1 /* 2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "ci/ciValueKlass.hpp" 28 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 29 #include "gc/g1/heapRegion.hpp" 30 #include "gc/shared/barrierSet.hpp" 31 #include "gc/shared/cardTableModRefBS.hpp" 32 #include "gc/shared/collectedHeap.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/graphKit.hpp" 38 #include "opto/idealKit.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/locknode.hpp" 41 #include "opto/machnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/valuetypenode.hpp" 47 #include "runtime/deoptimization.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 50 //----------------------------GraphKit----------------------------------------- 51 // Main utility constructor. 52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 53 : Phase(Phase::Parser), 54 _env(C->env()), 55 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()) 56 { 57 _exceptions = jvms->map()->next_exception(); 58 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 59 set_jvms(jvms); 60 #ifdef ASSERT 61 if (_gvn.is_IterGVN() != NULL) { 62 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 63 _worklist_size = _gvn.C->for_igvn()->size(); 64 } 65 #endif 66 } 67 68 // Private constructor for parser. 69 GraphKit::GraphKit() 70 : Phase(Phase::Parser), 71 _env(C->env()), 72 _gvn(*C->initial_gvn()) 73 { 74 _exceptions = NULL; 75 set_map(NULL); 76 debug_only(_sp = -99); 77 debug_only(set_bci(-99)); 78 } 79 80 81 82 //---------------------------clean_stack--------------------------------------- 83 // Clear away rubbish from the stack area of the JVM state. 84 // This destroys any arguments that may be waiting on the stack. 85 void GraphKit::clean_stack(int from_sp) { 86 SafePointNode* map = this->map(); 87 JVMState* jvms = this->jvms(); 88 int stk_size = jvms->stk_size(); 89 int stkoff = jvms->stkoff(); 90 Node* top = this->top(); 91 for (int i = from_sp; i < stk_size; i++) { 92 if (map->in(stkoff + i) != top) { 93 map->set_req(stkoff + i, top); 94 } 95 } 96 } 97 98 99 //--------------------------------sync_jvms----------------------------------- 100 // Make sure our current jvms agrees with our parse state. 101 JVMState* GraphKit::sync_jvms() const { 102 JVMState* jvms = this->jvms(); 103 jvms->set_bci(bci()); // Record the new bci in the JVMState 104 jvms->set_sp(sp()); // Record the new sp in the JVMState 105 assert(jvms_in_sync(), "jvms is now in sync"); 106 return jvms; 107 } 108 109 //--------------------------------sync_jvms_for_reexecute--------------------- 110 // Make sure our current jvms agrees with our parse state. This version 111 // uses the reexecute_sp for reexecuting bytecodes. 112 JVMState* GraphKit::sync_jvms_for_reexecute() { 113 JVMState* jvms = this->jvms(); 114 jvms->set_bci(bci()); // Record the new bci in the JVMState 115 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 116 return jvms; 117 } 118 119 #ifdef ASSERT 120 bool GraphKit::jvms_in_sync() const { 121 Parse* parse = is_Parse(); 122 if (parse == NULL) { 123 if (bci() != jvms()->bci()) return false; 124 if (sp() != (int)jvms()->sp()) return false; 125 return true; 126 } 127 if (jvms()->method() != parse->method()) return false; 128 if (jvms()->bci() != parse->bci()) return false; 129 int jvms_sp = jvms()->sp(); 130 if (jvms_sp != parse->sp()) return false; 131 int jvms_depth = jvms()->depth(); 132 if (jvms_depth != parse->depth()) return false; 133 return true; 134 } 135 136 // Local helper checks for special internal merge points 137 // used to accumulate and merge exception states. 138 // They are marked by the region's in(0) edge being the map itself. 139 // Such merge points must never "escape" into the parser at large, 140 // until they have been handed to gvn.transform. 141 static bool is_hidden_merge(Node* reg) { 142 if (reg == NULL) return false; 143 if (reg->is_Phi()) { 144 reg = reg->in(0); 145 if (reg == NULL) return false; 146 } 147 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 148 } 149 150 void GraphKit::verify_map() const { 151 if (map() == NULL) return; // null map is OK 152 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 153 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 154 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 155 } 156 157 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 158 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 159 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 160 } 161 #endif 162 163 //---------------------------stop_and_kill_map--------------------------------- 164 // Set _map to NULL, signalling a stop to further bytecode execution. 165 // First smash the current map's control to a constant, to mark it dead. 166 void GraphKit::stop_and_kill_map() { 167 SafePointNode* dead_map = stop(); 168 if (dead_map != NULL) { 169 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 170 assert(dead_map->is_killed(), "must be so marked"); 171 } 172 } 173 174 175 //--------------------------------stopped-------------------------------------- 176 // Tell if _map is NULL, or control is top. 177 bool GraphKit::stopped() { 178 if (map() == NULL) return true; 179 else if (control() == top()) return true; 180 else return false; 181 } 182 183 184 //-----------------------------has_ex_handler---------------------------------- 185 // Tell if this method or any caller method has exception handlers. 186 bool GraphKit::has_ex_handler() { 187 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 188 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 189 return true; 190 } 191 } 192 return false; 193 } 194 195 //------------------------------save_ex_oop------------------------------------ 196 // Save an exception without blowing stack contents or other JVM state. 197 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 198 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 199 ex_map->add_req(ex_oop); 200 debug_only(verify_exception_state(ex_map)); 201 } 202 203 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 204 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 205 Node* ex_oop = ex_map->in(ex_map->req()-1); 206 if (clear_it) ex_map->del_req(ex_map->req()-1); 207 return ex_oop; 208 } 209 210 //-----------------------------saved_ex_oop------------------------------------ 211 // Recover a saved exception from its map. 212 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 213 return common_saved_ex_oop(ex_map, false); 214 } 215 216 //--------------------------clear_saved_ex_oop--------------------------------- 217 // Erase a previously saved exception from its map. 218 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 219 return common_saved_ex_oop(ex_map, true); 220 } 221 222 #ifdef ASSERT 223 //---------------------------has_saved_ex_oop---------------------------------- 224 // Erase a previously saved exception from its map. 225 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 226 return ex_map->req() == ex_map->jvms()->endoff()+1; 227 } 228 #endif 229 230 //-------------------------make_exception_state-------------------------------- 231 // Turn the current JVM state into an exception state, appending the ex_oop. 232 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 233 sync_jvms(); 234 SafePointNode* ex_map = stop(); // do not manipulate this map any more 235 set_saved_ex_oop(ex_map, ex_oop); 236 return ex_map; 237 } 238 239 240 //--------------------------add_exception_state-------------------------------- 241 // Add an exception to my list of exceptions. 242 void GraphKit::add_exception_state(SafePointNode* ex_map) { 243 if (ex_map == NULL || ex_map->control() == top()) { 244 return; 245 } 246 #ifdef ASSERT 247 verify_exception_state(ex_map); 248 if (has_exceptions()) { 249 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 250 } 251 #endif 252 253 // If there is already an exception of exactly this type, merge with it. 254 // In particular, null-checks and other low-level exceptions common up here. 255 Node* ex_oop = saved_ex_oop(ex_map); 256 const Type* ex_type = _gvn.type(ex_oop); 257 if (ex_oop == top()) { 258 // No action needed. 259 return; 260 } 261 assert(ex_type->isa_instptr(), "exception must be an instance"); 262 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 263 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 264 // We check sp also because call bytecodes can generate exceptions 265 // both before and after arguments are popped! 266 if (ex_type2 == ex_type 267 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 268 combine_exception_states(ex_map, e2); 269 return; 270 } 271 } 272 273 // No pre-existing exception of the same type. Chain it on the list. 274 push_exception_state(ex_map); 275 } 276 277 //-----------------------add_exception_states_from----------------------------- 278 void GraphKit::add_exception_states_from(JVMState* jvms) { 279 SafePointNode* ex_map = jvms->map()->next_exception(); 280 if (ex_map != NULL) { 281 jvms->map()->set_next_exception(NULL); 282 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 283 next_map = ex_map->next_exception(); 284 ex_map->set_next_exception(NULL); 285 add_exception_state(ex_map); 286 } 287 } 288 } 289 290 //-----------------------transfer_exceptions_into_jvms------------------------- 291 JVMState* GraphKit::transfer_exceptions_into_jvms() { 292 if (map() == NULL) { 293 // We need a JVMS to carry the exceptions, but the map has gone away. 294 // Create a scratch JVMS, cloned from any of the exception states... 295 if (has_exceptions()) { 296 _map = _exceptions; 297 _map = clone_map(); 298 _map->set_next_exception(NULL); 299 clear_saved_ex_oop(_map); 300 debug_only(verify_map()); 301 } else { 302 // ...or created from scratch 303 JVMState* jvms = new (C) JVMState(_method, NULL); 304 jvms->set_bci(_bci); 305 jvms->set_sp(_sp); 306 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 307 set_jvms(jvms); 308 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 309 set_all_memory(top()); 310 while (map()->req() < jvms->endoff()) map()->add_req(top()); 311 } 312 // (This is a kludge, in case you didn't notice.) 313 set_control(top()); 314 } 315 JVMState* jvms = sync_jvms(); 316 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 317 jvms->map()->set_next_exception(_exceptions); 318 _exceptions = NULL; // done with this set of exceptions 319 return jvms; 320 } 321 322 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 323 assert(is_hidden_merge(dstphi), "must be a special merge node"); 324 assert(is_hidden_merge(srcphi), "must be a special merge node"); 325 uint limit = srcphi->req(); 326 for (uint i = PhiNode::Input; i < limit; i++) { 327 dstphi->add_req(srcphi->in(i)); 328 } 329 } 330 static inline void add_one_req(Node* dstphi, Node* src) { 331 assert(is_hidden_merge(dstphi), "must be a special merge node"); 332 assert(!is_hidden_merge(src), "must not be a special merge node"); 333 dstphi->add_req(src); 334 } 335 336 //-----------------------combine_exception_states------------------------------ 337 // This helper function combines exception states by building phis on a 338 // specially marked state-merging region. These regions and phis are 339 // untransformed, and can build up gradually. The region is marked by 340 // having a control input of its exception map, rather than NULL. Such 341 // regions do not appear except in this function, and in use_exception_state. 342 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 343 if (failing()) return; // dying anyway... 344 JVMState* ex_jvms = ex_map->_jvms; 345 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 346 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 347 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 348 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 349 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 350 assert(ex_map->req() == phi_map->req(), "matching maps"); 351 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 352 Node* hidden_merge_mark = root(); 353 Node* region = phi_map->control(); 354 MergeMemNode* phi_mem = phi_map->merged_memory(); 355 MergeMemNode* ex_mem = ex_map->merged_memory(); 356 if (region->in(0) != hidden_merge_mark) { 357 // The control input is not (yet) a specially-marked region in phi_map. 358 // Make it so, and build some phis. 359 region = new RegionNode(2); 360 _gvn.set_type(region, Type::CONTROL); 361 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 362 region->init_req(1, phi_map->control()); 363 phi_map->set_control(region); 364 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 365 record_for_igvn(io_phi); 366 _gvn.set_type(io_phi, Type::ABIO); 367 phi_map->set_i_o(io_phi); 368 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 369 Node* m = mms.memory(); 370 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 371 record_for_igvn(m_phi); 372 _gvn.set_type(m_phi, Type::MEMORY); 373 mms.set_memory(m_phi); 374 } 375 } 376 377 // Either or both of phi_map and ex_map might already be converted into phis. 378 Node* ex_control = ex_map->control(); 379 // if there is special marking on ex_map also, we add multiple edges from src 380 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 381 // how wide was the destination phi_map, originally? 382 uint orig_width = region->req(); 383 384 if (add_multiple) { 385 add_n_reqs(region, ex_control); 386 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 387 } else { 388 // ex_map has no merges, so we just add single edges everywhere 389 add_one_req(region, ex_control); 390 add_one_req(phi_map->i_o(), ex_map->i_o()); 391 } 392 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 393 if (mms.is_empty()) { 394 // get a copy of the base memory, and patch some inputs into it 395 const TypePtr* adr_type = mms.adr_type(C); 396 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 397 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 398 mms.set_memory(phi); 399 // Prepare to append interesting stuff onto the newly sliced phi: 400 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 401 } 402 // Append stuff from ex_map: 403 if (add_multiple) { 404 add_n_reqs(mms.memory(), mms.memory2()); 405 } else { 406 add_one_req(mms.memory(), mms.memory2()); 407 } 408 } 409 uint limit = ex_map->req(); 410 for (uint i = TypeFunc::Parms; i < limit; i++) { 411 // Skip everything in the JVMS after tos. (The ex_oop follows.) 412 if (i == tos) i = ex_jvms->monoff(); 413 Node* src = ex_map->in(i); 414 Node* dst = phi_map->in(i); 415 if (src != dst) { 416 PhiNode* phi; 417 if (dst->in(0) != region) { 418 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 419 record_for_igvn(phi); 420 _gvn.set_type(phi, phi->type()); 421 phi_map->set_req(i, dst); 422 // Prepare to append interesting stuff onto the new phi: 423 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 424 } else { 425 assert(dst->is_Phi(), "nobody else uses a hidden region"); 426 phi = dst->as_Phi(); 427 } 428 if (add_multiple && src->in(0) == ex_control) { 429 // Both are phis. 430 add_n_reqs(dst, src); 431 } else { 432 while (dst->req() < region->req()) add_one_req(dst, src); 433 } 434 const Type* srctype = _gvn.type(src); 435 if (phi->type() != srctype) { 436 const Type* dsttype = phi->type()->meet_speculative(srctype); 437 if (phi->type() != dsttype) { 438 phi->set_type(dsttype); 439 _gvn.set_type(phi, dsttype); 440 } 441 } 442 } 443 } 444 phi_map->merge_replaced_nodes_with(ex_map); 445 } 446 447 //--------------------------use_exception_state-------------------------------- 448 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 449 if (failing()) { stop(); return top(); } 450 Node* region = phi_map->control(); 451 Node* hidden_merge_mark = root(); 452 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 453 Node* ex_oop = clear_saved_ex_oop(phi_map); 454 if (region->in(0) == hidden_merge_mark) { 455 // Special marking for internal ex-states. Process the phis now. 456 region->set_req(0, region); // now it's an ordinary region 457 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 458 // Note: Setting the jvms also sets the bci and sp. 459 set_control(_gvn.transform(region)); 460 uint tos = jvms()->stkoff() + sp(); 461 for (uint i = 1; i < tos; i++) { 462 Node* x = phi_map->in(i); 463 if (x->in(0) == region) { 464 assert(x->is_Phi(), "expected a special phi"); 465 phi_map->set_req(i, _gvn.transform(x)); 466 } 467 } 468 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 469 Node* x = mms.memory(); 470 if (x->in(0) == region) { 471 assert(x->is_Phi(), "nobody else uses a hidden region"); 472 mms.set_memory(_gvn.transform(x)); 473 } 474 } 475 if (ex_oop->in(0) == region) { 476 assert(ex_oop->is_Phi(), "expected a special phi"); 477 ex_oop = _gvn.transform(ex_oop); 478 } 479 } else { 480 set_jvms(phi_map->jvms()); 481 } 482 483 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 484 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 485 return ex_oop; 486 } 487 488 //---------------------------------java_bc------------------------------------- 489 Bytecodes::Code GraphKit::java_bc() const { 490 ciMethod* method = this->method(); 491 int bci = this->bci(); 492 if (method != NULL && bci != InvocationEntryBci) 493 return method->java_code_at_bci(bci); 494 else 495 return Bytecodes::_illegal; 496 } 497 498 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 499 bool must_throw) { 500 // if the exception capability is set, then we will generate code 501 // to check the JavaThread.should_post_on_exceptions flag to see 502 // if we actually need to report exception events (for this 503 // thread). If we don't need to report exception events, we will 504 // take the normal fast path provided by add_exception_events. If 505 // exception event reporting is enabled for this thread, we will 506 // take the uncommon_trap in the BuildCutout below. 507 508 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 509 Node* jthread = _gvn.transform(new ThreadLocalNode()); 510 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 511 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 512 513 // Test the should_post_on_exceptions_flag vs. 0 514 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 515 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 516 517 // Branch to slow_path if should_post_on_exceptions_flag was true 518 { BuildCutout unless(this, tst, PROB_MAX); 519 // Do not try anything fancy if we're notifying the VM on every throw. 520 // Cf. case Bytecodes::_athrow in parse2.cpp. 521 uncommon_trap(reason, Deoptimization::Action_none, 522 (ciKlass*)NULL, (char*)NULL, must_throw); 523 } 524 525 } 526 527 //------------------------------builtin_throw---------------------------------- 528 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 529 bool must_throw = true; 530 531 if (env()->jvmti_can_post_on_exceptions()) { 532 // check if we must post exception events, take uncommon trap if so 533 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 534 // here if should_post_on_exceptions is false 535 // continue on with the normal codegen 536 } 537 538 // If this particular condition has not yet happened at this 539 // bytecode, then use the uncommon trap mechanism, and allow for 540 // a future recompilation if several traps occur here. 541 // If the throw is hot, try to use a more complicated inline mechanism 542 // which keeps execution inside the compiled code. 543 bool treat_throw_as_hot = false; 544 ciMethodData* md = method()->method_data(); 545 546 if (ProfileTraps) { 547 if (too_many_traps(reason)) { 548 treat_throw_as_hot = true; 549 } 550 // (If there is no MDO at all, assume it is early in 551 // execution, and that any deopts are part of the 552 // startup transient, and don't need to be remembered.) 553 554 // Also, if there is a local exception handler, treat all throws 555 // as hot if there has been at least one in this method. 556 if (C->trap_count(reason) != 0 557 && method()->method_data()->trap_count(reason) != 0 558 && has_ex_handler()) { 559 treat_throw_as_hot = true; 560 } 561 } 562 563 // If this throw happens frequently, an uncommon trap might cause 564 // a performance pothole. If there is a local exception handler, 565 // and if this particular bytecode appears to be deoptimizing often, 566 // let us handle the throw inline, with a preconstructed instance. 567 // Note: If the deopt count has blown up, the uncommon trap 568 // runtime is going to flush this nmethod, not matter what. 569 if (treat_throw_as_hot 570 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 571 // If the throw is local, we use a pre-existing instance and 572 // punt on the backtrace. This would lead to a missing backtrace 573 // (a repeat of 4292742) if the backtrace object is ever asked 574 // for its backtrace. 575 // Fixing this remaining case of 4292742 requires some flavor of 576 // escape analysis. Leave that for the future. 577 ciInstance* ex_obj = NULL; 578 switch (reason) { 579 case Deoptimization::Reason_null_check: 580 ex_obj = env()->NullPointerException_instance(); 581 break; 582 case Deoptimization::Reason_div0_check: 583 ex_obj = env()->ArithmeticException_instance(); 584 break; 585 case Deoptimization::Reason_range_check: 586 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 587 break; 588 case Deoptimization::Reason_class_check: 589 if (java_bc() == Bytecodes::_aastore || java_bc() == Bytecodes::_vastore) { 590 ex_obj = env()->ArrayStoreException_instance(); 591 } else { 592 ex_obj = env()->ClassCastException_instance(); 593 } 594 break; 595 default: 596 break; 597 } 598 if (failing()) { stop(); return; } // exception allocation might fail 599 if (ex_obj != NULL) { 600 // Cheat with a preallocated exception object. 601 if (C->log() != NULL) 602 C->log()->elem("hot_throw preallocated='1' reason='%s'", 603 Deoptimization::trap_reason_name(reason)); 604 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 605 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 606 607 // Clear the detail message of the preallocated exception object. 608 // Weblogic sometimes mutates the detail message of exceptions 609 // using reflection. 610 int offset = java_lang_Throwable::get_detailMessage_offset(); 611 const TypePtr* adr_typ = ex_con->add_offset(offset); 612 613 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 614 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 615 // Conservatively release stores of object references. 616 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 617 618 add_exception_state(make_exception_state(ex_node)); 619 return; 620 } 621 } 622 623 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 624 // It won't be much cheaper than bailing to the interp., since we'll 625 // have to pass up all the debug-info, and the runtime will have to 626 // create the stack trace. 627 628 // Usual case: Bail to interpreter. 629 // Reserve the right to recompile if we haven't seen anything yet. 630 631 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 632 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 633 if (treat_throw_as_hot 634 && (method()->method_data()->trap_recompiled_at(bci(), m) 635 || C->too_many_traps(reason))) { 636 // We cannot afford to take more traps here. Suffer in the interpreter. 637 if (C->log() != NULL) 638 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 639 Deoptimization::trap_reason_name(reason), 640 C->trap_count(reason)); 641 action = Deoptimization::Action_none; 642 } 643 644 // "must_throw" prunes the JVM state to include only the stack, if there 645 // are no local exception handlers. This should cut down on register 646 // allocation time and code size, by drastically reducing the number 647 // of in-edges on the call to the uncommon trap. 648 649 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 650 } 651 652 653 //----------------------------PreserveJVMState--------------------------------- 654 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 655 debug_only(kit->verify_map()); 656 _kit = kit; 657 _map = kit->map(); // preserve the map 658 _sp = kit->sp(); 659 kit->set_map(clone_map ? kit->clone_map() : NULL); 660 #ifdef ASSERT 661 _bci = kit->bci(); 662 Parse* parser = kit->is_Parse(); 663 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 664 _block = block; 665 #endif 666 } 667 PreserveJVMState::~PreserveJVMState() { 668 GraphKit* kit = _kit; 669 #ifdef ASSERT 670 assert(kit->bci() == _bci, "bci must not shift"); 671 Parse* parser = kit->is_Parse(); 672 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 673 assert(block == _block, "block must not shift"); 674 #endif 675 kit->set_map(_map); 676 kit->set_sp(_sp); 677 } 678 679 680 //-----------------------------BuildCutout------------------------------------- 681 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 682 : PreserveJVMState(kit) 683 { 684 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 685 SafePointNode* outer_map = _map; // preserved map is caller's 686 SafePointNode* inner_map = kit->map(); 687 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 688 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 689 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 690 } 691 BuildCutout::~BuildCutout() { 692 GraphKit* kit = _kit; 693 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 694 } 695 696 //---------------------------PreserveReexecuteState---------------------------- 697 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 698 assert(!kit->stopped(), "must call stopped() before"); 699 _kit = kit; 700 _sp = kit->sp(); 701 _reexecute = kit->jvms()->_reexecute; 702 } 703 PreserveReexecuteState::~PreserveReexecuteState() { 704 if (_kit->stopped()) return; 705 _kit->jvms()->_reexecute = _reexecute; 706 _kit->set_sp(_sp); 707 } 708 709 //------------------------------clone_map-------------------------------------- 710 // Implementation of PreserveJVMState 711 // 712 // Only clone_map(...) here. If this function is only used in the 713 // PreserveJVMState class we may want to get rid of this extra 714 // function eventually and do it all there. 715 716 SafePointNode* GraphKit::clone_map() { 717 if (map() == NULL) return NULL; 718 719 // Clone the memory edge first 720 Node* mem = MergeMemNode::make(map()->memory()); 721 gvn().set_type_bottom(mem); 722 723 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 724 JVMState* jvms = this->jvms(); 725 JVMState* clonejvms = jvms->clone_shallow(C); 726 clonemap->set_memory(mem); 727 clonemap->set_jvms(clonejvms); 728 clonejvms->set_map(clonemap); 729 record_for_igvn(clonemap); 730 gvn().set_type_bottom(clonemap); 731 return clonemap; 732 } 733 734 735 //-----------------------------set_map_clone----------------------------------- 736 void GraphKit::set_map_clone(SafePointNode* m) { 737 _map = m; 738 _map = clone_map(); 739 _map->set_next_exception(NULL); 740 debug_only(verify_map()); 741 } 742 743 744 //----------------------------kill_dead_locals--------------------------------- 745 // Detect any locals which are known to be dead, and force them to top. 746 void GraphKit::kill_dead_locals() { 747 // Consult the liveness information for the locals. If any 748 // of them are unused, then they can be replaced by top(). This 749 // should help register allocation time and cut down on the size 750 // of the deoptimization information. 751 752 // This call is made from many of the bytecode handling 753 // subroutines called from the Big Switch in do_one_bytecode. 754 // Every bytecode which might include a slow path is responsible 755 // for killing its dead locals. The more consistent we 756 // are about killing deads, the fewer useless phis will be 757 // constructed for them at various merge points. 758 759 // bci can be -1 (InvocationEntryBci). We return the entry 760 // liveness for the method. 761 762 if (method() == NULL || method()->code_size() == 0) { 763 // We are building a graph for a call to a native method. 764 // All locals are live. 765 return; 766 } 767 768 ResourceMark rm; 769 770 // Consult the liveness information for the locals. If any 771 // of them are unused, then they can be replaced by top(). This 772 // should help register allocation time and cut down on the size 773 // of the deoptimization information. 774 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 775 776 int len = (int)live_locals.size(); 777 assert(len <= jvms()->loc_size(), "too many live locals"); 778 for (int local = 0; local < len; local++) { 779 if (!live_locals.at(local)) { 780 set_local(local, top()); 781 } 782 } 783 } 784 785 #ifdef ASSERT 786 //-------------------------dead_locals_are_killed------------------------------ 787 // Return true if all dead locals are set to top in the map. 788 // Used to assert "clean" debug info at various points. 789 bool GraphKit::dead_locals_are_killed() { 790 if (method() == NULL || method()->code_size() == 0) { 791 // No locals need to be dead, so all is as it should be. 792 return true; 793 } 794 795 // Make sure somebody called kill_dead_locals upstream. 796 ResourceMark rm; 797 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 798 if (jvms->loc_size() == 0) continue; // no locals to consult 799 SafePointNode* map = jvms->map(); 800 ciMethod* method = jvms->method(); 801 int bci = jvms->bci(); 802 if (jvms == this->jvms()) { 803 bci = this->bci(); // it might not yet be synched 804 } 805 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 806 int len = (int)live_locals.size(); 807 if (!live_locals.is_valid() || len == 0) 808 // This method is trivial, or is poisoned by a breakpoint. 809 return true; 810 assert(len == jvms->loc_size(), "live map consistent with locals map"); 811 for (int local = 0; local < len; local++) { 812 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 813 if (PrintMiscellaneous && (Verbose || WizardMode)) { 814 tty->print_cr("Zombie local %d: ", local); 815 jvms->dump(); 816 } 817 return false; 818 } 819 } 820 } 821 return true; 822 } 823 824 #endif //ASSERT 825 826 // Helper function for adding JVMState and debug information to node 827 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 828 // Add the safepoint edges to the call (or other safepoint). 829 830 // Make sure dead locals are set to top. This 831 // should help register allocation time and cut down on the size 832 // of the deoptimization information. 833 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 834 835 // Walk the inline list to fill in the correct set of JVMState's 836 // Also fill in the associated edges for each JVMState. 837 838 // If the bytecode needs to be reexecuted we need to put 839 // the arguments back on the stack. 840 const bool should_reexecute = jvms()->should_reexecute(); 841 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 842 843 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 844 // undefined if the bci is different. This is normal for Parse but it 845 // should not happen for LibraryCallKit because only one bci is processed. 846 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 847 "in LibraryCallKit the reexecute bit should not change"); 848 849 // If we are guaranteed to throw, we can prune everything but the 850 // input to the current bytecode. 851 bool can_prune_locals = false; 852 uint stack_slots_not_pruned = 0; 853 int inputs = 0, depth = 0; 854 if (must_throw) { 855 assert(method() == youngest_jvms->method(), "sanity"); 856 if (compute_stack_effects(inputs, depth)) { 857 can_prune_locals = true; 858 stack_slots_not_pruned = inputs; 859 } 860 } 861 862 if (env()->should_retain_local_variables()) { 863 // At any safepoint, this method can get breakpointed, which would 864 // then require an immediate deoptimization. 865 can_prune_locals = false; // do not prune locals 866 stack_slots_not_pruned = 0; 867 } 868 869 C->add_safepoint_edges(call, youngest_jvms, can_prune_locals, stack_slots_not_pruned); 870 } 871 872 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 873 Bytecodes::Code code = java_bc(); 874 if (code == Bytecodes::_wide) { 875 code = method()->java_code_at_bci(bci() + 1); 876 } 877 878 BasicType rtype = T_ILLEGAL; 879 int rsize = 0; 880 881 if (code != Bytecodes::_illegal) { 882 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 883 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 884 if (rtype < T_CONFLICT) 885 rsize = type2size[rtype]; 886 } 887 888 switch (code) { 889 case Bytecodes::_illegal: 890 return false; 891 892 case Bytecodes::_ldc: 893 case Bytecodes::_ldc_w: 894 case Bytecodes::_ldc2_w: 895 inputs = 0; 896 break; 897 898 case Bytecodes::_dup: inputs = 1; break; 899 case Bytecodes::_dup_x1: inputs = 2; break; 900 case Bytecodes::_dup_x2: inputs = 3; break; 901 case Bytecodes::_dup2: inputs = 2; break; 902 case Bytecodes::_dup2_x1: inputs = 3; break; 903 case Bytecodes::_dup2_x2: inputs = 4; break; 904 case Bytecodes::_swap: inputs = 2; break; 905 case Bytecodes::_arraylength: inputs = 1; break; 906 907 case Bytecodes::_getstatic: 908 case Bytecodes::_putstatic: 909 case Bytecodes::_getfield: 910 case Bytecodes::_putfield: 911 { 912 bool ignored_will_link; 913 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 914 int size = field->type()->size(); 915 bool is_get = (depth >= 0), is_static = (depth & 1); 916 inputs = (is_static ? 0 : 1); 917 if (is_get) { 918 depth = size - inputs; 919 } else { 920 inputs += size; // putxxx pops the value from the stack 921 depth = - inputs; 922 } 923 } 924 break; 925 926 case Bytecodes::_invokevirtual: 927 case Bytecodes::_invokespecial: 928 case Bytecodes::_invokestatic: 929 case Bytecodes::_invokedynamic: 930 case Bytecodes::_invokeinterface: 931 { 932 bool ignored_will_link; 933 ciSignature* declared_signature = NULL; 934 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 935 assert(declared_signature != NULL, "cannot be null"); 936 inputs = declared_signature->arg_size_for_bc(code); 937 int size = declared_signature->return_type()->size(); 938 depth = size - inputs; 939 } 940 break; 941 942 case Bytecodes::_multianewarray: 943 { 944 ciBytecodeStream iter(method()); 945 iter.reset_to_bci(bci()); 946 iter.next(); 947 inputs = iter.get_dimensions(); 948 assert(rsize == 1, ""); 949 depth = rsize - inputs; 950 } 951 break; 952 953 case Bytecodes::_vwithfield: { 954 bool ignored_will_link; 955 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 956 int size = field->type()->size(); 957 inputs = size+1; 958 depth = rsize - inputs; 959 break; 960 } 961 962 case Bytecodes::_ireturn: 963 case Bytecodes::_lreturn: 964 case Bytecodes::_freturn: 965 case Bytecodes::_dreturn: 966 case Bytecodes::_areturn: 967 case Bytecodes::_vreturn: 968 assert(rsize == -depth, ""); 969 inputs = rsize; 970 break; 971 972 case Bytecodes::_jsr: 973 case Bytecodes::_jsr_w: 974 inputs = 0; 975 depth = 1; // S.B. depth=1, not zero 976 break; 977 978 default: 979 // bytecode produces a typed result 980 inputs = rsize - depth; 981 assert(inputs >= 0, ""); 982 break; 983 } 984 985 #ifdef ASSERT 986 // spot check 987 int outputs = depth + inputs; 988 assert(outputs >= 0, "sanity"); 989 switch (code) { 990 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 991 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 992 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 993 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 994 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 995 default: break; 996 } 997 #endif //ASSERT 998 999 return true; 1000 } 1001 1002 1003 1004 //------------------------------basic_plus_adr--------------------------------- 1005 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1006 // short-circuit a common case 1007 if (offset == intcon(0)) return ptr; 1008 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1009 } 1010 1011 Node* GraphKit::ConvI2L(Node* offset) { 1012 // short-circuit a common case 1013 jint offset_con = find_int_con(offset, Type::OffsetBot); 1014 if (offset_con != Type::OffsetBot) { 1015 return longcon((jlong) offset_con); 1016 } 1017 return _gvn.transform( new ConvI2LNode(offset)); 1018 } 1019 1020 Node* GraphKit::ConvI2UL(Node* offset) { 1021 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1022 if (offset_con != (juint) Type::OffsetBot) { 1023 return longcon((julong) offset_con); 1024 } 1025 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1026 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1027 return _gvn.transform( new AndLNode(conv, mask) ); 1028 } 1029 1030 Node* GraphKit::ConvL2I(Node* offset) { 1031 // short-circuit a common case 1032 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1033 if (offset_con != (jlong)Type::OffsetBot) { 1034 return intcon((int) offset_con); 1035 } 1036 return _gvn.transform( new ConvL2INode(offset)); 1037 } 1038 1039 //-------------------------load_object_klass----------------------------------- 1040 Node* GraphKit::load_object_klass(Node* obj) { 1041 // Special-case a fresh allocation to avoid building nodes: 1042 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1043 if (akls != NULL) return akls; 1044 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1045 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1046 } 1047 1048 //-------------------------load_array_length----------------------------------- 1049 Node* GraphKit::load_array_length(Node* array) { 1050 // Special-case a fresh allocation to avoid building nodes: 1051 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1052 Node *alen; 1053 if (alloc == NULL) { 1054 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1055 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1056 } else { 1057 alen = alloc->Ideal_length(); 1058 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1059 if (ccast != alen) { 1060 alen = _gvn.transform(ccast); 1061 } 1062 } 1063 return alen; 1064 } 1065 1066 //------------------------------do_null_check---------------------------------- 1067 // Helper function to do a NULL pointer check. Returned value is 1068 // the incoming address with NULL casted away. You are allowed to use the 1069 // not-null value only if you are control dependent on the test. 1070 #ifndef PRODUCT 1071 extern int explicit_null_checks_inserted, 1072 explicit_null_checks_elided; 1073 #endif 1074 Node* GraphKit::null_check_common(Node* value, BasicType type, 1075 // optional arguments for variations: 1076 bool assert_null, 1077 Node* *null_control, 1078 bool speculative) { 1079 assert(!assert_null || null_control == NULL, "not both at once"); 1080 if (stopped()) return top(); 1081 NOT_PRODUCT(explicit_null_checks_inserted++); 1082 1083 // Construct NULL check 1084 Node *chk = NULL; 1085 switch(type) { 1086 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1087 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1088 case T_VALUETYPE : // fall through 1089 case T_ARRAY : // fall through 1090 type = T_OBJECT; // simplify further tests 1091 case T_OBJECT : { 1092 const Type *t = _gvn.type( value ); 1093 1094 const TypeOopPtr* tp = t->isa_oopptr(); 1095 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1096 // Only for do_null_check, not any of its siblings: 1097 && !assert_null && null_control == NULL) { 1098 // Usually, any field access or invocation on an unloaded oop type 1099 // will simply fail to link, since the statically linked class is 1100 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1101 // the static class is loaded but the sharper oop type is not. 1102 // Rather than checking for this obscure case in lots of places, 1103 // we simply observe that a null check on an unloaded class 1104 // will always be followed by a nonsense operation, so we 1105 // can just issue the uncommon trap here. 1106 // Our access to the unloaded class will only be correct 1107 // after it has been loaded and initialized, which requires 1108 // a trip through the interpreter. 1109 #ifndef PRODUCT 1110 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1111 #endif 1112 uncommon_trap(Deoptimization::Reason_unloaded, 1113 Deoptimization::Action_reinterpret, 1114 tp->klass(), "!loaded"); 1115 return top(); 1116 } 1117 1118 if (assert_null) { 1119 // See if the type is contained in NULL_PTR. 1120 // If so, then the value is already null. 1121 if (t->higher_equal(TypePtr::NULL_PTR)) { 1122 NOT_PRODUCT(explicit_null_checks_elided++); 1123 return value; // Elided null assert quickly! 1124 } 1125 } else { 1126 // See if mixing in the NULL pointer changes type. 1127 // If so, then the NULL pointer was not allowed in the original 1128 // type. In other words, "value" was not-null. 1129 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1130 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1131 NOT_PRODUCT(explicit_null_checks_elided++); 1132 return value; // Elided null check quickly! 1133 } 1134 } 1135 chk = new CmpPNode( value, null() ); 1136 break; 1137 } 1138 1139 default: 1140 fatal("unexpected type: %s", type2name(type)); 1141 } 1142 assert(chk != NULL, "sanity check"); 1143 chk = _gvn.transform(chk); 1144 1145 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1146 BoolNode *btst = new BoolNode( chk, btest); 1147 Node *tst = _gvn.transform( btst ); 1148 1149 //----------- 1150 // if peephole optimizations occurred, a prior test existed. 1151 // If a prior test existed, maybe it dominates as we can avoid this test. 1152 if (tst != btst && type == T_OBJECT) { 1153 // At this point we want to scan up the CFG to see if we can 1154 // find an identical test (and so avoid this test altogether). 1155 Node *cfg = control(); 1156 int depth = 0; 1157 while( depth < 16 ) { // Limit search depth for speed 1158 if( cfg->Opcode() == Op_IfTrue && 1159 cfg->in(0)->in(1) == tst ) { 1160 // Found prior test. Use "cast_not_null" to construct an identical 1161 // CastPP (and hence hash to) as already exists for the prior test. 1162 // Return that casted value. 1163 if (assert_null) { 1164 replace_in_map(value, null()); 1165 return null(); // do not issue the redundant test 1166 } 1167 Node *oldcontrol = control(); 1168 set_control(cfg); 1169 Node *res = cast_not_null(value); 1170 set_control(oldcontrol); 1171 NOT_PRODUCT(explicit_null_checks_elided++); 1172 return res; 1173 } 1174 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1175 if (cfg == NULL) break; // Quit at region nodes 1176 depth++; 1177 } 1178 } 1179 1180 //----------- 1181 // Branch to failure if null 1182 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1183 Deoptimization::DeoptReason reason; 1184 if (assert_null) { 1185 reason = Deoptimization::reason_null_assert(speculative); 1186 } else if (type == T_OBJECT) { 1187 reason = Deoptimization::reason_null_check(speculative); 1188 } else { 1189 reason = Deoptimization::Reason_div0_check; 1190 } 1191 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1192 // ciMethodData::has_trap_at will return a conservative -1 if any 1193 // must-be-null assertion has failed. This could cause performance 1194 // problems for a method after its first do_null_assert failure. 1195 // Consider using 'Reason_class_check' instead? 1196 1197 // To cause an implicit null check, we set the not-null probability 1198 // to the maximum (PROB_MAX). For an explicit check the probability 1199 // is set to a smaller value. 1200 if (null_control != NULL || too_many_traps(reason)) { 1201 // probability is less likely 1202 ok_prob = PROB_LIKELY_MAG(3); 1203 } else if (!assert_null && 1204 (ImplicitNullCheckThreshold > 0) && 1205 method() != NULL && 1206 (method()->method_data()->trap_count(reason) 1207 >= (uint)ImplicitNullCheckThreshold)) { 1208 ok_prob = PROB_LIKELY_MAG(3); 1209 } 1210 1211 if (null_control != NULL) { 1212 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1213 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1214 set_control( _gvn.transform( new IfTrueNode(iff))); 1215 #ifndef PRODUCT 1216 if (null_true == top()) { 1217 explicit_null_checks_elided++; 1218 } 1219 #endif 1220 (*null_control) = null_true; 1221 } else { 1222 BuildCutout unless(this, tst, ok_prob); 1223 // Check for optimizer eliding test at parse time 1224 if (stopped()) { 1225 // Failure not possible; do not bother making uncommon trap. 1226 NOT_PRODUCT(explicit_null_checks_elided++); 1227 } else if (assert_null) { 1228 uncommon_trap(reason, 1229 Deoptimization::Action_make_not_entrant, 1230 NULL, "assert_null"); 1231 } else { 1232 replace_in_map(value, zerocon(type)); 1233 builtin_throw(reason); 1234 } 1235 } 1236 1237 // Must throw exception, fall-thru not possible? 1238 if (stopped()) { 1239 return top(); // No result 1240 } 1241 1242 if (assert_null) { 1243 // Cast obj to null on this path. 1244 replace_in_map(value, zerocon(type)); 1245 return zerocon(type); 1246 } 1247 1248 // Cast obj to not-null on this path, if there is no null_control. 1249 // (If there is a null_control, a non-null value may come back to haunt us.) 1250 if (type == T_OBJECT) { 1251 Node* cast = cast_not_null(value, false); 1252 if (null_control == NULL || (*null_control) == top()) 1253 replace_in_map(value, cast); 1254 value = cast; 1255 } 1256 1257 return value; 1258 } 1259 1260 1261 //------------------------------cast_not_null---------------------------------- 1262 // Cast obj to not-null on this path 1263 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1264 const Type *t = _gvn.type(obj); 1265 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1266 // Object is already not-null? 1267 if( t == t_not_null ) return obj; 1268 1269 Node *cast = new CastPPNode(obj,t_not_null); 1270 cast->init_req(0, control()); 1271 cast = _gvn.transform( cast ); 1272 1273 // Scan for instances of 'obj' in the current JVM mapping. 1274 // These instances are known to be not-null after the test. 1275 if (do_replace_in_map) 1276 replace_in_map(obj, cast); 1277 1278 return cast; // Return casted value 1279 } 1280 1281 // Sometimes in intrinsics, we implicitly know an object is not null 1282 // (there's no actual null check) so we can cast it to not null. In 1283 // the course of optimizations, the input to the cast can become null. 1284 // In that case that data path will die and we need the control path 1285 // to become dead as well to keep the graph consistent. So we have to 1286 // add a check for null for which one branch can't be taken. It uses 1287 // an Opaque4 node that will cause the check to be removed after loop 1288 // opts so the test goes away and the compiled code doesn't execute a 1289 // useless check. 1290 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1291 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1292 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1293 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1294 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1295 _gvn.set_type(iff, iff->Value(&_gvn)); 1296 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1297 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1298 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1299 C->root()->add_req(halt); 1300 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1301 set_control(if_t); 1302 return cast_not_null(value, do_replace_in_map); 1303 } 1304 1305 1306 //--------------------------replace_in_map------------------------------------- 1307 void GraphKit::replace_in_map(Node* old, Node* neww) { 1308 if (old == neww) { 1309 return; 1310 } 1311 1312 map()->replace_edge(old, neww); 1313 1314 // Note: This operation potentially replaces any edge 1315 // on the map. This includes locals, stack, and monitors 1316 // of the current (innermost) JVM state. 1317 1318 // don't let inconsistent types from profiling escape this 1319 // method 1320 1321 const Type* told = _gvn.type(old); 1322 const Type* tnew = _gvn.type(neww); 1323 1324 if (!tnew->higher_equal(told)) { 1325 return; 1326 } 1327 1328 map()->record_replaced_node(old, neww); 1329 } 1330 1331 1332 //============================================================================= 1333 //--------------------------------memory--------------------------------------- 1334 Node* GraphKit::memory(uint alias_idx) { 1335 MergeMemNode* mem = merged_memory(); 1336 Node* p = mem->memory_at(alias_idx); 1337 _gvn.set_type(p, Type::MEMORY); // must be mapped 1338 return p; 1339 } 1340 1341 //-----------------------------reset_memory------------------------------------ 1342 Node* GraphKit::reset_memory() { 1343 Node* mem = map()->memory(); 1344 // do not use this node for any more parsing! 1345 debug_only( map()->set_memory((Node*)NULL) ); 1346 return _gvn.transform( mem ); 1347 } 1348 1349 //------------------------------set_all_memory--------------------------------- 1350 void GraphKit::set_all_memory(Node* newmem) { 1351 Node* mergemem = MergeMemNode::make(newmem); 1352 gvn().set_type_bottom(mergemem); 1353 map()->set_memory(mergemem); 1354 } 1355 1356 //------------------------------set_all_memory_call---------------------------- 1357 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1358 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1359 set_all_memory(newmem); 1360 } 1361 1362 //============================================================================= 1363 // 1364 // parser factory methods for MemNodes 1365 // 1366 // These are layered on top of the factory methods in LoadNode and StoreNode, 1367 // and integrate with the parser's memory state and _gvn engine. 1368 // 1369 1370 // factory methods in "int adr_idx" 1371 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1372 int adr_idx, 1373 MemNode::MemOrd mo, 1374 LoadNode::ControlDependency control_dependency, 1375 bool require_atomic_access, 1376 bool unaligned, 1377 bool mismatched) { 1378 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1379 const TypePtr* adr_type = NULL; // debug-mode-only argument 1380 debug_only(adr_type = C->get_adr_type(adr_idx)); 1381 Node* mem = memory(adr_idx); 1382 Node* ld; 1383 if (require_atomic_access && bt == T_LONG) { 1384 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1385 } else if (require_atomic_access && bt == T_DOUBLE) { 1386 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1387 } else { 1388 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1389 } 1390 ld = _gvn.transform(ld); 1391 if (bt == T_VALUETYPE) { 1392 // Loading a non-flattened value type from memory requires a null check. 1393 ld = ValueTypeNode::make(this, ld, true /* null check */); 1394 } else if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1395 // Improve graph before escape analysis and boxing elimination. 1396 record_for_igvn(ld); 1397 } 1398 return ld; 1399 } 1400 1401 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1402 int adr_idx, 1403 MemNode::MemOrd mo, 1404 bool require_atomic_access, 1405 bool unaligned, 1406 bool mismatched) { 1407 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1408 const TypePtr* adr_type = NULL; 1409 debug_only(adr_type = C->get_adr_type(adr_idx)); 1410 Node *mem = memory(adr_idx); 1411 Node* st; 1412 if (require_atomic_access && bt == T_LONG) { 1413 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1414 } else if (require_atomic_access && bt == T_DOUBLE) { 1415 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1416 } else { 1417 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1418 } 1419 if (unaligned) { 1420 st->as_Store()->set_unaligned_access(); 1421 } 1422 if (mismatched) { 1423 st->as_Store()->set_mismatched_access(); 1424 } 1425 st = _gvn.transform(st); 1426 set_memory(st, adr_idx); 1427 // Back-to-back stores can only remove intermediate store with DU info 1428 // so push on worklist for optimizer. 1429 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1430 record_for_igvn(st); 1431 1432 return st; 1433 } 1434 1435 1436 void GraphKit::pre_barrier(bool do_load, 1437 Node* ctl, 1438 Node* obj, 1439 Node* adr, 1440 uint adr_idx, 1441 Node* val, 1442 const TypeOopPtr* val_type, 1443 Node* pre_val, 1444 BasicType bt) { 1445 1446 BarrierSet* bs = Universe::heap()->barrier_set(); 1447 set_control(ctl); 1448 switch (bs->kind()) { 1449 case BarrierSet::G1SATBCTLogging: 1450 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1451 break; 1452 1453 case BarrierSet::CardTableForRS: 1454 case BarrierSet::CardTableExtension: 1455 case BarrierSet::ModRef: 1456 break; 1457 1458 default : 1459 ShouldNotReachHere(); 1460 1461 } 1462 } 1463 1464 bool GraphKit::can_move_pre_barrier() const { 1465 BarrierSet* bs = Universe::heap()->barrier_set(); 1466 switch (bs->kind()) { 1467 case BarrierSet::G1SATBCTLogging: 1468 return true; // Can move it if no safepoint 1469 1470 case BarrierSet::CardTableForRS: 1471 case BarrierSet::CardTableExtension: 1472 case BarrierSet::ModRef: 1473 return true; // There is no pre-barrier 1474 1475 default : 1476 ShouldNotReachHere(); 1477 } 1478 return false; 1479 } 1480 1481 void GraphKit::post_barrier(Node* ctl, 1482 Node* store, 1483 Node* obj, 1484 Node* adr, 1485 uint adr_idx, 1486 Node* val, 1487 BasicType bt, 1488 bool use_precise) { 1489 BarrierSet* bs = Universe::heap()->barrier_set(); 1490 set_control(ctl); 1491 switch (bs->kind()) { 1492 case BarrierSet::G1SATBCTLogging: 1493 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1494 break; 1495 1496 case BarrierSet::CardTableForRS: 1497 case BarrierSet::CardTableExtension: 1498 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1499 break; 1500 1501 case BarrierSet::ModRef: 1502 break; 1503 1504 default : 1505 ShouldNotReachHere(); 1506 1507 } 1508 } 1509 1510 Node* GraphKit::store_oop(Node* ctl, 1511 Node* obj, 1512 Node* adr, 1513 const TypePtr* adr_type, 1514 Node* val, 1515 const TypeOopPtr* val_type, 1516 BasicType bt, 1517 bool use_precise, 1518 MemNode::MemOrd mo, 1519 bool mismatched) { 1520 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1521 // could be delayed during Parse (for example, in adjust_map_after_if()). 1522 // Execute transformation here to avoid barrier generation in such case. 1523 if (_gvn.type(val) == TypePtr::NULL_PTR) 1524 val = _gvn.makecon(TypePtr::NULL_PTR); 1525 1526 set_control(ctl); 1527 if (stopped()) return top(); // Dead path ? 1528 1529 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity"); 1530 assert(val != NULL, "not dead path"); 1531 uint adr_idx = C->get_alias_index(adr_type); 1532 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1533 1534 if (val->is_ValueType()) { 1535 // Allocate value type and get oop 1536 val = val->as_ValueType()->allocate(this)->get_oop(); 1537 } 1538 1539 pre_barrier(true /* do_load */, 1540 control(), obj, adr, adr_idx, val, val_type, 1541 NULL /* pre_val */, 1542 bt); 1543 1544 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1545 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1546 return store; 1547 } 1548 1549 // Could be an array or object we don't know at compile time (unsafe ref.) 1550 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1551 Node* obj, // containing obj 1552 Node* adr, // actual adress to store val at 1553 const TypePtr* adr_type, 1554 Node* val, 1555 BasicType bt, 1556 MemNode::MemOrd mo, 1557 bool mismatched) { 1558 Compile::AliasType* at = C->alias_type(adr_type); 1559 const TypeOopPtr* val_type = NULL; 1560 if (adr_type->isa_instptr()) { 1561 if (at->field() != NULL) { 1562 // known field. This code is a copy of the do_put_xxx logic. 1563 ciField* field = at->field(); 1564 if (!field->type()->is_loaded()) { 1565 val_type = TypeInstPtr::BOTTOM; 1566 } else { 1567 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1568 } 1569 } 1570 } else if (adr_type->isa_aryptr()) { 1571 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1572 } 1573 if (val_type == NULL) { 1574 val_type = TypeInstPtr::BOTTOM; 1575 } 1576 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1577 } 1578 1579 1580 //-------------------------array_element_address------------------------- 1581 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1582 const TypeInt* sizetype, Node* ctrl) { 1583 uint shift = exact_log2(type2aelembytes(elembt)); 1584 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1585 if (arytype_klass->is_value_array_klass()) { 1586 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass(); 1587 shift = vak->log2_element_size(); 1588 } 1589 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1590 1591 // short-circuit a common case (saves lots of confusing waste motion) 1592 jint idx_con = find_int_con(idx, -1); 1593 if (idx_con >= 0) { 1594 intptr_t offset = header + ((intptr_t)idx_con << shift); 1595 return basic_plus_adr(ary, offset); 1596 } 1597 1598 // must be correct type for alignment purposes 1599 Node* base = basic_plus_adr(ary, header); 1600 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1601 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1602 return basic_plus_adr(ary, base, scale); 1603 } 1604 1605 //-------------------------load_array_element------------------------- 1606 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1607 const Type* elemtype = arytype->elem(); 1608 BasicType elembt = elemtype->array_element_basic_type(); 1609 assert(elembt != T_VALUETYPE, "value types are not supported by this method"); 1610 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1611 if (elembt == T_NARROWOOP) { 1612 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1613 } 1614 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1615 return ld; 1616 } 1617 1618 //-------------------------set_arguments_for_java_call------------------------- 1619 // Arguments (pre-popped from the stack) are taken from the JVMS. 1620 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1621 // Add the call arguments: 1622 const TypeTuple* domain = call->tf()->domain_sig(); 1623 uint nargs = domain->cnt(); 1624 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1625 Node* arg = argument(i-TypeFunc::Parms); 1626 if (ValueTypePassFieldsAsArgs) { 1627 if (arg->is_ValueType()) { 1628 ValueTypeNode* vt = arg->as_ValueType(); 1629 if (!domain->field_at(i)->is_valuetypeptr()->is__Value()) { 1630 // We don't pass value type arguments by reference but instead 1631 // pass each field of the value type 1632 idx += vt->pass_fields(call, idx, *this); 1633 // If a value type argument is passed as fields, attach the Method* to the call site 1634 // to be able to access the extended signature later via attached_method_before_pc(). 1635 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1636 call->set_override_symbolic_info(true); 1637 } else { 1638 arg = arg->as_ValueType()->allocate(this)->get_oop(); 1639 call->init_req(idx, arg); 1640 idx++; 1641 } 1642 } else { 1643 call->init_req(idx, arg); 1644 idx++; 1645 } 1646 } else { 1647 if (arg->is_ValueType()) { 1648 // Pass value type argument via oop to callee 1649 arg = arg->as_ValueType()->allocate(this)->get_oop(); 1650 } 1651 call->init_req(i, arg); 1652 } 1653 } 1654 } 1655 1656 //---------------------------set_edges_for_java_call--------------------------- 1657 // Connect a newly created call into the current JVMS. 1658 // A return value node (if any) is returned from set_edges_for_java_call. 1659 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1660 1661 // Add the predefined inputs: 1662 call->init_req( TypeFunc::Control, control() ); 1663 call->init_req( TypeFunc::I_O , i_o() ); 1664 call->init_req( TypeFunc::Memory , reset_memory() ); 1665 call->init_req( TypeFunc::FramePtr, frameptr() ); 1666 call->init_req( TypeFunc::ReturnAdr, top() ); 1667 1668 add_safepoint_edges(call, must_throw); 1669 1670 Node* xcall = _gvn.transform(call); 1671 1672 if (xcall == top()) { 1673 set_control(top()); 1674 return; 1675 } 1676 assert(xcall == call, "call identity is stable"); 1677 1678 // Re-use the current map to produce the result. 1679 1680 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1681 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1682 set_all_memory_call(xcall, separate_io_proj); 1683 1684 //return xcall; // no need, caller already has it 1685 } 1686 1687 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1688 if (stopped()) return top(); // maybe the call folded up? 1689 1690 // Note: Since any out-of-line call can produce an exception, 1691 // we always insert an I_O projection from the call into the result. 1692 1693 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1694 1695 if (separate_io_proj) { 1696 // The caller requested separate projections be used by the fall 1697 // through and exceptional paths, so replace the projections for 1698 // the fall through path. 1699 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1700 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1701 } 1702 1703 // Capture the return value, if any. 1704 Node* ret; 1705 if (call->method() == NULL || 1706 call->method()->return_type()->basic_type() == T_VOID) { 1707 ret = top(); 1708 } else { 1709 if (!call->tf()->returns_value_type_as_fields()) { 1710 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1711 } else { 1712 // Return of multiple values (value type fields): we create a 1713 // ValueType node, each field is a projection from the call. 1714 const TypeTuple* range_sig = call->tf()->range_sig(); 1715 const Type* t = range_sig->field_at(TypeFunc::Parms); 1716 assert(t->isa_valuetypeptr(), "only value types for multiple return values"); 1717 ciValueKlass* vk = t->is_valuetypeptr()->value_type()->value_klass(); 1718 Node* ctl = control(); 1719 ret = ValueTypeNode::make(_gvn, ctl, merged_memory(), call, vk, TypeFunc::Parms+1, false); 1720 set_control(ctl); 1721 } 1722 } 1723 1724 return ret; 1725 } 1726 1727 //--------------------set_predefined_input_for_runtime_call-------------------- 1728 // Reading and setting the memory state is way conservative here. 1729 // The real problem is that I am not doing real Type analysis on memory, 1730 // so I cannot distinguish card mark stores from other stores. Across a GC 1731 // point the Store Barrier and the card mark memory has to agree. I cannot 1732 // have a card mark store and its barrier split across the GC point from 1733 // either above or below. Here I get that to happen by reading ALL of memory. 1734 // A better answer would be to separate out card marks from other memory. 1735 // For now, return the input memory state, so that it can be reused 1736 // after the call, if this call has restricted memory effects. 1737 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1738 // Set fixed predefined input arguments 1739 Node* memory = reset_memory(); 1740 call->init_req( TypeFunc::Control, control() ); 1741 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1742 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1743 call->init_req( TypeFunc::FramePtr, frameptr() ); 1744 call->init_req( TypeFunc::ReturnAdr, top() ); 1745 return memory; 1746 } 1747 1748 //-------------------set_predefined_output_for_runtime_call-------------------- 1749 // Set control and memory (not i_o) from the call. 1750 // If keep_mem is not NULL, use it for the output state, 1751 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1752 // If hook_mem is NULL, this call produces no memory effects at all. 1753 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1754 // then only that memory slice is taken from the call. 1755 // In the last case, we must put an appropriate memory barrier before 1756 // the call, so as to create the correct anti-dependencies on loads 1757 // preceding the call. 1758 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1759 Node* keep_mem, 1760 const TypePtr* hook_mem) { 1761 // no i/o 1762 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1763 if (keep_mem) { 1764 // First clone the existing memory state 1765 set_all_memory(keep_mem); 1766 if (hook_mem != NULL) { 1767 // Make memory for the call 1768 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1769 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1770 // We also use hook_mem to extract specific effects from arraycopy stubs. 1771 set_memory(mem, hook_mem); 1772 } 1773 // ...else the call has NO memory effects. 1774 1775 // Make sure the call advertises its memory effects precisely. 1776 // This lets us build accurate anti-dependences in gcm.cpp. 1777 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1778 "call node must be constructed correctly"); 1779 } else { 1780 assert(hook_mem == NULL, ""); 1781 // This is not a "slow path" call; all memory comes from the call. 1782 set_all_memory_call(call); 1783 } 1784 } 1785 1786 1787 // Replace the call with the current state of the kit. 1788 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1789 JVMState* ejvms = NULL; 1790 if (has_exceptions()) { 1791 ejvms = transfer_exceptions_into_jvms(); 1792 } 1793 1794 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1795 ReplacedNodes replaced_nodes_exception; 1796 Node* ex_ctl = top(); 1797 1798 SafePointNode* final_state = stop(); 1799 1800 // Find all the needed outputs of this call 1801 CallProjections callprojs; 1802 call->extract_projections(&callprojs, true); 1803 1804 Node* init_mem = call->in(TypeFunc::Memory); 1805 Node* final_mem = final_state->in(TypeFunc::Memory); 1806 Node* final_ctl = final_state->in(TypeFunc::Control); 1807 Node* final_io = final_state->in(TypeFunc::I_O); 1808 1809 // Replace all the old call edges with the edges from the inlining result 1810 if (callprojs.fallthrough_catchproj != NULL) { 1811 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1812 } 1813 if (callprojs.fallthrough_memproj != NULL) { 1814 if (final_mem->is_MergeMem()) { 1815 // Parser's exits MergeMem was not transformed but may be optimized 1816 final_mem = _gvn.transform(final_mem); 1817 } 1818 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1819 } 1820 if (callprojs.fallthrough_ioproj != NULL) { 1821 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1822 } 1823 1824 // Replace the result with the new result if it exists and is used 1825 if (callprojs.resproj != NULL && result != NULL) { 1826 C->gvn_replace_by(callprojs.resproj, result); 1827 } 1828 1829 if (ejvms == NULL) { 1830 // No exception edges to simply kill off those paths 1831 if (callprojs.catchall_catchproj != NULL) { 1832 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1833 } 1834 if (callprojs.catchall_memproj != NULL) { 1835 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1836 } 1837 if (callprojs.catchall_ioproj != NULL) { 1838 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1839 } 1840 // Replace the old exception object with top 1841 if (callprojs.exobj != NULL) { 1842 C->gvn_replace_by(callprojs.exobj, C->top()); 1843 } 1844 } else { 1845 GraphKit ekit(ejvms); 1846 1847 // Load my combined exception state into the kit, with all phis transformed: 1848 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1849 replaced_nodes_exception = ex_map->replaced_nodes(); 1850 1851 Node* ex_oop = ekit.use_exception_state(ex_map); 1852 1853 if (callprojs.catchall_catchproj != NULL) { 1854 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1855 ex_ctl = ekit.control(); 1856 } 1857 if (callprojs.catchall_memproj != NULL) { 1858 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1859 } 1860 if (callprojs.catchall_ioproj != NULL) { 1861 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1862 } 1863 1864 // Replace the old exception object with the newly created one 1865 if (callprojs.exobj != NULL) { 1866 C->gvn_replace_by(callprojs.exobj, ex_oop); 1867 } 1868 } 1869 1870 // Disconnect the call from the graph 1871 call->disconnect_inputs(NULL, C); 1872 C->gvn_replace_by(call, C->top()); 1873 1874 // Clean up any MergeMems that feed other MergeMems since the 1875 // optimizer doesn't like that. 1876 if (final_mem->is_MergeMem()) { 1877 Node_List wl; 1878 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1879 Node* m = i.get(); 1880 if (m->is_MergeMem() && !wl.contains(m)) { 1881 wl.push(m); 1882 } 1883 } 1884 while (wl.size() > 0) { 1885 _gvn.transform(wl.pop()); 1886 } 1887 } 1888 1889 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1890 replaced_nodes.apply(C, final_ctl); 1891 } 1892 if (!ex_ctl->is_top() && do_replaced_nodes) { 1893 replaced_nodes_exception.apply(C, ex_ctl); 1894 } 1895 } 1896 1897 1898 //------------------------------increment_counter------------------------------ 1899 // for statistics: increment a VM counter by 1 1900 1901 void GraphKit::increment_counter(address counter_addr) { 1902 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1903 increment_counter(adr1); 1904 } 1905 1906 void GraphKit::increment_counter(Node* counter_addr) { 1907 int adr_type = Compile::AliasIdxRaw; 1908 Node* ctrl = control(); 1909 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1910 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1911 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1912 } 1913 1914 1915 //------------------------------uncommon_trap---------------------------------- 1916 // Bail out to the interpreter in mid-method. Implemented by calling the 1917 // uncommon_trap blob. This helper function inserts a runtime call with the 1918 // right debug info. 1919 void GraphKit::uncommon_trap(int trap_request, 1920 ciKlass* klass, const char* comment, 1921 bool must_throw, 1922 bool keep_exact_action) { 1923 if (failing()) stop(); 1924 if (stopped()) return; // trap reachable? 1925 1926 // Note: If ProfileTraps is true, and if a deopt. actually 1927 // occurs here, the runtime will make sure an MDO exists. There is 1928 // no need to call method()->ensure_method_data() at this point. 1929 1930 // Set the stack pointer to the right value for reexecution: 1931 set_sp(reexecute_sp()); 1932 1933 #ifdef ASSERT 1934 if (!must_throw) { 1935 // Make sure the stack has at least enough depth to execute 1936 // the current bytecode. 1937 int inputs, ignored_depth; 1938 if (compute_stack_effects(inputs, ignored_depth)) { 1939 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1940 Bytecodes::name(java_bc()), sp(), inputs); 1941 } 1942 } 1943 #endif 1944 1945 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1946 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1947 1948 switch (action) { 1949 case Deoptimization::Action_maybe_recompile: 1950 case Deoptimization::Action_reinterpret: 1951 // Temporary fix for 6529811 to allow virtual calls to be sure they 1952 // get the chance to go from mono->bi->mega 1953 if (!keep_exact_action && 1954 Deoptimization::trap_request_index(trap_request) < 0 && 1955 too_many_recompiles(reason)) { 1956 // This BCI is causing too many recompilations. 1957 if (C->log() != NULL) { 1958 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1959 Deoptimization::trap_reason_name(reason), 1960 Deoptimization::trap_action_name(action)); 1961 } 1962 action = Deoptimization::Action_none; 1963 trap_request = Deoptimization::make_trap_request(reason, action); 1964 } else { 1965 C->set_trap_can_recompile(true); 1966 } 1967 break; 1968 case Deoptimization::Action_make_not_entrant: 1969 C->set_trap_can_recompile(true); 1970 break; 1971 case Deoptimization::Action_none: 1972 case Deoptimization::Action_make_not_compilable: 1973 break; 1974 default: 1975 #ifdef ASSERT 1976 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 1977 #endif 1978 break; 1979 } 1980 1981 if (TraceOptoParse) { 1982 char buf[100]; 1983 tty->print_cr("Uncommon trap %s at bci:%d", 1984 Deoptimization::format_trap_request(buf, sizeof(buf), 1985 trap_request), bci()); 1986 } 1987 1988 CompileLog* log = C->log(); 1989 if (log != NULL) { 1990 int kid = (klass == NULL)? -1: log->identify(klass); 1991 log->begin_elem("uncommon_trap bci='%d'", bci()); 1992 char buf[100]; 1993 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1994 trap_request)); 1995 if (kid >= 0) log->print(" klass='%d'", kid); 1996 if (comment != NULL) log->print(" comment='%s'", comment); 1997 log->end_elem(); 1998 } 1999 2000 // Make sure any guarding test views this path as very unlikely 2001 Node *i0 = control()->in(0); 2002 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2003 IfNode *iff = i0->as_If(); 2004 float f = iff->_prob; // Get prob 2005 if (control()->Opcode() == Op_IfTrue) { 2006 if (f > PROB_UNLIKELY_MAG(4)) 2007 iff->_prob = PROB_MIN; 2008 } else { 2009 if (f < PROB_LIKELY_MAG(4)) 2010 iff->_prob = PROB_MAX; 2011 } 2012 } 2013 2014 // Clear out dead values from the debug info. 2015 kill_dead_locals(); 2016 2017 // Now insert the uncommon trap subroutine call 2018 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2019 const TypePtr* no_memory_effects = NULL; 2020 // Pass the index of the class to be loaded 2021 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2022 (must_throw ? RC_MUST_THROW : 0), 2023 OptoRuntime::uncommon_trap_Type(), 2024 call_addr, "uncommon_trap", no_memory_effects, 2025 intcon(trap_request)); 2026 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2027 "must extract request correctly from the graph"); 2028 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2029 2030 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2031 // The debug info is the only real input to this call. 2032 2033 // Halt-and-catch fire here. The above call should never return! 2034 HaltNode* halt = new HaltNode(control(), frameptr()); 2035 _gvn.set_type_bottom(halt); 2036 root()->add_req(halt); 2037 2038 stop_and_kill_map(); 2039 } 2040 2041 2042 //--------------------------just_allocated_object------------------------------ 2043 // Report the object that was just allocated. 2044 // It must be the case that there are no intervening safepoints. 2045 // We use this to determine if an object is so "fresh" that 2046 // it does not require card marks. 2047 Node* GraphKit::just_allocated_object(Node* current_control) { 2048 if (C->recent_alloc_ctl() == current_control) 2049 return C->recent_alloc_obj(); 2050 return NULL; 2051 } 2052 2053 2054 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2055 // (Note: TypeFunc::make has a cache that makes this fast.) 2056 const TypeFunc* tf = TypeFunc::make(dest_method); 2057 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2058 for (int j = 0; j < nargs; j++) { 2059 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2060 if( targ->basic_type() == T_DOUBLE ) { 2061 // If any parameters are doubles, they must be rounded before 2062 // the call, dstore_rounding does gvn.transform 2063 Node *arg = argument(j); 2064 arg = dstore_rounding(arg); 2065 set_argument(j, arg); 2066 } 2067 } 2068 } 2069 2070 /** 2071 * Record profiling data exact_kls for Node n with the type system so 2072 * that it can propagate it (speculation) 2073 * 2074 * @param n node that the type applies to 2075 * @param exact_kls type from profiling 2076 * @param maybe_null did profiling see null? 2077 * 2078 * @return node with improved type 2079 */ 2080 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2081 const Type* current_type = _gvn.type(n); 2082 assert(UseTypeSpeculation, "type speculation must be on"); 2083 2084 const TypePtr* speculative = current_type->speculative(); 2085 2086 // Should the klass from the profile be recorded in the speculative type? 2087 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2088 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2089 const TypeOopPtr* xtype = tklass->as_instance_type(); 2090 assert(xtype->klass_is_exact(), "Should be exact"); 2091 // Any reason to believe n is not null (from this profiling or a previous one)? 2092 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2093 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2094 // record the new speculative type's depth 2095 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2096 speculative = speculative->with_inline_depth(jvms()->depth()); 2097 } else if (current_type->would_improve_ptr(ptr_kind)) { 2098 // Profiling report that null was never seen so we can change the 2099 // speculative type to non null ptr. 2100 if (ptr_kind == ProfileAlwaysNull) { 2101 speculative = TypePtr::NULL_PTR; 2102 } else { 2103 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2104 const TypePtr* ptr = TypePtr::NOTNULL; 2105 if (speculative != NULL) { 2106 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2107 } else { 2108 speculative = ptr; 2109 } 2110 } 2111 } 2112 2113 if (speculative != current_type->speculative()) { 2114 // Build a type with a speculative type (what we think we know 2115 // about the type but will need a guard when we use it) 2116 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2117 // We're changing the type, we need a new CheckCast node to carry 2118 // the new type. The new type depends on the control: what 2119 // profiling tells us is only valid from here as far as we can 2120 // tell. 2121 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2122 cast = _gvn.transform(cast); 2123 replace_in_map(n, cast); 2124 n = cast; 2125 } 2126 2127 return n; 2128 } 2129 2130 /** 2131 * Record profiling data from receiver profiling at an invoke with the 2132 * type system so that it can propagate it (speculation) 2133 * 2134 * @param n receiver node 2135 * 2136 * @return node with improved type 2137 */ 2138 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2139 if (!UseTypeSpeculation) { 2140 return n; 2141 } 2142 ciKlass* exact_kls = profile_has_unique_klass(); 2143 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2144 if ((java_bc() == Bytecodes::_checkcast || 2145 java_bc() == Bytecodes::_instanceof || 2146 java_bc() == Bytecodes::_aastore) && 2147 method()->method_data()->is_mature()) { 2148 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2149 if (data != NULL) { 2150 if (!data->as_BitData()->null_seen()) { 2151 ptr_kind = ProfileNeverNull; 2152 } else { 2153 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2154 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2155 uint i = 0; 2156 for (; i < call->row_limit(); i++) { 2157 ciKlass* receiver = call->receiver(i); 2158 if (receiver != NULL) { 2159 break; 2160 } 2161 } 2162 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2163 } 2164 } 2165 } 2166 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2167 } 2168 2169 /** 2170 * Record profiling data from argument profiling at an invoke with the 2171 * type system so that it can propagate it (speculation) 2172 * 2173 * @param dest_method target method for the call 2174 * @param bc what invoke bytecode is this? 2175 */ 2176 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2177 if (!UseTypeSpeculation) { 2178 return; 2179 } 2180 const TypeFunc* tf = TypeFunc::make(dest_method); 2181 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2182 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2183 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2184 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2185 if (targ->isa_oopptr() && !targ->isa_valuetypeptr()) { 2186 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2187 ciKlass* better_type = NULL; 2188 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2189 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2190 } 2191 i++; 2192 } 2193 } 2194 } 2195 2196 /** 2197 * Record profiling data from parameter profiling at an invoke with 2198 * the type system so that it can propagate it (speculation) 2199 */ 2200 void GraphKit::record_profiled_parameters_for_speculation() { 2201 if (!UseTypeSpeculation) { 2202 return; 2203 } 2204 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2205 if (_gvn.type(local(i))->isa_oopptr()) { 2206 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2207 ciKlass* better_type = NULL; 2208 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2209 record_profile_for_speculation(local(i), better_type, ptr_kind); 2210 } 2211 j++; 2212 } 2213 } 2214 } 2215 2216 /** 2217 * Record profiling data from return value profiling at an invoke with 2218 * the type system so that it can propagate it (speculation) 2219 */ 2220 void GraphKit::record_profiled_return_for_speculation() { 2221 if (!UseTypeSpeculation) { 2222 return; 2223 } 2224 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2225 ciKlass* better_type = NULL; 2226 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2227 // If profiling reports a single type for the return value, 2228 // feed it to the type system so it can propagate it as a 2229 // speculative type 2230 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2231 } 2232 } 2233 2234 void GraphKit::round_double_result(ciMethod* dest_method) { 2235 // A non-strict method may return a double value which has an extended 2236 // exponent, but this must not be visible in a caller which is 'strict' 2237 // If a strict caller invokes a non-strict callee, round a double result 2238 2239 BasicType result_type = dest_method->return_type()->basic_type(); 2240 assert( method() != NULL, "must have caller context"); 2241 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2242 // Destination method's return value is on top of stack 2243 // dstore_rounding() does gvn.transform 2244 Node *result = pop_pair(); 2245 result = dstore_rounding(result); 2246 push_pair(result); 2247 } 2248 } 2249 2250 // rounding for strict float precision conformance 2251 Node* GraphKit::precision_rounding(Node* n) { 2252 return UseStrictFP && _method->flags().is_strict() 2253 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2254 ? _gvn.transform( new RoundFloatNode(0, n) ) 2255 : n; 2256 } 2257 2258 // rounding for strict double precision conformance 2259 Node* GraphKit::dprecision_rounding(Node *n) { 2260 return UseStrictFP && _method->flags().is_strict() 2261 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2262 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2263 : n; 2264 } 2265 2266 // rounding for non-strict double stores 2267 Node* GraphKit::dstore_rounding(Node* n) { 2268 return Matcher::strict_fp_requires_explicit_rounding 2269 && UseSSE <= 1 2270 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2271 : n; 2272 } 2273 2274 //============================================================================= 2275 // Generate a fast path/slow path idiom. Graph looks like: 2276 // [foo] indicates that 'foo' is a parameter 2277 // 2278 // [in] NULL 2279 // \ / 2280 // CmpP 2281 // Bool ne 2282 // If 2283 // / \ 2284 // True False-<2> 2285 // / | 2286 // / cast_not_null 2287 // Load | | ^ 2288 // [fast_test] | | 2289 // gvn to opt_test | | 2290 // / \ | <1> 2291 // True False | 2292 // | \\ | 2293 // [slow_call] \[fast_result] 2294 // Ctl Val \ \ 2295 // | \ \ 2296 // Catch <1> \ \ 2297 // / \ ^ \ \ 2298 // Ex No_Ex | \ \ 2299 // | \ \ | \ <2> \ 2300 // ... \ [slow_res] | | \ [null_result] 2301 // \ \--+--+--- | | 2302 // \ | / \ | / 2303 // --------Region Phi 2304 // 2305 //============================================================================= 2306 // Code is structured as a series of driver functions all called 'do_XXX' that 2307 // call a set of helper functions. Helper functions first, then drivers. 2308 2309 //------------------------------null_check_oop--------------------------------- 2310 // Null check oop. Set null-path control into Region in slot 3. 2311 // Make a cast-not-nullness use the other not-null control. Return cast. 2312 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2313 bool never_see_null, 2314 bool safe_for_replace, 2315 bool speculative) { 2316 // Initial NULL check taken path 2317 (*null_control) = top(); 2318 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2319 2320 // Generate uncommon_trap: 2321 if (never_see_null && (*null_control) != top()) { 2322 // If we see an unexpected null at a check-cast we record it and force a 2323 // recompile; the offending check-cast will be compiled to handle NULLs. 2324 // If we see more than one offending BCI, then all checkcasts in the 2325 // method will be compiled to handle NULLs. 2326 PreserveJVMState pjvms(this); 2327 set_control(*null_control); 2328 replace_in_map(value, null()); 2329 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2330 uncommon_trap(reason, 2331 Deoptimization::Action_make_not_entrant); 2332 (*null_control) = top(); // NULL path is dead 2333 } 2334 if ((*null_control) == top() && safe_for_replace) { 2335 replace_in_map(value, cast); 2336 } 2337 2338 // Cast away null-ness on the result 2339 return cast; 2340 } 2341 2342 //------------------------------opt_iff---------------------------------------- 2343 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2344 // Return slow-path control. 2345 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2346 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2347 2348 // Fast path taken; set region slot 2 2349 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2350 region->init_req(2,fast_taken); // Capture fast-control 2351 2352 // Fast path not-taken, i.e. slow path 2353 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2354 return slow_taken; 2355 } 2356 2357 //-----------------------------make_runtime_call------------------------------- 2358 Node* GraphKit::make_runtime_call(int flags, 2359 const TypeFunc* call_type, address call_addr, 2360 const char* call_name, 2361 const TypePtr* adr_type, 2362 // The following parms are all optional. 2363 // The first NULL ends the list. 2364 Node* parm0, Node* parm1, 2365 Node* parm2, Node* parm3, 2366 Node* parm4, Node* parm5, 2367 Node* parm6, Node* parm7) { 2368 // Slow-path call 2369 bool is_leaf = !(flags & RC_NO_LEAF); 2370 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2371 if (call_name == NULL) { 2372 assert(!is_leaf, "must supply name for leaf"); 2373 call_name = OptoRuntime::stub_name(call_addr); 2374 } 2375 CallNode* call; 2376 if (!is_leaf) { 2377 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2378 bci(), adr_type); 2379 } else if (flags & RC_NO_FP) { 2380 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2381 } else { 2382 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2383 } 2384 2385 // The following is similar to set_edges_for_java_call, 2386 // except that the memory effects of the call are restricted to AliasIdxRaw. 2387 2388 // Slow path call has no side-effects, uses few values 2389 bool wide_in = !(flags & RC_NARROW_MEM); 2390 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2391 2392 Node* prev_mem = NULL; 2393 if (wide_in) { 2394 prev_mem = set_predefined_input_for_runtime_call(call); 2395 } else { 2396 assert(!wide_out, "narrow in => narrow out"); 2397 Node* narrow_mem = memory(adr_type); 2398 prev_mem = reset_memory(); 2399 map()->set_memory(narrow_mem); 2400 set_predefined_input_for_runtime_call(call); 2401 } 2402 2403 // Hook each parm in order. Stop looking at the first NULL. 2404 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2405 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2406 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2407 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2408 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2409 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2410 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2411 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2412 /* close each nested if ===> */ } } } } } } } } 2413 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2414 2415 if (!is_leaf) { 2416 // Non-leaves can block and take safepoints: 2417 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2418 } 2419 // Non-leaves can throw exceptions: 2420 if (has_io) { 2421 call->set_req(TypeFunc::I_O, i_o()); 2422 } 2423 2424 if (flags & RC_UNCOMMON) { 2425 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2426 // (An "if" probability corresponds roughly to an unconditional count. 2427 // Sort of.) 2428 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2429 } 2430 2431 Node* c = _gvn.transform(call); 2432 assert(c == call, "cannot disappear"); 2433 2434 if (wide_out) { 2435 // Slow path call has full side-effects. 2436 set_predefined_output_for_runtime_call(call); 2437 } else { 2438 // Slow path call has few side-effects, and/or sets few values. 2439 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2440 } 2441 2442 if (has_io) { 2443 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2444 } 2445 return call; 2446 2447 } 2448 2449 //------------------------------merge_memory----------------------------------- 2450 // Merge memory from one path into the current memory state. 2451 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2452 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2453 Node* old_slice = mms.force_memory(); 2454 Node* new_slice = mms.memory2(); 2455 if (old_slice != new_slice) { 2456 PhiNode* phi; 2457 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2458 if (mms.is_empty()) { 2459 // clone base memory Phi's inputs for this memory slice 2460 assert(old_slice == mms.base_memory(), "sanity"); 2461 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2462 _gvn.set_type(phi, Type::MEMORY); 2463 for (uint i = 1; i < phi->req(); i++) { 2464 phi->init_req(i, old_slice->in(i)); 2465 } 2466 } else { 2467 phi = old_slice->as_Phi(); // Phi was generated already 2468 } 2469 } else { 2470 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2471 _gvn.set_type(phi, Type::MEMORY); 2472 } 2473 phi->set_req(new_path, new_slice); 2474 mms.set_memory(phi); 2475 } 2476 } 2477 } 2478 2479 //------------------------------make_slow_call_ex------------------------------ 2480 // Make the exception handler hookups for the slow call 2481 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2482 if (stopped()) return; 2483 2484 // Make a catch node with just two handlers: fall-through and catch-all 2485 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2486 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2487 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2488 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2489 2490 { PreserveJVMState pjvms(this); 2491 set_control(excp); 2492 set_i_o(i_o); 2493 2494 if (excp != top()) { 2495 if (deoptimize) { 2496 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2497 uncommon_trap(Deoptimization::Reason_unhandled, 2498 Deoptimization::Action_none); 2499 } else { 2500 // Create an exception state also. 2501 // Use an exact type if the caller has specified a specific exception. 2502 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2503 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2504 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2505 } 2506 } 2507 } 2508 2509 // Get the no-exception control from the CatchNode. 2510 set_control(norm); 2511 } 2512 2513 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2514 Node* cmp = NULL; 2515 switch(bt) { 2516 case T_INT: cmp = new CmpINode(in1, in2); break; 2517 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2518 default: fatal("unexpected comparison type %s", type2name(bt)); 2519 } 2520 gvn->transform(cmp); 2521 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2522 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2523 gvn->transform(iff); 2524 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2525 return iff; 2526 } 2527 2528 2529 //-------------------------------gen_subtype_check----------------------------- 2530 // Generate a subtyping check. Takes as input the subtype and supertype. 2531 // Returns 2 values: sets the default control() to the true path and returns 2532 // the false path. Only reads invariant memory; sets no (visible) memory. 2533 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2534 // but that's not exposed to the optimizer. This call also doesn't take in an 2535 // Object; if you wish to check an Object you need to load the Object's class 2536 // prior to coming here. 2537 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2538 Compile* C = gvn->C; 2539 2540 if ((*ctrl)->is_top()) { 2541 return C->top(); 2542 } 2543 2544 // Fast check for identical types, perhaps identical constants. 2545 // The types can even be identical non-constants, in cases 2546 // involving Array.newInstance, Object.clone, etc. 2547 if (subklass == superklass) 2548 return C->top(); // false path is dead; no test needed. 2549 2550 if (gvn->type(superklass)->singleton()) { 2551 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2552 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2553 2554 // In the common case of an exact superklass, try to fold up the 2555 // test before generating code. You may ask, why not just generate 2556 // the code and then let it fold up? The answer is that the generated 2557 // code will necessarily include null checks, which do not always 2558 // completely fold away. If they are also needless, then they turn 2559 // into a performance loss. Example: 2560 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2561 // Here, the type of 'fa' is often exact, so the store check 2562 // of fa[1]=x will fold up, without testing the nullness of x. 2563 switch (C->static_subtype_check(superk, subk)) { 2564 case Compile::SSC_always_false: 2565 { 2566 Node* always_fail = *ctrl; 2567 *ctrl = gvn->C->top(); 2568 return always_fail; 2569 } 2570 case Compile::SSC_always_true: 2571 return C->top(); 2572 case Compile::SSC_easy_test: 2573 { 2574 // Just do a direct pointer compare and be done. 2575 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2576 *ctrl = gvn->transform(new IfTrueNode(iff)); 2577 return gvn->transform(new IfFalseNode(iff)); 2578 } 2579 case Compile::SSC_full_test: 2580 break; 2581 default: 2582 ShouldNotReachHere(); 2583 } 2584 } 2585 2586 // %%% Possible further optimization: Even if the superklass is not exact, 2587 // if the subklass is the unique subtype of the superklass, the check 2588 // will always succeed. We could leave a dependency behind to ensure this. 2589 2590 // First load the super-klass's check-offset 2591 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2592 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2593 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2594 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2595 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2596 2597 // Load from the sub-klass's super-class display list, or a 1-word cache of 2598 // the secondary superclass list, or a failing value with a sentinel offset 2599 // if the super-klass is an interface or exceptionally deep in the Java 2600 // hierarchy and we have to scan the secondary superclass list the hard way. 2601 // Worst-case type is a little odd: NULL is allowed as a result (usually 2602 // klass loads can never produce a NULL). 2603 Node *chk_off_X = chk_off; 2604 #ifdef _LP64 2605 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2606 #endif 2607 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2608 // For some types like interfaces the following loadKlass is from a 1-word 2609 // cache which is mutable so can't use immutable memory. Other 2610 // types load from the super-class display table which is immutable. 2611 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2612 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2613 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM)); 2614 2615 // Compile speed common case: ARE a subtype and we canNOT fail 2616 if( superklass == nkls ) 2617 return C->top(); // false path is dead; no test needed. 2618 2619 // See if we get an immediate positive hit. Happens roughly 83% of the 2620 // time. Test to see if the value loaded just previously from the subklass 2621 // is exactly the superklass. 2622 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2623 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2624 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2625 2626 // Compile speed common case: Check for being deterministic right now. If 2627 // chk_off is a constant and not equal to cacheoff then we are NOT a 2628 // subklass. In this case we need exactly the 1 test above and we can 2629 // return those results immediately. 2630 if (!might_be_cache) { 2631 Node* not_subtype_ctrl = *ctrl; 2632 *ctrl = iftrue1; // We need exactly the 1 test above 2633 return not_subtype_ctrl; 2634 } 2635 2636 // Gather the various success & failures here 2637 RegionNode *r_ok_subtype = new RegionNode(4); 2638 gvn->record_for_igvn(r_ok_subtype); 2639 RegionNode *r_not_subtype = new RegionNode(3); 2640 gvn->record_for_igvn(r_not_subtype); 2641 2642 r_ok_subtype->init_req(1, iftrue1); 2643 2644 // Check for immediate negative hit. Happens roughly 11% of the time (which 2645 // is roughly 63% of the remaining cases). Test to see if the loaded 2646 // check-offset points into the subklass display list or the 1-element 2647 // cache. If it points to the display (and NOT the cache) and the display 2648 // missed then it's not a subtype. 2649 Node *cacheoff = gvn->intcon(cacheoff_con); 2650 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2651 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2652 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2653 2654 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2655 // No performance impact (too rare) but allows sharing of secondary arrays 2656 // which has some footprint reduction. 2657 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2658 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2659 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2660 2661 // -- Roads not taken here: -- 2662 // We could also have chosen to perform the self-check at the beginning 2663 // of this code sequence, as the assembler does. This would not pay off 2664 // the same way, since the optimizer, unlike the assembler, can perform 2665 // static type analysis to fold away many successful self-checks. 2666 // Non-foldable self checks work better here in second position, because 2667 // the initial primary superclass check subsumes a self-check for most 2668 // types. An exception would be a secondary type like array-of-interface, 2669 // which does not appear in its own primary supertype display. 2670 // Finally, we could have chosen to move the self-check into the 2671 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2672 // dependent manner. But it is worthwhile to have the check here, 2673 // where it can be perhaps be optimized. The cost in code space is 2674 // small (register compare, branch). 2675 2676 // Now do a linear scan of the secondary super-klass array. Again, no real 2677 // performance impact (too rare) but it's gotta be done. 2678 // Since the code is rarely used, there is no penalty for moving it 2679 // out of line, and it can only improve I-cache density. 2680 // The decision to inline or out-of-line this final check is platform 2681 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2682 Node* psc = gvn->transform( 2683 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2684 2685 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2686 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2687 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2688 2689 // Return false path; set default control to true path. 2690 *ctrl = gvn->transform(r_ok_subtype); 2691 return gvn->transform(r_not_subtype); 2692 } 2693 2694 // Profile-driven exact type check: 2695 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2696 float prob, 2697 Node* *casted_receiver) { 2698 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2699 Node* recv_klass = load_object_klass(receiver); 2700 Node* fail = type_check(recv_klass, tklass, prob); 2701 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2702 assert(recv_xtype->klass_is_exact(), ""); 2703 2704 // Subsume downstream occurrences of receiver with a cast to 2705 // recv_xtype, since now we know what the type will be. 2706 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2707 (*casted_receiver) = _gvn.transform(cast); 2708 // (User must make the replace_in_map call.) 2709 2710 return fail; 2711 } 2712 2713 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 2714 float prob) { 2715 //const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2716 Node* want_klass = makecon(tklass); 2717 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass)); 2718 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2719 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2720 set_control( _gvn.transform( new IfTrueNode (iff))); 2721 Node* fail = _gvn.transform( new IfFalseNode(iff)); 2722 return fail; 2723 } 2724 2725 2726 //------------------------------seems_never_null------------------------------- 2727 // Use null_seen information if it is available from the profile. 2728 // If we see an unexpected null at a type check we record it and force a 2729 // recompile; the offending check will be recompiled to handle NULLs. 2730 // If we see several offending BCIs, then all checks in the 2731 // method will be recompiled. 2732 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2733 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2734 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2735 if (UncommonNullCast // Cutout for this technique 2736 && obj != null() // And not the -Xcomp stupid case? 2737 && !too_many_traps(reason) 2738 ) { 2739 if (speculating) { 2740 return true; 2741 } 2742 if (data == NULL) 2743 // Edge case: no mature data. Be optimistic here. 2744 return true; 2745 // If the profile has not seen a null, assume it won't happen. 2746 assert(java_bc() == Bytecodes::_checkcast || 2747 java_bc() == Bytecodes::_instanceof || 2748 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2749 return !data->as_BitData()->null_seen(); 2750 } 2751 speculating = false; 2752 return false; 2753 } 2754 2755 //------------------------maybe_cast_profiled_receiver------------------------- 2756 // If the profile has seen exactly one type, narrow to exactly that type. 2757 // Subsequent type checks will always fold up. 2758 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2759 ciKlass* require_klass, 2760 ciKlass* spec_klass, 2761 bool safe_for_replace) { 2762 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2763 2764 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2765 2766 // Make sure we haven't already deoptimized from this tactic. 2767 if (too_many_traps(reason) || too_many_recompiles(reason)) 2768 return NULL; 2769 2770 // (No, this isn't a call, but it's enough like a virtual call 2771 // to use the same ciMethod accessor to get the profile info...) 2772 // If we have a speculative type use it instead of profiling (which 2773 // may not help us) 2774 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2775 if (exact_kls != NULL) {// no cast failures here 2776 if (require_klass == NULL || 2777 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2778 // If we narrow the type to match what the type profile sees or 2779 // the speculative type, we can then remove the rest of the 2780 // cast. 2781 // This is a win, even if the exact_kls is very specific, 2782 // because downstream operations, such as method calls, 2783 // will often benefit from the sharper type. 2784 Node* exact_obj = not_null_obj; // will get updated in place... 2785 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2786 &exact_obj); 2787 { PreserveJVMState pjvms(this); 2788 set_control(slow_ctl); 2789 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2790 } 2791 if (safe_for_replace) { 2792 replace_in_map(not_null_obj, exact_obj); 2793 } 2794 return exact_obj; 2795 } 2796 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2797 } 2798 2799 return NULL; 2800 } 2801 2802 /** 2803 * Cast obj to type and emit guard unless we had too many traps here 2804 * already 2805 * 2806 * @param obj node being casted 2807 * @param type type to cast the node to 2808 * @param not_null true if we know node cannot be null 2809 */ 2810 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2811 ciKlass* type, 2812 bool not_null) { 2813 if (stopped()) { 2814 return obj; 2815 } 2816 2817 // type == NULL if profiling tells us this object is always null 2818 if (type != NULL) { 2819 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2820 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2821 2822 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2823 !too_many_traps(class_reason) && 2824 !too_many_recompiles(class_reason)) { 2825 Node* not_null_obj = NULL; 2826 // not_null is true if we know the object is not null and 2827 // there's no need for a null check 2828 if (!not_null) { 2829 Node* null_ctl = top(); 2830 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2831 assert(null_ctl->is_top(), "no null control here"); 2832 } else { 2833 not_null_obj = obj; 2834 } 2835 2836 Node* exact_obj = not_null_obj; 2837 ciKlass* exact_kls = type; 2838 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2839 &exact_obj); 2840 { 2841 PreserveJVMState pjvms(this); 2842 set_control(slow_ctl); 2843 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2844 } 2845 replace_in_map(not_null_obj, exact_obj); 2846 obj = exact_obj; 2847 } 2848 } else { 2849 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2850 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2851 Node* exact_obj = null_assert(obj); 2852 replace_in_map(obj, exact_obj); 2853 obj = exact_obj; 2854 } 2855 } 2856 return obj; 2857 } 2858 2859 //-------------------------------gen_instanceof-------------------------------- 2860 // Generate an instance-of idiom. Used by both the instance-of bytecode 2861 // and the reflective instance-of call. 2862 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2863 kill_dead_locals(); // Benefit all the uncommon traps 2864 assert( !stopped(), "dead parse path should be checked in callers" ); 2865 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2866 "must check for not-null not-dead klass in callers"); 2867 2868 // Make the merge point 2869 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2870 RegionNode* region = new RegionNode(PATH_LIMIT); 2871 Node* phi = new PhiNode(region, TypeInt::BOOL); 2872 C->set_has_split_ifs(true); // Has chance for split-if optimization 2873 2874 ciProfileData* data = NULL; 2875 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2876 data = method()->method_data()->bci_to_data(bci()); 2877 } 2878 bool speculative_not_null = false; 2879 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2880 && seems_never_null(obj, data, speculative_not_null)); 2881 2882 // Null check; get casted pointer; set region slot 3 2883 Node* null_ctl = top(); 2884 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2885 2886 // If not_null_obj is dead, only null-path is taken 2887 if (stopped()) { // Doing instance-of on a NULL? 2888 set_control(null_ctl); 2889 return intcon(0); 2890 } 2891 region->init_req(_null_path, null_ctl); 2892 phi ->init_req(_null_path, intcon(0)); // Set null path value 2893 if (null_ctl == top()) { 2894 // Do this eagerly, so that pattern matches like is_diamond_phi 2895 // will work even during parsing. 2896 assert(_null_path == PATH_LIMIT-1, "delete last"); 2897 region->del_req(_null_path); 2898 phi ->del_req(_null_path); 2899 } 2900 2901 // Do we know the type check always succeed? 2902 bool known_statically = false; 2903 if (_gvn.type(superklass)->singleton()) { 2904 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2905 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2906 if (subk != NULL && subk->is_loaded()) { 2907 int static_res = C->static_subtype_check(superk, subk); 2908 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2909 } 2910 } 2911 2912 if (!known_statically) { 2913 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2914 // We may not have profiling here or it may not help us. If we 2915 // have a speculative type use it to perform an exact cast. 2916 ciKlass* spec_obj_type = obj_type->speculative_type(); 2917 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2918 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2919 if (stopped()) { // Profile disagrees with this path. 2920 set_control(null_ctl); // Null is the only remaining possibility. 2921 return intcon(0); 2922 } 2923 if (cast_obj != NULL) { 2924 not_null_obj = cast_obj; 2925 } 2926 } 2927 } 2928 2929 // Load the object's klass 2930 Node* obj_klass = load_object_klass(not_null_obj); 2931 2932 // Generate the subtype check 2933 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2934 2935 // Plug in the success path to the general merge in slot 1. 2936 region->init_req(_obj_path, control()); 2937 phi ->init_req(_obj_path, intcon(1)); 2938 2939 // Plug in the failing path to the general merge in slot 2. 2940 region->init_req(_fail_path, not_subtype_ctrl); 2941 phi ->init_req(_fail_path, intcon(0)); 2942 2943 // Return final merged results 2944 set_control( _gvn.transform(region) ); 2945 record_for_igvn(region); 2946 2947 // If we know the type check always succeeds then we don't use the 2948 // profiling data at this bytecode. Don't lose it, feed it to the 2949 // type system as a speculative type. 2950 if (safe_for_replace) { 2951 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 2952 replace_in_map(obj, casted_obj); 2953 } 2954 2955 return _gvn.transform(phi); 2956 } 2957 2958 //-------------------------------gen_checkcast--------------------------------- 2959 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2960 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2961 // uncommon-trap paths work. Adjust stack after this call. 2962 // If failure_control is supplied and not null, it is filled in with 2963 // the control edge for the cast failure. Otherwise, an appropriate 2964 // uncommon trap or exception is thrown. 2965 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2966 Node* *failure_control) { 2967 kill_dead_locals(); // Benefit all the uncommon traps 2968 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2969 assert(tk->is_loaded(), "must be loaded"); 2970 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2971 2972 // Fast cutout: Check the case that the cast is vacuously true. 2973 // This detects the common cases where the test will short-circuit 2974 // away completely. We do this before we perform the null check, 2975 // because if the test is going to turn into zero code, we don't 2976 // want a residual null check left around. (Causes a slowdown, 2977 // for example, in some objArray manipulations, such as a[i]=a[j].) 2978 if (tk->singleton()) { 2979 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2980 if (objtp != NULL && objtp->klass() != NULL) { 2981 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2982 case Compile::SSC_always_true: 2983 // If we know the type check always succeed then we don't use 2984 // the profiling data at this bytecode. Don't lose it, feed it 2985 // to the type system as a speculative type. 2986 return record_profiled_receiver_for_speculation(obj); 2987 case Compile::SSC_always_false: 2988 // It needs a null check because a null will *pass* the cast check. 2989 // A non-null value will always produce an exception. 2990 return null_assert(obj); 2991 } 2992 } 2993 } 2994 2995 ciProfileData* data = NULL; 2996 bool safe_for_replace = false; 2997 if (failure_control == NULL) { // use MDO in regular case only 2998 assert(java_bc() == Bytecodes::_aastore || 2999 java_bc() == Bytecodes::_checkcast, 3000 "interpreter profiles type checks only for these BCs"); 3001 data = method()->method_data()->bci_to_data(bci()); 3002 safe_for_replace = true; 3003 } 3004 3005 // Make the merge point 3006 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3007 RegionNode* region = new RegionNode(PATH_LIMIT); 3008 Node* phi = new PhiNode(region, toop); 3009 C->set_has_split_ifs(true); // Has chance for split-if optimization 3010 3011 // Use null-cast information if it is available 3012 bool speculative_not_null = false; 3013 bool never_see_null = ((failure_control == NULL) // regular case only 3014 && seems_never_null(obj, data, speculative_not_null)); 3015 3016 // Null check; get casted pointer; set region slot 3 3017 Node* null_ctl = top(); 3018 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3019 3020 // If not_null_obj is dead, only null-path is taken 3021 if (stopped()) { // Doing instance-of on a NULL? 3022 set_control(null_ctl); 3023 return null(); 3024 } 3025 region->init_req(_null_path, null_ctl); 3026 phi ->init_req(_null_path, null()); // Set null path value 3027 if (null_ctl == top()) { 3028 // Do this eagerly, so that pattern matches like is_diamond_phi 3029 // will work even during parsing. 3030 assert(_null_path == PATH_LIMIT-1, "delete last"); 3031 region->del_req(_null_path); 3032 phi ->del_req(_null_path); 3033 } 3034 3035 Node* cast_obj = NULL; 3036 if (tk->klass_is_exact()) { 3037 // The following optimization tries to statically cast the speculative type of the object 3038 // (for example obtained during profiling) to the type of the superklass and then do a 3039 // dynamic check that the type of the object is what we expect. To work correctly 3040 // for checkcast and aastore the type of superklass should be exact. 3041 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3042 // We may not have profiling here or it may not help us. If we have 3043 // a speculative type use it to perform an exact cast. 3044 ciKlass* spec_obj_type = obj_type->speculative_type(); 3045 if (spec_obj_type != NULL || data != NULL) { 3046 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3047 if (cast_obj != NULL) { 3048 if (failure_control != NULL) // failure is now impossible 3049 (*failure_control) = top(); 3050 // adjust the type of the phi to the exact klass: 3051 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3052 } 3053 } 3054 } 3055 3056 if (cast_obj == NULL) { 3057 // Load the object's klass 3058 Node* obj_klass = load_object_klass(not_null_obj); 3059 3060 // Generate the subtype check 3061 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3062 3063 // Plug in success path into the merge 3064 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3065 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3066 if (failure_control == NULL) { 3067 if (not_subtype_ctrl != top()) { // If failure is possible 3068 PreserveJVMState pjvms(this); 3069 set_control(not_subtype_ctrl); 3070 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3071 } 3072 } else { 3073 (*failure_control) = not_subtype_ctrl; 3074 } 3075 } 3076 3077 region->init_req(_obj_path, control()); 3078 phi ->init_req(_obj_path, cast_obj); 3079 3080 // A merge of NULL or Casted-NotNull obj 3081 Node* res = _gvn.transform(phi); 3082 3083 // Note I do NOT always 'replace_in_map(obj,result)' here. 3084 // if( tk->klass()->can_be_primary_super() ) 3085 // This means that if I successfully store an Object into an array-of-String 3086 // I 'forget' that the Object is really now known to be a String. I have to 3087 // do this because we don't have true union types for interfaces - if I store 3088 // a Baz into an array-of-Interface and then tell the optimizer it's an 3089 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3090 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3091 // replace_in_map( obj, res ); 3092 3093 // Return final merged results 3094 set_control( _gvn.transform(region) ); 3095 record_for_igvn(region); 3096 3097 return record_profiled_receiver_for_speculation(res); 3098 } 3099 3100 //------------------------------next_monitor----------------------------------- 3101 // What number should be given to the next monitor? 3102 int GraphKit::next_monitor() { 3103 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3104 int next = current + C->sync_stack_slots(); 3105 // Keep the toplevel high water mark current: 3106 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3107 return current; 3108 } 3109 3110 //------------------------------insert_mem_bar--------------------------------- 3111 // Memory barrier to avoid floating things around 3112 // The membar serves as a pinch point between both control and all memory slices. 3113 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3114 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3115 mb->init_req(TypeFunc::Control, control()); 3116 mb->init_req(TypeFunc::Memory, reset_memory()); 3117 Node* membar = _gvn.transform(mb); 3118 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3119 set_all_memory_call(membar); 3120 return membar; 3121 } 3122 3123 //-------------------------insert_mem_bar_volatile---------------------------- 3124 // Memory barrier to avoid floating things around 3125 // The membar serves as a pinch point between both control and memory(alias_idx). 3126 // If you want to make a pinch point on all memory slices, do not use this 3127 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3128 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3129 // When Parse::do_put_xxx updates a volatile field, it appends a series 3130 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3131 // The first membar is on the same memory slice as the field store opcode. 3132 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3133 // All the other membars (for other volatile slices, including AliasIdxBot, 3134 // which stands for all unknown volatile slices) are control-dependent 3135 // on the first membar. This prevents later volatile loads or stores 3136 // from sliding up past the just-emitted store. 3137 3138 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3139 mb->set_req(TypeFunc::Control,control()); 3140 if (alias_idx == Compile::AliasIdxBot) { 3141 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3142 } else { 3143 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3144 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3145 } 3146 Node* membar = _gvn.transform(mb); 3147 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3148 if (alias_idx == Compile::AliasIdxBot) { 3149 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3150 } else { 3151 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3152 } 3153 return membar; 3154 } 3155 3156 //------------------------------shared_lock------------------------------------ 3157 // Emit locking code. 3158 FastLockNode* GraphKit::shared_lock(Node* obj) { 3159 // bci is either a monitorenter bc or InvocationEntryBci 3160 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3161 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3162 3163 if( !GenerateSynchronizationCode ) 3164 return NULL; // Not locking things? 3165 if (stopped()) // Dead monitor? 3166 return NULL; 3167 3168 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3169 3170 // Box the stack location 3171 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3172 Node* mem = reset_memory(); 3173 3174 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3175 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3176 // Create the counters for this fast lock. 3177 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3178 } 3179 3180 // Create the rtm counters for this fast lock if needed. 3181 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3182 3183 // Add monitor to debug info for the slow path. If we block inside the 3184 // slow path and de-opt, we need the monitor hanging around 3185 map()->push_monitor( flock ); 3186 3187 const TypeFunc *tf = LockNode::lock_type(); 3188 LockNode *lock = new LockNode(C, tf); 3189 3190 lock->init_req( TypeFunc::Control, control() ); 3191 lock->init_req( TypeFunc::Memory , mem ); 3192 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3193 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3194 lock->init_req( TypeFunc::ReturnAdr, top() ); 3195 3196 lock->init_req(TypeFunc::Parms + 0, obj); 3197 lock->init_req(TypeFunc::Parms + 1, box); 3198 lock->init_req(TypeFunc::Parms + 2, flock); 3199 add_safepoint_edges(lock); 3200 3201 lock = _gvn.transform( lock )->as_Lock(); 3202 3203 // lock has no side-effects, sets few values 3204 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3205 3206 insert_mem_bar(Op_MemBarAcquireLock); 3207 3208 // Add this to the worklist so that the lock can be eliminated 3209 record_for_igvn(lock); 3210 3211 #ifndef PRODUCT 3212 if (PrintLockStatistics) { 3213 // Update the counter for this lock. Don't bother using an atomic 3214 // operation since we don't require absolute accuracy. 3215 lock->create_lock_counter(map()->jvms()); 3216 increment_counter(lock->counter()->addr()); 3217 } 3218 #endif 3219 3220 return flock; 3221 } 3222 3223 3224 //------------------------------shared_unlock---------------------------------- 3225 // Emit unlocking code. 3226 void GraphKit::shared_unlock(Node* box, Node* obj) { 3227 // bci is either a monitorenter bc or InvocationEntryBci 3228 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3229 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3230 3231 if( !GenerateSynchronizationCode ) 3232 return; 3233 if (stopped()) { // Dead monitor? 3234 map()->pop_monitor(); // Kill monitor from debug info 3235 return; 3236 } 3237 3238 // Memory barrier to avoid floating things down past the locked region 3239 insert_mem_bar(Op_MemBarReleaseLock); 3240 3241 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3242 UnlockNode *unlock = new UnlockNode(C, tf); 3243 #ifdef ASSERT 3244 unlock->set_dbg_jvms(sync_jvms()); 3245 #endif 3246 uint raw_idx = Compile::AliasIdxRaw; 3247 unlock->init_req( TypeFunc::Control, control() ); 3248 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3249 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3250 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3251 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3252 3253 unlock->init_req(TypeFunc::Parms + 0, obj); 3254 unlock->init_req(TypeFunc::Parms + 1, box); 3255 unlock = _gvn.transform(unlock)->as_Unlock(); 3256 3257 Node* mem = reset_memory(); 3258 3259 // unlock has no side-effects, sets few values 3260 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3261 3262 // Kill monitor from debug info 3263 map()->pop_monitor( ); 3264 } 3265 3266 //-------------------------------get_layout_helper----------------------------- 3267 // If the given klass is a constant or known to be an array, 3268 // fetch the constant layout helper value into constant_value 3269 // and return (Node*)NULL. Otherwise, load the non-constant 3270 // layout helper value, and return the node which represents it. 3271 // This two-faced routine is useful because allocation sites 3272 // almost always feature constant types. 3273 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3274 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3275 if (!StressReflectiveCode && inst_klass != NULL) { 3276 ciKlass* klass = inst_klass->klass(); 3277 assert(klass != NULL, "klass should not be NULL"); 3278 bool xklass = inst_klass->klass_is_exact(); 3279 if (xklass || klass->is_array_klass()) { 3280 jint lhelper = klass->layout_helper(); 3281 if (lhelper != Klass::_lh_neutral_value) { 3282 constant_value = lhelper; 3283 return (Node*) NULL; 3284 } 3285 } 3286 } 3287 constant_value = Klass::_lh_neutral_value; // put in a known value 3288 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3289 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3290 } 3291 3292 // We just put in an allocate/initialize with a big raw-memory effect. 3293 // Hook selected additional alias categories on the initialization. 3294 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3295 MergeMemNode* init_in_merge, 3296 Node* init_out_raw) { 3297 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3298 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3299 3300 Node* prevmem = kit.memory(alias_idx); 3301 init_in_merge->set_memory_at(alias_idx, prevmem); 3302 kit.set_memory(init_out_raw, alias_idx); 3303 } 3304 3305 //---------------------------set_output_for_allocation------------------------- 3306 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3307 const TypeOopPtr* oop_type, 3308 bool deoptimize_on_exception) { 3309 int rawidx = Compile::AliasIdxRaw; 3310 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3311 add_safepoint_edges(alloc); 3312 Node* allocx = _gvn.transform(alloc); 3313 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3314 // create memory projection for i_o 3315 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3316 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3317 3318 // create a memory projection as for the normal control path 3319 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3320 set_memory(malloc, rawidx); 3321 3322 // a normal slow-call doesn't change i_o, but an allocation does 3323 // we create a separate i_o projection for the normal control path 3324 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3325 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3326 3327 // put in an initialization barrier 3328 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3329 rawoop)->as_Initialize(); 3330 assert(alloc->initialization() == init, "2-way macro link must work"); 3331 assert(init ->allocation() == alloc, "2-way macro link must work"); 3332 { 3333 // Extract memory strands which may participate in the new object's 3334 // initialization, and source them from the new InitializeNode. 3335 // This will allow us to observe initializations when they occur, 3336 // and link them properly (as a group) to the InitializeNode. 3337 assert(init->in(InitializeNode::Memory) == malloc, ""); 3338 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3339 init->set_req(InitializeNode::Memory, minit_in); 3340 record_for_igvn(minit_in); // fold it up later, if possible 3341 Node* minit_out = memory(rawidx); 3342 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3343 if (oop_type->isa_aryptr()) { 3344 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3345 int elemidx = C->get_alias_index(telemref); 3346 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3347 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) { 3348 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3349 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3350 ciField* field = ik->nonstatic_field_at(i); 3351 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3352 continue; // do not bother to track really large numbers of fields 3353 // Find (or create) the alias category for this field: 3354 int fieldidx = C->alias_type(field)->index(); 3355 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3356 } 3357 } 3358 } 3359 3360 // Cast raw oop to the real thing... 3361 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3362 javaoop = _gvn.transform(javaoop); 3363 C->set_recent_alloc(control(), javaoop); 3364 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3365 3366 #ifdef ASSERT 3367 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3368 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3369 "Ideal_allocation works"); 3370 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3371 "Ideal_allocation works"); 3372 if (alloc->is_AllocateArray()) { 3373 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3374 "Ideal_allocation works"); 3375 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3376 "Ideal_allocation works"); 3377 } else { 3378 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3379 } 3380 } 3381 #endif //ASSERT 3382 3383 return javaoop; 3384 } 3385 3386 //---------------------------new_instance-------------------------------------- 3387 // This routine takes a klass_node which may be constant (for a static type) 3388 // or may be non-constant (for reflective code). It will work equally well 3389 // for either, and the graph will fold nicely if the optimizer later reduces 3390 // the type to a constant. 3391 // The optional arguments are for specialized use by intrinsics: 3392 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3393 // - If 'return_size_val', report the the total object size to the caller. 3394 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3395 Node* GraphKit::new_instance(Node* klass_node, 3396 Node* extra_slow_test, 3397 Node* *return_size_val, 3398 bool deoptimize_on_exception, 3399 ValueTypeBaseNode* value_node) { 3400 // Compute size in doublewords 3401 // The size is always an integral number of doublewords, represented 3402 // as a positive bytewise size stored in the klass's layout_helper. 3403 // The layout_helper also encodes (in a low bit) the need for a slow path. 3404 jint layout_con = Klass::_lh_neutral_value; 3405 Node* layout_val = get_layout_helper(klass_node, layout_con); 3406 bool layout_is_con = (layout_val == NULL); 3407 3408 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3409 // Generate the initial go-slow test. It's either ALWAYS (return a 3410 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3411 // case) a computed value derived from the layout_helper. 3412 Node* initial_slow_test = NULL; 3413 if (layout_is_con) { 3414 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3415 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3416 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3417 } else { // reflective case 3418 // This reflective path is used by Unsafe.allocateInstance. 3419 // (It may be stress-tested by specifying StressReflectiveCode.) 3420 // Basically, we want to get into the VM is there's an illegal argument. 3421 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3422 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3423 if (extra_slow_test != intcon(0)) { 3424 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3425 } 3426 // (Macro-expander will further convert this to a Bool, if necessary.) 3427 } 3428 3429 // Find the size in bytes. This is easy; it's the layout_helper. 3430 // The size value must be valid even if the slow path is taken. 3431 Node* size = NULL; 3432 if (layout_is_con) { 3433 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3434 } else { // reflective case 3435 // This reflective path is used by clone and Unsafe.allocateInstance. 3436 size = ConvI2X(layout_val); 3437 3438 // Clear the low bits to extract layout_helper_size_in_bytes: 3439 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3440 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3441 size = _gvn.transform( new AndXNode(size, mask) ); 3442 } 3443 if (return_size_val != NULL) { 3444 (*return_size_val) = size; 3445 } 3446 3447 // This is a precise notnull oop of the klass. 3448 // (Actually, it need not be precise if this is a reflective allocation.) 3449 // It's what we cast the result to. 3450 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3451 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3452 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3453 3454 // Now generate allocation code 3455 3456 // The entire memory state is needed for slow path of the allocation 3457 // since GC and deoptimization can happen. 3458 Node *mem = reset_memory(); 3459 set_all_memory(mem); // Create new memory state 3460 3461 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3462 control(), mem, i_o(), 3463 size, klass_node, 3464 initial_slow_test, value_node); 3465 3466 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3467 } 3468 3469 //-------------------------------new_array------------------------------------- 3470 // helper for newarray and anewarray 3471 // The 'length' parameter is (obviously) the length of the array. 3472 // See comments on new_instance for the meaning of the other arguments. 3473 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3474 Node* length, // number of array elements 3475 int nargs, // number of arguments to push back for uncommon trap 3476 Node* *return_size_val, 3477 bool deoptimize_on_exception) { 3478 jint layout_con = Klass::_lh_neutral_value; 3479 Node* layout_val = get_layout_helper(klass_node, layout_con); 3480 bool layout_is_con = (layout_val == NULL); 3481 3482 if (!layout_is_con && !StressReflectiveCode && 3483 !too_many_traps(Deoptimization::Reason_class_check)) { 3484 // This is a reflective array creation site. 3485 // Optimistically assume that it is a subtype of Object[], 3486 // so that we can fold up all the address arithmetic. 3487 layout_con = Klass::array_layout_helper(T_OBJECT); 3488 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3489 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3490 { BuildCutout unless(this, bol_lh, PROB_MAX); 3491 inc_sp(nargs); 3492 uncommon_trap(Deoptimization::Reason_class_check, 3493 Deoptimization::Action_maybe_recompile); 3494 } 3495 layout_val = NULL; 3496 layout_is_con = true; 3497 } 3498 3499 // Generate the initial go-slow test. Make sure we do not overflow 3500 // if length is huge (near 2Gig) or negative! We do not need 3501 // exact double-words here, just a close approximation of needed 3502 // double-words. We can't add any offset or rounding bits, lest we 3503 // take a size -1 of bytes and make it positive. Use an unsigned 3504 // compare, so negative sizes look hugely positive. 3505 int fast_size_limit = FastAllocateSizeLimit; 3506 if (layout_is_con) { 3507 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3508 // Increase the size limit if we have exact knowledge of array type. 3509 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3510 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3511 } 3512 3513 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3514 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3515 3516 // --- Size Computation --- 3517 // array_size = round_to_heap(array_header + (length << elem_shift)); 3518 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3519 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3520 // The rounding mask is strength-reduced, if possible. 3521 int round_mask = MinObjAlignmentInBytes - 1; 3522 Node* header_size = NULL; 3523 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3524 // (T_BYTE has the weakest alignment and size restrictions...) 3525 if (layout_is_con) { 3526 int hsize = Klass::layout_helper_header_size(layout_con); 3527 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3528 BasicType etype = Klass::layout_helper_element_type(layout_con); 3529 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con); 3530 if ((round_mask & ~right_n_bits(eshift)) == 0) 3531 round_mask = 0; // strength-reduce it if it goes away completely 3532 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3533 assert(header_size_min <= hsize, "generic minimum is smallest"); 3534 header_size_min = hsize; 3535 header_size = intcon(hsize + round_mask); 3536 } else { 3537 Node* hss = intcon(Klass::_lh_header_size_shift); 3538 Node* hsm = intcon(Klass::_lh_header_size_mask); 3539 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3540 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3541 Node* mask = intcon(round_mask); 3542 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3543 } 3544 3545 Node* elem_shift = NULL; 3546 if (layout_is_con) { 3547 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3548 if (eshift != 0) 3549 elem_shift = intcon(eshift); 3550 } else { 3551 // There is no need to mask or shift this value. 3552 // The semantics of LShiftINode include an implicit mask to 0x1F. 3553 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3554 elem_shift = layout_val; 3555 } 3556 3557 // Transition to native address size for all offset calculations: 3558 Node* lengthx = ConvI2X(length); 3559 Node* headerx = ConvI2X(header_size); 3560 #ifdef _LP64 3561 { const TypeInt* tilen = _gvn.find_int_type(length); 3562 if (tilen != NULL && tilen->_lo < 0) { 3563 // Add a manual constraint to a positive range. Cf. array_element_address. 3564 jint size_max = fast_size_limit; 3565 if (size_max > tilen->_hi) size_max = tilen->_hi; 3566 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3567 3568 // Only do a narrow I2L conversion if the range check passed. 3569 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3570 _gvn.transform(iff); 3571 RegionNode* region = new RegionNode(3); 3572 _gvn.set_type(region, Type::CONTROL); 3573 lengthx = new PhiNode(region, TypeLong::LONG); 3574 _gvn.set_type(lengthx, TypeLong::LONG); 3575 3576 // Range check passed. Use ConvI2L node with narrow type. 3577 Node* passed = IfFalse(iff); 3578 region->init_req(1, passed); 3579 // Make I2L conversion control dependent to prevent it from 3580 // floating above the range check during loop optimizations. 3581 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3582 3583 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3584 region->init_req(2, IfTrue(iff)); 3585 lengthx->init_req(2, ConvI2X(length)); 3586 3587 set_control(region); 3588 record_for_igvn(region); 3589 record_for_igvn(lengthx); 3590 } 3591 } 3592 #endif 3593 3594 // Combine header size (plus rounding) and body size. Then round down. 3595 // This computation cannot overflow, because it is used only in two 3596 // places, one where the length is sharply limited, and the other 3597 // after a successful allocation. 3598 Node* abody = lengthx; 3599 if (elem_shift != NULL) 3600 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3601 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3602 if (round_mask != 0) { 3603 Node* mask = MakeConX(~round_mask); 3604 size = _gvn.transform( new AndXNode(size, mask) ); 3605 } 3606 // else if round_mask == 0, the size computation is self-rounding 3607 3608 if (return_size_val != NULL) { 3609 // This is the size 3610 (*return_size_val) = size; 3611 } 3612 3613 // Now generate allocation code 3614 3615 // The entire memory state is needed for slow path of the allocation 3616 // since GC and deoptimization can happen. 3617 Node *mem = reset_memory(); 3618 set_all_memory(mem); // Create new memory state 3619 3620 if (initial_slow_test->is_Bool()) { 3621 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3622 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3623 } 3624 3625 // Create the AllocateArrayNode and its result projections 3626 AllocateArrayNode* alloc 3627 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3628 control(), mem, i_o(), 3629 size, klass_node, 3630 initial_slow_test, 3631 length); 3632 3633 // Cast to correct type. Note that the klass_node may be constant or not, 3634 // and in the latter case the actual array type will be inexact also. 3635 // (This happens via a non-constant argument to inline_native_newArray.) 3636 // In any case, the value of klass_node provides the desired array type. 3637 const TypeInt* length_type = _gvn.find_int_type(length); 3638 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3639 if (ary_type->isa_aryptr() && length_type != NULL) { 3640 // Try to get a better type than POS for the size 3641 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3642 } 3643 3644 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3645 3646 // Cast length on remaining path to be as narrow as possible 3647 if (map()->find_edge(length) >= 0) { 3648 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3649 if (ccast != length) { 3650 _gvn.set_type_bottom(ccast); 3651 record_for_igvn(ccast); 3652 replace_in_map(length, ccast); 3653 } 3654 } 3655 3656 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 3657 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL; 3658 if (elem_klass != NULL && elem_klass->is_valuetype()) { 3659 ciValueKlass* vk = elem_klass->as_value_klass(); 3660 if (!vk->flatten_array()) { 3661 // Non-flattened value type arrays need to be initialized with default value type oops 3662 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs); 3663 InitializeNode* init = alloc->initialization(); 3664 init->set_complete_with_arraycopy(); 3665 } 3666 } 3667 3668 return javaoop; 3669 } 3670 3671 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) { 3672 // Check for zero length 3673 Node* null_ctl = top(); 3674 null_check_common(length, T_INT, false, &null_ctl, false); 3675 if (stopped()) { 3676 set_control(null_ctl); // Always zero 3677 return; 3678 } 3679 3680 // Prepare for merging control and IO 3681 RegionNode* res_ctl = new RegionNode(3); 3682 res_ctl->init_req(1, null_ctl); 3683 gvn().set_type(res_ctl, Type::CONTROL); 3684 record_for_igvn(res_ctl); 3685 Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO); 3686 gvn().set_type(res_io, Type::ABIO); 3687 record_for_igvn(res_io); 3688 3689 // TODO comment 3690 SafePointNode* loop_map = NULL; 3691 { 3692 PreserveJVMState pjvms(this); 3693 // Create default value type and store it to memory 3694 Node* oop = ValueTypeNode::make_default(gvn(), vk); 3695 oop = oop->as_ValueType()->allocate(this)->get_oop(); 3696 3697 length = SubI(length, intcon(1)); 3698 add_predicate(nargs); 3699 RegionNode* loop = new RegionNode(3); 3700 loop->init_req(1, control()); 3701 gvn().set_type(loop, Type::CONTROL); 3702 record_for_igvn(loop); 3703 3704 Node* index = new PhiNode(loop, TypeInt::INT); 3705 index->init_req(1, intcon(0)); 3706 gvn().set_type(index, TypeInt::INT); 3707 record_for_igvn(index); 3708 3709 // TODO explain why we need to capture all memory 3710 PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM); 3711 mem->init_req(1, reset_memory()); 3712 gvn().set_type(mem, Type::MEMORY); 3713 record_for_igvn(mem); 3714 set_control(loop); 3715 set_all_memory(mem); 3716 // Initialize array element 3717 Node* adr = array_element_address(array, index, T_OBJECT); 3718 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk); 3719 Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release); 3720 3721 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 3722 loop->init_req(2, IfTrue(iff)); 3723 mem->init_req(2, merged_memory()); 3724 index->init_req(2, AddI(index, intcon(1))); 3725 3726 res_ctl->init_req(2, IfFalse(iff)); 3727 res_io->set_req(2, i_o()); 3728 loop_map = stop(); 3729 } 3730 // Set merged control, IO and memory 3731 set_control(res_ctl); 3732 set_i_o(res_io); 3733 merge_memory(loop_map->merged_memory(), res_ctl, 2); 3734 3735 // Transform new memory Phis. 3736 for (MergeMemStream mms(merged_memory()); mms.next_non_empty();) { 3737 Node* phi = mms.memory(); 3738 if (phi->is_Phi() && phi->in(0) == res_ctl) { 3739 mms.set_memory(gvn().transform(phi)); 3740 } 3741 } 3742 } 3743 3744 // The following "Ideal_foo" functions are placed here because they recognize 3745 // the graph shapes created by the functions immediately above. 3746 3747 //---------------------------Ideal_allocation---------------------------------- 3748 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3749 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3750 if (ptr == NULL) { // reduce dumb test in callers 3751 return NULL; 3752 } 3753 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3754 ptr = ptr->in(1); 3755 if (ptr == NULL) return NULL; 3756 } 3757 // Return NULL for allocations with several casts: 3758 // j.l.reflect.Array.newInstance(jobject, jint) 3759 // Object.clone() 3760 // to keep more precise type from last cast. 3761 if (ptr->is_Proj()) { 3762 Node* allo = ptr->in(0); 3763 if (allo != NULL && allo->is_Allocate()) { 3764 return allo->as_Allocate(); 3765 } 3766 } 3767 // Report failure to match. 3768 return NULL; 3769 } 3770 3771 // Fancy version which also strips off an offset (and reports it to caller). 3772 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3773 intptr_t& offset) { 3774 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3775 if (base == NULL) return NULL; 3776 return Ideal_allocation(base, phase); 3777 } 3778 3779 // Trace Initialize <- Proj[Parm] <- Allocate 3780 AllocateNode* InitializeNode::allocation() { 3781 Node* rawoop = in(InitializeNode::RawAddress); 3782 if (rawoop->is_Proj()) { 3783 Node* alloc = rawoop->in(0); 3784 if (alloc->is_Allocate()) { 3785 return alloc->as_Allocate(); 3786 } 3787 } 3788 return NULL; 3789 } 3790 3791 // Trace Allocate -> Proj[Parm] -> Initialize 3792 InitializeNode* AllocateNode::initialization() { 3793 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3794 if (rawoop == NULL) return NULL; 3795 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3796 Node* init = rawoop->fast_out(i); 3797 if (init->is_Initialize()) { 3798 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3799 return init->as_Initialize(); 3800 } 3801 } 3802 return NULL; 3803 } 3804 3805 //----------------------------- loop predicates --------------------------- 3806 3807 //------------------------------add_predicate_impl---------------------------- 3808 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3809 // Too many traps seen? 3810 if (too_many_traps(reason)) { 3811 #ifdef ASSERT 3812 if (TraceLoopPredicate) { 3813 int tc = C->trap_count(reason); 3814 tty->print("too many traps=%s tcount=%d in ", 3815 Deoptimization::trap_reason_name(reason), tc); 3816 method()->print(); // which method has too many predicate traps 3817 tty->cr(); 3818 } 3819 #endif 3820 // We cannot afford to take more traps here, 3821 // do not generate predicate. 3822 return; 3823 } 3824 3825 Node *cont = _gvn.intcon(1); 3826 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3827 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3828 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3829 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3830 C->add_predicate_opaq(opq); 3831 { 3832 PreserveJVMState pjvms(this); 3833 set_control(iffalse); 3834 inc_sp(nargs); 3835 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3836 } 3837 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3838 set_control(iftrue); 3839 } 3840 3841 //------------------------------add_predicate--------------------------------- 3842 void GraphKit::add_predicate(int nargs) { 3843 if (UseLoopPredicate) { 3844 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3845 } 3846 // loop's limit check predicate should be near the loop. 3847 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3848 } 3849 3850 //----------------------------- store barriers ---------------------------- 3851 #define __ ideal. 3852 3853 void GraphKit::sync_kit(IdealKit& ideal) { 3854 set_all_memory(__ merged_memory()); 3855 set_i_o(__ i_o()); 3856 set_control(__ ctrl()); 3857 } 3858 3859 void GraphKit::final_sync(IdealKit& ideal) { 3860 // Final sync IdealKit and graphKit. 3861 sync_kit(ideal); 3862 } 3863 3864 Node* GraphKit::byte_map_base_node() { 3865 // Get base of card map 3866 CardTableModRefBS* ct = 3867 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3868 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3869 if (ct->byte_map_base != NULL) { 3870 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3871 } else { 3872 return null(); 3873 } 3874 } 3875 3876 // vanilla/CMS post barrier 3877 // Insert a write-barrier store. This is to let generational GC work; we have 3878 // to flag all oop-stores before the next GC point. 3879 void GraphKit::write_barrier_post(Node* oop_store, 3880 Node* obj, 3881 Node* adr, 3882 uint adr_idx, 3883 Node* val, 3884 bool use_precise) { 3885 // No store check needed if we're storing a NULL or an old object 3886 // (latter case is probably a string constant). The concurrent 3887 // mark sweep garbage collector, however, needs to have all nonNull 3888 // oop updates flagged via card-marks. 3889 if (val != NULL && val->is_Con()) { 3890 // must be either an oop or NULL 3891 const Type* t = val->bottom_type(); 3892 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3893 // stores of null never (?) need barriers 3894 return; 3895 } 3896 3897 if (use_ReduceInitialCardMarks() 3898 && obj == just_allocated_object(control())) { 3899 // We can skip marks on a freshly-allocated object in Eden. 3900 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3901 // That routine informs GC to take appropriate compensating steps, 3902 // upon a slow-path allocation, so as to make this card-mark 3903 // elision safe. 3904 return; 3905 } 3906 3907 if (!use_precise) { 3908 // All card marks for a (non-array) instance are in one place: 3909 adr = obj; 3910 } 3911 // (Else it's an array (or unknown), and we want more precise card marks.) 3912 assert(adr != NULL, ""); 3913 3914 IdealKit ideal(this, true); 3915 3916 // Convert the pointer to an int prior to doing math on it 3917 Node* cast = __ CastPX(__ ctrl(), adr); 3918 3919 // Divide by card size 3920 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3921 "Only one we handle so far."); 3922 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3923 3924 // Combine card table base and card offset 3925 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3926 3927 // Get the alias_index for raw card-mark memory 3928 int adr_type = Compile::AliasIdxRaw; 3929 Node* zero = __ ConI(0); // Dirty card value 3930 BasicType bt = T_BYTE; 3931 3932 if (UseConcMarkSweepGC && UseCondCardMark) { 3933 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3934 __ sync_kit(this); 3935 } 3936 3937 if (UseCondCardMark) { 3938 // The classic GC reference write barrier is typically implemented 3939 // as a store into the global card mark table. Unfortunately 3940 // unconditional stores can result in false sharing and excessive 3941 // coherence traffic as well as false transactional aborts. 3942 // UseCondCardMark enables MP "polite" conditional card mark 3943 // stores. In theory we could relax the load from ctrl() to 3944 // no_ctrl, but that doesn't buy much latitude. 3945 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3946 __ if_then(card_val, BoolTest::ne, zero); 3947 } 3948 3949 // Smash zero into card 3950 if( !UseConcMarkSweepGC ) { 3951 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3952 } else { 3953 // Specialized path for CM store barrier 3954 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3955 } 3956 3957 if (UseCondCardMark) { 3958 __ end_if(); 3959 } 3960 3961 // Final sync IdealKit and GraphKit. 3962 final_sync(ideal); 3963 } 3964 /* 3965 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3966 * required by SATB to make sure all objects live at the start of the 3967 * marking are kept alive, all reference updates need to any previous 3968 * reference stored before writing. 3969 * 3970 * If the previous value is NULL there is no need to save the old value. 3971 * References that are NULL are filtered during runtime by the barrier 3972 * code to avoid unnecessary queuing. 3973 * 3974 * However in the case of newly allocated objects it might be possible to 3975 * prove that the reference about to be overwritten is NULL during compile 3976 * time and avoid adding the barrier code completely. 3977 * 3978 * The compiler needs to determine that the object in which a field is about 3979 * to be written is newly allocated, and that no prior store to the same field 3980 * has happened since the allocation. 3981 * 3982 * Returns true if the pre-barrier can be removed 3983 */ 3984 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3985 BasicType bt, uint adr_idx) { 3986 intptr_t offset = 0; 3987 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3988 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3989 3990 if (offset == Type::OffsetBot) { 3991 return false; // cannot unalias unless there are precise offsets 3992 } 3993 3994 if (alloc == NULL) { 3995 return false; // No allocation found 3996 } 3997 3998 intptr_t size_in_bytes = type2aelembytes(bt); 3999 4000 Node* mem = memory(adr_idx); // start searching here... 4001 4002 for (int cnt = 0; cnt < 50; cnt++) { 4003 4004 if (mem->is_Store()) { 4005 4006 Node* st_adr = mem->in(MemNode::Address); 4007 intptr_t st_offset = 0; 4008 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 4009 4010 if (st_base == NULL) { 4011 break; // inscrutable pointer 4012 } 4013 4014 // Break we have found a store with same base and offset as ours so break 4015 if (st_base == base && st_offset == offset) { 4016 break; 4017 } 4018 4019 if (st_offset != offset && st_offset != Type::OffsetBot) { 4020 const int MAX_STORE = BytesPerLong; 4021 if (st_offset >= offset + size_in_bytes || 4022 st_offset <= offset - MAX_STORE || 4023 st_offset <= offset - mem->as_Store()->memory_size()) { 4024 // Success: The offsets are provably independent. 4025 // (You may ask, why not just test st_offset != offset and be done? 4026 // The answer is that stores of different sizes can co-exist 4027 // in the same sequence of RawMem effects. We sometimes initialize 4028 // a whole 'tile' of array elements with a single jint or jlong.) 4029 mem = mem->in(MemNode::Memory); 4030 continue; // advance through independent store memory 4031 } 4032 } 4033 4034 if (st_base != base 4035 && MemNode::detect_ptr_independence(base, alloc, st_base, 4036 AllocateNode::Ideal_allocation(st_base, phase), 4037 phase)) { 4038 // Success: The bases are provably independent. 4039 mem = mem->in(MemNode::Memory); 4040 continue; // advance through independent store memory 4041 } 4042 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4043 4044 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4045 AllocateNode* st_alloc = st_init->allocation(); 4046 4047 // Make sure that we are looking at the same allocation site. 4048 // The alloc variable is guaranteed to not be null here from earlier check. 4049 if (alloc == st_alloc) { 4050 // Check that the initialization is storing NULL so that no previous store 4051 // has been moved up and directly write a reference 4052 Node* captured_store = st_init->find_captured_store(offset, 4053 type2aelembytes(T_OBJECT), 4054 phase); 4055 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4056 return true; 4057 } 4058 } 4059 } 4060 4061 // Unless there is an explicit 'continue', we must bail out here, 4062 // because 'mem' is an inscrutable memory state (e.g., a call). 4063 break; 4064 } 4065 4066 return false; 4067 } 4068 4069 // G1 pre/post barriers 4070 void GraphKit::g1_write_barrier_pre(bool do_load, 4071 Node* obj, 4072 Node* adr, 4073 uint alias_idx, 4074 Node* val, 4075 const TypeOopPtr* val_type, 4076 Node* pre_val, 4077 BasicType bt) { 4078 4079 // Some sanity checks 4080 // Note: val is unused in this routine. 4081 4082 if (do_load) { 4083 // We need to generate the load of the previous value 4084 assert(obj != NULL, "must have a base"); 4085 assert(adr != NULL, "where are loading from?"); 4086 assert(pre_val == NULL, "loaded already?"); 4087 assert(val_type != NULL, "need a type"); 4088 4089 if (use_ReduceInitialCardMarks() 4090 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4091 return; 4092 } 4093 4094 } else { 4095 // In this case both val_type and alias_idx are unused. 4096 assert(pre_val != NULL, "must be loaded already"); 4097 // Nothing to be done if pre_val is null. 4098 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4099 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4100 } 4101 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here"); 4102 4103 IdealKit ideal(this, true); 4104 4105 Node* tls = __ thread(); // ThreadLocalStorage 4106 4107 Node* no_ctrl = NULL; 4108 Node* no_base = __ top(); 4109 Node* zero = __ ConI(0); 4110 Node* zeroX = __ ConX(0); 4111 4112 float likely = PROB_LIKELY(0.999); 4113 float unlikely = PROB_UNLIKELY(0.999); 4114 4115 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4116 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4117 4118 // Offsets into the thread 4119 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4120 SATBMarkQueue::byte_offset_of_active()); 4121 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4122 SATBMarkQueue::byte_offset_of_index()); 4123 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4124 SATBMarkQueue::byte_offset_of_buf()); 4125 4126 // Now the actual pointers into the thread 4127 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4128 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4129 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4130 4131 // Now some of the values 4132 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4133 4134 // if (!marking) 4135 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4136 BasicType index_bt = TypeX_X->basic_type(); 4137 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4138 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4139 4140 if (do_load) { 4141 // load original value 4142 // alias_idx correct?? 4143 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4144 } 4145 4146 // if (pre_val != NULL) 4147 __ if_then(pre_val, BoolTest::ne, null()); { 4148 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4149 4150 // is the queue for this thread full? 4151 __ if_then(index, BoolTest::ne, zeroX, likely); { 4152 4153 // decrement the index 4154 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4155 4156 // Now get the buffer location we will log the previous value into and store it 4157 Node *log_addr = __ AddP(no_base, buffer, next_index); 4158 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4159 // update the index 4160 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4161 4162 } __ else_(); { 4163 4164 // logging buffer is full, call the runtime 4165 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4166 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4167 } __ end_if(); // (!index) 4168 } __ end_if(); // (pre_val != NULL) 4169 } __ end_if(); // (!marking) 4170 4171 // Final sync IdealKit and GraphKit. 4172 final_sync(ideal); 4173 } 4174 4175 /* 4176 * G1 similar to any GC with a Young Generation requires a way to keep track of 4177 * references from Old Generation to Young Generation to make sure all live 4178 * objects are found. G1 also requires to keep track of object references 4179 * between different regions to enable evacuation of old regions, which is done 4180 * as part of mixed collections. References are tracked in remembered sets and 4181 * is continuously updated as reference are written to with the help of the 4182 * post-barrier. 4183 * 4184 * To reduce the number of updates to the remembered set the post-barrier 4185 * filters updates to fields in objects located in the Young Generation, 4186 * the same region as the reference, when the NULL is being written or 4187 * if the card is already marked as dirty by an earlier write. 4188 * 4189 * Under certain circumstances it is possible to avoid generating the 4190 * post-barrier completely if it is possible during compile time to prove 4191 * the object is newly allocated and that no safepoint exists between the 4192 * allocation and the store. 4193 * 4194 * In the case of slow allocation the allocation code must handle the barrier 4195 * as part of the allocation in the case the allocated object is not located 4196 * in the nursery, this would happen for humongous objects. This is similar to 4197 * how CMS is required to handle this case, see the comments for the method 4198 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4199 * A deferred card mark is required for these objects and handled in the above 4200 * mentioned methods. 4201 * 4202 * Returns true if the post barrier can be removed 4203 */ 4204 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4205 Node* adr) { 4206 intptr_t offset = 0; 4207 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4208 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4209 4210 if (offset == Type::OffsetBot) { 4211 return false; // cannot unalias unless there are precise offsets 4212 } 4213 4214 if (alloc == NULL) { 4215 return false; // No allocation found 4216 } 4217 4218 // Start search from Store node 4219 Node* mem = store->in(MemNode::Control); 4220 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4221 4222 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4223 AllocateNode* st_alloc = st_init->allocation(); 4224 4225 // Make sure we are looking at the same allocation 4226 if (alloc == st_alloc) { 4227 return true; 4228 } 4229 } 4230 4231 return false; 4232 } 4233 4234 // 4235 // Update the card table and add card address to the queue 4236 // 4237 void GraphKit::g1_mark_card(IdealKit& ideal, 4238 Node* card_adr, 4239 Node* oop_store, 4240 uint oop_alias_idx, 4241 Node* index, 4242 Node* index_adr, 4243 Node* buffer, 4244 const TypeFunc* tf) { 4245 4246 Node* zero = __ ConI(0); 4247 Node* zeroX = __ ConX(0); 4248 Node* no_base = __ top(); 4249 BasicType card_bt = T_BYTE; 4250 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4251 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4252 4253 // Now do the queue work 4254 __ if_then(index, BoolTest::ne, zeroX); { 4255 4256 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4257 Node* log_addr = __ AddP(no_base, buffer, next_index); 4258 4259 // Order, see storeCM. 4260 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4261 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4262 4263 } __ else_(); { 4264 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4265 } __ end_if(); 4266 4267 } 4268 4269 void GraphKit::g1_write_barrier_post(Node* oop_store, 4270 Node* obj, 4271 Node* adr, 4272 uint alias_idx, 4273 Node* val, 4274 BasicType bt, 4275 bool use_precise) { 4276 // If we are writing a NULL then we need no post barrier 4277 4278 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4279 // Must be NULL 4280 const Type* t = val->bottom_type(); 4281 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4282 // No post barrier if writing NULLx 4283 return; 4284 } 4285 4286 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4287 // We can skip marks on a freshly-allocated object in Eden. 4288 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4289 // That routine informs GC to take appropriate compensating steps, 4290 // upon a slow-path allocation, so as to make this card-mark 4291 // elision safe. 4292 return; 4293 } 4294 4295 if (use_ReduceInitialCardMarks() 4296 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4297 return; 4298 } 4299 4300 if (!use_precise) { 4301 // All card marks for a (non-array) instance are in one place: 4302 adr = obj; 4303 } 4304 // (Else it's an array (or unknown), and we want more precise card marks.) 4305 assert(adr != NULL, ""); 4306 4307 IdealKit ideal(this, true); 4308 4309 Node* tls = __ thread(); // ThreadLocalStorage 4310 4311 Node* no_base = __ top(); 4312 float likely = PROB_LIKELY(0.999); 4313 float unlikely = PROB_UNLIKELY(0.999); 4314 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4315 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4316 Node* zeroX = __ ConX(0); 4317 4318 // Get the alias_index for raw card-mark memory 4319 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4320 4321 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4322 4323 // Offsets into the thread 4324 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4325 DirtyCardQueue::byte_offset_of_index()); 4326 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4327 DirtyCardQueue::byte_offset_of_buf()); 4328 4329 // Pointers into the thread 4330 4331 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4332 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4333 4334 // Now some values 4335 // Use ctrl to avoid hoisting these values past a safepoint, which could 4336 // potentially reset these fields in the JavaThread. 4337 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4338 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4339 4340 // Convert the store obj pointer to an int prior to doing math on it 4341 // Must use ctrl to prevent "integerized oop" existing across safepoint 4342 Node* cast = __ CastPX(__ ctrl(), adr); 4343 4344 // Divide pointer by card size 4345 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4346 4347 // Combine card table base and card offset 4348 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4349 4350 // If we know the value being stored does it cross regions? 4351 4352 if (val != NULL) { 4353 // Does the store cause us to cross regions? 4354 4355 // Should be able to do an unsigned compare of region_size instead of 4356 // and extra shift. Do we have an unsigned compare?? 4357 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4358 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4359 4360 // if (xor_res == 0) same region so skip 4361 __ if_then(xor_res, BoolTest::ne, zeroX); { 4362 4363 // No barrier if we are storing a NULL 4364 __ if_then(val, BoolTest::ne, null(), unlikely); { 4365 4366 // Ok must mark the card if not already dirty 4367 4368 // load the original value of the card 4369 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4370 4371 __ if_then(card_val, BoolTest::ne, young_card); { 4372 sync_kit(ideal); 4373 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4374 insert_mem_bar(Op_MemBarVolatile, oop_store); 4375 __ sync_kit(this); 4376 4377 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4378 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4379 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4380 } __ end_if(); 4381 } __ end_if(); 4382 } __ end_if(); 4383 } __ end_if(); 4384 } else { 4385 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4386 // We don't need a barrier here if the destination is a newly allocated object 4387 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4388 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4389 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4390 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4391 __ if_then(card_val, BoolTest::ne, young_card); { 4392 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4393 } __ end_if(); 4394 } 4395 4396 // Final sync IdealKit and GraphKit. 4397 final_sync(ideal); 4398 } 4399 #undef __ 4400 4401 4402 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4403 Node* len = load_array_length(load_String_value(ctrl, str)); 4404 Node* coder = load_String_coder(ctrl, str); 4405 // Divide length by 2 if coder is UTF16 4406 return _gvn.transform(new RShiftINode(len, coder)); 4407 } 4408 4409 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4410 int value_offset = java_lang_String::value_offset_in_bytes(); 4411 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4412 false, NULL, Type::Offset(0)); 4413 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4414 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4415 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4416 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4417 int value_field_idx = C->get_alias_index(value_field_type); 4418 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4419 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4420 // String.value field is known to be @Stable. 4421 if (UseImplicitStableValues) { 4422 load = cast_array_to_stable(load, value_type); 4423 } 4424 return load; 4425 } 4426 4427 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4428 if (!CompactStrings) { 4429 return intcon(java_lang_String::CODER_UTF16); 4430 } 4431 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4432 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4433 false, NULL, Type::Offset(0)); 4434 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4435 int coder_field_idx = C->get_alias_index(coder_field_type); 4436 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4437 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4438 } 4439 4440 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4441 int value_offset = java_lang_String::value_offset_in_bytes(); 4442 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4443 false, NULL, Type::Offset(0)); 4444 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4445 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4446 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4447 } 4448 4449 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4450 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4451 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4452 false, NULL, Type::Offset(0)); 4453 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4454 int coder_field_idx = C->get_alias_index(coder_field_type); 4455 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4456 value, T_BYTE, coder_field_idx, MemNode::unordered); 4457 } 4458 4459 // Capture src and dst memory state with a MergeMemNode 4460 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4461 if (src_type == dst_type) { 4462 // Types are equal, we don't need a MergeMemNode 4463 return memory(src_type); 4464 } 4465 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4466 record_for_igvn(merge); // fold it up later, if possible 4467 int src_idx = C->get_alias_index(src_type); 4468 int dst_idx = C->get_alias_index(dst_type); 4469 merge->set_memory_at(src_idx, memory(src_idx)); 4470 merge->set_memory_at(dst_idx, memory(dst_idx)); 4471 return merge; 4472 } 4473 4474 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4475 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4476 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4477 // If input and output memory types differ, capture both states to preserve 4478 // the dependency between preceding and subsequent loads/stores. 4479 // For example, the following program: 4480 // StoreB 4481 // compress_string 4482 // LoadB 4483 // has this memory graph (use->def): 4484 // LoadB -> compress_string -> CharMem 4485 // ... -> StoreB -> ByteMem 4486 // The intrinsic hides the dependency between LoadB and StoreB, causing 4487 // the load to read from memory not containing the result of the StoreB. 4488 // The correct memory graph should look like this: 4489 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4490 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4491 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4492 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4493 set_memory(res_mem, TypeAryPtr::BYTES); 4494 return str; 4495 } 4496 4497 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4498 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4499 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4500 // Capture src and dst memory (see comment in 'compress_string'). 4501 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4502 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4503 set_memory(_gvn.transform(str), dst_type); 4504 } 4505 4506 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4507 /** 4508 * int i_char = start; 4509 * for (int i_byte = 0; i_byte < count; i_byte++) { 4510 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4511 * } 4512 */ 4513 add_predicate(); 4514 RegionNode* head = new RegionNode(3); 4515 head->init_req(1, control()); 4516 gvn().set_type(head, Type::CONTROL); 4517 record_for_igvn(head); 4518 4519 Node* i_byte = new PhiNode(head, TypeInt::INT); 4520 i_byte->init_req(1, intcon(0)); 4521 gvn().set_type(i_byte, TypeInt::INT); 4522 record_for_igvn(i_byte); 4523 4524 Node* i_char = new PhiNode(head, TypeInt::INT); 4525 i_char->init_req(1, start); 4526 gvn().set_type(i_char, TypeInt::INT); 4527 record_for_igvn(i_char); 4528 4529 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4530 gvn().set_type(mem, Type::MEMORY); 4531 record_for_igvn(mem); 4532 set_control(head); 4533 set_memory(mem, TypeAryPtr::BYTES); 4534 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4535 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4536 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4537 false, false, true /* mismatched */); 4538 4539 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4540 head->init_req(2, IfTrue(iff)); 4541 mem->init_req(2, st); 4542 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4543 i_char->init_req(2, AddI(i_char, intcon(2))); 4544 4545 set_control(IfFalse(iff)); 4546 set_memory(st, TypeAryPtr::BYTES); 4547 } 4548 4549 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4550 if (!field->is_constant()) { 4551 return NULL; // Field not marked as constant. 4552 } 4553 ciInstance* holder = NULL; 4554 if (!field->is_static()) { 4555 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4556 if (const_oop != NULL && const_oop->is_instance()) { 4557 holder = const_oop->as_instance(); 4558 } 4559 } 4560 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4561 /*is_unsigned_load=*/false); 4562 if (con_type != NULL) { 4563 Node* con = makecon(con_type); 4564 if (field->layout_type() == T_VALUETYPE) { 4565 // Load value type from constant oop 4566 con = ValueTypeNode::make(this, con); 4567 } 4568 return con; 4569 } 4570 return NULL; 4571 } 4572 4573 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4574 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4575 // assumption of CCP analysis. 4576 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4577 }