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/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc/g1/heapRegion.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/cardTableModRefBS.hpp" 31 #include "gc/shared/collectedHeap.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/intrinsicnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/shenandoahSupport.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 //----------------------------GraphKit----------------------------------------- 49 // Main utility constructor. 50 GraphKit::GraphKit(JVMState* jvms) 51 : Phase(Phase::Parser), 52 _env(C->env()), 53 _gvn(*C->initial_gvn()) 54 { 55 _exceptions = jvms->map()->next_exception(); 56 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 57 set_jvms(jvms); 58 } 59 60 // Private constructor for parser. 61 GraphKit::GraphKit() 62 : Phase(Phase::Parser), 63 _env(C->env()), 64 _gvn(*C->initial_gvn()) 65 { 66 _exceptions = NULL; 67 set_map(NULL); 68 debug_only(_sp = -99); 69 debug_only(set_bci(-99)); 70 } 71 72 73 74 //---------------------------clean_stack--------------------------------------- 75 // Clear away rubbish from the stack area of the JVM state. 76 // This destroys any arguments that may be waiting on the stack. 77 void GraphKit::clean_stack(int from_sp) { 78 SafePointNode* map = this->map(); 79 JVMState* jvms = this->jvms(); 80 int stk_size = jvms->stk_size(); 81 int stkoff = jvms->stkoff(); 82 Node* top = this->top(); 83 for (int i = from_sp; i < stk_size; i++) { 84 if (map->in(stkoff + i) != top) { 85 map->set_req(stkoff + i, top); 86 } 87 } 88 } 89 90 91 //--------------------------------sync_jvms----------------------------------- 92 // Make sure our current jvms agrees with our parse state. 93 JVMState* GraphKit::sync_jvms() const { 94 JVMState* jvms = this->jvms(); 95 jvms->set_bci(bci()); // Record the new bci in the JVMState 96 jvms->set_sp(sp()); // Record the new sp in the JVMState 97 assert(jvms_in_sync(), "jvms is now in sync"); 98 return jvms; 99 } 100 101 //--------------------------------sync_jvms_for_reexecute--------------------- 102 // Make sure our current jvms agrees with our parse state. This version 103 // uses the reexecute_sp for reexecuting bytecodes. 104 JVMState* GraphKit::sync_jvms_for_reexecute() { 105 JVMState* jvms = this->jvms(); 106 jvms->set_bci(bci()); // Record the new bci in the JVMState 107 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 108 return jvms; 109 } 110 111 #ifdef ASSERT 112 bool GraphKit::jvms_in_sync() const { 113 Parse* parse = is_Parse(); 114 if (parse == NULL) { 115 if (bci() != jvms()->bci()) return false; 116 if (sp() != (int)jvms()->sp()) return false; 117 return true; 118 } 119 if (jvms()->method() != parse->method()) return false; 120 if (jvms()->bci() != parse->bci()) return false; 121 int jvms_sp = jvms()->sp(); 122 if (jvms_sp != parse->sp()) return false; 123 int jvms_depth = jvms()->depth(); 124 if (jvms_depth != parse->depth()) return false; 125 return true; 126 } 127 128 // Local helper checks for special internal merge points 129 // used to accumulate and merge exception states. 130 // They are marked by the region's in(0) edge being the map itself. 131 // Such merge points must never "escape" into the parser at large, 132 // until they have been handed to gvn.transform. 133 static bool is_hidden_merge(Node* reg) { 134 if (reg == NULL) return false; 135 if (reg->is_Phi()) { 136 reg = reg->in(0); 137 if (reg == NULL) return false; 138 } 139 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 140 } 141 142 void GraphKit::verify_map() const { 143 if (map() == NULL) return; // null map is OK 144 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 145 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 146 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 147 } 148 149 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 150 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 151 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 152 } 153 #endif 154 155 //---------------------------stop_and_kill_map--------------------------------- 156 // Set _map to NULL, signalling a stop to further bytecode execution. 157 // First smash the current map's control to a constant, to mark it dead. 158 void GraphKit::stop_and_kill_map() { 159 SafePointNode* dead_map = stop(); 160 if (dead_map != NULL) { 161 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 162 assert(dead_map->is_killed(), "must be so marked"); 163 } 164 } 165 166 167 //--------------------------------stopped-------------------------------------- 168 // Tell if _map is NULL, or control is top. 169 bool GraphKit::stopped() { 170 if (map() == NULL) return true; 171 else if (control() == top()) return true; 172 else return false; 173 } 174 175 176 //-----------------------------has_ex_handler---------------------------------- 177 // Tell if this method or any caller method has exception handlers. 178 bool GraphKit::has_ex_handler() { 179 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 180 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 181 return true; 182 } 183 } 184 return false; 185 } 186 187 //------------------------------save_ex_oop------------------------------------ 188 // Save an exception without blowing stack contents or other JVM state. 189 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 190 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 191 ex_map->add_req(ex_oop); 192 debug_only(verify_exception_state(ex_map)); 193 } 194 195 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 196 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 197 Node* ex_oop = ex_map->in(ex_map->req()-1); 198 if (clear_it) ex_map->del_req(ex_map->req()-1); 199 return ex_oop; 200 } 201 202 //-----------------------------saved_ex_oop------------------------------------ 203 // Recover a saved exception from its map. 204 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 205 return common_saved_ex_oop(ex_map, false); 206 } 207 208 //--------------------------clear_saved_ex_oop--------------------------------- 209 // Erase a previously saved exception from its map. 210 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 211 return common_saved_ex_oop(ex_map, true); 212 } 213 214 #ifdef ASSERT 215 //---------------------------has_saved_ex_oop---------------------------------- 216 // Erase a previously saved exception from its map. 217 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 218 return ex_map->req() == ex_map->jvms()->endoff()+1; 219 } 220 #endif 221 222 //-------------------------make_exception_state-------------------------------- 223 // Turn the current JVM state into an exception state, appending the ex_oop. 224 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 225 sync_jvms(); 226 SafePointNode* ex_map = stop(); // do not manipulate this map any more 227 set_saved_ex_oop(ex_map, ex_oop); 228 return ex_map; 229 } 230 231 232 //--------------------------add_exception_state-------------------------------- 233 // Add an exception to my list of exceptions. 234 void GraphKit::add_exception_state(SafePointNode* ex_map) { 235 if (ex_map == NULL || ex_map->control() == top()) { 236 return; 237 } 238 #ifdef ASSERT 239 verify_exception_state(ex_map); 240 if (has_exceptions()) { 241 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 242 } 243 #endif 244 245 // If there is already an exception of exactly this type, merge with it. 246 // In particular, null-checks and other low-level exceptions common up here. 247 Node* ex_oop = saved_ex_oop(ex_map); 248 const Type* ex_type = _gvn.type(ex_oop); 249 if (ex_oop == top()) { 250 // No action needed. 251 return; 252 } 253 assert(ex_type->isa_instptr(), "exception must be an instance"); 254 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 255 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 256 // We check sp also because call bytecodes can generate exceptions 257 // both before and after arguments are popped! 258 if (ex_type2 == ex_type 259 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 260 combine_exception_states(ex_map, e2); 261 return; 262 } 263 } 264 265 // No pre-existing exception of the same type. Chain it on the list. 266 push_exception_state(ex_map); 267 } 268 269 //-----------------------add_exception_states_from----------------------------- 270 void GraphKit::add_exception_states_from(JVMState* jvms) { 271 SafePointNode* ex_map = jvms->map()->next_exception(); 272 if (ex_map != NULL) { 273 jvms->map()->set_next_exception(NULL); 274 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 275 next_map = ex_map->next_exception(); 276 ex_map->set_next_exception(NULL); 277 add_exception_state(ex_map); 278 } 279 } 280 } 281 282 //-----------------------transfer_exceptions_into_jvms------------------------- 283 JVMState* GraphKit::transfer_exceptions_into_jvms() { 284 if (map() == NULL) { 285 // We need a JVMS to carry the exceptions, but the map has gone away. 286 // Create a scratch JVMS, cloned from any of the exception states... 287 if (has_exceptions()) { 288 _map = _exceptions; 289 _map = clone_map(); 290 _map->set_next_exception(NULL); 291 clear_saved_ex_oop(_map); 292 debug_only(verify_map()); 293 } else { 294 // ...or created from scratch 295 JVMState* jvms = new (C) JVMState(_method, NULL); 296 jvms->set_bci(_bci); 297 jvms->set_sp(_sp); 298 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 299 set_jvms(jvms); 300 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 301 set_all_memory(top()); 302 while (map()->req() < jvms->endoff()) map()->add_req(top()); 303 } 304 // (This is a kludge, in case you didn't notice.) 305 set_control(top()); 306 } 307 JVMState* jvms = sync_jvms(); 308 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 309 jvms->map()->set_next_exception(_exceptions); 310 _exceptions = NULL; // done with this set of exceptions 311 return jvms; 312 } 313 314 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 315 assert(is_hidden_merge(dstphi), "must be a special merge node"); 316 assert(is_hidden_merge(srcphi), "must be a special merge node"); 317 uint limit = srcphi->req(); 318 for (uint i = PhiNode::Input; i < limit; i++) { 319 dstphi->add_req(srcphi->in(i)); 320 } 321 } 322 static inline void add_one_req(Node* dstphi, Node* src) { 323 assert(is_hidden_merge(dstphi), "must be a special merge node"); 324 assert(!is_hidden_merge(src), "must not be a special merge node"); 325 dstphi->add_req(src); 326 } 327 328 //-----------------------combine_exception_states------------------------------ 329 // This helper function combines exception states by building phis on a 330 // specially marked state-merging region. These regions and phis are 331 // untransformed, and can build up gradually. The region is marked by 332 // having a control input of its exception map, rather than NULL. Such 333 // regions do not appear except in this function, and in use_exception_state. 334 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 335 if (failing()) return; // dying anyway... 336 JVMState* ex_jvms = ex_map->_jvms; 337 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 338 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 339 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 340 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 341 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 342 assert(ex_map->req() == phi_map->req(), "matching maps"); 343 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 344 Node* hidden_merge_mark = root(); 345 Node* region = phi_map->control(); 346 MergeMemNode* phi_mem = phi_map->merged_memory(); 347 MergeMemNode* ex_mem = ex_map->merged_memory(); 348 if (region->in(0) != hidden_merge_mark) { 349 // The control input is not (yet) a specially-marked region in phi_map. 350 // Make it so, and build some phis. 351 region = new RegionNode(2); 352 _gvn.set_type(region, Type::CONTROL); 353 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 354 region->init_req(1, phi_map->control()); 355 phi_map->set_control(region); 356 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 357 record_for_igvn(io_phi); 358 _gvn.set_type(io_phi, Type::ABIO); 359 phi_map->set_i_o(io_phi); 360 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 361 Node* m = mms.memory(); 362 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 363 record_for_igvn(m_phi); 364 _gvn.set_type(m_phi, Type::MEMORY); 365 mms.set_memory(m_phi); 366 } 367 } 368 369 // Either or both of phi_map and ex_map might already be converted into phis. 370 Node* ex_control = ex_map->control(); 371 // if there is special marking on ex_map also, we add multiple edges from src 372 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 373 // how wide was the destination phi_map, originally? 374 uint orig_width = region->req(); 375 376 if (add_multiple) { 377 add_n_reqs(region, ex_control); 378 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 379 } else { 380 // ex_map has no merges, so we just add single edges everywhere 381 add_one_req(region, ex_control); 382 add_one_req(phi_map->i_o(), ex_map->i_o()); 383 } 384 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 385 if (mms.is_empty()) { 386 // get a copy of the base memory, and patch some inputs into it 387 const TypePtr* adr_type = mms.adr_type(C); 388 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 389 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 390 mms.set_memory(phi); 391 // Prepare to append interesting stuff onto the newly sliced phi: 392 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 393 } 394 // Append stuff from ex_map: 395 if (add_multiple) { 396 add_n_reqs(mms.memory(), mms.memory2()); 397 } else { 398 add_one_req(mms.memory(), mms.memory2()); 399 } 400 } 401 uint limit = ex_map->req(); 402 for (uint i = TypeFunc::Parms; i < limit; i++) { 403 // Skip everything in the JVMS after tos. (The ex_oop follows.) 404 if (i == tos) i = ex_jvms->monoff(); 405 Node* src = ex_map->in(i); 406 Node* dst = phi_map->in(i); 407 if (src != dst) { 408 PhiNode* phi; 409 if (dst->in(0) != region) { 410 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 411 record_for_igvn(phi); 412 _gvn.set_type(phi, phi->type()); 413 phi_map->set_req(i, dst); 414 // Prepare to append interesting stuff onto the new phi: 415 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 416 } else { 417 assert(dst->is_Phi(), "nobody else uses a hidden region"); 418 phi = dst->as_Phi(); 419 } 420 if (add_multiple && src->in(0) == ex_control) { 421 // Both are phis. 422 add_n_reqs(dst, src); 423 } else { 424 while (dst->req() < region->req()) add_one_req(dst, src); 425 } 426 const Type* srctype = _gvn.type(src); 427 if (phi->type() != srctype) { 428 const Type* dsttype = phi->type()->meet_speculative(srctype); 429 if (phi->type() != dsttype) { 430 phi->set_type(dsttype); 431 _gvn.set_type(phi, dsttype); 432 } 433 } 434 } 435 } 436 phi_map->merge_replaced_nodes_with(ex_map); 437 } 438 439 //--------------------------use_exception_state-------------------------------- 440 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 441 if (failing()) { stop(); return top(); } 442 Node* region = phi_map->control(); 443 Node* hidden_merge_mark = root(); 444 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 445 Node* ex_oop = clear_saved_ex_oop(phi_map); 446 if (region->in(0) == hidden_merge_mark) { 447 // Special marking for internal ex-states. Process the phis now. 448 region->set_req(0, region); // now it's an ordinary region 449 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 450 // Note: Setting the jvms also sets the bci and sp. 451 set_control(_gvn.transform(region)); 452 uint tos = jvms()->stkoff() + sp(); 453 for (uint i = 1; i < tos; i++) { 454 Node* x = phi_map->in(i); 455 if (x->in(0) == region) { 456 assert(x->is_Phi(), "expected a special phi"); 457 phi_map->set_req(i, _gvn.transform(x)); 458 } 459 } 460 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 461 Node* x = mms.memory(); 462 if (x->in(0) == region) { 463 assert(x->is_Phi(), "nobody else uses a hidden region"); 464 mms.set_memory(_gvn.transform(x)); 465 } 466 } 467 if (ex_oop->in(0) == region) { 468 assert(ex_oop->is_Phi(), "expected a special phi"); 469 ex_oop = _gvn.transform(ex_oop); 470 } 471 } else { 472 set_jvms(phi_map->jvms()); 473 } 474 475 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 476 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 477 return ex_oop; 478 } 479 480 //---------------------------------java_bc------------------------------------- 481 Bytecodes::Code GraphKit::java_bc() const { 482 ciMethod* method = this->method(); 483 int bci = this->bci(); 484 if (method != NULL && bci != InvocationEntryBci) 485 return method->java_code_at_bci(bci); 486 else 487 return Bytecodes::_illegal; 488 } 489 490 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 491 bool must_throw) { 492 // if the exception capability is set, then we will generate code 493 // to check the JavaThread.should_post_on_exceptions flag to see 494 // if we actually need to report exception events (for this 495 // thread). If we don't need to report exception events, we will 496 // take the normal fast path provided by add_exception_events. If 497 // exception event reporting is enabled for this thread, we will 498 // take the uncommon_trap in the BuildCutout below. 499 500 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 501 Node* jthread = _gvn.transform(new ThreadLocalNode()); 502 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 503 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 504 505 // Test the should_post_on_exceptions_flag vs. 0 506 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 507 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 508 509 // Branch to slow_path if should_post_on_exceptions_flag was true 510 { BuildCutout unless(this, tst, PROB_MAX); 511 // Do not try anything fancy if we're notifying the VM on every throw. 512 // Cf. case Bytecodes::_athrow in parse2.cpp. 513 uncommon_trap(reason, Deoptimization::Action_none, 514 (ciKlass*)NULL, (char*)NULL, must_throw); 515 } 516 517 } 518 519 //------------------------------builtin_throw---------------------------------- 520 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 521 bool must_throw = true; 522 523 if (env()->jvmti_can_post_on_exceptions()) { 524 // check if we must post exception events, take uncommon trap if so 525 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 526 // here if should_post_on_exceptions is false 527 // continue on with the normal codegen 528 } 529 530 // If this particular condition has not yet happened at this 531 // bytecode, then use the uncommon trap mechanism, and allow for 532 // a future recompilation if several traps occur here. 533 // If the throw is hot, try to use a more complicated inline mechanism 534 // which keeps execution inside the compiled code. 535 bool treat_throw_as_hot = false; 536 ciMethodData* md = method()->method_data(); 537 538 if (ProfileTraps) { 539 if (too_many_traps(reason)) { 540 treat_throw_as_hot = true; 541 } 542 // (If there is no MDO at all, assume it is early in 543 // execution, and that any deopts are part of the 544 // startup transient, and don't need to be remembered.) 545 546 // Also, if there is a local exception handler, treat all throws 547 // as hot if there has been at least one in this method. 548 if (C->trap_count(reason) != 0 549 && method()->method_data()->trap_count(reason) != 0 550 && has_ex_handler()) { 551 treat_throw_as_hot = true; 552 } 553 } 554 555 // If this throw happens frequently, an uncommon trap might cause 556 // a performance pothole. If there is a local exception handler, 557 // and if this particular bytecode appears to be deoptimizing often, 558 // let us handle the throw inline, with a preconstructed instance. 559 // Note: If the deopt count has blown up, the uncommon trap 560 // runtime is going to flush this nmethod, not matter what. 561 if (treat_throw_as_hot 562 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 563 // If the throw is local, we use a pre-existing instance and 564 // punt on the backtrace. This would lead to a missing backtrace 565 // (a repeat of 4292742) if the backtrace object is ever asked 566 // for its backtrace. 567 // Fixing this remaining case of 4292742 requires some flavor of 568 // escape analysis. Leave that for the future. 569 ciInstance* ex_obj = NULL; 570 switch (reason) { 571 case Deoptimization::Reason_null_check: 572 ex_obj = env()->NullPointerException_instance(); 573 break; 574 case Deoptimization::Reason_div0_check: 575 ex_obj = env()->ArithmeticException_instance(); 576 break; 577 case Deoptimization::Reason_range_check: 578 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 579 break; 580 case Deoptimization::Reason_class_check: 581 if (java_bc() == Bytecodes::_aastore) { 582 ex_obj = env()->ArrayStoreException_instance(); 583 } else { 584 ex_obj = env()->ClassCastException_instance(); 585 } 586 break; 587 } 588 if (failing()) { stop(); return; } // exception allocation might fail 589 if (ex_obj != NULL) { 590 // Cheat with a preallocated exception object. 591 if (C->log() != NULL) 592 C->log()->elem("hot_throw preallocated='1' reason='%s'", 593 Deoptimization::trap_reason_name(reason)); 594 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 595 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 596 597 // Clear the detail message of the preallocated exception object. 598 // Weblogic sometimes mutates the detail message of exceptions 599 // using reflection. 600 int offset = java_lang_Throwable::get_detailMessage_offset(); 601 const TypePtr* adr_typ = ex_con->add_offset(offset); 602 603 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 604 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 605 // Conservatively release stores of object references. 606 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 607 608 add_exception_state(make_exception_state(ex_node)); 609 return; 610 } 611 } 612 613 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 614 // It won't be much cheaper than bailing to the interp., since we'll 615 // have to pass up all the debug-info, and the runtime will have to 616 // create the stack trace. 617 618 // Usual case: Bail to interpreter. 619 // Reserve the right to recompile if we haven't seen anything yet. 620 621 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 622 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 623 if (treat_throw_as_hot 624 && (method()->method_data()->trap_recompiled_at(bci(), m) 625 || C->too_many_traps(reason))) { 626 // We cannot afford to take more traps here. Suffer in the interpreter. 627 if (C->log() != NULL) 628 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 629 Deoptimization::trap_reason_name(reason), 630 C->trap_count(reason)); 631 action = Deoptimization::Action_none; 632 } 633 634 // "must_throw" prunes the JVM state to include only the stack, if there 635 // are no local exception handlers. This should cut down on register 636 // allocation time and code size, by drastically reducing the number 637 // of in-edges on the call to the uncommon trap. 638 639 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 640 } 641 642 643 //----------------------------PreserveJVMState--------------------------------- 644 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 645 debug_only(kit->verify_map()); 646 _kit = kit; 647 _map = kit->map(); // preserve the map 648 _sp = kit->sp(); 649 kit->set_map(clone_map ? kit->clone_map() : NULL); 650 #ifdef ASSERT 651 _bci = kit->bci(); 652 Parse* parser = kit->is_Parse(); 653 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 654 _block = block; 655 #endif 656 } 657 PreserveJVMState::~PreserveJVMState() { 658 GraphKit* kit = _kit; 659 #ifdef ASSERT 660 assert(kit->bci() == _bci, "bci must not shift"); 661 Parse* parser = kit->is_Parse(); 662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 663 assert(block == _block, "block must not shift"); 664 #endif 665 kit->set_map(_map); 666 kit->set_sp(_sp); 667 } 668 669 670 //-----------------------------BuildCutout------------------------------------- 671 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 672 : PreserveJVMState(kit) 673 { 674 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 675 SafePointNode* outer_map = _map; // preserved map is caller's 676 SafePointNode* inner_map = kit->map(); 677 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 678 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 679 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 680 } 681 BuildCutout::~BuildCutout() { 682 GraphKit* kit = _kit; 683 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 684 } 685 686 //---------------------------PreserveReexecuteState---------------------------- 687 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 688 assert(!kit->stopped(), "must call stopped() before"); 689 _kit = kit; 690 _sp = kit->sp(); 691 _reexecute = kit->jvms()->_reexecute; 692 } 693 PreserveReexecuteState::~PreserveReexecuteState() { 694 if (_kit->stopped()) return; 695 _kit->jvms()->_reexecute = _reexecute; 696 _kit->set_sp(_sp); 697 } 698 699 //------------------------------clone_map-------------------------------------- 700 // Implementation of PreserveJVMState 701 // 702 // Only clone_map(...) here. If this function is only used in the 703 // PreserveJVMState class we may want to get rid of this extra 704 // function eventually and do it all there. 705 706 SafePointNode* GraphKit::clone_map() { 707 if (map() == NULL) return NULL; 708 709 // Clone the memory edge first 710 Node* mem = MergeMemNode::make(map()->memory()); 711 gvn().set_type_bottom(mem); 712 713 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 714 JVMState* jvms = this->jvms(); 715 JVMState* clonejvms = jvms->clone_shallow(C); 716 clonemap->set_memory(mem); 717 clonemap->set_jvms(clonejvms); 718 clonejvms->set_map(clonemap); 719 record_for_igvn(clonemap); 720 gvn().set_type_bottom(clonemap); 721 return clonemap; 722 } 723 724 725 //-----------------------------set_map_clone----------------------------------- 726 void GraphKit::set_map_clone(SafePointNode* m) { 727 _map = m; 728 _map = clone_map(); 729 _map->set_next_exception(NULL); 730 debug_only(verify_map()); 731 } 732 733 734 //----------------------------kill_dead_locals--------------------------------- 735 // Detect any locals which are known to be dead, and force them to top. 736 void GraphKit::kill_dead_locals() { 737 // Consult the liveness information for the locals. If any 738 // of them are unused, then they can be replaced by top(). This 739 // should help register allocation time and cut down on the size 740 // of the deoptimization information. 741 742 // This call is made from many of the bytecode handling 743 // subroutines called from the Big Switch in do_one_bytecode. 744 // Every bytecode which might include a slow path is responsible 745 // for killing its dead locals. The more consistent we 746 // are about killing deads, the fewer useless phis will be 747 // constructed for them at various merge points. 748 749 // bci can be -1 (InvocationEntryBci). We return the entry 750 // liveness for the method. 751 752 if (method() == NULL || method()->code_size() == 0) { 753 // We are building a graph for a call to a native method. 754 // All locals are live. 755 return; 756 } 757 758 ResourceMark rm; 759 760 // Consult the liveness information for the locals. If any 761 // of them are unused, then they can be replaced by top(). This 762 // should help register allocation time and cut down on the size 763 // of the deoptimization information. 764 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 765 766 int len = (int)live_locals.size(); 767 assert(len <= jvms()->loc_size(), "too many live locals"); 768 for (int local = 0; local < len; local++) { 769 if (!live_locals.at(local)) { 770 set_local(local, top()); 771 } 772 } 773 } 774 775 #ifdef ASSERT 776 //-------------------------dead_locals_are_killed------------------------------ 777 // Return true if all dead locals are set to top in the map. 778 // Used to assert "clean" debug info at various points. 779 bool GraphKit::dead_locals_are_killed() { 780 if (method() == NULL || method()->code_size() == 0) { 781 // No locals need to be dead, so all is as it should be. 782 return true; 783 } 784 785 // Make sure somebody called kill_dead_locals upstream. 786 ResourceMark rm; 787 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 788 if (jvms->loc_size() == 0) continue; // no locals to consult 789 SafePointNode* map = jvms->map(); 790 ciMethod* method = jvms->method(); 791 int bci = jvms->bci(); 792 if (jvms == this->jvms()) { 793 bci = this->bci(); // it might not yet be synched 794 } 795 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 796 int len = (int)live_locals.size(); 797 if (!live_locals.is_valid() || len == 0) 798 // This method is trivial, or is poisoned by a breakpoint. 799 return true; 800 assert(len == jvms->loc_size(), "live map consistent with locals map"); 801 for (int local = 0; local < len; local++) { 802 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 803 if (PrintMiscellaneous && (Verbose || WizardMode)) { 804 tty->print_cr("Zombie local %d: ", local); 805 jvms->dump(); 806 } 807 return false; 808 } 809 } 810 } 811 return true; 812 } 813 814 #endif //ASSERT 815 816 // Helper function for enforcing certain bytecodes to reexecute if 817 // deoptimization happens 818 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 819 ciMethod* cur_method = jvms->has_method() ? jvms->method() : NULL; 820 int cur_bci = jvms->bci(); 821 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 822 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 823 return Interpreter::bytecode_should_reexecute(code) || 824 is_anewarray && code == Bytecodes::_multianewarray; 825 // Reexecute _multianewarray bytecode which was replaced with 826 // sequence of [a]newarray. See Parse::do_multianewarray(). 827 // 828 // Note: interpreter should not have it set since this optimization 829 // is limited by dimensions and guarded by flag so in some cases 830 // multianewarray() runtime calls will be generated and 831 // the bytecode should not be reexecutes (stack will not be reset). 832 } else 833 return false; 834 } 835 836 // Helper function for adding JVMState and debug information to node 837 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 838 // Add the safepoint edges to the call (or other safepoint). 839 840 // Make sure dead locals are set to top. This 841 // should help register allocation time and cut down on the size 842 // of the deoptimization information. 843 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 844 845 // Walk the inline list to fill in the correct set of JVMState's 846 // Also fill in the associated edges for each JVMState. 847 848 // If the bytecode needs to be reexecuted we need to put 849 // the arguments back on the stack. 850 const bool should_reexecute = jvms()->should_reexecute(); 851 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 852 853 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 854 // undefined if the bci is different. This is normal for Parse but it 855 // should not happen for LibraryCallKit because only one bci is processed. 856 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 857 "in LibraryCallKit the reexecute bit should not change"); 858 859 // If we are guaranteed to throw, we can prune everything but the 860 // input to the current bytecode. 861 bool can_prune_locals = false; 862 uint stack_slots_not_pruned = 0; 863 int inputs = 0, depth = 0; 864 if (must_throw) { 865 assert(method() == youngest_jvms->method(), "sanity"); 866 if (compute_stack_effects(inputs, depth)) { 867 can_prune_locals = true; 868 stack_slots_not_pruned = inputs; 869 } 870 } 871 872 if (env()->should_retain_local_variables()) { 873 // At any safepoint, this method can get breakpointed, which would 874 // then require an immediate deoptimization. 875 can_prune_locals = false; // do not prune locals 876 stack_slots_not_pruned = 0; 877 } 878 879 // do not scribble on the input jvms 880 JVMState* out_jvms = youngest_jvms->clone_deep(C); 881 call->set_jvms(out_jvms); // Start jvms list for call node 882 883 // For a known set of bytecodes, the interpreter should reexecute them if 884 // deoptimization happens. We set the reexecute state for them here 885 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 886 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 887 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 888 } 889 890 // Presize the call: 891 DEBUG_ONLY(uint non_debug_edges = call->req()); 892 call->add_req_batch(top(), youngest_jvms->debug_depth()); 893 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 894 895 // Set up edges so that the call looks like this: 896 // Call [state:] ctl io mem fptr retadr 897 // [parms:] parm0 ... parmN 898 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 899 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 900 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 901 // Note that caller debug info precedes callee debug info. 902 903 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 904 uint debug_ptr = call->req(); 905 906 // Loop over the map input edges associated with jvms, add them 907 // to the call node, & reset all offsets to match call node array. 908 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 909 uint debug_end = debug_ptr; 910 uint debug_start = debug_ptr - in_jvms->debug_size(); 911 debug_ptr = debug_start; // back up the ptr 912 913 uint p = debug_start; // walks forward in [debug_start, debug_end) 914 uint j, k, l; 915 SafePointNode* in_map = in_jvms->map(); 916 out_jvms->set_map(call); 917 918 if (can_prune_locals) { 919 assert(in_jvms->method() == out_jvms->method(), "sanity"); 920 // If the current throw can reach an exception handler in this JVMS, 921 // then we must keep everything live that can reach that handler. 922 // As a quick and dirty approximation, we look for any handlers at all. 923 if (in_jvms->method()->has_exception_handlers()) { 924 can_prune_locals = false; 925 } 926 } 927 928 // Add the Locals 929 k = in_jvms->locoff(); 930 l = in_jvms->loc_size(); 931 out_jvms->set_locoff(p); 932 if (!can_prune_locals) { 933 for (j = 0; j < l; j++) 934 call->set_req(p++, in_map->in(k+j)); 935 } else { 936 p += l; // already set to top above by add_req_batch 937 } 938 939 // Add the Expression Stack 940 k = in_jvms->stkoff(); 941 l = in_jvms->sp(); 942 out_jvms->set_stkoff(p); 943 if (!can_prune_locals) { 944 for (j = 0; j < l; j++) 945 call->set_req(p++, in_map->in(k+j)); 946 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 947 // Divide stack into {S0,...,S1}, where S0 is set to top. 948 uint s1 = stack_slots_not_pruned; 949 stack_slots_not_pruned = 0; // for next iteration 950 if (s1 > l) s1 = l; 951 uint s0 = l - s1; 952 p += s0; // skip the tops preinstalled by add_req_batch 953 for (j = s0; j < l; j++) 954 call->set_req(p++, in_map->in(k+j)); 955 } else { 956 p += l; // already set to top above by add_req_batch 957 } 958 959 // Add the Monitors 960 k = in_jvms->monoff(); 961 l = in_jvms->mon_size(); 962 out_jvms->set_monoff(p); 963 for (j = 0; j < l; j++) 964 call->set_req(p++, in_map->in(k+j)); 965 966 // Copy any scalar object fields. 967 k = in_jvms->scloff(); 968 l = in_jvms->scl_size(); 969 out_jvms->set_scloff(p); 970 for (j = 0; j < l; j++) 971 call->set_req(p++, in_map->in(k+j)); 972 973 // Finish the new jvms. 974 out_jvms->set_endoff(p); 975 976 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 977 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 978 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 979 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 980 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 981 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 982 983 // Update the two tail pointers in parallel. 984 out_jvms = out_jvms->caller(); 985 in_jvms = in_jvms->caller(); 986 } 987 988 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 989 990 // Test the correctness of JVMState::debug_xxx accessors: 991 assert(call->jvms()->debug_start() == non_debug_edges, ""); 992 assert(call->jvms()->debug_end() == call->req(), ""); 993 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 994 } 995 996 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 997 Bytecodes::Code code = java_bc(); 998 if (code == Bytecodes::_wide) { 999 code = method()->java_code_at_bci(bci() + 1); 1000 } 1001 1002 BasicType rtype = T_ILLEGAL; 1003 int rsize = 0; 1004 1005 if (code != Bytecodes::_illegal) { 1006 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1007 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1008 if (rtype < T_CONFLICT) 1009 rsize = type2size[rtype]; 1010 } 1011 1012 switch (code) { 1013 case Bytecodes::_illegal: 1014 return false; 1015 1016 case Bytecodes::_ldc: 1017 case Bytecodes::_ldc_w: 1018 case Bytecodes::_ldc2_w: 1019 inputs = 0; 1020 break; 1021 1022 case Bytecodes::_dup: inputs = 1; break; 1023 case Bytecodes::_dup_x1: inputs = 2; break; 1024 case Bytecodes::_dup_x2: inputs = 3; break; 1025 case Bytecodes::_dup2: inputs = 2; break; 1026 case Bytecodes::_dup2_x1: inputs = 3; break; 1027 case Bytecodes::_dup2_x2: inputs = 4; break; 1028 case Bytecodes::_swap: inputs = 2; break; 1029 case Bytecodes::_arraylength: inputs = 1; break; 1030 1031 case Bytecodes::_getstatic: 1032 case Bytecodes::_putstatic: 1033 case Bytecodes::_getfield: 1034 case Bytecodes::_putfield: 1035 { 1036 bool ignored_will_link; 1037 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1038 int size = field->type()->size(); 1039 bool is_get = (depth >= 0), is_static = (depth & 1); 1040 inputs = (is_static ? 0 : 1); 1041 if (is_get) { 1042 depth = size - inputs; 1043 } else { 1044 inputs += size; // putxxx pops the value from the stack 1045 depth = - inputs; 1046 } 1047 } 1048 break; 1049 1050 case Bytecodes::_invokevirtual: 1051 case Bytecodes::_invokespecial: 1052 case Bytecodes::_invokestatic: 1053 case Bytecodes::_invokedynamic: 1054 case Bytecodes::_invokeinterface: 1055 { 1056 bool ignored_will_link; 1057 ciSignature* declared_signature = NULL; 1058 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1059 assert(declared_signature != NULL, "cannot be null"); 1060 inputs = declared_signature->arg_size_for_bc(code); 1061 int size = declared_signature->return_type()->size(); 1062 depth = size - inputs; 1063 } 1064 break; 1065 1066 case Bytecodes::_multianewarray: 1067 { 1068 ciBytecodeStream iter(method()); 1069 iter.reset_to_bci(bci()); 1070 iter.next(); 1071 inputs = iter.get_dimensions(); 1072 assert(rsize == 1, ""); 1073 depth = rsize - inputs; 1074 } 1075 break; 1076 1077 case Bytecodes::_ireturn: 1078 case Bytecodes::_lreturn: 1079 case Bytecodes::_freturn: 1080 case Bytecodes::_dreturn: 1081 case Bytecodes::_areturn: 1082 assert(rsize = -depth, ""); 1083 inputs = rsize; 1084 break; 1085 1086 case Bytecodes::_jsr: 1087 case Bytecodes::_jsr_w: 1088 inputs = 0; 1089 depth = 1; // S.B. depth=1, not zero 1090 break; 1091 1092 default: 1093 // bytecode produces a typed result 1094 inputs = rsize - depth; 1095 assert(inputs >= 0, ""); 1096 break; 1097 } 1098 1099 #ifdef ASSERT 1100 // spot check 1101 int outputs = depth + inputs; 1102 assert(outputs >= 0, "sanity"); 1103 switch (code) { 1104 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1105 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1106 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1107 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1108 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1109 } 1110 #endif //ASSERT 1111 1112 return true; 1113 } 1114 1115 1116 1117 //------------------------------basic_plus_adr--------------------------------- 1118 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1119 // short-circuit a common case 1120 if (offset == intcon(0)) return ptr; 1121 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1122 } 1123 1124 Node* GraphKit::ConvI2L(Node* offset) { 1125 // short-circuit a common case 1126 jint offset_con = find_int_con(offset, Type::OffsetBot); 1127 if (offset_con != Type::OffsetBot) { 1128 return longcon((jlong) offset_con); 1129 } 1130 return _gvn.transform( new ConvI2LNode(offset)); 1131 } 1132 1133 Node* GraphKit::ConvI2UL(Node* offset) { 1134 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1135 if (offset_con != (juint) Type::OffsetBot) { 1136 return longcon((julong) offset_con); 1137 } 1138 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1139 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1140 return _gvn.transform( new AndLNode(conv, mask) ); 1141 } 1142 1143 Node* GraphKit::ConvL2I(Node* offset) { 1144 // short-circuit a common case 1145 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1146 if (offset_con != (jlong)Type::OffsetBot) { 1147 return intcon((int) offset_con); 1148 } 1149 return _gvn.transform( new ConvL2INode(offset)); 1150 } 1151 1152 //-------------------------load_object_klass----------------------------------- 1153 Node* GraphKit::load_object_klass(Node* obj) { 1154 // Special-case a fresh allocation to avoid building nodes: 1155 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1156 if (akls != NULL) return akls; 1157 if (ShenandoahVerifyReadsToFromSpace) { 1158 obj = shenandoah_read_barrier(obj); 1159 } 1160 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1161 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1162 } 1163 1164 //-------------------------load_array_length----------------------------------- 1165 Node* GraphKit::load_array_length(Node* array) { 1166 // Special-case a fresh allocation to avoid building nodes: 1167 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1168 Node *alen; 1169 if (alloc == NULL) { 1170 if (ShenandoahVerifyReadsToFromSpace) { 1171 array = shenandoah_read_barrier(array); 1172 } 1173 1174 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1175 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1176 } else { 1177 alen = alloc->Ideal_length(); 1178 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1179 if (ccast != alen) { 1180 alen = _gvn.transform(ccast); 1181 } 1182 } 1183 return alen; 1184 } 1185 1186 //------------------------------do_null_check---------------------------------- 1187 // Helper function to do a NULL pointer check. Returned value is 1188 // the incoming address with NULL casted away. You are allowed to use the 1189 // not-null value only if you are control dependent on the test. 1190 extern int explicit_null_checks_inserted, 1191 explicit_null_checks_elided; 1192 Node* GraphKit::null_check_common(Node* value, BasicType type, 1193 // optional arguments for variations: 1194 bool assert_null, 1195 Node* *null_control, 1196 bool speculative) { 1197 assert(!assert_null || null_control == NULL, "not both at once"); 1198 if (stopped()) return top(); 1199 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1200 // For some performance testing, we may wish to suppress null checking. 1201 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1202 return value; 1203 } 1204 explicit_null_checks_inserted++; 1205 1206 // Construct NULL check 1207 Node *chk = NULL; 1208 switch(type) { 1209 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1210 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1211 case T_ARRAY : // fall through 1212 type = T_OBJECT; // simplify further tests 1213 case T_OBJECT : { 1214 const Type *t = _gvn.type( value ); 1215 1216 const TypeOopPtr* tp = t->isa_oopptr(); 1217 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1218 // Only for do_null_check, not any of its siblings: 1219 && !assert_null && null_control == NULL) { 1220 // Usually, any field access or invocation on an unloaded oop type 1221 // will simply fail to link, since the statically linked class is 1222 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1223 // the static class is loaded but the sharper oop type is not. 1224 // Rather than checking for this obscure case in lots of places, 1225 // we simply observe that a null check on an unloaded class 1226 // will always be followed by a nonsense operation, so we 1227 // can just issue the uncommon trap here. 1228 // Our access to the unloaded class will only be correct 1229 // after it has been loaded and initialized, which requires 1230 // a trip through the interpreter. 1231 #ifndef PRODUCT 1232 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1233 #endif 1234 uncommon_trap(Deoptimization::Reason_unloaded, 1235 Deoptimization::Action_reinterpret, 1236 tp->klass(), "!loaded"); 1237 return top(); 1238 } 1239 1240 if (assert_null) { 1241 // See if the type is contained in NULL_PTR. 1242 // If so, then the value is already null. 1243 if (t->higher_equal(TypePtr::NULL_PTR)) { 1244 explicit_null_checks_elided++; 1245 return value; // Elided null assert quickly! 1246 } 1247 } else { 1248 // See if mixing in the NULL pointer changes type. 1249 // If so, then the NULL pointer was not allowed in the original 1250 // type. In other words, "value" was not-null. 1251 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1252 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1253 explicit_null_checks_elided++; 1254 return value; // Elided null check quickly! 1255 } 1256 } 1257 chk = new CmpPNode( value, null() ); 1258 break; 1259 } 1260 1261 default: 1262 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1263 } 1264 assert(chk != NULL, "sanity check"); 1265 chk = _gvn.transform(chk); 1266 1267 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1268 BoolNode *btst = new BoolNode( chk, btest); 1269 Node *tst = _gvn.transform( btst ); 1270 1271 //----------- 1272 // if peephole optimizations occurred, a prior test existed. 1273 // If a prior test existed, maybe it dominates as we can avoid this test. 1274 if (tst != btst && type == T_OBJECT) { 1275 // At this point we want to scan up the CFG to see if we can 1276 // find an identical test (and so avoid this test altogether). 1277 Node *cfg = control(); 1278 int depth = 0; 1279 while( depth < 16 ) { // Limit search depth for speed 1280 if( cfg->Opcode() == Op_IfTrue && 1281 cfg->in(0)->in(1) == tst ) { 1282 // Found prior test. Use "cast_not_null" to construct an identical 1283 // CastPP (and hence hash to) as already exists for the prior test. 1284 // Return that casted value. 1285 if (assert_null) { 1286 replace_in_map(value, null()); 1287 return null(); // do not issue the redundant test 1288 } 1289 Node *oldcontrol = control(); 1290 set_control(cfg); 1291 Node *res = cast_not_null(value); 1292 set_control(oldcontrol); 1293 explicit_null_checks_elided++; 1294 return res; 1295 } 1296 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1297 if (cfg == NULL) break; // Quit at region nodes 1298 depth++; 1299 } 1300 } 1301 1302 //----------- 1303 // Branch to failure if null 1304 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1305 Deoptimization::DeoptReason reason; 1306 if (assert_null) { 1307 reason = Deoptimization::Reason_null_assert; 1308 } else if (type == T_OBJECT) { 1309 reason = Deoptimization::reason_null_check(speculative); 1310 } else { 1311 reason = Deoptimization::Reason_div0_check; 1312 } 1313 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1314 // ciMethodData::has_trap_at will return a conservative -1 if any 1315 // must-be-null assertion has failed. This could cause performance 1316 // problems for a method after its first do_null_assert failure. 1317 // Consider using 'Reason_class_check' instead? 1318 1319 // To cause an implicit null check, we set the not-null probability 1320 // to the maximum (PROB_MAX). For an explicit check the probability 1321 // is set to a smaller value. 1322 if (null_control != NULL || too_many_traps(reason)) { 1323 // probability is less likely 1324 ok_prob = PROB_LIKELY_MAG(3); 1325 } else if (!assert_null && 1326 (ImplicitNullCheckThreshold > 0) && 1327 method() != NULL && 1328 (method()->method_data()->trap_count(reason) 1329 >= (uint)ImplicitNullCheckThreshold)) { 1330 ok_prob = PROB_LIKELY_MAG(3); 1331 } 1332 1333 if (null_control != NULL) { 1334 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1335 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1336 set_control( _gvn.transform( new IfTrueNode(iff))); 1337 if (null_true == top()) 1338 explicit_null_checks_elided++; 1339 (*null_control) = null_true; 1340 } else { 1341 BuildCutout unless(this, tst, ok_prob); 1342 // Check for optimizer eliding test at parse time 1343 if (stopped()) { 1344 // Failure not possible; do not bother making uncommon trap. 1345 explicit_null_checks_elided++; 1346 } else if (assert_null) { 1347 uncommon_trap(reason, 1348 Deoptimization::Action_make_not_entrant, 1349 NULL, "assert_null"); 1350 } else { 1351 replace_in_map(value, zerocon(type)); 1352 builtin_throw(reason); 1353 } 1354 } 1355 1356 // Must throw exception, fall-thru not possible? 1357 if (stopped()) { 1358 return top(); // No result 1359 } 1360 1361 if (assert_null) { 1362 // Cast obj to null on this path. 1363 replace_in_map(value, zerocon(type)); 1364 return zerocon(type); 1365 } 1366 1367 // Cast obj to not-null on this path, if there is no null_control. 1368 // (If there is a null_control, a non-null value may come back to haunt us.) 1369 if (type == T_OBJECT) { 1370 Node* cast = cast_not_null(value, false); 1371 if (null_control == NULL || (*null_control) == top()) 1372 replace_in_map(value, cast); 1373 value = cast; 1374 } 1375 1376 return value; 1377 } 1378 1379 1380 //------------------------------cast_not_null---------------------------------- 1381 // Cast obj to not-null on this path 1382 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1383 const Type *t = _gvn.type(obj); 1384 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1385 // Object is already not-null? 1386 if( t == t_not_null ) return obj; 1387 1388 Node *cast = new CastPPNode(obj,t_not_null); 1389 cast->init_req(0, control()); 1390 cast = _gvn.transform( cast ); 1391 1392 // Scan for instances of 'obj' in the current JVM mapping. 1393 // These instances are known to be not-null after the test. 1394 if (do_replace_in_map) 1395 replace_in_map(obj, cast); 1396 1397 return cast; // Return casted value 1398 } 1399 1400 1401 //--------------------------replace_in_map------------------------------------- 1402 void GraphKit::replace_in_map(Node* old, Node* neww) { 1403 if (old == neww) { 1404 return; 1405 } 1406 1407 map()->replace_edge(old, neww); 1408 1409 // Note: This operation potentially replaces any edge 1410 // on the map. This includes locals, stack, and monitors 1411 // of the current (innermost) JVM state. 1412 1413 // don't let inconsistent types from profiling escape this 1414 // method 1415 1416 const Type* told = _gvn.type(old); 1417 const Type* tnew = _gvn.type(neww); 1418 1419 if (!tnew->higher_equal(told)) { 1420 return; 1421 } 1422 1423 map()->record_replaced_node(old, neww); 1424 } 1425 1426 1427 //============================================================================= 1428 //--------------------------------memory--------------------------------------- 1429 Node* GraphKit::memory(uint alias_idx) { 1430 MergeMemNode* mem = merged_memory(); 1431 Node* p = mem->memory_at(alias_idx); 1432 _gvn.set_type(p, Type::MEMORY); // must be mapped 1433 return p; 1434 } 1435 1436 //-----------------------------reset_memory------------------------------------ 1437 Node* GraphKit::reset_memory() { 1438 Node* mem = map()->memory(); 1439 // do not use this node for any more parsing! 1440 debug_only( map()->set_memory((Node*)NULL) ); 1441 return _gvn.transform( mem ); 1442 } 1443 1444 //------------------------------set_all_memory--------------------------------- 1445 void GraphKit::set_all_memory(Node* newmem) { 1446 Node* mergemem = MergeMemNode::make(newmem); 1447 gvn().set_type_bottom(mergemem); 1448 map()->set_memory(mergemem); 1449 } 1450 1451 //------------------------------set_all_memory_call---------------------------- 1452 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1453 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1454 set_all_memory(newmem); 1455 } 1456 1457 //============================================================================= 1458 // 1459 // parser factory methods for MemNodes 1460 // 1461 // These are layered on top of the factory methods in LoadNode and StoreNode, 1462 // and integrate with the parser's memory state and _gvn engine. 1463 // 1464 1465 // factory methods in "int adr_idx" 1466 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1467 int adr_idx, 1468 MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency, bool require_atomic_access) { 1469 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1470 const TypePtr* adr_type = NULL; // debug-mode-only argument 1471 debug_only(adr_type = C->get_adr_type(adr_idx)); 1472 Node* mem = memory(adr_idx); 1473 Node* ld; 1474 if (require_atomic_access && bt == T_LONG) { 1475 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency); 1476 } else if (require_atomic_access && bt == T_DOUBLE) { 1477 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency); 1478 } else { 1479 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency); 1480 } 1481 ld = _gvn.transform(ld); 1482 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1483 // Improve graph before escape analysis and boxing elimination. 1484 record_for_igvn(ld); 1485 } 1486 return ld; 1487 } 1488 1489 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1490 int adr_idx, 1491 MemNode::MemOrd mo, 1492 bool require_atomic_access) { 1493 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1494 const TypePtr* adr_type = NULL; 1495 debug_only(adr_type = C->get_adr_type(adr_idx)); 1496 Node *mem = memory(adr_idx); 1497 Node* st; 1498 if (require_atomic_access && bt == T_LONG) { 1499 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1500 } else if (require_atomic_access && bt == T_DOUBLE) { 1501 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1502 } else { 1503 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1504 } 1505 st = _gvn.transform(st); 1506 set_memory(st, adr_idx); 1507 // Back-to-back stores can only remove intermediate store with DU info 1508 // so push on worklist for optimizer. 1509 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1510 record_for_igvn(st); 1511 1512 return st; 1513 } 1514 1515 1516 void GraphKit::pre_barrier(bool do_load, 1517 Node* ctl, 1518 Node* obj, 1519 Node* adr, 1520 uint adr_idx, 1521 Node* val, 1522 const TypeOopPtr* val_type, 1523 Node* pre_val, 1524 BasicType bt) { 1525 1526 BarrierSet* bs = Universe::heap()->barrier_set(); 1527 set_control(ctl); 1528 switch (bs->kind()) { 1529 case BarrierSet::G1SATBCTLogging: 1530 case BarrierSet::ShenandoahBarrierSet: 1531 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1532 break; 1533 1534 case BarrierSet::CardTableForRS: 1535 case BarrierSet::CardTableExtension: 1536 case BarrierSet::ModRef: 1537 break; 1538 1539 default : 1540 ShouldNotReachHere(); 1541 1542 } 1543 } 1544 1545 bool GraphKit::can_move_pre_barrier() const { 1546 BarrierSet* bs = Universe::heap()->barrier_set(); 1547 switch (bs->kind()) { 1548 case BarrierSet::G1SATBCTLogging: 1549 case BarrierSet::ShenandoahBarrierSet: 1550 return true; // Can move it if no safepoint 1551 1552 case BarrierSet::CardTableForRS: 1553 case BarrierSet::CardTableExtension: 1554 case BarrierSet::ModRef: 1555 return true; // There is no pre-barrier 1556 1557 default : 1558 ShouldNotReachHere(); 1559 } 1560 return false; 1561 } 1562 1563 void GraphKit::post_barrier(Node* ctl, 1564 Node* store, 1565 Node* obj, 1566 Node* adr, 1567 uint adr_idx, 1568 Node* val, 1569 BasicType bt, 1570 bool use_precise) { 1571 BarrierSet* bs = Universe::heap()->barrier_set(); 1572 set_control(ctl); 1573 switch (bs->kind()) { 1574 case BarrierSet::G1SATBCTLogging: 1575 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1576 break; 1577 1578 case BarrierSet::CardTableForRS: 1579 case BarrierSet::CardTableExtension: 1580 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1581 break; 1582 1583 case BarrierSet::ModRef: 1584 case BarrierSet::ShenandoahBarrierSet: 1585 break; 1586 1587 default : 1588 ShouldNotReachHere(); 1589 1590 } 1591 } 1592 1593 Node* GraphKit::store_oop(Node* ctl, 1594 Node* obj, 1595 Node* adr, 1596 const TypePtr* adr_type, 1597 Node* val, 1598 const TypeOopPtr* val_type, 1599 BasicType bt, 1600 bool use_precise, 1601 MemNode::MemOrd mo) { 1602 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1603 // could be delayed during Parse (for example, in adjust_map_after_if()). 1604 // Execute transformation here to avoid barrier generation in such case. 1605 if (_gvn.type(val) == TypePtr::NULL_PTR) 1606 val = _gvn.makecon(TypePtr::NULL_PTR); 1607 1608 set_control(ctl); 1609 if (stopped()) return top(); // Dead path ? 1610 1611 assert(bt == T_OBJECT, "sanity"); 1612 assert(val != NULL, "not dead path"); 1613 uint adr_idx = C->get_alias_index(adr_type); 1614 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1615 1616 pre_barrier(true /* do_load */, 1617 control(), obj, adr, adr_idx, val, val_type, 1618 NULL /* pre_val */, 1619 bt); 1620 1621 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1622 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1623 return store; 1624 } 1625 1626 // Could be an array or object we don't know at compile time (unsafe ref.) 1627 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1628 Node* obj, // containing obj 1629 Node* adr, // actual adress to store val at 1630 const TypePtr* adr_type, 1631 Node* val, 1632 BasicType bt, 1633 MemNode::MemOrd mo) { 1634 Compile::AliasType* at = C->alias_type(adr_type); 1635 const TypeOopPtr* val_type = NULL; 1636 if (adr_type->isa_instptr()) { 1637 if (at->field() != NULL) { 1638 // known field. This code is a copy of the do_put_xxx logic. 1639 ciField* field = at->field(); 1640 if (!field->type()->is_loaded()) { 1641 val_type = TypeInstPtr::BOTTOM; 1642 } else { 1643 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1644 } 1645 } 1646 } else if (adr_type->isa_aryptr()) { 1647 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1648 } 1649 if (val_type == NULL) { 1650 val_type = TypeInstPtr::BOTTOM; 1651 } 1652 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1653 } 1654 1655 1656 //-------------------------array_element_address------------------------- 1657 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1658 const TypeInt* sizetype) { 1659 uint shift = exact_log2(type2aelembytes(elembt)); 1660 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1661 1662 // short-circuit a common case (saves lots of confusing waste motion) 1663 jint idx_con = find_int_con(idx, -1); 1664 if (idx_con >= 0) { 1665 intptr_t offset = header + ((intptr_t)idx_con << shift); 1666 return basic_plus_adr(ary, offset); 1667 } 1668 1669 // must be correct type for alignment purposes 1670 Node* base = basic_plus_adr(ary, header); 1671 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype); 1672 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1673 return basic_plus_adr(ary, base, scale); 1674 } 1675 1676 //-------------------------load_array_element------------------------- 1677 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1678 const Type* elemtype = arytype->elem(); 1679 BasicType elembt = elemtype->array_element_basic_type(); 1680 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1681 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1682 return ld; 1683 } 1684 1685 //-------------------------set_arguments_for_java_call------------------------- 1686 // Arguments (pre-popped from the stack) are taken from the JVMS. 1687 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1688 // Add the call arguments: 1689 uint nargs = call->method()->arg_size(); 1690 for (uint i = 0; i < nargs; i++) { 1691 Node* arg = argument(i); 1692 if (ShenandoahVerifyReadsToFromSpace && call->is_CallDynamicJava() && i == 0) { 1693 arg = shenandoah_read_barrier(arg); 1694 } 1695 call->init_req(i + TypeFunc::Parms, arg); 1696 } 1697 } 1698 1699 //---------------------------set_edges_for_java_call--------------------------- 1700 // Connect a newly created call into the current JVMS. 1701 // A return value node (if any) is returned from set_edges_for_java_call. 1702 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1703 1704 // Add the predefined inputs: 1705 call->init_req( TypeFunc::Control, control() ); 1706 call->init_req( TypeFunc::I_O , i_o() ); 1707 call->init_req( TypeFunc::Memory , reset_memory() ); 1708 call->init_req( TypeFunc::FramePtr, frameptr() ); 1709 call->init_req( TypeFunc::ReturnAdr, top() ); 1710 1711 add_safepoint_edges(call, must_throw); 1712 1713 Node* xcall = _gvn.transform(call); 1714 1715 if (xcall == top()) { 1716 set_control(top()); 1717 return; 1718 } 1719 assert(xcall == call, "call identity is stable"); 1720 1721 // Re-use the current map to produce the result. 1722 1723 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1724 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1725 set_all_memory_call(xcall, separate_io_proj); 1726 1727 //return xcall; // no need, caller already has it 1728 } 1729 1730 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1731 if (stopped()) return top(); // maybe the call folded up? 1732 1733 // Capture the return value, if any. 1734 Node* ret; 1735 if (call->method() == NULL || 1736 call->method()->return_type()->basic_type() == T_VOID) 1737 ret = top(); 1738 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1739 1740 // Note: Since any out-of-line call can produce an exception, 1741 // we always insert an I_O projection from the call into the result. 1742 1743 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1744 1745 if (separate_io_proj) { 1746 // The caller requested separate projections be used by the fall 1747 // through and exceptional paths, so replace the projections for 1748 // the fall through path. 1749 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1750 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1751 } 1752 return ret; 1753 } 1754 1755 //--------------------set_predefined_input_for_runtime_call-------------------- 1756 // Reading and setting the memory state is way conservative here. 1757 // The real problem is that I am not doing real Type analysis on memory, 1758 // so I cannot distinguish card mark stores from other stores. Across a GC 1759 // point the Store Barrier and the card mark memory has to agree. I cannot 1760 // have a card mark store and its barrier split across the GC point from 1761 // either above or below. Here I get that to happen by reading ALL of memory. 1762 // A better answer would be to separate out card marks from other memory. 1763 // For now, return the input memory state, so that it can be reused 1764 // after the call, if this call has restricted memory effects. 1765 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1766 // Set fixed predefined input arguments 1767 Node* memory = reset_memory(); 1768 call->init_req( TypeFunc::Control, control() ); 1769 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1770 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1771 call->init_req( TypeFunc::FramePtr, frameptr() ); 1772 call->init_req( TypeFunc::ReturnAdr, top() ); 1773 return memory; 1774 } 1775 1776 //-------------------set_predefined_output_for_runtime_call-------------------- 1777 // Set control and memory (not i_o) from the call. 1778 // If keep_mem is not NULL, use it for the output state, 1779 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1780 // If hook_mem is NULL, this call produces no memory effects at all. 1781 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1782 // then only that memory slice is taken from the call. 1783 // In the last case, we must put an appropriate memory barrier before 1784 // the call, so as to create the correct anti-dependencies on loads 1785 // preceding the call. 1786 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1787 Node* keep_mem, 1788 const TypePtr* hook_mem) { 1789 // no i/o 1790 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1791 if (keep_mem) { 1792 // First clone the existing memory state 1793 set_all_memory(keep_mem); 1794 if (hook_mem != NULL) { 1795 // Make memory for the call 1796 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1797 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1798 // We also use hook_mem to extract specific effects from arraycopy stubs. 1799 set_memory(mem, hook_mem); 1800 } 1801 // ...else the call has NO memory effects. 1802 1803 // Make sure the call advertises its memory effects precisely. 1804 // This lets us build accurate anti-dependences in gcm.cpp. 1805 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1806 "call node must be constructed correctly"); 1807 } else { 1808 assert(hook_mem == NULL, ""); 1809 // This is not a "slow path" call; all memory comes from the call. 1810 set_all_memory_call(call); 1811 } 1812 } 1813 1814 1815 // Replace the call with the current state of the kit. 1816 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1817 JVMState* ejvms = NULL; 1818 if (has_exceptions()) { 1819 ejvms = transfer_exceptions_into_jvms(); 1820 } 1821 1822 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1823 ReplacedNodes replaced_nodes_exception; 1824 Node* ex_ctl = top(); 1825 1826 SafePointNode* final_state = stop(); 1827 1828 // Find all the needed outputs of this call 1829 CallProjections callprojs; 1830 call->extract_projections(&callprojs, true); 1831 1832 Node* init_mem = call->in(TypeFunc::Memory); 1833 Node* final_mem = final_state->in(TypeFunc::Memory); 1834 Node* final_ctl = final_state->in(TypeFunc::Control); 1835 Node* final_io = final_state->in(TypeFunc::I_O); 1836 1837 // Replace all the old call edges with the edges from the inlining result 1838 if (callprojs.fallthrough_catchproj != NULL) { 1839 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1840 } 1841 if (callprojs.fallthrough_memproj != NULL) { 1842 if (final_mem->is_MergeMem()) { 1843 // Parser's exits MergeMem was not transformed but may be optimized 1844 final_mem = _gvn.transform(final_mem); 1845 } 1846 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1847 } 1848 if (callprojs.fallthrough_ioproj != NULL) { 1849 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1850 } 1851 1852 // Replace the result with the new result if it exists and is used 1853 if (callprojs.resproj != NULL && result != NULL) { 1854 C->gvn_replace_by(callprojs.resproj, result); 1855 } 1856 1857 if (ejvms == NULL) { 1858 // No exception edges to simply kill off those paths 1859 if (callprojs.catchall_catchproj != NULL) { 1860 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1861 } 1862 if (callprojs.catchall_memproj != NULL) { 1863 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1864 } 1865 if (callprojs.catchall_ioproj != NULL) { 1866 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1867 } 1868 // Replace the old exception object with top 1869 if (callprojs.exobj != NULL) { 1870 C->gvn_replace_by(callprojs.exobj, C->top()); 1871 } 1872 } else { 1873 GraphKit ekit(ejvms); 1874 1875 // Load my combined exception state into the kit, with all phis transformed: 1876 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1877 replaced_nodes_exception = ex_map->replaced_nodes(); 1878 1879 Node* ex_oop = ekit.use_exception_state(ex_map); 1880 1881 if (callprojs.catchall_catchproj != NULL) { 1882 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1883 ex_ctl = ekit.control(); 1884 } 1885 if (callprojs.catchall_memproj != NULL) { 1886 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1887 } 1888 if (callprojs.catchall_ioproj != NULL) { 1889 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1890 } 1891 1892 // Replace the old exception object with the newly created one 1893 if (callprojs.exobj != NULL) { 1894 C->gvn_replace_by(callprojs.exobj, ex_oop); 1895 } 1896 } 1897 1898 // Disconnect the call from the graph 1899 call->disconnect_inputs(NULL, C); 1900 C->gvn_replace_by(call, C->top()); 1901 1902 // Clean up any MergeMems that feed other MergeMems since the 1903 // optimizer doesn't like that. 1904 if (final_mem->is_MergeMem()) { 1905 Node_List wl; 1906 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1907 Node* m = i.get(); 1908 if (m->is_MergeMem() && !wl.contains(m)) { 1909 wl.push(m); 1910 } 1911 } 1912 while (wl.size() > 0) { 1913 _gvn.transform(wl.pop()); 1914 } 1915 } 1916 1917 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1918 replaced_nodes.apply(C, final_ctl); 1919 } 1920 if (!ex_ctl->is_top() && do_replaced_nodes) { 1921 replaced_nodes_exception.apply(C, ex_ctl); 1922 } 1923 } 1924 1925 1926 //------------------------------increment_counter------------------------------ 1927 // for statistics: increment a VM counter by 1 1928 1929 void GraphKit::increment_counter(address counter_addr) { 1930 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1931 increment_counter(adr1); 1932 } 1933 1934 void GraphKit::increment_counter(Node* counter_addr) { 1935 int adr_type = Compile::AliasIdxRaw; 1936 Node* ctrl = control(); 1937 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1938 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1939 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1940 } 1941 1942 1943 //------------------------------uncommon_trap---------------------------------- 1944 // Bail out to the interpreter in mid-method. Implemented by calling the 1945 // uncommon_trap blob. This helper function inserts a runtime call with the 1946 // right debug info. 1947 void GraphKit::uncommon_trap(int trap_request, 1948 ciKlass* klass, const char* comment, 1949 bool must_throw, 1950 bool keep_exact_action) { 1951 if (failing()) stop(); 1952 if (stopped()) return; // trap reachable? 1953 1954 // Note: If ProfileTraps is true, and if a deopt. actually 1955 // occurs here, the runtime will make sure an MDO exists. There is 1956 // no need to call method()->ensure_method_data() at this point. 1957 1958 // Set the stack pointer to the right value for reexecution: 1959 set_sp(reexecute_sp()); 1960 1961 #ifdef ASSERT 1962 if (!must_throw) { 1963 // Make sure the stack has at least enough depth to execute 1964 // the current bytecode. 1965 int inputs, ignored_depth; 1966 if (compute_stack_effects(inputs, ignored_depth)) { 1967 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1968 Bytecodes::name(java_bc()), sp(), inputs)); 1969 } 1970 } 1971 #endif 1972 1973 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1974 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1975 1976 switch (action) { 1977 case Deoptimization::Action_maybe_recompile: 1978 case Deoptimization::Action_reinterpret: 1979 // Temporary fix for 6529811 to allow virtual calls to be sure they 1980 // get the chance to go from mono->bi->mega 1981 if (!keep_exact_action && 1982 Deoptimization::trap_request_index(trap_request) < 0 && 1983 too_many_recompiles(reason)) { 1984 // This BCI is causing too many recompilations. 1985 if (C->log() != NULL) { 1986 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1987 Deoptimization::trap_reason_name(reason), 1988 Deoptimization::trap_action_name(action)); 1989 } 1990 action = Deoptimization::Action_none; 1991 trap_request = Deoptimization::make_trap_request(reason, action); 1992 } else { 1993 C->set_trap_can_recompile(true); 1994 } 1995 break; 1996 case Deoptimization::Action_make_not_entrant: 1997 C->set_trap_can_recompile(true); 1998 break; 1999 #ifdef ASSERT 2000 case Deoptimization::Action_none: 2001 case Deoptimization::Action_make_not_compilable: 2002 break; 2003 default: 2004 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2005 break; 2006 #endif 2007 } 2008 2009 if (TraceOptoParse) { 2010 char buf[100]; 2011 tty->print_cr("Uncommon trap %s at bci:%d", 2012 Deoptimization::format_trap_request(buf, sizeof(buf), 2013 trap_request), bci()); 2014 } 2015 2016 CompileLog* log = C->log(); 2017 if (log != NULL) { 2018 int kid = (klass == NULL)? -1: log->identify(klass); 2019 log->begin_elem("uncommon_trap bci='%d'", bci()); 2020 char buf[100]; 2021 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2022 trap_request)); 2023 if (kid >= 0) log->print(" klass='%d'", kid); 2024 if (comment != NULL) log->print(" comment='%s'", comment); 2025 log->end_elem(); 2026 } 2027 2028 // Make sure any guarding test views this path as very unlikely 2029 Node *i0 = control()->in(0); 2030 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2031 IfNode *iff = i0->as_If(); 2032 float f = iff->_prob; // Get prob 2033 if (control()->Opcode() == Op_IfTrue) { 2034 if (f > PROB_UNLIKELY_MAG(4)) 2035 iff->_prob = PROB_MIN; 2036 } else { 2037 if (f < PROB_LIKELY_MAG(4)) 2038 iff->_prob = PROB_MAX; 2039 } 2040 } 2041 2042 // Clear out dead values from the debug info. 2043 kill_dead_locals(); 2044 2045 // Now insert the uncommon trap subroutine call 2046 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2047 const TypePtr* no_memory_effects = NULL; 2048 // Pass the index of the class to be loaded 2049 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2050 (must_throw ? RC_MUST_THROW : 0), 2051 OptoRuntime::uncommon_trap_Type(), 2052 call_addr, "uncommon_trap", no_memory_effects, 2053 intcon(trap_request)); 2054 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2055 "must extract request correctly from the graph"); 2056 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2057 2058 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2059 // The debug info is the only real input to this call. 2060 2061 // Halt-and-catch fire here. The above call should never return! 2062 HaltNode* halt = new HaltNode(control(), frameptr()); 2063 _gvn.set_type_bottom(halt); 2064 root()->add_req(halt); 2065 2066 stop_and_kill_map(); 2067 } 2068 2069 2070 //--------------------------just_allocated_object------------------------------ 2071 // Report the object that was just allocated. 2072 // It must be the case that there are no intervening safepoints. 2073 // We use this to determine if an object is so "fresh" that 2074 // it does not require card marks. 2075 Node* GraphKit::just_allocated_object(Node* current_control) { 2076 if (C->recent_alloc_ctl() == current_control) 2077 return C->recent_alloc_obj(); 2078 return NULL; 2079 } 2080 2081 2082 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2083 // (Note: TypeFunc::make has a cache that makes this fast.) 2084 const TypeFunc* tf = TypeFunc::make(dest_method); 2085 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2086 for (int j = 0; j < nargs; j++) { 2087 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2088 if( targ->basic_type() == T_DOUBLE ) { 2089 // If any parameters are doubles, they must be rounded before 2090 // the call, dstore_rounding does gvn.transform 2091 Node *arg = argument(j); 2092 arg = dstore_rounding(arg); 2093 set_argument(j, arg); 2094 } 2095 } 2096 } 2097 2098 /** 2099 * Record profiling data exact_kls for Node n with the type system so 2100 * that it can propagate it (speculation) 2101 * 2102 * @param n node that the type applies to 2103 * @param exact_kls type from profiling 2104 * @param maybe_null did profiling see null? 2105 * 2106 * @return node with improved type 2107 */ 2108 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2109 const Type* current_type = _gvn.type(n); 2110 assert(UseTypeSpeculation, "type speculation must be on"); 2111 2112 const TypePtr* speculative = current_type->speculative(); 2113 2114 // Should the klass from the profile be recorded in the speculative type? 2115 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2116 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2117 const TypeOopPtr* xtype = tklass->as_instance_type(); 2118 assert(xtype->klass_is_exact(), "Should be exact"); 2119 // Any reason to believe n is not null (from this profiling or a previous one)? 2120 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2121 // record the new speculative type's depth 2122 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2123 speculative = speculative->with_inline_depth(jvms()->depth()); 2124 } else if (current_type->would_improve_ptr(maybe_null)) { 2125 // Profiling report that null was never seen so we can change the 2126 // speculative type to non null ptr. 2127 assert(!maybe_null, "nothing to improve"); 2128 if (speculative == NULL) { 2129 speculative = TypePtr::NOTNULL; 2130 } else { 2131 const TypePtr* ptr = TypePtr::NOTNULL; 2132 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2133 } 2134 } 2135 2136 if (speculative != current_type->speculative()) { 2137 // Build a type with a speculative type (what we think we know 2138 // about the type but will need a guard when we use it) 2139 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2140 // We're changing the type, we need a new CheckCast node to carry 2141 // the new type. The new type depends on the control: what 2142 // profiling tells us is only valid from here as far as we can 2143 // tell. 2144 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2145 cast = _gvn.transform(cast); 2146 replace_in_map(n, cast); 2147 n = cast; 2148 } 2149 2150 return n; 2151 } 2152 2153 /** 2154 * Record profiling data from receiver profiling at an invoke with the 2155 * type system so that it can propagate it (speculation) 2156 * 2157 * @param n receiver node 2158 * 2159 * @return node with improved type 2160 */ 2161 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2162 if (!UseTypeSpeculation) { 2163 return n; 2164 } 2165 ciKlass* exact_kls = profile_has_unique_klass(); 2166 bool maybe_null = true; 2167 if (java_bc() == Bytecodes::_checkcast || 2168 java_bc() == Bytecodes::_instanceof || 2169 java_bc() == Bytecodes::_aastore) { 2170 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2171 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2172 } 2173 return record_profile_for_speculation(n, exact_kls, maybe_null); 2174 return n; 2175 } 2176 2177 /** 2178 * Record profiling data from argument profiling at an invoke with the 2179 * type system so that it can propagate it (speculation) 2180 * 2181 * @param dest_method target method for the call 2182 * @param bc what invoke bytecode is this? 2183 */ 2184 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2185 if (!UseTypeSpeculation) { 2186 return; 2187 } 2188 const TypeFunc* tf = TypeFunc::make(dest_method); 2189 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2190 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2191 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2192 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2193 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2194 bool maybe_null = true; 2195 ciKlass* better_type = NULL; 2196 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2197 record_profile_for_speculation(argument(j), better_type, maybe_null); 2198 } 2199 i++; 2200 } 2201 } 2202 } 2203 2204 /** 2205 * Record profiling data from parameter profiling at an invoke with 2206 * the type system so that it can propagate it (speculation) 2207 */ 2208 void GraphKit::record_profiled_parameters_for_speculation() { 2209 if (!UseTypeSpeculation) { 2210 return; 2211 } 2212 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2213 if (_gvn.type(local(i))->isa_oopptr()) { 2214 bool maybe_null = true; 2215 ciKlass* better_type = NULL; 2216 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2217 record_profile_for_speculation(local(i), better_type, maybe_null); 2218 } 2219 j++; 2220 } 2221 } 2222 } 2223 2224 /** 2225 * Record profiling data from return value profiling at an invoke with 2226 * the type system so that it can propagate it (speculation) 2227 */ 2228 void GraphKit::record_profiled_return_for_speculation() { 2229 if (!UseTypeSpeculation) { 2230 return; 2231 } 2232 bool maybe_null = true; 2233 ciKlass* better_type = NULL; 2234 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2235 // If profiling reports a single type for the return value, 2236 // feed it to the type system so it can propagate it as a 2237 // speculative type 2238 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2239 } 2240 } 2241 2242 void GraphKit::round_double_result(ciMethod* dest_method) { 2243 // A non-strict method may return a double value which has an extended 2244 // exponent, but this must not be visible in a caller which is 'strict' 2245 // If a strict caller invokes a non-strict callee, round a double result 2246 2247 BasicType result_type = dest_method->return_type()->basic_type(); 2248 assert( method() != NULL, "must have caller context"); 2249 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2250 // Destination method's return value is on top of stack 2251 // dstore_rounding() does gvn.transform 2252 Node *result = pop_pair(); 2253 result = dstore_rounding(result); 2254 push_pair(result); 2255 } 2256 } 2257 2258 // rounding for strict float precision conformance 2259 Node* GraphKit::precision_rounding(Node* n) { 2260 return UseStrictFP && _method->flags().is_strict() 2261 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2262 ? _gvn.transform( new RoundFloatNode(0, n) ) 2263 : n; 2264 } 2265 2266 // rounding for strict double precision conformance 2267 Node* GraphKit::dprecision_rounding(Node *n) { 2268 return UseStrictFP && _method->flags().is_strict() 2269 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2270 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2271 : n; 2272 } 2273 2274 // rounding for non-strict double stores 2275 Node* GraphKit::dstore_rounding(Node* n) { 2276 return Matcher::strict_fp_requires_explicit_rounding 2277 && UseSSE <= 1 2278 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2279 : n; 2280 } 2281 2282 //============================================================================= 2283 // Generate a fast path/slow path idiom. Graph looks like: 2284 // [foo] indicates that 'foo' is a parameter 2285 // 2286 // [in] NULL 2287 // \ / 2288 // CmpP 2289 // Bool ne 2290 // If 2291 // / \ 2292 // True False-<2> 2293 // / | 2294 // / cast_not_null 2295 // Load | | ^ 2296 // [fast_test] | | 2297 // gvn to opt_test | | 2298 // / \ | <1> 2299 // True False | 2300 // | \\ | 2301 // [slow_call] \[fast_result] 2302 // Ctl Val \ \ 2303 // | \ \ 2304 // Catch <1> \ \ 2305 // / \ ^ \ \ 2306 // Ex No_Ex | \ \ 2307 // | \ \ | \ <2> \ 2308 // ... \ [slow_res] | | \ [null_result] 2309 // \ \--+--+--- | | 2310 // \ | / \ | / 2311 // --------Region Phi 2312 // 2313 //============================================================================= 2314 // Code is structured as a series of driver functions all called 'do_XXX' that 2315 // call a set of helper functions. Helper functions first, then drivers. 2316 2317 //------------------------------null_check_oop--------------------------------- 2318 // Null check oop. Set null-path control into Region in slot 3. 2319 // Make a cast-not-nullness use the other not-null control. Return cast. 2320 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2321 bool never_see_null, 2322 bool safe_for_replace, 2323 bool speculative) { 2324 // Initial NULL check taken path 2325 (*null_control) = top(); 2326 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2327 2328 // Generate uncommon_trap: 2329 if (never_see_null && (*null_control) != top()) { 2330 // If we see an unexpected null at a check-cast we record it and force a 2331 // recompile; the offending check-cast will be compiled to handle NULLs. 2332 // If we see more than one offending BCI, then all checkcasts in the 2333 // method will be compiled to handle NULLs. 2334 PreserveJVMState pjvms(this); 2335 set_control(*null_control); 2336 replace_in_map(value, null()); 2337 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2338 uncommon_trap(reason, 2339 Deoptimization::Action_make_not_entrant); 2340 (*null_control) = top(); // NULL path is dead 2341 } 2342 if ((*null_control) == top() && safe_for_replace) { 2343 replace_in_map(value, cast); 2344 } 2345 2346 // Cast away null-ness on the result 2347 return cast; 2348 } 2349 2350 //------------------------------opt_iff---------------------------------------- 2351 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2352 // Return slow-path control. 2353 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2354 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2355 2356 // Fast path taken; set region slot 2 2357 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2358 region->init_req(2,fast_taken); // Capture fast-control 2359 2360 // Fast path not-taken, i.e. slow path 2361 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2362 return slow_taken; 2363 } 2364 2365 //-----------------------------make_runtime_call------------------------------- 2366 Node* GraphKit::make_runtime_call(int flags, 2367 const TypeFunc* call_type, address call_addr, 2368 const char* call_name, 2369 const TypePtr* adr_type, 2370 // The following parms are all optional. 2371 // The first NULL ends the list. 2372 Node* parm0, Node* parm1, 2373 Node* parm2, Node* parm3, 2374 Node* parm4, Node* parm5, 2375 Node* parm6, Node* parm7) { 2376 // Slow-path call 2377 bool is_leaf = !(flags & RC_NO_LEAF); 2378 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2379 if (call_name == NULL) { 2380 assert(!is_leaf, "must supply name for leaf"); 2381 call_name = OptoRuntime::stub_name(call_addr); 2382 } 2383 CallNode* call; 2384 if (!is_leaf) { 2385 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2386 bci(), adr_type); 2387 } else if (flags & RC_NO_FP) { 2388 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2389 } else { 2390 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2391 } 2392 2393 // The following is similar to set_edges_for_java_call, 2394 // except that the memory effects of the call are restricted to AliasIdxRaw. 2395 2396 // Slow path call has no side-effects, uses few values 2397 bool wide_in = !(flags & RC_NARROW_MEM); 2398 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2399 2400 Node* prev_mem = NULL; 2401 if (wide_in) { 2402 prev_mem = set_predefined_input_for_runtime_call(call); 2403 } else { 2404 assert(!wide_out, "narrow in => narrow out"); 2405 Node* narrow_mem = memory(adr_type); 2406 prev_mem = reset_memory(); 2407 map()->set_memory(narrow_mem); 2408 set_predefined_input_for_runtime_call(call); 2409 } 2410 2411 // Hook each parm in order. Stop looking at the first NULL. 2412 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2413 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2414 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2415 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2416 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2417 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2418 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2419 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2420 /* close each nested if ===> */ } } } } } } } } 2421 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2422 2423 if (!is_leaf) { 2424 // Non-leaves can block and take safepoints: 2425 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2426 } 2427 // Non-leaves can throw exceptions: 2428 if (has_io) { 2429 call->set_req(TypeFunc::I_O, i_o()); 2430 } 2431 2432 if (flags & RC_UNCOMMON) { 2433 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2434 // (An "if" probability corresponds roughly to an unconditional count. 2435 // Sort of.) 2436 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2437 } 2438 2439 Node* c = _gvn.transform(call); 2440 assert(c == call, "cannot disappear"); 2441 2442 if (wide_out) { 2443 // Slow path call has full side-effects. 2444 set_predefined_output_for_runtime_call(call); 2445 } else { 2446 // Slow path call has few side-effects, and/or sets few values. 2447 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2448 } 2449 2450 if (has_io) { 2451 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2452 } 2453 return call; 2454 2455 } 2456 2457 //------------------------------merge_memory----------------------------------- 2458 // Merge memory from one path into the current memory state. 2459 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2460 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2461 Node* old_slice = mms.force_memory(); 2462 Node* new_slice = mms.memory2(); 2463 if (old_slice != new_slice) { 2464 PhiNode* phi; 2465 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2466 if (mms.is_empty()) { 2467 // clone base memory Phi's inputs for this memory slice 2468 assert(old_slice == mms.base_memory(), "sanity"); 2469 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2470 _gvn.set_type(phi, Type::MEMORY); 2471 for (uint i = 1; i < phi->req(); i++) { 2472 phi->init_req(i, old_slice->in(i)); 2473 } 2474 } else { 2475 phi = old_slice->as_Phi(); // Phi was generated already 2476 } 2477 } else { 2478 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2479 _gvn.set_type(phi, Type::MEMORY); 2480 } 2481 phi->set_req(new_path, new_slice); 2482 mms.set_memory(phi); 2483 } 2484 } 2485 } 2486 2487 //------------------------------make_slow_call_ex------------------------------ 2488 // Make the exception handler hookups for the slow call 2489 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2490 if (stopped()) return; 2491 2492 // Make a catch node with just two handlers: fall-through and catch-all 2493 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2494 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2495 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2496 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2497 2498 { PreserveJVMState pjvms(this); 2499 set_control(excp); 2500 set_i_o(i_o); 2501 2502 if (excp != top()) { 2503 if (deoptimize) { 2504 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2505 uncommon_trap(Deoptimization::Reason_unhandled, 2506 Deoptimization::Action_none); 2507 } else { 2508 // Create an exception state also. 2509 // Use an exact type if the caller has specified a specific exception. 2510 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2511 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2512 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2513 } 2514 } 2515 } 2516 2517 // Get the no-exception control from the CatchNode. 2518 set_control(norm); 2519 } 2520 2521 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2522 Node* cmp = NULL; 2523 switch(bt) { 2524 case T_INT: cmp = new CmpINode(in1, in2); break; 2525 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2526 default: fatal(err_msg("unexpected comparison type %s", type2name(bt))); 2527 } 2528 gvn->transform(cmp); 2529 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2530 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2531 gvn->transform(iff); 2532 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2533 return iff; 2534 } 2535 2536 2537 //-------------------------------gen_subtype_check----------------------------- 2538 // Generate a subtyping check. Takes as input the subtype and supertype. 2539 // Returns 2 values: sets the default control() to the true path and returns 2540 // the false path. Only reads invariant memory; sets no (visible) memory. 2541 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2542 // but that's not exposed to the optimizer. This call also doesn't take in an 2543 // Object; if you wish to check an Object you need to load the Object's class 2544 // prior to coming here. 2545 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2546 Compile* C = gvn->C; 2547 2548 if ((*ctrl)->is_top()) { 2549 return C->top(); 2550 } 2551 2552 // Fast check for identical types, perhaps identical constants. 2553 // The types can even be identical non-constants, in cases 2554 // involving Array.newInstance, Object.clone, etc. 2555 if (subklass == superklass) 2556 return C->top(); // false path is dead; no test needed. 2557 2558 if (gvn->type(superklass)->singleton()) { 2559 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2560 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2561 2562 // In the common case of an exact superklass, try to fold up the 2563 // test before generating code. You may ask, why not just generate 2564 // the code and then let it fold up? The answer is that the generated 2565 // code will necessarily include null checks, which do not always 2566 // completely fold away. If they are also needless, then they turn 2567 // into a performance loss. Example: 2568 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2569 // Here, the type of 'fa' is often exact, so the store check 2570 // of fa[1]=x will fold up, without testing the nullness of x. 2571 switch (C->static_subtype_check(superk, subk)) { 2572 case Compile::SSC_always_false: 2573 { 2574 Node* always_fail = *ctrl; 2575 *ctrl = gvn->C->top(); 2576 return always_fail; 2577 } 2578 case Compile::SSC_always_true: 2579 return C->top(); 2580 case Compile::SSC_easy_test: 2581 { 2582 // Just do a direct pointer compare and be done. 2583 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2584 *ctrl = gvn->transform(new IfTrueNode(iff)); 2585 return gvn->transform(new IfFalseNode(iff)); 2586 } 2587 case Compile::SSC_full_test: 2588 break; 2589 default: 2590 ShouldNotReachHere(); 2591 } 2592 } 2593 2594 // %%% Possible further optimization: Even if the superklass is not exact, 2595 // if the subklass is the unique subtype of the superklass, the check 2596 // will always succeed. We could leave a dependency behind to ensure this. 2597 2598 // First load the super-klass's check-offset 2599 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2600 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2601 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2602 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2603 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2604 2605 // Load from the sub-klass's super-class display list, or a 1-word cache of 2606 // the secondary superclass list, or a failing value with a sentinel offset 2607 // if the super-klass is an interface or exceptionally deep in the Java 2608 // hierarchy and we have to scan the secondary superclass list the hard way. 2609 // Worst-case type is a little odd: NULL is allowed as a result (usually 2610 // klass loads can never produce a NULL). 2611 Node *chk_off_X = chk_off; 2612 #ifdef _LP64 2613 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2614 #endif 2615 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2616 // For some types like interfaces the following loadKlass is from a 1-word 2617 // cache which is mutable so can't use immutable memory. Other 2618 // types load from the super-class display table which is immutable. 2619 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2620 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2621 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2622 2623 // Compile speed common case: ARE a subtype and we canNOT fail 2624 if( superklass == nkls ) 2625 return C->top(); // false path is dead; no test needed. 2626 2627 // See if we get an immediate positive hit. Happens roughly 83% of the 2628 // time. Test to see if the value loaded just previously from the subklass 2629 // is exactly the superklass. 2630 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2631 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2632 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2633 2634 // Compile speed common case: Check for being deterministic right now. If 2635 // chk_off is a constant and not equal to cacheoff then we are NOT a 2636 // subklass. In this case we need exactly the 1 test above and we can 2637 // return those results immediately. 2638 if (!might_be_cache) { 2639 Node* not_subtype_ctrl = *ctrl; 2640 *ctrl = iftrue1; // We need exactly the 1 test above 2641 return not_subtype_ctrl; 2642 } 2643 2644 // Gather the various success & failures here 2645 RegionNode *r_ok_subtype = new RegionNode(4); 2646 gvn->record_for_igvn(r_ok_subtype); 2647 RegionNode *r_not_subtype = new RegionNode(3); 2648 gvn->record_for_igvn(r_not_subtype); 2649 2650 r_ok_subtype->init_req(1, iftrue1); 2651 2652 // Check for immediate negative hit. Happens roughly 11% of the time (which 2653 // is roughly 63% of the remaining cases). Test to see if the loaded 2654 // check-offset points into the subklass display list or the 1-element 2655 // cache. If it points to the display (and NOT the cache) and the display 2656 // missed then it's not a subtype. 2657 Node *cacheoff = gvn->intcon(cacheoff_con); 2658 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2659 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2660 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2661 2662 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2663 // No performance impact (too rare) but allows sharing of secondary arrays 2664 // which has some footprint reduction. 2665 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2666 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2667 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2668 2669 // -- Roads not taken here: -- 2670 // We could also have chosen to perform the self-check at the beginning 2671 // of this code sequence, as the assembler does. This would not pay off 2672 // the same way, since the optimizer, unlike the assembler, can perform 2673 // static type analysis to fold away many successful self-checks. 2674 // Non-foldable self checks work better here in second position, because 2675 // the initial primary superclass check subsumes a self-check for most 2676 // types. An exception would be a secondary type like array-of-interface, 2677 // which does not appear in its own primary supertype display. 2678 // Finally, we could have chosen to move the self-check into the 2679 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2680 // dependent manner. But it is worthwhile to have the check here, 2681 // where it can be perhaps be optimized. The cost in code space is 2682 // small (register compare, branch). 2683 2684 // Now do a linear scan of the secondary super-klass array. Again, no real 2685 // performance impact (too rare) but it's gotta be done. 2686 // Since the code is rarely used, there is no penalty for moving it 2687 // out of line, and it can only improve I-cache density. 2688 // The decision to inline or out-of-line this final check is platform 2689 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2690 Node* psc = gvn->transform( 2691 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2692 2693 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2694 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2695 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2696 2697 // Return false path; set default control to true path. 2698 *ctrl = gvn->transform(r_ok_subtype); 2699 return gvn->transform(r_not_subtype); 2700 } 2701 2702 // Profile-driven exact type check: 2703 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2704 float prob, 2705 Node* *casted_receiver) { 2706 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2707 Node* recv_klass = load_object_klass(receiver); 2708 Node* want_klass = makecon(tklass); 2709 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2710 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2711 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2712 set_control( _gvn.transform( new IfTrueNode (iff) )); 2713 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2714 2715 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2716 assert(recv_xtype->klass_is_exact(), ""); 2717 2718 // Subsume downstream occurrences of receiver with a cast to 2719 // recv_xtype, since now we know what the type will be. 2720 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2721 (*casted_receiver) = _gvn.transform(cast); 2722 // (User must make the replace_in_map call.) 2723 2724 return fail; 2725 } 2726 2727 2728 //------------------------------seems_never_null------------------------------- 2729 // Use null_seen information if it is available from the profile. 2730 // If we see an unexpected null at a type check we record it and force a 2731 // recompile; the offending check will be recompiled to handle NULLs. 2732 // If we see several offending BCIs, then all checks in the 2733 // method will be recompiled. 2734 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2735 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2736 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2737 if (UncommonNullCast // Cutout for this technique 2738 && obj != null() // And not the -Xcomp stupid case? 2739 && !too_many_traps(reason) 2740 ) { 2741 if (speculating) { 2742 return true; 2743 } 2744 if (data == NULL) 2745 // Edge case: no mature data. Be optimistic here. 2746 return true; 2747 // If the profile has not seen a null, assume it won't happen. 2748 assert(java_bc() == Bytecodes::_checkcast || 2749 java_bc() == Bytecodes::_instanceof || 2750 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2751 return !data->as_BitData()->null_seen(); 2752 } 2753 speculating = false; 2754 return false; 2755 } 2756 2757 //------------------------maybe_cast_profiled_receiver------------------------- 2758 // If the profile has seen exactly one type, narrow to exactly that type. 2759 // Subsequent type checks will always fold up. 2760 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2761 ciKlass* require_klass, 2762 ciKlass* spec_klass, 2763 bool safe_for_replace) { 2764 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2765 2766 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2767 2768 // Make sure we haven't already deoptimized from this tactic. 2769 if (too_many_traps(reason) || too_many_recompiles(reason)) 2770 return NULL; 2771 2772 // (No, this isn't a call, but it's enough like a virtual call 2773 // to use the same ciMethod accessor to get the profile info...) 2774 // If we have a speculative type use it instead of profiling (which 2775 // may not help us) 2776 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2777 if (exact_kls != NULL) {// no cast failures here 2778 if (require_klass == NULL || 2779 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2780 // If we narrow the type to match what the type profile sees or 2781 // the speculative type, we can then remove the rest of the 2782 // cast. 2783 // This is a win, even if the exact_kls is very specific, 2784 // because downstream operations, such as method calls, 2785 // will often benefit from the sharper type. 2786 Node* exact_obj = not_null_obj; // will get updated in place... 2787 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2788 &exact_obj); 2789 { PreserveJVMState pjvms(this); 2790 set_control(slow_ctl); 2791 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2792 } 2793 if (safe_for_replace) { 2794 replace_in_map(not_null_obj, exact_obj); 2795 } 2796 return exact_obj; 2797 } 2798 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2799 } 2800 2801 return NULL; 2802 } 2803 2804 /** 2805 * Cast obj to type and emit guard unless we had too many traps here 2806 * already 2807 * 2808 * @param obj node being casted 2809 * @param type type to cast the node to 2810 * @param not_null true if we know node cannot be null 2811 */ 2812 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2813 ciKlass* type, 2814 bool not_null) { 2815 if (stopped()) { 2816 return obj; 2817 } 2818 2819 // type == NULL if profiling tells us this object is always null 2820 if (type != NULL) { 2821 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2822 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2823 2824 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2825 !too_many_traps(class_reason) && 2826 !too_many_recompiles(class_reason)) { 2827 Node* not_null_obj = NULL; 2828 // not_null is true if we know the object is not null and 2829 // there's no need for a null check 2830 if (!not_null) { 2831 Node* null_ctl = top(); 2832 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2833 assert(null_ctl->is_top(), "no null control here"); 2834 } else { 2835 not_null_obj = obj; 2836 } 2837 2838 Node* exact_obj = not_null_obj; 2839 ciKlass* exact_kls = type; 2840 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2841 &exact_obj); 2842 { 2843 PreserveJVMState pjvms(this); 2844 set_control(slow_ctl); 2845 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2846 } 2847 replace_in_map(not_null_obj, exact_obj); 2848 obj = exact_obj; 2849 } 2850 } else { 2851 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2852 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2853 Node* exact_obj = null_assert(obj); 2854 replace_in_map(obj, exact_obj); 2855 obj = exact_obj; 2856 } 2857 } 2858 return obj; 2859 } 2860 2861 //-------------------------------gen_instanceof-------------------------------- 2862 // Generate an instance-of idiom. Used by both the instance-of bytecode 2863 // and the reflective instance-of call. 2864 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2865 kill_dead_locals(); // Benefit all the uncommon traps 2866 assert( !stopped(), "dead parse path should be checked in callers" ); 2867 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2868 "must check for not-null not-dead klass in callers"); 2869 2870 // Make the merge point 2871 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2872 RegionNode* region = new RegionNode(PATH_LIMIT); 2873 Node* phi = new PhiNode(region, TypeInt::BOOL); 2874 C->set_has_split_ifs(true); // Has chance for split-if optimization 2875 2876 ciProfileData* data = NULL; 2877 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2878 data = method()->method_data()->bci_to_data(bci()); 2879 } 2880 bool speculative_not_null = false; 2881 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2882 && seems_never_null(obj, data, speculative_not_null)); 2883 2884 // Null check; get casted pointer; set region slot 3 2885 Node* null_ctl = top(); 2886 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2887 2888 // If not_null_obj is dead, only null-path is taken 2889 if (stopped()) { // Doing instance-of on a NULL? 2890 set_control(null_ctl); 2891 return intcon(0); 2892 } 2893 region->init_req(_null_path, null_ctl); 2894 phi ->init_req(_null_path, intcon(0)); // Set null path value 2895 if (null_ctl == top()) { 2896 // Do this eagerly, so that pattern matches like is_diamond_phi 2897 // will work even during parsing. 2898 assert(_null_path == PATH_LIMIT-1, "delete last"); 2899 region->del_req(_null_path); 2900 phi ->del_req(_null_path); 2901 } 2902 2903 // Do we know the type check always succeed? 2904 bool known_statically = false; 2905 if (_gvn.type(superklass)->singleton()) { 2906 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2907 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2908 if (subk != NULL && subk->is_loaded()) { 2909 int static_res = C->static_subtype_check(superk, subk); 2910 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2911 } 2912 } 2913 2914 if (known_statically && UseTypeSpeculation) { 2915 // If we know the type check always succeeds then we don't use the 2916 // profiling data at this bytecode. Don't lose it, feed it to the 2917 // type system as a speculative type. 2918 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2919 } else { 2920 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2921 // We may not have profiling here or it may not help us. If we 2922 // have a speculative type use it to perform an exact cast. 2923 ciKlass* spec_obj_type = obj_type->speculative_type(); 2924 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2925 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2926 if (stopped()) { // Profile disagrees with this path. 2927 set_control(null_ctl); // Null is the only remaining possibility. 2928 return intcon(0); 2929 } 2930 if (cast_obj != NULL) { 2931 not_null_obj = cast_obj; 2932 } 2933 } 2934 } 2935 2936 if (ShenandoahVerifyReadsToFromSpace) { 2937 not_null_obj = shenandoah_read_barrier(not_null_obj); 2938 } 2939 2940 // Load the object's klass 2941 Node* obj_klass = load_object_klass(not_null_obj); 2942 2943 // Generate the subtype check 2944 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2945 2946 // Plug in the success path to the general merge in slot 1. 2947 region->init_req(_obj_path, control()); 2948 phi ->init_req(_obj_path, intcon(1)); 2949 2950 // Plug in the failing path to the general merge in slot 2. 2951 region->init_req(_fail_path, not_subtype_ctrl); 2952 phi ->init_req(_fail_path, intcon(0)); 2953 2954 // Return final merged results 2955 set_control( _gvn.transform(region) ); 2956 record_for_igvn(region); 2957 return _gvn.transform(phi); 2958 } 2959 2960 //-------------------------------gen_checkcast--------------------------------- 2961 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2962 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2963 // uncommon-trap paths work. Adjust stack after this call. 2964 // If failure_control is supplied and not null, it is filled in with 2965 // the control edge for the cast failure. Otherwise, an appropriate 2966 // uncommon trap or exception is thrown. 2967 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2968 Node* *failure_control) { 2969 kill_dead_locals(); // Benefit all the uncommon traps 2970 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2971 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2972 2973 // Fast cutout: Check the case that the cast is vacuously true. 2974 // This detects the common cases where the test will short-circuit 2975 // away completely. We do this before we perform the null check, 2976 // because if the test is going to turn into zero code, we don't 2977 // want a residual null check left around. (Causes a slowdown, 2978 // for example, in some objArray manipulations, such as a[i]=a[j].) 2979 if (tk->singleton()) { 2980 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2981 if (objtp != NULL && objtp->klass() != NULL) { 2982 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2983 case Compile::SSC_always_true: 2984 // If we know the type check always succeed then we don't use 2985 // the profiling data at this bytecode. Don't lose it, feed it 2986 // to the type system as a speculative type. 2987 return record_profiled_receiver_for_speculation(obj); 2988 case Compile::SSC_always_false: 2989 // It needs a null check because a null will *pass* the cast check. 2990 // A non-null value will always produce an exception. 2991 return null_assert(obj); 2992 } 2993 } 2994 } 2995 2996 ciProfileData* data = NULL; 2997 bool safe_for_replace = false; 2998 if (failure_control == NULL) { // use MDO in regular case only 2999 assert(java_bc() == Bytecodes::_aastore || 3000 java_bc() == Bytecodes::_checkcast, 3001 "interpreter profiles type checks only for these BCs"); 3002 data = method()->method_data()->bci_to_data(bci()); 3003 safe_for_replace = true; 3004 } 3005 3006 // Make the merge point 3007 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3008 RegionNode* region = new RegionNode(PATH_LIMIT); 3009 Node* phi = new PhiNode(region, toop); 3010 C->set_has_split_ifs(true); // Has chance for split-if optimization 3011 3012 // Use null-cast information if it is available 3013 bool speculative_not_null = false; 3014 bool never_see_null = ((failure_control == NULL) // regular case only 3015 && seems_never_null(obj, data, speculative_not_null)); 3016 3017 // Null check; get casted pointer; set region slot 3 3018 Node* null_ctl = top(); 3019 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3020 3021 if (ShenandoahVerifyReadsToFromSpace) { 3022 not_null_obj = shenandoah_read_barrier(not_null_obj); 3023 } 3024 3025 // If not_null_obj is dead, only null-path is taken 3026 if (stopped()) { // Doing instance-of on a NULL? 3027 set_control(null_ctl); 3028 return null(); 3029 } 3030 region->init_req(_null_path, null_ctl); 3031 phi ->init_req(_null_path, null()); // Set null path value 3032 if (null_ctl == top()) { 3033 // Do this eagerly, so that pattern matches like is_diamond_phi 3034 // will work even during parsing. 3035 assert(_null_path == PATH_LIMIT-1, "delete last"); 3036 region->del_req(_null_path); 3037 phi ->del_req(_null_path); 3038 } 3039 3040 Node* cast_obj = NULL; 3041 if (tk->klass_is_exact()) { 3042 // The following optimization tries to statically cast the speculative type of the object 3043 // (for example obtained during profiling) to the type of the superklass and then do a 3044 // dynamic check that the type of the object is what we expect. To work correctly 3045 // for checkcast and aastore the type of superklass should be exact. 3046 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3047 // We may not have profiling here or it may not help us. If we have 3048 // a speculative type use it to perform an exact cast. 3049 ciKlass* spec_obj_type = obj_type->speculative_type(); 3050 if (spec_obj_type != NULL || data != NULL) { 3051 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3052 if (cast_obj != NULL) { 3053 if (failure_control != NULL) // failure is now impossible 3054 (*failure_control) = top(); 3055 // adjust the type of the phi to the exact klass: 3056 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3057 } 3058 } 3059 } 3060 3061 if (cast_obj == NULL) { 3062 // Load the object's klass 3063 Node* obj_klass = load_object_klass(not_null_obj); 3064 3065 // Generate the subtype check 3066 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3067 3068 // Plug in success path into the merge 3069 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3070 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3071 if (failure_control == NULL) { 3072 if (not_subtype_ctrl != top()) { // If failure is possible 3073 PreserveJVMState pjvms(this); 3074 set_control(not_subtype_ctrl); 3075 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3076 } 3077 } else { 3078 (*failure_control) = not_subtype_ctrl; 3079 } 3080 } 3081 3082 region->init_req(_obj_path, control()); 3083 phi ->init_req(_obj_path, cast_obj); 3084 3085 // A merge of NULL or Casted-NotNull obj 3086 Node* res = _gvn.transform(phi); 3087 3088 // Note I do NOT always 'replace_in_map(obj,result)' here. 3089 // if( tk->klass()->can_be_primary_super() ) 3090 // This means that if I successfully store an Object into an array-of-String 3091 // I 'forget' that the Object is really now known to be a String. I have to 3092 // do this because we don't have true union types for interfaces - if I store 3093 // a Baz into an array-of-Interface and then tell the optimizer it's an 3094 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3095 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3096 // replace_in_map( obj, res ); 3097 3098 // Return final merged results 3099 set_control( _gvn.transform(region) ); 3100 record_for_igvn(region); 3101 return res; 3102 } 3103 3104 //------------------------------next_monitor----------------------------------- 3105 // What number should be given to the next monitor? 3106 int GraphKit::next_monitor() { 3107 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3108 int next = current + C->sync_stack_slots(); 3109 // Keep the toplevel high water mark current: 3110 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3111 return current; 3112 } 3113 3114 //------------------------------insert_mem_bar--------------------------------- 3115 // Memory barrier to avoid floating things around 3116 // The membar serves as a pinch point between both control and all memory slices. 3117 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3118 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3119 mb->init_req(TypeFunc::Control, control()); 3120 mb->init_req(TypeFunc::Memory, reset_memory()); 3121 Node* membar = _gvn.transform(mb); 3122 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3123 set_all_memory_call(membar); 3124 return membar; 3125 } 3126 3127 //-------------------------insert_mem_bar_volatile---------------------------- 3128 // Memory barrier to avoid floating things around 3129 // The membar serves as a pinch point between both control and memory(alias_idx). 3130 // If you want to make a pinch point on all memory slices, do not use this 3131 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3132 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3133 // When Parse::do_put_xxx updates a volatile field, it appends a series 3134 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3135 // The first membar is on the same memory slice as the field store opcode. 3136 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3137 // All the other membars (for other volatile slices, including AliasIdxBot, 3138 // which stands for all unknown volatile slices) are control-dependent 3139 // on the first membar. This prevents later volatile loads or stores 3140 // from sliding up past the just-emitted store. 3141 3142 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3143 mb->set_req(TypeFunc::Control,control()); 3144 if (alias_idx == Compile::AliasIdxBot) { 3145 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3146 } else { 3147 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3148 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3149 } 3150 Node* membar = _gvn.transform(mb); 3151 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3152 if (alias_idx == Compile::AliasIdxBot) { 3153 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3154 } else { 3155 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3156 } 3157 return membar; 3158 } 3159 3160 //------------------------------shared_lock------------------------------------ 3161 // Emit locking code. 3162 FastLockNode* GraphKit::shared_lock(Node* obj) { 3163 // bci is either a monitorenter bc or InvocationEntryBci 3164 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3165 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3166 3167 if( !GenerateSynchronizationCode ) 3168 return NULL; // Not locking things? 3169 if (stopped()) // Dead monitor? 3170 return NULL; 3171 3172 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3173 3174 obj = shenandoah_write_barrier(obj); 3175 3176 // Box the stack location 3177 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3178 Node* mem = reset_memory(); 3179 3180 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3181 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3182 // Create the counters for this fast lock. 3183 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3184 } 3185 3186 // Create the rtm counters for this fast lock if needed. 3187 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3188 3189 // Add monitor to debug info for the slow path. If we block inside the 3190 // slow path and de-opt, we need the monitor hanging around 3191 map()->push_monitor( flock ); 3192 3193 const TypeFunc *tf = LockNode::lock_type(); 3194 LockNode *lock = new LockNode(C, tf); 3195 3196 lock->init_req( TypeFunc::Control, control() ); 3197 lock->init_req( TypeFunc::Memory , mem ); 3198 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3199 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3200 lock->init_req( TypeFunc::ReturnAdr, top() ); 3201 3202 lock->init_req(TypeFunc::Parms + 0, obj); 3203 lock->init_req(TypeFunc::Parms + 1, box); 3204 lock->init_req(TypeFunc::Parms + 2, flock); 3205 add_safepoint_edges(lock); 3206 3207 lock = _gvn.transform( lock )->as_Lock(); 3208 3209 // lock has no side-effects, sets few values 3210 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3211 3212 insert_mem_bar(Op_MemBarAcquireLock); 3213 3214 // Add this to the worklist so that the lock can be eliminated 3215 record_for_igvn(lock); 3216 3217 #ifndef PRODUCT 3218 if (PrintLockStatistics) { 3219 // Update the counter for this lock. Don't bother using an atomic 3220 // operation since we don't require absolute accuracy. 3221 lock->create_lock_counter(map()->jvms()); 3222 increment_counter(lock->counter()->addr()); 3223 } 3224 #endif 3225 3226 return flock; 3227 } 3228 3229 3230 //------------------------------shared_unlock---------------------------------- 3231 // Emit unlocking code. 3232 void GraphKit::shared_unlock(Node* box, Node* obj) { 3233 // bci is either a monitorenter bc or InvocationEntryBci 3234 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3235 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3236 3237 if( !GenerateSynchronizationCode ) 3238 return; 3239 if (stopped()) { // Dead monitor? 3240 map()->pop_monitor(); // Kill monitor from debug info 3241 return; 3242 } 3243 3244 // Memory barrier to avoid floating things down past the locked region 3245 insert_mem_bar(Op_MemBarReleaseLock); 3246 3247 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3248 UnlockNode *unlock = new UnlockNode(C, tf); 3249 #ifdef ASSERT 3250 unlock->set_dbg_jvms(sync_jvms()); 3251 #endif 3252 uint raw_idx = Compile::AliasIdxRaw; 3253 unlock->init_req( TypeFunc::Control, control() ); 3254 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3255 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3256 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3257 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3258 3259 unlock->init_req(TypeFunc::Parms + 0, obj); 3260 unlock->init_req(TypeFunc::Parms + 1, box); 3261 unlock = _gvn.transform(unlock)->as_Unlock(); 3262 3263 Node* mem = reset_memory(); 3264 3265 // unlock has no side-effects, sets few values 3266 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3267 3268 // Kill monitor from debug info 3269 map()->pop_monitor( ); 3270 } 3271 3272 //-------------------------------get_layout_helper----------------------------- 3273 // If the given klass is a constant or known to be an array, 3274 // fetch the constant layout helper value into constant_value 3275 // and return (Node*)NULL. Otherwise, load the non-constant 3276 // layout helper value, and return the node which represents it. 3277 // This two-faced routine is useful because allocation sites 3278 // almost always feature constant types. 3279 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3280 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3281 if (!StressReflectiveCode && inst_klass != NULL) { 3282 ciKlass* klass = inst_klass->klass(); 3283 bool xklass = inst_klass->klass_is_exact(); 3284 if (xklass || klass->is_array_klass()) { 3285 jint lhelper = klass->layout_helper(); 3286 if (lhelper != Klass::_lh_neutral_value) { 3287 constant_value = lhelper; 3288 return (Node*) NULL; 3289 } 3290 } 3291 } 3292 constant_value = Klass::_lh_neutral_value; // put in a known value 3293 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3294 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3295 } 3296 3297 // We just put in an allocate/initialize with a big raw-memory effect. 3298 // Hook selected additional alias categories on the initialization. 3299 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3300 MergeMemNode* init_in_merge, 3301 Node* init_out_raw) { 3302 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3303 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3304 3305 Node* prevmem = kit.memory(alias_idx); 3306 init_in_merge->set_memory_at(alias_idx, prevmem); 3307 kit.set_memory(init_out_raw, alias_idx); 3308 } 3309 3310 //---------------------------set_output_for_allocation------------------------- 3311 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3312 const TypeOopPtr* oop_type, 3313 bool deoptimize_on_exception) { 3314 int rawidx = Compile::AliasIdxRaw; 3315 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3316 add_safepoint_edges(alloc); 3317 Node* allocx = _gvn.transform(alloc); 3318 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3319 // create memory projection for i_o 3320 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3321 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3322 3323 // create a memory projection as for the normal control path 3324 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3325 set_memory(malloc, rawidx); 3326 3327 // a normal slow-call doesn't change i_o, but an allocation does 3328 // we create a separate i_o projection for the normal control path 3329 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3330 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3331 3332 // put in an initialization barrier 3333 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3334 rawoop)->as_Initialize(); 3335 assert(alloc->initialization() == init, "2-way macro link must work"); 3336 assert(init ->allocation() == alloc, "2-way macro link must work"); 3337 { 3338 // Extract memory strands which may participate in the new object's 3339 // initialization, and source them from the new InitializeNode. 3340 // This will allow us to observe initializations when they occur, 3341 // and link them properly (as a group) to the InitializeNode. 3342 assert(init->in(InitializeNode::Memory) == malloc, ""); 3343 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3344 init->set_req(InitializeNode::Memory, minit_in); 3345 record_for_igvn(minit_in); // fold it up later, if possible 3346 Node* minit_out = memory(rawidx); 3347 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3348 if (oop_type->isa_aryptr()) { 3349 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3350 int elemidx = C->get_alias_index(telemref); 3351 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3352 } else if (oop_type->isa_instptr()) { 3353 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3354 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3355 ciField* field = ik->nonstatic_field_at(i); 3356 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3357 continue; // do not bother to track really large numbers of fields 3358 // Find (or create) the alias category for this field: 3359 int fieldidx = C->alias_type(field)->index(); 3360 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3361 } 3362 } 3363 } 3364 3365 // Cast raw oop to the real thing... 3366 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3367 javaoop = _gvn.transform(javaoop); 3368 C->set_recent_alloc(control(), javaoop); 3369 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3370 3371 #ifdef ASSERT 3372 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3373 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3374 "Ideal_allocation works"); 3375 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3376 "Ideal_allocation works"); 3377 if (alloc->is_AllocateArray()) { 3378 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3379 "Ideal_allocation works"); 3380 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3381 "Ideal_allocation works"); 3382 } else { 3383 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3384 } 3385 } 3386 #endif //ASSERT 3387 3388 return javaoop; 3389 } 3390 3391 //---------------------------new_instance-------------------------------------- 3392 // This routine takes a klass_node which may be constant (for a static type) 3393 // or may be non-constant (for reflective code). It will work equally well 3394 // for either, and the graph will fold nicely if the optimizer later reduces 3395 // the type to a constant. 3396 // The optional arguments are for specialized use by intrinsics: 3397 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3398 // - If 'return_size_val', report the the total object size to the caller. 3399 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3400 Node* GraphKit::new_instance(Node* klass_node, 3401 Node* extra_slow_test, 3402 Node* *return_size_val, 3403 bool deoptimize_on_exception) { 3404 // Compute size in doublewords 3405 // The size is always an integral number of doublewords, represented 3406 // as a positive bytewise size stored in the klass's layout_helper. 3407 // The layout_helper also encodes (in a low bit) the need for a slow path. 3408 jint layout_con = Klass::_lh_neutral_value; 3409 Node* layout_val = get_layout_helper(klass_node, layout_con); 3410 int layout_is_con = (layout_val == NULL); 3411 3412 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3413 // Generate the initial go-slow test. It's either ALWAYS (return a 3414 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3415 // case) a computed value derived from the layout_helper. 3416 Node* initial_slow_test = NULL; 3417 if (layout_is_con) { 3418 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3419 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3420 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3421 3422 } else { // reflective case 3423 // This reflective path is used by Unsafe.allocateInstance. 3424 // (It may be stress-tested by specifying StressReflectiveCode.) 3425 // Basically, we want to get into the VM is there's an illegal argument. 3426 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3427 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3428 if (extra_slow_test != intcon(0)) { 3429 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3430 } 3431 // (Macro-expander will further convert this to a Bool, if necessary.) 3432 } 3433 3434 // Find the size in bytes. This is easy; it's the layout_helper. 3435 // The size value must be valid even if the slow path is taken. 3436 Node* size = NULL; 3437 if (layout_is_con) { 3438 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3439 } else { // reflective case 3440 // This reflective path is used by clone and Unsafe.allocateInstance. 3441 size = ConvI2X(layout_val); 3442 3443 // Clear the low bits to extract layout_helper_size_in_bytes: 3444 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3445 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3446 size = _gvn.transform( new AndXNode(size, mask) ); 3447 } 3448 if (return_size_val != NULL) { 3449 (*return_size_val) = size; 3450 } 3451 3452 // This is a precise notnull oop of the klass. 3453 // (Actually, it need not be precise if this is a reflective allocation.) 3454 // It's what we cast the result to. 3455 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3456 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3457 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3458 3459 // Now generate allocation code 3460 3461 // The entire memory state is needed for slow path of the allocation 3462 // since GC and deoptimization can happened. 3463 Node *mem = reset_memory(); 3464 set_all_memory(mem); // Create new memory state 3465 3466 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3467 control(), mem, i_o(), 3468 size, klass_node, 3469 initial_slow_test); 3470 3471 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3472 } 3473 3474 //-------------------------------new_array------------------------------------- 3475 // helper for both newarray and anewarray 3476 // The 'length' parameter is (obviously) the length of the array. 3477 // See comments on new_instance for the meaning of the other arguments. 3478 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3479 Node* length, // number of array elements 3480 int nargs, // number of arguments to push back for uncommon trap 3481 Node* *return_size_val, 3482 bool deoptimize_on_exception) { 3483 jint layout_con = Klass::_lh_neutral_value; 3484 Node* layout_val = get_layout_helper(klass_node, layout_con); 3485 int layout_is_con = (layout_val == NULL); 3486 3487 if (!layout_is_con && !StressReflectiveCode && 3488 !too_many_traps(Deoptimization::Reason_class_check)) { 3489 // This is a reflective array creation site. 3490 // Optimistically assume that it is a subtype of Object[], 3491 // so that we can fold up all the address arithmetic. 3492 layout_con = Klass::array_layout_helper(T_OBJECT); 3493 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3494 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3495 { BuildCutout unless(this, bol_lh, PROB_MAX); 3496 inc_sp(nargs); 3497 uncommon_trap(Deoptimization::Reason_class_check, 3498 Deoptimization::Action_maybe_recompile); 3499 } 3500 layout_val = NULL; 3501 layout_is_con = true; 3502 } 3503 3504 // Generate the initial go-slow test. Make sure we do not overflow 3505 // if length is huge (near 2Gig) or negative! We do not need 3506 // exact double-words here, just a close approximation of needed 3507 // double-words. We can't add any offset or rounding bits, lest we 3508 // take a size -1 of bytes and make it positive. Use an unsigned 3509 // compare, so negative sizes look hugely positive. 3510 int fast_size_limit = FastAllocateSizeLimit; 3511 if (layout_is_con) { 3512 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3513 // Increase the size limit if we have exact knowledge of array type. 3514 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3515 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3516 } 3517 3518 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3519 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3520 if (initial_slow_test->is_Bool()) { 3521 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3522 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3523 } 3524 3525 // --- Size Computation --- 3526 // array_size = round_to_heap(array_header + (length << elem_shift)); 3527 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3528 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3529 // The rounding mask is strength-reduced, if possible. 3530 int round_mask = MinObjAlignmentInBytes - 1; 3531 Node* header_size = NULL; 3532 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3533 // (T_BYTE has the weakest alignment and size restrictions...) 3534 if (layout_is_con) { 3535 int hsize = Klass::layout_helper_header_size(layout_con); 3536 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3537 BasicType etype = Klass::layout_helper_element_type(layout_con); 3538 if ((round_mask & ~right_n_bits(eshift)) == 0) 3539 round_mask = 0; // strength-reduce it if it goes away completely 3540 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3541 assert(header_size_min <= hsize, "generic minimum is smallest"); 3542 header_size_min = hsize; 3543 header_size = intcon(hsize + round_mask); 3544 } else { 3545 Node* hss = intcon(Klass::_lh_header_size_shift); 3546 Node* hsm = intcon(Klass::_lh_header_size_mask); 3547 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3548 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3549 Node* mask = intcon(round_mask); 3550 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3551 } 3552 3553 Node* elem_shift = NULL; 3554 if (layout_is_con) { 3555 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3556 if (eshift != 0) 3557 elem_shift = intcon(eshift); 3558 } else { 3559 // There is no need to mask or shift this value. 3560 // The semantics of LShiftINode include an implicit mask to 0x1F. 3561 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3562 elem_shift = layout_val; 3563 } 3564 3565 // Transition to native address size for all offset calculations: 3566 Node* lengthx = ConvI2X(length); 3567 Node* headerx = ConvI2X(header_size); 3568 #ifdef _LP64 3569 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3570 if (tllen != NULL && tllen->_lo < 0) { 3571 // Add a manual constraint to a positive range. Cf. array_element_address. 3572 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3573 if (size_max > tllen->_hi) size_max = tllen->_hi; 3574 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3575 lengthx = _gvn.transform( new ConvI2LNode(length, tlcon)); 3576 } 3577 } 3578 #endif 3579 3580 // Combine header size (plus rounding) and body size. Then round down. 3581 // This computation cannot overflow, because it is used only in two 3582 // places, one where the length is sharply limited, and the other 3583 // after a successful allocation. 3584 Node* abody = lengthx; 3585 if (elem_shift != NULL) 3586 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3587 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3588 if (round_mask != 0) { 3589 Node* mask = MakeConX(~round_mask); 3590 size = _gvn.transform( new AndXNode(size, mask) ); 3591 } 3592 // else if round_mask == 0, the size computation is self-rounding 3593 3594 if (return_size_val != NULL) { 3595 // This is the size 3596 (*return_size_val) = size; 3597 } 3598 3599 // Now generate allocation code 3600 3601 // The entire memory state is needed for slow path of the allocation 3602 // since GC and deoptimization can happened. 3603 Node *mem = reset_memory(); 3604 set_all_memory(mem); // Create new memory state 3605 3606 // Create the AllocateArrayNode and its result projections 3607 AllocateArrayNode* alloc 3608 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3609 control(), mem, i_o(), 3610 size, klass_node, 3611 initial_slow_test, 3612 length); 3613 3614 // Cast to correct type. Note that the klass_node may be constant or not, 3615 // and in the latter case the actual array type will be inexact also. 3616 // (This happens via a non-constant argument to inline_native_newArray.) 3617 // In any case, the value of klass_node provides the desired array type. 3618 const TypeInt* length_type = _gvn.find_int_type(length); 3619 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3620 if (ary_type->isa_aryptr() && length_type != NULL) { 3621 // Try to get a better type than POS for the size 3622 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3623 } 3624 3625 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3626 3627 // Cast length on remaining path to be as narrow as possible 3628 if (map()->find_edge(length) >= 0) { 3629 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3630 if (ccast != length) { 3631 _gvn.set_type_bottom(ccast); 3632 record_for_igvn(ccast); 3633 replace_in_map(length, ccast); 3634 } 3635 } 3636 3637 return javaoop; 3638 } 3639 3640 // The following "Ideal_foo" functions are placed here because they recognize 3641 // the graph shapes created by the functions immediately above. 3642 3643 //---------------------------Ideal_allocation---------------------------------- 3644 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3645 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3646 if (ptr == NULL) { // reduce dumb test in callers 3647 return NULL; 3648 } 3649 3650 // Attempt to see through Shenandoah barriers. 3651 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr); 3652 3653 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3654 ptr = ptr->in(1); 3655 if (ptr == NULL) return NULL; 3656 } 3657 // Return NULL for allocations with several casts: 3658 // j.l.reflect.Array.newInstance(jobject, jint) 3659 // Object.clone() 3660 // to keep more precise type from last cast. 3661 if (ptr->is_Proj()) { 3662 Node* allo = ptr->in(0); 3663 if (allo != NULL && allo->is_Allocate()) { 3664 return allo->as_Allocate(); 3665 } 3666 } 3667 // Report failure to match. 3668 return NULL; 3669 } 3670 3671 // Fancy version which also strips off an offset (and reports it to caller). 3672 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3673 intptr_t& offset) { 3674 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3675 if (base == NULL) return NULL; 3676 return Ideal_allocation(base, phase); 3677 } 3678 3679 // Trace Initialize <- Proj[Parm] <- Allocate 3680 AllocateNode* InitializeNode::allocation() { 3681 Node* rawoop = in(InitializeNode::RawAddress); 3682 if (rawoop->is_Proj()) { 3683 Node* alloc = rawoop->in(0); 3684 if (alloc->is_Allocate()) { 3685 return alloc->as_Allocate(); 3686 } 3687 } 3688 return NULL; 3689 } 3690 3691 // Trace Allocate -> Proj[Parm] -> Initialize 3692 InitializeNode* AllocateNode::initialization() { 3693 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3694 if (rawoop == NULL) return NULL; 3695 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3696 Node* init = rawoop->fast_out(i); 3697 if (init->is_Initialize()) { 3698 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3699 return init->as_Initialize(); 3700 } 3701 } 3702 return NULL; 3703 } 3704 3705 //----------------------------- loop predicates --------------------------- 3706 3707 //------------------------------add_predicate_impl---------------------------- 3708 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3709 // Too many traps seen? 3710 if (too_many_traps(reason)) { 3711 #ifdef ASSERT 3712 if (TraceLoopPredicate) { 3713 int tc = C->trap_count(reason); 3714 tty->print("too many traps=%s tcount=%d in ", 3715 Deoptimization::trap_reason_name(reason), tc); 3716 method()->print(); // which method has too many predicate traps 3717 tty->cr(); 3718 } 3719 #endif 3720 // We cannot afford to take more traps here, 3721 // do not generate predicate. 3722 return; 3723 } 3724 3725 Node *cont = _gvn.intcon(1); 3726 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3727 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3728 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3729 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3730 C->add_predicate_opaq(opq); 3731 { 3732 PreserveJVMState pjvms(this); 3733 set_control(iffalse); 3734 inc_sp(nargs); 3735 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3736 } 3737 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3738 set_control(iftrue); 3739 } 3740 3741 //------------------------------add_predicate--------------------------------- 3742 void GraphKit::add_predicate(int nargs) { 3743 if (UseLoopPredicate) { 3744 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3745 } 3746 // loop's limit check predicate should be near the loop. 3747 if (LoopLimitCheck) { 3748 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3749 } 3750 } 3751 3752 //----------------------------- store barriers ---------------------------- 3753 #define __ ideal. 3754 3755 void GraphKit::sync_kit(IdealKit& ideal) { 3756 set_all_memory(__ merged_memory()); 3757 set_i_o(__ i_o()); 3758 set_control(__ ctrl()); 3759 } 3760 3761 void GraphKit::final_sync(IdealKit& ideal) { 3762 // Final sync IdealKit and graphKit. 3763 sync_kit(ideal); 3764 } 3765 3766 Node* GraphKit::byte_map_base_node() { 3767 // Get base of card map 3768 CardTableModRefBS* ct = 3769 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3770 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3771 if (ct->byte_map_base != NULL) { 3772 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3773 } else { 3774 return null(); 3775 } 3776 } 3777 3778 // vanilla/CMS post barrier 3779 // Insert a write-barrier store. This is to let generational GC work; we have 3780 // to flag all oop-stores before the next GC point. 3781 void GraphKit::write_barrier_post(Node* oop_store, 3782 Node* obj, 3783 Node* adr, 3784 uint adr_idx, 3785 Node* val, 3786 bool use_precise) { 3787 // No store check needed if we're storing a NULL or an old object 3788 // (latter case is probably a string constant). The concurrent 3789 // mark sweep garbage collector, however, needs to have all nonNull 3790 // oop updates flagged via card-marks. 3791 if (val != NULL && val->is_Con()) { 3792 // must be either an oop or NULL 3793 const Type* t = val->bottom_type(); 3794 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3795 // stores of null never (?) need barriers 3796 return; 3797 } 3798 3799 if (use_ReduceInitialCardMarks() 3800 && obj == just_allocated_object(control())) { 3801 // We can skip marks on a freshly-allocated object in Eden. 3802 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3803 // That routine informs GC to take appropriate compensating steps, 3804 // upon a slow-path allocation, so as to make this card-mark 3805 // elision safe. 3806 return; 3807 } 3808 3809 if (!use_precise) { 3810 // All card marks for a (non-array) instance are in one place: 3811 adr = obj; 3812 } 3813 // (Else it's an array (or unknown), and we want more precise card marks.) 3814 assert(adr != NULL, ""); 3815 3816 IdealKit ideal(this, true); 3817 3818 // Convert the pointer to an int prior to doing math on it 3819 Node* cast = __ CastPX(__ ctrl(), adr); 3820 3821 // Divide by card size 3822 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3823 "Only one we handle so far."); 3824 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3825 3826 // Combine card table base and card offset 3827 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3828 3829 // Get the alias_index for raw card-mark memory 3830 int adr_type = Compile::AliasIdxRaw; 3831 Node* zero = __ ConI(0); // Dirty card value 3832 BasicType bt = T_BYTE; 3833 3834 if (UseConcMarkSweepGC && UseCondCardMark) { 3835 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3836 __ sync_kit(this); 3837 } 3838 3839 if (UseCondCardMark) { 3840 // The classic GC reference write barrier is typically implemented 3841 // as a store into the global card mark table. Unfortunately 3842 // unconditional stores can result in false sharing and excessive 3843 // coherence traffic as well as false transactional aborts. 3844 // UseCondCardMark enables MP "polite" conditional card mark 3845 // stores. In theory we could relax the load from ctrl() to 3846 // no_ctrl, but that doesn't buy much latitude. 3847 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3848 __ if_then(card_val, BoolTest::ne, zero); 3849 } 3850 3851 // Smash zero into card 3852 if( !UseConcMarkSweepGC ) { 3853 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3854 } else { 3855 // Specialized path for CM store barrier 3856 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3857 } 3858 3859 if (UseCondCardMark) { 3860 __ end_if(); 3861 } 3862 3863 // Final sync IdealKit and GraphKit. 3864 final_sync(ideal); 3865 } 3866 /* 3867 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3868 * required by SATB to make sure all objects live at the start of the 3869 * marking are kept alive, all reference updates need to any previous 3870 * reference stored before writing. 3871 * 3872 * If the previous value is NULL there is no need to save the old value. 3873 * References that are NULL are filtered during runtime by the barrier 3874 * code to avoid unnecessary queuing. 3875 * 3876 * However in the case of newly allocated objects it might be possible to 3877 * prove that the reference about to be overwritten is NULL during compile 3878 * time and avoid adding the barrier code completely. 3879 * 3880 * The compiler needs to determine that the object in which a field is about 3881 * to be written is newly allocated, and that no prior store to the same field 3882 * has happened since the allocation. 3883 * 3884 * Returns true if the pre-barrier can be removed 3885 */ 3886 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3887 BasicType bt, uint adr_idx) { 3888 intptr_t offset = 0; 3889 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3890 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3891 3892 if (offset == Type::OffsetBot) { 3893 return false; // cannot unalias unless there are precise offsets 3894 } 3895 3896 if (alloc == NULL) { 3897 return false; // No allocation found 3898 } 3899 3900 intptr_t size_in_bytes = type2aelembytes(bt); 3901 3902 Node* mem = memory(adr_idx); // start searching here... 3903 3904 for (int cnt = 0; cnt < 50; cnt++) { 3905 3906 if (mem->is_Store()) { 3907 3908 Node* st_adr = mem->in(MemNode::Address); 3909 intptr_t st_offset = 0; 3910 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3911 3912 if (st_base == NULL) { 3913 break; // inscrutable pointer 3914 } 3915 3916 // Break we have found a store with same base and offset as ours so break 3917 if (st_base == base && st_offset == offset) { 3918 break; 3919 } 3920 3921 if (st_offset != offset && st_offset != Type::OffsetBot) { 3922 const int MAX_STORE = BytesPerLong; 3923 if (st_offset >= offset + size_in_bytes || 3924 st_offset <= offset - MAX_STORE || 3925 st_offset <= offset - mem->as_Store()->memory_size()) { 3926 // Success: The offsets are provably independent. 3927 // (You may ask, why not just test st_offset != offset and be done? 3928 // The answer is that stores of different sizes can co-exist 3929 // in the same sequence of RawMem effects. We sometimes initialize 3930 // a whole 'tile' of array elements with a single jint or jlong.) 3931 mem = mem->in(MemNode::Memory); 3932 continue; // advance through independent store memory 3933 } 3934 } 3935 3936 if (st_base != base 3937 && MemNode::detect_ptr_independence(base, alloc, st_base, 3938 AllocateNode::Ideal_allocation(st_base, phase), 3939 phase)) { 3940 // Success: The bases are provably independent. 3941 mem = mem->in(MemNode::Memory); 3942 continue; // advance through independent store memory 3943 } 3944 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3945 3946 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3947 AllocateNode* st_alloc = st_init->allocation(); 3948 3949 // Make sure that we are looking at the same allocation site. 3950 // The alloc variable is guaranteed to not be null here from earlier check. 3951 if (alloc == st_alloc) { 3952 // Check that the initialization is storing NULL so that no previous store 3953 // has been moved up and directly write a reference 3954 Node* captured_store = st_init->find_captured_store(offset, 3955 type2aelembytes(T_OBJECT), 3956 phase); 3957 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3958 return true; 3959 } 3960 } 3961 } 3962 3963 // Unless there is an explicit 'continue', we must bail out here, 3964 // because 'mem' is an inscrutable memory state (e.g., a call). 3965 break; 3966 } 3967 3968 return false; 3969 } 3970 3971 // G1 pre/post barriers 3972 void GraphKit::g1_write_barrier_pre(bool do_load, 3973 Node* obj, 3974 Node* adr, 3975 uint alias_idx, 3976 Node* val, 3977 const TypeOopPtr* val_type, 3978 Node* pre_val, 3979 BasicType bt) { 3980 3981 // Some sanity checks 3982 // Note: val is unused in this routine. 3983 3984 if (do_load) { 3985 // We need to generate the load of the previous value 3986 assert(obj != NULL, "must have a base"); 3987 assert(adr != NULL, "where are loading from?"); 3988 assert(pre_val == NULL, "loaded already?"); 3989 assert(val_type != NULL, "need a type"); 3990 3991 if (use_ReduceInitialCardMarks() 3992 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 3993 return; 3994 } 3995 3996 } else { 3997 // In this case both val_type and alias_idx are unused. 3998 assert(pre_val != NULL, "must be loaded already"); 3999 // Nothing to be done if pre_val is null. 4000 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4001 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4002 } 4003 assert(bt == T_OBJECT, "or we shouldn't be here"); 4004 4005 IdealKit ideal(this, true); 4006 4007 Node* tls = __ thread(); // ThreadLocalStorage 4008 4009 Node* no_ctrl = NULL; 4010 Node* no_base = __ top(); 4011 Node* zero = __ ConI(0); 4012 Node* zeroX = __ ConX(0); 4013 4014 float likely = PROB_LIKELY(0.999); 4015 float unlikely = PROB_UNLIKELY(0.999); 4016 4017 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4018 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 4019 4020 // Offsets into the thread 4021 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4022 PtrQueue::byte_offset_of_active()); 4023 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4024 PtrQueue::byte_offset_of_index()); 4025 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4026 PtrQueue::byte_offset_of_buf()); 4027 4028 // Now the actual pointers into the thread 4029 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4030 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4031 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4032 4033 // Now some of the values 4034 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4035 4036 // if (!marking) 4037 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4038 BasicType index_bt = TypeX_X->basic_type(); 4039 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 4040 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4041 4042 if (do_load) { 4043 // load original value 4044 // alias_idx correct?? 4045 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4046 } 4047 4048 // if (pre_val != NULL) 4049 __ if_then(pre_val, BoolTest::ne, null()); { 4050 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4051 4052 // is the queue for this thread full? 4053 __ if_then(index, BoolTest::ne, zeroX, likely); { 4054 4055 // decrement the index 4056 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4057 4058 // Now get the buffer location we will log the previous value into and store it 4059 Node *log_addr = __ AddP(no_base, buffer, next_index); 4060 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4061 // update the index 4062 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4063 4064 } __ else_(); { 4065 4066 // logging buffer is full, call the runtime 4067 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4068 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4069 } __ end_if(); // (!index) 4070 } __ end_if(); // (pre_val != NULL) 4071 } __ end_if(); // (!marking) 4072 4073 // Final sync IdealKit and GraphKit. 4074 final_sync(ideal); 4075 } 4076 4077 /* 4078 * G1 similar to any GC with a Young Generation requires a way to keep track of 4079 * references from Old Generation to Young Generation to make sure all live 4080 * objects are found. G1 also requires to keep track of object references 4081 * between different regions to enable evacuation of old regions, which is done 4082 * as part of mixed collections. References are tracked in remembered sets and 4083 * is continuously updated as reference are written to with the help of the 4084 * post-barrier. 4085 * 4086 * To reduce the number of updates to the remembered set the post-barrier 4087 * filters updates to fields in objects located in the Young Generation, 4088 * the same region as the reference, when the NULL is being written or 4089 * if the card is already marked as dirty by an earlier write. 4090 * 4091 * Under certain circumstances it is possible to avoid generating the 4092 * post-barrier completely if it is possible during compile time to prove 4093 * the object is newly allocated and that no safepoint exists between the 4094 * allocation and the store. 4095 * 4096 * In the case of slow allocation the allocation code must handle the barrier 4097 * as part of the allocation in the case the allocated object is not located 4098 * in the nursery, this would happen for humongous objects. This is similar to 4099 * how CMS is required to handle this case, see the comments for the method 4100 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4101 * A deferred card mark is required for these objects and handled in the above 4102 * mentioned methods. 4103 * 4104 * Returns true if the post barrier can be removed 4105 */ 4106 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4107 Node* adr) { 4108 intptr_t offset = 0; 4109 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4110 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4111 4112 if (offset == Type::OffsetBot) { 4113 return false; // cannot unalias unless there are precise offsets 4114 } 4115 4116 if (alloc == NULL) { 4117 return false; // No allocation found 4118 } 4119 4120 // Start search from Store node 4121 Node* mem = store->in(MemNode::Control); 4122 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4123 4124 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4125 AllocateNode* st_alloc = st_init->allocation(); 4126 4127 // Make sure we are looking at the same allocation 4128 if (alloc == st_alloc) { 4129 return true; 4130 } 4131 } 4132 4133 return false; 4134 } 4135 4136 // 4137 // Update the card table and add card address to the queue 4138 // 4139 void GraphKit::g1_mark_card(IdealKit& ideal, 4140 Node* card_adr, 4141 Node* oop_store, 4142 uint oop_alias_idx, 4143 Node* index, 4144 Node* index_adr, 4145 Node* buffer, 4146 const TypeFunc* tf) { 4147 4148 Node* zero = __ ConI(0); 4149 Node* zeroX = __ ConX(0); 4150 Node* no_base = __ top(); 4151 BasicType card_bt = T_BYTE; 4152 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4153 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4154 4155 // Now do the queue work 4156 __ if_then(index, BoolTest::ne, zeroX); { 4157 4158 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4159 Node* log_addr = __ AddP(no_base, buffer, next_index); 4160 4161 // Order, see storeCM. 4162 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4163 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4164 4165 } __ else_(); { 4166 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4167 } __ end_if(); 4168 4169 } 4170 4171 void GraphKit::g1_write_barrier_post(Node* oop_store, 4172 Node* obj, 4173 Node* adr, 4174 uint alias_idx, 4175 Node* val, 4176 BasicType bt, 4177 bool use_precise) { 4178 // If we are writing a NULL then we need no post barrier 4179 4180 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4181 // Must be NULL 4182 const Type* t = val->bottom_type(); 4183 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4184 // No post barrier if writing NULLx 4185 return; 4186 } 4187 4188 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4189 // We can skip marks on a freshly-allocated object in Eden. 4190 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4191 // That routine informs GC to take appropriate compensating steps, 4192 // upon a slow-path allocation, so as to make this card-mark 4193 // elision safe. 4194 return; 4195 } 4196 4197 if (use_ReduceInitialCardMarks() 4198 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4199 return; 4200 } 4201 4202 if (!use_precise) { 4203 // All card marks for a (non-array) instance are in one place: 4204 adr = obj; 4205 } 4206 // (Else it's an array (or unknown), and we want more precise card marks.) 4207 assert(adr != NULL, ""); 4208 4209 IdealKit ideal(this, true); 4210 4211 Node* tls = __ thread(); // ThreadLocalStorage 4212 4213 Node* no_base = __ top(); 4214 float likely = PROB_LIKELY(0.999); 4215 float unlikely = PROB_UNLIKELY(0.999); 4216 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4217 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4218 Node* zeroX = __ ConX(0); 4219 4220 // Get the alias_index for raw card-mark memory 4221 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4222 4223 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4224 4225 // Offsets into the thread 4226 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4227 PtrQueue::byte_offset_of_index()); 4228 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4229 PtrQueue::byte_offset_of_buf()); 4230 4231 // Pointers into the thread 4232 4233 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4234 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4235 4236 // Now some values 4237 // Use ctrl to avoid hoisting these values past a safepoint, which could 4238 // potentially reset these fields in the JavaThread. 4239 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4240 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4241 4242 // Convert the store obj pointer to an int prior to doing math on it 4243 // Must use ctrl to prevent "integerized oop" existing across safepoint 4244 Node* cast = __ CastPX(__ ctrl(), adr); 4245 4246 // Divide pointer by card size 4247 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4248 4249 // Combine card table base and card offset 4250 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4251 4252 // If we know the value being stored does it cross regions? 4253 4254 if (val != NULL) { 4255 // Does the store cause us to cross regions? 4256 4257 // Should be able to do an unsigned compare of region_size instead of 4258 // and extra shift. Do we have an unsigned compare?? 4259 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4260 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4261 4262 // if (xor_res == 0) same region so skip 4263 __ if_then(xor_res, BoolTest::ne, zeroX); { 4264 4265 // No barrier if we are storing a NULL 4266 __ if_then(val, BoolTest::ne, null(), unlikely); { 4267 4268 // Ok must mark the card if not already dirty 4269 4270 // load the original value of the card 4271 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4272 4273 __ if_then(card_val, BoolTest::ne, young_card); { 4274 sync_kit(ideal); 4275 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4276 insert_mem_bar(Op_MemBarVolatile, oop_store); 4277 __ sync_kit(this); 4278 4279 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4280 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4281 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4282 } __ end_if(); 4283 } __ end_if(); 4284 } __ end_if(); 4285 } __ end_if(); 4286 } else { 4287 // Object.clone() instrinsic uses this path. 4288 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4289 } 4290 4291 // Final sync IdealKit and GraphKit. 4292 final_sync(ideal); 4293 } 4294 #undef __ 4295 4296 4297 4298 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4299 if (java_lang_String::has_offset_field()) { 4300 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4301 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4302 false, NULL, 0); 4303 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4304 int offset_field_idx = C->get_alias_index(offset_field_type); 4305 4306 str = shenandoah_read_barrier(str); 4307 4308 return make_load(ctrl, 4309 basic_plus_adr(str, str, offset_offset), 4310 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4311 } else { 4312 return intcon(0); 4313 } 4314 } 4315 4316 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4317 if (java_lang_String::has_count_field()) { 4318 int count_offset = java_lang_String::count_offset_in_bytes(); 4319 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4320 false, NULL, 0); 4321 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4322 int count_field_idx = C->get_alias_index(count_field_type); 4323 4324 str = shenandoah_read_barrier(str); 4325 4326 return make_load(ctrl, 4327 basic_plus_adr(str, str, count_offset), 4328 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4329 } else { 4330 return load_array_length(load_String_value(ctrl, str)); 4331 } 4332 } 4333 4334 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4335 int value_offset = java_lang_String::value_offset_in_bytes(); 4336 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4337 false, NULL, 0); 4338 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4339 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4340 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4341 ciTypeArrayKlass::make(T_CHAR), true, 0); 4342 int value_field_idx = C->get_alias_index(value_field_type); 4343 4344 str = shenandoah_read_barrier(str); 4345 4346 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4347 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4348 // String.value field is known to be @Stable. 4349 if (UseImplicitStableValues) { 4350 load = cast_array_to_stable(load, value_type); 4351 } 4352 return load; 4353 } 4354 4355 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4356 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4357 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4358 false, NULL, 0); 4359 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4360 int offset_field_idx = C->get_alias_index(offset_field_type); 4361 4362 // TODO: Use incoming ctrl. 4363 str = shenandoah_write_barrier(str); 4364 4365 store_to_memory(UseShenandoahGC ? control() : ctrl, basic_plus_adr(str, offset_offset), 4366 value, T_INT, offset_field_idx, MemNode::unordered); 4367 } 4368 4369 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4370 int value_offset = java_lang_String::value_offset_in_bytes(); 4371 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4372 false, NULL, 0); 4373 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4374 4375 // TODO: Use incoming ctrl. 4376 str = shenandoah_write_barrier(str); 4377 value = shenandoah_read_barrier_nomem(value); 4378 4379 store_oop_to_object(UseShenandoahGC ? control() : ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4380 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4381 } 4382 4383 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4384 int count_offset = java_lang_String::count_offset_in_bytes(); 4385 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4386 false, NULL, 0); 4387 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4388 int count_field_idx = C->get_alias_index(count_field_type); 4389 4390 // TODO: Use incoming ctrl. 4391 str = shenandoah_write_barrier(str); 4392 4393 store_to_memory(UseShenandoahGC ? control() : ctrl, basic_plus_adr(str, count_offset), 4394 value, T_INT, count_field_idx, MemNode::unordered); 4395 } 4396 4397 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4398 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4399 // assumption of CCP analysis. 4400 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4401 } 4402 4403 Node* GraphKit::shenandoah_read_barrier(Node* obj) { 4404 return shenandoah_read_barrier_impl(obj, false, true); 4405 } 4406 4407 Node* GraphKit::shenandoah_read_barrier_nomem(Node* obj) { 4408 return shenandoah_read_barrier_impl(obj, false, false); 4409 } 4410 4411 Node* GraphKit::shenandoah_read_barrier_impl(Node* obj, bool use_ctrl, bool use_mem) { 4412 4413 if (UseShenandoahGC && ShenandoahReadBarrier) { 4414 const Type* obj_type = obj->bottom_type(); 4415 if (obj_type->higher_equal(TypePtr::NULL_PTR)) { 4416 // tty->print_cr("killed barrier for NULL object"); 4417 return obj; 4418 } 4419 const TypePtr* adr_type = obj_type->is_ptr()->add_offset(-8); 4420 Node* mem = use_mem ? memory(adr_type) : immutable_memory(); 4421 4422 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, mem)) { 4423 // We know it is null, no barrier needed. 4424 return obj; 4425 } 4426 4427 4428 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4429 4430 // We don't know if it's null or not. Need null-check. 4431 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4432 RegionNode* region = new RegionNode(PATH_LIMIT); 4433 Node* phi = new PhiNode(region, obj_type); 4434 Node* null_ctrl = top(); 4435 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4436 4437 region->init_req(_null_path, null_ctrl); 4438 phi ->init_req(_null_path, obj); 4439 4440 Node* ctrl = use_ctrl ? control() : NULL; 4441 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj); 4442 Node* n = _gvn.transform(rb); 4443 4444 region->init_req(_not_null_path, control()); 4445 phi ->init_req(_not_null_path, n); 4446 4447 set_control(_gvn.transform(region)); 4448 record_for_igvn(region); 4449 return _gvn.transform(phi); 4450 4451 } else { 4452 // We know it is not null. Simple barrier is sufficient. 4453 Node* ctrl = use_ctrl ? control() : NULL; 4454 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj); 4455 Node* n = _gvn.transform(rb); 4456 record_for_igvn(n); 4457 return n; 4458 } 4459 4460 } else { 4461 return obj; 4462 } 4463 } 4464 4465 Node* GraphKit::shenandoah_write_barrier(Node* obj) { 4466 4467 if (UseShenandoahGC && ShenandoahWriteBarrier) { 4468 4469 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, NULL)) { 4470 return obj; 4471 } 4472 const Type* obj_type = obj->bottom_type(); 4473 const TypePtr* adr_type = obj_type->is_ptr()->add_offset(-8); 4474 // tty->print_cr("memory at:"); 4475 // adr_type->dump(); 4476 // tty->print_cr("\n"); 4477 // memory(adr_type)->dump(); 4478 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) { 4479 // We don't know if it's null or not. Need null-check. 4480 enum { _not_null_path = 1, _null_path, PATH_LIMIT }; 4481 RegionNode* region = new RegionNode(PATH_LIMIT); 4482 Node* phi = new PhiNode(region, obj_type); 4483 Node* memphi = PhiNode::make(region, memory(adr_type), Type::MEMORY, C->alias_type(adr_type)->adr_type()); 4484 4485 Node* prev_mem = memory(adr_type); 4486 Node* null_ctrl = top(); 4487 Node* not_null_obj = null_check_oop(obj, &null_ctrl); 4488 4489 region->init_req(_null_path, null_ctrl); 4490 phi ->init_req(_null_path, null()); 4491 memphi->init_req(_null_path, prev_mem); 4492 4493 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(NULL, memory(adr_type), not_null_obj); 4494 Node* n = _gvn.transform(wb); 4495 if (n == wb) { // New barrier needs memory projection. 4496 Node* proj = _gvn.transform(new ShenandoahWBMemProjNode(n)); 4497 set_memory(proj, adr_type); 4498 } 4499 4500 region->init_req(_not_null_path, control()); 4501 phi ->init_req(_not_null_path, n); 4502 memphi->init_req(_not_null_path, memory(adr_type)); 4503 4504 set_control(_gvn.transform(region)); 4505 record_for_igvn(region); 4506 set_memory(_gvn.transform(memphi), adr_type); 4507 4508 Node* res_val = _gvn.transform(phi); 4509 // replace_in_map(obj, res_val); 4510 return res_val; 4511 } else { 4512 // We know it is not null. Simple barrier is sufficient. 4513 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(NULL, memory(adr_type), obj); 4514 Node* n = _gvn.transform(wb); 4515 if (n == wb) { 4516 Node* proj = _gvn.transform(new ShenandoahWBMemProjNode(wb)); 4517 set_memory(proj, adr_type); 4518 } 4519 // replace_in_map(obj, n); 4520 record_for_igvn(n); 4521 return n; 4522 } 4523 4524 } else { 4525 return obj; 4526 } 4527 } 4528 4529 void GraphKit::shenandoah_acmp_barrier(Node*& a, Node*& b) { 4530 if (UseShenandoahGC) { 4531 const Type* a_type = a->bottom_type(); 4532 const Type* b_type = b->bottom_type(); 4533 if (a_type->higher_equal(TypePtr::NULL_PTR) || b_type->higher_equal(TypePtr::NULL_PTR)) { 4534 // We know one arg is gonna be null. No need for barriers. 4535 // tty->print_cr("eliminate acmp barrier on null"); 4536 return; 4537 } 4538 /* 4539 if ((!a_type->isa_oopptr()) || (!b_type->isa_oopptr())) { 4540 a_type->dump(); 4541 b_type->dump(); 4542 } 4543 */ 4544 if (a_type->is_oopptr()->const_oop() != NULL && b_type->is_oopptr()->const_oop() != NULL ) { 4545 // We know one arg is inlined constant. No need for barriers. 4546 // tty->print_cr("eliminate acmp barrier on constant"); 4547 return; 4548 } 4549 if (a->Opcode() == Op_ShenandoahWriteBarrier && b->Opcode() == Op_ShenandoahWriteBarrier) { 4550 // We know one arg is already write-barrier'd. No need for barriers. 4551 // tty->print_cr("eliminate acmp barrier on write barrier"); 4552 return; 4553 } 4554 if (AllocateNode::Ideal_allocation(a, &_gvn) != NULL || AllocateNode::Ideal_allocation(b, &_gvn) != NULL) { 4555 // We know one arg is already in to-space. No need for barriers. 4556 // tty->print_cr("eliminate acmp barrier on new obj"); 4557 return; 4558 } 4559 4560 enum { _equal = 1, _not_equal, PATH_LIMIT }; 4561 RegionNode* region = new RegionNode(PATH_LIMIT); 4562 PhiNode* phiA = PhiNode::make(region, a); 4563 PhiNode* phiB = PhiNode::make(region, b); 4564 4565 Node* cmp = _gvn.transform(new CmpPNode(b, a)); 4566 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 4567 4568 // TODO: Use profiling data. 4569 IfNode* iff = create_and_map_if(control(), tst, PROB_FAIR, COUNT_UNKNOWN); 4570 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4571 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4572 4573 // Equal path: Use original values. 4574 region->init_req(_equal, iftrue); 4575 phiA->init_req(_equal, a); 4576 phiB->init_req(_equal, b); 4577 4578 // Unequal path: retry after read barriers. 4579 set_control(iffalse); 4580 a = shenandoah_read_barrier_impl(a, true, true); 4581 b = shenandoah_read_barrier_impl(b, true, true); 4582 4583 region->init_req(_not_equal, control()); 4584 phiA->init_req(_not_equal, a); 4585 phiB->init_req(_not_equal, b); 4586 4587 set_control(_gvn.transform(region)); 4588 record_for_igvn(region); 4589 4590 a = _gvn.transform(phiA); 4591 b = _gvn.transform(phiB); 4592 } 4593 }