1 /* 2 * Copyright (c) 2001, 2017, 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 default: 593 break; 594 } 595 if (failing()) { stop(); return; } // exception allocation might fail 596 if (ex_obj != NULL) { 597 // Cheat with a preallocated exception object. 598 if (C->log() != NULL) 599 C->log()->elem("hot_throw preallocated='1' reason='%s'", 600 Deoptimization::trap_reason_name(reason)); 601 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 602 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 603 604 ex_node = shenandoah_write_barrier(ex_node); 605 606 // Clear the detail message of the preallocated exception object. 607 // Weblogic sometimes mutates the detail message of exceptions 608 // using reflection. 609 int offset = java_lang_Throwable::get_detailMessage_offset(); 610 const TypePtr* adr_typ = ex_con->add_offset(offset); 611 612 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 613 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 614 // Conservatively release stores of object references. 615 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 616 617 add_exception_state(make_exception_state(ex_node)); 618 return; 619 } 620 } 621 622 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 623 // It won't be much cheaper than bailing to the interp., since we'll 624 // have to pass up all the debug-info, and the runtime will have to 625 // create the stack trace. 626 627 // Usual case: Bail to interpreter. 628 // Reserve the right to recompile if we haven't seen anything yet. 629 630 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 631 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 632 if (treat_throw_as_hot 633 && (method()->method_data()->trap_recompiled_at(bci(), m) 634 || C->too_many_traps(reason))) { 635 // We cannot afford to take more traps here. Suffer in the interpreter. 636 if (C->log() != NULL) 637 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 638 Deoptimization::trap_reason_name(reason), 639 C->trap_count(reason)); 640 action = Deoptimization::Action_none; 641 } 642 643 // "must_throw" prunes the JVM state to include only the stack, if there 644 // are no local exception handlers. This should cut down on register 645 // allocation time and code size, by drastically reducing the number 646 // of in-edges on the call to the uncommon trap. 647 648 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 649 } 650 651 652 //----------------------------PreserveJVMState--------------------------------- 653 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 654 debug_only(kit->verify_map()); 655 _kit = kit; 656 _map = kit->map(); // preserve the map 657 _sp = kit->sp(); 658 kit->set_map(clone_map ? kit->clone_map() : NULL); 659 #ifdef ASSERT 660 _bci = kit->bci(); 661 Parse* parser = kit->is_Parse(); 662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 663 _block = block; 664 #endif 665 } 666 PreserveJVMState::~PreserveJVMState() { 667 GraphKit* kit = _kit; 668 #ifdef ASSERT 669 assert(kit->bci() == _bci, "bci must not shift"); 670 Parse* parser = kit->is_Parse(); 671 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 672 assert(block == _block, "block must not shift"); 673 #endif 674 kit->set_map(_map); 675 kit->set_sp(_sp); 676 } 677 678 679 //-----------------------------BuildCutout------------------------------------- 680 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 681 : PreserveJVMState(kit) 682 { 683 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 684 SafePointNode* outer_map = _map; // preserved map is caller's 685 SafePointNode* inner_map = kit->map(); 686 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 687 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 688 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 689 } 690 BuildCutout::~BuildCutout() { 691 GraphKit* kit = _kit; 692 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 693 } 694 695 //---------------------------PreserveReexecuteState---------------------------- 696 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 697 assert(!kit->stopped(), "must call stopped() before"); 698 _kit = kit; 699 _sp = kit->sp(); 700 _reexecute = kit->jvms()->_reexecute; 701 } 702 PreserveReexecuteState::~PreserveReexecuteState() { 703 if (_kit->stopped()) return; 704 _kit->jvms()->_reexecute = _reexecute; 705 _kit->set_sp(_sp); 706 } 707 708 //------------------------------clone_map-------------------------------------- 709 // Implementation of PreserveJVMState 710 // 711 // Only clone_map(...) here. If this function is only used in the 712 // PreserveJVMState class we may want to get rid of this extra 713 // function eventually and do it all there. 714 715 SafePointNode* GraphKit::clone_map() { 716 if (map() == NULL) return NULL; 717 718 // Clone the memory edge first 719 Node* mem = MergeMemNode::make(map()->memory()); 720 gvn().set_type_bottom(mem); 721 722 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 723 JVMState* jvms = this->jvms(); 724 JVMState* clonejvms = jvms->clone_shallow(C); 725 clonemap->set_memory(mem); 726 clonemap->set_jvms(clonejvms); 727 clonejvms->set_map(clonemap); 728 record_for_igvn(clonemap); 729 gvn().set_type_bottom(clonemap); 730 return clonemap; 731 } 732 733 734 //-----------------------------set_map_clone----------------------------------- 735 void GraphKit::set_map_clone(SafePointNode* m) { 736 _map = m; 737 _map = clone_map(); 738 _map->set_next_exception(NULL); 739 debug_only(verify_map()); 740 } 741 742 743 //----------------------------kill_dead_locals--------------------------------- 744 // Detect any locals which are known to be dead, and force them to top. 745 void GraphKit::kill_dead_locals() { 746 // Consult the liveness information for the locals. If any 747 // of them are unused, then they can be replaced by top(). This 748 // should help register allocation time and cut down on the size 749 // of the deoptimization information. 750 751 // This call is made from many of the bytecode handling 752 // subroutines called from the Big Switch in do_one_bytecode. 753 // Every bytecode which might include a slow path is responsible 754 // for killing its dead locals. The more consistent we 755 // are about killing deads, the fewer useless phis will be 756 // constructed for them at various merge points. 757 758 // bci can be -1 (InvocationEntryBci). We return the entry 759 // liveness for the method. 760 761 if (method() == NULL || method()->code_size() == 0) { 762 // We are building a graph for a call to a native method. 763 // All locals are live. 764 return; 765 } 766 767 ResourceMark rm; 768 769 // Consult the liveness information for the locals. If any 770 // of them are unused, then they can be replaced by top(). This 771 // should help register allocation time and cut down on the size 772 // of the deoptimization information. 773 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 774 775 int len = (int)live_locals.size(); 776 assert(len <= jvms()->loc_size(), "too many live locals"); 777 for (int local = 0; local < len; local++) { 778 if (!live_locals.at(local)) { 779 set_local(local, top()); 780 } 781 } 782 } 783 784 #ifdef ASSERT 785 //-------------------------dead_locals_are_killed------------------------------ 786 // Return true if all dead locals are set to top in the map. 787 // Used to assert "clean" debug info at various points. 788 bool GraphKit::dead_locals_are_killed() { 789 if (method() == NULL || method()->code_size() == 0) { 790 // No locals need to be dead, so all is as it should be. 791 return true; 792 } 793 794 // Make sure somebody called kill_dead_locals upstream. 795 ResourceMark rm; 796 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 797 if (jvms->loc_size() == 0) continue; // no locals to consult 798 SafePointNode* map = jvms->map(); 799 ciMethod* method = jvms->method(); 800 int bci = jvms->bci(); 801 if (jvms == this->jvms()) { 802 bci = this->bci(); // it might not yet be synched 803 } 804 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 805 int len = (int)live_locals.size(); 806 if (!live_locals.is_valid() || len == 0) 807 // This method is trivial, or is poisoned by a breakpoint. 808 return true; 809 assert(len == jvms->loc_size(), "live map consistent with locals map"); 810 for (int local = 0; local < len; local++) { 811 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 812 if (PrintMiscellaneous && (Verbose || WizardMode)) { 813 tty->print_cr("Zombie local %d: ", local); 814 jvms->dump(); 815 } 816 return false; 817 } 818 } 819 } 820 return true; 821 } 822 823 #endif //ASSERT 824 825 // Helper function for enforcing certain bytecodes to reexecute if 826 // deoptimization happens 827 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 828 ciMethod* cur_method = jvms->method(); 829 int cur_bci = jvms->bci(); 830 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 831 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 832 return Interpreter::bytecode_should_reexecute(code) || 833 (is_anewarray && code == Bytecodes::_multianewarray); 834 // Reexecute _multianewarray bytecode which was replaced with 835 // sequence of [a]newarray. See Parse::do_multianewarray(). 836 // 837 // Note: interpreter should not have it set since this optimization 838 // is limited by dimensions and guarded by flag so in some cases 839 // multianewarray() runtime calls will be generated and 840 // the bytecode should not be reexecutes (stack will not be reset). 841 } else 842 return false; 843 } 844 845 // Helper function for adding JVMState and debug information to node 846 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 847 // Add the safepoint edges to the call (or other safepoint). 848 849 // Make sure dead locals are set to top. This 850 // should help register allocation time and cut down on the size 851 // of the deoptimization information. 852 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 853 854 // Walk the inline list to fill in the correct set of JVMState's 855 // Also fill in the associated edges for each JVMState. 856 857 // If the bytecode needs to be reexecuted we need to put 858 // the arguments back on the stack. 859 const bool should_reexecute = jvms()->should_reexecute(); 860 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 861 862 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 863 // undefined if the bci is different. This is normal for Parse but it 864 // should not happen for LibraryCallKit because only one bci is processed. 865 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 866 "in LibraryCallKit the reexecute bit should not change"); 867 868 // If we are guaranteed to throw, we can prune everything but the 869 // input to the current bytecode. 870 bool can_prune_locals = false; 871 uint stack_slots_not_pruned = 0; 872 int inputs = 0, depth = 0; 873 if (must_throw) { 874 assert(method() == youngest_jvms->method(), "sanity"); 875 if (compute_stack_effects(inputs, depth)) { 876 can_prune_locals = true; 877 stack_slots_not_pruned = inputs; 878 } 879 } 880 881 if (env()->should_retain_local_variables()) { 882 // At any safepoint, this method can get breakpointed, which would 883 // then require an immediate deoptimization. 884 can_prune_locals = false; // do not prune locals 885 stack_slots_not_pruned = 0; 886 } 887 888 // do not scribble on the input jvms 889 JVMState* out_jvms = youngest_jvms->clone_deep(C); 890 call->set_jvms(out_jvms); // Start jvms list for call node 891 892 // For a known set of bytecodes, the interpreter should reexecute them if 893 // deoptimization happens. We set the reexecute state for them here 894 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 895 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 896 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 897 } 898 899 // Presize the call: 900 DEBUG_ONLY(uint non_debug_edges = call->req()); 901 call->add_req_batch(top(), youngest_jvms->debug_depth()); 902 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 903 904 // Set up edges so that the call looks like this: 905 // Call [state:] ctl io mem fptr retadr 906 // [parms:] parm0 ... parmN 907 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 908 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 909 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 910 // Note that caller debug info precedes callee debug info. 911 912 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 913 uint debug_ptr = call->req(); 914 915 // Loop over the map input edges associated with jvms, add them 916 // to the call node, & reset all offsets to match call node array. 917 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 918 uint debug_end = debug_ptr; 919 uint debug_start = debug_ptr - in_jvms->debug_size(); 920 debug_ptr = debug_start; // back up the ptr 921 922 uint p = debug_start; // walks forward in [debug_start, debug_end) 923 uint j, k, l; 924 SafePointNode* in_map = in_jvms->map(); 925 out_jvms->set_map(call); 926 927 if (can_prune_locals) { 928 assert(in_jvms->method() == out_jvms->method(), "sanity"); 929 // If the current throw can reach an exception handler in this JVMS, 930 // then we must keep everything live that can reach that handler. 931 // As a quick and dirty approximation, we look for any handlers at all. 932 if (in_jvms->method()->has_exception_handlers()) { 933 can_prune_locals = false; 934 } 935 } 936 937 // Add the Locals 938 k = in_jvms->locoff(); 939 l = in_jvms->loc_size(); 940 out_jvms->set_locoff(p); 941 if (!can_prune_locals) { 942 for (j = 0; j < l; j++) 943 call->set_req(p++, in_map->in(k+j)); 944 } else { 945 p += l; // already set to top above by add_req_batch 946 } 947 948 // Add the Expression Stack 949 k = in_jvms->stkoff(); 950 l = in_jvms->sp(); 951 out_jvms->set_stkoff(p); 952 if (!can_prune_locals) { 953 for (j = 0; j < l; j++) 954 call->set_req(p++, in_map->in(k+j)); 955 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 956 // Divide stack into {S0,...,S1}, where S0 is set to top. 957 uint s1 = stack_slots_not_pruned; 958 stack_slots_not_pruned = 0; // for next iteration 959 if (s1 > l) s1 = l; 960 uint s0 = l - s1; 961 p += s0; // skip the tops preinstalled by add_req_batch 962 for (j = s0; j < l; j++) 963 call->set_req(p++, in_map->in(k+j)); 964 } else { 965 p += l; // already set to top above by add_req_batch 966 } 967 968 // Add the Monitors 969 k = in_jvms->monoff(); 970 l = in_jvms->mon_size(); 971 out_jvms->set_monoff(p); 972 for (j = 0; j < l; j++) 973 call->set_req(p++, in_map->in(k+j)); 974 975 // Copy any scalar object fields. 976 k = in_jvms->scloff(); 977 l = in_jvms->scl_size(); 978 out_jvms->set_scloff(p); 979 for (j = 0; j < l; j++) 980 call->set_req(p++, in_map->in(k+j)); 981 982 // Finish the new jvms. 983 out_jvms->set_endoff(p); 984 985 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 986 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 987 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 988 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 989 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 990 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 991 992 // Update the two tail pointers in parallel. 993 out_jvms = out_jvms->caller(); 994 in_jvms = in_jvms->caller(); 995 } 996 997 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 998 999 // Test the correctness of JVMState::debug_xxx accessors: 1000 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1001 assert(call->jvms()->debug_end() == call->req(), ""); 1002 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1003 } 1004 1005 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1006 Bytecodes::Code code = java_bc(); 1007 if (code == Bytecodes::_wide) { 1008 code = method()->java_code_at_bci(bci() + 1); 1009 } 1010 1011 BasicType rtype = T_ILLEGAL; 1012 int rsize = 0; 1013 1014 if (code != Bytecodes::_illegal) { 1015 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1016 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1017 if (rtype < T_CONFLICT) 1018 rsize = type2size[rtype]; 1019 } 1020 1021 switch (code) { 1022 case Bytecodes::_illegal: 1023 return false; 1024 1025 case Bytecodes::_ldc: 1026 case Bytecodes::_ldc_w: 1027 case Bytecodes::_ldc2_w: 1028 inputs = 0; 1029 break; 1030 1031 case Bytecodes::_dup: inputs = 1; break; 1032 case Bytecodes::_dup_x1: inputs = 2; break; 1033 case Bytecodes::_dup_x2: inputs = 3; break; 1034 case Bytecodes::_dup2: inputs = 2; break; 1035 case Bytecodes::_dup2_x1: inputs = 3; break; 1036 case Bytecodes::_dup2_x2: inputs = 4; break; 1037 case Bytecodes::_swap: inputs = 2; break; 1038 case Bytecodes::_arraylength: inputs = 1; break; 1039 1040 case Bytecodes::_getstatic: 1041 case Bytecodes::_putstatic: 1042 case Bytecodes::_getfield: 1043 case Bytecodes::_putfield: 1044 { 1045 bool ignored_will_link; 1046 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1047 int size = field->type()->size(); 1048 bool is_get = (depth >= 0), is_static = (depth & 1); 1049 inputs = (is_static ? 0 : 1); 1050 if (is_get) { 1051 depth = size - inputs; 1052 } else { 1053 inputs += size; // putxxx pops the value from the stack 1054 depth = - inputs; 1055 } 1056 } 1057 break; 1058 1059 case Bytecodes::_invokevirtual: 1060 case Bytecodes::_invokespecial: 1061 case Bytecodes::_invokestatic: 1062 case Bytecodes::_invokedynamic: 1063 case Bytecodes::_invokeinterface: 1064 { 1065 bool ignored_will_link; 1066 ciSignature* declared_signature = NULL; 1067 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1068 assert(declared_signature != NULL, "cannot be null"); 1069 inputs = declared_signature->arg_size_for_bc(code); 1070 int size = declared_signature->return_type()->size(); 1071 depth = size - inputs; 1072 } 1073 break; 1074 1075 case Bytecodes::_multianewarray: 1076 { 1077 ciBytecodeStream iter(method()); 1078 iter.reset_to_bci(bci()); 1079 iter.next(); 1080 inputs = iter.get_dimensions(); 1081 assert(rsize == 1, ""); 1082 depth = rsize - inputs; 1083 } 1084 break; 1085 1086 case Bytecodes::_ireturn: 1087 case Bytecodes::_lreturn: 1088 case Bytecodes::_freturn: 1089 case Bytecodes::_dreturn: 1090 case Bytecodes::_areturn: 1091 assert(rsize == -depth, ""); 1092 inputs = rsize; 1093 break; 1094 1095 case Bytecodes::_jsr: 1096 case Bytecodes::_jsr_w: 1097 inputs = 0; 1098 depth = 1; // S.B. depth=1, not zero 1099 break; 1100 1101 default: 1102 // bytecode produces a typed result 1103 inputs = rsize - depth; 1104 assert(inputs >= 0, ""); 1105 break; 1106 } 1107 1108 #ifdef ASSERT 1109 // spot check 1110 int outputs = depth + inputs; 1111 assert(outputs >= 0, "sanity"); 1112 switch (code) { 1113 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1114 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1115 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1116 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1117 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1118 default: break; 1119 } 1120 #endif //ASSERT 1121 1122 return true; 1123 } 1124 1125 1126 1127 //------------------------------basic_plus_adr--------------------------------- 1128 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1129 // short-circuit a common case 1130 if (offset == intcon(0)) return ptr; 1131 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1132 } 1133 1134 Node* GraphKit::ConvI2L(Node* offset) { 1135 // short-circuit a common case 1136 jint offset_con = find_int_con(offset, Type::OffsetBot); 1137 if (offset_con != Type::OffsetBot) { 1138 return longcon((jlong) offset_con); 1139 } 1140 return _gvn.transform( new ConvI2LNode(offset)); 1141 } 1142 1143 Node* GraphKit::ConvI2UL(Node* offset) { 1144 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1145 if (offset_con != (juint) Type::OffsetBot) { 1146 return longcon((julong) offset_con); 1147 } 1148 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1149 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1150 return _gvn.transform( new AndLNode(conv, mask) ); 1151 } 1152 1153 Node* GraphKit::ConvL2I(Node* offset) { 1154 // short-circuit a common case 1155 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1156 if (offset_con != (jlong)Type::OffsetBot) { 1157 return intcon((int) offset_con); 1158 } 1159 return _gvn.transform( new ConvL2INode(offset)); 1160 } 1161 1162 //-------------------------load_object_klass----------------------------------- 1163 Node* GraphKit::load_object_klass(Node* obj) { 1164 // Special-case a fresh allocation to avoid building nodes: 1165 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1166 if (akls != NULL) return akls; 1167 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1168 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1169 } 1170 1171 //-------------------------load_array_length----------------------------------- 1172 Node* GraphKit::load_array_length(Node* array) { 1173 // Special-case a fresh allocation to avoid building nodes: 1174 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1175 Node *alen; 1176 if (alloc == NULL) { 1177 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1178 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1179 } else { 1180 alen = alloc->Ideal_length(); 1181 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1182 if (ccast != alen) { 1183 alen = _gvn.transform(ccast); 1184 } 1185 } 1186 return alen; 1187 } 1188 1189 //------------------------------do_null_check---------------------------------- 1190 // Helper function to do a NULL pointer check. Returned value is 1191 // the incoming address with NULL casted away. You are allowed to use the 1192 // not-null value only if you are control dependent on the test. 1193 #ifndef PRODUCT 1194 extern int explicit_null_checks_inserted, 1195 explicit_null_checks_elided; 1196 #endif 1197 Node* GraphKit::null_check_common(Node* value, BasicType type, 1198 // optional arguments for variations: 1199 bool assert_null, 1200 Node* *null_control, 1201 bool speculative) { 1202 assert(!assert_null || null_control == NULL, "not both at once"); 1203 if (stopped()) return top(); 1204 NOT_PRODUCT(explicit_null_checks_inserted++); 1205 1206 // Construct NULL check 1207 Node *chk = NULL; 1208 switch(type) { 1209 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1210 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1211 case T_ARRAY : // fall through 1212 type = T_OBJECT; // simplify further tests 1213 case T_OBJECT : { 1214 const Type *t = _gvn.type( value ); 1215 1216 const TypeOopPtr* tp = t->isa_oopptr(); 1217 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1218 // Only for do_null_check, not any of its siblings: 1219 && !assert_null && null_control == NULL) { 1220 // Usually, any field access or invocation on an unloaded oop type 1221 // will simply fail to link, since the statically linked class is 1222 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1223 // the static class is loaded but the sharper oop type is not. 1224 // Rather than checking for this obscure case in lots of places, 1225 // we simply observe that a null check on an unloaded class 1226 // will always be followed by a nonsense operation, so we 1227 // can just issue the uncommon trap here. 1228 // Our access to the unloaded class will only be correct 1229 // after it has been loaded and initialized, which requires 1230 // a trip through the interpreter. 1231 #ifndef PRODUCT 1232 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1233 #endif 1234 uncommon_trap(Deoptimization::Reason_unloaded, 1235 Deoptimization::Action_reinterpret, 1236 tp->klass(), "!loaded"); 1237 return top(); 1238 } 1239 1240 if (assert_null) { 1241 // See if the type is contained in NULL_PTR. 1242 // If so, then the value is already null. 1243 if (t->higher_equal(TypePtr::NULL_PTR)) { 1244 NOT_PRODUCT(explicit_null_checks_elided++); 1245 return value; // Elided null assert quickly! 1246 } 1247 } else { 1248 // See if mixing in the NULL pointer changes type. 1249 // If so, then the NULL pointer was not allowed in the original 1250 // type. In other words, "value" was not-null. 1251 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1252 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1253 NOT_PRODUCT(explicit_null_checks_elided++); 1254 return value; // Elided null check quickly! 1255 } 1256 } 1257 chk = new CmpPNode( value, null() ); 1258 break; 1259 } 1260 1261 default: 1262 fatal("unexpected type: %s", type2name(type)); 1263 } 1264 assert(chk != NULL, "sanity check"); 1265 chk = _gvn.transform(chk); 1266 1267 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1268 BoolNode *btst = new BoolNode( chk, btest); 1269 Node *tst = _gvn.transform( btst ); 1270 1271 //----------- 1272 // if peephole optimizations occurred, a prior test existed. 1273 // If a prior test existed, maybe it dominates as we can avoid this test. 1274 if (tst != btst && type == T_OBJECT) { 1275 // At this point we want to scan up the CFG to see if we can 1276 // find an identical test (and so avoid this test altogether). 1277 Node *cfg = control(); 1278 int depth = 0; 1279 while( depth < 16 ) { // Limit search depth for speed 1280 if( cfg->Opcode() == Op_IfTrue && 1281 cfg->in(0)->in(1) == tst ) { 1282 // Found prior test. Use "cast_not_null" to construct an identical 1283 // CastPP (and hence hash to) as already exists for the prior test. 1284 // Return that casted value. 1285 if (assert_null) { 1286 replace_in_map(value, null()); 1287 return null(); // do not issue the redundant test 1288 } 1289 Node *oldcontrol = control(); 1290 set_control(cfg); 1291 Node *res = cast_not_null(value); 1292 set_control(oldcontrol); 1293 NOT_PRODUCT(explicit_null_checks_elided++); 1294 return res; 1295 } 1296 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1297 if (cfg == NULL) break; // Quit at region nodes 1298 depth++; 1299 } 1300 } 1301 1302 //----------- 1303 // Branch to failure if null 1304 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1305 Deoptimization::DeoptReason reason; 1306 if (assert_null) { 1307 reason = Deoptimization::reason_null_assert(speculative); 1308 } else if (type == T_OBJECT) { 1309 reason = Deoptimization::reason_null_check(speculative); 1310 } else { 1311 reason = Deoptimization::Reason_div0_check; 1312 } 1313 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1314 // ciMethodData::has_trap_at will return a conservative -1 if any 1315 // must-be-null assertion has failed. This could cause performance 1316 // problems for a method after its first do_null_assert failure. 1317 // Consider using 'Reason_class_check' instead? 1318 1319 // To cause an implicit null check, we set the not-null probability 1320 // to the maximum (PROB_MAX). For an explicit check the probability 1321 // is set to a smaller value. 1322 if (null_control != NULL || too_many_traps(reason)) { 1323 // probability is less likely 1324 ok_prob = PROB_LIKELY_MAG(3); 1325 } else if (!assert_null && 1326 (ImplicitNullCheckThreshold > 0) && 1327 method() != NULL && 1328 (method()->method_data()->trap_count(reason) 1329 >= (uint)ImplicitNullCheckThreshold)) { 1330 ok_prob = PROB_LIKELY_MAG(3); 1331 } 1332 1333 if (null_control != NULL) { 1334 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1335 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1336 set_control( _gvn.transform( new IfTrueNode(iff))); 1337 #ifndef PRODUCT 1338 if (null_true == top()) { 1339 explicit_null_checks_elided++; 1340 } 1341 #endif 1342 (*null_control) = null_true; 1343 } else { 1344 BuildCutout unless(this, tst, ok_prob); 1345 // Check for optimizer eliding test at parse time 1346 if (stopped()) { 1347 // Failure not possible; do not bother making uncommon trap. 1348 NOT_PRODUCT(explicit_null_checks_elided++); 1349 } else if (assert_null) { 1350 uncommon_trap(reason, 1351 Deoptimization::Action_make_not_entrant, 1352 NULL, "assert_null"); 1353 } else { 1354 replace_in_map(value, zerocon(type)); 1355 builtin_throw(reason); 1356 } 1357 } 1358 1359 // Must throw exception, fall-thru not possible? 1360 if (stopped()) { 1361 return top(); // No result 1362 } 1363 1364 if (assert_null) { 1365 // Cast obj to null on this path. 1366 replace_in_map(value, zerocon(type)); 1367 return zerocon(type); 1368 } 1369 1370 // Cast obj to not-null on this path, if there is no null_control. 1371 // (If there is a null_control, a non-null value may come back to haunt us.) 1372 if (type == T_OBJECT) { 1373 Node* cast = cast_not_null(value, false); 1374 if (null_control == NULL || (*null_control) == top()) 1375 replace_in_map(value, cast); 1376 value = cast; 1377 } 1378 1379 return value; 1380 } 1381 1382 1383 //------------------------------cast_not_null---------------------------------- 1384 // Cast obj to not-null on this path 1385 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1386 const Type *t = _gvn.type(obj); 1387 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1388 // Object is already not-null? 1389 if( t == t_not_null ) return obj; 1390 1391 Node *cast = new CastPPNode(obj,t_not_null); 1392 cast->init_req(0, control()); 1393 cast = _gvn.transform( cast ); 1394 1395 // Scan for instances of 'obj' in the current JVM mapping. 1396 // These instances are known to be not-null after the test. 1397 if (do_replace_in_map) 1398 replace_in_map(obj, cast); 1399 1400 return cast; // Return casted value 1401 } 1402 1403 // Sometimes in intrinsics, we implicitly know an object is not null 1404 // (there's no actual null check) so we can cast it to not null. In 1405 // the course of optimizations, the input to the cast can become null. 1406 // In that case that data path will die and we need the control path 1407 // to become dead as well to keep the graph consistent. So we have to 1408 // add a check for null for which one branch can't be taken. It uses 1409 // an Opaque4 node that will cause the check to be removed after loop 1410 // opts so the test goes away and the compiled code doesn't execute a 1411 // useless check. 1412 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1413 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1414 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1415 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1416 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1417 _gvn.set_type(iff, iff->Value(&_gvn)); 1418 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1419 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1420 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1421 C->root()->add_req(halt); 1422 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1423 set_control(if_t); 1424 return cast_not_null(value, do_replace_in_map); 1425 } 1426 1427 1428 //--------------------------replace_in_map------------------------------------- 1429 void GraphKit::replace_in_map(Node* old, Node* neww) { 1430 if (old == neww) { 1431 return; 1432 } 1433 1434 map()->replace_edge(old, neww); 1435 1436 // Note: This operation potentially replaces any edge 1437 // on the map. This includes locals, stack, and monitors 1438 // of the current (innermost) JVM state. 1439 1440 // don't let inconsistent types from profiling escape this 1441 // method 1442 1443 const Type* told = _gvn.type(old); 1444 const Type* tnew = _gvn.type(neww); 1445 1446 if (!tnew->higher_equal(told)) { 1447 return; 1448 } 1449 1450 map()->record_replaced_node(old, neww); 1451 } 1452 1453 1454 //============================================================================= 1455 //--------------------------------memory--------------------------------------- 1456 Node* GraphKit::memory(uint alias_idx) { 1457 MergeMemNode* mem = merged_memory(); 1458 Node* p = mem->memory_at(alias_idx); 1459 _gvn.set_type(p, Type::MEMORY); // must be mapped 1460 return p; 1461 } 1462 1463 //-----------------------------reset_memory------------------------------------ 1464 Node* GraphKit::reset_memory() { 1465 Node* mem = map()->memory(); 1466 // do not use this node for any more parsing! 1467 debug_only( map()->set_memory((Node*)NULL) ); 1468 return _gvn.transform( mem ); 1469 } 1470 1471 //------------------------------set_all_memory--------------------------------- 1472 void GraphKit::set_all_memory(Node* newmem) { 1473 Node* mergemem = MergeMemNode::make(newmem); 1474 gvn().set_type_bottom(mergemem); 1475 map()->set_memory(mergemem); 1476 } 1477 1478 //------------------------------set_all_memory_call---------------------------- 1479 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1480 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1481 set_all_memory(newmem); 1482 } 1483 1484 //============================================================================= 1485 // 1486 // parser factory methods for MemNodes 1487 // 1488 // These are layered on top of the factory methods in LoadNode and StoreNode, 1489 // and integrate with the parser's memory state and _gvn engine. 1490 // 1491 1492 // factory methods in "int adr_idx" 1493 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1494 int adr_idx, 1495 MemNode::MemOrd mo, 1496 LoadNode::ControlDependency control_dependency, 1497 bool require_atomic_access, 1498 bool unaligned, 1499 bool mismatched) { 1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1501 const TypePtr* adr_type = NULL; // debug-mode-only argument 1502 debug_only(adr_type = C->get_adr_type(adr_idx)); 1503 Node* mem = memory(adr_idx); 1504 Node* ld; 1505 if (require_atomic_access && bt == T_LONG) { 1506 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1507 } else if (require_atomic_access && bt == T_DOUBLE) { 1508 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1509 } else { 1510 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1511 } 1512 ld = _gvn.transform(ld); 1513 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1514 // Improve graph before escape analysis and boxing elimination. 1515 record_for_igvn(ld); 1516 } 1517 return ld; 1518 } 1519 1520 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1521 int adr_idx, 1522 MemNode::MemOrd mo, 1523 bool require_atomic_access, 1524 bool unaligned, 1525 bool mismatched) { 1526 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1527 const TypePtr* adr_type = NULL; 1528 debug_only(adr_type = C->get_adr_type(adr_idx)); 1529 Node *mem = memory(adr_idx); 1530 Node* st; 1531 if (require_atomic_access && bt == T_LONG) { 1532 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1533 } else if (require_atomic_access && bt == T_DOUBLE) { 1534 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1535 } else { 1536 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1537 } 1538 if (unaligned) { 1539 st->as_Store()->set_unaligned_access(); 1540 } 1541 if (mismatched) { 1542 st->as_Store()->set_mismatched_access(); 1543 } 1544 st = _gvn.transform(st); 1545 set_memory(st, adr_idx); 1546 // Back-to-back stores can only remove intermediate store with DU info 1547 // so push on worklist for optimizer. 1548 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1549 record_for_igvn(st); 1550 1551 return st; 1552 } 1553 1554 1555 Node* GraphKit::pre_barrier(bool do_load, 1556 Node* ctl, 1557 Node* obj, 1558 Node* adr, 1559 uint adr_idx, 1560 Node* val, 1561 const TypeOopPtr* val_type, 1562 Node* pre_val, 1563 BasicType bt) { 1564 1565 BarrierSet* bs = Universe::heap()->barrier_set(); 1566 set_control(ctl); 1567 switch (bs->kind()) { 1568 case BarrierSet::G1SATBCTLogging: 1569 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1570 return val; 1571 case BarrierSet::ShenandoahBarrierSet: 1572 return shenandoah_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1573 1574 case BarrierSet::CardTableForRS: 1575 case BarrierSet::CardTableExtension: 1576 case BarrierSet::ModRef: 1577 break; 1578 1579 default : 1580 ShouldNotReachHere(); 1581 1582 } 1583 return val; 1584 } 1585 1586 bool GraphKit::can_move_pre_barrier() const { 1587 BarrierSet* bs = Universe::heap()->barrier_set(); 1588 switch (bs->kind()) { 1589 case BarrierSet::G1SATBCTLogging: 1590 case BarrierSet::ShenandoahBarrierSet: 1591 return true; // Can move it if no safepoint 1592 1593 case BarrierSet::CardTableForRS: 1594 case BarrierSet::CardTableExtension: 1595 case BarrierSet::ModRef: 1596 return true; // There is no pre-barrier 1597 1598 default : 1599 ShouldNotReachHere(); 1600 } 1601 return false; 1602 } 1603 1604 void GraphKit::post_barrier(Node* ctl, 1605 Node* store, 1606 Node* obj, 1607 Node* adr, 1608 uint adr_idx, 1609 Node* val, 1610 BasicType bt, 1611 bool use_precise) { 1612 BarrierSet* bs = Universe::heap()->barrier_set(); 1613 set_control(ctl); 1614 switch (bs->kind()) { 1615 case BarrierSet::G1SATBCTLogging: 1616 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1617 break; 1618 1619 case BarrierSet::CardTableForRS: 1620 case BarrierSet::CardTableExtension: 1621 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1622 break; 1623 1624 case BarrierSet::ModRef: 1625 case BarrierSet::ShenandoahBarrierSet: 1626 break; 1627 1628 default : 1629 ShouldNotReachHere(); 1630 1631 } 1632 } 1633 1634 Node* GraphKit::store_oop(Node* ctl, 1635 Node* obj, 1636 Node* adr, 1637 const TypePtr* adr_type, 1638 Node* val, 1639 const TypeOopPtr* val_type, 1640 BasicType bt, 1641 bool use_precise, 1642 MemNode::MemOrd mo, 1643 bool mismatched) { 1644 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1645 // could be delayed during Parse (for example, in adjust_map_after_if()). 1646 // Execute transformation here to avoid barrier generation in such case. 1647 if (_gvn.type(val) == TypePtr::NULL_PTR) 1648 val = _gvn.makecon(TypePtr::NULL_PTR); 1649 1650 set_control(ctl); 1651 if (stopped()) return top(); // Dead path ? 1652 1653 assert(bt == T_OBJECT, "sanity"); 1654 assert(val != NULL, "not dead path"); 1655 uint adr_idx = C->get_alias_index(adr_type); 1656 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1657 1658 val = pre_barrier(true /* do_load */, 1659 control(), obj, adr, adr_idx, val, val_type, 1660 NULL /* pre_val */, 1661 bt); 1662 1663 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1664 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1665 return store; 1666 } 1667 1668 // Could be an array or object we don't know at compile time (unsafe ref.) 1669 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1670 Node* obj, // containing obj 1671 Node* adr, // actual adress to store val at 1672 const TypePtr* adr_type, 1673 Node* val, 1674 BasicType bt, 1675 MemNode::MemOrd mo, 1676 bool mismatched) { 1677 Compile::AliasType* at = C->alias_type(adr_type); 1678 const TypeOopPtr* val_type = NULL; 1679 if (adr_type->isa_instptr()) { 1680 if (at->field() != NULL) { 1681 // known field. This code is a copy of the do_put_xxx logic. 1682 ciField* field = at->field(); 1683 if (!field->type()->is_loaded()) { 1684 val_type = TypeInstPtr::BOTTOM; 1685 } else { 1686 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1687 } 1688 } 1689 } else if (adr_type->isa_aryptr()) { 1690 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1691 } 1692 if (val_type == NULL) { 1693 val_type = TypeInstPtr::BOTTOM; 1694 } 1695 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1696 } 1697 1698 1699 //-------------------------array_element_address------------------------- 1700 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1701 const TypeInt* sizetype, Node* ctrl) { 1702 uint shift = exact_log2(type2aelembytes(elembt)); 1703 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1704 1705 // short-circuit a common case (saves lots of confusing waste motion) 1706 jint idx_con = find_int_con(idx, -1); 1707 if (idx_con >= 0) { 1708 intptr_t offset = header + ((intptr_t)idx_con << shift); 1709 return basic_plus_adr(ary, offset); 1710 } 1711 1712 // must be correct type for alignment purposes 1713 Node* base = basic_plus_adr(ary, header); 1714 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1715 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1716 return basic_plus_adr(ary, base, scale); 1717 } 1718 1719 //-------------------------load_array_element------------------------- 1720 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1721 const Type* elemtype = arytype->elem(); 1722 BasicType elembt = elemtype->array_element_basic_type(); 1723 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1724 if (elembt == T_NARROWOOP) { 1725 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1726 } 1727 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1728 return ld; 1729 } 1730 1731 //-------------------------set_arguments_for_java_call------------------------- 1732 // Arguments (pre-popped from the stack) are taken from the JVMS. 1733 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1734 // Add the call arguments: 1735 uint nargs = call->method()->arg_size(); 1736 for (uint i = 0; i < nargs; i++) { 1737 Node* arg = argument(i); 1738 call->init_req(i + TypeFunc::Parms, arg); 1739 } 1740 } 1741 1742 //---------------------------set_edges_for_java_call--------------------------- 1743 // Connect a newly created call into the current JVMS. 1744 // A return value node (if any) is returned from set_edges_for_java_call. 1745 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1746 1747 // Add the predefined inputs: 1748 call->init_req( TypeFunc::Control, control() ); 1749 call->init_req( TypeFunc::I_O , i_o() ); 1750 call->init_req( TypeFunc::Memory , reset_memory() ); 1751 call->init_req( TypeFunc::FramePtr, frameptr() ); 1752 call->init_req( TypeFunc::ReturnAdr, top() ); 1753 1754 add_safepoint_edges(call, must_throw); 1755 1756 Node* xcall = _gvn.transform(call); 1757 1758 if (xcall == top()) { 1759 set_control(top()); 1760 return; 1761 } 1762 assert(xcall == call, "call identity is stable"); 1763 1764 // Re-use the current map to produce the result. 1765 1766 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1767 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1768 set_all_memory_call(xcall, separate_io_proj); 1769 1770 //return xcall; // no need, caller already has it 1771 } 1772 1773 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1774 if (stopped()) return top(); // maybe the call folded up? 1775 1776 // Capture the return value, if any. 1777 Node* ret; 1778 if (call->method() == NULL || 1779 call->method()->return_type()->basic_type() == T_VOID) 1780 ret = top(); 1781 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1782 1783 // Note: Since any out-of-line call can produce an exception, 1784 // we always insert an I_O projection from the call into the result. 1785 1786 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1787 1788 if (separate_io_proj) { 1789 // The caller requested separate projections be used by the fall 1790 // through and exceptional paths, so replace the projections for 1791 // the fall through path. 1792 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1793 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1794 } 1795 return ret; 1796 } 1797 1798 //--------------------set_predefined_input_for_runtime_call-------------------- 1799 // Reading and setting the memory state is way conservative here. 1800 // The real problem is that I am not doing real Type analysis on memory, 1801 // so I cannot distinguish card mark stores from other stores. Across a GC 1802 // point the Store Barrier and the card mark memory has to agree. I cannot 1803 // have a card mark store and its barrier split across the GC point from 1804 // either above or below. Here I get that to happen by reading ALL of memory. 1805 // A better answer would be to separate out card marks from other memory. 1806 // For now, return the input memory state, so that it can be reused 1807 // after the call, if this call has restricted memory effects. 1808 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1809 // Set fixed predefined input arguments 1810 Node* memory = reset_memory(); 1811 call->init_req( TypeFunc::Control, control() ); 1812 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1813 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1814 call->init_req( TypeFunc::FramePtr, frameptr() ); 1815 call->init_req( TypeFunc::ReturnAdr, top() ); 1816 return memory; 1817 } 1818 1819 //-------------------set_predefined_output_for_runtime_call-------------------- 1820 // Set control and memory (not i_o) from the call. 1821 // If keep_mem is not NULL, use it for the output state, 1822 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1823 // If hook_mem is NULL, this call produces no memory effects at all. 1824 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1825 // then only that memory slice is taken from the call. 1826 // In the last case, we must put an appropriate memory barrier before 1827 // the call, so as to create the correct anti-dependencies on loads 1828 // preceding the call. 1829 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1830 Node* keep_mem, 1831 const TypePtr* hook_mem) { 1832 // no i/o 1833 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1834 if (keep_mem) { 1835 // First clone the existing memory state 1836 set_all_memory(keep_mem); 1837 if (hook_mem != NULL) { 1838 // Make memory for the call 1839 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1840 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1841 // We also use hook_mem to extract specific effects from arraycopy stubs. 1842 set_memory(mem, hook_mem); 1843 } 1844 // ...else the call has NO memory effects. 1845 1846 // Make sure the call advertises its memory effects precisely. 1847 // This lets us build accurate anti-dependences in gcm.cpp. 1848 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1849 "call node must be constructed correctly"); 1850 } else { 1851 assert(hook_mem == NULL, ""); 1852 // This is not a "slow path" call; all memory comes from the call. 1853 set_all_memory_call(call); 1854 } 1855 } 1856 1857 1858 // Replace the call with the current state of the kit. 1859 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1860 JVMState* ejvms = NULL; 1861 if (has_exceptions()) { 1862 ejvms = transfer_exceptions_into_jvms(); 1863 } 1864 1865 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1866 ReplacedNodes replaced_nodes_exception; 1867 Node* ex_ctl = top(); 1868 1869 SafePointNode* final_state = stop(); 1870 1871 // Find all the needed outputs of this call 1872 CallProjections callprojs; 1873 call->extract_projections(&callprojs, true); 1874 1875 Node* init_mem = call->in(TypeFunc::Memory); 1876 Node* final_mem = final_state->in(TypeFunc::Memory); 1877 Node* final_ctl = final_state->in(TypeFunc::Control); 1878 Node* final_io = final_state->in(TypeFunc::I_O); 1879 1880 // Replace all the old call edges with the edges from the inlining result 1881 if (callprojs.fallthrough_catchproj != NULL) { 1882 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1883 } 1884 if (callprojs.fallthrough_memproj != NULL) { 1885 if (final_mem->is_MergeMem()) { 1886 // Parser's exits MergeMem was not transformed but may be optimized 1887 final_mem = _gvn.transform(final_mem); 1888 } 1889 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1890 } 1891 if (callprojs.fallthrough_ioproj != NULL) { 1892 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1893 } 1894 1895 // Replace the result with the new result if it exists and is used 1896 if (callprojs.resproj != NULL && result != NULL) { 1897 C->gvn_replace_by(callprojs.resproj, result); 1898 } 1899 1900 if (ejvms == NULL) { 1901 // No exception edges to simply kill off those paths 1902 if (callprojs.catchall_catchproj != NULL) { 1903 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1904 } 1905 if (callprojs.catchall_memproj != NULL) { 1906 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1907 } 1908 if (callprojs.catchall_ioproj != NULL) { 1909 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1910 } 1911 // Replace the old exception object with top 1912 if (callprojs.exobj != NULL) { 1913 C->gvn_replace_by(callprojs.exobj, C->top()); 1914 } 1915 } else { 1916 GraphKit ekit(ejvms); 1917 1918 // Load my combined exception state into the kit, with all phis transformed: 1919 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1920 replaced_nodes_exception = ex_map->replaced_nodes(); 1921 1922 Node* ex_oop = ekit.use_exception_state(ex_map); 1923 1924 if (callprojs.catchall_catchproj != NULL) { 1925 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1926 ex_ctl = ekit.control(); 1927 } 1928 if (callprojs.catchall_memproj != NULL) { 1929 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1930 } 1931 if (callprojs.catchall_ioproj != NULL) { 1932 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1933 } 1934 1935 // Replace the old exception object with the newly created one 1936 if (callprojs.exobj != NULL) { 1937 C->gvn_replace_by(callprojs.exobj, ex_oop); 1938 } 1939 } 1940 1941 // Disconnect the call from the graph 1942 call->disconnect_inputs(NULL, C); 1943 C->gvn_replace_by(call, C->top()); 1944 1945 // Clean up any MergeMems that feed other MergeMems since the 1946 // optimizer doesn't like that. 1947 if (final_mem->is_MergeMem()) { 1948 Node_List wl; 1949 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1950 Node* m = i.get(); 1951 if (m->is_MergeMem() && !wl.contains(m)) { 1952 wl.push(m); 1953 } 1954 } 1955 while (wl.size() > 0) { 1956 _gvn.transform(wl.pop()); 1957 } 1958 } 1959 1960 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1961 replaced_nodes.apply(C, final_ctl); 1962 } 1963 if (!ex_ctl->is_top() && do_replaced_nodes) { 1964 replaced_nodes_exception.apply(C, ex_ctl); 1965 } 1966 } 1967 1968 1969 //------------------------------increment_counter------------------------------ 1970 // for statistics: increment a VM counter by 1 1971 1972 void GraphKit::increment_counter(address counter_addr) { 1973 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1974 increment_counter(adr1); 1975 } 1976 1977 void GraphKit::increment_counter(Node* counter_addr) { 1978 int adr_type = Compile::AliasIdxRaw; 1979 Node* ctrl = control(); 1980 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1981 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1982 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1983 } 1984 1985 1986 //------------------------------uncommon_trap---------------------------------- 1987 // Bail out to the interpreter in mid-method. Implemented by calling the 1988 // uncommon_trap blob. This helper function inserts a runtime call with the 1989 // right debug info. 1990 void GraphKit::uncommon_trap(int trap_request, 1991 ciKlass* klass, const char* comment, 1992 bool must_throw, 1993 bool keep_exact_action) { 1994 if (failing()) stop(); 1995 if (stopped()) return; // trap reachable? 1996 1997 // Note: If ProfileTraps is true, and if a deopt. actually 1998 // occurs here, the runtime will make sure an MDO exists. There is 1999 // no need to call method()->ensure_method_data() at this point. 2000 2001 // Set the stack pointer to the right value for reexecution: 2002 set_sp(reexecute_sp()); 2003 2004 #ifdef ASSERT 2005 if (!must_throw) { 2006 // Make sure the stack has at least enough depth to execute 2007 // the current bytecode. 2008 int inputs, ignored_depth; 2009 if (compute_stack_effects(inputs, ignored_depth)) { 2010 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2011 Bytecodes::name(java_bc()), sp(), inputs); 2012 } 2013 } 2014 #endif 2015 2016 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2017 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2018 2019 switch (action) { 2020 case Deoptimization::Action_maybe_recompile: 2021 case Deoptimization::Action_reinterpret: 2022 // Temporary fix for 6529811 to allow virtual calls to be sure they 2023 // get the chance to go from mono->bi->mega 2024 if (!keep_exact_action && 2025 Deoptimization::trap_request_index(trap_request) < 0 && 2026 too_many_recompiles(reason)) { 2027 // This BCI is causing too many recompilations. 2028 if (C->log() != NULL) { 2029 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2030 Deoptimization::trap_reason_name(reason), 2031 Deoptimization::trap_action_name(action)); 2032 } 2033 action = Deoptimization::Action_none; 2034 trap_request = Deoptimization::make_trap_request(reason, action); 2035 } else { 2036 C->set_trap_can_recompile(true); 2037 } 2038 break; 2039 case Deoptimization::Action_make_not_entrant: 2040 C->set_trap_can_recompile(true); 2041 break; 2042 case Deoptimization::Action_none: 2043 case Deoptimization::Action_make_not_compilable: 2044 break; 2045 default: 2046 #ifdef ASSERT 2047 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2048 #endif 2049 break; 2050 } 2051 2052 if (TraceOptoParse) { 2053 char buf[100]; 2054 tty->print_cr("Uncommon trap %s at bci:%d", 2055 Deoptimization::format_trap_request(buf, sizeof(buf), 2056 trap_request), bci()); 2057 } 2058 2059 CompileLog* log = C->log(); 2060 if (log != NULL) { 2061 int kid = (klass == NULL)? -1: log->identify(klass); 2062 log->begin_elem("uncommon_trap bci='%d'", bci()); 2063 char buf[100]; 2064 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2065 trap_request)); 2066 if (kid >= 0) log->print(" klass='%d'", kid); 2067 if (comment != NULL) log->print(" comment='%s'", comment); 2068 log->end_elem(); 2069 } 2070 2071 // Make sure any guarding test views this path as very unlikely 2072 Node *i0 = control()->in(0); 2073 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2074 IfNode *iff = i0->as_If(); 2075 float f = iff->_prob; // Get prob 2076 if (control()->Opcode() == Op_IfTrue) { 2077 if (f > PROB_UNLIKELY_MAG(4)) 2078 iff->_prob = PROB_MIN; 2079 } else { 2080 if (f < PROB_LIKELY_MAG(4)) 2081 iff->_prob = PROB_MAX; 2082 } 2083 } 2084 2085 // Clear out dead values from the debug info. 2086 kill_dead_locals(); 2087 2088 // Now insert the uncommon trap subroutine call 2089 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2090 const TypePtr* no_memory_effects = NULL; 2091 // Pass the index of the class to be loaded 2092 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2093 (must_throw ? RC_MUST_THROW : 0), 2094 OptoRuntime::uncommon_trap_Type(), 2095 call_addr, "uncommon_trap", no_memory_effects, 2096 intcon(trap_request)); 2097 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2098 "must extract request correctly from the graph"); 2099 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2100 2101 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2102 // The debug info is the only real input to this call. 2103 2104 // Halt-and-catch fire here. The above call should never return! 2105 HaltNode* halt = new HaltNode(control(), frameptr()); 2106 _gvn.set_type_bottom(halt); 2107 root()->add_req(halt); 2108 2109 stop_and_kill_map(); 2110 } 2111 2112 2113 //--------------------------just_allocated_object------------------------------ 2114 // Report the object that was just allocated. 2115 // It must be the case that there are no intervening safepoints. 2116 // We use this to determine if an object is so "fresh" that 2117 // it does not require card marks. 2118 Node* GraphKit::just_allocated_object(Node* current_control) { 2119 if (C->recent_alloc_ctl() == current_control) 2120 return C->recent_alloc_obj(); 2121 return NULL; 2122 } 2123 2124 2125 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2126 // (Note: TypeFunc::make has a cache that makes this fast.) 2127 const TypeFunc* tf = TypeFunc::make(dest_method); 2128 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2129 for (int j = 0; j < nargs; j++) { 2130 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2131 if( targ->basic_type() == T_DOUBLE ) { 2132 // If any parameters are doubles, they must be rounded before 2133 // the call, dstore_rounding does gvn.transform 2134 Node *arg = argument(j); 2135 arg = dstore_rounding(arg); 2136 set_argument(j, arg); 2137 } 2138 } 2139 } 2140 2141 /** 2142 * Record profiling data exact_kls for Node n with the type system so 2143 * that it can propagate it (speculation) 2144 * 2145 * @param n node that the type applies to 2146 * @param exact_kls type from profiling 2147 * @param maybe_null did profiling see null? 2148 * 2149 * @return node with improved type 2150 */ 2151 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2152 const Type* current_type = _gvn.type(n); 2153 assert(UseTypeSpeculation, "type speculation must be on"); 2154 2155 const TypePtr* speculative = current_type->speculative(); 2156 2157 // Should the klass from the profile be recorded in the speculative type? 2158 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2159 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2160 const TypeOopPtr* xtype = tklass->as_instance_type(); 2161 assert(xtype->klass_is_exact(), "Should be exact"); 2162 // Any reason to believe n is not null (from this profiling or a previous one)? 2163 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2164 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2165 // record the new speculative type's depth 2166 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2167 speculative = speculative->with_inline_depth(jvms()->depth()); 2168 } else if (current_type->would_improve_ptr(ptr_kind)) { 2169 // Profiling report that null was never seen so we can change the 2170 // speculative type to non null ptr. 2171 if (ptr_kind == ProfileAlwaysNull) { 2172 speculative = TypePtr::NULL_PTR; 2173 } else { 2174 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2175 const TypePtr* ptr = TypePtr::NOTNULL; 2176 if (speculative != NULL) { 2177 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2178 } else { 2179 speculative = ptr; 2180 } 2181 } 2182 } 2183 2184 if (speculative != current_type->speculative()) { 2185 // Build a type with a speculative type (what we think we know 2186 // about the type but will need a guard when we use it) 2187 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2188 // We're changing the type, we need a new CheckCast node to carry 2189 // the new type. The new type depends on the control: what 2190 // profiling tells us is only valid from here as far as we can 2191 // tell. 2192 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2193 cast = _gvn.transform(cast); 2194 replace_in_map(n, cast); 2195 n = cast; 2196 } 2197 2198 return n; 2199 } 2200 2201 /** 2202 * Record profiling data from receiver profiling at an invoke with the 2203 * type system so that it can propagate it (speculation) 2204 * 2205 * @param n receiver node 2206 * 2207 * @return node with improved type 2208 */ 2209 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2210 if (!UseTypeSpeculation) { 2211 return n; 2212 } 2213 ciKlass* exact_kls = profile_has_unique_klass(); 2214 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2215 if ((java_bc() == Bytecodes::_checkcast || 2216 java_bc() == Bytecodes::_instanceof || 2217 java_bc() == Bytecodes::_aastore) && 2218 method()->method_data()->is_mature()) { 2219 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2220 if (data != NULL) { 2221 if (!data->as_BitData()->null_seen()) { 2222 ptr_kind = ProfileNeverNull; 2223 } else { 2224 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2225 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2226 uint i = 0; 2227 for (; i < call->row_limit(); i++) { 2228 ciKlass* receiver = call->receiver(i); 2229 if (receiver != NULL) { 2230 break; 2231 } 2232 } 2233 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2234 } 2235 } 2236 } 2237 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2238 } 2239 2240 /** 2241 * Record profiling data from argument profiling at an invoke with the 2242 * type system so that it can propagate it (speculation) 2243 * 2244 * @param dest_method target method for the call 2245 * @param bc what invoke bytecode is this? 2246 */ 2247 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2248 if (!UseTypeSpeculation) { 2249 return; 2250 } 2251 const TypeFunc* tf = TypeFunc::make(dest_method); 2252 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2253 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2254 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2255 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2256 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2257 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2258 ciKlass* better_type = NULL; 2259 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2260 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2261 } 2262 i++; 2263 } 2264 } 2265 } 2266 2267 /** 2268 * Record profiling data from parameter profiling at an invoke with 2269 * the type system so that it can propagate it (speculation) 2270 */ 2271 void GraphKit::record_profiled_parameters_for_speculation() { 2272 if (!UseTypeSpeculation) { 2273 return; 2274 } 2275 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2276 if (_gvn.type(local(i))->isa_oopptr()) { 2277 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2278 ciKlass* better_type = NULL; 2279 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2280 record_profile_for_speculation(local(i), better_type, ptr_kind); 2281 } 2282 j++; 2283 } 2284 } 2285 } 2286 2287 /** 2288 * Record profiling data from return value profiling at an invoke with 2289 * the type system so that it can propagate it (speculation) 2290 */ 2291 void GraphKit::record_profiled_return_for_speculation() { 2292 if (!UseTypeSpeculation) { 2293 return; 2294 } 2295 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2296 ciKlass* better_type = NULL; 2297 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2298 // If profiling reports a single type for the return value, 2299 // feed it to the type system so it can propagate it as a 2300 // speculative type 2301 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2302 } 2303 } 2304 2305 void GraphKit::round_double_result(ciMethod* dest_method) { 2306 // A non-strict method may return a double value which has an extended 2307 // exponent, but this must not be visible in a caller which is 'strict' 2308 // If a strict caller invokes a non-strict callee, round a double result 2309 2310 BasicType result_type = dest_method->return_type()->basic_type(); 2311 assert( method() != NULL, "must have caller context"); 2312 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2313 // Destination method's return value is on top of stack 2314 // dstore_rounding() does gvn.transform 2315 Node *result = pop_pair(); 2316 result = dstore_rounding(result); 2317 push_pair(result); 2318 } 2319 } 2320 2321 // rounding for strict float precision conformance 2322 Node* GraphKit::precision_rounding(Node* n) { 2323 return UseStrictFP && _method->flags().is_strict() 2324 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2325 ? _gvn.transform( new RoundFloatNode(0, n) ) 2326 : n; 2327 } 2328 2329 // rounding for strict double precision conformance 2330 Node* GraphKit::dprecision_rounding(Node *n) { 2331 return UseStrictFP && _method->flags().is_strict() 2332 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2333 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2334 : n; 2335 } 2336 2337 // rounding for non-strict double stores 2338 Node* GraphKit::dstore_rounding(Node* n) { 2339 return Matcher::strict_fp_requires_explicit_rounding 2340 && UseSSE <= 1 2341 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2342 : n; 2343 } 2344 2345 //============================================================================= 2346 // Generate a fast path/slow path idiom. Graph looks like: 2347 // [foo] indicates that 'foo' is a parameter 2348 // 2349 // [in] NULL 2350 // \ / 2351 // CmpP 2352 // Bool ne 2353 // If 2354 // / \ 2355 // True False-<2> 2356 // / | 2357 // / cast_not_null 2358 // Load | | ^ 2359 // [fast_test] | | 2360 // gvn to opt_test | | 2361 // / \ | <1> 2362 // True False | 2363 // | \\ | 2364 // [slow_call] \[fast_result] 2365 // Ctl Val \ \ 2366 // | \ \ 2367 // Catch <1> \ \ 2368 // / \ ^ \ \ 2369 // Ex No_Ex | \ \ 2370 // | \ \ | \ <2> \ 2371 // ... \ [slow_res] | | \ [null_result] 2372 // \ \--+--+--- | | 2373 // \ | / \ | / 2374 // --------Region Phi 2375 // 2376 //============================================================================= 2377 // Code is structured as a series of driver functions all called 'do_XXX' that 2378 // call a set of helper functions. Helper functions first, then drivers. 2379 2380 //------------------------------null_check_oop--------------------------------- 2381 // Null check oop. Set null-path control into Region in slot 3. 2382 // Make a cast-not-nullness use the other not-null control. Return cast. 2383 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2384 bool never_see_null, 2385 bool safe_for_replace, 2386 bool speculative) { 2387 // Initial NULL check taken path 2388 (*null_control) = top(); 2389 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2390 2391 // Generate uncommon_trap: 2392 if (never_see_null && (*null_control) != top()) { 2393 // If we see an unexpected null at a check-cast we record it and force a 2394 // recompile; the offending check-cast will be compiled to handle NULLs. 2395 // If we see more than one offending BCI, then all checkcasts in the 2396 // method will be compiled to handle NULLs. 2397 PreserveJVMState pjvms(this); 2398 set_control(*null_control); 2399 replace_in_map(value, null()); 2400 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2401 uncommon_trap(reason, 2402 Deoptimization::Action_make_not_entrant); 2403 (*null_control) = top(); // NULL path is dead 2404 } 2405 if ((*null_control) == top() && safe_for_replace) { 2406 replace_in_map(value, cast); 2407 } 2408 2409 // Cast away null-ness on the result 2410 return cast; 2411 } 2412 2413 //------------------------------opt_iff---------------------------------------- 2414 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2415 // Return slow-path control. 2416 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2417 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2418 2419 // Fast path taken; set region slot 2 2420 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2421 region->init_req(2,fast_taken); // Capture fast-control 2422 2423 // Fast path not-taken, i.e. slow path 2424 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2425 return slow_taken; 2426 } 2427 2428 //-----------------------------make_runtime_call------------------------------- 2429 Node* GraphKit::make_runtime_call(int flags, 2430 const TypeFunc* call_type, address call_addr, 2431 const char* call_name, 2432 const TypePtr* adr_type, 2433 // The following parms are all optional. 2434 // The first NULL ends the list. 2435 Node* parm0, Node* parm1, 2436 Node* parm2, Node* parm3, 2437 Node* parm4, Node* parm5, 2438 Node* parm6, Node* parm7) { 2439 // Slow-path call 2440 bool is_leaf = !(flags & RC_NO_LEAF); 2441 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2442 if (call_name == NULL) { 2443 assert(!is_leaf, "must supply name for leaf"); 2444 call_name = OptoRuntime::stub_name(call_addr); 2445 } 2446 CallNode* call; 2447 if (!is_leaf) { 2448 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2449 bci(), adr_type); 2450 } else if (flags & RC_NO_FP) { 2451 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2452 } else { 2453 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2454 } 2455 2456 // The following is similar to set_edges_for_java_call, 2457 // except that the memory effects of the call are restricted to AliasIdxRaw. 2458 2459 // Slow path call has no side-effects, uses few values 2460 bool wide_in = !(flags & RC_NARROW_MEM); 2461 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2462 2463 Node* prev_mem = NULL; 2464 if (wide_in) { 2465 prev_mem = set_predefined_input_for_runtime_call(call); 2466 } else { 2467 assert(!wide_out, "narrow in => narrow out"); 2468 Node* narrow_mem = memory(adr_type); 2469 prev_mem = reset_memory(); 2470 map()->set_memory(narrow_mem); 2471 set_predefined_input_for_runtime_call(call); 2472 } 2473 2474 // Hook each parm in order. Stop looking at the first NULL. 2475 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2476 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2477 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2478 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2479 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2480 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2481 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2482 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2483 /* close each nested if ===> */ } } } } } } } } 2484 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2485 2486 if (!is_leaf) { 2487 // Non-leaves can block and take safepoints: 2488 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2489 } 2490 // Non-leaves can throw exceptions: 2491 if (has_io) { 2492 call->set_req(TypeFunc::I_O, i_o()); 2493 } 2494 2495 if (flags & RC_UNCOMMON) { 2496 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2497 // (An "if" probability corresponds roughly to an unconditional count. 2498 // Sort of.) 2499 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2500 } 2501 2502 Node* c = _gvn.transform(call); 2503 assert(c == call, "cannot disappear"); 2504 2505 if (wide_out) { 2506 // Slow path call has full side-effects. 2507 set_predefined_output_for_runtime_call(call); 2508 } else { 2509 // Slow path call has few side-effects, and/or sets few values. 2510 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2511 } 2512 2513 if (has_io) { 2514 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2515 } 2516 return call; 2517 2518 } 2519 2520 //------------------------------merge_memory----------------------------------- 2521 // Merge memory from one path into the current memory state. 2522 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2523 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2524 Node* old_slice = mms.force_memory(); 2525 Node* new_slice = mms.memory2(); 2526 if (old_slice != new_slice) { 2527 PhiNode* phi; 2528 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2529 if (mms.is_empty()) { 2530 // clone base memory Phi's inputs for this memory slice 2531 assert(old_slice == mms.base_memory(), "sanity"); 2532 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2533 _gvn.set_type(phi, Type::MEMORY); 2534 for (uint i = 1; i < phi->req(); i++) { 2535 phi->init_req(i, old_slice->in(i)); 2536 } 2537 } else { 2538 phi = old_slice->as_Phi(); // Phi was generated already 2539 } 2540 } else { 2541 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2542 _gvn.set_type(phi, Type::MEMORY); 2543 } 2544 phi->set_req(new_path, new_slice); 2545 mms.set_memory(phi); 2546 } 2547 } 2548 } 2549 2550 //------------------------------make_slow_call_ex------------------------------ 2551 // Make the exception handler hookups for the slow call 2552 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2553 if (stopped()) return; 2554 2555 // Make a catch node with just two handlers: fall-through and catch-all 2556 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2557 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2558 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2559 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2560 2561 { PreserveJVMState pjvms(this); 2562 set_control(excp); 2563 set_i_o(i_o); 2564 2565 if (excp != top()) { 2566 if (deoptimize) { 2567 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2568 uncommon_trap(Deoptimization::Reason_unhandled, 2569 Deoptimization::Action_none); 2570 } else { 2571 // Create an exception state also. 2572 // Use an exact type if the caller has specified a specific exception. 2573 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2574 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2575 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2576 } 2577 } 2578 } 2579 2580 // Get the no-exception control from the CatchNode. 2581 set_control(norm); 2582 } 2583 2584 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2585 Node* cmp = NULL; 2586 switch(bt) { 2587 case T_INT: cmp = new CmpINode(in1, in2); break; 2588 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2589 default: fatal("unexpected comparison type %s", type2name(bt)); 2590 } 2591 gvn->transform(cmp); 2592 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2593 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2594 gvn->transform(iff); 2595 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2596 return iff; 2597 } 2598 2599 2600 //-------------------------------gen_subtype_check----------------------------- 2601 // Generate a subtyping check. Takes as input the subtype and supertype. 2602 // Returns 2 values: sets the default control() to the true path and returns 2603 // the false path. Only reads invariant memory; sets no (visible) memory. 2604 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2605 // but that's not exposed to the optimizer. This call also doesn't take in an 2606 // Object; if you wish to check an Object you need to load the Object's class 2607 // prior to coming here. 2608 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2609 Compile* C = gvn->C; 2610 2611 if ((*ctrl)->is_top()) { 2612 return C->top(); 2613 } 2614 2615 // Fast check for identical types, perhaps identical constants. 2616 // The types can even be identical non-constants, in cases 2617 // involving Array.newInstance, Object.clone, etc. 2618 if (subklass == superklass) 2619 return C->top(); // false path is dead; no test needed. 2620 2621 if (gvn->type(superklass)->singleton()) { 2622 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2623 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2624 2625 // In the common case of an exact superklass, try to fold up the 2626 // test before generating code. You may ask, why not just generate 2627 // the code and then let it fold up? The answer is that the generated 2628 // code will necessarily include null checks, which do not always 2629 // completely fold away. If they are also needless, then they turn 2630 // into a performance loss. Example: 2631 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2632 // Here, the type of 'fa' is often exact, so the store check 2633 // of fa[1]=x will fold up, without testing the nullness of x. 2634 switch (C->static_subtype_check(superk, subk)) { 2635 case Compile::SSC_always_false: 2636 { 2637 Node* always_fail = *ctrl; 2638 *ctrl = gvn->C->top(); 2639 return always_fail; 2640 } 2641 case Compile::SSC_always_true: 2642 return C->top(); 2643 case Compile::SSC_easy_test: 2644 { 2645 // Just do a direct pointer compare and be done. 2646 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2647 *ctrl = gvn->transform(new IfTrueNode(iff)); 2648 return gvn->transform(new IfFalseNode(iff)); 2649 } 2650 case Compile::SSC_full_test: 2651 break; 2652 default: 2653 ShouldNotReachHere(); 2654 } 2655 } 2656 2657 // %%% Possible further optimization: Even if the superklass is not exact, 2658 // if the subklass is the unique subtype of the superklass, the check 2659 // will always succeed. We could leave a dependency behind to ensure this. 2660 2661 // First load the super-klass's check-offset 2662 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2663 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2664 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2665 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2666 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2667 2668 // Load from the sub-klass's super-class display list, or a 1-word cache of 2669 // the secondary superclass list, or a failing value with a sentinel offset 2670 // if the super-klass is an interface or exceptionally deep in the Java 2671 // hierarchy and we have to scan the secondary superclass list the hard way. 2672 // Worst-case type is a little odd: NULL is allowed as a result (usually 2673 // klass loads can never produce a NULL). 2674 Node *chk_off_X = chk_off; 2675 #ifdef _LP64 2676 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2677 #endif 2678 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2679 // For some types like interfaces the following loadKlass is from a 1-word 2680 // cache which is mutable so can't use immutable memory. Other 2681 // types load from the super-class display table which is immutable. 2682 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2683 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2684 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2685 2686 // Compile speed common case: ARE a subtype and we canNOT fail 2687 if( superklass == nkls ) 2688 return C->top(); // false path is dead; no test needed. 2689 2690 // See if we get an immediate positive hit. Happens roughly 83% of the 2691 // time. Test to see if the value loaded just previously from the subklass 2692 // is exactly the superklass. 2693 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2694 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2695 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2696 2697 // Compile speed common case: Check for being deterministic right now. If 2698 // chk_off is a constant and not equal to cacheoff then we are NOT a 2699 // subklass. In this case we need exactly the 1 test above and we can 2700 // return those results immediately. 2701 if (!might_be_cache) { 2702 Node* not_subtype_ctrl = *ctrl; 2703 *ctrl = iftrue1; // We need exactly the 1 test above 2704 return not_subtype_ctrl; 2705 } 2706 2707 // Gather the various success & failures here 2708 RegionNode *r_ok_subtype = new RegionNode(4); 2709 gvn->record_for_igvn(r_ok_subtype); 2710 RegionNode *r_not_subtype = new RegionNode(3); 2711 gvn->record_for_igvn(r_not_subtype); 2712 2713 r_ok_subtype->init_req(1, iftrue1); 2714 2715 // Check for immediate negative hit. Happens roughly 11% of the time (which 2716 // is roughly 63% of the remaining cases). Test to see if the loaded 2717 // check-offset points into the subklass display list or the 1-element 2718 // cache. If it points to the display (and NOT the cache) and the display 2719 // missed then it's not a subtype. 2720 Node *cacheoff = gvn->intcon(cacheoff_con); 2721 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2722 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2723 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2724 2725 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2726 // No performance impact (too rare) but allows sharing of secondary arrays 2727 // which has some footprint reduction. 2728 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2729 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2730 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2731 2732 // -- Roads not taken here: -- 2733 // We could also have chosen to perform the self-check at the beginning 2734 // of this code sequence, as the assembler does. This would not pay off 2735 // the same way, since the optimizer, unlike the assembler, can perform 2736 // static type analysis to fold away many successful self-checks. 2737 // Non-foldable self checks work better here in second position, because 2738 // the initial primary superclass check subsumes a self-check for most 2739 // types. An exception would be a secondary type like array-of-interface, 2740 // which does not appear in its own primary supertype display. 2741 // Finally, we could have chosen to move the self-check into the 2742 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2743 // dependent manner. But it is worthwhile to have the check here, 2744 // where it can be perhaps be optimized. The cost in code space is 2745 // small (register compare, branch). 2746 2747 // Now do a linear scan of the secondary super-klass array. Again, no real 2748 // performance impact (too rare) but it's gotta be done. 2749 // Since the code is rarely used, there is no penalty for moving it 2750 // out of line, and it can only improve I-cache density. 2751 // The decision to inline or out-of-line this final check is platform 2752 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2753 Node* psc = gvn->transform( 2754 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2755 2756 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2757 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2758 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2759 2760 // Return false path; set default control to true path. 2761 *ctrl = gvn->transform(r_ok_subtype); 2762 return gvn->transform(r_not_subtype); 2763 } 2764 2765 // Profile-driven exact type check: 2766 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2767 float prob, 2768 Node* *casted_receiver) { 2769 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2770 Node* recv_klass = load_object_klass(receiver); 2771 Node* want_klass = makecon(tklass); 2772 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2773 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2774 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2775 set_control( _gvn.transform( new IfTrueNode (iff) )); 2776 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2777 2778 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2779 assert(recv_xtype->klass_is_exact(), ""); 2780 2781 // Subsume downstream occurrences of receiver with a cast to 2782 // recv_xtype, since now we know what the type will be. 2783 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2784 (*casted_receiver) = _gvn.transform(cast); 2785 // (User must make the replace_in_map call.) 2786 2787 return fail; 2788 } 2789 2790 2791 //------------------------------seems_never_null------------------------------- 2792 // Use null_seen information if it is available from the profile. 2793 // If we see an unexpected null at a type check we record it and force a 2794 // recompile; the offending check will be recompiled to handle NULLs. 2795 // If we see several offending BCIs, then all checks in the 2796 // method will be recompiled. 2797 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2798 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2799 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2800 if (UncommonNullCast // Cutout for this technique 2801 && obj != null() // And not the -Xcomp stupid case? 2802 && !too_many_traps(reason) 2803 ) { 2804 if (speculating) { 2805 return true; 2806 } 2807 if (data == NULL) 2808 // Edge case: no mature data. Be optimistic here. 2809 return true; 2810 // If the profile has not seen a null, assume it won't happen. 2811 assert(java_bc() == Bytecodes::_checkcast || 2812 java_bc() == Bytecodes::_instanceof || 2813 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2814 return !data->as_BitData()->null_seen(); 2815 } 2816 speculating = false; 2817 return false; 2818 } 2819 2820 //------------------------maybe_cast_profiled_receiver------------------------- 2821 // If the profile has seen exactly one type, narrow to exactly that type. 2822 // Subsequent type checks will always fold up. 2823 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2824 ciKlass* require_klass, 2825 ciKlass* spec_klass, 2826 bool safe_for_replace) { 2827 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2828 2829 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2830 2831 // Make sure we haven't already deoptimized from this tactic. 2832 if (too_many_traps(reason) || too_many_recompiles(reason)) 2833 return NULL; 2834 2835 // (No, this isn't a call, but it's enough like a virtual call 2836 // to use the same ciMethod accessor to get the profile info...) 2837 // If we have a speculative type use it instead of profiling (which 2838 // may not help us) 2839 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2840 if (exact_kls != NULL) {// no cast failures here 2841 if (require_klass == NULL || 2842 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2843 // If we narrow the type to match what the type profile sees or 2844 // the speculative type, we can then remove the rest of the 2845 // cast. 2846 // This is a win, even if the exact_kls is very specific, 2847 // because downstream operations, such as method calls, 2848 // will often benefit from the sharper type. 2849 Node* exact_obj = not_null_obj; // will get updated in place... 2850 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2851 &exact_obj); 2852 { PreserveJVMState pjvms(this); 2853 set_control(slow_ctl); 2854 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2855 } 2856 if (safe_for_replace) { 2857 replace_in_map(not_null_obj, exact_obj); 2858 } 2859 return exact_obj; 2860 } 2861 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2862 } 2863 2864 return NULL; 2865 } 2866 2867 /** 2868 * Cast obj to type and emit guard unless we had too many traps here 2869 * already 2870 * 2871 * @param obj node being casted 2872 * @param type type to cast the node to 2873 * @param not_null true if we know node cannot be null 2874 */ 2875 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2876 ciKlass* type, 2877 bool not_null) { 2878 if (stopped()) { 2879 return obj; 2880 } 2881 2882 // type == NULL if profiling tells us this object is always null 2883 if (type != NULL) { 2884 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2885 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2886 2887 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2888 !too_many_traps(class_reason) && 2889 !too_many_recompiles(class_reason)) { 2890 Node* not_null_obj = NULL; 2891 // not_null is true if we know the object is not null and 2892 // there's no need for a null check 2893 if (!not_null) { 2894 Node* null_ctl = top(); 2895 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2896 assert(null_ctl->is_top(), "no null control here"); 2897 } else { 2898 not_null_obj = obj; 2899 } 2900 2901 Node* exact_obj = not_null_obj; 2902 ciKlass* exact_kls = type; 2903 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2904 &exact_obj); 2905 { 2906 PreserveJVMState pjvms(this); 2907 set_control(slow_ctl); 2908 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2909 } 2910 replace_in_map(not_null_obj, exact_obj); 2911 obj = exact_obj; 2912 } 2913 } else { 2914 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2915 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2916 Node* exact_obj = null_assert(obj); 2917 replace_in_map(obj, exact_obj); 2918 obj = exact_obj; 2919 } 2920 } 2921 return obj; 2922 } 2923 2924 //-------------------------------gen_instanceof-------------------------------- 2925 // Generate an instance-of idiom. Used by both the instance-of bytecode 2926 // and the reflective instance-of call. 2927 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2928 kill_dead_locals(); // Benefit all the uncommon traps 2929 assert( !stopped(), "dead parse path should be checked in callers" ); 2930 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2931 "must check for not-null not-dead klass in callers"); 2932 2933 // Make the merge point 2934 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2935 RegionNode* region = new RegionNode(PATH_LIMIT); 2936 Node* phi = new PhiNode(region, TypeInt::BOOL); 2937 C->set_has_split_ifs(true); // Has chance for split-if optimization 2938 2939 ciProfileData* data = NULL; 2940 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2941 data = method()->method_data()->bci_to_data(bci()); 2942 } 2943 bool speculative_not_null = false; 2944 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2945 && seems_never_null(obj, data, speculative_not_null)); 2946 2947 // Null check; get casted pointer; set region slot 3 2948 Node* null_ctl = top(); 2949 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2950 2951 // If not_null_obj is dead, only null-path is taken 2952 if (stopped()) { // Doing instance-of on a NULL? 2953 set_control(null_ctl); 2954 return intcon(0); 2955 } 2956 region->init_req(_null_path, null_ctl); 2957 phi ->init_req(_null_path, intcon(0)); // Set null path value 2958 if (null_ctl == top()) { 2959 // Do this eagerly, so that pattern matches like is_diamond_phi 2960 // will work even during parsing. 2961 assert(_null_path == PATH_LIMIT-1, "delete last"); 2962 region->del_req(_null_path); 2963 phi ->del_req(_null_path); 2964 } 2965 2966 // Do we know the type check always succeed? 2967 bool known_statically = false; 2968 if (_gvn.type(superklass)->singleton()) { 2969 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2970 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2971 if (subk != NULL && subk->is_loaded()) { 2972 int static_res = C->static_subtype_check(superk, subk); 2973 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2974 } 2975 } 2976 2977 if (!known_statically) { 2978 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2979 // We may not have profiling here or it may not help us. If we 2980 // have a speculative type use it to perform an exact cast. 2981 ciKlass* spec_obj_type = obj_type->speculative_type(); 2982 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2983 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2984 if (stopped()) { // Profile disagrees with this path. 2985 set_control(null_ctl); // Null is the only remaining possibility. 2986 return intcon(0); 2987 } 2988 if (cast_obj != NULL) { 2989 not_null_obj = cast_obj; 2990 } 2991 } 2992 } 2993 2994 // Load the object's klass 2995 Node* obj_klass = load_object_klass(not_null_obj); 2996 2997 // Generate the subtype check 2998 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2999 3000 // Plug in the success path to the general merge in slot 1. 3001 region->init_req(_obj_path, control()); 3002 phi ->init_req(_obj_path, intcon(1)); 3003 3004 // Plug in the failing path to the general merge in slot 2. 3005 region->init_req(_fail_path, not_subtype_ctrl); 3006 phi ->init_req(_fail_path, intcon(0)); 3007 3008 // Return final merged results 3009 set_control( _gvn.transform(region) ); 3010 record_for_igvn(region); 3011 3012 // If we know the type check always succeeds then we don't use the 3013 // profiling data at this bytecode. Don't lose it, feed it to the 3014 // type system as a speculative type. 3015 if (safe_for_replace) { 3016 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3017 replace_in_map(obj, casted_obj); 3018 } 3019 3020 return _gvn.transform(phi); 3021 } 3022 3023 //-------------------------------gen_checkcast--------------------------------- 3024 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3025 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3026 // uncommon-trap paths work. Adjust stack after this call. 3027 // If failure_control is supplied and not null, it is filled in with 3028 // the control edge for the cast failure. Otherwise, an appropriate 3029 // uncommon trap or exception is thrown. 3030 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3031 Node* *failure_control) { 3032 kill_dead_locals(); // Benefit all the uncommon traps 3033 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3034 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3035 3036 // Fast cutout: Check the case that the cast is vacuously true. 3037 // This detects the common cases where the test will short-circuit 3038 // away completely. We do this before we perform the null check, 3039 // because if the test is going to turn into zero code, we don't 3040 // want a residual null check left around. (Causes a slowdown, 3041 // for example, in some objArray manipulations, such as a[i]=a[j].) 3042 if (tk->singleton()) { 3043 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3044 if (objtp != NULL && objtp->klass() != NULL) { 3045 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3046 case Compile::SSC_always_true: 3047 // If we know the type check always succeed then we don't use 3048 // the profiling data at this bytecode. Don't lose it, feed it 3049 // to the type system as a speculative type. 3050 return record_profiled_receiver_for_speculation(obj); 3051 case Compile::SSC_always_false: 3052 // It needs a null check because a null will *pass* the cast check. 3053 // A non-null value will always produce an exception. 3054 return null_assert(obj); 3055 } 3056 } 3057 } 3058 3059 ciProfileData* data = NULL; 3060 bool safe_for_replace = false; 3061 if (failure_control == NULL) { // use MDO in regular case only 3062 assert(java_bc() == Bytecodes::_aastore || 3063 java_bc() == Bytecodes::_checkcast, 3064 "interpreter profiles type checks only for these BCs"); 3065 data = method()->method_data()->bci_to_data(bci()); 3066 safe_for_replace = true; 3067 } 3068 3069 // Make the merge point 3070 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3071 RegionNode* region = new RegionNode(PATH_LIMIT); 3072 Node* phi = new PhiNode(region, toop); 3073 C->set_has_split_ifs(true); // Has chance for split-if optimization 3074 3075 // Use null-cast information if it is available 3076 bool speculative_not_null = false; 3077 bool never_see_null = ((failure_control == NULL) // regular case only 3078 && seems_never_null(obj, data, speculative_not_null)); 3079 3080 // Null check; get casted pointer; set region slot 3 3081 Node* null_ctl = top(); 3082 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3083 3084 // If not_null_obj is dead, only null-path is taken 3085 if (stopped()) { // Doing instance-of on a NULL? 3086 set_control(null_ctl); 3087 return null(); 3088 } 3089 region->init_req(_null_path, null_ctl); 3090 phi ->init_req(_null_path, null()); // Set null path value 3091 if (null_ctl == top()) { 3092 // Do this eagerly, so that pattern matches like is_diamond_phi 3093 // will work even during parsing. 3094 assert(_null_path == PATH_LIMIT-1, "delete last"); 3095 region->del_req(_null_path); 3096 phi ->del_req(_null_path); 3097 } 3098 3099 Node* cast_obj = NULL; 3100 if (tk->klass_is_exact()) { 3101 // The following optimization tries to statically cast the speculative type of the object 3102 // (for example obtained during profiling) to the type of the superklass and then do a 3103 // dynamic check that the type of the object is what we expect. To work correctly 3104 // for checkcast and aastore the type of superklass should be exact. 3105 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3106 // We may not have profiling here or it may not help us. If we have 3107 // a speculative type use it to perform an exact cast. 3108 ciKlass* spec_obj_type = obj_type->speculative_type(); 3109 if (spec_obj_type != NULL || data != NULL) { 3110 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3111 if (cast_obj != NULL) { 3112 if (failure_control != NULL) // failure is now impossible 3113 (*failure_control) = top(); 3114 // adjust the type of the phi to the exact klass: 3115 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3116 } 3117 } 3118 } 3119 3120 if (cast_obj == NULL) { 3121 // Load the object's klass 3122 Node* obj_klass = load_object_klass(not_null_obj); 3123 3124 // Generate the subtype check 3125 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3126 3127 // Plug in success path into the merge 3128 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3129 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3130 if (failure_control == NULL) { 3131 if (not_subtype_ctrl != top()) { // If failure is possible 3132 PreserveJVMState pjvms(this); 3133 set_control(not_subtype_ctrl); 3134 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3135 } 3136 } else { 3137 (*failure_control) = not_subtype_ctrl; 3138 } 3139 } 3140 3141 region->init_req(_obj_path, control()); 3142 phi ->init_req(_obj_path, cast_obj); 3143 3144 // A merge of NULL or Casted-NotNull obj 3145 Node* res = _gvn.transform(phi); 3146 3147 // Note I do NOT always 'replace_in_map(obj,result)' here. 3148 // if( tk->klass()->can_be_primary_super() ) 3149 // This means that if I successfully store an Object into an array-of-String 3150 // I 'forget' that the Object is really now known to be a String. I have to 3151 // do this because we don't have true union types for interfaces - if I store 3152 // a Baz into an array-of-Interface and then tell the optimizer it's an 3153 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3154 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3155 // replace_in_map( obj, res ); 3156 3157 // Return final merged results 3158 set_control( _gvn.transform(region) ); 3159 record_for_igvn(region); 3160 3161 return record_profiled_receiver_for_speculation(res); 3162 } 3163 3164 //------------------------------next_monitor----------------------------------- 3165 // What number should be given to the next monitor? 3166 int GraphKit::next_monitor() { 3167 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3168 int next = current + C->sync_stack_slots(); 3169 // Keep the toplevel high water mark current: 3170 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3171 return current; 3172 } 3173 3174 //------------------------------insert_mem_bar--------------------------------- 3175 // Memory barrier to avoid floating things around 3176 // The membar serves as a pinch point between both control and all memory slices. 3177 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3178 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3179 mb->init_req(TypeFunc::Control, control()); 3180 mb->init_req(TypeFunc::Memory, reset_memory()); 3181 Node* membar = _gvn.transform(mb); 3182 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3183 set_all_memory_call(membar); 3184 return membar; 3185 } 3186 3187 //-------------------------insert_mem_bar_volatile---------------------------- 3188 // Memory barrier to avoid floating things around 3189 // The membar serves as a pinch point between both control and memory(alias_idx). 3190 // If you want to make a pinch point on all memory slices, do not use this 3191 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3192 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3193 // When Parse::do_put_xxx updates a volatile field, it appends a series 3194 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3195 // The first membar is on the same memory slice as the field store opcode. 3196 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3197 // All the other membars (for other volatile slices, including AliasIdxBot, 3198 // which stands for all unknown volatile slices) are control-dependent 3199 // on the first membar. This prevents later volatile loads or stores 3200 // from sliding up past the just-emitted store. 3201 3202 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3203 mb->set_req(TypeFunc::Control,control()); 3204 if (alias_idx == Compile::AliasIdxBot) { 3205 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3206 } else { 3207 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3208 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3209 } 3210 Node* membar = _gvn.transform(mb); 3211 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3212 if (alias_idx == Compile::AliasIdxBot) { 3213 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3214 } else { 3215 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3216 } 3217 return membar; 3218 } 3219 3220 //------------------------------shared_lock------------------------------------ 3221 // Emit locking code. 3222 FastLockNode* GraphKit::shared_lock(Node* obj) { 3223 // bci is either a monitorenter bc or InvocationEntryBci 3224 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3225 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3226 3227 if( !GenerateSynchronizationCode ) 3228 return NULL; // Not locking things? 3229 if (stopped()) // Dead monitor? 3230 return NULL; 3231 3232 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3233 3234 obj = shenandoah_write_barrier(obj); 3235 3236 // Box the stack location 3237 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3238 Node* mem = reset_memory(); 3239 3240 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3241 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3242 // Create the counters for this fast lock. 3243 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3244 } 3245 3246 // Create the rtm counters for this fast lock if needed. 3247 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3248 3249 // Add monitor to debug info for the slow path. If we block inside the 3250 // slow path and de-opt, we need the monitor hanging around 3251 map()->push_monitor( flock ); 3252 3253 const TypeFunc *tf = LockNode::lock_type(); 3254 LockNode *lock = new LockNode(C, tf); 3255 3256 lock->init_req( TypeFunc::Control, control() ); 3257 lock->init_req( TypeFunc::Memory , mem ); 3258 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3259 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3260 lock->init_req( TypeFunc::ReturnAdr, top() ); 3261 3262 lock->init_req(TypeFunc::Parms + 0, obj); 3263 lock->init_req(TypeFunc::Parms + 1, box); 3264 lock->init_req(TypeFunc::Parms + 2, flock); 3265 add_safepoint_edges(lock); 3266 3267 lock = _gvn.transform( lock )->as_Lock(); 3268 3269 // lock has no side-effects, sets few values 3270 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3271 3272 insert_mem_bar(Op_MemBarAcquireLock); 3273 3274 // Add this to the worklist so that the lock can be eliminated 3275 record_for_igvn(lock); 3276 3277 #ifndef PRODUCT 3278 if (PrintLockStatistics) { 3279 // Update the counter for this lock. Don't bother using an atomic 3280 // operation since we don't require absolute accuracy. 3281 lock->create_lock_counter(map()->jvms()); 3282 increment_counter(lock->counter()->addr()); 3283 } 3284 #endif 3285 3286 return flock; 3287 } 3288 3289 3290 //------------------------------shared_unlock---------------------------------- 3291 // Emit unlocking code. 3292 void GraphKit::shared_unlock(Node* box, Node* obj) { 3293 // bci is either a monitorenter bc or InvocationEntryBci 3294 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3295 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3296 3297 if( !GenerateSynchronizationCode ) 3298 return; 3299 if (stopped()) { // Dead monitor? 3300 map()->pop_monitor(); // Kill monitor from debug info 3301 return; 3302 } 3303 3304 // Memory barrier to avoid floating things down past the locked region 3305 insert_mem_bar(Op_MemBarReleaseLock); 3306 3307 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3308 UnlockNode *unlock = new UnlockNode(C, tf); 3309 #ifdef ASSERT 3310 unlock->set_dbg_jvms(sync_jvms()); 3311 #endif 3312 uint raw_idx = Compile::AliasIdxRaw; 3313 unlock->init_req( TypeFunc::Control, control() ); 3314 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3315 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3316 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3317 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3318 3319 unlock->init_req(TypeFunc::Parms + 0, obj); 3320 unlock->init_req(TypeFunc::Parms + 1, box); 3321 unlock = _gvn.transform(unlock)->as_Unlock(); 3322 3323 Node* mem = reset_memory(); 3324 3325 // unlock has no side-effects, sets few values 3326 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3327 3328 // Kill monitor from debug info 3329 map()->pop_monitor( ); 3330 } 3331 3332 //-------------------------------get_layout_helper----------------------------- 3333 // If the given klass is a constant or known to be an array, 3334 // fetch the constant layout helper value into constant_value 3335 // and return (Node*)NULL. Otherwise, load the non-constant 3336 // layout helper value, and return the node which represents it. 3337 // This two-faced routine is useful because allocation sites 3338 // almost always feature constant types. 3339 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3340 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3341 if (!StressReflectiveCode && inst_klass != NULL) { 3342 ciKlass* klass = inst_klass->klass(); 3343 bool xklass = inst_klass->klass_is_exact(); 3344 if (xklass || klass->is_array_klass()) { 3345 jint lhelper = klass->layout_helper(); 3346 if (lhelper != Klass::_lh_neutral_value) { 3347 constant_value = lhelper; 3348 return (Node*) NULL; 3349 } 3350 } 3351 } 3352 constant_value = Klass::_lh_neutral_value; // put in a known value 3353 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3354 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3355 } 3356 3357 // We just put in an allocate/initialize with a big raw-memory effect. 3358 // Hook selected additional alias categories on the initialization. 3359 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3360 MergeMemNode* init_in_merge, 3361 Node* init_out_raw) { 3362 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3363 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3364 3365 Node* prevmem = kit.memory(alias_idx); 3366 init_in_merge->set_memory_at(alias_idx, prevmem); 3367 kit.set_memory(init_out_raw, alias_idx); 3368 } 3369 3370 //---------------------------set_output_for_allocation------------------------- 3371 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3372 const TypeOopPtr* oop_type, 3373 bool deoptimize_on_exception) { 3374 int rawidx = Compile::AliasIdxRaw; 3375 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3376 add_safepoint_edges(alloc); 3377 Node* allocx = _gvn.transform(alloc); 3378 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3379 // create memory projection for i_o 3380 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3381 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3382 3383 // create a memory projection as for the normal control path 3384 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3385 set_memory(malloc, rawidx); 3386 3387 // a normal slow-call doesn't change i_o, but an allocation does 3388 // we create a separate i_o projection for the normal control path 3389 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3390 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3391 3392 // put in an initialization barrier 3393 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3394 rawoop)->as_Initialize(); 3395 assert(alloc->initialization() == init, "2-way macro link must work"); 3396 assert(init ->allocation() == alloc, "2-way macro link must work"); 3397 { 3398 // Extract memory strands which may participate in the new object's 3399 // initialization, and source them from the new InitializeNode. 3400 // This will allow us to observe initializations when they occur, 3401 // and link them properly (as a group) to the InitializeNode. 3402 assert(init->in(InitializeNode::Memory) == malloc, ""); 3403 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3404 init->set_req(InitializeNode::Memory, minit_in); 3405 record_for_igvn(minit_in); // fold it up later, if possible 3406 Node* minit_out = memory(rawidx); 3407 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3408 if (oop_type->isa_aryptr()) { 3409 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3410 int elemidx = C->get_alias_index(telemref); 3411 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3412 } else if (oop_type->isa_instptr()) { 3413 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3414 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3415 ciField* field = ik->nonstatic_field_at(i); 3416 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3417 continue; // do not bother to track really large numbers of fields 3418 // Find (or create) the alias category for this field: 3419 int fieldidx = C->alias_type(field)->index(); 3420 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3421 } 3422 } 3423 } 3424 3425 // Cast raw oop to the real thing... 3426 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3427 javaoop = _gvn.transform(javaoop); 3428 C->set_recent_alloc(control(), javaoop); 3429 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3430 3431 #ifdef ASSERT 3432 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3433 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3434 "Ideal_allocation works"); 3435 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3436 "Ideal_allocation works"); 3437 if (alloc->is_AllocateArray()) { 3438 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3439 "Ideal_allocation works"); 3440 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3441 "Ideal_allocation works"); 3442 } else { 3443 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3444 } 3445 } 3446 #endif //ASSERT 3447 3448 return javaoop; 3449 } 3450 3451 //---------------------------new_instance-------------------------------------- 3452 // This routine takes a klass_node which may be constant (for a static type) 3453 // or may be non-constant (for reflective code). It will work equally well 3454 // for either, and the graph will fold nicely if the optimizer later reduces 3455 // the type to a constant. 3456 // The optional arguments are for specialized use by intrinsics: 3457 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3458 // - If 'return_size_val', report the the total object size to the caller. 3459 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3460 Node* GraphKit::new_instance(Node* klass_node, 3461 Node* extra_slow_test, 3462 Node* *return_size_val, 3463 bool deoptimize_on_exception) { 3464 // Compute size in doublewords 3465 // The size is always an integral number of doublewords, represented 3466 // as a positive bytewise size stored in the klass's layout_helper. 3467 // The layout_helper also encodes (in a low bit) the need for a slow path. 3468 jint layout_con = Klass::_lh_neutral_value; 3469 Node* layout_val = get_layout_helper(klass_node, layout_con); 3470 int layout_is_con = (layout_val == NULL); 3471 3472 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3473 // Generate the initial go-slow test. It's either ALWAYS (return a 3474 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3475 // case) a computed value derived from the layout_helper. 3476 Node* initial_slow_test = NULL; 3477 if (layout_is_con) { 3478 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3479 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3480 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3481 } else { // reflective case 3482 // This reflective path is used by Unsafe.allocateInstance. 3483 // (It may be stress-tested by specifying StressReflectiveCode.) 3484 // Basically, we want to get into the VM is there's an illegal argument. 3485 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3486 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3487 if (extra_slow_test != intcon(0)) { 3488 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3489 } 3490 // (Macro-expander will further convert this to a Bool, if necessary.) 3491 } 3492 3493 // Find the size in bytes. This is easy; it's the layout_helper. 3494 // The size value must be valid even if the slow path is taken. 3495 Node* size = NULL; 3496 if (layout_is_con) { 3497 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3498 } else { // reflective case 3499 // This reflective path is used by clone and Unsafe.allocateInstance. 3500 size = ConvI2X(layout_val); 3501 3502 // Clear the low bits to extract layout_helper_size_in_bytes: 3503 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3504 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3505 size = _gvn.transform( new AndXNode(size, mask) ); 3506 } 3507 if (return_size_val != NULL) { 3508 (*return_size_val) = size; 3509 } 3510 3511 // This is a precise notnull oop of the klass. 3512 // (Actually, it need not be precise if this is a reflective allocation.) 3513 // It's what we cast the result to. 3514 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3515 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3516 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3517 3518 // Now generate allocation code 3519 3520 // The entire memory state is needed for slow path of the allocation 3521 // since GC and deoptimization can happened. 3522 Node *mem = reset_memory(); 3523 set_all_memory(mem); // Create new memory state 3524 3525 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3526 control(), mem, i_o(), 3527 size, klass_node, 3528 initial_slow_test); 3529 3530 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3531 } 3532 3533 //-------------------------------new_array------------------------------------- 3534 // helper for both newarray and anewarray 3535 // The 'length' parameter is (obviously) the length of the array. 3536 // See comments on new_instance for the meaning of the other arguments. 3537 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3538 Node* length, // number of array elements 3539 int nargs, // number of arguments to push back for uncommon trap 3540 Node* *return_size_val, 3541 bool deoptimize_on_exception) { 3542 jint layout_con = Klass::_lh_neutral_value; 3543 Node* layout_val = get_layout_helper(klass_node, layout_con); 3544 int layout_is_con = (layout_val == NULL); 3545 3546 if (!layout_is_con && !StressReflectiveCode && 3547 !too_many_traps(Deoptimization::Reason_class_check)) { 3548 // This is a reflective array creation site. 3549 // Optimistically assume that it is a subtype of Object[], 3550 // so that we can fold up all the address arithmetic. 3551 layout_con = Klass::array_layout_helper(T_OBJECT); 3552 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3553 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3554 { BuildCutout unless(this, bol_lh, PROB_MAX); 3555 inc_sp(nargs); 3556 uncommon_trap(Deoptimization::Reason_class_check, 3557 Deoptimization::Action_maybe_recompile); 3558 } 3559 layout_val = NULL; 3560 layout_is_con = true; 3561 } 3562 3563 // Generate the initial go-slow test. Make sure we do not overflow 3564 // if length is huge (near 2Gig) or negative! We do not need 3565 // exact double-words here, just a close approximation of needed 3566 // double-words. We can't add any offset or rounding bits, lest we 3567 // take a size -1 of bytes and make it positive. Use an unsigned 3568 // compare, so negative sizes look hugely positive. 3569 int fast_size_limit = FastAllocateSizeLimit; 3570 if (layout_is_con) { 3571 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3572 // Increase the size limit if we have exact knowledge of array type. 3573 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3574 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3575 } 3576 3577 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3578 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3579 3580 // --- Size Computation --- 3581 // array_size = round_to_heap(array_header + (length << elem_shift)); 3582 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3583 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3584 // The rounding mask is strength-reduced, if possible. 3585 int round_mask = MinObjAlignmentInBytes - 1; 3586 Node* header_size = NULL; 3587 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3588 // (T_BYTE has the weakest alignment and size restrictions...) 3589 if (layout_is_con) { 3590 int hsize = Klass::layout_helper_header_size(layout_con); 3591 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3592 BasicType etype = Klass::layout_helper_element_type(layout_con); 3593 if ((round_mask & ~right_n_bits(eshift)) == 0) 3594 round_mask = 0; // strength-reduce it if it goes away completely 3595 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3596 assert(header_size_min <= hsize, "generic minimum is smallest"); 3597 header_size_min = hsize; 3598 header_size = intcon(hsize + round_mask); 3599 } else { 3600 Node* hss = intcon(Klass::_lh_header_size_shift); 3601 Node* hsm = intcon(Klass::_lh_header_size_mask); 3602 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3603 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3604 Node* mask = intcon(round_mask); 3605 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3606 } 3607 3608 Node* elem_shift = NULL; 3609 if (layout_is_con) { 3610 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3611 if (eshift != 0) 3612 elem_shift = intcon(eshift); 3613 } else { 3614 // There is no need to mask or shift this value. 3615 // The semantics of LShiftINode include an implicit mask to 0x1F. 3616 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3617 elem_shift = layout_val; 3618 } 3619 3620 // Transition to native address size for all offset calculations: 3621 Node* lengthx = ConvI2X(length); 3622 Node* headerx = ConvI2X(header_size); 3623 #ifdef _LP64 3624 { const TypeInt* tilen = _gvn.find_int_type(length); 3625 if (tilen != NULL && tilen->_lo < 0) { 3626 // Add a manual constraint to a positive range. Cf. array_element_address. 3627 jint size_max = fast_size_limit; 3628 if (size_max > tilen->_hi) size_max = tilen->_hi; 3629 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3630 3631 // Only do a narrow I2L conversion if the range check passed. 3632 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3633 _gvn.transform(iff); 3634 RegionNode* region = new RegionNode(3); 3635 _gvn.set_type(region, Type::CONTROL); 3636 lengthx = new PhiNode(region, TypeLong::LONG); 3637 _gvn.set_type(lengthx, TypeLong::LONG); 3638 3639 // Range check passed. Use ConvI2L node with narrow type. 3640 Node* passed = IfFalse(iff); 3641 region->init_req(1, passed); 3642 // Make I2L conversion control dependent to prevent it from 3643 // floating above the range check during loop optimizations. 3644 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3645 3646 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3647 region->init_req(2, IfTrue(iff)); 3648 lengthx->init_req(2, ConvI2X(length)); 3649 3650 set_control(region); 3651 record_for_igvn(region); 3652 record_for_igvn(lengthx); 3653 } 3654 } 3655 #endif 3656 3657 // Combine header size (plus rounding) and body size. Then round down. 3658 // This computation cannot overflow, because it is used only in two 3659 // places, one where the length is sharply limited, and the other 3660 // after a successful allocation. 3661 Node* abody = lengthx; 3662 if (elem_shift != NULL) 3663 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3664 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3665 if (round_mask != 0) { 3666 Node* mask = MakeConX(~round_mask); 3667 size = _gvn.transform( new AndXNode(size, mask) ); 3668 } 3669 // else if round_mask == 0, the size computation is self-rounding 3670 3671 if (return_size_val != NULL) { 3672 // This is the size 3673 (*return_size_val) = size; 3674 } 3675 3676 // Now generate allocation code 3677 3678 // The entire memory state is needed for slow path of the allocation 3679 // since GC and deoptimization can happened. 3680 Node *mem = reset_memory(); 3681 set_all_memory(mem); // Create new memory state 3682 3683 if (initial_slow_test->is_Bool()) { 3684 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3685 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3686 } 3687 3688 // Create the AllocateArrayNode and its result projections 3689 AllocateArrayNode* alloc 3690 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3691 control(), mem, i_o(), 3692 size, klass_node, 3693 initial_slow_test, 3694 length); 3695 3696 // Cast to correct type. Note that the klass_node may be constant or not, 3697 // and in the latter case the actual array type will be inexact also. 3698 // (This happens via a non-constant argument to inline_native_newArray.) 3699 // In any case, the value of klass_node provides the desired array type. 3700 const TypeInt* length_type = _gvn.find_int_type(length); 3701 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3702 if (ary_type->isa_aryptr() && length_type != NULL) { 3703 // Try to get a better type than POS for the size 3704 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3705 } 3706 3707 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3708 3709 // Cast length on remaining path to be as narrow as possible 3710 if (map()->find_edge(length) >= 0) { 3711 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3712 if (ccast != length) { 3713 _gvn.set_type_bottom(ccast); 3714 record_for_igvn(ccast); 3715 replace_in_map(length, ccast); 3716 } 3717 } 3718 3719 return javaoop; 3720 } 3721 3722 // The following "Ideal_foo" functions are placed here because they recognize 3723 // the graph shapes created by the functions immediately above. 3724 3725 //---------------------------Ideal_allocation---------------------------------- 3726 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3727 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3728 if (ptr == NULL) { // reduce dumb test in callers 3729 return NULL; 3730 } 3731 3732 // Attempt to see through Shenandoah barriers. 3733 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr); 3734 3735 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3736 ptr = ptr->in(1); 3737 if (ptr == NULL) return NULL; 3738 } 3739 // Return NULL for allocations with several casts: 3740 // j.l.reflect.Array.newInstance(jobject, jint) 3741 // Object.clone() 3742 // to keep more precise type from last cast. 3743 if (ptr->is_Proj()) { 3744 Node* allo = ptr->in(0); 3745 if (allo != NULL && allo->is_Allocate()) { 3746 return allo->as_Allocate(); 3747 } 3748 } 3749 // Report failure to match. 3750 return NULL; 3751 } 3752 3753 // Fancy version which also strips off an offset (and reports it to caller). 3754 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3755 intptr_t& offset) { 3756 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3757 if (base == NULL) return NULL; 3758 return Ideal_allocation(base, phase); 3759 } 3760 3761 // Trace Initialize <- Proj[Parm] <- Allocate 3762 AllocateNode* InitializeNode::allocation() { 3763 Node* rawoop = in(InitializeNode::RawAddress); 3764 if (rawoop->is_Proj()) { 3765 Node* alloc = rawoop->in(0); 3766 if (alloc->is_Allocate()) { 3767 return alloc->as_Allocate(); 3768 } 3769 } 3770 return NULL; 3771 } 3772 3773 // Trace Allocate -> Proj[Parm] -> Initialize 3774 InitializeNode* AllocateNode::initialization() { 3775 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3776 if (rawoop == NULL) return NULL; 3777 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3778 Node* init = rawoop->fast_out(i); 3779 if (init->is_Initialize()) { 3780 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3781 return init->as_Initialize(); 3782 } 3783 } 3784 return NULL; 3785 } 3786 3787 //----------------------------- loop predicates --------------------------- 3788 3789 //------------------------------add_predicate_impl---------------------------- 3790 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3791 // Too many traps seen? 3792 if (too_many_traps(reason)) { 3793 #ifdef ASSERT 3794 if (TraceLoopPredicate) { 3795 int tc = C->trap_count(reason); 3796 tty->print("too many traps=%s tcount=%d in ", 3797 Deoptimization::trap_reason_name(reason), tc); 3798 method()->print(); // which method has too many predicate traps 3799 tty->cr(); 3800 } 3801 #endif 3802 // We cannot afford to take more traps here, 3803 // do not generate predicate. 3804 return; 3805 } 3806 3807 Node *cont = _gvn.intcon(1); 3808 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3809 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3810 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3811 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3812 C->add_predicate_opaq(opq); 3813 { 3814 PreserveJVMState pjvms(this); 3815 set_control(iffalse); 3816 inc_sp(nargs); 3817 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3818 } 3819 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3820 set_control(iftrue); 3821 } 3822 3823 //------------------------------add_predicate--------------------------------- 3824 void GraphKit::add_predicate(int nargs) { 3825 if (UseLoopPredicate) { 3826 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3827 } 3828 // loop's limit check predicate should be near the loop. 3829 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3830 } 3831 3832 //----------------------------- store barriers ---------------------------- 3833 #define __ ideal. 3834 3835 void GraphKit::sync_kit(IdealKit& ideal) { 3836 set_all_memory(__ merged_memory()); 3837 set_i_o(__ i_o()); 3838 set_control(__ ctrl()); 3839 } 3840 3841 void GraphKit::final_sync(IdealKit& ideal) { 3842 // Final sync IdealKit and graphKit. 3843 sync_kit(ideal); 3844 } 3845 3846 Node* GraphKit::byte_map_base_node() { 3847 // Get base of card map 3848 CardTableModRefBS* ct = 3849 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3850 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3851 if (ct->byte_map_base != NULL) { 3852 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3853 } else { 3854 return null(); 3855 } 3856 } 3857 3858 // vanilla/CMS post barrier 3859 // Insert a write-barrier store. This is to let generational GC work; we have 3860 // to flag all oop-stores before the next GC point. 3861 void GraphKit::write_barrier_post(Node* oop_store, 3862 Node* obj, 3863 Node* adr, 3864 uint adr_idx, 3865 Node* val, 3866 bool use_precise) { 3867 // No store check needed if we're storing a NULL or an old object 3868 // (latter case is probably a string constant). The concurrent 3869 // mark sweep garbage collector, however, needs to have all nonNull 3870 // oop updates flagged via card-marks. 3871 if (val != NULL && val->is_Con()) { 3872 // must be either an oop or NULL 3873 const Type* t = val->bottom_type(); 3874 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3875 // stores of null never (?) need barriers 3876 return; 3877 } 3878 3879 if (use_ReduceInitialCardMarks() 3880 && obj == just_allocated_object(control())) { 3881 // We can skip marks on a freshly-allocated object in Eden. 3882 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3883 // That routine informs GC to take appropriate compensating steps, 3884 // upon a slow-path allocation, so as to make this card-mark 3885 // elision safe. 3886 return; 3887 } 3888 3889 if (!use_precise) { 3890 // All card marks for a (non-array) instance are in one place: 3891 adr = obj; 3892 } 3893 // (Else it's an array (or unknown), and we want more precise card marks.) 3894 assert(adr != NULL, ""); 3895 3896 IdealKit ideal(this, true); 3897 3898 // Convert the pointer to an int prior to doing math on it 3899 Node* cast = __ CastPX(__ ctrl(), adr); 3900 3901 // Divide by card size 3902 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3903 "Only one we handle so far."); 3904 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3905 3906 // Combine card table base and card offset 3907 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3908 3909 // Get the alias_index for raw card-mark memory 3910 int adr_type = Compile::AliasIdxRaw; 3911 Node* zero = __ ConI(0); // Dirty card value 3912 BasicType bt = T_BYTE; 3913 3914 if (UseConcMarkSweepGC && UseCondCardMark) { 3915 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3916 __ sync_kit(this); 3917 } 3918 3919 if (UseCondCardMark) { 3920 // The classic GC reference write barrier is typically implemented 3921 // as a store into the global card mark table. Unfortunately 3922 // unconditional stores can result in false sharing and excessive 3923 // coherence traffic as well as false transactional aborts. 3924 // UseCondCardMark enables MP "polite" conditional card mark 3925 // stores. In theory we could relax the load from ctrl() to 3926 // no_ctrl, but that doesn't buy much latitude. 3927 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3928 __ if_then(card_val, BoolTest::ne, zero); 3929 } 3930 3931 // Smash zero into card 3932 if( !UseConcMarkSweepGC ) { 3933 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3934 } else { 3935 // Specialized path for CM store barrier 3936 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3937 } 3938 3939 if (UseCondCardMark) { 3940 __ end_if(); 3941 } 3942 3943 // Final sync IdealKit and GraphKit. 3944 final_sync(ideal); 3945 } 3946 /* 3947 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3948 * required by SATB to make sure all objects live at the start of the 3949 * marking are kept alive, all reference updates need to any previous 3950 * reference stored before writing. 3951 * 3952 * If the previous value is NULL there is no need to save the old value. 3953 * References that are NULL are filtered during runtime by the barrier 3954 * code to avoid unnecessary queuing. 3955 * 3956 * However in the case of newly allocated objects it might be possible to 3957 * prove that the reference about to be overwritten is NULL during compile 3958 * time and avoid adding the barrier code completely. 3959 * 3960 * The compiler needs to determine that the object in which a field is about 3961 * to be written is newly allocated, and that no prior store to the same field 3962 * has happened since the allocation. 3963 * 3964 * Returns true if the pre-barrier can be removed 3965 */ 3966 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3967 BasicType bt, uint adr_idx) { 3968 intptr_t offset = 0; 3969 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3970 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3971 3972 if (offset == Type::OffsetBot) { 3973 return false; // cannot unalias unless there are precise offsets 3974 } 3975 3976 if (alloc == NULL) { 3977 return false; // No allocation found 3978 } 3979 3980 intptr_t size_in_bytes = type2aelembytes(bt); 3981 3982 Node* mem = memory(adr_idx); // start searching here... 3983 3984 for (int cnt = 0; cnt < 50; cnt++) { 3985 3986 if (mem->is_Store()) { 3987 3988 Node* st_adr = mem->in(MemNode::Address); 3989 intptr_t st_offset = 0; 3990 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3991 3992 if (st_base == NULL) { 3993 break; // inscrutable pointer 3994 } 3995 3996 // Break we have found a store with same base and offset as ours so break 3997 if (st_base == base && st_offset == offset) { 3998 break; 3999 } 4000 4001 if (st_offset != offset && st_offset != Type::OffsetBot) { 4002 const int MAX_STORE = BytesPerLong; 4003 if (st_offset >= offset + size_in_bytes || 4004 st_offset <= offset - MAX_STORE || 4005 st_offset <= offset - mem->as_Store()->memory_size()) { 4006 // Success: The offsets are provably independent. 4007 // (You may ask, why not just test st_offset != offset and be done? 4008 // The answer is that stores of different sizes can co-exist 4009 // in the same sequence of RawMem effects. We sometimes initialize 4010 // a whole 'tile' of array elements with a single jint or jlong.) 4011 mem = mem->in(MemNode::Memory); 4012 continue; // advance through independent store memory 4013 } 4014 } 4015 4016 if (st_base != base 4017 && MemNode::detect_ptr_independence(base, alloc, st_base, 4018 AllocateNode::Ideal_allocation(st_base, phase), 4019 phase)) { 4020 // Success: The bases are provably independent. 4021 mem = mem->in(MemNode::Memory); 4022 continue; // advance through independent store memory 4023 } 4024 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4025 4026 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4027 AllocateNode* st_alloc = st_init->allocation(); 4028 4029 // Make sure that we are looking at the same allocation site. 4030 // The alloc variable is guaranteed to not be null here from earlier check. 4031 if (alloc == st_alloc) { 4032 // Check that the initialization is storing NULL so that no previous store 4033 // has been moved up and directly write a reference 4034 Node* captured_store = st_init->find_captured_store(offset, 4035 type2aelembytes(T_OBJECT), 4036 phase); 4037 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4038 return true; 4039 } 4040 } 4041 } 4042 4043 // Unless there is an explicit 'continue', we must bail out here, 4044 // because 'mem' is an inscrutable memory state (e.g., a call). 4045 break; 4046 } 4047 4048 return false; 4049 } 4050 4051 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) { 4052 if (UseShenandoahGC && ShenandoahSATBBarrier && adr != NULL) { 4053 Node* c = kit.control(); 4054 Node* call = c->in(1)->in(1)->in(1)->in(0); 4055 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected"); 4056 call->add_req(adr); 4057 } 4058 } 4059 4060 // G1 pre/post barriers 4061 void GraphKit::g1_write_barrier_pre(bool do_load, 4062 Node* obj, 4063 Node* adr, 4064 uint alias_idx, 4065 Node* val, 4066 const TypeOopPtr* val_type, 4067 Node* pre_val, 4068 BasicType bt) { 4069 4070 // Some sanity checks 4071 // Note: val is unused in this routine. 4072 4073 if (do_load) { 4074 // We need to generate the load of the previous value 4075 assert(obj != NULL, "must have a base"); 4076 assert(adr != NULL, "where are loading from?"); 4077 assert(pre_val == NULL, "loaded already?"); 4078 assert(val_type != NULL, "need a type"); 4079 4080 if (use_ReduceInitialCardMarks() 4081 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4082 return; 4083 } 4084 4085 } else { 4086 // In this case both val_type and alias_idx are unused. 4087 assert(pre_val != NULL, "must be loaded already"); 4088 // Nothing to be done if pre_val is null. 4089 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4090 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4091 } 4092 assert(bt == T_OBJECT, "or we shouldn't be here"); 4093 4094 IdealKit ideal(this, true); 4095 4096 Node* tls = __ thread(); // ThreadLocalStorage 4097 4098 Node* no_ctrl = NULL; 4099 Node* no_base = __ top(); 4100 Node* zero = __ ConI(0); 4101 Node* zeroX = __ ConX(0); 4102 4103 float likely = PROB_LIKELY(0.999); 4104 float unlikely = PROB_UNLIKELY(0.999); 4105 4106 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4107 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4108 4109 // Offsets into the thread 4110 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4111 SATBMarkQueue::byte_offset_of_active()); 4112 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4113 SATBMarkQueue::byte_offset_of_index()); 4114 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4115 SATBMarkQueue::byte_offset_of_buf()); 4116 4117 // Now the actual pointers into the thread 4118 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4119 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4120 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4121 4122 // Now some of the values 4123 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4124 4125 // if (!marking) 4126 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4127 BasicType index_bt = TypeX_X->basic_type(); 4128 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4129 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4130 4131 if (do_load) { 4132 // load original value 4133 // alias_idx correct?? 4134 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4135 } 4136 4137 // if (pre_val != NULL) 4138 __ if_then(pre_val, BoolTest::ne, null()); { 4139 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4140 4141 // is the queue for this thread full? 4142 __ if_then(index, BoolTest::ne, zeroX, likely); { 4143 4144 // decrement the index 4145 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4146 4147 // Now get the buffer location we will log the previous value into and store it 4148 Node *log_addr = __ AddP(no_base, buffer, next_index); 4149 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4150 // update the index 4151 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4152 4153 } __ else_(); { 4154 4155 // logging buffer is full, call the runtime 4156 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4157 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4158 } __ end_if(); // (!index) 4159 } __ end_if(); // (pre_val != NULL) 4160 } __ end_if(); // (!marking) 4161 4162 // Final sync IdealKit and GraphKit. 4163 final_sync(ideal); 4164 g1_write_barrier_pre_helper(*this, adr); 4165 } 4166 4167 Node* GraphKit::shenandoah_write_barrier_pre(bool do_load, 4168 Node* obj, 4169 Node* adr, 4170 uint alias_idx, 4171 Node* val, 4172 const TypeOopPtr* val_type, 4173 Node* pre_val, 4174 BasicType bt) { 4175 4176 // Some sanity checks 4177 // Note: val is unused in this routine. 4178 4179 if (!ShenandoahSATBBarrier) { 4180 if (val != NULL) { 4181 shenandoah_update_matrix(adr, val); 4182 } 4183 return val; 4184 } 4185 4186 if (val == NULL) { 4187 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 4188 return NULL; 4189 } 4190 4191 if (! ShenandoahReduceStoreValBarrier) { 4192 val = shenandoah_storeval_barrier(val); 4193 shenandoah_update_matrix(adr, val); 4194 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt); 4195 return val; 4196 } 4197 4198 if (do_load) { 4199 // We need to generate the load of the previous value 4200 assert(obj != NULL, "must have a base"); 4201 assert(adr != NULL, "where are loading from?"); 4202 assert(pre_val == NULL, "loaded already?"); 4203 assert(val_type != NULL, "need a type"); 4204 4205 if (use_ReduceInitialCardMarks() 4206 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4207 return shenandoah_storeval_barrier(val); 4208 } 4209 4210 } else { 4211 // In this case both val_type and alias_idx are unused. 4212 assert(pre_val != NULL, "must be loaded already"); 4213 // Nothing to be done if pre_val is null. 4214 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return val; 4215 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4216 } 4217 assert(bt == T_OBJECT, "or we shouldn't be here"); 4218 4219 IdealKit ideal(this, true, true); 4220 IdealVariable ival(ideal); 4221 __ declarations_done(); 4222 __ set(ival, val); 4223 Node* tls = __ thread(); // ThreadLocalStorage 4224 4225 Node* no_ctrl = NULL; 4226 Node* no_base = __ top(); 4227 Node* zero = __ ConI(0); 4228 Node* zeroX = __ ConX(0); 4229 4230 float likely = PROB_LIKELY(0.999); 4231 float unlikely = PROB_UNLIKELY(0.999); 4232 4233 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4234 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4235 4236 // Offsets into the thread 4237 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4238 SATBMarkQueue::byte_offset_of_active()); 4239 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4240 SATBMarkQueue::byte_offset_of_index()); 4241 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4242 SATBMarkQueue::byte_offset_of_buf()); 4243 4244 // Now the actual pointers into the thread 4245 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4246 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4247 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4248 4249 // Now some of the values 4250 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4251 4252 // if (!marking) 4253 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4254 4255 Node* storeval = ideal.value(ival); 4256 sync_kit(ideal); 4257 storeval = shenandoah_storeval_barrier(storeval); 4258 __ sync_kit(this); 4259 __ set(ival, storeval); 4260 4261 BasicType index_bt = TypeX_X->basic_type(); 4262 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4263 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4264 4265 if (do_load) { 4266 // load original value 4267 // alias_idx correct?? 4268 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4269 } 4270 4271 // if (pre_val != NULL) 4272 __ if_then(pre_val, BoolTest::ne, null()); { 4273 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4274 4275 // is the queue for this thread full? 4276 __ if_then(index, BoolTest::ne, zeroX, likely); { 4277 4278 // decrement the index 4279 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4280 4281 // Now get the buffer location we will log the previous value into and store it 4282 Node *log_addr = __ AddP(no_base, buffer, next_index); 4283 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4284 // update the index 4285 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4286 4287 } __ else_(); { 4288 4289 // logging buffer is full, call the runtime 4290 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4291 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4292 } __ end_if(); // (!index) 4293 } __ end_if(); // (pre_val != NULL) 4294 } __ end_if(); // (!marking) 4295 Node* new_val = __ value(ival); 4296 // IdealKit generates a Phi with very conservative type, and even 4297 // turns array types into TypeInstPtr (see type.cpp, _const_basic_type[T_ARRAY]). 4298 // We're forcing the result to be the original type. 4299 if (new_val != val) { 4300 const Type* t = _gvn.type(val); 4301 if (new_val->isa_Type()) { 4302 new_val->as_Type()->set_type(t); 4303 } 4304 _gvn.set_type(new_val, t); 4305 } 4306 val = new_val; 4307 __ dead(ival); 4308 // Final sync IdealKit and GraphKit. 4309 final_sync(ideal); 4310 g1_write_barrier_pre_helper(*this, adr); 4311 return val; 4312 } 4313 4314 void GraphKit::shenandoah_update_matrix(Node* adr, Node* val) { 4315 if (!UseShenandoahMatrix) { 4316 return; 4317 } 4318 4319 assert(val != NULL, "checked before"); 4320 if (adr == NULL) { 4321 return; // Nothing to do 4322 } 4323 assert(adr != NULL, "must not happen"); 4324 if (val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) { 4325 // Nothing to do. 4326 return; 4327 } 4328 4329 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix(); 4330 4331 enum { _set_path = 1, _already_set_path, _val_null_path, PATH_LIMIT }; 4332 RegionNode* region = new RegionNode(PATH_LIMIT); 4333 Node* prev_mem = memory(Compile::AliasIdxRaw); 4334 Node* memphi = PhiNode::make(region, prev_mem, Type::MEMORY, TypeRawPtr::BOTTOM); 4335 Node* null_ctrl = top(); 4336 Node* not_null_val = null_check_oop(val, &null_ctrl); 4337 4338 // Null path: nothing to do. 4339 region->init_req(_val_null_path, null_ctrl); 4340 memphi->init_req(_val_null_path, prev_mem); 4341 4342 // Not null path. Update the matrix. 4343 4344 // This uses a fast calculation for the matrix address. For a description, 4345 // see src/share/vm/gc/shenandoah/shenandoahConnectionMatrix.inline.hpp, 4346 // ShenandoahConnectionMatrix::compute_address(const void* from, const void* to). 4347 address heap_base = ShenandoahHeap::heap()->base(); 4348 jint stride = matrix->stride_jint(); 4349 jint rs = ShenandoahHeapRegion::region_size_bytes_shift_jint(); 4350 4351 guarantee(stride < ShenandoahHeapRegion::region_size_bytes_jint(), "sanity"); 4352 guarantee(is_aligned(heap_base, ShenandoahHeapRegion::region_size_bytes()), "sanity"); 4353 4354 Node* magic_con = MakeConX((jlong) matrix->matrix_addr() - ((jlong) heap_base >> rs) * (stride + 1)); 4355 4356 // Compute addr part 4357 Node* adr_idx = _gvn.transform(new CastP2XNode(control(), adr)); 4358 adr_idx = _gvn.transform(new URShiftXNode(adr_idx, intcon(rs))); 4359 4360 // Compute new_val part 4361 Node* val_idx = _gvn.transform(new CastP2XNode(control(), not_null_val)); 4362 val_idx = _gvn.transform(new URShiftXNode(val_idx, intcon(rs))); 4363 val_idx = _gvn.transform(new MulXNode(val_idx, MakeConX(stride))); 4364 4365 // Add everything up 4366 adr_idx = _gvn.transform(new AddXNode(adr_idx, val_idx)); 4367 adr_idx = _gvn.transform(new CastX2PNode(adr_idx)); 4368 Node* matrix_adr = _gvn.transform(new AddPNode(top(), adr_idx, magic_con)); 4369 4370 // Load current value 4371 const TypePtr* adr_type = TypeRawPtr::BOTTOM; 4372 Node* current = _gvn.transform(LoadNode::make(_gvn, control(), memory(Compile::AliasIdxRaw), 4373 matrix_adr, adr_type, TypeInt::INT, T_BYTE, MemNode::unordered)); 4374 4375 // Check if already set 4376 Node* cmp_set = _gvn.transform(new CmpINode(current, intcon(0))); 4377 Node* cmp_set_bool = _gvn.transform(new BoolNode(cmp_set, BoolTest::eq)); 4378 IfNode* cmp_iff = create_and_map_if(control(), cmp_set_bool, PROB_MIN, COUNT_UNKNOWN); 4379 4380 Node* if_not_set = _gvn.transform(new IfTrueNode(cmp_iff)); 4381 Node* if_set = _gvn.transform(new IfFalseNode(cmp_iff)); 4382 4383 // Already set, exit 4384 set_control(if_set); 4385 region->init_req(_already_set_path, control()); 4386 memphi->init_req(_already_set_path, prev_mem); 4387 4388 // Not set: do the store, and finish up 4389 set_control(if_not_set); 4390 Node* store = _gvn.transform(StoreNode::make(_gvn, control(), memory(Compile::AliasIdxRaw), 4391 matrix_adr, adr_type, intcon(1), T_BYTE, MemNode::unordered)); 4392 region->init_req(_set_path, control()); 4393 memphi->init_req(_set_path, store); 4394 4395 // Merge control flows and memory. 4396 set_control(_gvn.transform(region)); 4397 record_for_igvn(region); 4398 set_memory(_gvn.transform(memphi), Compile::AliasIdxRaw); 4399 } 4400 4401 /* 4402 * G1 similar to any GC with a Young Generation requires a way to keep track of 4403 * references from Old Generation to Young Generation to make sure all live 4404 * objects are found. G1 also requires to keep track of object references 4405 * between different regions to enable evacuation of old regions, which is done 4406 * as part of mixed collections. References are tracked in remembered sets and 4407 * is continuously updated as reference are written to with the help of the 4408 * post-barrier. 4409 * 4410 * To reduce the number of updates to the remembered set the post-barrier 4411 * filters updates to fields in objects located in the Young Generation, 4412 * the same region as the reference, when the NULL is being written or 4413 * if the card is already marked as dirty by an earlier write. 4414 * 4415 * Under certain circumstances it is possible to avoid generating the 4416 * post-barrier completely if it is possible during compile time to prove 4417 * the object is newly allocated and that no safepoint exists between the 4418 * allocation and the store. 4419 * 4420 * In the case of slow allocation the allocation code must handle the barrier 4421 * as part of the allocation in the case the allocated object is not located 4422 * in the nursery, this would happen for humongous objects. This is similar to 4423 * how CMS is required to handle this case, see the comments for the method 4424 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4425 * A deferred card mark is required for these objects and handled in the above 4426 * mentioned methods. 4427 * 4428 * Returns true if the post barrier can be removed 4429 */ 4430 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4431 Node* adr) { 4432 intptr_t offset = 0; 4433 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4434 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4435 4436 if (offset == Type::OffsetBot) { 4437 return false; // cannot unalias unless there are precise offsets 4438 } 4439 4440 if (alloc == NULL) { 4441 return false; // No allocation found 4442 } 4443 4444 // Start search from Store node 4445 Node* mem = store->in(MemNode::Control); 4446 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4447 4448 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4449 AllocateNode* st_alloc = st_init->allocation(); 4450 4451 // Make sure we are looking at the same allocation 4452 if (alloc == st_alloc) { 4453 return true; 4454 } 4455 } 4456 4457 return false; 4458 } 4459 4460 // 4461 // Update the card table and add card address to the queue 4462 // 4463 void GraphKit::g1_mark_card(IdealKit& ideal, 4464 Node* card_adr, 4465 Node* oop_store, 4466 uint oop_alias_idx, 4467 Node* index, 4468 Node* index_adr, 4469 Node* buffer, 4470 const TypeFunc* tf) { 4471 4472 Node* zero = __ ConI(0); 4473 Node* zeroX = __ ConX(0); 4474 Node* no_base = __ top(); 4475 BasicType card_bt = T_BYTE; 4476 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4477 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4478 4479 // Now do the queue work 4480 __ if_then(index, BoolTest::ne, zeroX); { 4481 4482 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4483 Node* log_addr = __ AddP(no_base, buffer, next_index); 4484 4485 // Order, see storeCM. 4486 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4487 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4488 4489 } __ else_(); { 4490 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4491 } __ end_if(); 4492 4493 } 4494 4495 void GraphKit::g1_write_barrier_post(Node* oop_store, 4496 Node* obj, 4497 Node* adr, 4498 uint alias_idx, 4499 Node* val, 4500 BasicType bt, 4501 bool use_precise) { 4502 // If we are writing a NULL then we need no post barrier 4503 4504 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4505 // Must be NULL 4506 const Type* t = val->bottom_type(); 4507 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4508 // No post barrier if writing NULLx 4509 return; 4510 } 4511 4512 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4513 // We can skip marks on a freshly-allocated object in Eden. 4514 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4515 // That routine informs GC to take appropriate compensating steps, 4516 // upon a slow-path allocation, so as to make this card-mark 4517 // elision safe. 4518 return; 4519 } 4520 4521 if (use_ReduceInitialCardMarks() 4522 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4523 return; 4524 } 4525 4526 if (!use_precise) { 4527 // All card marks for a (non-array) instance are in one place: 4528 adr = obj; 4529 } 4530 // (Else it's an array (or unknown), and we want more precise card marks.) 4531 assert(adr != NULL, ""); 4532 4533 IdealKit ideal(this, true); 4534 4535 Node* tls = __ thread(); // ThreadLocalStorage 4536 4537 Node* no_base = __ top(); 4538 float likely = PROB_LIKELY(0.999); 4539 float unlikely = PROB_UNLIKELY(0.999); 4540 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4541 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4542 Node* zeroX = __ ConX(0); 4543 4544 // Get the alias_index for raw card-mark memory 4545 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4546 4547 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4548 4549 // Offsets into the thread 4550 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4551 DirtyCardQueue::byte_offset_of_index()); 4552 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4553 DirtyCardQueue::byte_offset_of_buf()); 4554 4555 // Pointers into the thread 4556 4557 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4558 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4559 4560 // Now some values 4561 // Use ctrl to avoid hoisting these values past a safepoint, which could 4562 // potentially reset these fields in the JavaThread. 4563 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4564 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4565 4566 // Convert the store obj pointer to an int prior to doing math on it 4567 // Must use ctrl to prevent "integerized oop" existing across safepoint 4568 Node* cast = __ CastPX(__ ctrl(), adr); 4569 4570 // Divide pointer by card size 4571 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4572 4573 // Combine card table base and card offset 4574 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4575 4576 // If we know the value being stored does it cross regions? 4577 4578 if (val != NULL) { 4579 // Does the store cause us to cross regions? 4580 4581 // Should be able to do an unsigned compare of region_size instead of 4582 // and extra shift. Do we have an unsigned compare?? 4583 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4584 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4585 4586 // if (xor_res == 0) same region so skip 4587 __ if_then(xor_res, BoolTest::ne, zeroX); { 4588 4589 // No barrier if we are storing a NULL 4590 __ if_then(val, BoolTest::ne, null(), unlikely); { 4591 4592 // Ok must mark the card if not already dirty 4593 4594 // load the original value of the card 4595 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4596 4597 __ if_then(card_val, BoolTest::ne, young_card); { 4598 sync_kit(ideal); 4599 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4600 insert_mem_bar(Op_MemBarVolatile, oop_store); 4601 __ sync_kit(this); 4602 4603 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4604 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4605 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4606 } __ end_if(); 4607 } __ end_if(); 4608 } __ end_if(); 4609 } __ end_if(); 4610 } else { 4611 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4612 // We don't need a barrier here if the destination is a newly allocated object 4613 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4614 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4615 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4616 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4617 __ if_then(card_val, BoolTest::ne, young_card); { 4618 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4619 } __ end_if(); 4620 } 4621 4622 // Final sync IdealKit and GraphKit. 4623 final_sync(ideal); 4624 } 4625 #undef __ 4626 4627 4628 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4629 Node* len = load_array_length(load_String_value(ctrl, str)); 4630 Node* coder = load_String_coder(ctrl, str); 4631 // Divide length by 2 if coder is UTF16 4632 return _gvn.transform(new RShiftINode(len, coder)); 4633 } 4634 4635 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4636 int value_offset = java_lang_String::value_offset_in_bytes(); 4637 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4638 false, NULL, 0); 4639 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4640 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4641 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4642 ciTypeArrayKlass::make(T_BYTE), true, 0); 4643 int value_field_idx = C->get_alias_index(value_field_type); 4644 4645 if (! ShenandoahOptimizeFinals) { 4646 str = shenandoah_read_barrier(str); 4647 } 4648 4649 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4650 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4651 // String.value field is known to be @Stable. 4652 if (UseImplicitStableValues) { 4653 load = cast_array_to_stable(load, value_type); 4654 } 4655 return load; 4656 } 4657 4658 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4659 if (!CompactStrings) { 4660 return intcon(java_lang_String::CODER_UTF16); 4661 } 4662 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4663 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4664 false, NULL, 0); 4665 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4666 int coder_field_idx = C->get_alias_index(coder_field_type); 4667 4668 if (! ShenandoahOptimizeFinals) { 4669 str = shenandoah_read_barrier(str); 4670 } 4671 4672 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4673 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4674 } 4675 4676 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4677 int value_offset = java_lang_String::value_offset_in_bytes(); 4678 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4679 false, NULL, 0); 4680 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4681 4682 str = shenandoah_write_barrier(str); 4683 4684 store_oop_to_object(control(), str, basic_plus_adr(str, value_offset), value_field_type, 4685 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4686 } 4687 4688 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4689 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4690 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4691 false, NULL, 0); 4692 4693 str = shenandoah_write_barrier(str); 4694 4695 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4696 int coder_field_idx = C->get_alias_index(coder_field_type); 4697 store_to_memory(control(), basic_plus_adr(str, coder_offset), 4698 value, T_BYTE, coder_field_idx, MemNode::unordered); 4699 } 4700 4701 // Capture src and dst memory state with a MergeMemNode 4702 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4703 if (src_type == dst_type) { 4704 // Types are equal, we don't need a MergeMemNode 4705 return memory(src_type); 4706 } 4707 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4708 record_for_igvn(merge); // fold it up later, if possible 4709 int src_idx = C->get_alias_index(src_type); 4710 int dst_idx = C->get_alias_index(dst_type); 4711 merge->set_memory_at(src_idx, memory(src_idx)); 4712 merge->set_memory_at(dst_idx, memory(dst_idx)); 4713 return merge; 4714 } 4715 4716 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4717 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4718 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4719 // If input and output memory types differ, capture both states to preserve 4720 // the dependency between preceding and subsequent loads/stores. 4721 // For example, the following program: 4722 // StoreB 4723 // compress_string 4724 // LoadB 4725 // has this memory graph (use->def): 4726 // LoadB -> compress_string -> CharMem 4727 // ... -> StoreB -> ByteMem 4728 // The intrinsic hides the dependency between LoadB and StoreB, causing 4729 // the load to read from memory not containing the result of the StoreB. 4730 // The correct memory graph should look like this: 4731 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4732 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4733 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4734 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4735 set_memory(res_mem, TypeAryPtr::BYTES); 4736 return str; 4737 } 4738 4739 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4740 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4741 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4742 // Capture src and dst memory (see comment in 'compress_string'). 4743 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4744 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4745 set_memory(_gvn.transform(str), dst_type); 4746 } 4747 4748 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4749 4750 src = shenandoah_read_barrier(src); 4751 dst = shenandoah_write_barrier(dst); 4752 4753 /** 4754 * int i_char = start; 4755 * for (int i_byte = 0; i_byte < count; i_byte++) { 4756 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4757 * } 4758 */ 4759 add_predicate(); 4760 RegionNode* head = new RegionNode(3); 4761 head->init_req(1, control()); 4762 gvn().set_type(head, Type::CONTROL); 4763 record_for_igvn(head); 4764 4765 Node* i_byte = new PhiNode(head, TypeInt::INT); 4766 i_byte->init_req(1, intcon(0)); 4767 gvn().set_type(i_byte, TypeInt::INT); 4768 record_for_igvn(i_byte); 4769 4770 Node* i_char = new PhiNode(head, TypeInt::INT); 4771 i_char->init_req(1, start); 4772 gvn().set_type(i_char, TypeInt::INT); 4773 record_for_igvn(i_char); 4774 4775 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4776 gvn().set_type(mem, Type::MEMORY); 4777 record_for_igvn(mem); 4778 set_control(head); 4779 set_memory(mem, TypeAryPtr::BYTES); 4780 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4781 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4782 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4783 false, false, true /* mismatched */); 4784 4785 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4786 head->init_req(2, IfTrue(iff)); 4787 mem->init_req(2, st); 4788 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4789 i_char->init_req(2, AddI(i_char, intcon(2))); 4790 4791 set_control(IfFalse(iff)); 4792 set_memory(st, TypeAryPtr::BYTES); 4793 } 4794 4795 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4796 if (!field->is_constant()) { 4797 return NULL; // Field not marked as constant. 4798 } 4799 ciInstance* holder = NULL; 4800 if (!field->is_static()) { 4801 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4802 if (const_oop != NULL && const_oop->is_instance()) { 4803 holder = const_oop->as_instance(); 4804 } 4805 } 4806 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4807 /*is_unsigned_load=*/false); 4808 if (con_type != NULL) { 4809 return makecon(con_type); 4810 } 4811 return NULL; 4812 } 4813 4814 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4815 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4816 // assumption of CCP analysis. 4817 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4818 } 4819 4820 Node* GraphKit::shenandoah_read_barrier(Node* obj) { 4821 if (UseShenandoahGC && ShenandoahReadBarrier) { 4822 return shenandoah_read_barrier_impl(obj, false, true, true); 4823 } else { 4824 return obj; 4825 } 4826 } 4827 4828 Node* GraphKit::shenandoah_storeval_barrier(Node* obj) { 4829 if (UseShenandoahGC && ShenandoahStoreValBarrier) { 4830 return shenandoah_read_barrier_impl(obj, true, false, false); 4831 } else { 4832 return obj; 4833 } 4834 } 4835 4836 Node* GraphKit::shenandoah_read_barrier_acmp(Node* obj) { 4837 return shenandoah_read_barrier_impl(obj, true, true, false); 4838 } 4839 4840 Node* GraphKit::shenandoah_read_barrier_impl(Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) { 4841 4842 const Type* obj_type = obj->bottom_type(); 4843 if (obj_type->higher_equal(TypePtr::NULL_PTR)) { 4844 return obj; 4845 } 4846 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 4847 Node* mem = use_mem ? memory(adr_type) : immutable_memory(); 4848 4849 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, mem, allow_fromspace)) { 4850 // We know it is null, no barrier needed. 4851 return obj; 4852 } 4853 4854 4855 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4856 4857 // We don't know if it's null or not. Need null-check. 4858 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4859 RegionNode* region = new RegionNode(PATH_LIMIT); 4860 Node* phi = new PhiNode(region, obj_type); 4861 Node* null_ctrl = top(); 4862 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4863 4864 region->init_req(_null_path, null_ctrl); 4865 phi ->init_req(_null_path, zerocon(T_OBJECT)); 4866 4867 Node* ctrl = use_ctrl ? control() : NULL; 4868 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace); 4869 Node* n = _gvn.transform(rb); 4870 4871 region->init_req(_not_null_path, control()); 4872 phi ->init_req(_not_null_path, n); 4873 4874 set_control(_gvn.transform(region)); 4875 record_for_igvn(region); 4876 return _gvn.transform(phi); 4877 4878 } else { 4879 // We know it is not null. Simple barrier is sufficient. 4880 Node* ctrl = use_ctrl ? control() : NULL; 4881 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace); 4882 Node* n = _gvn.transform(rb); 4883 record_for_igvn(n); 4884 return n; 4885 } 4886 } 4887 4888 static Node* shenandoah_write_barrier_helper(GraphKit& kit, Node* obj, const TypePtr* adr_type) { 4889 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit.C, kit.control(), kit.memory(adr_type), obj); 4890 Node* n = kit.gvn().transform(wb); 4891 if (n == wb) { // New barrier needs memory projection. 4892 Node* proj = kit.gvn().transform(new ShenandoahWBMemProjNode(n)); 4893 kit.set_memory(proj, adr_type); 4894 } 4895 4896 return n; 4897 } 4898 4899 Node* GraphKit::shenandoah_write_barrier(Node* obj) { 4900 4901 if (UseShenandoahGC && ShenandoahWriteBarrier) { 4902 4903 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, NULL, true)) { 4904 return obj; 4905 } 4906 const Type* obj_type = obj->bottom_type(); 4907 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type); 4908 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4909 // We don't know if it's null or not. Need null-check. 4910 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4911 RegionNode* region = new RegionNode(PATH_LIMIT); 4912 Node* phi = new PhiNode(region, obj_type); 4913 Node* memphi = PhiNode::make(region, memory(adr_type), Type::MEMORY, C->alias_type(adr_type)->adr_type()); 4914 4915 Node* prev_mem = memory(adr_type); 4916 Node* null_ctrl = top(); 4917 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4918 4919 region->init_req(_null_path, null_ctrl); 4920 phi ->init_req(_null_path, zerocon(T_OBJECT)); 4921 memphi->init_req(_null_path, prev_mem); 4922 4923 Node* n = shenandoah_write_barrier_helper(*this, not_null_obj, adr_type); 4924 4925 region->init_req(_not_null_path, control()); 4926 phi ->init_req(_not_null_path, n); 4927 memphi->init_req(_not_null_path, memory(adr_type)); 4928 4929 set_control(_gvn.transform(region)); 4930 record_for_igvn(region); 4931 set_memory(_gvn.transform(memphi), adr_type); 4932 4933 Node* res_val = _gvn.transform(phi); 4934 // replace_in_map(obj, res_val); 4935 return res_val; 4936 } else { 4937 // We know it is not null. Simple barrier is sufficient. 4938 Node* n = shenandoah_write_barrier_helper(*this, obj, adr_type); 4939 // replace_in_map(obj, n); 4940 record_for_igvn(n); 4941 return n; 4942 } 4943 4944 } else { 4945 return obj; 4946 } 4947 }