1 /* 2 * Copyright (c) 2001, 2015, 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_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "gc_implementation/shenandoah/shenandoahHeap.hpp" 31 #include "memory/barrierSet.hpp" 32 #include "memory/cardTableModRefBS.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/graphKit.hpp" 35 #include "opto/idealKit.hpp" 36 #include "opto/locknode.hpp" 37 #include "opto/machnode.hpp" 38 #include "opto/parse.hpp" 39 #include "opto/rootnode.hpp" 40 #include "opto/runtime.hpp" 41 #include "runtime/deoptimization.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 44 #if INCLUDE_ALL_GCS 45 #include "gc_implementation/shenandoah/shenandoahBarrierSetC2.hpp" 46 #include "gc_implementation/shenandoah/shenandoahSupport.hpp" 47 #endif 48 49 //----------------------------GraphKit----------------------------------------- 50 // Main utility constructor. 51 GraphKit::GraphKit(JVMState* jvms) 52 : Phase(Phase::Parser), 53 _env(C->env()), 54 _gvn(*C->initial_gvn()) 55 { 56 _exceptions = jvms->map()->next_exception(); 57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 58 set_jvms(jvms); 59 } 60 61 // Private constructor for parser. 62 GraphKit::GraphKit() 63 : Phase(Phase::Parser), 64 _env(C->env()), 65 _gvn(*C->initial_gvn()) 66 { 67 _exceptions = NULL; 68 set_map(NULL); 69 debug_only(_sp = -99); 70 debug_only(set_bci(-99)); 71 } 72 73 74 75 //---------------------------clean_stack--------------------------------------- 76 // Clear away rubbish from the stack area of the JVM state. 77 // This destroys any arguments that may be waiting on the stack. 78 void GraphKit::clean_stack(int from_sp) { 79 SafePointNode* map = this->map(); 80 JVMState* jvms = this->jvms(); 81 int stk_size = jvms->stk_size(); 82 int stkoff = jvms->stkoff(); 83 Node* top = this->top(); 84 for (int i = from_sp; i < stk_size; i++) { 85 if (map->in(stkoff + i) != top) { 86 map->set_req(stkoff + i, top); 87 } 88 } 89 } 90 91 92 //--------------------------------sync_jvms----------------------------------- 93 // Make sure our current jvms agrees with our parse state. 94 JVMState* GraphKit::sync_jvms() const { 95 JVMState* jvms = this->jvms(); 96 jvms->set_bci(bci()); // Record the new bci in the JVMState 97 jvms->set_sp(sp()); // Record the new sp in the JVMState 98 assert(jvms_in_sync(), "jvms is now in sync"); 99 return jvms; 100 } 101 102 //--------------------------------sync_jvms_for_reexecute--------------------- 103 // Make sure our current jvms agrees with our parse state. This version 104 // uses the reexecute_sp for reexecuting bytecodes. 105 JVMState* GraphKit::sync_jvms_for_reexecute() { 106 JVMState* jvms = this->jvms(); 107 jvms->set_bci(bci()); // Record the new bci in the JVMState 108 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 109 return jvms; 110 } 111 112 #ifdef ASSERT 113 bool GraphKit::jvms_in_sync() const { 114 Parse* parse = is_Parse(); 115 if (parse == NULL) { 116 if (bci() != jvms()->bci()) return false; 117 if (sp() != (int)jvms()->sp()) return false; 118 return true; 119 } 120 if (jvms()->method() != parse->method()) return false; 121 if (jvms()->bci() != parse->bci()) return false; 122 int jvms_sp = jvms()->sp(); 123 if (jvms_sp != parse->sp()) return false; 124 int jvms_depth = jvms()->depth(); 125 if (jvms_depth != parse->depth()) return false; 126 return true; 127 } 128 129 // Local helper checks for special internal merge points 130 // used to accumulate and merge exception states. 131 // They are marked by the region's in(0) edge being the map itself. 132 // Such merge points must never "escape" into the parser at large, 133 // until they have been handed to gvn.transform. 134 static bool is_hidden_merge(Node* reg) { 135 if (reg == NULL) return false; 136 if (reg->is_Phi()) { 137 reg = reg->in(0); 138 if (reg == NULL) return false; 139 } 140 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 141 } 142 143 void GraphKit::verify_map() const { 144 if (map() == NULL) return; // null map is OK 145 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 146 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 147 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 148 } 149 150 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 151 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 152 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 153 } 154 #endif 155 156 //---------------------------stop_and_kill_map--------------------------------- 157 // Set _map to NULL, signalling a stop to further bytecode execution. 158 // First smash the current map's control to a constant, to mark it dead. 159 void GraphKit::stop_and_kill_map() { 160 SafePointNode* dead_map = stop(); 161 if (dead_map != NULL) { 162 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 163 assert(dead_map->is_killed(), "must be so marked"); 164 } 165 } 166 167 168 //--------------------------------stopped-------------------------------------- 169 // Tell if _map is NULL, or control is top. 170 bool GraphKit::stopped() { 171 if (map() == NULL) return true; 172 else if (control() == top()) return true; 173 else return false; 174 } 175 176 177 //-----------------------------has_ex_handler---------------------------------- 178 // Tell if this method or any caller method has exception handlers. 179 bool GraphKit::has_ex_handler() { 180 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 181 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 182 return true; 183 } 184 } 185 return false; 186 } 187 188 //------------------------------save_ex_oop------------------------------------ 189 // Save an exception without blowing stack contents or other JVM state. 190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 191 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 192 ex_map->add_req(ex_oop); 193 debug_only(verify_exception_state(ex_map)); 194 } 195 196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 197 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 198 Node* ex_oop = ex_map->in(ex_map->req()-1); 199 if (clear_it) ex_map->del_req(ex_map->req()-1); 200 return ex_oop; 201 } 202 203 //-----------------------------saved_ex_oop------------------------------------ 204 // Recover a saved exception from its map. 205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 206 return common_saved_ex_oop(ex_map, false); 207 } 208 209 //--------------------------clear_saved_ex_oop--------------------------------- 210 // Erase a previously saved exception from its map. 211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 212 return common_saved_ex_oop(ex_map, true); 213 } 214 215 #ifdef ASSERT 216 //---------------------------has_saved_ex_oop---------------------------------- 217 // Erase a previously saved exception from its map. 218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 219 return ex_map->req() == ex_map->jvms()->endoff()+1; 220 } 221 #endif 222 223 //-------------------------make_exception_state-------------------------------- 224 // Turn the current JVM state into an exception state, appending the ex_oop. 225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 226 sync_jvms(); 227 SafePointNode* ex_map = stop(); // do not manipulate this map any more 228 set_saved_ex_oop(ex_map, ex_oop); 229 return ex_map; 230 } 231 232 233 //--------------------------add_exception_state-------------------------------- 234 // Add an exception to my list of exceptions. 235 void GraphKit::add_exception_state(SafePointNode* ex_map) { 236 if (ex_map == NULL || ex_map->control() == top()) { 237 return; 238 } 239 #ifdef ASSERT 240 verify_exception_state(ex_map); 241 if (has_exceptions()) { 242 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 243 } 244 #endif 245 246 // If there is already an exception of exactly this type, merge with it. 247 // In particular, null-checks and other low-level exceptions common up here. 248 Node* ex_oop = saved_ex_oop(ex_map); 249 const Type* ex_type = _gvn.type(ex_oop); 250 if (ex_oop == top()) { 251 // No action needed. 252 return; 253 } 254 assert(ex_type->isa_instptr(), "exception must be an instance"); 255 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 256 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 257 // We check sp also because call bytecodes can generate exceptions 258 // both before and after arguments are popped! 259 if (ex_type2 == ex_type 260 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 261 combine_exception_states(ex_map, e2); 262 return; 263 } 264 } 265 266 // No pre-existing exception of the same type. Chain it on the list. 267 push_exception_state(ex_map); 268 } 269 270 //-----------------------add_exception_states_from----------------------------- 271 void GraphKit::add_exception_states_from(JVMState* jvms) { 272 SafePointNode* ex_map = jvms->map()->next_exception(); 273 if (ex_map != NULL) { 274 jvms->map()->set_next_exception(NULL); 275 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 276 next_map = ex_map->next_exception(); 277 ex_map->set_next_exception(NULL); 278 add_exception_state(ex_map); 279 } 280 } 281 } 282 283 //-----------------------transfer_exceptions_into_jvms------------------------- 284 JVMState* GraphKit::transfer_exceptions_into_jvms() { 285 if (map() == NULL) { 286 // We need a JVMS to carry the exceptions, but the map has gone away. 287 // Create a scratch JVMS, cloned from any of the exception states... 288 if (has_exceptions()) { 289 _map = _exceptions; 290 _map = clone_map(); 291 _map->set_next_exception(NULL); 292 clear_saved_ex_oop(_map); 293 debug_only(verify_map()); 294 } else { 295 // ...or created from scratch 296 JVMState* jvms = new (C) JVMState(_method, NULL); 297 jvms->set_bci(_bci); 298 jvms->set_sp(_sp); 299 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); 300 set_jvms(jvms); 301 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 302 set_all_memory(top()); 303 while (map()->req() < jvms->endoff()) map()->add_req(top()); 304 } 305 // (This is a kludge, in case you didn't notice.) 306 set_control(top()); 307 } 308 JVMState* jvms = sync_jvms(); 309 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 310 jvms->map()->set_next_exception(_exceptions); 311 _exceptions = NULL; // done with this set of exceptions 312 return jvms; 313 } 314 315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 316 assert(is_hidden_merge(dstphi), "must be a special merge node"); 317 assert(is_hidden_merge(srcphi), "must be a special merge node"); 318 uint limit = srcphi->req(); 319 for (uint i = PhiNode::Input; i < limit; i++) { 320 dstphi->add_req(srcphi->in(i)); 321 } 322 } 323 static inline void add_one_req(Node* dstphi, Node* src) { 324 assert(is_hidden_merge(dstphi), "must be a special merge node"); 325 assert(!is_hidden_merge(src), "must not be a special merge node"); 326 dstphi->add_req(src); 327 } 328 329 //-----------------------combine_exception_states------------------------------ 330 // This helper function combines exception states by building phis on a 331 // specially marked state-merging region. These regions and phis are 332 // untransformed, and can build up gradually. The region is marked by 333 // having a control input of its exception map, rather than NULL. Such 334 // regions do not appear except in this function, and in use_exception_state. 335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 336 if (failing()) return; // dying anyway... 337 JVMState* ex_jvms = ex_map->_jvms; 338 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 339 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 340 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 341 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 342 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 343 assert(ex_map->req() == phi_map->req(), "matching maps"); 344 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 345 Node* hidden_merge_mark = root(); 346 Node* region = phi_map->control(); 347 MergeMemNode* phi_mem = phi_map->merged_memory(); 348 MergeMemNode* ex_mem = ex_map->merged_memory(); 349 if (region->in(0) != hidden_merge_mark) { 350 // The control input is not (yet) a specially-marked region in phi_map. 351 // Make it so, and build some phis. 352 region = new (C) RegionNode(2); 353 _gvn.set_type(region, Type::CONTROL); 354 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 355 region->init_req(1, phi_map->control()); 356 phi_map->set_control(region); 357 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 358 record_for_igvn(io_phi); 359 _gvn.set_type(io_phi, Type::ABIO); 360 phi_map->set_i_o(io_phi); 361 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 362 Node* m = mms.memory(); 363 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 364 record_for_igvn(m_phi); 365 _gvn.set_type(m_phi, Type::MEMORY); 366 mms.set_memory(m_phi); 367 } 368 } 369 370 // Either or both of phi_map and ex_map might already be converted into phis. 371 Node* ex_control = ex_map->control(); 372 // if there is special marking on ex_map also, we add multiple edges from src 373 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 374 // how wide was the destination phi_map, originally? 375 uint orig_width = region->req(); 376 377 if (add_multiple) { 378 add_n_reqs(region, ex_control); 379 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 380 } else { 381 // ex_map has no merges, so we just add single edges everywhere 382 add_one_req(region, ex_control); 383 add_one_req(phi_map->i_o(), ex_map->i_o()); 384 } 385 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 386 if (mms.is_empty()) { 387 // get a copy of the base memory, and patch some inputs into it 388 const TypePtr* adr_type = mms.adr_type(C); 389 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 390 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 391 mms.set_memory(phi); 392 // Prepare to append interesting stuff onto the newly sliced phi: 393 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 394 } 395 // Append stuff from ex_map: 396 if (add_multiple) { 397 add_n_reqs(mms.memory(), mms.memory2()); 398 } else { 399 add_one_req(mms.memory(), mms.memory2()); 400 } 401 } 402 uint limit = ex_map->req(); 403 for (uint i = TypeFunc::Parms; i < limit; i++) { 404 // Skip everything in the JVMS after tos. (The ex_oop follows.) 405 if (i == tos) i = ex_jvms->monoff(); 406 Node* src = ex_map->in(i); 407 Node* dst = phi_map->in(i); 408 if (src != dst) { 409 PhiNode* phi; 410 if (dst->in(0) != region) { 411 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 412 record_for_igvn(phi); 413 _gvn.set_type(phi, phi->type()); 414 phi_map->set_req(i, dst); 415 // Prepare to append interesting stuff onto the new phi: 416 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 417 } else { 418 assert(dst->is_Phi(), "nobody else uses a hidden region"); 419 phi = dst->as_Phi(); 420 } 421 if (add_multiple && src->in(0) == ex_control) { 422 // Both are phis. 423 add_n_reqs(dst, src); 424 } else { 425 while (dst->req() < region->req()) add_one_req(dst, src); 426 } 427 const Type* srctype = _gvn.type(src); 428 if (phi->type() != srctype) { 429 const Type* dsttype = phi->type()->meet_speculative(srctype); 430 if (phi->type() != dsttype) { 431 phi->set_type(dsttype); 432 _gvn.set_type(phi, dsttype); 433 } 434 } 435 } 436 } 437 phi_map->merge_replaced_nodes_with(ex_map); 438 } 439 440 //--------------------------use_exception_state-------------------------------- 441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 442 if (failing()) { stop(); return top(); } 443 Node* region = phi_map->control(); 444 Node* hidden_merge_mark = root(); 445 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 446 Node* ex_oop = clear_saved_ex_oop(phi_map); 447 if (region->in(0) == hidden_merge_mark) { 448 // Special marking for internal ex-states. Process the phis now. 449 region->set_req(0, region); // now it's an ordinary region 450 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 451 // Note: Setting the jvms also sets the bci and sp. 452 set_control(_gvn.transform(region)); 453 uint tos = jvms()->stkoff() + sp(); 454 for (uint i = 1; i < tos; i++) { 455 Node* x = phi_map->in(i); 456 if (x->in(0) == region) { 457 assert(x->is_Phi(), "expected a special phi"); 458 phi_map->set_req(i, _gvn.transform(x)); 459 } 460 } 461 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 462 Node* x = mms.memory(); 463 if (x->in(0) == region) { 464 assert(x->is_Phi(), "nobody else uses a hidden region"); 465 mms.set_memory(_gvn.transform(x)); 466 } 467 } 468 if (ex_oop->in(0) == region) { 469 assert(ex_oop->is_Phi(), "expected a special phi"); 470 ex_oop = _gvn.transform(ex_oop); 471 } 472 } else { 473 set_jvms(phi_map->jvms()); 474 } 475 476 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 477 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 478 return ex_oop; 479 } 480 481 //---------------------------------java_bc------------------------------------- 482 Bytecodes::Code GraphKit::java_bc() const { 483 ciMethod* method = this->method(); 484 int bci = this->bci(); 485 if (method != NULL && bci != InvocationEntryBci) 486 return method->java_code_at_bci(bci); 487 else 488 return Bytecodes::_illegal; 489 } 490 491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 492 bool must_throw) { 493 // if the exception capability is set, then we will generate code 494 // to check the JavaThread.should_post_on_exceptions flag to see 495 // if we actually need to report exception events (for this 496 // thread). If we don't need to report exception events, we will 497 // take the normal fast path provided by add_exception_events. If 498 // exception event reporting is enabled for this thread, we will 499 // take the uncommon_trap in the BuildCutout below. 500 501 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 502 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 503 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 504 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 505 506 // Test the should_post_on_exceptions_flag vs. 0 507 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 508 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 509 510 // Branch to slow_path if should_post_on_exceptions_flag was true 511 { BuildCutout unless(this, tst, PROB_MAX); 512 // Do not try anything fancy if we're notifying the VM on every throw. 513 // Cf. case Bytecodes::_athrow in parse2.cpp. 514 uncommon_trap(reason, Deoptimization::Action_none, 515 (ciKlass*)NULL, (char*)NULL, must_throw); 516 } 517 518 } 519 520 //------------------------------builtin_throw---------------------------------- 521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 522 bool must_throw = true; 523 524 if (env()->jvmti_can_post_on_exceptions()) { 525 // check if we must post exception events, take uncommon trap if so 526 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 527 // here if should_post_on_exceptions is false 528 // continue on with the normal codegen 529 } 530 531 // If this particular condition has not yet happened at this 532 // bytecode, then use the uncommon trap mechanism, and allow for 533 // a future recompilation if several traps occur here. 534 // If the throw is hot, try to use a more complicated inline mechanism 535 // which keeps execution inside the compiled code. 536 bool treat_throw_as_hot = false; 537 ciMethodData* md = method()->method_data(); 538 539 if (ProfileTraps) { 540 if (too_many_traps(reason)) { 541 treat_throw_as_hot = true; 542 } 543 // (If there is no MDO at all, assume it is early in 544 // execution, and that any deopts are part of the 545 // startup transient, and don't need to be remembered.) 546 547 // Also, if there is a local exception handler, treat all throws 548 // as hot if there has been at least one in this method. 549 if (C->trap_count(reason) != 0 550 && method()->method_data()->trap_count(reason) != 0 551 && has_ex_handler()) { 552 treat_throw_as_hot = true; 553 } 554 } 555 556 // If this throw happens frequently, an uncommon trap might cause 557 // a performance pothole. If there is a local exception handler, 558 // and if this particular bytecode appears to be deoptimizing often, 559 // let us handle the throw inline, with a preconstructed instance. 560 // Note: If the deopt count has blown up, the uncommon trap 561 // runtime is going to flush this nmethod, not matter what. 562 if (treat_throw_as_hot 563 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 564 // If the throw is local, we use a pre-existing instance and 565 // punt on the backtrace. This would lead to a missing backtrace 566 // (a repeat of 4292742) if the backtrace object is ever asked 567 // for its backtrace. 568 // Fixing this remaining case of 4292742 requires some flavor of 569 // escape analysis. Leave that for the future. 570 ciInstance* ex_obj = NULL; 571 switch (reason) { 572 case Deoptimization::Reason_null_check: 573 ex_obj = env()->NullPointerException_instance(); 574 break; 575 case Deoptimization::Reason_div0_check: 576 ex_obj = env()->ArithmeticException_instance(); 577 break; 578 case Deoptimization::Reason_range_check: 579 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 580 break; 581 case Deoptimization::Reason_class_check: 582 if (java_bc() == Bytecodes::_aastore) { 583 ex_obj = env()->ArrayStoreException_instance(); 584 } else { 585 ex_obj = env()->ClassCastException_instance(); 586 } 587 break; 588 } 589 if (failing()) { stop(); return; } // exception allocation might fail 590 if (ex_obj != NULL) { 591 // Cheat with a preallocated exception object. 592 if (C->log() != NULL) 593 C->log()->elem("hot_throw preallocated='1' reason='%s'", 594 Deoptimization::trap_reason_name(reason)); 595 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 596 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 597 598 // Clear the detail message of the preallocated exception object. 599 // Weblogic sometimes mutates the detail message of exceptions 600 // using reflection. 601 int offset = java_lang_Throwable::get_detailMessage_offset(); 602 const TypePtr* adr_typ = ex_con->add_offset(offset); 603 604 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 605 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 606 // Conservatively release stores of object references. 607 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 608 609 add_exception_state(make_exception_state(ex_node)); 610 return; 611 } 612 } 613 614 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 615 // It won't be much cheaper than bailing to the interp., since we'll 616 // have to pass up all the debug-info, and the runtime will have to 617 // create the stack trace. 618 619 // Usual case: Bail to interpreter. 620 // Reserve the right to recompile if we haven't seen anything yet. 621 622 assert(!Deoptimization::reason_is_speculate(reason), "unsupported"); 623 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 624 if (treat_throw_as_hot 625 && (method()->method_data()->trap_recompiled_at(bci(), NULL) 626 || C->too_many_traps(reason))) { 627 // We cannot afford to take more traps here. Suffer in the interpreter. 628 if (C->log() != NULL) 629 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 630 Deoptimization::trap_reason_name(reason), 631 C->trap_count(reason)); 632 action = Deoptimization::Action_none; 633 } 634 635 // "must_throw" prunes the JVM state to include only the stack, if there 636 // are no local exception handlers. This should cut down on register 637 // allocation time and code size, by drastically reducing the number 638 // of in-edges on the call to the uncommon trap. 639 640 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 641 } 642 643 644 //----------------------------PreserveJVMState--------------------------------- 645 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 646 debug_only(kit->verify_map()); 647 _kit = kit; 648 _map = kit->map(); // preserve the map 649 _sp = kit->sp(); 650 kit->set_map(clone_map ? kit->clone_map() : NULL); 651 #ifdef ASSERT 652 _bci = kit->bci(); 653 Parse* parser = kit->is_Parse(); 654 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 655 _block = block; 656 #endif 657 } 658 PreserveJVMState::~PreserveJVMState() { 659 GraphKit* kit = _kit; 660 #ifdef ASSERT 661 assert(kit->bci() == _bci, "bci must not shift"); 662 Parse* parser = kit->is_Parse(); 663 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 664 assert(block == _block, "block must not shift"); 665 #endif 666 kit->set_map(_map); 667 kit->set_sp(_sp); 668 } 669 670 671 //-----------------------------BuildCutout------------------------------------- 672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 673 : PreserveJVMState(kit) 674 { 675 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 676 SafePointNode* outer_map = _map; // preserved map is caller's 677 SafePointNode* inner_map = kit->map(); 678 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 679 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 680 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 681 } 682 BuildCutout::~BuildCutout() { 683 GraphKit* kit = _kit; 684 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 685 } 686 687 //---------------------------PreserveReexecuteState---------------------------- 688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 689 assert(!kit->stopped(), "must call stopped() before"); 690 _kit = kit; 691 _sp = kit->sp(); 692 _reexecute = kit->jvms()->_reexecute; 693 } 694 PreserveReexecuteState::~PreserveReexecuteState() { 695 if (_kit->stopped()) return; 696 _kit->jvms()->_reexecute = _reexecute; 697 _kit->set_sp(_sp); 698 } 699 700 //------------------------------clone_map-------------------------------------- 701 // Implementation of PreserveJVMState 702 // 703 // Only clone_map(...) here. If this function is only used in the 704 // PreserveJVMState class we may want to get rid of this extra 705 // function eventually and do it all there. 706 707 SafePointNode* GraphKit::clone_map() { 708 if (map() == NULL) return NULL; 709 710 // Clone the memory edge first 711 Node* mem = MergeMemNode::make(C, map()->memory()); 712 gvn().set_type_bottom(mem); 713 714 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 715 JVMState* jvms = this->jvms(); 716 JVMState* clonejvms = jvms->clone_shallow(C); 717 clonemap->set_memory(mem); 718 clonemap->set_jvms(clonejvms); 719 clonejvms->set_map(clonemap); 720 record_for_igvn(clonemap); 721 gvn().set_type_bottom(clonemap); 722 return clonemap; 723 } 724 725 726 //-----------------------------set_map_clone----------------------------------- 727 void GraphKit::set_map_clone(SafePointNode* m) { 728 _map = m; 729 _map = clone_map(); 730 _map->set_next_exception(NULL); 731 debug_only(verify_map()); 732 } 733 734 735 //----------------------------kill_dead_locals--------------------------------- 736 // Detect any locals which are known to be dead, and force them to top. 737 void GraphKit::kill_dead_locals() { 738 // Consult the liveness information for the locals. If any 739 // of them are unused, then they can be replaced by top(). This 740 // should help register allocation time and cut down on the size 741 // of the deoptimization information. 742 743 // This call is made from many of the bytecode handling 744 // subroutines called from the Big Switch in do_one_bytecode. 745 // Every bytecode which might include a slow path is responsible 746 // for killing its dead locals. The more consistent we 747 // are about killing deads, the fewer useless phis will be 748 // constructed for them at various merge points. 749 750 // bci can be -1 (InvocationEntryBci). We return the entry 751 // liveness for the method. 752 753 if (method() == NULL || method()->code_size() == 0) { 754 // We are building a graph for a call to a native method. 755 // All locals are live. 756 return; 757 } 758 759 ResourceMark rm; 760 761 // Consult the liveness information for the locals. If any 762 // of them are unused, then they can be replaced by top(). This 763 // should help register allocation time and cut down on the size 764 // of the deoptimization information. 765 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 766 767 int len = (int)live_locals.size(); 768 assert(len <= jvms()->loc_size(), "too many live locals"); 769 for (int local = 0; local < len; local++) { 770 if (!live_locals.at(local)) { 771 set_local(local, top()); 772 } 773 } 774 } 775 776 #ifdef ASSERT 777 //-------------------------dead_locals_are_killed------------------------------ 778 // Return true if all dead locals are set to top in the map. 779 // Used to assert "clean" debug info at various points. 780 bool GraphKit::dead_locals_are_killed() { 781 if (method() == NULL || method()->code_size() == 0) { 782 // No locals need to be dead, so all is as it should be. 783 return true; 784 } 785 786 // Make sure somebody called kill_dead_locals upstream. 787 ResourceMark rm; 788 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 789 if (jvms->loc_size() == 0) continue; // no locals to consult 790 SafePointNode* map = jvms->map(); 791 ciMethod* method = jvms->method(); 792 int bci = jvms->bci(); 793 if (jvms == this->jvms()) { 794 bci = this->bci(); // it might not yet be synched 795 } 796 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 797 int len = (int)live_locals.size(); 798 if (!live_locals.is_valid() || len == 0) 799 // This method is trivial, or is poisoned by a breakpoint. 800 return true; 801 assert(len == jvms->loc_size(), "live map consistent with locals map"); 802 for (int local = 0; local < len; local++) { 803 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 804 if (PrintMiscellaneous && (Verbose || WizardMode)) { 805 tty->print_cr("Zombie local %d: ", local); 806 jvms->dump(); 807 } 808 return false; 809 } 810 } 811 } 812 return true; 813 } 814 815 #endif //ASSERT 816 817 // Helper function for enforcing certain bytecodes to reexecute if 818 // deoptimization happens 819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 820 ciMethod* cur_method = jvms->method(); 821 int cur_bci = jvms->bci(); 822 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 823 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 824 return Interpreter::bytecode_should_reexecute(code) || 825 is_anewarray && code == Bytecodes::_multianewarray; 826 // Reexecute _multianewarray bytecode which was replaced with 827 // sequence of [a]newarray. See Parse::do_multianewarray(). 828 // 829 // Note: interpreter should not have it set since this optimization 830 // is limited by dimensions and guarded by flag so in some cases 831 // multianewarray() runtime calls will be generated and 832 // the bytecode should not be reexecutes (stack will not be reset). 833 } else 834 return false; 835 } 836 837 // Helper function for adding JVMState and debug information to node 838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 839 // Add the safepoint edges to the call (or other safepoint). 840 841 // Make sure dead locals are set to top. This 842 // should help register allocation time and cut down on the size 843 // of the deoptimization information. 844 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 845 846 // Walk the inline list to fill in the correct set of JVMState's 847 // Also fill in the associated edges for each JVMState. 848 849 // If the bytecode needs to be reexecuted we need to put 850 // the arguments back on the stack. 851 const bool should_reexecute = jvms()->should_reexecute(); 852 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 853 854 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 855 // undefined if the bci is different. This is normal for Parse but it 856 // should not happen for LibraryCallKit because only one bci is processed. 857 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 858 "in LibraryCallKit the reexecute bit should not change"); 859 860 // If we are guaranteed to throw, we can prune everything but the 861 // input to the current bytecode. 862 bool can_prune_locals = false; 863 uint stack_slots_not_pruned = 0; 864 int inputs = 0, depth = 0; 865 if (must_throw) { 866 assert(method() == youngest_jvms->method(), "sanity"); 867 if (compute_stack_effects(inputs, depth)) { 868 can_prune_locals = true; 869 stack_slots_not_pruned = inputs; 870 } 871 } 872 873 if (env()->jvmti_can_access_local_variables()) { 874 // At any safepoint, this method can get breakpointed, which would 875 // then require an immediate deoptimization. 876 can_prune_locals = false; // do not prune locals 877 stack_slots_not_pruned = 0; 878 } 879 880 // do not scribble on the input jvms 881 JVMState* out_jvms = youngest_jvms->clone_deep(C); 882 call->set_jvms(out_jvms); // Start jvms list for call node 883 884 // For a known set of bytecodes, the interpreter should reexecute them if 885 // deoptimization happens. We set the reexecute state for them here 886 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 887 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 888 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 889 } 890 891 // Presize the call: 892 DEBUG_ONLY(uint non_debug_edges = call->req()); 893 call->add_req_batch(top(), youngest_jvms->debug_depth()); 894 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 895 896 // Set up edges so that the call looks like this: 897 // Call [state:] ctl io mem fptr retadr 898 // [parms:] parm0 ... parmN 899 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 900 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 901 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 902 // Note that caller debug info precedes callee debug info. 903 904 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 905 uint debug_ptr = call->req(); 906 907 // Loop over the map input edges associated with jvms, add them 908 // to the call node, & reset all offsets to match call node array. 909 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 910 uint debug_end = debug_ptr; 911 uint debug_start = debug_ptr - in_jvms->debug_size(); 912 debug_ptr = debug_start; // back up the ptr 913 914 uint p = debug_start; // walks forward in [debug_start, debug_end) 915 uint j, k, l; 916 SafePointNode* in_map = in_jvms->map(); 917 out_jvms->set_map(call); 918 919 if (can_prune_locals) { 920 assert(in_jvms->method() == out_jvms->method(), "sanity"); 921 // If the current throw can reach an exception handler in this JVMS, 922 // then we must keep everything live that can reach that handler. 923 // As a quick and dirty approximation, we look for any handlers at all. 924 if (in_jvms->method()->has_exception_handlers()) { 925 can_prune_locals = false; 926 } 927 } 928 929 // Add the Locals 930 k = in_jvms->locoff(); 931 l = in_jvms->loc_size(); 932 out_jvms->set_locoff(p); 933 if (!can_prune_locals) { 934 for (j = 0; j < l; j++) 935 call->set_req(p++, in_map->in(k+j)); 936 } else { 937 p += l; // already set to top above by add_req_batch 938 } 939 940 // Add the Expression Stack 941 k = in_jvms->stkoff(); 942 l = in_jvms->sp(); 943 out_jvms->set_stkoff(p); 944 if (!can_prune_locals) { 945 for (j = 0; j < l; j++) 946 call->set_req(p++, in_map->in(k+j)); 947 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 948 // Divide stack into {S0,...,S1}, where S0 is set to top. 949 uint s1 = stack_slots_not_pruned; 950 stack_slots_not_pruned = 0; // for next iteration 951 if (s1 > l) s1 = l; 952 uint s0 = l - s1; 953 p += s0; // skip the tops preinstalled by add_req_batch 954 for (j = s0; j < l; j++) 955 call->set_req(p++, in_map->in(k+j)); 956 } else { 957 p += l; // already set to top above by add_req_batch 958 } 959 960 // Add the Monitors 961 k = in_jvms->monoff(); 962 l = in_jvms->mon_size(); 963 out_jvms->set_monoff(p); 964 for (j = 0; j < l; j++) 965 call->set_req(p++, in_map->in(k+j)); 966 967 // Copy any scalar object fields. 968 k = in_jvms->scloff(); 969 l = in_jvms->scl_size(); 970 out_jvms->set_scloff(p); 971 for (j = 0; j < l; j++) 972 call->set_req(p++, in_map->in(k+j)); 973 974 // Finish the new jvms. 975 out_jvms->set_endoff(p); 976 977 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 978 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 979 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 980 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 981 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 982 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 983 984 // Update the two tail pointers in parallel. 985 out_jvms = out_jvms->caller(); 986 in_jvms = in_jvms->caller(); 987 } 988 989 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 990 991 // Test the correctness of JVMState::debug_xxx accessors: 992 assert(call->jvms()->debug_start() == non_debug_edges, ""); 993 assert(call->jvms()->debug_end() == call->req(), ""); 994 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 995 } 996 997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 998 Bytecodes::Code code = java_bc(); 999 if (code == Bytecodes::_wide) { 1000 code = method()->java_code_at_bci(bci() + 1); 1001 } 1002 1003 BasicType rtype = T_ILLEGAL; 1004 int rsize = 0; 1005 1006 if (code != Bytecodes::_illegal) { 1007 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1008 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1009 if (rtype < T_CONFLICT) 1010 rsize = type2size[rtype]; 1011 } 1012 1013 switch (code) { 1014 case Bytecodes::_illegal: 1015 return false; 1016 1017 case Bytecodes::_ldc: 1018 case Bytecodes::_ldc_w: 1019 case Bytecodes::_ldc2_w: 1020 inputs = 0; 1021 break; 1022 1023 case Bytecodes::_dup: inputs = 1; break; 1024 case Bytecodes::_dup_x1: inputs = 2; break; 1025 case Bytecodes::_dup_x2: inputs = 3; break; 1026 case Bytecodes::_dup2: inputs = 2; break; 1027 case Bytecodes::_dup2_x1: inputs = 3; break; 1028 case Bytecodes::_dup2_x2: inputs = 4; break; 1029 case Bytecodes::_swap: inputs = 2; break; 1030 case Bytecodes::_arraylength: inputs = 1; break; 1031 1032 case Bytecodes::_getstatic: 1033 case Bytecodes::_putstatic: 1034 case Bytecodes::_getfield: 1035 case Bytecodes::_putfield: 1036 { 1037 bool ignored_will_link; 1038 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1039 int size = field->type()->size(); 1040 bool is_get = (depth >= 0), is_static = (depth & 1); 1041 inputs = (is_static ? 0 : 1); 1042 if (is_get) { 1043 depth = size - inputs; 1044 } else { 1045 inputs += size; // putxxx pops the value from the stack 1046 depth = - inputs; 1047 } 1048 } 1049 break; 1050 1051 case Bytecodes::_invokevirtual: 1052 case Bytecodes::_invokespecial: 1053 case Bytecodes::_invokestatic: 1054 case Bytecodes::_invokedynamic: 1055 case Bytecodes::_invokeinterface: 1056 { 1057 bool ignored_will_link; 1058 ciSignature* declared_signature = NULL; 1059 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1060 assert(declared_signature != NULL, "cannot be null"); 1061 inputs = declared_signature->arg_size_for_bc(code); 1062 int size = declared_signature->return_type()->size(); 1063 depth = size - inputs; 1064 } 1065 break; 1066 1067 case Bytecodes::_multianewarray: 1068 { 1069 ciBytecodeStream iter(method()); 1070 iter.reset_to_bci(bci()); 1071 iter.next(); 1072 inputs = iter.get_dimensions(); 1073 assert(rsize == 1, ""); 1074 depth = rsize - inputs; 1075 } 1076 break; 1077 1078 case Bytecodes::_ireturn: 1079 case Bytecodes::_lreturn: 1080 case Bytecodes::_freturn: 1081 case Bytecodes::_dreturn: 1082 case Bytecodes::_areturn: 1083 assert(rsize = -depth, ""); 1084 inputs = rsize; 1085 break; 1086 1087 case Bytecodes::_jsr: 1088 case Bytecodes::_jsr_w: 1089 inputs = 0; 1090 depth = 1; // S.B. depth=1, not zero 1091 break; 1092 1093 default: 1094 // bytecode produces a typed result 1095 inputs = rsize - depth; 1096 assert(inputs >= 0, ""); 1097 break; 1098 } 1099 1100 #ifdef ASSERT 1101 // spot check 1102 int outputs = depth + inputs; 1103 assert(outputs >= 0, "sanity"); 1104 switch (code) { 1105 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1106 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1107 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1108 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1109 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1110 } 1111 #endif //ASSERT 1112 1113 return true; 1114 } 1115 1116 1117 1118 //------------------------------basic_plus_adr--------------------------------- 1119 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1120 // short-circuit a common case 1121 if (offset == intcon(0)) return ptr; 1122 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1123 } 1124 1125 Node* GraphKit::ConvI2L(Node* offset) { 1126 // short-circuit a common case 1127 jint offset_con = find_int_con(offset, Type::OffsetBot); 1128 if (offset_con != Type::OffsetBot) { 1129 return longcon((jlong) offset_con); 1130 } 1131 return _gvn.transform( new (C) ConvI2LNode(offset)); 1132 } 1133 1134 Node* GraphKit::ConvI2UL(Node* offset) { 1135 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1136 if (offset_con != (juint) Type::OffsetBot) { 1137 return longcon((julong) offset_con); 1138 } 1139 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset)); 1140 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) ); 1141 return _gvn.transform( new (C) AndLNode(conv, mask) ); 1142 } 1143 1144 Node* GraphKit::ConvL2I(Node* offset) { 1145 // short-circuit a common case 1146 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1147 if (offset_con != (jlong)Type::OffsetBot) { 1148 return intcon((int) offset_con); 1149 } 1150 return _gvn.transform( new (C) ConvL2INode(offset)); 1151 } 1152 1153 //-------------------------load_object_klass----------------------------------- 1154 Node* GraphKit::load_object_klass(Node* obj) { 1155 // Special-case a fresh allocation to avoid building nodes: 1156 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1157 if (akls != NULL) return akls; 1158 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1159 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1160 } 1161 1162 //-------------------------load_array_length----------------------------------- 1163 Node* GraphKit::load_array_length(Node* array) { 1164 // Special-case a fresh allocation to avoid building nodes: 1165 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1166 Node *alen; 1167 if (alloc == NULL) { 1168 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1169 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1170 } else { 1171 alen = alloc->Ideal_length(); 1172 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1173 if (ccast != alen) { 1174 alen = _gvn.transform(ccast); 1175 } 1176 } 1177 return alen; 1178 } 1179 1180 //------------------------------do_null_check---------------------------------- 1181 // Helper function to do a NULL pointer check. Returned value is 1182 // the incoming address with NULL casted away. You are allowed to use the 1183 // not-null value only if you are control dependent on the test. 1184 extern int explicit_null_checks_inserted, 1185 explicit_null_checks_elided; 1186 Node* GraphKit::null_check_common(Node* value, BasicType type, 1187 // optional arguments for variations: 1188 bool assert_null, 1189 Node* *null_control) { 1190 assert(!assert_null || null_control == NULL, "not both at once"); 1191 if (stopped()) return top(); 1192 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1193 // For some performance testing, we may wish to suppress null checking. 1194 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1195 return value; 1196 } 1197 explicit_null_checks_inserted++; 1198 1199 // Construct NULL check 1200 Node *chk = NULL; 1201 switch(type) { 1202 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1203 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1204 case T_ARRAY : // fall through 1205 type = T_OBJECT; // simplify further tests 1206 case T_OBJECT : { 1207 const Type *t = _gvn.type( value ); 1208 1209 const TypeOopPtr* tp = t->isa_oopptr(); 1210 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1211 // Only for do_null_check, not any of its siblings: 1212 && !assert_null && null_control == NULL) { 1213 // Usually, any field access or invocation on an unloaded oop type 1214 // will simply fail to link, since the statically linked class is 1215 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1216 // the static class is loaded but the sharper oop type is not. 1217 // Rather than checking for this obscure case in lots of places, 1218 // we simply observe that a null check on an unloaded class 1219 // will always be followed by a nonsense operation, so we 1220 // can just issue the uncommon trap here. 1221 // Our access to the unloaded class will only be correct 1222 // after it has been loaded and initialized, which requires 1223 // a trip through the interpreter. 1224 #ifndef PRODUCT 1225 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1226 #endif 1227 uncommon_trap(Deoptimization::Reason_unloaded, 1228 Deoptimization::Action_reinterpret, 1229 tp->klass(), "!loaded"); 1230 return top(); 1231 } 1232 1233 if (assert_null) { 1234 // See if the type is contained in NULL_PTR. 1235 // If so, then the value is already null. 1236 if (t->higher_equal(TypePtr::NULL_PTR)) { 1237 explicit_null_checks_elided++; 1238 return value; // Elided null assert quickly! 1239 } 1240 } else { 1241 // See if mixing in the NULL pointer changes type. 1242 // If so, then the NULL pointer was not allowed in the original 1243 // type. In other words, "value" was not-null. 1244 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1245 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1246 explicit_null_checks_elided++; 1247 return value; // Elided null check quickly! 1248 } 1249 } 1250 chk = new (C) CmpPNode( value, null() ); 1251 break; 1252 } 1253 1254 default: 1255 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1256 } 1257 assert(chk != NULL, "sanity check"); 1258 chk = _gvn.transform(chk); 1259 1260 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1261 BoolNode *btst = new (C) BoolNode( chk, btest); 1262 Node *tst = _gvn.transform( btst ); 1263 1264 //----------- 1265 // if peephole optimizations occurred, a prior test existed. 1266 // If a prior test existed, maybe it dominates as we can avoid this test. 1267 if (tst != btst && type == T_OBJECT) { 1268 // At this point we want to scan up the CFG to see if we can 1269 // find an identical test (and so avoid this test altogether). 1270 Node *cfg = control(); 1271 int depth = 0; 1272 while( depth < 16 ) { // Limit search depth for speed 1273 if( cfg->Opcode() == Op_IfTrue && 1274 cfg->in(0)->in(1) == tst ) { 1275 // Found prior test. Use "cast_not_null" to construct an identical 1276 // CastPP (and hence hash to) as already exists for the prior test. 1277 // Return that casted value. 1278 if (assert_null) { 1279 replace_in_map(value, null()); 1280 return null(); // do not issue the redundant test 1281 } 1282 Node *oldcontrol = control(); 1283 set_control(cfg); 1284 Node *res = cast_not_null(value); 1285 set_control(oldcontrol); 1286 explicit_null_checks_elided++; 1287 return res; 1288 } 1289 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1290 if (cfg == NULL) break; // Quit at region nodes 1291 depth++; 1292 } 1293 } 1294 1295 //----------- 1296 // Branch to failure if null 1297 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1298 Deoptimization::DeoptReason reason; 1299 if (assert_null) 1300 reason = Deoptimization::Reason_null_assert; 1301 else if (type == T_OBJECT) 1302 reason = Deoptimization::Reason_null_check; 1303 else 1304 reason = Deoptimization::Reason_div0_check; 1305 1306 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1307 // ciMethodData::has_trap_at will return a conservative -1 if any 1308 // must-be-null assertion has failed. This could cause performance 1309 // problems for a method after its first do_null_assert failure. 1310 // Consider using 'Reason_class_check' instead? 1311 1312 // To cause an implicit null check, we set the not-null probability 1313 // to the maximum (PROB_MAX). For an explicit check the probability 1314 // is set to a smaller value. 1315 if (null_control != NULL || too_many_traps(reason)) { 1316 // probability is less likely 1317 ok_prob = PROB_LIKELY_MAG(3); 1318 } else if (!assert_null && 1319 (ImplicitNullCheckThreshold > 0) && 1320 method() != NULL && 1321 (method()->method_data()->trap_count(reason) 1322 >= (uint)ImplicitNullCheckThreshold)) { 1323 ok_prob = PROB_LIKELY_MAG(3); 1324 } 1325 1326 if (null_control != NULL) { 1327 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1328 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1329 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1330 if (null_true == top()) 1331 explicit_null_checks_elided++; 1332 (*null_control) = null_true; 1333 } else { 1334 BuildCutout unless(this, tst, ok_prob); 1335 // Check for optimizer eliding test at parse time 1336 if (stopped()) { 1337 // Failure not possible; do not bother making uncommon trap. 1338 explicit_null_checks_elided++; 1339 } else if (assert_null) { 1340 uncommon_trap(reason, 1341 Deoptimization::Action_make_not_entrant, 1342 NULL, "assert_null"); 1343 } else { 1344 replace_in_map(value, zerocon(type)); 1345 builtin_throw(reason); 1346 } 1347 } 1348 1349 // Must throw exception, fall-thru not possible? 1350 if (stopped()) { 1351 return top(); // No result 1352 } 1353 1354 if (assert_null) { 1355 // Cast obj to null on this path. 1356 replace_in_map(value, zerocon(type)); 1357 return zerocon(type); 1358 } 1359 1360 // Cast obj to not-null on this path, if there is no null_control. 1361 // (If there is a null_control, a non-null value may come back to haunt us.) 1362 if (type == T_OBJECT) { 1363 Node* cast = cast_not_null(value, false); 1364 if (null_control == NULL || (*null_control) == top()) 1365 replace_in_map(value, cast); 1366 value = cast; 1367 } 1368 1369 return value; 1370 } 1371 1372 1373 //------------------------------cast_not_null---------------------------------- 1374 // Cast obj to not-null on this path 1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1376 const Type *t = _gvn.type(obj); 1377 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1378 // Object is already not-null? 1379 if( t == t_not_null ) return obj; 1380 1381 Node *cast = new (C) CastPPNode(obj,t_not_null); 1382 cast->init_req(0, control()); 1383 cast = _gvn.transform( cast ); 1384 1385 // Scan for instances of 'obj' in the current JVM mapping. 1386 // These instances are known to be not-null after the test. 1387 if (do_replace_in_map) 1388 replace_in_map(obj, cast); 1389 1390 return cast; // Return casted value 1391 } 1392 1393 1394 //--------------------------replace_in_map------------------------------------- 1395 void GraphKit::replace_in_map(Node* old, Node* neww) { 1396 if (old == neww) { 1397 return; 1398 } 1399 1400 map()->replace_edge(old, neww); 1401 1402 // Note: This operation potentially replaces any edge 1403 // on the map. This includes locals, stack, and monitors 1404 // of the current (innermost) JVM state. 1405 1406 // don't let inconsistent types from profiling escape this 1407 // method 1408 1409 const Type* told = _gvn.type(old); 1410 const Type* tnew = _gvn.type(neww); 1411 1412 if (!tnew->higher_equal(told)) { 1413 return; 1414 } 1415 1416 map()->record_replaced_node(old, neww); 1417 } 1418 1419 1420 //============================================================================= 1421 //--------------------------------memory--------------------------------------- 1422 Node* GraphKit::memory(uint alias_idx) { 1423 MergeMemNode* mem = merged_memory(); 1424 Node* p = mem->memory_at(alias_idx); 1425 _gvn.set_type(p, Type::MEMORY); // must be mapped 1426 return p; 1427 } 1428 1429 //-----------------------------reset_memory------------------------------------ 1430 Node* GraphKit::reset_memory() { 1431 Node* mem = map()->memory(); 1432 // do not use this node for any more parsing! 1433 debug_only( map()->set_memory((Node*)NULL) ); 1434 return _gvn.transform( mem ); 1435 } 1436 1437 //------------------------------set_all_memory--------------------------------- 1438 void GraphKit::set_all_memory(Node* newmem) { 1439 Node* mergemem = MergeMemNode::make(C, newmem); 1440 gvn().set_type_bottom(mergemem); 1441 map()->set_memory(mergemem); 1442 } 1443 1444 //------------------------------set_all_memory_call---------------------------- 1445 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1446 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1447 set_all_memory(newmem); 1448 } 1449 1450 //============================================================================= 1451 // 1452 // parser factory methods for MemNodes 1453 // 1454 // These are layered on top of the factory methods in LoadNode and StoreNode, 1455 // and integrate with the parser's memory state and _gvn engine. 1456 // 1457 1458 // factory methods in "int adr_idx" 1459 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1460 int adr_idx, 1461 MemNode::MemOrd mo, 1462 LoadNode::ControlDependency control_dependency, 1463 bool require_atomic_access, 1464 bool unaligned, 1465 bool mismatched) { 1466 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1467 const TypePtr* adr_type = NULL; // debug-mode-only argument 1468 debug_only(adr_type = C->get_adr_type(adr_idx)); 1469 Node* mem = memory(adr_idx); 1470 Node* ld; 1471 if (require_atomic_access && bt == T_LONG) { 1472 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency); 1473 } else if (require_atomic_access && bt == T_DOUBLE) { 1474 ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency); 1475 } else { 1476 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency); 1477 } 1478 if (unaligned) { 1479 ld->as_Load()->set_unaligned_access(); 1480 } 1481 if (mismatched) { 1482 ld->as_Load()->set_mismatched_access(); 1483 } 1484 ld = _gvn.transform(ld); 1485 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1486 // Improve graph before escape analysis and boxing elimination. 1487 record_for_igvn(ld); 1488 } 1489 return ld; 1490 } 1491 1492 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1493 int adr_idx, 1494 MemNode::MemOrd mo, 1495 bool require_atomic_access, 1496 bool unaligned, 1497 bool mismatched) { 1498 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1499 const TypePtr* adr_type = NULL; 1500 debug_only(adr_type = C->get_adr_type(adr_idx)); 1501 Node *mem = memory(adr_idx); 1502 Node* st; 1503 if (require_atomic_access && bt == T_LONG) { 1504 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1505 } else if (require_atomic_access && bt == T_DOUBLE) { 1506 st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1507 } else { 1508 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1509 } 1510 if (unaligned) { 1511 st->as_Store()->set_unaligned_access(); 1512 } 1513 if (mismatched) { 1514 st->as_Store()->set_mismatched_access(); 1515 } 1516 st = _gvn.transform(st); 1517 set_memory(st, adr_idx); 1518 // Back-to-back stores can only remove intermediate store with DU info 1519 // so push on worklist for optimizer. 1520 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1521 record_for_igvn(st); 1522 1523 return st; 1524 } 1525 1526 1527 void GraphKit::pre_barrier(bool do_load, 1528 Node* ctl, 1529 Node* obj, 1530 Node* adr, 1531 uint adr_idx, 1532 Node* val, 1533 const TypeOopPtr* val_type, 1534 Node* pre_val, 1535 BasicType bt) { 1536 1537 BarrierSet* bs = Universe::heap()->barrier_set(); 1538 set_control(ctl); 1539 switch (bs->kind()) { 1540 case BarrierSet::G1SATBCT: 1541 case BarrierSet::G1SATBCTLogging: 1542 case BarrierSet::ShenandoahBarrierSet: 1543 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1544 break; 1545 1546 case BarrierSet::CardTableModRef: 1547 case BarrierSet::CardTableExtension: 1548 case BarrierSet::ModRef: 1549 break; 1550 1551 case BarrierSet::Other: 1552 default : 1553 ShouldNotReachHere(); 1554 1555 } 1556 } 1557 1558 bool GraphKit::can_move_pre_barrier() const { 1559 BarrierSet* bs = Universe::heap()->barrier_set(); 1560 switch (bs->kind()) { 1561 case BarrierSet::G1SATBCT: 1562 case BarrierSet::G1SATBCTLogging: 1563 case BarrierSet::ShenandoahBarrierSet: 1564 return true; // Can move it if no safepoint 1565 1566 case BarrierSet::CardTableModRef: 1567 case BarrierSet::CardTableExtension: 1568 case BarrierSet::ModRef: 1569 return true; // There is no pre-barrier 1570 1571 case BarrierSet::Other: 1572 default : 1573 ShouldNotReachHere(); 1574 } 1575 return false; 1576 } 1577 1578 void GraphKit::post_barrier(Node* ctl, 1579 Node* store, 1580 Node* obj, 1581 Node* adr, 1582 uint adr_idx, 1583 Node* val, 1584 BasicType bt, 1585 bool use_precise) { 1586 BarrierSet* bs = Universe::heap()->barrier_set(); 1587 set_control(ctl); 1588 switch (bs->kind()) { 1589 case BarrierSet::G1SATBCT: 1590 case BarrierSet::G1SATBCTLogging: 1591 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1592 break; 1593 1594 case BarrierSet::CardTableModRef: 1595 case BarrierSet::CardTableExtension: 1596 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1597 break; 1598 1599 case BarrierSet::ModRef: 1600 case BarrierSet::ShenandoahBarrierSet: 1601 break; 1602 1603 case BarrierSet::Other: 1604 default : 1605 ShouldNotReachHere(); 1606 1607 } 1608 } 1609 1610 Node* GraphKit::store_oop(Node* ctl, 1611 Node* obj, 1612 Node* adr, 1613 const TypePtr* adr_type, 1614 Node* val, 1615 const TypeOopPtr* val_type, 1616 BasicType bt, 1617 bool use_precise, 1618 MemNode::MemOrd mo, 1619 bool mismatched) { 1620 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1621 // could be delayed during Parse (for example, in adjust_map_after_if()). 1622 // Execute transformation here to avoid barrier generation in such case. 1623 if (_gvn.type(val) == TypePtr::NULL_PTR) 1624 val = _gvn.makecon(TypePtr::NULL_PTR); 1625 1626 set_control(ctl); 1627 if (stopped()) return top(); // Dead path ? 1628 1629 assert(bt == T_OBJECT, "sanity"); 1630 assert(val != NULL, "not dead path"); 1631 uint adr_idx = C->get_alias_index(adr_type); 1632 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1633 1634 pre_barrier(true /* do_load */, 1635 control(), obj, adr, adr_idx, val, val_type, 1636 NULL /* pre_val */, 1637 bt); 1638 1639 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1640 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1641 return store; 1642 } 1643 1644 // Could be an array or object we don't know at compile time (unsafe ref.) 1645 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1646 Node* obj, // containing obj 1647 Node* adr, // actual adress to store val at 1648 const TypePtr* adr_type, 1649 Node* val, 1650 BasicType bt, 1651 MemNode::MemOrd mo, 1652 bool mismatched) { 1653 Compile::AliasType* at = C->alias_type(adr_type); 1654 const TypeOopPtr* val_type = NULL; 1655 if (adr_type->isa_instptr()) { 1656 if (at->field() != NULL) { 1657 // known field. This code is a copy of the do_put_xxx logic. 1658 ciField* field = at->field(); 1659 if (!field->type()->is_loaded()) { 1660 val_type = TypeInstPtr::BOTTOM; 1661 } else { 1662 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1663 } 1664 } 1665 } else if (adr_type->isa_aryptr()) { 1666 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1667 } 1668 if (val_type == NULL) { 1669 val_type = TypeInstPtr::BOTTOM; 1670 } 1671 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1672 } 1673 1674 1675 //-------------------------array_element_address------------------------- 1676 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1677 const TypeInt* sizetype, Node* ctrl) { 1678 uint shift = exact_log2(type2aelembytes(elembt)); 1679 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1680 1681 // short-circuit a common case (saves lots of confusing waste motion) 1682 jint idx_con = find_int_con(idx, -1); 1683 if (idx_con >= 0) { 1684 intptr_t offset = header + ((intptr_t)idx_con << shift); 1685 return basic_plus_adr(ary, offset); 1686 } 1687 1688 // must be correct type for alignment purposes 1689 Node* base = basic_plus_adr(ary, header); 1690 #ifdef _LP64 1691 // The scaled index operand to AddP must be a clean 64-bit value. 1692 // Java allows a 32-bit int to be incremented to a negative 1693 // value, which appears in a 64-bit register as a large 1694 // positive number. Using that large positive number as an 1695 // operand in pointer arithmetic has bad consequences. 1696 // On the other hand, 32-bit overflow is rare, and the possibility 1697 // can often be excluded, if we annotate the ConvI2L node with 1698 // a type assertion that its value is known to be a small positive 1699 // number. (The prior range check has ensured this.) 1700 // This assertion is used by ConvI2LNode::Ideal. 1701 int index_max = max_jint - 1; // array size is max_jint, index is one less 1702 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1703 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 1704 idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl); 1705 #endif 1706 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1707 return basic_plus_adr(ary, base, scale); 1708 } 1709 1710 //-------------------------load_array_element------------------------- 1711 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1712 const Type* elemtype = arytype->elem(); 1713 BasicType elembt = elemtype->array_element_basic_type(); 1714 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1715 if (elembt == T_NARROWOOP) { 1716 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1717 } 1718 assert(elembt != T_OBJECT && elembt != T_ARRAY, "sanity"); 1719 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1720 return ld; 1721 } 1722 1723 //-------------------------set_arguments_for_java_call------------------------- 1724 // Arguments (pre-popped from the stack) are taken from the JVMS. 1725 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1726 // Add the call arguments: 1727 uint nargs = call->method()->arg_size(); 1728 for (uint i = 0; i < nargs; i++) { 1729 Node* arg = argument(i); 1730 call->init_req(i + TypeFunc::Parms, arg); 1731 } 1732 } 1733 1734 //---------------------------set_edges_for_java_call--------------------------- 1735 // Connect a newly created call into the current JVMS. 1736 // A return value node (if any) is returned from set_edges_for_java_call. 1737 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1738 1739 // Add the predefined inputs: 1740 call->init_req( TypeFunc::Control, control() ); 1741 call->init_req( TypeFunc::I_O , i_o() ); 1742 call->init_req( TypeFunc::Memory , reset_memory() ); 1743 call->init_req( TypeFunc::FramePtr, frameptr() ); 1744 call->init_req( TypeFunc::ReturnAdr, top() ); 1745 1746 add_safepoint_edges(call, must_throw); 1747 1748 Node* xcall = _gvn.transform(call); 1749 1750 if (xcall == top()) { 1751 set_control(top()); 1752 return; 1753 } 1754 assert(xcall == call, "call identity is stable"); 1755 1756 // Re-use the current map to produce the result. 1757 1758 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1759 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1760 set_all_memory_call(xcall, separate_io_proj); 1761 1762 //return xcall; // no need, caller already has it 1763 } 1764 1765 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1766 if (stopped()) return top(); // maybe the call folded up? 1767 1768 // Capture the return value, if any. 1769 Node* ret; 1770 if (call->method() == NULL || 1771 call->method()->return_type()->basic_type() == T_VOID) 1772 ret = top(); 1773 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1774 1775 // Note: Since any out-of-line call can produce an exception, 1776 // we always insert an I_O projection from the call into the result. 1777 1778 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1779 1780 if (separate_io_proj) { 1781 // The caller requested separate projections be used by the fall 1782 // through and exceptional paths, so replace the projections for 1783 // the fall through path. 1784 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1785 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1786 } 1787 return ret; 1788 } 1789 1790 //--------------------set_predefined_input_for_runtime_call-------------------- 1791 // Reading and setting the memory state is way conservative here. 1792 // The real problem is that I am not doing real Type analysis on memory, 1793 // so I cannot distinguish card mark stores from other stores. Across a GC 1794 // point the Store Barrier and the card mark memory has to agree. I cannot 1795 // have a card mark store and its barrier split across the GC point from 1796 // either above or below. Here I get that to happen by reading ALL of memory. 1797 // A better answer would be to separate out card marks from other memory. 1798 // For now, return the input memory state, so that it can be reused 1799 // after the call, if this call has restricted memory effects. 1800 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1801 // Set fixed predefined input arguments 1802 Node* memory = reset_memory(); 1803 Node* m = narrow_mem == NULL ? memory : narrow_mem; 1804 call->init_req( TypeFunc::Control, control() ); 1805 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1806 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1807 call->init_req( TypeFunc::FramePtr, frameptr() ); 1808 call->init_req( TypeFunc::ReturnAdr, top() ); 1809 return memory; 1810 } 1811 1812 //-------------------set_predefined_output_for_runtime_call-------------------- 1813 // Set control and memory (not i_o) from the call. 1814 // If keep_mem is not NULL, use it for the output state, 1815 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1816 // If hook_mem is NULL, this call produces no memory effects at all. 1817 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1818 // then only that memory slice is taken from the call. 1819 // In the last case, we must put an appropriate memory barrier before 1820 // the call, so as to create the correct anti-dependencies on loads 1821 // preceding the call. 1822 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1823 Node* keep_mem, 1824 const TypePtr* hook_mem) { 1825 // no i/o 1826 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1827 if (keep_mem) { 1828 // First clone the existing memory state 1829 set_all_memory(keep_mem); 1830 if (hook_mem != NULL) { 1831 // Make memory for the call 1832 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1833 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1834 // We also use hook_mem to extract specific effects from arraycopy stubs. 1835 set_memory(mem, hook_mem); 1836 } 1837 // ...else the call has NO memory effects. 1838 1839 // Make sure the call advertises its memory effects precisely. 1840 // This lets us build accurate anti-dependences in gcm.cpp. 1841 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1842 "call node must be constructed correctly"); 1843 } else { 1844 assert(hook_mem == NULL, ""); 1845 // This is not a "slow path" call; all memory comes from the call. 1846 set_all_memory_call(call); 1847 } 1848 } 1849 1850 1851 // Replace the call with the current state of the kit. 1852 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1853 JVMState* ejvms = NULL; 1854 if (has_exceptions()) { 1855 ejvms = transfer_exceptions_into_jvms(); 1856 } 1857 1858 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1859 ReplacedNodes replaced_nodes_exception; 1860 Node* ex_ctl = top(); 1861 1862 SafePointNode* final_state = stop(); 1863 1864 // Find all the needed outputs of this call 1865 CallProjections callprojs; 1866 call->extract_projections(&callprojs, true); 1867 1868 Node* init_mem = call->in(TypeFunc::Memory); 1869 Node* final_mem = final_state->in(TypeFunc::Memory); 1870 Node* final_ctl = final_state->in(TypeFunc::Control); 1871 Node* final_io = final_state->in(TypeFunc::I_O); 1872 1873 // Replace all the old call edges with the edges from the inlining result 1874 if (callprojs.fallthrough_catchproj != NULL) { 1875 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1876 } 1877 if (callprojs.fallthrough_memproj != NULL) { 1878 if (final_mem->is_MergeMem()) { 1879 // Parser's exits MergeMem was not transformed but may be optimized 1880 final_mem = _gvn.transform(final_mem); 1881 } 1882 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1883 } 1884 if (callprojs.fallthrough_ioproj != NULL) { 1885 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1886 } 1887 1888 // Replace the result with the new result if it exists and is used 1889 if (callprojs.resproj != NULL && result != NULL) { 1890 C->gvn_replace_by(callprojs.resproj, result); 1891 } 1892 1893 if (ejvms == NULL) { 1894 // No exception edges to simply kill off those paths 1895 if (callprojs.catchall_catchproj != NULL) { 1896 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1897 } 1898 if (callprojs.catchall_memproj != NULL) { 1899 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1900 } 1901 if (callprojs.catchall_ioproj != NULL) { 1902 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1903 } 1904 // Replace the old exception object with top 1905 if (callprojs.exobj != NULL) { 1906 C->gvn_replace_by(callprojs.exobj, C->top()); 1907 } 1908 } else { 1909 GraphKit ekit(ejvms); 1910 1911 // Load my combined exception state into the kit, with all phis transformed: 1912 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1913 replaced_nodes_exception = ex_map->replaced_nodes(); 1914 1915 Node* ex_oop = ekit.use_exception_state(ex_map); 1916 1917 if (callprojs.catchall_catchproj != NULL) { 1918 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1919 ex_ctl = ekit.control(); 1920 } 1921 if (callprojs.catchall_memproj != NULL) { 1922 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1923 } 1924 if (callprojs.catchall_ioproj != NULL) { 1925 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1926 } 1927 1928 // Replace the old exception object with the newly created one 1929 if (callprojs.exobj != NULL) { 1930 C->gvn_replace_by(callprojs.exobj, ex_oop); 1931 } 1932 } 1933 1934 // Disconnect the call from the graph 1935 call->disconnect_inputs(NULL, C); 1936 C->gvn_replace_by(call, C->top()); 1937 1938 // Clean up any MergeMems that feed other MergeMems since the 1939 // optimizer doesn't like that. 1940 if (final_mem->is_MergeMem()) { 1941 Node_List wl; 1942 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1943 Node* m = i.get(); 1944 if (m->is_MergeMem() && !wl.contains(m)) { 1945 wl.push(m); 1946 } 1947 } 1948 while (wl.size() > 0) { 1949 _gvn.transform(wl.pop()); 1950 } 1951 } 1952 1953 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1954 replaced_nodes.apply(C, final_ctl); 1955 } 1956 if (!ex_ctl->is_top() && do_replaced_nodes) { 1957 replaced_nodes_exception.apply(C, ex_ctl); 1958 } 1959 } 1960 1961 1962 //------------------------------increment_counter------------------------------ 1963 // for statistics: increment a VM counter by 1 1964 1965 void GraphKit::increment_counter(address counter_addr) { 1966 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1967 increment_counter(adr1); 1968 } 1969 1970 void GraphKit::increment_counter(Node* counter_addr) { 1971 int adr_type = Compile::AliasIdxRaw; 1972 Node* ctrl = control(); 1973 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1974 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1975 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1976 } 1977 1978 1979 //------------------------------uncommon_trap---------------------------------- 1980 // Bail out to the interpreter in mid-method. Implemented by calling the 1981 // uncommon_trap blob. This helper function inserts a runtime call with the 1982 // right debug info. 1983 void GraphKit::uncommon_trap(int trap_request, 1984 ciKlass* klass, const char* comment, 1985 bool must_throw, 1986 bool keep_exact_action) { 1987 if (failing()) stop(); 1988 if (stopped()) return; // trap reachable? 1989 1990 // Note: If ProfileTraps is true, and if a deopt. actually 1991 // occurs here, the runtime will make sure an MDO exists. There is 1992 // no need to call method()->ensure_method_data() at this point. 1993 1994 // Set the stack pointer to the right value for reexecution: 1995 set_sp(reexecute_sp()); 1996 1997 #ifdef ASSERT 1998 if (!must_throw) { 1999 // Make sure the stack has at least enough depth to execute 2000 // the current bytecode. 2001 int inputs, ignored_depth; 2002 if (compute_stack_effects(inputs, ignored_depth)) { 2003 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2004 Bytecodes::name(java_bc()), sp(), inputs)); 2005 } 2006 } 2007 #endif 2008 2009 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2010 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2011 2012 switch (action) { 2013 case Deoptimization::Action_maybe_recompile: 2014 case Deoptimization::Action_reinterpret: 2015 // Temporary fix for 6529811 to allow virtual calls to be sure they 2016 // get the chance to go from mono->bi->mega 2017 if (!keep_exact_action && 2018 Deoptimization::trap_request_index(trap_request) < 0 && 2019 too_many_recompiles(reason)) { 2020 // This BCI is causing too many recompilations. 2021 if (C->log() != NULL) { 2022 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2023 Deoptimization::trap_reason_name(reason), 2024 Deoptimization::trap_action_name(action)); 2025 } 2026 action = Deoptimization::Action_none; 2027 trap_request = Deoptimization::make_trap_request(reason, action); 2028 } else { 2029 C->set_trap_can_recompile(true); 2030 } 2031 break; 2032 case Deoptimization::Action_make_not_entrant: 2033 C->set_trap_can_recompile(true); 2034 break; 2035 #ifdef ASSERT 2036 case Deoptimization::Action_none: 2037 case Deoptimization::Action_make_not_compilable: 2038 break; 2039 default: 2040 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2041 break; 2042 #endif 2043 } 2044 2045 if (TraceOptoParse) { 2046 char buf[100]; 2047 tty->print_cr("Uncommon trap %s at bci:%d", 2048 Deoptimization::format_trap_request(buf, sizeof(buf), 2049 trap_request), bci()); 2050 } 2051 2052 CompileLog* log = C->log(); 2053 if (log != NULL) { 2054 int kid = (klass == NULL)? -1: log->identify(klass); 2055 log->begin_elem("uncommon_trap bci='%d'", bci()); 2056 char buf[100]; 2057 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2058 trap_request)); 2059 if (kid >= 0) log->print(" klass='%d'", kid); 2060 if (comment != NULL) log->print(" comment='%s'", comment); 2061 log->end_elem(); 2062 } 2063 2064 // Make sure any guarding test views this path as very unlikely 2065 Node *i0 = control()->in(0); 2066 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2067 IfNode *iff = i0->as_If(); 2068 float f = iff->_prob; // Get prob 2069 if (control()->Opcode() == Op_IfTrue) { 2070 if (f > PROB_UNLIKELY_MAG(4)) 2071 iff->_prob = PROB_MIN; 2072 } else { 2073 if (f < PROB_LIKELY_MAG(4)) 2074 iff->_prob = PROB_MAX; 2075 } 2076 } 2077 2078 // Clear out dead values from the debug info. 2079 kill_dead_locals(); 2080 2081 // Now insert the uncommon trap subroutine call 2082 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2083 const TypePtr* no_memory_effects = NULL; 2084 // Pass the index of the class to be loaded 2085 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2086 (must_throw ? RC_MUST_THROW : 0), 2087 OptoRuntime::uncommon_trap_Type(), 2088 call_addr, "uncommon_trap", no_memory_effects, 2089 intcon(trap_request)); 2090 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2091 "must extract request correctly from the graph"); 2092 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2093 2094 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2095 // The debug info is the only real input to this call. 2096 2097 // Halt-and-catch fire here. The above call should never return! 2098 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2099 _gvn.set_type_bottom(halt); 2100 root()->add_req(halt); 2101 2102 stop_and_kill_map(); 2103 } 2104 2105 2106 //--------------------------just_allocated_object------------------------------ 2107 // Report the object that was just allocated. 2108 // It must be the case that there are no intervening safepoints. 2109 // We use this to determine if an object is so "fresh" that 2110 // it does not require card marks. 2111 Node* GraphKit::just_allocated_object(Node* current_control) { 2112 if (C->recent_alloc_ctl() == current_control) 2113 return C->recent_alloc_obj(); 2114 return NULL; 2115 } 2116 2117 2118 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2119 // (Note: TypeFunc::make has a cache that makes this fast.) 2120 const TypeFunc* tf = TypeFunc::make(dest_method); 2121 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2122 for (int j = 0; j < nargs; j++) { 2123 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2124 if( targ->basic_type() == T_DOUBLE ) { 2125 // If any parameters are doubles, they must be rounded before 2126 // the call, dstore_rounding does gvn.transform 2127 Node *arg = argument(j); 2128 arg = dstore_rounding(arg); 2129 set_argument(j, arg); 2130 } 2131 } 2132 } 2133 2134 /** 2135 * Record profiling data exact_kls for Node n with the type system so 2136 * that it can propagate it (speculation) 2137 * 2138 * @param n node that the type applies to 2139 * @param exact_kls type from profiling 2140 * 2141 * @return node with improved type 2142 */ 2143 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) { 2144 const Type* current_type = _gvn.type(n); 2145 assert(UseTypeSpeculation, "type speculation must be on"); 2146 2147 const TypeOopPtr* speculative = current_type->speculative(); 2148 2149 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2150 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2151 const TypeOopPtr* xtype = tklass->as_instance_type(); 2152 assert(xtype->klass_is_exact(), "Should be exact"); 2153 // record the new speculative type's depth 2154 speculative = xtype->with_inline_depth(jvms()->depth()); 2155 } 2156 2157 if (speculative != current_type->speculative()) { 2158 // Build a type with a speculative type (what we think we know 2159 // about the type but will need a guard when we use it) 2160 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2161 // We're changing the type, we need a new CheckCast node to carry 2162 // the new type. The new type depends on the control: what 2163 // profiling tells us is only valid from here as far as we can 2164 // tell. 2165 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2166 cast = _gvn.transform(cast); 2167 replace_in_map(n, cast); 2168 n = cast; 2169 } 2170 2171 return n; 2172 } 2173 2174 /** 2175 * Record profiling data from receiver profiling at an invoke with the 2176 * type system so that it can propagate it (speculation) 2177 * 2178 * @param n receiver node 2179 * 2180 * @return node with improved type 2181 */ 2182 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2183 if (!UseTypeSpeculation) { 2184 return n; 2185 } 2186 ciKlass* exact_kls = profile_has_unique_klass(); 2187 return record_profile_for_speculation(n, exact_kls); 2188 } 2189 2190 /** 2191 * Record profiling data from argument profiling at an invoke with the 2192 * type system so that it can propagate it (speculation) 2193 * 2194 * @param dest_method target method for the call 2195 * @param bc what invoke bytecode is this? 2196 */ 2197 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2198 if (!UseTypeSpeculation) { 2199 return; 2200 } 2201 const TypeFunc* tf = TypeFunc::make(dest_method); 2202 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2203 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2204 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2205 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2206 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2207 ciKlass* better_type = method()->argument_profiled_type(bci(), i); 2208 if (better_type != NULL) { 2209 record_profile_for_speculation(argument(j), better_type); 2210 } 2211 i++; 2212 } 2213 } 2214 } 2215 2216 /** 2217 * Record profiling data from parameter profiling at an invoke with 2218 * the type system so that it can propagate it (speculation) 2219 */ 2220 void GraphKit::record_profiled_parameters_for_speculation() { 2221 if (!UseTypeSpeculation) { 2222 return; 2223 } 2224 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2225 if (_gvn.type(local(i))->isa_oopptr()) { 2226 ciKlass* better_type = method()->parameter_profiled_type(j); 2227 if (better_type != NULL) { 2228 record_profile_for_speculation(local(i), better_type); 2229 } 2230 j++; 2231 } 2232 } 2233 } 2234 2235 void GraphKit::round_double_result(ciMethod* dest_method) { 2236 // A non-strict method may return a double value which has an extended 2237 // exponent, but this must not be visible in a caller which is 'strict' 2238 // If a strict caller invokes a non-strict callee, round a double result 2239 2240 BasicType result_type = dest_method->return_type()->basic_type(); 2241 assert( method() != NULL, "must have caller context"); 2242 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2243 // Destination method's return value is on top of stack 2244 // dstore_rounding() does gvn.transform 2245 Node *result = pop_pair(); 2246 result = dstore_rounding(result); 2247 push_pair(result); 2248 } 2249 } 2250 2251 // rounding for strict float precision conformance 2252 Node* GraphKit::precision_rounding(Node* n) { 2253 return UseStrictFP && _method->flags().is_strict() 2254 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2255 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2256 : n; 2257 } 2258 2259 // rounding for strict double precision conformance 2260 Node* GraphKit::dprecision_rounding(Node *n) { 2261 return UseStrictFP && _method->flags().is_strict() 2262 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2263 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2264 : n; 2265 } 2266 2267 // rounding for non-strict double stores 2268 Node* GraphKit::dstore_rounding(Node* n) { 2269 return Matcher::strict_fp_requires_explicit_rounding 2270 && UseSSE <= 1 2271 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2272 : n; 2273 } 2274 2275 //============================================================================= 2276 // Generate a fast path/slow path idiom. Graph looks like: 2277 // [foo] indicates that 'foo' is a parameter 2278 // 2279 // [in] NULL 2280 // \ / 2281 // CmpP 2282 // Bool ne 2283 // If 2284 // / \ 2285 // True False-<2> 2286 // / | 2287 // / cast_not_null 2288 // Load | | ^ 2289 // [fast_test] | | 2290 // gvn to opt_test | | 2291 // / \ | <1> 2292 // True False | 2293 // | \\ | 2294 // [slow_call] \[fast_result] 2295 // Ctl Val \ \ 2296 // | \ \ 2297 // Catch <1> \ \ 2298 // / \ ^ \ \ 2299 // Ex No_Ex | \ \ 2300 // | \ \ | \ <2> \ 2301 // ... \ [slow_res] | | \ [null_result] 2302 // \ \--+--+--- | | 2303 // \ | / \ | / 2304 // --------Region Phi 2305 // 2306 //============================================================================= 2307 // Code is structured as a series of driver functions all called 'do_XXX' that 2308 // call a set of helper functions. Helper functions first, then drivers. 2309 2310 //------------------------------null_check_oop--------------------------------- 2311 // Null check oop. Set null-path control into Region in slot 3. 2312 // Make a cast-not-nullness use the other not-null control. Return cast. 2313 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2314 bool never_see_null, bool safe_for_replace) { 2315 // Initial NULL check taken path 2316 (*null_control) = top(); 2317 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2318 2319 // Generate uncommon_trap: 2320 if (never_see_null && (*null_control) != top()) { 2321 // If we see an unexpected null at a check-cast we record it and force a 2322 // recompile; the offending check-cast will be compiled to handle NULLs. 2323 // If we see more than one offending BCI, then all checkcasts in the 2324 // method will be compiled to handle NULLs. 2325 PreserveJVMState pjvms(this); 2326 set_control(*null_control); 2327 replace_in_map(value, null()); 2328 uncommon_trap(Deoptimization::Reason_null_check, 2329 Deoptimization::Action_make_not_entrant); 2330 (*null_control) = top(); // NULL path is dead 2331 } 2332 if ((*null_control) == top() && safe_for_replace) { 2333 replace_in_map(value, cast); 2334 } 2335 2336 // Cast away null-ness on the result 2337 return cast; 2338 } 2339 2340 //------------------------------opt_iff---------------------------------------- 2341 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2342 // Return slow-path control. 2343 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2344 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2345 2346 // Fast path taken; set region slot 2 2347 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2348 region->init_req(2,fast_taken); // Capture fast-control 2349 2350 // Fast path not-taken, i.e. slow path 2351 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2352 return slow_taken; 2353 } 2354 2355 //-----------------------------make_runtime_call------------------------------- 2356 Node* GraphKit::make_runtime_call(int flags, 2357 const TypeFunc* call_type, address call_addr, 2358 const char* call_name, 2359 const TypePtr* adr_type, 2360 // The following parms are all optional. 2361 // The first NULL ends the list. 2362 Node* parm0, Node* parm1, 2363 Node* parm2, Node* parm3, 2364 Node* parm4, Node* parm5, 2365 Node* parm6, Node* parm7) { 2366 // Slow-path call 2367 bool is_leaf = !(flags & RC_NO_LEAF); 2368 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2369 if (call_name == NULL) { 2370 assert(!is_leaf, "must supply name for leaf"); 2371 call_name = OptoRuntime::stub_name(call_addr); 2372 } 2373 CallNode* call; 2374 if (!is_leaf) { 2375 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2376 bci(), adr_type); 2377 } else if (flags & RC_NO_FP) { 2378 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2379 } else { 2380 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2381 } 2382 2383 // The following is similar to set_edges_for_java_call, 2384 // except that the memory effects of the call are restricted to AliasIdxRaw. 2385 2386 // Slow path call has no side-effects, uses few values 2387 bool wide_in = !(flags & RC_NARROW_MEM); 2388 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2389 2390 Node* prev_mem = NULL; 2391 if (wide_in) { 2392 prev_mem = set_predefined_input_for_runtime_call(call); 2393 } else { 2394 assert(!wide_out, "narrow in => narrow out"); 2395 Node* narrow_mem = memory(adr_type); 2396 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2397 } 2398 2399 // Hook each parm in order. Stop looking at the first NULL. 2400 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2401 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2402 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2403 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2404 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2405 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2406 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2407 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2408 /* close each nested if ===> */ } } } } } } } } 2409 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2410 2411 if (!is_leaf) { 2412 // Non-leaves can block and take safepoints: 2413 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2414 } 2415 // Non-leaves can throw exceptions: 2416 if (has_io) { 2417 call->set_req(TypeFunc::I_O, i_o()); 2418 } 2419 2420 if (flags & RC_UNCOMMON) { 2421 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2422 // (An "if" probability corresponds roughly to an unconditional count. 2423 // Sort of.) 2424 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2425 } 2426 2427 Node* c = _gvn.transform(call); 2428 assert(c == call, "cannot disappear"); 2429 2430 if (wide_out) { 2431 // Slow path call has full side-effects. 2432 set_predefined_output_for_runtime_call(call); 2433 } else { 2434 // Slow path call has few side-effects, and/or sets few values. 2435 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2436 } 2437 2438 if (has_io) { 2439 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2440 } 2441 return call; 2442 2443 } 2444 2445 //------------------------------merge_memory----------------------------------- 2446 // Merge memory from one path into the current memory state. 2447 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2448 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2449 Node* old_slice = mms.force_memory(); 2450 Node* new_slice = mms.memory2(); 2451 if (old_slice != new_slice) { 2452 PhiNode* phi; 2453 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2454 if (mms.is_empty()) { 2455 // clone base memory Phi's inputs for this memory slice 2456 assert(old_slice == mms.base_memory(), "sanity"); 2457 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2458 _gvn.set_type(phi, Type::MEMORY); 2459 for (uint i = 1; i < phi->req(); i++) { 2460 phi->init_req(i, old_slice->in(i)); 2461 } 2462 } else { 2463 phi = old_slice->as_Phi(); // Phi was generated already 2464 } 2465 } else { 2466 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2467 _gvn.set_type(phi, Type::MEMORY); 2468 } 2469 phi->set_req(new_path, new_slice); 2470 mms.set_memory(phi); 2471 } 2472 } 2473 } 2474 2475 //------------------------------make_slow_call_ex------------------------------ 2476 // Make the exception handler hookups for the slow call 2477 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2478 if (stopped()) return; 2479 2480 // Make a catch node with just two handlers: fall-through and catch-all 2481 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2482 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2483 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2484 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2485 2486 { PreserveJVMState pjvms(this); 2487 set_control(excp); 2488 set_i_o(i_o); 2489 2490 if (excp != top()) { 2491 if (deoptimize) { 2492 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2493 uncommon_trap(Deoptimization::Reason_unhandled, 2494 Deoptimization::Action_none); 2495 } else { 2496 // Create an exception state also. 2497 // Use an exact type if the caller has specified a specific exception. 2498 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2499 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2500 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2501 } 2502 } 2503 } 2504 2505 // Get the no-exception control from the CatchNode. 2506 set_control(norm); 2507 } 2508 2509 2510 //-------------------------------gen_subtype_check----------------------------- 2511 // Generate a subtyping check. Takes as input the subtype and supertype. 2512 // Returns 2 values: sets the default control() to the true path and returns 2513 // the false path. Only reads invariant memory; sets no (visible) memory. 2514 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2515 // but that's not exposed to the optimizer. This call also doesn't take in an 2516 // Object; if you wish to check an Object you need to load the Object's class 2517 // prior to coming here. 2518 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2519 // Fast check for identical types, perhaps identical constants. 2520 // The types can even be identical non-constants, in cases 2521 // involving Array.newInstance, Object.clone, etc. 2522 if (subklass == superklass) 2523 return top(); // false path is dead; no test needed. 2524 2525 if (_gvn.type(superklass)->singleton()) { 2526 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2527 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2528 2529 // In the common case of an exact superklass, try to fold up the 2530 // test before generating code. You may ask, why not just generate 2531 // the code and then let it fold up? The answer is that the generated 2532 // code will necessarily include null checks, which do not always 2533 // completely fold away. If they are also needless, then they turn 2534 // into a performance loss. Example: 2535 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2536 // Here, the type of 'fa' is often exact, so the store check 2537 // of fa[1]=x will fold up, without testing the nullness of x. 2538 switch (static_subtype_check(superk, subk)) { 2539 case SSC_always_false: 2540 { 2541 Node* always_fail = control(); 2542 set_control(top()); 2543 return always_fail; 2544 } 2545 case SSC_always_true: 2546 return top(); 2547 case SSC_easy_test: 2548 { 2549 // Just do a direct pointer compare and be done. 2550 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2551 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2552 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2553 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2554 return _gvn.transform( new(C) IfFalseNode(iff) ); 2555 } 2556 case SSC_full_test: 2557 break; 2558 default: 2559 ShouldNotReachHere(); 2560 } 2561 } 2562 2563 // %%% Possible further optimization: Even if the superklass is not exact, 2564 // if the subklass is the unique subtype of the superklass, the check 2565 // will always succeed. We could leave a dependency behind to ensure this. 2566 2567 // First load the super-klass's check-offset 2568 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2569 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), 2570 TypeInt::INT, MemNode::unordered)); 2571 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2572 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2573 2574 // Load from the sub-klass's super-class display list, or a 1-word cache of 2575 // the secondary superclass list, or a failing value with a sentinel offset 2576 // if the super-klass is an interface or exceptionally deep in the Java 2577 // hierarchy and we have to scan the secondary superclass list the hard way. 2578 // Worst-case type is a little odd: NULL is allowed as a result (usually 2579 // klass loads can never produce a NULL). 2580 Node *chk_off_X = ConvI2X(chk_off); 2581 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2582 // For some types like interfaces the following loadKlass is from a 1-word 2583 // cache which is mutable so can't use immutable memory. Other 2584 // types load from the super-class display table which is immutable. 2585 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2586 Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2587 2588 // Compile speed common case: ARE a subtype and we canNOT fail 2589 if( superklass == nkls ) 2590 return top(); // false path is dead; no test needed. 2591 2592 // See if we get an immediate positive hit. Happens roughly 83% of the 2593 // time. Test to see if the value loaded just previously from the subklass 2594 // is exactly the superklass. 2595 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2596 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2597 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2598 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2599 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2600 2601 // Compile speed common case: Check for being deterministic right now. If 2602 // chk_off is a constant and not equal to cacheoff then we are NOT a 2603 // subklass. In this case we need exactly the 1 test above and we can 2604 // return those results immediately. 2605 if (!might_be_cache) { 2606 Node* not_subtype_ctrl = control(); 2607 set_control(iftrue1); // We need exactly the 1 test above 2608 return not_subtype_ctrl; 2609 } 2610 2611 // Gather the various success & failures here 2612 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2613 record_for_igvn(r_ok_subtype); 2614 RegionNode *r_not_subtype = new (C) RegionNode(3); 2615 record_for_igvn(r_not_subtype); 2616 2617 r_ok_subtype->init_req(1, iftrue1); 2618 2619 // Check for immediate negative hit. Happens roughly 11% of the time (which 2620 // is roughly 63% of the remaining cases). Test to see if the loaded 2621 // check-offset points into the subklass display list or the 1-element 2622 // cache. If it points to the display (and NOT the cache) and the display 2623 // missed then it's not a subtype. 2624 Node *cacheoff = _gvn.intcon(cacheoff_con); 2625 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2626 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2627 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2628 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2629 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2630 2631 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2632 // No performance impact (too rare) but allows sharing of secondary arrays 2633 // which has some footprint reduction. 2634 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2635 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2636 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2637 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2638 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2639 2640 // -- Roads not taken here: -- 2641 // We could also have chosen to perform the self-check at the beginning 2642 // of this code sequence, as the assembler does. This would not pay off 2643 // the same way, since the optimizer, unlike the assembler, can perform 2644 // static type analysis to fold away many successful self-checks. 2645 // Non-foldable self checks work better here in second position, because 2646 // the initial primary superclass check subsumes a self-check for most 2647 // types. An exception would be a secondary type like array-of-interface, 2648 // which does not appear in its own primary supertype display. 2649 // Finally, we could have chosen to move the self-check into the 2650 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2651 // dependent manner. But it is worthwhile to have the check here, 2652 // where it can be perhaps be optimized. The cost in code space is 2653 // small (register compare, branch). 2654 2655 // Now do a linear scan of the secondary super-klass array. Again, no real 2656 // performance impact (too rare) but it's gotta be done. 2657 // Since the code is rarely used, there is no penalty for moving it 2658 // out of line, and it can only improve I-cache density. 2659 // The decision to inline or out-of-line this final check is platform 2660 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2661 Node* psc = _gvn.transform( 2662 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2663 2664 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2665 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2666 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2667 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2668 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2669 2670 // Return false path; set default control to true path. 2671 set_control( _gvn.transform(r_ok_subtype) ); 2672 return _gvn.transform(r_not_subtype); 2673 } 2674 2675 //----------------------------static_subtype_check----------------------------- 2676 // Shortcut important common cases when superklass is exact: 2677 // (0) superklass is java.lang.Object (can occur in reflective code) 2678 // (1) subklass is already limited to a subtype of superklass => always ok 2679 // (2) subklass does not overlap with superklass => always fail 2680 // (3) superklass has NO subtypes and we can check with a simple compare. 2681 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2682 if (StressReflectiveCode) { 2683 return SSC_full_test; // Let caller generate the general case. 2684 } 2685 2686 if (superk == env()->Object_klass()) { 2687 return SSC_always_true; // (0) this test cannot fail 2688 } 2689 2690 ciType* superelem = superk; 2691 if (superelem->is_array_klass()) 2692 superelem = superelem->as_array_klass()->base_element_type(); 2693 2694 if (!subk->is_interface()) { // cannot trust static interface types yet 2695 if (subk->is_subtype_of(superk)) { 2696 return SSC_always_true; // (1) false path dead; no dynamic test needed 2697 } 2698 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2699 !superk->is_subtype_of(subk)) { 2700 return SSC_always_false; 2701 } 2702 } 2703 2704 // If casting to an instance klass, it must have no subtypes 2705 if (superk->is_interface()) { 2706 // Cannot trust interfaces yet. 2707 // %%% S.B. superk->nof_implementors() == 1 2708 } else if (superelem->is_instance_klass()) { 2709 ciInstanceKlass* ik = superelem->as_instance_klass(); 2710 if (!ik->has_subklass() && !ik->is_interface()) { 2711 if (!ik->is_final()) { 2712 // Add a dependency if there is a chance of a later subclass. 2713 C->dependencies()->assert_leaf_type(ik); 2714 } 2715 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2716 } 2717 } else { 2718 // A primitive array type has no subtypes. 2719 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2720 } 2721 2722 return SSC_full_test; 2723 } 2724 2725 // Profile-driven exact type check: 2726 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2727 float prob, 2728 Node* *casted_receiver) { 2729 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2730 Node* recv_klass = load_object_klass(receiver); 2731 Node* want_klass = makecon(tklass); 2732 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2733 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2734 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2735 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2736 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2737 2738 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2739 assert(recv_xtype->klass_is_exact(), ""); 2740 2741 // Subsume downstream occurrences of receiver with a cast to 2742 // recv_xtype, since now we know what the type will be. 2743 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2744 (*casted_receiver) = _gvn.transform(cast); 2745 // (User must make the replace_in_map call.) 2746 2747 return fail; 2748 } 2749 2750 2751 //------------------------------seems_never_null------------------------------- 2752 // Use null_seen information if it is available from the profile. 2753 // If we see an unexpected null at a type check we record it and force a 2754 // recompile; the offending check will be recompiled to handle NULLs. 2755 // If we see several offending BCIs, then all checks in the 2756 // method will be recompiled. 2757 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2758 if (UncommonNullCast // Cutout for this technique 2759 && obj != null() // And not the -Xcomp stupid case? 2760 && !too_many_traps(Deoptimization::Reason_null_check) 2761 ) { 2762 if (data == NULL) 2763 // Edge case: no mature data. Be optimistic here. 2764 return true; 2765 // If the profile has not seen a null, assume it won't happen. 2766 assert(java_bc() == Bytecodes::_checkcast || 2767 java_bc() == Bytecodes::_instanceof || 2768 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2769 return !data->as_BitData()->null_seen(); 2770 } 2771 return false; 2772 } 2773 2774 //------------------------maybe_cast_profiled_receiver------------------------- 2775 // If the profile has seen exactly one type, narrow to exactly that type. 2776 // Subsequent type checks will always fold up. 2777 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2778 ciKlass* require_klass, 2779 ciKlass* spec_klass, 2780 bool safe_for_replace) { 2781 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2782 2783 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check; 2784 2785 // Make sure we haven't already deoptimized from this tactic. 2786 if (too_many_traps(reason) || too_many_recompiles(reason)) 2787 return NULL; 2788 2789 // (No, this isn't a call, but it's enough like a virtual call 2790 // to use the same ciMethod accessor to get the profile info...) 2791 // If we have a speculative type use it instead of profiling (which 2792 // may not help us) 2793 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2794 if (exact_kls != NULL) {// no cast failures here 2795 if (require_klass == NULL || 2796 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2797 // If we narrow the type to match what the type profile sees or 2798 // the speculative type, we can then remove the rest of the 2799 // cast. 2800 // This is a win, even if the exact_kls is very specific, 2801 // because downstream operations, such as method calls, 2802 // will often benefit from the sharper type. 2803 Node* exact_obj = not_null_obj; // will get updated in place... 2804 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2805 &exact_obj); 2806 { PreserveJVMState pjvms(this); 2807 set_control(slow_ctl); 2808 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2809 } 2810 if (safe_for_replace) { 2811 replace_in_map(not_null_obj, exact_obj); 2812 } 2813 return exact_obj; 2814 } 2815 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2816 } 2817 2818 return NULL; 2819 } 2820 2821 /** 2822 * Cast obj to type and emit guard unless we had too many traps here 2823 * already 2824 * 2825 * @param obj node being casted 2826 * @param type type to cast the node to 2827 * @param not_null true if we know node cannot be null 2828 */ 2829 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2830 ciKlass* type, 2831 bool not_null) { 2832 // type == NULL if profiling tells us this object is always null 2833 if (type != NULL) { 2834 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2835 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check; 2836 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2837 !too_many_traps(class_reason) && !too_many_recompiles(class_reason)) { 2838 Node* not_null_obj = NULL; 2839 // not_null is true if we know the object is not null and 2840 // there's no need for a null check 2841 if (!not_null) { 2842 Node* null_ctl = top(); 2843 not_null_obj = null_check_oop(obj, &null_ctl, true, true); 2844 assert(null_ctl->is_top(), "no null control here"); 2845 } else { 2846 not_null_obj = obj; 2847 } 2848 2849 Node* exact_obj = not_null_obj; 2850 ciKlass* exact_kls = type; 2851 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2852 &exact_obj); 2853 { 2854 PreserveJVMState pjvms(this); 2855 set_control(slow_ctl); 2856 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2857 } 2858 replace_in_map(not_null_obj, exact_obj); 2859 obj = exact_obj; 2860 } 2861 } else { 2862 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2863 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2864 Node* exact_obj = null_assert(obj); 2865 replace_in_map(obj, exact_obj); 2866 obj = exact_obj; 2867 } 2868 } 2869 return obj; 2870 } 2871 2872 //-------------------------------gen_instanceof-------------------------------- 2873 // Generate an instance-of idiom. Used by both the instance-of bytecode 2874 // and the reflective instance-of call. 2875 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2876 kill_dead_locals(); // Benefit all the uncommon traps 2877 assert( !stopped(), "dead parse path should be checked in callers" ); 2878 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2879 "must check for not-null not-dead klass in callers"); 2880 2881 // Make the merge point 2882 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2883 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2884 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2885 C->set_has_split_ifs(true); // Has chance for split-if optimization 2886 2887 ciProfileData* data = NULL; 2888 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2889 data = method()->method_data()->bci_to_data(bci()); 2890 } 2891 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2892 && seems_never_null(obj, data)); 2893 2894 // Null check; get casted pointer; set region slot 3 2895 Node* null_ctl = top(); 2896 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2897 2898 // If not_null_obj is dead, only null-path is taken 2899 if (stopped()) { // Doing instance-of on a NULL? 2900 set_control(null_ctl); 2901 return intcon(0); 2902 } 2903 region->init_req(_null_path, null_ctl); 2904 phi ->init_req(_null_path, intcon(0)); // Set null path value 2905 if (null_ctl == top()) { 2906 // Do this eagerly, so that pattern matches like is_diamond_phi 2907 // will work even during parsing. 2908 assert(_null_path == PATH_LIMIT-1, "delete last"); 2909 region->del_req(_null_path); 2910 phi ->del_req(_null_path); 2911 } 2912 2913 // Do we know the type check always succeed? 2914 bool known_statically = false; 2915 if (_gvn.type(superklass)->singleton()) { 2916 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2917 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2918 if (subk != NULL && subk->is_loaded()) { 2919 int static_res = static_subtype_check(superk, subk); 2920 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2921 } 2922 } 2923 2924 if (known_statically && UseTypeSpeculation) { 2925 // If we know the type check always succeeds then we don't use the 2926 // profiling data at this bytecode. Don't lose it, feed it to the 2927 // type system as a speculative type. 2928 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2929 } else { 2930 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2931 // We may not have profiling here or it may not help us. If we 2932 // have a speculative type use it to perform an exact cast. 2933 ciKlass* spec_obj_type = obj_type->speculative_type(); 2934 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2935 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2936 if (stopped()) { // Profile disagrees with this path. 2937 set_control(null_ctl); // Null is the only remaining possibility. 2938 return intcon(0); 2939 } 2940 if (cast_obj != NULL) { 2941 not_null_obj = cast_obj; 2942 } 2943 } 2944 } 2945 2946 // Load the object's klass 2947 Node* obj_klass = load_object_klass(not_null_obj); 2948 2949 // Generate the subtype check 2950 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2951 2952 // Plug in the success path to the general merge in slot 1. 2953 region->init_req(_obj_path, control()); 2954 phi ->init_req(_obj_path, intcon(1)); 2955 2956 // Plug in the failing path to the general merge in slot 2. 2957 region->init_req(_fail_path, not_subtype_ctrl); 2958 phi ->init_req(_fail_path, intcon(0)); 2959 2960 // Return final merged results 2961 set_control( _gvn.transform(region) ); 2962 record_for_igvn(region); 2963 return _gvn.transform(phi); 2964 } 2965 2966 //-------------------------------gen_checkcast--------------------------------- 2967 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2968 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2969 // uncommon-trap paths work. Adjust stack after this call. 2970 // If failure_control is supplied and not null, it is filled in with 2971 // the control edge for the cast failure. Otherwise, an appropriate 2972 // uncommon trap or exception is thrown. 2973 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2974 Node* *failure_control) { 2975 kill_dead_locals(); // Benefit all the uncommon traps 2976 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2977 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2978 2979 // Fast cutout: Check the case that the cast is vacuously true. 2980 // This detects the common cases where the test will short-circuit 2981 // away completely. We do this before we perform the null check, 2982 // because if the test is going to turn into zero code, we don't 2983 // want a residual null check left around. (Causes a slowdown, 2984 // for example, in some objArray manipulations, such as a[i]=a[j].) 2985 if (tk->singleton()) { 2986 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2987 if (objtp != NULL && objtp->klass() != NULL) { 2988 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2989 case SSC_always_true: 2990 // If we know the type check always succeed then we don't use 2991 // the profiling data at this bytecode. Don't lose it, feed it 2992 // to the type system as a speculative type. 2993 return record_profiled_receiver_for_speculation(obj); 2994 case SSC_always_false: 2995 // It needs a null check because a null will *pass* the cast check. 2996 // A non-null value will always produce an exception. 2997 return null_assert(obj); 2998 } 2999 } 3000 } 3001 3002 ciProfileData* data = NULL; 3003 bool safe_for_replace = false; 3004 if (failure_control == NULL) { // use MDO in regular case only 3005 assert(java_bc() == Bytecodes::_aastore || 3006 java_bc() == Bytecodes::_checkcast, 3007 "interpreter profiles type checks only for these BCs"); 3008 data = method()->method_data()->bci_to_data(bci()); 3009 safe_for_replace = true; 3010 } 3011 3012 // Make the merge point 3013 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3014 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3015 Node* phi = new (C) PhiNode(region, toop); 3016 C->set_has_split_ifs(true); // Has chance for split-if optimization 3017 3018 // Use null-cast information if it is available 3019 bool never_see_null = ((failure_control == NULL) // regular case only 3020 && seems_never_null(obj, data)); 3021 3022 // Null check; get casted pointer; set region slot 3 3023 Node* null_ctl = top(); 3024 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 3025 3026 // If not_null_obj is dead, only null-path is taken 3027 if (stopped()) { // Doing instance-of on a NULL? 3028 set_control(null_ctl); 3029 return null(); 3030 } 3031 region->init_req(_null_path, null_ctl); 3032 phi ->init_req(_null_path, null()); // Set null path value 3033 if (null_ctl == top()) { 3034 // Do this eagerly, so that pattern matches like is_diamond_phi 3035 // will work even during parsing. 3036 assert(_null_path == PATH_LIMIT-1, "delete last"); 3037 region->del_req(_null_path); 3038 phi ->del_req(_null_path); 3039 } 3040 3041 Node* cast_obj = NULL; 3042 if (tk->klass_is_exact()) { 3043 // The following optimization tries to statically cast the speculative type of the object 3044 // (for example obtained during profiling) to the type of the superklass and then do a 3045 // dynamic check that the type of the object is what we expect. To work correctly 3046 // for checkcast and aastore the type of superklass should be exact. 3047 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3048 // We may not have profiling here or it may not help us. If we have 3049 // a speculative type use it to perform an exact cast. 3050 ciKlass* spec_obj_type = obj_type->speculative_type(); 3051 if (spec_obj_type != NULL || 3052 (data != NULL && 3053 // Counter has never been decremented (due to cast failure). 3054 // ...This is a reasonable thing to expect. It is true of 3055 // all casts inserted by javac to implement generic types. 3056 data->as_CounterData()->count() >= 0)) { 3057 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3058 if (cast_obj != NULL) { 3059 if (failure_control != NULL) // failure is now impossible 3060 (*failure_control) = top(); 3061 // adjust the type of the phi to the exact klass: 3062 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3063 } 3064 } 3065 } 3066 3067 if (cast_obj == NULL) { 3068 // Load the object's klass 3069 Node* obj_klass = load_object_klass(not_null_obj); 3070 3071 // Generate the subtype check 3072 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3073 3074 // Plug in success path into the merge 3075 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 3076 not_null_obj, toop)); 3077 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3078 if (failure_control == NULL) { 3079 if (not_subtype_ctrl != top()) { // If failure is possible 3080 PreserveJVMState pjvms(this); 3081 set_control(not_subtype_ctrl); 3082 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3083 } 3084 } else { 3085 (*failure_control) = not_subtype_ctrl; 3086 } 3087 } 3088 3089 region->init_req(_obj_path, control()); 3090 phi ->init_req(_obj_path, cast_obj); 3091 3092 // A merge of NULL or Casted-NotNull obj 3093 Node* res = _gvn.transform(phi); 3094 3095 // Note I do NOT always 'replace_in_map(obj,result)' here. 3096 // if( tk->klass()->can_be_primary_super() ) 3097 // This means that if I successfully store an Object into an array-of-String 3098 // I 'forget' that the Object is really now known to be a String. I have to 3099 // do this because we don't have true union types for interfaces - if I store 3100 // a Baz into an array-of-Interface and then tell the optimizer it's an 3101 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3102 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3103 // replace_in_map( obj, res ); 3104 3105 // Return final merged results 3106 set_control( _gvn.transform(region) ); 3107 record_for_igvn(region); 3108 return res; 3109 } 3110 3111 //------------------------------next_monitor----------------------------------- 3112 // What number should be given to the next monitor? 3113 int GraphKit::next_monitor() { 3114 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3115 int next = current + C->sync_stack_slots(); 3116 // Keep the toplevel high water mark current: 3117 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3118 return current; 3119 } 3120 3121 //------------------------------insert_mem_bar--------------------------------- 3122 // Memory barrier to avoid floating things around 3123 // The membar serves as a pinch point between both control and all memory slices. 3124 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3125 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3126 mb->init_req(TypeFunc::Control, control()); 3127 mb->init_req(TypeFunc::Memory, reset_memory()); 3128 Node* membar = _gvn.transform(mb); 3129 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3130 set_all_memory_call(membar); 3131 return membar; 3132 } 3133 3134 //-------------------------insert_mem_bar_volatile---------------------------- 3135 // Memory barrier to avoid floating things around 3136 // The membar serves as a pinch point between both control and memory(alias_idx). 3137 // If you want to make a pinch point on all memory slices, do not use this 3138 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3139 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3140 // When Parse::do_put_xxx updates a volatile field, it appends a series 3141 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3142 // The first membar is on the same memory slice as the field store opcode. 3143 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3144 // All the other membars (for other volatile slices, including AliasIdxBot, 3145 // which stands for all unknown volatile slices) are control-dependent 3146 // on the first membar. This prevents later volatile loads or stores 3147 // from sliding up past the just-emitted store. 3148 3149 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3150 mb->set_req(TypeFunc::Control,control()); 3151 if (alias_idx == Compile::AliasIdxBot) { 3152 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3153 } else { 3154 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3155 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3156 } 3157 Node* membar = _gvn.transform(mb); 3158 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3159 if (alias_idx == Compile::AliasIdxBot) { 3160 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 3161 } else { 3162 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 3163 } 3164 return membar; 3165 } 3166 3167 //------------------------------shared_lock------------------------------------ 3168 // Emit locking code. 3169 FastLockNode* GraphKit::shared_lock(Node* obj) { 3170 // bci is either a monitorenter bc or InvocationEntryBci 3171 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3172 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3173 3174 if( !GenerateSynchronizationCode ) 3175 return NULL; // Not locking things? 3176 if (stopped()) // Dead monitor? 3177 return NULL; 3178 3179 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3180 3181 // Box the stack location 3182 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 3183 Node* mem = reset_memory(); 3184 3185 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 3186 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3187 // Create the counters for this fast lock. 3188 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3189 } 3190 3191 // Create the rtm counters for this fast lock if needed. 3192 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3193 3194 // Add monitor to debug info for the slow path. If we block inside the 3195 // slow path and de-opt, we need the monitor hanging around 3196 map()->push_monitor( flock ); 3197 3198 const TypeFunc *tf = LockNode::lock_type(); 3199 LockNode *lock = new (C) LockNode(C, tf); 3200 3201 lock->init_req( TypeFunc::Control, control() ); 3202 lock->init_req( TypeFunc::Memory , mem ); 3203 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3204 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3205 lock->init_req( TypeFunc::ReturnAdr, top() ); 3206 3207 lock->init_req(TypeFunc::Parms + 0, obj); 3208 lock->init_req(TypeFunc::Parms + 1, box); 3209 lock->init_req(TypeFunc::Parms + 2, flock); 3210 add_safepoint_edges(lock); 3211 3212 lock = _gvn.transform( lock )->as_Lock(); 3213 3214 // lock has no side-effects, sets few values 3215 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3216 3217 insert_mem_bar(Op_MemBarAcquireLock); 3218 3219 // Add this to the worklist so that the lock can be eliminated 3220 record_for_igvn(lock); 3221 3222 #ifndef PRODUCT 3223 if (PrintLockStatistics) { 3224 // Update the counter for this lock. Don't bother using an atomic 3225 // operation since we don't require absolute accuracy. 3226 lock->create_lock_counter(map()->jvms()); 3227 increment_counter(lock->counter()->addr()); 3228 } 3229 #endif 3230 3231 return flock; 3232 } 3233 3234 3235 //------------------------------shared_unlock---------------------------------- 3236 // Emit unlocking code. 3237 void GraphKit::shared_unlock(Node* box, Node* obj) { 3238 // bci is either a monitorenter bc or InvocationEntryBci 3239 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3240 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3241 3242 if( !GenerateSynchronizationCode ) 3243 return; 3244 if (stopped()) { // Dead monitor? 3245 map()->pop_monitor(); // Kill monitor from debug info 3246 return; 3247 } 3248 3249 // Memory barrier to avoid floating things down past the locked region 3250 insert_mem_bar(Op_MemBarReleaseLock); 3251 3252 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3253 UnlockNode *unlock = new (C) UnlockNode(C, tf); 3254 #ifdef ASSERT 3255 unlock->set_dbg_jvms(sync_jvms()); 3256 #endif 3257 uint raw_idx = Compile::AliasIdxRaw; 3258 unlock->init_req( TypeFunc::Control, control() ); 3259 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3260 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3261 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3262 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3263 3264 unlock->init_req(TypeFunc::Parms + 0, obj); 3265 unlock->init_req(TypeFunc::Parms + 1, box); 3266 unlock = _gvn.transform(unlock)->as_Unlock(); 3267 3268 Node* mem = reset_memory(); 3269 3270 // unlock has no side-effects, sets few values 3271 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3272 3273 // Kill monitor from debug info 3274 map()->pop_monitor( ); 3275 } 3276 3277 //-------------------------------get_layout_helper----------------------------- 3278 // If the given klass is a constant or known to be an array, 3279 // fetch the constant layout helper value into constant_value 3280 // and return (Node*)NULL. Otherwise, load the non-constant 3281 // layout helper value, and return the node which represents it. 3282 // This two-faced routine is useful because allocation sites 3283 // almost always feature constant types. 3284 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3285 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3286 if (!StressReflectiveCode && inst_klass != NULL) { 3287 ciKlass* klass = inst_klass->klass(); 3288 bool xklass = inst_klass->klass_is_exact(); 3289 if (xklass || klass->is_array_klass()) { 3290 jint lhelper = klass->layout_helper(); 3291 if (lhelper != Klass::_lh_neutral_value) { 3292 constant_value = lhelper; 3293 return (Node*) NULL; 3294 } 3295 } 3296 } 3297 constant_value = Klass::_lh_neutral_value; // put in a known value 3298 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3299 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3300 } 3301 3302 // We just put in an allocate/initialize with a big raw-memory effect. 3303 // Hook selected additional alias categories on the initialization. 3304 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3305 MergeMemNode* init_in_merge, 3306 Node* init_out_raw) { 3307 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3308 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3309 3310 Node* prevmem = kit.memory(alias_idx); 3311 init_in_merge->set_memory_at(alias_idx, prevmem); 3312 kit.set_memory(init_out_raw, alias_idx); 3313 } 3314 3315 //---------------------------set_output_for_allocation------------------------- 3316 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3317 const TypeOopPtr* oop_type, 3318 bool deoptimize_on_exception) { 3319 int rawidx = Compile::AliasIdxRaw; 3320 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3321 add_safepoint_edges(alloc); 3322 Node* allocx = _gvn.transform(alloc); 3323 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3324 // create memory projection for i_o 3325 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3326 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3327 3328 // create a memory projection as for the normal control path 3329 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3330 set_memory(malloc, rawidx); 3331 3332 // a normal slow-call doesn't change i_o, but an allocation does 3333 // we create a separate i_o projection for the normal control path 3334 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3335 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3336 3337 // put in an initialization barrier 3338 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3339 rawoop)->as_Initialize(); 3340 assert(alloc->initialization() == init, "2-way macro link must work"); 3341 assert(init ->allocation() == alloc, "2-way macro link must work"); 3342 { 3343 // Extract memory strands which may participate in the new object's 3344 // initialization, and source them from the new InitializeNode. 3345 // This will allow us to observe initializations when they occur, 3346 // and link them properly (as a group) to the InitializeNode. 3347 assert(init->in(InitializeNode::Memory) == malloc, ""); 3348 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3349 init->set_req(InitializeNode::Memory, minit_in); 3350 record_for_igvn(minit_in); // fold it up later, if possible 3351 Node* minit_out = memory(rawidx); 3352 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3353 if (oop_type->isa_aryptr()) { 3354 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3355 int elemidx = C->get_alias_index(telemref); 3356 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3357 } else if (oop_type->isa_instptr()) { 3358 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3359 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3360 ciField* field = ik->nonstatic_field_at(i); 3361 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3362 continue; // do not bother to track really large numbers of fields 3363 // Find (or create) the alias category for this field: 3364 int fieldidx = C->alias_type(field)->index(); 3365 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3366 } 3367 } 3368 } 3369 3370 // Cast raw oop to the real thing... 3371 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3372 javaoop = _gvn.transform(javaoop); 3373 C->set_recent_alloc(control(), javaoop); 3374 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3375 3376 #ifdef ASSERT 3377 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3378 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3379 "Ideal_allocation works"); 3380 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3381 "Ideal_allocation works"); 3382 if (alloc->is_AllocateArray()) { 3383 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3384 "Ideal_allocation works"); 3385 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3386 "Ideal_allocation works"); 3387 } else { 3388 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3389 } 3390 } 3391 #endif //ASSERT 3392 3393 return javaoop; 3394 } 3395 3396 //---------------------------new_instance-------------------------------------- 3397 // This routine takes a klass_node which may be constant (for a static type) 3398 // or may be non-constant (for reflective code). It will work equally well 3399 // for either, and the graph will fold nicely if the optimizer later reduces 3400 // the type to a constant. 3401 // The optional arguments are for specialized use by intrinsics: 3402 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3403 // - If 'return_size_val', report the the total object size to the caller. 3404 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3405 Node* GraphKit::new_instance(Node* klass_node, 3406 Node* extra_slow_test, 3407 Node* *return_size_val, 3408 bool deoptimize_on_exception) { 3409 // Compute size in doublewords 3410 // The size is always an integral number of doublewords, represented 3411 // as a positive bytewise size stored in the klass's layout_helper. 3412 // The layout_helper also encodes (in a low bit) the need for a slow path. 3413 jint layout_con = Klass::_lh_neutral_value; 3414 Node* layout_val = get_layout_helper(klass_node, layout_con); 3415 int layout_is_con = (layout_val == NULL); 3416 3417 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3418 // Generate the initial go-slow test. It's either ALWAYS (return a 3419 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3420 // case) a computed value derived from the layout_helper. 3421 Node* initial_slow_test = NULL; 3422 if (layout_is_con) { 3423 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3424 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3425 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3426 3427 } else { // reflective case 3428 // This reflective path is used by Unsafe.allocateInstance. 3429 // (It may be stress-tested by specifying StressReflectiveCode.) 3430 // Basically, we want to get into the VM is there's an illegal argument. 3431 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3432 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3433 if (extra_slow_test != intcon(0)) { 3434 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3435 } 3436 // (Macro-expander will further convert this to a Bool, if necessary.) 3437 } 3438 3439 // Find the size in bytes. This is easy; it's the layout_helper. 3440 // The size value must be valid even if the slow path is taken. 3441 Node* size = NULL; 3442 if (layout_is_con) { 3443 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3444 } else { // reflective case 3445 // This reflective path is used by clone and Unsafe.allocateInstance. 3446 size = ConvI2X(layout_val); 3447 3448 // Clear the low bits to extract layout_helper_size_in_bytes: 3449 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3450 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3451 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3452 } 3453 if (return_size_val != NULL) { 3454 (*return_size_val) = size; 3455 } 3456 3457 // This is a precise notnull oop of the klass. 3458 // (Actually, it need not be precise if this is a reflective allocation.) 3459 // It's what we cast the result to. 3460 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3461 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3462 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3463 3464 // Now generate allocation code 3465 3466 // The entire memory state is needed for slow path of the allocation 3467 // since GC and deoptimization can happened. 3468 Node *mem = reset_memory(); 3469 set_all_memory(mem); // Create new memory state 3470 3471 AllocateNode* alloc 3472 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3473 control(), mem, i_o(), 3474 size, klass_node, 3475 initial_slow_test); 3476 3477 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3478 } 3479 3480 //-------------------------------new_array------------------------------------- 3481 // helper for both newarray and anewarray 3482 // The 'length' parameter is (obviously) the length of the array. 3483 // See comments on new_instance for the meaning of the other arguments. 3484 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3485 Node* length, // number of array elements 3486 int nargs, // number of arguments to push back for uncommon trap 3487 Node* *return_size_val, 3488 bool deoptimize_on_exception) { 3489 jint layout_con = Klass::_lh_neutral_value; 3490 Node* layout_val = get_layout_helper(klass_node, layout_con); 3491 int layout_is_con = (layout_val == NULL); 3492 3493 if (!layout_is_con && !StressReflectiveCode && 3494 !too_many_traps(Deoptimization::Reason_class_check)) { 3495 // This is a reflective array creation site. 3496 // Optimistically assume that it is a subtype of Object[], 3497 // so that we can fold up all the address arithmetic. 3498 layout_con = Klass::array_layout_helper(T_OBJECT); 3499 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3500 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3501 { BuildCutout unless(this, bol_lh, PROB_MAX); 3502 inc_sp(nargs); 3503 uncommon_trap(Deoptimization::Reason_class_check, 3504 Deoptimization::Action_maybe_recompile); 3505 } 3506 layout_val = NULL; 3507 layout_is_con = true; 3508 } 3509 3510 // Generate the initial go-slow test. Make sure we do not overflow 3511 // if length is huge (near 2Gig) or negative! We do not need 3512 // exact double-words here, just a close approximation of needed 3513 // double-words. We can't add any offset or rounding bits, lest we 3514 // take a size -1 of bytes and make it positive. Use an unsigned 3515 // compare, so negative sizes look hugely positive. 3516 int fast_size_limit = FastAllocateSizeLimit; 3517 if (layout_is_con) { 3518 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3519 // Increase the size limit if we have exact knowledge of array type. 3520 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3521 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3522 } 3523 3524 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3525 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3526 3527 // --- Size Computation --- 3528 // array_size = round_to_heap(array_header + (length << elem_shift)); 3529 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3530 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3531 // The rounding mask is strength-reduced, if possible. 3532 int round_mask = MinObjAlignmentInBytes - 1; 3533 Node* header_size = NULL; 3534 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3535 // (T_BYTE has the weakest alignment and size restrictions...) 3536 if (layout_is_con) { 3537 int hsize = Klass::layout_helper_header_size(layout_con); 3538 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3539 BasicType etype = Klass::layout_helper_element_type(layout_con); 3540 if ((round_mask & ~right_n_bits(eshift)) == 0) 3541 round_mask = 0; // strength-reduce it if it goes away completely 3542 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3543 assert(header_size_min <= hsize, "generic minimum is smallest"); 3544 header_size_min = hsize; 3545 header_size = intcon(hsize + round_mask); 3546 } else { 3547 Node* hss = intcon(Klass::_lh_header_size_shift); 3548 Node* hsm = intcon(Klass::_lh_header_size_mask); 3549 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3550 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3551 Node* mask = intcon(round_mask); 3552 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3553 } 3554 3555 Node* elem_shift = NULL; 3556 if (layout_is_con) { 3557 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3558 if (eshift != 0) 3559 elem_shift = intcon(eshift); 3560 } else { 3561 // There is no need to mask or shift this value. 3562 // The semantics of LShiftINode include an implicit mask to 0x1F. 3563 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3564 elem_shift = layout_val; 3565 } 3566 3567 // Transition to native address size for all offset calculations: 3568 Node* lengthx = ConvI2X(length); 3569 Node* headerx = ConvI2X(header_size); 3570 #ifdef _LP64 3571 { const TypeInt* tilen = _gvn.find_int_type(length); 3572 if (tilen != NULL && tilen->_lo < 0) { 3573 // Add a manual constraint to a positive range. Cf. array_element_address. 3574 jlong size_max = fast_size_limit; 3575 if (size_max > tilen->_hi) size_max = tilen->_hi; 3576 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3577 3578 // Only do a narrow I2L conversion if the range check passed. 3579 IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3580 _gvn.transform(iff); 3581 RegionNode* region = new (C) RegionNode(3); 3582 _gvn.set_type(region, Type::CONTROL); 3583 lengthx = new (C) PhiNode(region, TypeLong::LONG); 3584 _gvn.set_type(lengthx, TypeLong::LONG); 3585 3586 // Range check passed. Use ConvI2L node with narrow type. 3587 Node* passed = IfFalse(iff); 3588 region->init_req(1, passed); 3589 // Make I2L conversion control dependent to prevent it from 3590 // floating above the range check during loop optimizations. 3591 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3592 3593 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3594 region->init_req(2, IfTrue(iff)); 3595 lengthx->init_req(2, ConvI2X(length)); 3596 3597 set_control(region); 3598 record_for_igvn(region); 3599 record_for_igvn(lengthx); 3600 } 3601 } 3602 #endif 3603 3604 // Combine header size (plus rounding) and body size. Then round down. 3605 // This computation cannot overflow, because it is used only in two 3606 // places, one where the length is sharply limited, and the other 3607 // after a successful allocation. 3608 Node* abody = lengthx; 3609 if (elem_shift != NULL) 3610 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3611 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3612 if (round_mask != 0) { 3613 Node* mask = MakeConX(~round_mask); 3614 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3615 } 3616 // else if round_mask == 0, the size computation is self-rounding 3617 3618 if (return_size_val != NULL) { 3619 // This is the size 3620 (*return_size_val) = size; 3621 } 3622 3623 // Now generate allocation code 3624 3625 // The entire memory state is needed for slow path of the allocation 3626 // since GC and deoptimization can happened. 3627 Node *mem = reset_memory(); 3628 set_all_memory(mem); // Create new memory state 3629 3630 if (initial_slow_test->is_Bool()) { 3631 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3632 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3633 } 3634 3635 // Create the AllocateArrayNode and its result projections 3636 AllocateArrayNode* alloc 3637 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3638 control(), mem, i_o(), 3639 size, klass_node, 3640 initial_slow_test, 3641 length); 3642 3643 // Cast to correct type. Note that the klass_node may be constant or not, 3644 // and in the latter case the actual array type will be inexact also. 3645 // (This happens via a non-constant argument to inline_native_newArray.) 3646 // In any case, the value of klass_node provides the desired array type. 3647 const TypeInt* length_type = _gvn.find_int_type(length); 3648 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3649 if (ary_type->isa_aryptr() && length_type != NULL) { 3650 // Try to get a better type than POS for the size 3651 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3652 } 3653 3654 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3655 3656 // Cast length on remaining path to be as narrow as possible 3657 if (map()->find_edge(length) >= 0) { 3658 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3659 if (ccast != length) { 3660 _gvn.set_type_bottom(ccast); 3661 record_for_igvn(ccast); 3662 replace_in_map(length, ccast); 3663 } 3664 } 3665 3666 return javaoop; 3667 } 3668 3669 // The following "Ideal_foo" functions are placed here because they recognize 3670 // the graph shapes created by the functions immediately above. 3671 3672 //---------------------------Ideal_allocation---------------------------------- 3673 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3674 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3675 if (ptr == NULL) { // reduce dumb test in callers 3676 return NULL; 3677 } 3678 3679 #if INCLUDE_ALL_GCS 3680 if (UseShenandoahGC) { 3681 ptr = ShenandoahBarrierSetC2::bsc2()->step_over_gc_barrier(ptr); 3682 } 3683 #endif 3684 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3685 ptr = ptr->in(1); 3686 if (ptr == NULL) return NULL; 3687 } 3688 // Return NULL for allocations with several casts: 3689 // j.l.reflect.Array.newInstance(jobject, jint) 3690 // Object.clone() 3691 // to keep more precise type from last cast. 3692 if (ptr->is_Proj()) { 3693 Node* allo = ptr->in(0); 3694 if (allo != NULL && allo->is_Allocate()) { 3695 return allo->as_Allocate(); 3696 } 3697 } 3698 // Report failure to match. 3699 return NULL; 3700 } 3701 3702 // Fancy version which also strips off an offset (and reports it to caller). 3703 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3704 intptr_t& offset) { 3705 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3706 if (base == NULL) return NULL; 3707 return Ideal_allocation(base, phase); 3708 } 3709 3710 // Trace Initialize <- Proj[Parm] <- Allocate 3711 AllocateNode* InitializeNode::allocation() { 3712 Node* rawoop = in(InitializeNode::RawAddress); 3713 if (rawoop->is_Proj()) { 3714 Node* alloc = rawoop->in(0); 3715 if (alloc->is_Allocate()) { 3716 return alloc->as_Allocate(); 3717 } 3718 } 3719 return NULL; 3720 } 3721 3722 // Trace Allocate -> Proj[Parm] -> Initialize 3723 InitializeNode* AllocateNode::initialization() { 3724 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3725 if (rawoop == NULL) return NULL; 3726 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3727 Node* init = rawoop->fast_out(i); 3728 if (init->is_Initialize()) { 3729 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3730 return init->as_Initialize(); 3731 } 3732 } 3733 return NULL; 3734 } 3735 3736 //----------------------------- loop predicates --------------------------- 3737 3738 //------------------------------add_predicate_impl---------------------------- 3739 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3740 // Too many traps seen? 3741 if (too_many_traps(reason)) { 3742 #ifdef ASSERT 3743 if (TraceLoopPredicate) { 3744 int tc = C->trap_count(reason); 3745 tty->print("too many traps=%s tcount=%d in ", 3746 Deoptimization::trap_reason_name(reason), tc); 3747 method()->print(); // which method has too many predicate traps 3748 tty->cr(); 3749 } 3750 #endif 3751 // We cannot afford to take more traps here, 3752 // do not generate predicate. 3753 return; 3754 } 3755 3756 Node *cont = _gvn.intcon(1); 3757 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3758 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3759 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3760 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3761 C->add_predicate_opaq(opq); 3762 { 3763 PreserveJVMState pjvms(this); 3764 set_control(iffalse); 3765 inc_sp(nargs); 3766 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3767 } 3768 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3769 set_control(iftrue); 3770 } 3771 3772 //------------------------------add_predicate--------------------------------- 3773 void GraphKit::add_predicate(int nargs) { 3774 if (UseLoopPredicate) { 3775 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3776 } 3777 // loop's limit check predicate should be near the loop. 3778 if (LoopLimitCheck) { 3779 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3780 } 3781 } 3782 3783 //----------------------------- store barriers ---------------------------- 3784 #define __ ideal. 3785 3786 void GraphKit::sync_kit(IdealKit& ideal) { 3787 set_all_memory(__ merged_memory()); 3788 set_i_o(__ i_o()); 3789 set_control(__ ctrl()); 3790 } 3791 3792 void GraphKit::final_sync(IdealKit& ideal) { 3793 // Final sync IdealKit and graphKit. 3794 sync_kit(ideal); 3795 } 3796 3797 // vanilla/CMS post barrier 3798 // Insert a write-barrier store. This is to let generational GC work; we have 3799 // to flag all oop-stores before the next GC point. 3800 void GraphKit::write_barrier_post(Node* oop_store, 3801 Node* obj, 3802 Node* adr, 3803 uint adr_idx, 3804 Node* val, 3805 bool use_precise) { 3806 // No store check needed if we're storing a NULL or an old object 3807 // (latter case is probably a string constant). The concurrent 3808 // mark sweep garbage collector, however, needs to have all nonNull 3809 // oop updates flagged via card-marks. 3810 if (val != NULL && val->is_Con()) { 3811 // must be either an oop or NULL 3812 const Type* t = val->bottom_type(); 3813 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3814 // stores of null never (?) need barriers 3815 return; 3816 } 3817 3818 if (use_ReduceInitialCardMarks() 3819 && obj == just_allocated_object(control())) { 3820 // We can skip marks on a freshly-allocated object in Eden. 3821 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3822 // That routine informs GC to take appropriate compensating steps, 3823 // upon a slow-path allocation, so as to make this card-mark 3824 // elision safe. 3825 return; 3826 } 3827 3828 if (!use_precise) { 3829 // All card marks for a (non-array) instance are in one place: 3830 adr = obj; 3831 } 3832 // (Else it's an array (or unknown), and we want more precise card marks.) 3833 assert(adr != NULL, ""); 3834 3835 IdealKit ideal(this, true); 3836 3837 // Convert the pointer to an int prior to doing math on it 3838 Node* cast = __ CastPX(__ ctrl(), adr); 3839 3840 // Divide by card size 3841 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3842 "Only one we handle so far."); 3843 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3844 3845 // Combine card table base and card offset 3846 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3847 3848 // Get the alias_index for raw card-mark memory 3849 int adr_type = Compile::AliasIdxRaw; 3850 Node* zero = __ ConI(0); // Dirty card value 3851 BasicType bt = T_BYTE; 3852 3853 if (UseCondCardMark) { 3854 // The classic GC reference write barrier is typically implemented 3855 // as a store into the global card mark table. Unfortunately 3856 // unconditional stores can result in false sharing and excessive 3857 // coherence traffic as well as false transactional aborts. 3858 // UseCondCardMark enables MP "polite" conditional card mark 3859 // stores. In theory we could relax the load from ctrl() to 3860 // no_ctrl, but that doesn't buy much latitude. 3861 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3862 __ if_then(card_val, BoolTest::ne, zero); 3863 } 3864 3865 // Smash zero into card 3866 if( !UseConcMarkSweepGC ) { 3867 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3868 } else { 3869 // Specialized path for CM store barrier 3870 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3871 } 3872 3873 if (UseCondCardMark) { 3874 __ end_if(); 3875 } 3876 3877 // Final sync IdealKit and GraphKit. 3878 final_sync(ideal); 3879 } 3880 3881 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) { 3882 if (UseShenandoahGC && adr != NULL) { 3883 Node* c = kit.control(); 3884 Node* call = c->in(1)->in(1)->in(1)->in(0); 3885 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected"); 3886 call->add_req(adr); 3887 } 3888 } 3889 3890 // G1 pre/post barriers 3891 void GraphKit::g1_write_barrier_pre(bool do_load, 3892 Node* obj, 3893 Node* adr, 3894 uint alias_idx, 3895 Node* val, 3896 const TypeOopPtr* val_type, 3897 Node* pre_val, 3898 BasicType bt) { 3899 3900 // Some sanity checks 3901 // Note: val is unused in this routine. 3902 3903 if (do_load) { 3904 // We need to generate the load of the previous value 3905 assert(obj != NULL, "must have a base"); 3906 assert(adr != NULL, "where are loading from?"); 3907 assert(pre_val == NULL, "loaded already?"); 3908 assert(val_type != NULL, "need a type"); 3909 } else { 3910 // In this case both val_type and alias_idx are unused. 3911 assert(pre_val != NULL, "must be loaded already"); 3912 // Nothing to be done if pre_val is null. 3913 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3914 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3915 } 3916 assert(bt == T_OBJECT, "or we shouldn't be here"); 3917 3918 IdealKit ideal(this, true); 3919 3920 Node* tls = __ thread(); // ThreadLocalStorage 3921 3922 Node* no_ctrl = NULL; 3923 Node* no_base = __ top(); 3924 Node* zero = __ ConI(0); 3925 Node* zeroX = __ ConX(0); 3926 3927 float likely = PROB_LIKELY(0.999); 3928 float unlikely = PROB_UNLIKELY(0.999); 3929 3930 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3931 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3932 3933 // Offsets into the thread 3934 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3935 PtrQueue::byte_offset_of_active()); 3936 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3937 PtrQueue::byte_offset_of_index()); 3938 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3939 PtrQueue::byte_offset_of_buf()); 3940 3941 // Now the actual pointers into the thread 3942 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3943 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3944 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3945 3946 // Now some of the values 3947 Node* marking; 3948 if (UseShenandoahGC) { 3949 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(JavaThread::gc_state_offset()))); 3950 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw); 3951 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING)); 3952 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape"); 3953 } else { 3954 assert(UseG1GC, "should be"); 3955 marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3956 } 3957 3958 // if (!marking) 3959 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3960 BasicType index_bt = TypeX_X->basic_type(); 3961 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3962 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3963 3964 if (do_load) { 3965 // load original value 3966 // alias_idx correct?? 3967 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3968 } 3969 3970 // if (pre_val != NULL) 3971 __ if_then(pre_val, BoolTest::ne, null()); { 3972 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3973 3974 // is the queue for this thread full? 3975 __ if_then(index, BoolTest::ne, zeroX, likely); { 3976 3977 // decrement the index 3978 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3979 3980 // Now get the buffer location we will log the previous value into and store it 3981 Node *log_addr = __ AddP(no_base, buffer, next_index); 3982 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 3983 // update the index 3984 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 3985 3986 } __ else_(); { 3987 3988 // logging buffer is full, call the runtime 3989 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3990 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3991 } __ end_if(); // (!index) 3992 } __ end_if(); // (pre_val != NULL) 3993 } __ end_if(); // (!marking) 3994 3995 // Final sync IdealKit and GraphKit. 3996 final_sync(ideal); 3997 g1_write_barrier_pre_helper(*this, adr); 3998 } 3999 4000 // 4001 // Update the card table and add card address to the queue 4002 // 4003 void GraphKit::g1_mark_card(IdealKit& ideal, 4004 Node* card_adr, 4005 Node* oop_store, 4006 uint oop_alias_idx, 4007 Node* index, 4008 Node* index_adr, 4009 Node* buffer, 4010 const TypeFunc* tf) { 4011 4012 Node* zero = __ ConI(0); 4013 Node* zeroX = __ ConX(0); 4014 Node* no_base = __ top(); 4015 BasicType card_bt = T_BYTE; 4016 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4017 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4018 4019 // Now do the queue work 4020 __ if_then(index, BoolTest::ne, zeroX); { 4021 4022 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 4023 Node* log_addr = __ AddP(no_base, buffer, next_index); 4024 4025 // Order, see storeCM. 4026 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4027 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4028 4029 } __ else_(); { 4030 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4031 } __ end_if(); 4032 4033 } 4034 4035 void GraphKit::g1_write_barrier_post(Node* oop_store, 4036 Node* obj, 4037 Node* adr, 4038 uint alias_idx, 4039 Node* val, 4040 BasicType bt, 4041 bool use_precise) { 4042 // If we are writing a NULL then we need no post barrier 4043 4044 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4045 // Must be NULL 4046 const Type* t = val->bottom_type(); 4047 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4048 // No post barrier if writing NULLx 4049 return; 4050 } 4051 4052 if (!use_precise) { 4053 // All card marks for a (non-array) instance are in one place: 4054 adr = obj; 4055 } 4056 // (Else it's an array (or unknown), and we want more precise card marks.) 4057 assert(adr != NULL, ""); 4058 4059 IdealKit ideal(this, true); 4060 4061 Node* tls = __ thread(); // ThreadLocalStorage 4062 4063 Node* no_base = __ top(); 4064 float likely = PROB_LIKELY(0.999); 4065 float unlikely = PROB_UNLIKELY(0.999); 4066 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4067 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4068 Node* zeroX = __ ConX(0); 4069 4070 // Get the alias_index for raw card-mark memory 4071 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4072 4073 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4074 4075 // Offsets into the thread 4076 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4077 PtrQueue::byte_offset_of_index()); 4078 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4079 PtrQueue::byte_offset_of_buf()); 4080 4081 // Pointers into the thread 4082 4083 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4084 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4085 4086 // Now some values 4087 // Use ctrl to avoid hoisting these values past a safepoint, which could 4088 // potentially reset these fields in the JavaThread. 4089 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4090 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4091 4092 // Convert the store obj pointer to an int prior to doing math on it 4093 // Must use ctrl to prevent "integerized oop" existing across safepoint 4094 Node* cast = __ CastPX(__ ctrl(), adr); 4095 4096 // Divide pointer by card size 4097 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4098 4099 // Combine card table base and card offset 4100 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4101 4102 // If we know the value being stored does it cross regions? 4103 4104 if (val != NULL) { 4105 // Does the store cause us to cross regions? 4106 4107 // Should be able to do an unsigned compare of region_size instead of 4108 // and extra shift. Do we have an unsigned compare?? 4109 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4110 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4111 4112 // if (xor_res == 0) same region so skip 4113 __ if_then(xor_res, BoolTest::ne, zeroX); { 4114 4115 // No barrier if we are storing a NULL 4116 __ if_then(val, BoolTest::ne, null(), unlikely); { 4117 4118 // Ok must mark the card if not already dirty 4119 4120 // load the original value of the card 4121 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4122 4123 __ if_then(card_val, BoolTest::ne, young_card); { 4124 sync_kit(ideal); 4125 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4126 insert_mem_bar(Op_MemBarVolatile, oop_store); 4127 __ sync_kit(this); 4128 4129 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4130 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4131 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4132 } __ end_if(); 4133 } __ end_if(); 4134 } __ end_if(); 4135 } __ end_if(); 4136 } else { 4137 // Object.clone() instrinsic uses this path. 4138 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4139 } 4140 4141 // Final sync IdealKit and GraphKit. 4142 final_sync(ideal); 4143 } 4144 #undef __ 4145 4146 4147 4148 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4149 if (java_lang_String::has_offset_field()) { 4150 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4151 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4152 false, NULL, 0); 4153 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4154 int offset_field_idx = C->get_alias_index(offset_field_type); 4155 return make_load(ctrl, 4156 basic_plus_adr(str, str, offset_offset), 4157 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4158 } else { 4159 return intcon(0); 4160 } 4161 } 4162 4163 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4164 if (java_lang_String::has_count_field()) { 4165 int count_offset = java_lang_String::count_offset_in_bytes(); 4166 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4167 false, NULL, 0); 4168 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4169 int count_field_idx = C->get_alias_index(count_field_type); 4170 return make_load(ctrl, 4171 basic_plus_adr(str, str, count_offset), 4172 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4173 } else { 4174 return load_array_length(load_String_value(ctrl, str)); 4175 } 4176 } 4177 4178 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4179 int value_offset = java_lang_String::value_offset_in_bytes(); 4180 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4181 false, NULL, 0); 4182 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4183 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4184 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4185 ciTypeArrayKlass::make(T_CHAR), true, 0); 4186 int value_field_idx = C->get_alias_index(value_field_type); 4187 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4188 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4189 #if INCLUDE_ALL_GCS 4190 if (UseShenandoahGC) { 4191 load = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, load); 4192 } 4193 #endif 4194 // String.value field is known to be @Stable. 4195 if (UseImplicitStableValues) { 4196 load = cast_array_to_stable(load, value_type); 4197 } 4198 return load; 4199 } 4200 4201 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4202 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4203 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4204 false, NULL, 0); 4205 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4206 int offset_field_idx = C->get_alias_index(offset_field_type); 4207 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4208 value, T_INT, offset_field_idx, MemNode::unordered); 4209 } 4210 4211 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4212 int value_offset = java_lang_String::value_offset_in_bytes(); 4213 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4214 false, NULL, 0); 4215 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4216 4217 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4218 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4219 } 4220 4221 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4222 int count_offset = java_lang_String::count_offset_in_bytes(); 4223 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4224 false, NULL, 0); 4225 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4226 int count_field_idx = C->get_alias_index(count_field_type); 4227 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4228 value, T_INT, count_field_idx, MemNode::unordered); 4229 } 4230 4231 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4232 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4233 // assumption of CCP analysis. 4234 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); 4235 }