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