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