3049 // dynamic check that the type of the object is what we expect. To work correctly 3050 // for checkcast and aastore the type of superklass should be exact. 3051 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3052 // We may not have profiling here or it may not help us. If we have 3053 // a speculative type use it to perform an exact cast. 3054 ciKlass* spec_obj_type = obj_type->speculative_type(); 3055 if (spec_obj_type != NULL || data != NULL) { 3056 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3057 if (cast_obj != NULL) { 3058 if (failure_control != NULL) // failure is now impossible 3059 (*failure_control) = top(); 3060 // adjust the type of the phi to the exact klass: 3061 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3062 } 3063 } 3064 } 3065 3066 if (cast_obj == NULL) { 3067 // Load the object's klass 3068 Node* obj_klass = load_object_klass(not_null_obj); 3069 3070 // Generate the subtype check 3071 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3072 3073 // Plug in success path into the merge 3074 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3075 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3076 if (failure_control == NULL) { 3077 if (not_subtype_ctrl != top()) { // If failure is possible 3078 PreserveJVMState pjvms(this); 3079 set_control(not_subtype_ctrl); 3080 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3081 } 3082 } else { 3083 (*failure_control) = not_subtype_ctrl; 3084 } 3085 } 3086 3087 region->init_req(_obj_path, control()); 3088 phi ->init_req(_obj_path, cast_obj); 3089 3090 // A merge of NULL or Casted-NotNull obj 3091 Node* res = _gvn.transform(phi); 3092 3093 // Note I do NOT always 'replace_in_map(obj,result)' here. 3094 // if( tk->klass()->can_be_primary_super() ) 3095 // This means that if I successfully store an Object into an array-of-String 3096 // I 'forget' that the Object is really now known to be a String. I have to 3097 // do this because we don't have true union types for interfaces - if I store 3098 // a Baz into an array-of-Interface and then tell the optimizer it's an 3099 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3100 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3101 // replace_in_map( obj, res ); 3102 3103 // Return final merged results 3104 set_control( _gvn.transform(region) ); 3105 record_for_igvn(region); 3106 return res; 3107 } 3108 3109 //------------------------------next_monitor----------------------------------- 3110 // What number should be given to the next monitor? 3111 int GraphKit::next_monitor() { 3112 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3113 int next = current + C->sync_stack_slots(); 3114 // Keep the toplevel high water mark current: 3115 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3116 return current; 3117 } 3118 3119 //------------------------------insert_mem_bar--------------------------------- 3120 // Memory barrier to avoid floating things around 3121 // The membar serves as a pinch point between both control and all memory slices. 3122 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3123 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3124 mb->init_req(TypeFunc::Control, control()); 3125 mb->init_req(TypeFunc::Memory, reset_memory()); 3126 Node* membar = _gvn.transform(mb); 3127 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3128 set_all_memory_call(membar); 3129 return membar; 3130 } 3131 3132 //-------------------------insert_mem_bar_volatile---------------------------- 3133 // Memory barrier to avoid floating things around 3134 // The membar serves as a pinch point between both control and memory(alias_idx). 3135 // If you want to make a pinch point on all memory slices, do not use this 3136 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3137 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3138 // When Parse::do_put_xxx updates a volatile field, it appends a series 3139 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3140 // The first membar is on the same memory slice as the field store opcode. 3141 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3142 // All the other membars (for other volatile slices, including AliasIdxBot, 3143 // which stands for all unknown volatile slices) are control-dependent 3144 // on the first membar. This prevents later volatile loads or stores 3145 // from sliding up past the just-emitted store. 3146 3147 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3148 mb->set_req(TypeFunc::Control,control()); 3149 if (alias_idx == Compile::AliasIdxBot) { 3150 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3151 } else { 3152 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3153 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3154 } 3155 Node* membar = _gvn.transform(mb); 3156 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3157 if (alias_idx == Compile::AliasIdxBot) { 3158 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3159 } else { 3160 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3161 } 3162 return membar; 3163 } 3164 3165 //------------------------------shared_lock------------------------------------ 3166 // Emit locking code. 3167 FastLockNode* GraphKit::shared_lock(Node* obj) { 3168 // bci is either a monitorenter bc or InvocationEntryBci 3169 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3170 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3171 3172 if( !GenerateSynchronizationCode ) 3173 return NULL; // Not locking things? 3174 if (stopped()) // Dead monitor? 3175 return NULL; 3176 3177 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3178 3179 obj = shenandoah_write_barrier(obj); 3180 3181 // Box the stack location 3182 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3183 Node* mem = reset_memory(); 3184 3185 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3186 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3187 // Create the counters for this fast lock. 3188 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3189 } 3190 3191 // Create the rtm counters for this fast lock if needed. 3192 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3193 3194 // Add monitor to debug info for the slow path. If we block inside the 3195 // slow path and de-opt, we need the monitor hanging around 3196 map()->push_monitor( flock ); 3197 3198 const TypeFunc *tf = LockNode::lock_type(); 3199 LockNode *lock = new LockNode(C, tf); 3200 3201 lock->init_req( TypeFunc::Control, control() ); 3202 lock->init_req( TypeFunc::Memory , mem ); 3203 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3204 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3205 lock->init_req( TypeFunc::ReturnAdr, top() ); 3206 3207 lock->init_req(TypeFunc::Parms + 0, obj); 3208 lock->init_req(TypeFunc::Parms + 1, box); 3209 lock->init_req(TypeFunc::Parms + 2, flock); 3210 add_safepoint_edges(lock); 3211 3212 lock = _gvn.transform( lock )->as_Lock(); 3213 3214 // lock has no side-effects, sets few values 3215 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3216 3217 insert_mem_bar(Op_MemBarAcquireLock); 3218 3219 // Add this to the worklist so that the lock can be eliminated 3220 record_for_igvn(lock); 3221 3222 #ifndef PRODUCT 3223 if (PrintLockStatistics) { 3224 // Update the counter for this lock. Don't bother using an atomic 3225 // operation since we don't require absolute accuracy. 3226 lock->create_lock_counter(map()->jvms()); 3227 increment_counter(lock->counter()->addr()); 3228 } 3229 #endif 3230 3231 return flock; 3232 } 3233 3234 3235 //------------------------------shared_unlock---------------------------------- 3236 // Emit unlocking code. 3237 void GraphKit::shared_unlock(Node* box, Node* obj) { 3238 // bci is either a monitorenter bc or InvocationEntryBci 3239 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3240 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3241 3242 if( !GenerateSynchronizationCode ) 3243 return; 3244 if (stopped()) { // Dead monitor? 3245 map()->pop_monitor(); // Kill monitor from debug info 3246 return; 3247 } 3248 3249 // Memory barrier to avoid floating things down past the locked region 3250 insert_mem_bar(Op_MemBarReleaseLock); 3251 3252 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3253 UnlockNode *unlock = new UnlockNode(C, tf); 3254 #ifdef ASSERT 3255 unlock->set_dbg_jvms(sync_jvms()); 3256 #endif 3257 uint raw_idx = Compile::AliasIdxRaw; 3258 unlock->init_req( TypeFunc::Control, control() ); 3259 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3260 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3261 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3262 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3263 3264 unlock->init_req(TypeFunc::Parms + 0, obj); 3265 unlock->init_req(TypeFunc::Parms + 1, box); 3266 unlock = _gvn.transform(unlock)->as_Unlock(); 3267 3268 Node* mem = reset_memory(); 3269 3270 // unlock has no side-effects, sets few values 3271 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3272 3273 // Kill monitor from debug info 3274 map()->pop_monitor( ); 3275 } 3276 3277 //-------------------------------get_layout_helper----------------------------- 3278 // If the given klass is a constant or known to be an array, 3279 // fetch the constant layout helper value into constant_value 3280 // and return (Node*)NULL. Otherwise, load the non-constant 3281 // layout helper value, and return the node which represents it. 3282 // This two-faced routine is useful because allocation sites 3283 // almost always feature constant types. 3284 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3285 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3286 if (!StressReflectiveCode && inst_klass != NULL) { 3287 ciKlass* klass = inst_klass->klass(); 3288 bool xklass = inst_klass->klass_is_exact(); 3289 if (xklass || klass->is_array_klass()) { 3290 jint lhelper = klass->layout_helper(); 3291 if (lhelper != Klass::_lh_neutral_value) { 3292 constant_value = lhelper; 3293 return (Node*) NULL; 3294 } 3295 } 3296 } 3297 constant_value = Klass::_lh_neutral_value; // put in a known value 3298 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3299 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3300 } 3301 3302 // We just put in an allocate/initialize with a big raw-memory effect. 3303 // Hook selected additional alias categories on the initialization. 3304 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3305 MergeMemNode* init_in_merge, 3306 Node* init_out_raw) { 3307 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3308 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3309 3310 Node* prevmem = kit.memory(alias_idx); 3311 init_in_merge->set_memory_at(alias_idx, prevmem); 3312 kit.set_memory(init_out_raw, alias_idx); 3313 } 3314 3315 //---------------------------set_output_for_allocation------------------------- 3316 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3317 const TypeOopPtr* oop_type, 3318 bool deoptimize_on_exception) { 3319 int rawidx = Compile::AliasIdxRaw; 3320 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3321 add_safepoint_edges(alloc); 3322 Node* allocx = _gvn.transform(alloc); 3323 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3324 // create memory projection for i_o 3325 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3326 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3327 3328 // create a memory projection as for the normal control path 3329 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3330 set_memory(malloc, rawidx); 3331 3332 // a normal slow-call doesn't change i_o, but an allocation does 3333 // we create a separate i_o projection for the normal control path 3334 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3335 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3336 3337 // put in an initialization barrier 3338 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3339 rawoop)->as_Initialize(); 3340 assert(alloc->initialization() == init, "2-way macro link must work"); 3341 assert(init ->allocation() == alloc, "2-way macro link must work"); 3342 { 3343 // Extract memory strands which may participate in the new object's 3344 // initialization, and source them from the new InitializeNode. 3345 // This will allow us to observe initializations when they occur, 3346 // and link them properly (as a group) to the InitializeNode. 3347 assert(init->in(InitializeNode::Memory) == malloc, ""); 3348 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3349 init->set_req(InitializeNode::Memory, minit_in); 3350 record_for_igvn(minit_in); // fold it up later, if possible 3351 Node* minit_out = memory(rawidx); 3352 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3353 if (oop_type->isa_aryptr()) { 3354 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3355 int elemidx = C->get_alias_index(telemref); 3356 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3357 } else if (oop_type->isa_instptr()) { 3358 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3359 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3360 ciField* field = ik->nonstatic_field_at(i); 3361 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3362 continue; // do not bother to track really large numbers of fields 3363 // Find (or create) the alias category for this field: 3364 int fieldidx = C->alias_type(field)->index(); 3365 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3366 } 3367 } 3368 } 3369 3370 // Cast raw oop to the real thing... 3371 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3372 javaoop = _gvn.transform(javaoop); 3373 C->set_recent_alloc(control(), javaoop); 3374 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3375 3376 #ifdef ASSERT 3377 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3378 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3379 "Ideal_allocation works"); 3380 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3381 "Ideal_allocation works"); 3382 if (alloc->is_AllocateArray()) { 3383 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3384 "Ideal_allocation works"); 3385 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3386 "Ideal_allocation works"); 3387 } else { 3388 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3389 } 3390 } 3391 #endif //ASSERT 3392 3393 return javaoop; 3394 } 3395 3396 //---------------------------new_instance-------------------------------------- 3397 // This routine takes a klass_node which may be constant (for a static type) 3398 // or may be non-constant (for reflective code). It will work equally well 3399 // for either, and the graph will fold nicely if the optimizer later reduces 3400 // the type to a constant. 3401 // The optional arguments are for specialized use by intrinsics: 3402 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3403 // - If 'return_size_val', report the the total object size to the caller. 3404 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3405 Node* GraphKit::new_instance(Node* klass_node, 3406 Node* extra_slow_test, 3407 Node* *return_size_val, 3408 bool deoptimize_on_exception) { 3409 // Compute size in doublewords 3410 // The size is always an integral number of doublewords, represented 3411 // as a positive bytewise size stored in the klass's layout_helper. 3412 // The layout_helper also encodes (in a low bit) the need for a slow path. 3413 jint layout_con = Klass::_lh_neutral_value; 3414 Node* layout_val = get_layout_helper(klass_node, layout_con); 3415 int layout_is_con = (layout_val == NULL); 3416 3417 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3418 // Generate the initial go-slow test. It's either ALWAYS (return a 3419 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3420 // case) a computed value derived from the layout_helper. 3421 Node* initial_slow_test = NULL; 3422 if (layout_is_con) { 3423 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3424 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3425 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3426 } else { // reflective case 3427 // This reflective path is used by Unsafe.allocateInstance. 3428 // (It may be stress-tested by specifying StressReflectiveCode.) 3429 // Basically, we want to get into the VM is there's an illegal argument. 3430 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3431 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3432 if (extra_slow_test != intcon(0)) { 3433 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3434 } 3435 // (Macro-expander will further convert this to a Bool, if necessary.) 3436 } 3437 3438 // Find the size in bytes. This is easy; it's the layout_helper. 3439 // The size value must be valid even if the slow path is taken. 3440 Node* size = NULL; 3441 if (layout_is_con) { 3442 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3443 } else { // reflective case 3444 // This reflective path is used by clone and Unsafe.allocateInstance. 3445 size = ConvI2X(layout_val); 3446 3447 // Clear the low bits to extract layout_helper_size_in_bytes: 3448 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3449 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3450 size = _gvn.transform( new AndXNode(size, mask) ); 3451 } 3452 if (return_size_val != NULL) { 3453 (*return_size_val) = size; 3454 } 3455 3456 // This is a precise notnull oop of the klass. 3457 // (Actually, it need not be precise if this is a reflective allocation.) 3458 // It's what we cast the result to. 3459 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3460 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3461 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3462 3463 // Now generate allocation code 3464 3465 // The entire memory state is needed for slow path of the allocation 3466 // since GC and deoptimization can happened. 3467 Node *mem = reset_memory(); 3468 set_all_memory(mem); // Create new memory state 3469 3470 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3471 control(), mem, i_o(), 3472 size, klass_node, 3473 initial_slow_test); 3474 3475 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3476 } 3477 3478 //-------------------------------new_array------------------------------------- 3479 // helper for both newarray and anewarray 3480 // The 'length' parameter is (obviously) the length of the array. 3481 // See comments on new_instance for the meaning of the other arguments. 3482 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3483 Node* length, // number of array elements 3484 int nargs, // number of arguments to push back for uncommon trap 3485 Node* *return_size_val, 3486 bool deoptimize_on_exception) { 3487 jint layout_con = Klass::_lh_neutral_value; 3488 Node* layout_val = get_layout_helper(klass_node, layout_con); 3489 int layout_is_con = (layout_val == NULL); 3490 3491 if (!layout_is_con && !StressReflectiveCode && 3492 !too_many_traps(Deoptimization::Reason_class_check)) { 3493 // This is a reflective array creation site. 3494 // Optimistically assume that it is a subtype of Object[], 3495 // so that we can fold up all the address arithmetic. 3496 layout_con = Klass::array_layout_helper(T_OBJECT); 3497 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3498 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3499 { BuildCutout unless(this, bol_lh, PROB_MAX); 3500 inc_sp(nargs); 3501 uncommon_trap(Deoptimization::Reason_class_check, 3502 Deoptimization::Action_maybe_recompile); 3503 } 3504 layout_val = NULL; 3505 layout_is_con = true; 3506 } 3507 3508 // Generate the initial go-slow test. Make sure we do not overflow 3509 // if length is huge (near 2Gig) or negative! We do not need 3510 // exact double-words here, just a close approximation of needed 3511 // double-words. We can't add any offset or rounding bits, lest we 3512 // take a size -1 of bytes and make it positive. Use an unsigned 3513 // compare, so negative sizes look hugely positive. 3514 int fast_size_limit = FastAllocateSizeLimit; 3515 if (layout_is_con) { 3516 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3517 // Increase the size limit if we have exact knowledge of array type. 3518 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3519 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3520 } 3521 3522 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3523 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3524 3525 // --- Size Computation --- 3526 // array_size = round_to_heap(array_header + (length << elem_shift)); 3527 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3528 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3529 // The rounding mask is strength-reduced, if possible. 3530 int round_mask = MinObjAlignmentInBytes - 1; 3531 Node* header_size = NULL; 3532 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3533 // (T_BYTE has the weakest alignment and size restrictions...) 3534 if (layout_is_con) { 3535 int hsize = Klass::layout_helper_header_size(layout_con); 3536 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3537 BasicType etype = Klass::layout_helper_element_type(layout_con); 3538 if ((round_mask & ~right_n_bits(eshift)) == 0) 3539 round_mask = 0; // strength-reduce it if it goes away completely 3540 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3541 assert(header_size_min <= hsize, "generic minimum is smallest"); 3542 header_size_min = hsize; 3543 header_size = intcon(hsize + round_mask); 3544 } else { 3545 Node* hss = intcon(Klass::_lh_header_size_shift); 3546 Node* hsm = intcon(Klass::_lh_header_size_mask); 3547 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3548 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3549 Node* mask = intcon(round_mask); 3550 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3551 } 3552 3553 Node* elem_shift = NULL; 3554 if (layout_is_con) { 3555 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3556 if (eshift != 0) 3557 elem_shift = intcon(eshift); 3558 } else { 3559 // There is no need to mask or shift this value. 3560 // The semantics of LShiftINode include an implicit mask to 0x1F. 3561 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3562 elem_shift = layout_val; 3563 } 3564 3565 // Transition to native address size for all offset calculations: 3566 Node* lengthx = ConvI2X(length); 3567 Node* headerx = ConvI2X(header_size); 3568 #ifdef _LP64 3569 { const TypeInt* tilen = _gvn.find_int_type(length); 3570 if (tilen != NULL && tilen->_lo < 0) { 3571 // Add a manual constraint to a positive range. Cf. array_element_address. 3572 jint size_max = fast_size_limit; 3573 if (size_max > tilen->_hi) size_max = tilen->_hi; 3574 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3575 3576 // Only do a narrow I2L conversion if the range check passed. 3577 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3578 _gvn.transform(iff); 3579 RegionNode* region = new RegionNode(3); 3580 _gvn.set_type(region, Type::CONTROL); 3581 lengthx = new PhiNode(region, TypeLong::LONG); 3582 _gvn.set_type(lengthx, TypeLong::LONG); 3583 3584 // Range check passed. Use ConvI2L node with narrow type. 3585 Node* passed = IfFalse(iff); 3586 region->init_req(1, passed); 3587 // Make I2L conversion control dependent to prevent it from 3588 // floating above the range check during loop optimizations. 3589 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3590 3591 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3592 region->init_req(2, IfTrue(iff)); 3593 lengthx->init_req(2, ConvI2X(length)); 3594 3595 set_control(region); 3596 record_for_igvn(region); 3597 record_for_igvn(lengthx); 3598 } 3599 } 3600 #endif 3601 3602 // Combine header size (plus rounding) and body size. Then round down. 3603 // This computation cannot overflow, because it is used only in two 3604 // places, one where the length is sharply limited, and the other 3605 // after a successful allocation. 3606 Node* abody = lengthx; 3607 if (elem_shift != NULL) 3608 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3609 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3610 if (round_mask != 0) { 3611 Node* mask = MakeConX(~round_mask); 3612 size = _gvn.transform( new AndXNode(size, mask) ); 3613 } 3614 // else if round_mask == 0, the size computation is self-rounding 3615 3616 if (return_size_val != NULL) { 3617 // This is the size 3618 (*return_size_val) = size; 3619 } 3620 3621 // Now generate allocation code 3622 3623 // The entire memory state is needed for slow path of the allocation 3624 // since GC and deoptimization can happened. 3625 Node *mem = reset_memory(); 3626 set_all_memory(mem); // Create new memory state 3627 3628 if (initial_slow_test->is_Bool()) { 3629 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3630 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3631 } 3632 3633 // Create the AllocateArrayNode and its result projections 3634 AllocateArrayNode* alloc 3635 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3636 control(), mem, i_o(), 3637 size, klass_node, 3638 initial_slow_test, 3639 length); 3640 3641 // Cast to correct type. Note that the klass_node may be constant or not, 3642 // and in the latter case the actual array type will be inexact also. 3643 // (This happens via a non-constant argument to inline_native_newArray.) 3644 // In any case, the value of klass_node provides the desired array type. 3645 const TypeInt* length_type = _gvn.find_int_type(length); 3646 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3647 if (ary_type->isa_aryptr() && length_type != NULL) { 3648 // Try to get a better type than POS for the size 3649 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3650 } 3651 3652 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3653 3654 // Cast length on remaining path to be as narrow as possible 3655 if (map()->find_edge(length) >= 0) { 3656 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3657 if (ccast != length) { 3658 _gvn.set_type_bottom(ccast); 3659 record_for_igvn(ccast); 3660 replace_in_map(length, ccast); 3661 } 3662 } 3663 3664 return javaoop; 3665 } 3666 3667 // The following "Ideal_foo" functions are placed here because they recognize 3668 // the graph shapes created by the functions immediately above. 3669 3670 //---------------------------Ideal_allocation---------------------------------- 3671 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3672 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3673 if (ptr == NULL) { // reduce dumb test in callers 3674 return NULL; 3675 } 3676 3677 // Attempt to see through Shenandoah barriers. 3678 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr); 3679 3680 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3681 ptr = ptr->in(1); 3682 if (ptr == NULL) return NULL; 3683 } 3684 // Return NULL for allocations with several casts: 3685 // j.l.reflect.Array.newInstance(jobject, jint) 3686 // Object.clone() 3687 // to keep more precise type from last cast. 3688 if (ptr->is_Proj()) { 3689 Node* allo = ptr->in(0); 3690 if (allo != NULL && allo->is_Allocate()) { 3691 return allo->as_Allocate(); 3692 } 3693 } 3694 // Report failure to match. 3695 return NULL; 3696 } 3697 3698 // Fancy version which also strips off an offset (and reports it to caller). 3699 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3700 intptr_t& offset) { 3701 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3702 if (base == NULL) return NULL; 3703 return Ideal_allocation(base, phase); 3704 } 3705 3706 // Trace Initialize <- Proj[Parm] <- Allocate 3707 AllocateNode* InitializeNode::allocation() { 3708 Node* rawoop = in(InitializeNode::RawAddress); 3709 if (rawoop->is_Proj()) { 3710 Node* alloc = rawoop->in(0); 3711 if (alloc->is_Allocate()) { 3712 return alloc->as_Allocate(); 3713 } 3714 } 3715 return NULL; 3716 } 3717 3718 // Trace Allocate -> Proj[Parm] -> Initialize 3719 InitializeNode* AllocateNode::initialization() { 3720 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3721 if (rawoop == NULL) return NULL; 3722 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3723 Node* init = rawoop->fast_out(i); 3724 if (init->is_Initialize()) { 3725 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3726 return init->as_Initialize(); 3727 } 3728 } 3729 return NULL; 3730 } 3731 3732 //----------------------------- loop predicates --------------------------- 3733 3734 //------------------------------add_predicate_impl---------------------------- 3735 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3736 // Too many traps seen? 3737 if (too_many_traps(reason)) { 3738 #ifdef ASSERT 3739 if (TraceLoopPredicate) { 3740 int tc = C->trap_count(reason); 3741 tty->print("too many traps=%s tcount=%d in ", 3742 Deoptimization::trap_reason_name(reason), tc); 3743 method()->print(); // which method has too many predicate traps 3744 tty->cr(); 3745 } 3746 #endif 3747 // We cannot afford to take more traps here, 3748 // do not generate predicate. 3749 return; 3750 } 3751 3752 Node *cont = _gvn.intcon(1); 3753 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3754 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3755 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3756 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3757 C->add_predicate_opaq(opq); 3758 { 3759 PreserveJVMState pjvms(this); 3760 set_control(iffalse); 3761 inc_sp(nargs); 3762 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3763 } 3764 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3765 set_control(iftrue); 3766 } 3767 3768 //------------------------------add_predicate--------------------------------- 3769 void GraphKit::add_predicate(int nargs) { 3770 if (UseLoopPredicate) { 3771 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3772 } 3773 // loop's limit check predicate should be near the loop. 3774 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3775 } 3776 3777 //----------------------------- store barriers ---------------------------- 3778 #define __ ideal. 3779 3780 void GraphKit::sync_kit(IdealKit& ideal) { 3781 set_all_memory(__ merged_memory()); 3782 set_i_o(__ i_o()); 3783 set_control(__ ctrl()); 3784 } 3785 3786 void GraphKit::final_sync(IdealKit& ideal) { 3787 // Final sync IdealKit and graphKit. 3788 sync_kit(ideal); 3789 } 3790 3791 Node* GraphKit::byte_map_base_node() { 3792 // Get base of card map 3793 CardTableModRefBS* ct = 3794 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3795 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3796 if (ct->byte_map_base != NULL) { 3797 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3798 } else { 3799 return null(); 3800 } 3801 } 3802 3803 // vanilla/CMS post barrier 3804 // Insert a write-barrier store. This is to let generational GC work; we have 3805 // to flag all oop-stores before the next GC point. 3806 void GraphKit::write_barrier_post(Node* oop_store, 3807 Node* obj, 3808 Node* adr, 3809 uint adr_idx, 3810 Node* val, 3811 bool use_precise) { 3812 // No store check needed if we're storing a NULL or an old object 3813 // (latter case is probably a string constant). The concurrent 3814 // mark sweep garbage collector, however, needs to have all nonNull 3815 // oop updates flagged via card-marks. 3816 if (val != NULL && val->is_Con()) { 3817 // must be either an oop or NULL 3818 const Type* t = val->bottom_type(); 3819 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3820 // stores of null never (?) need barriers 3821 return; 3822 } 3823 3824 if (use_ReduceInitialCardMarks() 3825 && obj == just_allocated_object(control())) { 3826 // We can skip marks on a freshly-allocated object in Eden. 3827 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3828 // That routine informs GC to take appropriate compensating steps, 3829 // upon a slow-path allocation, so as to make this card-mark 3830 // elision safe. 3831 return; 3832 } 3833 3834 if (!use_precise) { 3835 // All card marks for a (non-array) instance are in one place: 3836 adr = obj; 3837 } 3838 // (Else it's an array (or unknown), and we want more precise card marks.) 3839 assert(adr != NULL, ""); 3840 3841 IdealKit ideal(this, true); 3842 3843 // Convert the pointer to an int prior to doing math on it 3844 Node* cast = __ CastPX(__ ctrl(), adr); 3845 3846 // Divide by card size 3847 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3848 "Only one we handle so far."); 3849 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3850 3851 // Combine card table base and card offset 3852 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3853 3854 // Get the alias_index for raw card-mark memory 3855 int adr_type = Compile::AliasIdxRaw; 3856 Node* zero = __ ConI(0); // Dirty card value 3857 BasicType bt = T_BYTE; 3858 3859 if (UseConcMarkSweepGC && UseCondCardMark) { 3860 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3861 __ sync_kit(this); 3862 } 3863 3864 if (UseCondCardMark) { 3865 // The classic GC reference write barrier is typically implemented 3866 // as a store into the global card mark table. Unfortunately 3867 // unconditional stores can result in false sharing and excessive 3868 // coherence traffic as well as false transactional aborts. 3869 // UseCondCardMark enables MP "polite" conditional card mark 3870 // stores. In theory we could relax the load from ctrl() to 3871 // no_ctrl, but that doesn't buy much latitude. 3872 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3873 __ if_then(card_val, BoolTest::ne, zero); 3874 } 3875 3876 // Smash zero into card 3877 if( !UseConcMarkSweepGC ) { 3878 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3879 } else { 3880 // Specialized path for CM store barrier 3881 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3882 } 3883 3884 if (UseCondCardMark) { 3885 __ end_if(); 3886 } 3887 3888 // Final sync IdealKit and GraphKit. 3889 final_sync(ideal); 3890 } 3891 /* 3892 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3893 * required by SATB to make sure all objects live at the start of the 3894 * marking are kept alive, all reference updates need to any previous 3895 * reference stored before writing. 3896 * 3897 * If the previous value is NULL there is no need to save the old value. 3898 * References that are NULL are filtered during runtime by the barrier 3899 * code to avoid unnecessary queuing. 3900 * 3901 * However in the case of newly allocated objects it might be possible to 3902 * prove that the reference about to be overwritten is NULL during compile 3903 * time and avoid adding the barrier code completely. 3904 * 3905 * The compiler needs to determine that the object in which a field is about 3906 * to be written is newly allocated, and that no prior store to the same field 3907 * has happened since the allocation. 3908 * 3909 * Returns true if the pre-barrier can be removed 3910 */ 3911 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3912 BasicType bt, uint adr_idx) { 3913 intptr_t offset = 0; 3914 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3915 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3916 3917 if (offset == Type::OffsetBot) { 3918 return false; // cannot unalias unless there are precise offsets 3919 } 3920 3921 if (alloc == NULL) { 3922 return false; // No allocation found 3923 } 3924 3925 intptr_t size_in_bytes = type2aelembytes(bt); 3926 3927 Node* mem = memory(adr_idx); // start searching here... 3928 3929 for (int cnt = 0; cnt < 50; cnt++) { 3930 3931 if (mem->is_Store()) { 3932 3933 Node* st_adr = mem->in(MemNode::Address); 3934 intptr_t st_offset = 0; 3935 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3936 3937 if (st_base == NULL) { 3938 break; // inscrutable pointer 3939 } 3940 3941 // Break we have found a store with same base and offset as ours so break 3942 if (st_base == base && st_offset == offset) { 3943 break; 3944 } 3945 3946 if (st_offset != offset && st_offset != Type::OffsetBot) { 3947 const int MAX_STORE = BytesPerLong; 3948 if (st_offset >= offset + size_in_bytes || 3949 st_offset <= offset - MAX_STORE || 3950 st_offset <= offset - mem->as_Store()->memory_size()) { 3951 // Success: The offsets are provably independent. 3952 // (You may ask, why not just test st_offset != offset and be done? 3953 // The answer is that stores of different sizes can co-exist 3954 // in the same sequence of RawMem effects. We sometimes initialize 3955 // a whole 'tile' of array elements with a single jint or jlong.) 3956 mem = mem->in(MemNode::Memory); 3957 continue; // advance through independent store memory 3958 } 3959 } 3960 3961 if (st_base != base 3962 && MemNode::detect_ptr_independence(base, alloc, st_base, 3963 AllocateNode::Ideal_allocation(st_base, phase), 3964 phase)) { 3965 // Success: The bases are provably independent. 3966 mem = mem->in(MemNode::Memory); 3967 continue; // advance through independent store memory 3968 } 3969 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3970 3971 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3972 AllocateNode* st_alloc = st_init->allocation(); 3973 3974 // Make sure that we are looking at the same allocation site. 3975 // The alloc variable is guaranteed to not be null here from earlier check. 3976 if (alloc == st_alloc) { 3977 // Check that the initialization is storing NULL so that no previous store 3978 // has been moved up and directly write a reference 3979 Node* captured_store = st_init->find_captured_store(offset, 3980 type2aelembytes(T_OBJECT), 3981 phase); 3982 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3983 return true; 3984 } 3985 } 3986 } 3987 3988 // Unless there is an explicit 'continue', we must bail out here, 3989 // because 'mem' is an inscrutable memory state (e.g., a call). 3990 break; 3991 } 3992 3993 return false; 3994 } 3995 3996 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) { 3997 if (UseShenandoahGC && ShenandoahSATBBarrier && adr != NULL) { 3998 Node* c = kit.control(); 3999 Node* call = c->in(1)->in(1)->in(1)->in(0); 4000 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected"); 4001 call->add_req(adr); 4002 } 4003 } 4004 4005 // G1 pre/post barriers 4006 void GraphKit::g1_write_barrier_pre(bool do_load, 4007 Node* obj, 4008 Node* adr, 4009 uint alias_idx, 4010 Node* val, 4011 const TypeOopPtr* val_type, 4012 Node* pre_val, 4013 BasicType bt) { 4014 4015 // Some sanity checks 4016 // Note: val is unused in this routine. 4017 4018 if (do_load) { 4019 // We need to generate the load of the previous value 4020 assert(obj != NULL, "must have a base"); 4021 assert(adr != NULL, "where are loading from?"); 4022 assert(pre_val == NULL, "loaded already?"); 4023 assert(val_type != NULL, "need a type"); 4024 4025 if (use_ReduceInitialCardMarks() 4026 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4027 return; 4028 } 4029 4030 } else { 4031 // In this case both val_type and alias_idx are unused. 4032 assert(pre_val != NULL, "must be loaded already"); 4033 // Nothing to be done if pre_val is null. 4034 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4035 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4036 } 4037 assert(bt == T_OBJECT, "or we shouldn't be here"); 4038 4039 IdealKit ideal(this, true); 4040 4041 Node* tls = __ thread(); // ThreadLocalStorage 4042 4043 Node* no_ctrl = NULL; 4044 Node* no_base = __ top(); 4045 Node* zero = __ ConI(0); 4046 Node* zeroX = __ ConX(0); 4047 4048 float likely = PROB_LIKELY(0.999); 4049 float unlikely = PROB_UNLIKELY(0.999); 4050 4051 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4052 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4053 4054 // Offsets into the thread 4055 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4056 SATBMarkQueue::byte_offset_of_active()); 4057 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4058 SATBMarkQueue::byte_offset_of_index()); 4059 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4060 SATBMarkQueue::byte_offset_of_buf()); 4061 4062 // Now the actual pointers into the thread 4063 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4064 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4065 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4066 4067 // Now some of the values 4068 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4069 4070 // if (!marking) 4071 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4072 BasicType index_bt = TypeX_X->basic_type(); 4073 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4074 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4075 4076 if (do_load) { 4077 // load original value 4078 // alias_idx correct?? 4079 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4080 } 4081 4082 // if (pre_val != NULL) 4083 __ if_then(pre_val, BoolTest::ne, null()); { 4084 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4085 4086 // is the queue for this thread full? 4087 __ if_then(index, BoolTest::ne, zeroX, likely); { 4088 4089 // decrement the index 4090 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4091 4092 // Now get the buffer location we will log the previous value into and store it 4093 Node *log_addr = __ AddP(no_base, buffer, next_index); 4094 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4095 // update the index 4096 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4097 4098 } __ else_(); { 4099 4100 // logging buffer is full, call the runtime 4101 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4102 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4103 } __ end_if(); // (!index) 4104 } __ end_if(); // (pre_val != NULL) 4105 } __ end_if(); // (!marking) 4106 4107 // Final sync IdealKit and GraphKit. 4108 final_sync(ideal); 4109 g1_write_barrier_pre_helper(*this, adr); 4110 } 4111 4112 Node* GraphKit::shenandoah_write_barrier_pre(bool do_load, 4113 Node* obj, 4114 Node* adr, 4115 uint alias_idx, 4116 Node* val, 4117 const TypeOopPtr* val_type, 4118 Node* pre_val, 4119 BasicType bt) { 4120 4121 // Some sanity checks 4122 // Note: val is unused in this routine. 4123 4124 if (!ShenandoahSATBBarrier) { 4125 if (val != NULL) { 4126 shenandoah_update_matrix(adr, val); 4127 } 4128 return val; 4129 } 4130 4131 if (val == NULL) { 4132 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 4133 return NULL; 4134 } 4135 4136 if (! ShenandoahReduceStoreValBarrier) { 4137 val = shenandoah_read_barrier_storeval(val); 4138 shenandoah_update_matrix(adr, val); 4139 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 4140 return val; 4141 } 4142 4143 if (do_load) { 4144 // We need to generate the load of the previous value 4145 assert(obj != NULL, "must have a base"); 4146 assert(adr != NULL, "where are loading from?"); 4147 assert(pre_val == NULL, "loaded already?"); 4148 assert(val_type != NULL, "need a type"); 4149 4150 if (use_ReduceInitialCardMarks() 4151 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4152 return shenandoah_read_barrier_storeval(val); 4153 } 4154 4155 } else { 4156 // In this case both val_type and alias_idx are unused. 4157 assert(pre_val != NULL, "must be loaded already"); 4158 // Nothing to be done if pre_val is null. 4159 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return val; 4160 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4161 } 4162 assert(bt == T_OBJECT, "or we shouldn't be here"); 4163 4164 IdealKit ideal(this, true, true); 4165 IdealVariable ival(ideal); 4166 __ declarations_done(); 4167 __ set(ival, val); 4168 Node* tls = __ thread(); // ThreadLocalStorage 4169 4170 Node* no_ctrl = NULL; 4171 Node* no_base = __ top(); 4172 Node* zero = __ ConI(0); 4173 Node* zeroX = __ ConX(0); 4174 4175 float likely = PROB_LIKELY(0.999); 4176 float unlikely = PROB_UNLIKELY(0.999); 4177 4178 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4179 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4180 4181 // Offsets into the thread 4182 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4183 SATBMarkQueue::byte_offset_of_active()); 4184 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4185 SATBMarkQueue::byte_offset_of_index()); 4186 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4187 SATBMarkQueue::byte_offset_of_buf()); 4188 4189 // Now the actual pointers into the thread 4190 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4191 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4192 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4193 4194 // Now some of the values 4195 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4196 4197 // if (!marking) 4198 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4199 4200 Node* storeval = ideal.value(ival); 4201 sync_kit(ideal); 4202 storeval = shenandoah_read_barrier_storeval(storeval); 4203 __ sync_kit(this); 4204 __ set(ival, storeval); 4205 4206 BasicType index_bt = TypeX_X->basic_type(); 4207 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4208 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4209 4210 if (do_load) { 4211 // load original value 4212 // alias_idx correct?? 4213 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4214 } 4215 4216 // if (pre_val != NULL) 4217 __ if_then(pre_val, BoolTest::ne, null()); { 4218 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4219 4220 // is the queue for this thread full? 4221 __ if_then(index, BoolTest::ne, zeroX, likely); { 4222 4223 // decrement the index 4224 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4225 4226 // Now get the buffer location we will log the previous value into and store it 4227 Node *log_addr = __ AddP(no_base, buffer, next_index); 4228 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4229 // update the index 4230 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4231 4232 } __ else_(); { 4233 4234 // logging buffer is full, call the runtime 4235 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4236 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4237 } __ end_if(); // (!index) 4238 } __ end_if(); // (pre_val != NULL) 4239 } __ end_if(); // (!marking) 4240 Node* new_val = __ value(ival); 4241 // IdealKit generates a Phi with very conservative type, and even 4242 // turns array types into TypeInstPtr (see type.cpp, _const_basic_type[T_ARRAY]). 4243 // We're forcing the result to be the original type. 4244 if (new_val != val) { 4245 const Type* t = _gvn.type(val); 4246 if (new_val->isa_Type()) { 4247 new_val->as_Type()->set_type(t); 4248 } 4249 _gvn.set_type(new_val, t); 4250 } 4251 val = new_val; 4252 __ dead(ival); 4253 // Final sync IdealKit and GraphKit. 4254 final_sync(ideal); 4255 g1_write_barrier_pre_helper(*this, adr); 4256 return val; 4257 } 4258 4259 void GraphKit::shenandoah_update_matrix(Node* adr, Node* val) { 4260 if (!UseShenandoahMatrix) { 4261 return; 4262 } 4263 4264 assert(val != NULL, "checked before"); 4265 if (adr == NULL) { 4266 return; // Nothing to do 4267 } 4268 assert(adr != NULL, "must not happen"); 4269 if (val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) { 4270 // Nothing to do. 4271 return; 4272 } 4273 4274 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix(); 4275 4276 enum { _set_path = 1, _already_set_path, _val_null_path, PATH_LIMIT }; 4277 RegionNode* region = new RegionNode(PATH_LIMIT); 4278 Node* prev_mem = memory(Compile::AliasIdxRaw); 4279 Node* memphi = PhiNode::make(region, prev_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 4280 Node* null_ctrl = top(); 4281 Node* not_null_val = null_check_oop(val, &null_ctrl); 4282 4283 // Null path: nothing to do. 4284 region->init_req(_val_null_path, null_ctrl); 4285 memphi->init_req(_val_null_path, prev_mem); 4286 4287 // Not null path. Update the matrix. 4288 4289 // This uses a fast calculation for the matrix address. For a description, 4290 // see src/share/vm/gc/shenandoah/shenandoahConnectionMatrix.inline.hpp, 4291 // ShenandoahConnectionMatrix::compute_address(const void* from, const void* to). 4292 address heap_base = ShenandoahHeap::heap()->base(); 4293 jint stride = matrix->stride_jint(); 4294 jint rs = ShenandoahHeapRegion::region_size_shift_jint(); 4295 4296 guarantee(stride < ShenandoahHeapRegion::region_size_bytes_jint(), "sanity"); 4297 guarantee(is_ptr_aligned(heap_base, ShenandoahHeapRegion::region_size_bytes()), "sanity"); 4298 4299 Node* magic_con = MakeConX((jlong) matrix->matrix_addr() - ((jlong) heap_base >> rs) * (stride + 1)); 4300 4301 // Compute addr part 4302 Node* adr_idx = _gvn.transform(new CastP2XNode(control(), adr)); 4303 adr_idx = _gvn.transform(new URShiftXNode(adr_idx, intcon(rs))); 4304 4305 // Compute new_val part 4306 Node* val_idx = _gvn.transform(new CastP2XNode(control(), not_null_val)); 4307 val_idx = _gvn.transform(new URShiftXNode(val_idx, intcon(rs))); 4308 val_idx = _gvn.transform(new MulXNode(val_idx, MakeConX(stride))); 4309 4310 // Add everything up 4311 adr_idx = _gvn.transform(new AddXNode(adr_idx, val_idx)); 4312 adr_idx = _gvn.transform(new CastX2PNode(adr_idx)); 4313 Node* matrix_adr = _gvn.transform(new AddPNode(top(), adr_idx, magic_con)); 4314 4315 // Load current value 4316 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 4317 Node* current = _gvn.transform(LoadNode::make(_gvn, control(), memory(Compile::AliasIdxRaw), 4318 matrix_adr, adr_type, TypeInt::INT, T_BYTE, MemNode::unordered)); 4319 4320 // Check if already set 4321 Node* cmp_set = _gvn.transform(new CmpINode(current, intcon(0))); 4322 Node* cmp_set_bool = _gvn.transform(new BoolNode(cmp_set, BoolTest::eq)); 4323 IfNode* cmp_iff = create_and_map_if(control(), cmp_set_bool, PROB_MIN, COUNT_UNKNOWN); 4324 4325 Node* if_not_set = _gvn.transform(new IfTrueNode(cmp_iff)); 4326 Node* if_set = _gvn.transform(new IfFalseNode(cmp_iff)); 4327 4328 // Already set, exit 4329 set_control(if_set); 4330 region->init_req(_already_set_path, control()); 4331 memphi->init_req(_already_set_path, prev_mem); 4332 4333 // Not set: do the store, and finish up 4334 set_control(if_not_set); 4335 Node* store = _gvn.transform(StoreNode::make(_gvn, control(), memory(Compile::AliasIdxRaw), 4336 matrix_adr, adr_type, intcon(1), T_BYTE, MemNode::unordered)); 4337 region->init_req(_set_path, control()); 4338 memphi->init_req(_set_path, store); 4339 4340 // Merge control flows and memory. 4341 set_control(_gvn.transform(region)); 4342 record_for_igvn(region); 4343 set_memory(_gvn.transform(memphi), Compile::AliasIdxRaw); 4344 } 4345 4346 /* 4347 * G1 similar to any GC with a Young Generation requires a way to keep track of 4348 * references from Old Generation to Young Generation to make sure all live 4349 * objects are found. G1 also requires to keep track of object references 4350 * between different regions to enable evacuation of old regions, which is done 4351 * as part of mixed collections. References are tracked in remembered sets and 4352 * is continuously updated as reference are written to with the help of the 4353 * post-barrier. 4354 * 4355 * To reduce the number of updates to the remembered set the post-barrier 4356 * filters updates to fields in objects located in the Young Generation, 4357 * the same region as the reference, when the NULL is being written or 4358 * if the card is already marked as dirty by an earlier write. 4359 * 4360 * Under certain circumstances it is possible to avoid generating the 4361 * post-barrier completely if it is possible during compile time to prove 4362 * the object is newly allocated and that no safepoint exists between the 4363 * allocation and the store. 4364 * 4365 * In the case of slow allocation the allocation code must handle the barrier 4366 * as part of the allocation in the case the allocated object is not located 4367 * in the nursery, this would happen for humongous objects. This is similar to 4368 * how CMS is required to handle this case, see the comments for the method 4369 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4370 * A deferred card mark is required for these objects and handled in the above 4371 * mentioned methods. 4372 * 4373 * Returns true if the post barrier can be removed 4374 */ 4375 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4376 Node* adr) { 4377 intptr_t offset = 0; 4378 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4379 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4380 4381 if (offset == Type::OffsetBot) { 4382 return false; // cannot unalias unless there are precise offsets 4383 } 4384 4385 if (alloc == NULL) { 4386 return false; // No allocation found 4387 } 4388 4389 // Start search from Store node 4390 Node* mem = store->in(MemNode::Control); 4391 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4392 4393 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4394 AllocateNode* st_alloc = st_init->allocation(); 4395 4396 // Make sure we are looking at the same allocation 4397 if (alloc == st_alloc) { 4398 return true; 4399 } 4400 } 4401 4402 return false; 4403 } 4404 4405 // 4406 // Update the card table and add card address to the queue 4407 // 4408 void GraphKit::g1_mark_card(IdealKit& ideal, 4409 Node* card_adr, 4410 Node* oop_store, 4411 uint oop_alias_idx, 4412 Node* index, 4413 Node* index_adr, 4414 Node* buffer, 4415 const TypeFunc* tf) { 4416 4417 Node* zero = __ ConI(0); 4418 Node* zeroX = __ ConX(0); 4419 Node* no_base = __ top(); 4420 BasicType card_bt = T_BYTE; 4421 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4422 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4423 4424 // Now do the queue work 4425 __ if_then(index, BoolTest::ne, zeroX); { 4426 4427 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4428 Node* log_addr = __ AddP(no_base, buffer, next_index); 4429 4430 // Order, see storeCM. 4431 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4432 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4433 4434 } __ else_(); { 4435 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4436 } __ end_if(); 4437 4438 } 4439 4440 void GraphKit::g1_write_barrier_post(Node* oop_store, 4441 Node* obj, 4442 Node* adr, 4443 uint alias_idx, 4444 Node* val, 4445 BasicType bt, 4446 bool use_precise) { 4447 // If we are writing a NULL then we need no post barrier 4448 4449 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4450 // Must be NULL 4451 const Type* t = val->bottom_type(); 4452 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4453 // No post barrier if writing NULLx 4454 return; 4455 } 4456 4457 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4458 // We can skip marks on a freshly-allocated object in Eden. 4459 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4460 // That routine informs GC to take appropriate compensating steps, 4461 // upon a slow-path allocation, so as to make this card-mark 4462 // elision safe. 4463 return; 4464 } 4465 4466 if (use_ReduceInitialCardMarks() 4467 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4468 return; 4469 } 4470 4471 if (!use_precise) { 4472 // All card marks for a (non-array) instance are in one place: 4473 adr = obj; 4474 } 4475 // (Else it's an array (or unknown), and we want more precise card marks.) 4476 assert(adr != NULL, ""); 4477 4478 IdealKit ideal(this, true); 4479 4480 Node* tls = __ thread(); // ThreadLocalStorage 4481 4482 Node* no_base = __ top(); 4483 float likely = PROB_LIKELY(0.999); 4484 float unlikely = PROB_UNLIKELY(0.999); 4485 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4486 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4487 Node* zeroX = __ ConX(0); 4488 4489 // Get the alias_index for raw card-mark memory 4490 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4491 4492 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4493 4494 // Offsets into the thread 4495 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4496 DirtyCardQueue::byte_offset_of_index()); 4497 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4498 DirtyCardQueue::byte_offset_of_buf()); 4499 4500 // Pointers into the thread 4501 4502 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4503 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4504 4505 // Now some values 4506 // Use ctrl to avoid hoisting these values past a safepoint, which could 4507 // potentially reset these fields in the JavaThread. 4508 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4509 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4510 4511 // Convert the store obj pointer to an int prior to doing math on it 4512 // Must use ctrl to prevent "integerized oop" existing across safepoint 4513 Node* cast = __ CastPX(__ ctrl(), adr); 4514 4515 // Divide pointer by card size 4516 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4517 4518 // Combine card table base and card offset 4519 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4520 4521 // If we know the value being stored does it cross regions? 4522 4523 if (val != NULL) { 4524 // Does the store cause us to cross regions? 4525 4526 // Should be able to do an unsigned compare of region_size instead of 4527 // and extra shift. Do we have an unsigned compare?? 4528 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4529 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4530 4531 // if (xor_res == 0) same region so skip 4532 __ if_then(xor_res, BoolTest::ne, zeroX); { 4533 4534 // No barrier if we are storing a NULL 4535 __ if_then(val, BoolTest::ne, null(), unlikely); { 4536 4537 // Ok must mark the card if not already dirty 4538 4539 // load the original value of the card 4540 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4541 4542 __ if_then(card_val, BoolTest::ne, young_card); { 4543 sync_kit(ideal); 4544 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4545 insert_mem_bar(Op_MemBarVolatile, oop_store); 4546 __ sync_kit(this); 4547 4548 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4549 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4550 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4551 } __ end_if(); 4552 } __ end_if(); 4553 } __ end_if(); 4554 } __ end_if(); 4555 } else { 4556 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4557 // We don't need a barrier here if the destination is a newly allocated object 4558 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4559 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4560 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4561 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4562 __ if_then(card_val, BoolTest::ne, young_card); { 4563 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4564 } __ end_if(); 4565 } 4566 4567 // Final sync IdealKit and GraphKit. 4568 final_sync(ideal); 4569 } 4570 #undef __ 4571 4572 4573 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4574 Node* len = load_array_length(load_String_value(ctrl, str)); 4575 Node* coder = load_String_coder(ctrl, str); 4576 // Divide length by 2 if coder is UTF16 4577 return _gvn.transform(new RShiftINode(len, coder)); 4578 } 4579 4580 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4581 int value_offset = java_lang_String::value_offset_in_bytes(); 4582 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4583 false, NULL, 0); 4584 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4585 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4586 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4587 ciTypeArrayKlass::make(T_BYTE), true, 0); 4588 int value_field_idx = C->get_alias_index(value_field_type); 4589 4590 if (! ShenandoahOptimizeFinals) { 4591 str = shenandoah_read_barrier(str); 4592 } 4593 4594 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4595 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4596 // String.value field is known to be @Stable. 4597 if (UseImplicitStableValues) { 4598 load = cast_array_to_stable(load, value_type); 4599 } 4600 return load; 4601 } 4602 4603 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4604 if (!CompactStrings) { 4605 return intcon(java_lang_String::CODER_UTF16); 4606 } 4607 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4608 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4609 false, NULL, 0); 4610 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4611 int coder_field_idx = C->get_alias_index(coder_field_type); 4612 4613 if (! ShenandoahOptimizeFinals) { 4614 str = shenandoah_read_barrier(str); 4615 } 4616 4617 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4618 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4619 } 4620 4621 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4622 int value_offset = java_lang_String::value_offset_in_bytes(); 4623 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4624 false, NULL, 0); 4625 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4626 4627 str = shenandoah_write_barrier(str); 4628 4629 store_oop_to_object(control(), str, basic_plus_adr(str, value_offset), value_field_type, 4630 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4631 } 4632 4633 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4634 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4635 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4636 false, NULL, 0); 4637 4638 str = shenandoah_write_barrier(str); 4639 4640 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4641 int coder_field_idx = C->get_alias_index(coder_field_type); 4642 store_to_memory(control(), basic_plus_adr(str, coder_offset), 4643 value, T_BYTE, coder_field_idx, MemNode::unordered); 4644 } 4645 4646 // Capture src and dst memory state with a MergeMemNode 4647 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4648 if (src_type == dst_type) { 4649 // Types are equal, we don't need a MergeMemNode 4650 return memory(src_type); 4651 } 4652 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4653 record_for_igvn(merge); // fold it up later, if possible 4654 int src_idx = C->get_alias_index(src_type); 4655 int dst_idx = C->get_alias_index(dst_type); 4656 merge->set_memory_at(src_idx, memory(src_idx)); 4657 merge->set_memory_at(dst_idx, memory(dst_idx)); 4658 return merge; 4659 } 4660 4661 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4662 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4663 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4664 // If input and output memory types differ, capture both states to preserve 4665 // the dependency between preceding and subsequent loads/stores. 4666 // For example, the following program: 4667 // StoreB 4668 // compress_string 4669 // LoadB 4670 // has this memory graph (use->def): 4671 // LoadB -> compress_string -> CharMem 4672 // ... -> StoreB -> ByteMem 4673 // The intrinsic hides the dependency between LoadB and StoreB, causing 4674 // the load to read from memory not containing the result of the StoreB. 4675 // The correct memory graph should look like this: 4676 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4677 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4678 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4679 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4680 set_memory(res_mem, TypeAryPtr::BYTES); 4681 return str; 4682 } 4683 4684 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4685 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4686 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4687 // Capture src and dst memory (see comment in 'compress_string'). 4688 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4689 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4690 set_memory(_gvn.transform(str), dst_type); 4691 } 4692 4693 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4694 4695 src = shenandoah_read_barrier(src); 4696 dst = shenandoah_write_barrier(dst); 4697 4698 /** 4699 * int i_char = start; 4700 * for (int i_byte = 0; i_byte < count; i_byte++) { 4701 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4702 * } 4703 */ 4704 add_predicate(); 4705 RegionNode* head = new RegionNode(3); 4706 head->init_req(1, control()); 4707 gvn().set_type(head, Type::CONTROL); 4708 record_for_igvn(head); 4709 4710 Node* i_byte = new PhiNode(head, TypeInt::INT); 4711 i_byte->init_req(1, intcon(0)); 4712 gvn().set_type(i_byte, TypeInt::INT); 4713 record_for_igvn(i_byte); 4714 4715 Node* i_char = new PhiNode(head, TypeInt::INT); 4716 i_char->init_req(1, start); 4717 gvn().set_type(i_char, TypeInt::INT); 4718 record_for_igvn(i_char); 4719 4720 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4721 gvn().set_type(mem, Type::MEMORY); 4722 record_for_igvn(mem); 4723 set_control(head); 4724 set_memory(mem, TypeAryPtr::BYTES); 4725 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4726 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4727 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4728 false, false, true /* mismatched */); 4729 4730 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4731 head->init_req(2, IfTrue(iff)); 4732 mem->init_req(2, st); 4733 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4734 i_char->init_req(2, AddI(i_char, intcon(2))); 4735 4736 set_control(IfFalse(iff)); 4737 set_memory(st, TypeAryPtr::BYTES); 4738 } 4739 4740 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4741 if (!field->is_constant()) { 4742 return NULL; // Field not marked as constant. 4743 } 4744 ciInstance* holder = NULL; 4745 if (!field->is_static()) { 4746 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4747 if (const_oop != NULL && const_oop->is_instance()) { 4748 holder = const_oop->as_instance(); 4749 } 4750 } 4751 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4752 /*is_unsigned_load=*/false); 4753 if (con_type != NULL) { 4754 return makecon(con_type); 4755 } 4756 return NULL; 4757 } 4758 4759 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4760 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4761 // assumption of CCP analysis. 4762 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4763 } 4764 4765 Node* GraphKit::shenandoah_read_barrier(Node* obj) { 4766 return shenandoah_read_barrier_impl(obj, false, true, true); 4767 } 4768 4769 Node* GraphKit::shenandoah_read_barrier_storeval(Node* obj) { 4770 return shenandoah_read_barrier_impl(obj, true, false, false); 4771 } 4772 4773 Node* GraphKit::shenandoah_read_barrier_impl(Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) { 4774 4775 if (UseShenandoahGC && ShenandoahReadBarrier) { 4776 const Type* obj_type = obj->bottom_type(); 4777 if (obj_type->higher_equal(TypePtr::NULL_PTR)) { 4778 return obj; 4779 } 4780 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 4781 Node* mem = use_mem ? memory(adr_type) : immutable_memory(); 4782 4783 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, mem, allow_fromspace)) { 4784 // We know it is null, no barrier needed. 4785 return obj; 4786 } 4787 4788 4789 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4790 4791 // We don't know if it's null or not. Need null-check. 4792 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4793 RegionNode* region = new RegionNode(PATH_LIMIT); 4794 Node* phi = new PhiNode(region, obj_type); 4795 Node* null_ctrl = top(); 4796 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4797 4798 region->init_req(_null_path, null_ctrl); 4799 phi ->init_req(_null_path, zerocon(T_OBJECT)); 4800 4801 Node* ctrl = use_ctrl ? control() : NULL; 4802 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace); 4803 Node* n = _gvn.transform(rb); 4804 4805 region->init_req(_not_null_path, control()); 4806 phi ->init_req(_not_null_path, n); 4807 4808 set_control(_gvn.transform(region)); 4809 record_for_igvn(region); 4810 return _gvn.transform(phi); 4811 4812 } else { 4813 // We know it is not null. Simple barrier is sufficient. 4814 Node* ctrl = use_ctrl ? control() : NULL; 4815 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace); 4816 Node* n = _gvn.transform(rb); 4817 record_for_igvn(n); 4818 return n; 4819 } 4820 4821 } else { 4822 return obj; 4823 } 4824 } 4825 4826 static Node* shenandoah_write_barrier_helper(GraphKit& kit, Node* obj, const TypePtr* adr_type) { 4827 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit.C, kit.control(), kit.memory(adr_type), obj); 4828 Node* n = kit.gvn().transform(wb); 4829 if (n == wb) { // New barrier needs memory projection. 4830 Node* proj = kit.gvn().transform(new ShenandoahWBMemProjNode(n)); 4831 kit.set_memory(proj, adr_type); 4832 } 4833 4834 return n; 4835 } 4836 4837 Node* GraphKit::shenandoah_write_barrier(Node* obj) { 4838 4839 if (UseShenandoahGC && ShenandoahWriteBarrier) { 4840 4841 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, NULL, true)) { 4842 return obj; 4843 } 4844 const Type* obj_type = obj->bottom_type(); 4845 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 4846 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4847 // We don't know if it's null or not. Need null-check. 4848 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4849 RegionNode* region = new RegionNode(PATH_LIMIT); 4850 Node* phi = new PhiNode(region, obj_type); 4851 Node* memphi = PhiNode::make(region, memory(adr_type), Type::MEMORY, C->alias_type(adr_type)->adr_type()); 4852 4853 Node* prev_mem = memory(adr_type); 4854 Node* null_ctrl = top(); 4855 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4856 4857 region->init_req(_null_path, null_ctrl); 4858 phi ->init_req(_null_path, zerocon(T_OBJECT)); 4859 memphi->init_req(_null_path, prev_mem); 4860 4861 Node* n = shenandoah_write_barrier_helper(*this, not_null_obj, adr_type); 4862 4863 region->init_req(_not_null_path, control()); 4864 phi ->init_req(_not_null_path, n); 4865 memphi->init_req(_not_null_path, memory(adr_type)); 4866 4867 set_control(_gvn.transform(region)); 4868 record_for_igvn(region); 4869 set_memory(_gvn.transform(memphi), adr_type); 4870 4871 Node* res_val = _gvn.transform(phi); 4872 // replace_in_map(obj, res_val); 4873 return res_val; 4874 } else { 4875 // We know it is not null. Simple barrier is sufficient. 4876 Node* n = shenandoah_write_barrier_helper(*this, obj, adr_type); 4877 // replace_in_map(obj, n); 4878 record_for_igvn(n); 4879 return n; 4880 } 4881 4882 } else { 4883 return obj; 4884 } 4885 } 4886 4887 /** 4888 * In Shenandoah, we need barriers on acmp (and similar instructions that compare two 4889 * oops) to avoid false negatives. If it compares a from-space and a to-space 4890 * copy of an object, a regular acmp would return false, even though both are 4891 * the same. The acmp barrier compares the two objects, and when they are 4892 * *not equal* it does a read-barrier on both, and compares them again. When it 4893 * failed because of different copies of the object, we know that the object 4894 * must already have been evacuated (and therefore doesn't require a write-barrier). 4895 */ 4896 Node* GraphKit::cmp_objects(Node* a, Node* b) { 4897 // TODO: Refactor into proper GC interface. 4898 if (UseShenandoahGC && ShenandoahAcmpBarrier) { 4899 const Type* a_type = a->bottom_type(); 4900 const Type* b_type = b->bottom_type(); 4901 if (a_type->higher_equal(TypePtr::NULL_PTR) || b_type->higher_equal(TypePtr::NULL_PTR)) { 4902 // We know one arg is gonna be null. No need for barriers. 4903 return _gvn.transform(new CmpPNode(b, a)); 4904 } 4905 4906 const TypePtr* a_adr_type = ShenandoahBarrierNode::brooks_pointer_type(a_type); 4907 const TypePtr* b_adr_type = ShenandoahBarrierNode::brooks_pointer_type(b_type); 4908 if ((! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, a, memory(a_adr_type), false)) && 4909 (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, b, memory(b_adr_type), false))) { 4910 // We know both args are in to-space already. No acmp barrier needed. 4911 return _gvn.transform(new CmpPNode(b, a)); 4912 } 4913 4914 C->set_has_split_ifs(true); 4915 4916 if (ShenandoahVerifyOptoBarriers) { 4917 a = shenandoah_write_barrier(a); 4918 b = shenandoah_write_barrier(b); 4919 return _gvn.transform(new CmpPNode(b, a)); 4920 } 4921 4922 enum { _equal = 1, _not_equal, PATH_LIMIT }; 4923 RegionNode* region = new RegionNode(PATH_LIMIT); 4924 PhiNode* phiA = PhiNode::make(region, a, _gvn.type(a)->is_oopptr()->cast_to_nonconst()); 4925 PhiNode* phiB = PhiNode::make(region, b, _gvn.type(b)->is_oopptr()->cast_to_nonconst()); 4926 4927 Node* cmp = _gvn.transform(new CmpPNode(b, a)); 4928 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 4929 4930 // TODO: Use profiling data. 4931 IfNode* iff = create_and_map_if(control(), tst, PROB_FAIR, COUNT_UNKNOWN); 4932 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4933 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4934 4935 // Equal path: Use original values. 4936 region->init_req(_equal, iftrue); 4937 phiA->init_req(_equal, a); 4938 phiB->init_req(_equal, b); 4939 4940 uint alias_a = C->get_alias_index(a_adr_type); 4941 uint alias_b = C->get_alias_index(b_adr_type); 4942 PhiNode* mem_phi = NULL; 4943 if (alias_a == alias_b) { 4944 mem_phi = PhiNode::make(region, memory(alias_a), Type::MEMORY, C->get_adr_type(alias_a)); 4945 } else { 4946 mem_phi = PhiNode::make(region, map()->memory(), Type::MEMORY, TypePtr::BOTTOM); 4947 } 4948 4949 // Unequal path: retry after read barriers. 4950 set_control(iffalse); 4951 if (!iffalse->is_top()) { 4952 Node* mb = NULL; 4953 if (alias_a == alias_b) { 4954 Node* mem = reset_memory(); 4955 mb = MemBarNode::make(C, Op_MemBarAcquire, alias_a); 4956 mb->init_req(TypeFunc::Control, control()); 4957 mb->init_req(TypeFunc::Memory, mem); 4958 Node* membar = _gvn.transform(mb); 4959 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 4960 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory)); 4961 set_all_memory(mem); 4962 set_memory(newmem, alias_a); 4963 } else { 4964 mb = insert_mem_bar(Op_MemBarAcquire); 4965 } 4966 } else { 4967 a = top(); 4968 b = top(); 4969 } 4970 4971 a = shenandoah_read_barrier_impl(a, true, true, false); 4972 b = shenandoah_read_barrier_impl(b, true, true, false); 4973 4974 region->init_req(_not_equal, control()); 4975 phiA->init_req(_not_equal, a); 4976 phiB->init_req(_not_equal, b); 4977 if (alias_a == alias_b) { 4978 mem_phi->init_req(_not_equal, memory(alias_a)); 4979 set_memory(mem_phi, alias_a); 4980 } else { 4981 mem_phi->init_req(_not_equal, reset_memory()); 4982 set_all_memory(mem_phi); 4983 } 4984 record_for_igvn(mem_phi); 4985 _gvn.set_type(mem_phi, Type::MEMORY); 4986 set_control(_gvn.transform(region)); 4987 record_for_igvn(region); 4988 4989 a = _gvn.transform(phiA); 4990 b = _gvn.transform(phiB); 4991 return _gvn.transform(new CmpPNode(b, a)); 4992 } else { 4993 return _gvn.transform(new CmpPNode(b, a)); 4994 } 4995 }