1 /* 2 * Copyright (c) 2001, 2018, 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 "ci/ciUtilities.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/g1/g1BarrierSet.hpp" 29 #include "gc/g1/g1CardTable.hpp" 30 #include "gc/g1/heapRegion.hpp" 31 #include "gc/shared/barrierSet.hpp" 32 #include "gc/shared/cardTable.hpp" 33 #include "gc/shared/cardTableBarrierSet.hpp" 34 #include "gc/shared/collectedHeap.hpp" 35 #include "interpreter/interpreter.hpp" 36 #include "memory/resourceArea.hpp" 37 #include "opto/addnode.hpp" 38 #include "opto/castnode.hpp" 39 #include "opto/convertnode.hpp" 40 #include "opto/graphKit.hpp" 41 #include "opto/idealKit.hpp" 42 #include "opto/intrinsicnode.hpp" 43 #include "opto/locknode.hpp" 44 #include "opto/machnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/rootnode.hpp" 48 #include "opto/runtime.hpp" 49 #include "runtime/deoptimization.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #if INCLUDE_ALL_GCS 52 #include "gc/g1/g1ThreadLocalData.hpp" 53 #endif // INCLUDE_ALL_GCS 54 55 //----------------------------GraphKit----------------------------------------- 56 // Main utility constructor. 57 GraphKit::GraphKit(JVMState* jvms) 58 : Phase(Phase::Parser), 59 _env(C->env()), 60 _gvn(*C->initial_gvn()) 61 { 62 _exceptions = jvms->map()->next_exception(); 63 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 64 set_jvms(jvms); 65 } 66 67 // Private constructor for parser. 68 GraphKit::GraphKit() 69 : Phase(Phase::Parser), 70 _env(C->env()), 71 _gvn(*C->initial_gvn()) 72 { 73 _exceptions = NULL; 74 set_map(NULL); 75 debug_only(_sp = -99); 76 debug_only(set_bci(-99)); 77 } 78 79 80 81 //---------------------------clean_stack--------------------------------------- 82 // Clear away rubbish from the stack area of the JVM state. 83 // This destroys any arguments that may be waiting on the stack. 84 void GraphKit::clean_stack(int from_sp) { 85 SafePointNode* map = this->map(); 86 JVMState* jvms = this->jvms(); 87 int stk_size = jvms->stk_size(); 88 int stkoff = jvms->stkoff(); 89 Node* top = this->top(); 90 for (int i = from_sp; i < stk_size; i++) { 91 if (map->in(stkoff + i) != top) { 92 map->set_req(stkoff + i, top); 93 } 94 } 95 } 96 97 98 //--------------------------------sync_jvms----------------------------------- 99 // Make sure our current jvms agrees with our parse state. 100 JVMState* GraphKit::sync_jvms() const { 101 JVMState* jvms = this->jvms(); 102 jvms->set_bci(bci()); // Record the new bci in the JVMState 103 jvms->set_sp(sp()); // Record the new sp in the JVMState 104 assert(jvms_in_sync(), "jvms is now in sync"); 105 return jvms; 106 } 107 108 //--------------------------------sync_jvms_for_reexecute--------------------- 109 // Make sure our current jvms agrees with our parse state. This version 110 // uses the reexecute_sp for reexecuting bytecodes. 111 JVMState* GraphKit::sync_jvms_for_reexecute() { 112 JVMState* jvms = this->jvms(); 113 jvms->set_bci(bci()); // Record the new bci in the JVMState 114 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 115 return jvms; 116 } 117 118 #ifdef ASSERT 119 bool GraphKit::jvms_in_sync() const { 120 Parse* parse = is_Parse(); 121 if (parse == NULL) { 122 if (bci() != jvms()->bci()) return false; 123 if (sp() != (int)jvms()->sp()) return false; 124 return true; 125 } 126 if (jvms()->method() != parse->method()) return false; 127 if (jvms()->bci() != parse->bci()) return false; 128 int jvms_sp = jvms()->sp(); 129 if (jvms_sp != parse->sp()) return false; 130 int jvms_depth = jvms()->depth(); 131 if (jvms_depth != parse->depth()) return false; 132 return true; 133 } 134 135 // Local helper checks for special internal merge points 136 // used to accumulate and merge exception states. 137 // They are marked by the region's in(0) edge being the map itself. 138 // Such merge points must never "escape" into the parser at large, 139 // until they have been handed to gvn.transform. 140 static bool is_hidden_merge(Node* reg) { 141 if (reg == NULL) return false; 142 if (reg->is_Phi()) { 143 reg = reg->in(0); 144 if (reg == NULL) return false; 145 } 146 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 147 } 148 149 void GraphKit::verify_map() const { 150 if (map() == NULL) return; // null map is OK 151 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 152 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 153 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 154 } 155 156 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 157 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 158 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 159 } 160 #endif 161 162 //---------------------------stop_and_kill_map--------------------------------- 163 // Set _map to NULL, signalling a stop to further bytecode execution. 164 // First smash the current map's control to a constant, to mark it dead. 165 void GraphKit::stop_and_kill_map() { 166 SafePointNode* dead_map = stop(); 167 if (dead_map != NULL) { 168 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 169 assert(dead_map->is_killed(), "must be so marked"); 170 } 171 } 172 173 174 //--------------------------------stopped-------------------------------------- 175 // Tell if _map is NULL, or control is top. 176 bool GraphKit::stopped() { 177 if (map() == NULL) return true; 178 else if (control() == top()) return true; 179 else return false; 180 } 181 182 183 //-----------------------------has_ex_handler---------------------------------- 184 // Tell if this method or any caller method has exception handlers. 185 bool GraphKit::has_ex_handler() { 186 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 187 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 188 return true; 189 } 190 } 191 return false; 192 } 193 194 //------------------------------save_ex_oop------------------------------------ 195 // Save an exception without blowing stack contents or other JVM state. 196 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 197 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 198 ex_map->add_req(ex_oop); 199 debug_only(verify_exception_state(ex_map)); 200 } 201 202 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 203 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 204 Node* ex_oop = ex_map->in(ex_map->req()-1); 205 if (clear_it) ex_map->del_req(ex_map->req()-1); 206 return ex_oop; 207 } 208 209 //-----------------------------saved_ex_oop------------------------------------ 210 // Recover a saved exception from its map. 211 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 212 return common_saved_ex_oop(ex_map, false); 213 } 214 215 //--------------------------clear_saved_ex_oop--------------------------------- 216 // Erase a previously saved exception from its map. 217 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 218 return common_saved_ex_oop(ex_map, true); 219 } 220 221 #ifdef ASSERT 222 //---------------------------has_saved_ex_oop---------------------------------- 223 // Erase a previously saved exception from its map. 224 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 225 return ex_map->req() == ex_map->jvms()->endoff()+1; 226 } 227 #endif 228 229 //-------------------------make_exception_state-------------------------------- 230 // Turn the current JVM state into an exception state, appending the ex_oop. 231 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 232 sync_jvms(); 233 SafePointNode* ex_map = stop(); // do not manipulate this map any more 234 set_saved_ex_oop(ex_map, ex_oop); 235 return ex_map; 236 } 237 238 239 //--------------------------add_exception_state-------------------------------- 240 // Add an exception to my list of exceptions. 241 void GraphKit::add_exception_state(SafePointNode* ex_map) { 242 if (ex_map == NULL || ex_map->control() == top()) { 243 return; 244 } 245 #ifdef ASSERT 246 verify_exception_state(ex_map); 247 if (has_exceptions()) { 248 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 249 } 250 #endif 251 252 // If there is already an exception of exactly this type, merge with it. 253 // In particular, null-checks and other low-level exceptions common up here. 254 Node* ex_oop = saved_ex_oop(ex_map); 255 const Type* ex_type = _gvn.type(ex_oop); 256 if (ex_oop == top()) { 257 // No action needed. 258 return; 259 } 260 assert(ex_type->isa_instptr(), "exception must be an instance"); 261 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 262 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 263 // We check sp also because call bytecodes can generate exceptions 264 // both before and after arguments are popped! 265 if (ex_type2 == ex_type 266 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 267 combine_exception_states(ex_map, e2); 268 return; 269 } 270 } 271 272 // No pre-existing exception of the same type. Chain it on the list. 273 push_exception_state(ex_map); 274 } 275 276 //-----------------------add_exception_states_from----------------------------- 277 void GraphKit::add_exception_states_from(JVMState* jvms) { 278 SafePointNode* ex_map = jvms->map()->next_exception(); 279 if (ex_map != NULL) { 280 jvms->map()->set_next_exception(NULL); 281 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 282 next_map = ex_map->next_exception(); 283 ex_map->set_next_exception(NULL); 284 add_exception_state(ex_map); 285 } 286 } 287 } 288 289 //-----------------------transfer_exceptions_into_jvms------------------------- 290 JVMState* GraphKit::transfer_exceptions_into_jvms() { 291 if (map() == NULL) { 292 // We need a JVMS to carry the exceptions, but the map has gone away. 293 // Create a scratch JVMS, cloned from any of the exception states... 294 if (has_exceptions()) { 295 _map = _exceptions; 296 _map = clone_map(); 297 _map->set_next_exception(NULL); 298 clear_saved_ex_oop(_map); 299 debug_only(verify_map()); 300 } else { 301 // ...or created from scratch 302 JVMState* jvms = new (C) JVMState(_method, NULL); 303 jvms->set_bci(_bci); 304 jvms->set_sp(_sp); 305 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 306 set_jvms(jvms); 307 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 308 set_all_memory(top()); 309 while (map()->req() < jvms->endoff()) map()->add_req(top()); 310 } 311 // (This is a kludge, in case you didn't notice.) 312 set_control(top()); 313 } 314 JVMState* jvms = sync_jvms(); 315 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 316 jvms->map()->set_next_exception(_exceptions); 317 _exceptions = NULL; // done with this set of exceptions 318 return jvms; 319 } 320 321 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 322 assert(is_hidden_merge(dstphi), "must be a special merge node"); 323 assert(is_hidden_merge(srcphi), "must be a special merge node"); 324 uint limit = srcphi->req(); 325 for (uint i = PhiNode::Input; i < limit; i++) { 326 dstphi->add_req(srcphi->in(i)); 327 } 328 } 329 static inline void add_one_req(Node* dstphi, Node* src) { 330 assert(is_hidden_merge(dstphi), "must be a special merge node"); 331 assert(!is_hidden_merge(src), "must not be a special merge node"); 332 dstphi->add_req(src); 333 } 334 335 //-----------------------combine_exception_states------------------------------ 336 // This helper function combines exception states by building phis on a 337 // specially marked state-merging region. These regions and phis are 338 // untransformed, and can build up gradually. The region is marked by 339 // having a control input of its exception map, rather than NULL. Such 340 // regions do not appear except in this function, and in use_exception_state. 341 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 342 if (failing()) return; // dying anyway... 343 JVMState* ex_jvms = ex_map->_jvms; 344 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 345 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 346 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 347 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 348 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 349 assert(ex_map->req() == phi_map->req(), "matching maps"); 350 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 351 Node* hidden_merge_mark = root(); 352 Node* region = phi_map->control(); 353 MergeMemNode* phi_mem = phi_map->merged_memory(); 354 MergeMemNode* ex_mem = ex_map->merged_memory(); 355 if (region->in(0) != hidden_merge_mark) { 356 // The control input is not (yet) a specially-marked region in phi_map. 357 // Make it so, and build some phis. 358 region = new RegionNode(2); 359 _gvn.set_type(region, Type::CONTROL); 360 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 361 region->init_req(1, phi_map->control()); 362 phi_map->set_control(region); 363 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 364 record_for_igvn(io_phi); 365 _gvn.set_type(io_phi, Type::ABIO); 366 phi_map->set_i_o(io_phi); 367 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 368 Node* m = mms.memory(); 369 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 370 record_for_igvn(m_phi); 371 _gvn.set_type(m_phi, Type::MEMORY); 372 mms.set_memory(m_phi); 373 } 374 } 375 376 // Either or both of phi_map and ex_map might already be converted into phis. 377 Node* ex_control = ex_map->control(); 378 // if there is special marking on ex_map also, we add multiple edges from src 379 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 380 // how wide was the destination phi_map, originally? 381 uint orig_width = region->req(); 382 383 if (add_multiple) { 384 add_n_reqs(region, ex_control); 385 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 386 } else { 387 // ex_map has no merges, so we just add single edges everywhere 388 add_one_req(region, ex_control); 389 add_one_req(phi_map->i_o(), ex_map->i_o()); 390 } 391 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 392 if (mms.is_empty()) { 393 // get a copy of the base memory, and patch some inputs into it 394 const TypePtr* adr_type = mms.adr_type(C); 395 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 396 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 397 mms.set_memory(phi); 398 // Prepare to append interesting stuff onto the newly sliced phi: 399 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 400 } 401 // Append stuff from ex_map: 402 if (add_multiple) { 403 add_n_reqs(mms.memory(), mms.memory2()); 404 } else { 405 add_one_req(mms.memory(), mms.memory2()); 406 } 407 } 408 uint limit = ex_map->req(); 409 for (uint i = TypeFunc::Parms; i < limit; i++) { 410 // Skip everything in the JVMS after tos. (The ex_oop follows.) 411 if (i == tos) i = ex_jvms->monoff(); 412 Node* src = ex_map->in(i); 413 Node* dst = phi_map->in(i); 414 if (src != dst) { 415 PhiNode* phi; 416 if (dst->in(0) != region) { 417 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 418 record_for_igvn(phi); 419 _gvn.set_type(phi, phi->type()); 420 phi_map->set_req(i, dst); 421 // Prepare to append interesting stuff onto the new phi: 422 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 423 } else { 424 assert(dst->is_Phi(), "nobody else uses a hidden region"); 425 phi = dst->as_Phi(); 426 } 427 if (add_multiple && src->in(0) == ex_control) { 428 // Both are phis. 429 add_n_reqs(dst, src); 430 } else { 431 while (dst->req() < region->req()) add_one_req(dst, src); 432 } 433 const Type* srctype = _gvn.type(src); 434 if (phi->type() != srctype) { 435 const Type* dsttype = phi->type()->meet_speculative(srctype); 436 if (phi->type() != dsttype) { 437 phi->set_type(dsttype); 438 _gvn.set_type(phi, dsttype); 439 } 440 } 441 } 442 } 443 phi_map->merge_replaced_nodes_with(ex_map); 444 } 445 446 //--------------------------use_exception_state-------------------------------- 447 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 448 if (failing()) { stop(); return top(); } 449 Node* region = phi_map->control(); 450 Node* hidden_merge_mark = root(); 451 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 452 Node* ex_oop = clear_saved_ex_oop(phi_map); 453 if (region->in(0) == hidden_merge_mark) { 454 // Special marking for internal ex-states. Process the phis now. 455 region->set_req(0, region); // now it's an ordinary region 456 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 457 // Note: Setting the jvms also sets the bci and sp. 458 set_control(_gvn.transform(region)); 459 uint tos = jvms()->stkoff() + sp(); 460 for (uint i = 1; i < tos; i++) { 461 Node* x = phi_map->in(i); 462 if (x->in(0) == region) { 463 assert(x->is_Phi(), "expected a special phi"); 464 phi_map->set_req(i, _gvn.transform(x)); 465 } 466 } 467 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 468 Node* x = mms.memory(); 469 if (x->in(0) == region) { 470 assert(x->is_Phi(), "nobody else uses a hidden region"); 471 mms.set_memory(_gvn.transform(x)); 472 } 473 } 474 if (ex_oop->in(0) == region) { 475 assert(ex_oop->is_Phi(), "expected a special phi"); 476 ex_oop = _gvn.transform(ex_oop); 477 } 478 } else { 479 set_jvms(phi_map->jvms()); 480 } 481 482 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 483 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 484 return ex_oop; 485 } 486 487 //---------------------------------java_bc------------------------------------- 488 Bytecodes::Code GraphKit::java_bc() const { 489 ciMethod* method = this->method(); 490 int bci = this->bci(); 491 if (method != NULL && bci != InvocationEntryBci) 492 return method->java_code_at_bci(bci); 493 else 494 return Bytecodes::_illegal; 495 } 496 497 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 498 bool must_throw) { 499 // if the exception capability is set, then we will generate code 500 // to check the JavaThread.should_post_on_exceptions flag to see 501 // if we actually need to report exception events (for this 502 // thread). If we don't need to report exception events, we will 503 // take the normal fast path provided by add_exception_events. If 504 // exception event reporting is enabled for this thread, we will 505 // take the uncommon_trap in the BuildCutout below. 506 507 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 508 Node* jthread = _gvn.transform(new ThreadLocalNode()); 509 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 510 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 511 512 // Test the should_post_on_exceptions_flag vs. 0 513 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 514 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 515 516 // Branch to slow_path if should_post_on_exceptions_flag was true 517 { BuildCutout unless(this, tst, PROB_MAX); 518 // Do not try anything fancy if we're notifying the VM on every throw. 519 // Cf. case Bytecodes::_athrow in parse2.cpp. 520 uncommon_trap(reason, Deoptimization::Action_none, 521 (ciKlass*)NULL, (char*)NULL, must_throw); 522 } 523 524 } 525 526 //------------------------------builtin_throw---------------------------------- 527 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 528 bool must_throw = true; 529 530 if (env()->jvmti_can_post_on_exceptions()) { 531 // check if we must post exception events, take uncommon trap if so 532 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 533 // here if should_post_on_exceptions is false 534 // continue on with the normal codegen 535 } 536 537 // If this particular condition has not yet happened at this 538 // bytecode, then use the uncommon trap mechanism, and allow for 539 // a future recompilation if several traps occur here. 540 // If the throw is hot, try to use a more complicated inline mechanism 541 // which keeps execution inside the compiled code. 542 bool treat_throw_as_hot = false; 543 ciMethodData* md = method()->method_data(); 544 545 if (ProfileTraps) { 546 if (too_many_traps(reason)) { 547 treat_throw_as_hot = true; 548 } 549 // (If there is no MDO at all, assume it is early in 550 // execution, and that any deopts are part of the 551 // startup transient, and don't need to be remembered.) 552 553 // Also, if there is a local exception handler, treat all throws 554 // as hot if there has been at least one in this method. 555 if (C->trap_count(reason) != 0 556 && method()->method_data()->trap_count(reason) != 0 557 && has_ex_handler()) { 558 treat_throw_as_hot = true; 559 } 560 } 561 562 // If this throw happens frequently, an uncommon trap might cause 563 // a performance pothole. If there is a local exception handler, 564 // and if this particular bytecode appears to be deoptimizing often, 565 // let us handle the throw inline, with a preconstructed instance. 566 // Note: If the deopt count has blown up, the uncommon trap 567 // runtime is going to flush this nmethod, not matter what. 568 if (treat_throw_as_hot 569 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 570 // If the throw is local, we use a pre-existing instance and 571 // punt on the backtrace. This would lead to a missing backtrace 572 // (a repeat of 4292742) if the backtrace object is ever asked 573 // for its backtrace. 574 // Fixing this remaining case of 4292742 requires some flavor of 575 // escape analysis. Leave that for the future. 576 ciInstance* ex_obj = NULL; 577 switch (reason) { 578 case Deoptimization::Reason_null_check: 579 ex_obj = env()->NullPointerException_instance(); 580 break; 581 case Deoptimization::Reason_div0_check: 582 ex_obj = env()->ArithmeticException_instance(); 583 break; 584 case Deoptimization::Reason_range_check: 585 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 586 break; 587 case Deoptimization::Reason_class_check: 588 if (java_bc() == Bytecodes::_aastore) { 589 ex_obj = env()->ArrayStoreException_instance(); 590 } else { 591 ex_obj = env()->ClassCastException_instance(); 592 } 593 break; 594 default: 595 break; 596 } 597 if (failing()) { stop(); return; } // exception allocation might fail 598 if (ex_obj != NULL) { 599 // Cheat with a preallocated exception object. 600 if (C->log() != NULL) 601 C->log()->elem("hot_throw preallocated='1' reason='%s'", 602 Deoptimization::trap_reason_name(reason)); 603 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 604 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 605 606 // Clear the detail message of the preallocated exception object. 607 // Weblogic sometimes mutates the detail message of exceptions 608 // using reflection. 609 int offset = java_lang_Throwable::get_detailMessage_offset(); 610 const TypePtr* adr_typ = ex_con->add_offset(offset); 611 612 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 613 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 614 // Conservatively release stores of object references. 615 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 616 617 add_exception_state(make_exception_state(ex_node)); 618 return; 619 } 620 } 621 622 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 623 // It won't be much cheaper than bailing to the interp., since we'll 624 // have to pass up all the debug-info, and the runtime will have to 625 // create the stack trace. 626 627 // Usual case: Bail to interpreter. 628 // Reserve the right to recompile if we haven't seen anything yet. 629 630 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 631 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 632 if (treat_throw_as_hot 633 && (method()->method_data()->trap_recompiled_at(bci(), m) 634 || C->too_many_traps(reason))) { 635 // We cannot afford to take more traps here. Suffer in the interpreter. 636 if (C->log() != NULL) 637 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 638 Deoptimization::trap_reason_name(reason), 639 C->trap_count(reason)); 640 action = Deoptimization::Action_none; 641 } 642 643 // "must_throw" prunes the JVM state to include only the stack, if there 644 // are no local exception handlers. This should cut down on register 645 // allocation time and code size, by drastically reducing the number 646 // of in-edges on the call to the uncommon trap. 647 648 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 649 } 650 651 652 //----------------------------PreserveJVMState--------------------------------- 653 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 654 debug_only(kit->verify_map()); 655 _kit = kit; 656 _map = kit->map(); // preserve the map 657 _sp = kit->sp(); 658 kit->set_map(clone_map ? kit->clone_map() : NULL); 659 #ifdef ASSERT 660 _bci = kit->bci(); 661 Parse* parser = kit->is_Parse(); 662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 663 _block = block; 664 #endif 665 } 666 PreserveJVMState::~PreserveJVMState() { 667 GraphKit* kit = _kit; 668 #ifdef ASSERT 669 assert(kit->bci() == _bci, "bci must not shift"); 670 Parse* parser = kit->is_Parse(); 671 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 672 assert(block == _block, "block must not shift"); 673 #endif 674 kit->set_map(_map); 675 kit->set_sp(_sp); 676 } 677 678 679 //-----------------------------BuildCutout------------------------------------- 680 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 681 : PreserveJVMState(kit) 682 { 683 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 684 SafePointNode* outer_map = _map; // preserved map is caller's 685 SafePointNode* inner_map = kit->map(); 686 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 687 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 688 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 689 } 690 BuildCutout::~BuildCutout() { 691 GraphKit* kit = _kit; 692 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 693 } 694 695 //---------------------------PreserveReexecuteState---------------------------- 696 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 697 assert(!kit->stopped(), "must call stopped() before"); 698 _kit = kit; 699 _sp = kit->sp(); 700 _reexecute = kit->jvms()->_reexecute; 701 } 702 PreserveReexecuteState::~PreserveReexecuteState() { 703 if (_kit->stopped()) return; 704 _kit->jvms()->_reexecute = _reexecute; 705 _kit->set_sp(_sp); 706 } 707 708 //------------------------------clone_map-------------------------------------- 709 // Implementation of PreserveJVMState 710 // 711 // Only clone_map(...) here. If this function is only used in the 712 // PreserveJVMState class we may want to get rid of this extra 713 // function eventually and do it all there. 714 715 SafePointNode* GraphKit::clone_map() { 716 if (map() == NULL) return NULL; 717 718 // Clone the memory edge first 719 Node* mem = MergeMemNode::make(map()->memory()); 720 gvn().set_type_bottom(mem); 721 722 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 723 JVMState* jvms = this->jvms(); 724 JVMState* clonejvms = jvms->clone_shallow(C); 725 clonemap->set_memory(mem); 726 clonemap->set_jvms(clonejvms); 727 clonejvms->set_map(clonemap); 728 record_for_igvn(clonemap); 729 gvn().set_type_bottom(clonemap); 730 return clonemap; 731 } 732 733 734 //-----------------------------set_map_clone----------------------------------- 735 void GraphKit::set_map_clone(SafePointNode* m) { 736 _map = m; 737 _map = clone_map(); 738 _map->set_next_exception(NULL); 739 debug_only(verify_map()); 740 } 741 742 743 //----------------------------kill_dead_locals--------------------------------- 744 // Detect any locals which are known to be dead, and force them to top. 745 void GraphKit::kill_dead_locals() { 746 // Consult the liveness information for the locals. If any 747 // of them are unused, then they can be replaced by top(). This 748 // should help register allocation time and cut down on the size 749 // of the deoptimization information. 750 751 // This call is made from many of the bytecode handling 752 // subroutines called from the Big Switch in do_one_bytecode. 753 // Every bytecode which might include a slow path is responsible 754 // for killing its dead locals. The more consistent we 755 // are about killing deads, the fewer useless phis will be 756 // constructed for them at various merge points. 757 758 // bci can be -1 (InvocationEntryBci). We return the entry 759 // liveness for the method. 760 761 if (method() == NULL || method()->code_size() == 0) { 762 // We are building a graph for a call to a native method. 763 // All locals are live. 764 return; 765 } 766 767 ResourceMark rm; 768 769 // Consult the liveness information for the locals. If any 770 // of them are unused, then they can be replaced by top(). This 771 // should help register allocation time and cut down on the size 772 // of the deoptimization information. 773 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 774 775 int len = (int)live_locals.size(); 776 assert(len <= jvms()->loc_size(), "too many live locals"); 777 for (int local = 0; local < len; local++) { 778 if (!live_locals.at(local)) { 779 set_local(local, top()); 780 } 781 } 782 } 783 784 #ifdef ASSERT 785 //-------------------------dead_locals_are_killed------------------------------ 786 // Return true if all dead locals are set to top in the map. 787 // Used to assert "clean" debug info at various points. 788 bool GraphKit::dead_locals_are_killed() { 789 if (method() == NULL || method()->code_size() == 0) { 790 // No locals need to be dead, so all is as it should be. 791 return true; 792 } 793 794 // Make sure somebody called kill_dead_locals upstream. 795 ResourceMark rm; 796 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 797 if (jvms->loc_size() == 0) continue; // no locals to consult 798 SafePointNode* map = jvms->map(); 799 ciMethod* method = jvms->method(); 800 int bci = jvms->bci(); 801 if (jvms == this->jvms()) { 802 bci = this->bci(); // it might not yet be synched 803 } 804 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 805 int len = (int)live_locals.size(); 806 if (!live_locals.is_valid() || len == 0) 807 // This method is trivial, or is poisoned by a breakpoint. 808 return true; 809 assert(len == jvms->loc_size(), "live map consistent with locals map"); 810 for (int local = 0; local < len; local++) { 811 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 812 if (PrintMiscellaneous && (Verbose || WizardMode)) { 813 tty->print_cr("Zombie local %d: ", local); 814 jvms->dump(); 815 } 816 return false; 817 } 818 } 819 } 820 return true; 821 } 822 823 #endif //ASSERT 824 825 // Helper function for enforcing certain bytecodes to reexecute if 826 // deoptimization happens 827 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 828 ciMethod* cur_method = jvms->method(); 829 int cur_bci = jvms->bci(); 830 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 831 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 832 return Interpreter::bytecode_should_reexecute(code) || 833 (is_anewarray && code == Bytecodes::_multianewarray); 834 // Reexecute _multianewarray bytecode which was replaced with 835 // sequence of [a]newarray. See Parse::do_multianewarray(). 836 // 837 // Note: interpreter should not have it set since this optimization 838 // is limited by dimensions and guarded by flag so in some cases 839 // multianewarray() runtime calls will be generated and 840 // the bytecode should not be reexecutes (stack will not be reset). 841 } else 842 return false; 843 } 844 845 // Helper function for adding JVMState and debug information to node 846 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 847 // Add the safepoint edges to the call (or other safepoint). 848 849 // Make sure dead locals are set to top. This 850 // should help register allocation time and cut down on the size 851 // of the deoptimization information. 852 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 853 854 // Walk the inline list to fill in the correct set of JVMState's 855 // Also fill in the associated edges for each JVMState. 856 857 // If the bytecode needs to be reexecuted we need to put 858 // the arguments back on the stack. 859 const bool should_reexecute = jvms()->should_reexecute(); 860 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 861 862 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 863 // undefined if the bci is different. This is normal for Parse but it 864 // should not happen for LibraryCallKit because only one bci is processed. 865 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 866 "in LibraryCallKit the reexecute bit should not change"); 867 868 // If we are guaranteed to throw, we can prune everything but the 869 // input to the current bytecode. 870 bool can_prune_locals = false; 871 uint stack_slots_not_pruned = 0; 872 int inputs = 0, depth = 0; 873 if (must_throw) { 874 assert(method() == youngest_jvms->method(), "sanity"); 875 if (compute_stack_effects(inputs, depth)) { 876 can_prune_locals = true; 877 stack_slots_not_pruned = inputs; 878 } 879 } 880 881 if (env()->should_retain_local_variables()) { 882 // At any safepoint, this method can get breakpointed, which would 883 // then require an immediate deoptimization. 884 can_prune_locals = false; // do not prune locals 885 stack_slots_not_pruned = 0; 886 } 887 888 // do not scribble on the input jvms 889 JVMState* out_jvms = youngest_jvms->clone_deep(C); 890 call->set_jvms(out_jvms); // Start jvms list for call node 891 892 // For a known set of bytecodes, the interpreter should reexecute them if 893 // deoptimization happens. We set the reexecute state for them here 894 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 895 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 896 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 897 } 898 899 // Presize the call: 900 DEBUG_ONLY(uint non_debug_edges = call->req()); 901 call->add_req_batch(top(), youngest_jvms->debug_depth()); 902 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 903 904 // Set up edges so that the call looks like this: 905 // Call [state:] ctl io mem fptr retadr 906 // [parms:] parm0 ... parmN 907 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 908 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 909 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 910 // Note that caller debug info precedes callee debug info. 911 912 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 913 uint debug_ptr = call->req(); 914 915 // Loop over the map input edges associated with jvms, add them 916 // to the call node, & reset all offsets to match call node array. 917 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 918 uint debug_end = debug_ptr; 919 uint debug_start = debug_ptr - in_jvms->debug_size(); 920 debug_ptr = debug_start; // back up the ptr 921 922 uint p = debug_start; // walks forward in [debug_start, debug_end) 923 uint j, k, l; 924 SafePointNode* in_map = in_jvms->map(); 925 out_jvms->set_map(call); 926 927 if (can_prune_locals) { 928 assert(in_jvms->method() == out_jvms->method(), "sanity"); 929 // If the current throw can reach an exception handler in this JVMS, 930 // then we must keep everything live that can reach that handler. 931 // As a quick and dirty approximation, we look for any handlers at all. 932 if (in_jvms->method()->has_exception_handlers()) { 933 can_prune_locals = false; 934 } 935 } 936 937 // Add the Locals 938 k = in_jvms->locoff(); 939 l = in_jvms->loc_size(); 940 out_jvms->set_locoff(p); 941 if (!can_prune_locals) { 942 for (j = 0; j < l; j++) 943 call->set_req(p++, in_map->in(k+j)); 944 } else { 945 p += l; // already set to top above by add_req_batch 946 } 947 948 // Add the Expression Stack 949 k = in_jvms->stkoff(); 950 l = in_jvms->sp(); 951 out_jvms->set_stkoff(p); 952 if (!can_prune_locals) { 953 for (j = 0; j < l; j++) 954 call->set_req(p++, in_map->in(k+j)); 955 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 956 // Divide stack into {S0,...,S1}, where S0 is set to top. 957 uint s1 = stack_slots_not_pruned; 958 stack_slots_not_pruned = 0; // for next iteration 959 if (s1 > l) s1 = l; 960 uint s0 = l - s1; 961 p += s0; // skip the tops preinstalled by add_req_batch 962 for (j = s0; j < l; j++) 963 call->set_req(p++, in_map->in(k+j)); 964 } else { 965 p += l; // already set to top above by add_req_batch 966 } 967 968 // Add the Monitors 969 k = in_jvms->monoff(); 970 l = in_jvms->mon_size(); 971 out_jvms->set_monoff(p); 972 for (j = 0; j < l; j++) 973 call->set_req(p++, in_map->in(k+j)); 974 975 // Copy any scalar object fields. 976 k = in_jvms->scloff(); 977 l = in_jvms->scl_size(); 978 out_jvms->set_scloff(p); 979 for (j = 0; j < l; j++) 980 call->set_req(p++, in_map->in(k+j)); 981 982 // Finish the new jvms. 983 out_jvms->set_endoff(p); 984 985 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 986 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 987 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 988 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 989 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 990 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 991 992 // Update the two tail pointers in parallel. 993 out_jvms = out_jvms->caller(); 994 in_jvms = in_jvms->caller(); 995 } 996 997 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 998 999 // Test the correctness of JVMState::debug_xxx accessors: 1000 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1001 assert(call->jvms()->debug_end() == call->req(), ""); 1002 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1003 } 1004 1005 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1006 Bytecodes::Code code = java_bc(); 1007 if (code == Bytecodes::_wide) { 1008 code = method()->java_code_at_bci(bci() + 1); 1009 } 1010 1011 BasicType rtype = T_ILLEGAL; 1012 int rsize = 0; 1013 1014 if (code != Bytecodes::_illegal) { 1015 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1016 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1017 if (rtype < T_CONFLICT) 1018 rsize = type2size[rtype]; 1019 } 1020 1021 switch (code) { 1022 case Bytecodes::_illegal: 1023 return false; 1024 1025 case Bytecodes::_ldc: 1026 case Bytecodes::_ldc_w: 1027 case Bytecodes::_ldc2_w: 1028 inputs = 0; 1029 break; 1030 1031 case Bytecodes::_dup: inputs = 1; break; 1032 case Bytecodes::_dup_x1: inputs = 2; break; 1033 case Bytecodes::_dup_x2: inputs = 3; break; 1034 case Bytecodes::_dup2: inputs = 2; break; 1035 case Bytecodes::_dup2_x1: inputs = 3; break; 1036 case Bytecodes::_dup2_x2: inputs = 4; break; 1037 case Bytecodes::_swap: inputs = 2; break; 1038 case Bytecodes::_arraylength: inputs = 1; break; 1039 1040 case Bytecodes::_getstatic: 1041 case Bytecodes::_putstatic: 1042 case Bytecodes::_getfield: 1043 case Bytecodes::_putfield: 1044 { 1045 bool ignored_will_link; 1046 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1047 int size = field->type()->size(); 1048 bool is_get = (depth >= 0), is_static = (depth & 1); 1049 inputs = (is_static ? 0 : 1); 1050 if (is_get) { 1051 depth = size - inputs; 1052 } else { 1053 inputs += size; // putxxx pops the value from the stack 1054 depth = - inputs; 1055 } 1056 } 1057 break; 1058 1059 case Bytecodes::_invokevirtual: 1060 case Bytecodes::_invokespecial: 1061 case Bytecodes::_invokestatic: 1062 case Bytecodes::_invokedynamic: 1063 case Bytecodes::_invokeinterface: 1064 { 1065 bool ignored_will_link; 1066 ciSignature* declared_signature = NULL; 1067 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1068 assert(declared_signature != NULL, "cannot be null"); 1069 inputs = declared_signature->arg_size_for_bc(code); 1070 int size = declared_signature->return_type()->size(); 1071 depth = size - inputs; 1072 } 1073 break; 1074 1075 case Bytecodes::_multianewarray: 1076 { 1077 ciBytecodeStream iter(method()); 1078 iter.reset_to_bci(bci()); 1079 iter.next(); 1080 inputs = iter.get_dimensions(); 1081 assert(rsize == 1, ""); 1082 depth = rsize - inputs; 1083 } 1084 break; 1085 1086 case Bytecodes::_ireturn: 1087 case Bytecodes::_lreturn: 1088 case Bytecodes::_freturn: 1089 case Bytecodes::_dreturn: 1090 case Bytecodes::_areturn: 1091 assert(rsize == -depth, ""); 1092 inputs = rsize; 1093 break; 1094 1095 case Bytecodes::_jsr: 1096 case Bytecodes::_jsr_w: 1097 inputs = 0; 1098 depth = 1; // S.B. depth=1, not zero 1099 break; 1100 1101 default: 1102 // bytecode produces a typed result 1103 inputs = rsize - depth; 1104 assert(inputs >= 0, ""); 1105 break; 1106 } 1107 1108 #ifdef ASSERT 1109 // spot check 1110 int outputs = depth + inputs; 1111 assert(outputs >= 0, "sanity"); 1112 switch (code) { 1113 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1114 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1115 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1116 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1117 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1118 default: break; 1119 } 1120 #endif //ASSERT 1121 1122 return true; 1123 } 1124 1125 1126 1127 //------------------------------basic_plus_adr--------------------------------- 1128 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1129 // short-circuit a common case 1130 if (offset == intcon(0)) return ptr; 1131 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1132 } 1133 1134 Node* GraphKit::ConvI2L(Node* offset) { 1135 // short-circuit a common case 1136 jint offset_con = find_int_con(offset, Type::OffsetBot); 1137 if (offset_con != Type::OffsetBot) { 1138 return longcon((jlong) offset_con); 1139 } 1140 return _gvn.transform( new ConvI2LNode(offset)); 1141 } 1142 1143 Node* GraphKit::ConvI2UL(Node* offset) { 1144 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1145 if (offset_con != (juint) Type::OffsetBot) { 1146 return longcon((julong) offset_con); 1147 } 1148 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1149 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1150 return _gvn.transform( new AndLNode(conv, mask) ); 1151 } 1152 1153 Node* GraphKit::ConvL2I(Node* offset) { 1154 // short-circuit a common case 1155 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1156 if (offset_con != (jlong)Type::OffsetBot) { 1157 return intcon((int) offset_con); 1158 } 1159 return _gvn.transform( new ConvL2INode(offset)); 1160 } 1161 1162 //-------------------------load_object_klass----------------------------------- 1163 Node* GraphKit::load_object_klass(Node* obj) { 1164 // Special-case a fresh allocation to avoid building nodes: 1165 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1166 if (akls != NULL) return akls; 1167 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1168 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1169 } 1170 1171 //-------------------------load_array_length----------------------------------- 1172 Node* GraphKit::load_array_length(Node* array) { 1173 // Special-case a fresh allocation to avoid building nodes: 1174 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1175 Node *alen; 1176 if (alloc == NULL) { 1177 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1178 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1179 } else { 1180 alen = alloc->Ideal_length(); 1181 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1182 if (ccast != alen) { 1183 alen = _gvn.transform(ccast); 1184 } 1185 } 1186 return alen; 1187 } 1188 1189 //------------------------------do_null_check---------------------------------- 1190 // Helper function to do a NULL pointer check. Returned value is 1191 // the incoming address with NULL casted away. You are allowed to use the 1192 // not-null value only if you are control dependent on the test. 1193 #ifndef PRODUCT 1194 extern int explicit_null_checks_inserted, 1195 explicit_null_checks_elided; 1196 #endif 1197 Node* GraphKit::null_check_common(Node* value, BasicType type, 1198 // optional arguments for variations: 1199 bool assert_null, 1200 Node* *null_control, 1201 bool speculative) { 1202 assert(!assert_null || null_control == NULL, "not both at once"); 1203 if (stopped()) return top(); 1204 NOT_PRODUCT(explicit_null_checks_inserted++); 1205 1206 // Construct NULL check 1207 Node *chk = NULL; 1208 switch(type) { 1209 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1210 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1211 case T_ARRAY : // fall through 1212 type = T_OBJECT; // simplify further tests 1213 case T_OBJECT : { 1214 const Type *t = _gvn.type( value ); 1215 1216 const TypeOopPtr* tp = t->isa_oopptr(); 1217 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1218 // Only for do_null_check, not any of its siblings: 1219 && !assert_null && null_control == NULL) { 1220 // Usually, any field access or invocation on an unloaded oop type 1221 // will simply fail to link, since the statically linked class is 1222 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1223 // the static class is loaded but the sharper oop type is not. 1224 // Rather than checking for this obscure case in lots of places, 1225 // we simply observe that a null check on an unloaded class 1226 // will always be followed by a nonsense operation, so we 1227 // can just issue the uncommon trap here. 1228 // Our access to the unloaded class will only be correct 1229 // after it has been loaded and initialized, which requires 1230 // a trip through the interpreter. 1231 #ifndef PRODUCT 1232 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1233 #endif 1234 uncommon_trap(Deoptimization::Reason_unloaded, 1235 Deoptimization::Action_reinterpret, 1236 tp->klass(), "!loaded"); 1237 return top(); 1238 } 1239 1240 if (assert_null) { 1241 // See if the type is contained in NULL_PTR. 1242 // If so, then the value is already null. 1243 if (t->higher_equal(TypePtr::NULL_PTR)) { 1244 NOT_PRODUCT(explicit_null_checks_elided++); 1245 return value; // Elided null assert quickly! 1246 } 1247 } else { 1248 // See if mixing in the NULL pointer changes type. 1249 // If so, then the NULL pointer was not allowed in the original 1250 // type. In other words, "value" was not-null. 1251 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1252 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1253 NOT_PRODUCT(explicit_null_checks_elided++); 1254 return value; // Elided null check quickly! 1255 } 1256 } 1257 chk = new CmpPNode( value, null() ); 1258 break; 1259 } 1260 1261 default: 1262 fatal("unexpected type: %s", type2name(type)); 1263 } 1264 assert(chk != NULL, "sanity check"); 1265 chk = _gvn.transform(chk); 1266 1267 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1268 BoolNode *btst = new BoolNode( chk, btest); 1269 Node *tst = _gvn.transform( btst ); 1270 1271 //----------- 1272 // if peephole optimizations occurred, a prior test existed. 1273 // If a prior test existed, maybe it dominates as we can avoid this test. 1274 if (tst != btst && type == T_OBJECT) { 1275 // At this point we want to scan up the CFG to see if we can 1276 // find an identical test (and so avoid this test altogether). 1277 Node *cfg = control(); 1278 int depth = 0; 1279 while( depth < 16 ) { // Limit search depth for speed 1280 if( cfg->Opcode() == Op_IfTrue && 1281 cfg->in(0)->in(1) == tst ) { 1282 // Found prior test. Use "cast_not_null" to construct an identical 1283 // CastPP (and hence hash to) as already exists for the prior test. 1284 // Return that casted value. 1285 if (assert_null) { 1286 replace_in_map(value, null()); 1287 return null(); // do not issue the redundant test 1288 } 1289 Node *oldcontrol = control(); 1290 set_control(cfg); 1291 Node *res = cast_not_null(value); 1292 set_control(oldcontrol); 1293 NOT_PRODUCT(explicit_null_checks_elided++); 1294 return res; 1295 } 1296 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1297 if (cfg == NULL) break; // Quit at region nodes 1298 depth++; 1299 } 1300 } 1301 1302 //----------- 1303 // Branch to failure if null 1304 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1305 Deoptimization::DeoptReason reason; 1306 if (assert_null) { 1307 reason = Deoptimization::reason_null_assert(speculative); 1308 } else if (type == T_OBJECT) { 1309 reason = Deoptimization::reason_null_check(speculative); 1310 } else { 1311 reason = Deoptimization::Reason_div0_check; 1312 } 1313 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1314 // ciMethodData::has_trap_at will return a conservative -1 if any 1315 // must-be-null assertion has failed. This could cause performance 1316 // problems for a method after its first do_null_assert failure. 1317 // Consider using 'Reason_class_check' instead? 1318 1319 // To cause an implicit null check, we set the not-null probability 1320 // to the maximum (PROB_MAX). For an explicit check the probability 1321 // is set to a smaller value. 1322 if (null_control != NULL || too_many_traps(reason)) { 1323 // probability is less likely 1324 ok_prob = PROB_LIKELY_MAG(3); 1325 } else if (!assert_null && 1326 (ImplicitNullCheckThreshold > 0) && 1327 method() != NULL && 1328 (method()->method_data()->trap_count(reason) 1329 >= (uint)ImplicitNullCheckThreshold)) { 1330 ok_prob = PROB_LIKELY_MAG(3); 1331 } 1332 1333 if (null_control != NULL) { 1334 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1335 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1336 set_control( _gvn.transform( new IfTrueNode(iff))); 1337 #ifndef PRODUCT 1338 if (null_true == top()) { 1339 explicit_null_checks_elided++; 1340 } 1341 #endif 1342 (*null_control) = null_true; 1343 } else { 1344 BuildCutout unless(this, tst, ok_prob); 1345 // Check for optimizer eliding test at parse time 1346 if (stopped()) { 1347 // Failure not possible; do not bother making uncommon trap. 1348 NOT_PRODUCT(explicit_null_checks_elided++); 1349 } else if (assert_null) { 1350 uncommon_trap(reason, 1351 Deoptimization::Action_make_not_entrant, 1352 NULL, "assert_null"); 1353 } else { 1354 replace_in_map(value, zerocon(type)); 1355 builtin_throw(reason); 1356 } 1357 } 1358 1359 // Must throw exception, fall-thru not possible? 1360 if (stopped()) { 1361 return top(); // No result 1362 } 1363 1364 if (assert_null) { 1365 // Cast obj to null on this path. 1366 replace_in_map(value, zerocon(type)); 1367 return zerocon(type); 1368 } 1369 1370 // Cast obj to not-null on this path, if there is no null_control. 1371 // (If there is a null_control, a non-null value may come back to haunt us.) 1372 if (type == T_OBJECT) { 1373 Node* cast = cast_not_null(value, false); 1374 if (null_control == NULL || (*null_control) == top()) 1375 replace_in_map(value, cast); 1376 value = cast; 1377 } 1378 1379 return value; 1380 } 1381 1382 1383 //------------------------------cast_not_null---------------------------------- 1384 // Cast obj to not-null on this path 1385 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1386 const Type *t = _gvn.type(obj); 1387 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1388 // Object is already not-null? 1389 if( t == t_not_null ) return obj; 1390 1391 Node *cast = new CastPPNode(obj,t_not_null); 1392 cast->init_req(0, control()); 1393 cast = _gvn.transform( cast ); 1394 1395 // Scan for instances of 'obj' in the current JVM mapping. 1396 // These instances are known to be not-null after the test. 1397 if (do_replace_in_map) 1398 replace_in_map(obj, cast); 1399 1400 return cast; // Return casted value 1401 } 1402 1403 // Sometimes in intrinsics, we implicitly know an object is not null 1404 // (there's no actual null check) so we can cast it to not null. In 1405 // the course of optimizations, the input to the cast can become null. 1406 // In that case that data path will die and we need the control path 1407 // to become dead as well to keep the graph consistent. So we have to 1408 // add a check for null for which one branch can't be taken. It uses 1409 // an Opaque4 node that will cause the check to be removed after loop 1410 // opts so the test goes away and the compiled code doesn't execute a 1411 // useless check. 1412 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1413 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1414 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1415 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1416 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1417 _gvn.set_type(iff, iff->Value(&_gvn)); 1418 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1419 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1420 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1421 C->root()->add_req(halt); 1422 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1423 set_control(if_t); 1424 return cast_not_null(value, do_replace_in_map); 1425 } 1426 1427 1428 //--------------------------replace_in_map------------------------------------- 1429 void GraphKit::replace_in_map(Node* old, Node* neww) { 1430 if (old == neww) { 1431 return; 1432 } 1433 1434 map()->replace_edge(old, neww); 1435 1436 // Note: This operation potentially replaces any edge 1437 // on the map. This includes locals, stack, and monitors 1438 // of the current (innermost) JVM state. 1439 1440 // don't let inconsistent types from profiling escape this 1441 // method 1442 1443 const Type* told = _gvn.type(old); 1444 const Type* tnew = _gvn.type(neww); 1445 1446 if (!tnew->higher_equal(told)) { 1447 return; 1448 } 1449 1450 map()->record_replaced_node(old, neww); 1451 } 1452 1453 1454 //============================================================================= 1455 //--------------------------------memory--------------------------------------- 1456 Node* GraphKit::memory(uint alias_idx) { 1457 MergeMemNode* mem = merged_memory(); 1458 Node* p = mem->memory_at(alias_idx); 1459 _gvn.set_type(p, Type::MEMORY); // must be mapped 1460 return p; 1461 } 1462 1463 //-----------------------------reset_memory------------------------------------ 1464 Node* GraphKit::reset_memory() { 1465 Node* mem = map()->memory(); 1466 // do not use this node for any more parsing! 1467 debug_only( map()->set_memory((Node*)NULL) ); 1468 return _gvn.transform( mem ); 1469 } 1470 1471 //------------------------------set_all_memory--------------------------------- 1472 void GraphKit::set_all_memory(Node* newmem) { 1473 Node* mergemem = MergeMemNode::make(newmem); 1474 gvn().set_type_bottom(mergemem); 1475 map()->set_memory(mergemem); 1476 } 1477 1478 //------------------------------set_all_memory_call---------------------------- 1479 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1480 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1481 set_all_memory(newmem); 1482 } 1483 1484 //============================================================================= 1485 // 1486 // parser factory methods for MemNodes 1487 // 1488 // These are layered on top of the factory methods in LoadNode and StoreNode, 1489 // and integrate with the parser's memory state and _gvn engine. 1490 // 1491 1492 // factory methods in "int adr_idx" 1493 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1494 int adr_idx, 1495 MemNode::MemOrd mo, 1496 LoadNode::ControlDependency control_dependency, 1497 bool require_atomic_access, 1498 bool unaligned, 1499 bool mismatched) { 1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1501 const TypePtr* adr_type = NULL; // debug-mode-only argument 1502 debug_only(adr_type = C->get_adr_type(adr_idx)); 1503 Node* mem = memory(adr_idx); 1504 Node* ld; 1505 if (require_atomic_access && bt == T_LONG) { 1506 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1507 } else if (require_atomic_access && bt == T_DOUBLE) { 1508 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1509 } else { 1510 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1511 } 1512 ld = _gvn.transform(ld); 1513 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1514 // Improve graph before escape analysis and boxing elimination. 1515 record_for_igvn(ld); 1516 } 1517 return ld; 1518 } 1519 1520 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1521 int adr_idx, 1522 MemNode::MemOrd mo, 1523 bool require_atomic_access, 1524 bool unaligned, 1525 bool mismatched) { 1526 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1527 const TypePtr* adr_type = NULL; 1528 debug_only(adr_type = C->get_adr_type(adr_idx)); 1529 Node *mem = memory(adr_idx); 1530 Node* st; 1531 if (require_atomic_access && bt == T_LONG) { 1532 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1533 } else if (require_atomic_access && bt == T_DOUBLE) { 1534 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1535 } else { 1536 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1537 } 1538 if (unaligned) { 1539 st->as_Store()->set_unaligned_access(); 1540 } 1541 if (mismatched) { 1542 st->as_Store()->set_mismatched_access(); 1543 } 1544 st = _gvn.transform(st); 1545 set_memory(st, adr_idx); 1546 // Back-to-back stores can only remove intermediate store with DU info 1547 // so push on worklist for optimizer. 1548 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1549 record_for_igvn(st); 1550 1551 return st; 1552 } 1553 1554 1555 void GraphKit::pre_barrier(bool do_load, 1556 Node* ctl, 1557 Node* obj, 1558 Node* adr, 1559 uint adr_idx, 1560 Node* val, 1561 const TypeOopPtr* val_type, 1562 Node* pre_val, 1563 BasicType bt) { 1564 1565 BarrierSet* bs = BarrierSet::barrier_set(); 1566 set_control(ctl); 1567 switch (bs->kind()) { 1568 case BarrierSet::G1BarrierSet: 1569 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1570 break; 1571 1572 case BarrierSet::CardTableBarrierSet: 1573 break; 1574 1575 default : 1576 ShouldNotReachHere(); 1577 1578 } 1579 } 1580 1581 bool GraphKit::can_move_pre_barrier() const { 1582 BarrierSet* bs = BarrierSet::barrier_set(); 1583 switch (bs->kind()) { 1584 case BarrierSet::G1BarrierSet: 1585 return true; // Can move it if no safepoint 1586 1587 case BarrierSet::CardTableBarrierSet: 1588 return true; // There is no pre-barrier 1589 1590 default : 1591 ShouldNotReachHere(); 1592 } 1593 return false; 1594 } 1595 1596 void GraphKit::post_barrier(Node* ctl, 1597 Node* store, 1598 Node* obj, 1599 Node* adr, 1600 uint adr_idx, 1601 Node* val, 1602 BasicType bt, 1603 bool use_precise) { 1604 BarrierSet* bs = BarrierSet::barrier_set(); 1605 set_control(ctl); 1606 switch (bs->kind()) { 1607 case BarrierSet::G1BarrierSet: 1608 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1609 break; 1610 1611 case BarrierSet::CardTableBarrierSet: 1612 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1613 break; 1614 1615 default : 1616 ShouldNotReachHere(); 1617 1618 } 1619 } 1620 1621 Node* GraphKit::store_oop(Node* ctl, 1622 Node* obj, 1623 Node* adr, 1624 const TypePtr* adr_type, 1625 Node* val, 1626 const TypeOopPtr* val_type, 1627 BasicType bt, 1628 bool use_precise, 1629 MemNode::MemOrd mo, 1630 bool mismatched) { 1631 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1632 // could be delayed during Parse (for example, in adjust_map_after_if()). 1633 // Execute transformation here to avoid barrier generation in such case. 1634 if (_gvn.type(val) == TypePtr::NULL_PTR) 1635 val = _gvn.makecon(TypePtr::NULL_PTR); 1636 1637 set_control(ctl); 1638 if (stopped()) return top(); // Dead path ? 1639 1640 assert(bt == T_OBJECT, "sanity"); 1641 assert(val != NULL, "not dead path"); 1642 uint adr_idx = C->get_alias_index(adr_type); 1643 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1644 1645 pre_barrier(true /* do_load */, 1646 control(), obj, adr, adr_idx, val, val_type, 1647 NULL /* pre_val */, 1648 bt); 1649 1650 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1651 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1652 return store; 1653 } 1654 1655 // Could be an array or object we don't know at compile time (unsafe ref.) 1656 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1657 Node* obj, // containing obj 1658 Node* adr, // actual adress to store val at 1659 const TypePtr* adr_type, 1660 Node* val, 1661 BasicType bt, 1662 MemNode::MemOrd mo, 1663 bool mismatched) { 1664 Compile::AliasType* at = C->alias_type(adr_type); 1665 const TypeOopPtr* val_type = NULL; 1666 if (adr_type->isa_instptr()) { 1667 if (at->field() != NULL) { 1668 // known field. This code is a copy of the do_put_xxx logic. 1669 ciField* field = at->field(); 1670 if (!field->type()->is_loaded()) { 1671 val_type = TypeInstPtr::BOTTOM; 1672 } else { 1673 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1674 } 1675 } 1676 } else if (adr_type->isa_aryptr()) { 1677 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1678 } 1679 if (val_type == NULL) { 1680 val_type = TypeInstPtr::BOTTOM; 1681 } 1682 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1683 } 1684 1685 1686 //-------------------------array_element_address------------------------- 1687 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1688 const TypeInt* sizetype, Node* ctrl) { 1689 uint shift = exact_log2(type2aelembytes(elembt)); 1690 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1691 1692 // short-circuit a common case (saves lots of confusing waste motion) 1693 jint idx_con = find_int_con(idx, -1); 1694 if (idx_con >= 0) { 1695 intptr_t offset = header + ((intptr_t)idx_con << shift); 1696 return basic_plus_adr(ary, offset); 1697 } 1698 1699 // must be correct type for alignment purposes 1700 Node* base = basic_plus_adr(ary, header); 1701 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1702 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1703 return basic_plus_adr(ary, base, scale); 1704 } 1705 1706 //-------------------------load_array_element------------------------- 1707 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1708 const Type* elemtype = arytype->elem(); 1709 BasicType elembt = elemtype->array_element_basic_type(); 1710 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1711 if (elembt == T_NARROWOOP) { 1712 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1713 } 1714 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1715 return ld; 1716 } 1717 1718 //-------------------------set_arguments_for_java_call------------------------- 1719 // Arguments (pre-popped from the stack) are taken from the JVMS. 1720 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1721 // Add the call arguments: 1722 uint nargs = call->method()->arg_size(); 1723 for (uint i = 0; i < nargs; i++) { 1724 Node* arg = argument(i); 1725 call->init_req(i + TypeFunc::Parms, arg); 1726 } 1727 } 1728 1729 //---------------------------set_edges_for_java_call--------------------------- 1730 // Connect a newly created call into the current JVMS. 1731 // A return value node (if any) is returned from set_edges_for_java_call. 1732 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1733 1734 // Add the predefined inputs: 1735 call->init_req( TypeFunc::Control, control() ); 1736 call->init_req( TypeFunc::I_O , i_o() ); 1737 call->init_req( TypeFunc::Memory , reset_memory() ); 1738 call->init_req( TypeFunc::FramePtr, frameptr() ); 1739 call->init_req( TypeFunc::ReturnAdr, top() ); 1740 1741 add_safepoint_edges(call, must_throw); 1742 1743 Node* xcall = _gvn.transform(call); 1744 1745 if (xcall == top()) { 1746 set_control(top()); 1747 return; 1748 } 1749 assert(xcall == call, "call identity is stable"); 1750 1751 // Re-use the current map to produce the result. 1752 1753 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1754 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1755 set_all_memory_call(xcall, separate_io_proj); 1756 1757 //return xcall; // no need, caller already has it 1758 } 1759 1760 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1761 if (stopped()) return top(); // maybe the call folded up? 1762 1763 // Capture the return value, if any. 1764 Node* ret; 1765 if (call->method() == NULL || 1766 call->method()->return_type()->basic_type() == T_VOID) 1767 ret = top(); 1768 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1769 1770 // Note: Since any out-of-line call can produce an exception, 1771 // we always insert an I_O projection from the call into the result. 1772 1773 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1774 1775 if (separate_io_proj) { 1776 // The caller requested separate projections be used by the fall 1777 // through and exceptional paths, so replace the projections for 1778 // the fall through path. 1779 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1780 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1781 } 1782 return ret; 1783 } 1784 1785 //--------------------set_predefined_input_for_runtime_call-------------------- 1786 // Reading and setting the memory state is way conservative here. 1787 // The real problem is that I am not doing real Type analysis on memory, 1788 // so I cannot distinguish card mark stores from other stores. Across a GC 1789 // point the Store Barrier and the card mark memory has to agree. I cannot 1790 // have a card mark store and its barrier split across the GC point from 1791 // either above or below. Here I get that to happen by reading ALL of memory. 1792 // A better answer would be to separate out card marks from other memory. 1793 // For now, return the input memory state, so that it can be reused 1794 // after the call, if this call has restricted memory effects. 1795 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1796 // Set fixed predefined input arguments 1797 Node* memory = reset_memory(); 1798 call->init_req( TypeFunc::Control, control() ); 1799 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1800 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1801 call->init_req( TypeFunc::FramePtr, frameptr() ); 1802 call->init_req( TypeFunc::ReturnAdr, top() ); 1803 return memory; 1804 } 1805 1806 //-------------------set_predefined_output_for_runtime_call-------------------- 1807 // Set control and memory (not i_o) from the call. 1808 // If keep_mem is not NULL, use it for the output state, 1809 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1810 // If hook_mem is NULL, this call produces no memory effects at all. 1811 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1812 // then only that memory slice is taken from the call. 1813 // In the last case, we must put an appropriate memory barrier before 1814 // the call, so as to create the correct anti-dependencies on loads 1815 // preceding the call. 1816 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1817 Node* keep_mem, 1818 const TypePtr* hook_mem) { 1819 // no i/o 1820 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1821 if (keep_mem) { 1822 // First clone the existing memory state 1823 set_all_memory(keep_mem); 1824 if (hook_mem != NULL) { 1825 // Make memory for the call 1826 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1827 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1828 // We also use hook_mem to extract specific effects from arraycopy stubs. 1829 set_memory(mem, hook_mem); 1830 } 1831 // ...else the call has NO memory effects. 1832 1833 // Make sure the call advertises its memory effects precisely. 1834 // This lets us build accurate anti-dependences in gcm.cpp. 1835 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1836 "call node must be constructed correctly"); 1837 } else { 1838 assert(hook_mem == NULL, ""); 1839 // This is not a "slow path" call; all memory comes from the call. 1840 set_all_memory_call(call); 1841 } 1842 } 1843 1844 1845 // Replace the call with the current state of the kit. 1846 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1847 JVMState* ejvms = NULL; 1848 if (has_exceptions()) { 1849 ejvms = transfer_exceptions_into_jvms(); 1850 } 1851 1852 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1853 ReplacedNodes replaced_nodes_exception; 1854 Node* ex_ctl = top(); 1855 1856 SafePointNode* final_state = stop(); 1857 1858 // Find all the needed outputs of this call 1859 CallProjections callprojs; 1860 call->extract_projections(&callprojs, true); 1861 1862 Node* init_mem = call->in(TypeFunc::Memory); 1863 Node* final_mem = final_state->in(TypeFunc::Memory); 1864 Node* final_ctl = final_state->in(TypeFunc::Control); 1865 Node* final_io = final_state->in(TypeFunc::I_O); 1866 1867 // Replace all the old call edges with the edges from the inlining result 1868 if (callprojs.fallthrough_catchproj != NULL) { 1869 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1870 } 1871 if (callprojs.fallthrough_memproj != NULL) { 1872 if (final_mem->is_MergeMem()) { 1873 // Parser's exits MergeMem was not transformed but may be optimized 1874 final_mem = _gvn.transform(final_mem); 1875 } 1876 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1877 } 1878 if (callprojs.fallthrough_ioproj != NULL) { 1879 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1880 } 1881 1882 // Replace the result with the new result if it exists and is used 1883 if (callprojs.resproj != NULL && result != NULL) { 1884 C->gvn_replace_by(callprojs.resproj, result); 1885 } 1886 1887 if (ejvms == NULL) { 1888 // No exception edges to simply kill off those paths 1889 if (callprojs.catchall_catchproj != NULL) { 1890 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1891 } 1892 if (callprojs.catchall_memproj != NULL) { 1893 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1894 } 1895 if (callprojs.catchall_ioproj != NULL) { 1896 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1897 } 1898 // Replace the old exception object with top 1899 if (callprojs.exobj != NULL) { 1900 C->gvn_replace_by(callprojs.exobj, C->top()); 1901 } 1902 } else { 1903 GraphKit ekit(ejvms); 1904 1905 // Load my combined exception state into the kit, with all phis transformed: 1906 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1907 replaced_nodes_exception = ex_map->replaced_nodes(); 1908 1909 Node* ex_oop = ekit.use_exception_state(ex_map); 1910 1911 if (callprojs.catchall_catchproj != NULL) { 1912 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1913 ex_ctl = ekit.control(); 1914 } 1915 if (callprojs.catchall_memproj != NULL) { 1916 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1917 } 1918 if (callprojs.catchall_ioproj != NULL) { 1919 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1920 } 1921 1922 // Replace the old exception object with the newly created one 1923 if (callprojs.exobj != NULL) { 1924 C->gvn_replace_by(callprojs.exobj, ex_oop); 1925 } 1926 } 1927 1928 // Disconnect the call from the graph 1929 call->disconnect_inputs(NULL, C); 1930 C->gvn_replace_by(call, C->top()); 1931 1932 // Clean up any MergeMems that feed other MergeMems since the 1933 // optimizer doesn't like that. 1934 if (final_mem->is_MergeMem()) { 1935 Node_List wl; 1936 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1937 Node* m = i.get(); 1938 if (m->is_MergeMem() && !wl.contains(m)) { 1939 wl.push(m); 1940 } 1941 } 1942 while (wl.size() > 0) { 1943 _gvn.transform(wl.pop()); 1944 } 1945 } 1946 1947 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1948 replaced_nodes.apply(C, final_ctl); 1949 } 1950 if (!ex_ctl->is_top() && do_replaced_nodes) { 1951 replaced_nodes_exception.apply(C, ex_ctl); 1952 } 1953 } 1954 1955 1956 //------------------------------increment_counter------------------------------ 1957 // for statistics: increment a VM counter by 1 1958 1959 void GraphKit::increment_counter(address counter_addr) { 1960 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1961 increment_counter(adr1); 1962 } 1963 1964 void GraphKit::increment_counter(Node* counter_addr) { 1965 int adr_type = Compile::AliasIdxRaw; 1966 Node* ctrl = control(); 1967 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1968 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1969 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1970 } 1971 1972 1973 //------------------------------uncommon_trap---------------------------------- 1974 // Bail out to the interpreter in mid-method. Implemented by calling the 1975 // uncommon_trap blob. This helper function inserts a runtime call with the 1976 // right debug info. 1977 void GraphKit::uncommon_trap(int trap_request, 1978 ciKlass* klass, const char* comment, 1979 bool must_throw, 1980 bool keep_exact_action) { 1981 if (failing()) stop(); 1982 if (stopped()) return; // trap reachable? 1983 1984 // Note: If ProfileTraps is true, and if a deopt. actually 1985 // occurs here, the runtime will make sure an MDO exists. There is 1986 // no need to call method()->ensure_method_data() at this point. 1987 1988 // Set the stack pointer to the right value for reexecution: 1989 set_sp(reexecute_sp()); 1990 1991 #ifdef ASSERT 1992 if (!must_throw) { 1993 // Make sure the stack has at least enough depth to execute 1994 // the current bytecode. 1995 int inputs, ignored_depth; 1996 if (compute_stack_effects(inputs, ignored_depth)) { 1997 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1998 Bytecodes::name(java_bc()), sp(), inputs); 1999 } 2000 } 2001 #endif 2002 2003 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2004 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2005 2006 switch (action) { 2007 case Deoptimization::Action_maybe_recompile: 2008 case Deoptimization::Action_reinterpret: 2009 // Temporary fix for 6529811 to allow virtual calls to be sure they 2010 // get the chance to go from mono->bi->mega 2011 if (!keep_exact_action && 2012 Deoptimization::trap_request_index(trap_request) < 0 && 2013 too_many_recompiles(reason)) { 2014 // This BCI is causing too many recompilations. 2015 if (C->log() != NULL) { 2016 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2017 Deoptimization::trap_reason_name(reason), 2018 Deoptimization::trap_action_name(action)); 2019 } 2020 action = Deoptimization::Action_none; 2021 trap_request = Deoptimization::make_trap_request(reason, action); 2022 } else { 2023 C->set_trap_can_recompile(true); 2024 } 2025 break; 2026 case Deoptimization::Action_make_not_entrant: 2027 C->set_trap_can_recompile(true); 2028 break; 2029 case Deoptimization::Action_none: 2030 case Deoptimization::Action_make_not_compilable: 2031 break; 2032 default: 2033 #ifdef ASSERT 2034 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2035 #endif 2036 break; 2037 } 2038 2039 if (TraceOptoParse) { 2040 char buf[100]; 2041 tty->print_cr("Uncommon trap %s at bci:%d", 2042 Deoptimization::format_trap_request(buf, sizeof(buf), 2043 trap_request), bci()); 2044 } 2045 2046 CompileLog* log = C->log(); 2047 if (log != NULL) { 2048 int kid = (klass == NULL)? -1: log->identify(klass); 2049 log->begin_elem("uncommon_trap bci='%d'", bci()); 2050 char buf[100]; 2051 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2052 trap_request)); 2053 if (kid >= 0) log->print(" klass='%d'", kid); 2054 if (comment != NULL) log->print(" comment='%s'", comment); 2055 log->end_elem(); 2056 } 2057 2058 // Make sure any guarding test views this path as very unlikely 2059 Node *i0 = control()->in(0); 2060 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2061 IfNode *iff = i0->as_If(); 2062 float f = iff->_prob; // Get prob 2063 if (control()->Opcode() == Op_IfTrue) { 2064 if (f > PROB_UNLIKELY_MAG(4)) 2065 iff->_prob = PROB_MIN; 2066 } else { 2067 if (f < PROB_LIKELY_MAG(4)) 2068 iff->_prob = PROB_MAX; 2069 } 2070 } 2071 2072 // Clear out dead values from the debug info. 2073 kill_dead_locals(); 2074 2075 // Now insert the uncommon trap subroutine call 2076 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2077 const TypePtr* no_memory_effects = NULL; 2078 // Pass the index of the class to be loaded 2079 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2080 (must_throw ? RC_MUST_THROW : 0), 2081 OptoRuntime::uncommon_trap_Type(), 2082 call_addr, "uncommon_trap", no_memory_effects, 2083 intcon(trap_request)); 2084 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2085 "must extract request correctly from the graph"); 2086 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2087 2088 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2089 // The debug info is the only real input to this call. 2090 2091 // Halt-and-catch fire here. The above call should never return! 2092 HaltNode* halt = new HaltNode(control(), frameptr()); 2093 _gvn.set_type_bottom(halt); 2094 root()->add_req(halt); 2095 2096 stop_and_kill_map(); 2097 } 2098 2099 2100 //--------------------------just_allocated_object------------------------------ 2101 // Report the object that was just allocated. 2102 // It must be the case that there are no intervening safepoints. 2103 // We use this to determine if an object is so "fresh" that 2104 // it does not require card marks. 2105 Node* GraphKit::just_allocated_object(Node* current_control) { 2106 if (C->recent_alloc_ctl() == current_control) 2107 return C->recent_alloc_obj(); 2108 return NULL; 2109 } 2110 2111 2112 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2113 // (Note: TypeFunc::make has a cache that makes this fast.) 2114 const TypeFunc* tf = TypeFunc::make(dest_method); 2115 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2116 for (int j = 0; j < nargs; j++) { 2117 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2118 if( targ->basic_type() == T_DOUBLE ) { 2119 // If any parameters are doubles, they must be rounded before 2120 // the call, dstore_rounding does gvn.transform 2121 Node *arg = argument(j); 2122 arg = dstore_rounding(arg); 2123 set_argument(j, arg); 2124 } 2125 } 2126 } 2127 2128 /** 2129 * Record profiling data exact_kls for Node n with the type system so 2130 * that it can propagate it (speculation) 2131 * 2132 * @param n node that the type applies to 2133 * @param exact_kls type from profiling 2134 * @param maybe_null did profiling see null? 2135 * 2136 * @return node with improved type 2137 */ 2138 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2139 const Type* current_type = _gvn.type(n); 2140 assert(UseTypeSpeculation, "type speculation must be on"); 2141 2142 const TypePtr* speculative = current_type->speculative(); 2143 2144 // Should the klass from the profile be recorded in the speculative type? 2145 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2146 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2147 const TypeOopPtr* xtype = tklass->as_instance_type(); 2148 assert(xtype->klass_is_exact(), "Should be exact"); 2149 // Any reason to believe n is not null (from this profiling or a previous one)? 2150 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2151 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2152 // record the new speculative type's depth 2153 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2154 speculative = speculative->with_inline_depth(jvms()->depth()); 2155 } else if (current_type->would_improve_ptr(ptr_kind)) { 2156 // Profiling report that null was never seen so we can change the 2157 // speculative type to non null ptr. 2158 if (ptr_kind == ProfileAlwaysNull) { 2159 speculative = TypePtr::NULL_PTR; 2160 } else { 2161 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2162 const TypePtr* ptr = TypePtr::NOTNULL; 2163 if (speculative != NULL) { 2164 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2165 } else { 2166 speculative = ptr; 2167 } 2168 } 2169 } 2170 2171 if (speculative != current_type->speculative()) { 2172 // Build a type with a speculative type (what we think we know 2173 // about the type but will need a guard when we use it) 2174 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2175 // We're changing the type, we need a new CheckCast node to carry 2176 // the new type. The new type depends on the control: what 2177 // profiling tells us is only valid from here as far as we can 2178 // tell. 2179 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2180 cast = _gvn.transform(cast); 2181 replace_in_map(n, cast); 2182 n = cast; 2183 } 2184 2185 return n; 2186 } 2187 2188 /** 2189 * Record profiling data from receiver profiling at an invoke with the 2190 * type system so that it can propagate it (speculation) 2191 * 2192 * @param n receiver node 2193 * 2194 * @return node with improved type 2195 */ 2196 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2197 if (!UseTypeSpeculation) { 2198 return n; 2199 } 2200 ciKlass* exact_kls = profile_has_unique_klass(); 2201 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2202 if ((java_bc() == Bytecodes::_checkcast || 2203 java_bc() == Bytecodes::_instanceof || 2204 java_bc() == Bytecodes::_aastore) && 2205 method()->method_data()->is_mature()) { 2206 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2207 if (data != NULL) { 2208 if (!data->as_BitData()->null_seen()) { 2209 ptr_kind = ProfileNeverNull; 2210 } else { 2211 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2212 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2213 uint i = 0; 2214 for (; i < call->row_limit(); i++) { 2215 ciKlass* receiver = call->receiver(i); 2216 if (receiver != NULL) { 2217 break; 2218 } 2219 } 2220 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2221 } 2222 } 2223 } 2224 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2225 } 2226 2227 /** 2228 * Record profiling data from argument profiling at an invoke with the 2229 * type system so that it can propagate it (speculation) 2230 * 2231 * @param dest_method target method for the call 2232 * @param bc what invoke bytecode is this? 2233 */ 2234 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2235 if (!UseTypeSpeculation) { 2236 return; 2237 } 2238 const TypeFunc* tf = TypeFunc::make(dest_method); 2239 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2240 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2241 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2242 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2243 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2244 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2245 ciKlass* better_type = NULL; 2246 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2247 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2248 } 2249 i++; 2250 } 2251 } 2252 } 2253 2254 /** 2255 * Record profiling data from parameter profiling at an invoke with 2256 * the type system so that it can propagate it (speculation) 2257 */ 2258 void GraphKit::record_profiled_parameters_for_speculation() { 2259 if (!UseTypeSpeculation) { 2260 return; 2261 } 2262 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2263 if (_gvn.type(local(i))->isa_oopptr()) { 2264 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2265 ciKlass* better_type = NULL; 2266 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2267 record_profile_for_speculation(local(i), better_type, ptr_kind); 2268 } 2269 j++; 2270 } 2271 } 2272 } 2273 2274 /** 2275 * Record profiling data from return value profiling at an invoke with 2276 * the type system so that it can propagate it (speculation) 2277 */ 2278 void GraphKit::record_profiled_return_for_speculation() { 2279 if (!UseTypeSpeculation) { 2280 return; 2281 } 2282 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2283 ciKlass* better_type = NULL; 2284 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2285 // If profiling reports a single type for the return value, 2286 // feed it to the type system so it can propagate it as a 2287 // speculative type 2288 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2289 } 2290 } 2291 2292 void GraphKit::round_double_result(ciMethod* dest_method) { 2293 // A non-strict method may return a double value which has an extended 2294 // exponent, but this must not be visible in a caller which is 'strict' 2295 // If a strict caller invokes a non-strict callee, round a double result 2296 2297 BasicType result_type = dest_method->return_type()->basic_type(); 2298 assert( method() != NULL, "must have caller context"); 2299 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2300 // Destination method's return value is on top of stack 2301 // dstore_rounding() does gvn.transform 2302 Node *result = pop_pair(); 2303 result = dstore_rounding(result); 2304 push_pair(result); 2305 } 2306 } 2307 2308 // rounding for strict float precision conformance 2309 Node* GraphKit::precision_rounding(Node* n) { 2310 return UseStrictFP && _method->flags().is_strict() 2311 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2312 ? _gvn.transform( new RoundFloatNode(0, n) ) 2313 : n; 2314 } 2315 2316 // rounding for strict double precision conformance 2317 Node* GraphKit::dprecision_rounding(Node *n) { 2318 return UseStrictFP && _method->flags().is_strict() 2319 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2320 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2321 : n; 2322 } 2323 2324 // rounding for non-strict double stores 2325 Node* GraphKit::dstore_rounding(Node* n) { 2326 return Matcher::strict_fp_requires_explicit_rounding 2327 && UseSSE <= 1 2328 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2329 : n; 2330 } 2331 2332 //============================================================================= 2333 // Generate a fast path/slow path idiom. Graph looks like: 2334 // [foo] indicates that 'foo' is a parameter 2335 // 2336 // [in] NULL 2337 // \ / 2338 // CmpP 2339 // Bool ne 2340 // If 2341 // / \ 2342 // True False-<2> 2343 // / | 2344 // / cast_not_null 2345 // Load | | ^ 2346 // [fast_test] | | 2347 // gvn to opt_test | | 2348 // / \ | <1> 2349 // True False | 2350 // | \\ | 2351 // [slow_call] \[fast_result] 2352 // Ctl Val \ \ 2353 // | \ \ 2354 // Catch <1> \ \ 2355 // / \ ^ \ \ 2356 // Ex No_Ex | \ \ 2357 // | \ \ | \ <2> \ 2358 // ... \ [slow_res] | | \ [null_result] 2359 // \ \--+--+--- | | 2360 // \ | / \ | / 2361 // --------Region Phi 2362 // 2363 //============================================================================= 2364 // Code is structured as a series of driver functions all called 'do_XXX' that 2365 // call a set of helper functions. Helper functions first, then drivers. 2366 2367 //------------------------------null_check_oop--------------------------------- 2368 // Null check oop. Set null-path control into Region in slot 3. 2369 // Make a cast-not-nullness use the other not-null control. Return cast. 2370 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2371 bool never_see_null, 2372 bool safe_for_replace, 2373 bool speculative) { 2374 // Initial NULL check taken path 2375 (*null_control) = top(); 2376 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2377 2378 // Generate uncommon_trap: 2379 if (never_see_null && (*null_control) != top()) { 2380 // If we see an unexpected null at a check-cast we record it and force a 2381 // recompile; the offending check-cast will be compiled to handle NULLs. 2382 // If we see more than one offending BCI, then all checkcasts in the 2383 // method will be compiled to handle NULLs. 2384 PreserveJVMState pjvms(this); 2385 set_control(*null_control); 2386 replace_in_map(value, null()); 2387 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2388 uncommon_trap(reason, 2389 Deoptimization::Action_make_not_entrant); 2390 (*null_control) = top(); // NULL path is dead 2391 } 2392 if ((*null_control) == top() && safe_for_replace) { 2393 replace_in_map(value, cast); 2394 } 2395 2396 // Cast away null-ness on the result 2397 return cast; 2398 } 2399 2400 //------------------------------opt_iff---------------------------------------- 2401 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2402 // Return slow-path control. 2403 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2404 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2405 2406 // Fast path taken; set region slot 2 2407 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2408 region->init_req(2,fast_taken); // Capture fast-control 2409 2410 // Fast path not-taken, i.e. slow path 2411 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2412 return slow_taken; 2413 } 2414 2415 //-----------------------------make_runtime_call------------------------------- 2416 Node* GraphKit::make_runtime_call(int flags, 2417 const TypeFunc* call_type, address call_addr, 2418 const char* call_name, 2419 const TypePtr* adr_type, 2420 // The following parms are all optional. 2421 // The first NULL ends the list. 2422 Node* parm0, Node* parm1, 2423 Node* parm2, Node* parm3, 2424 Node* parm4, Node* parm5, 2425 Node* parm6, Node* parm7) { 2426 // Slow-path call 2427 bool is_leaf = !(flags & RC_NO_LEAF); 2428 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2429 if (call_name == NULL) { 2430 assert(!is_leaf, "must supply name for leaf"); 2431 call_name = OptoRuntime::stub_name(call_addr); 2432 } 2433 CallNode* call; 2434 if (!is_leaf) { 2435 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2436 bci(), adr_type); 2437 } else if (flags & RC_NO_FP) { 2438 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2439 } else { 2440 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2441 } 2442 2443 // The following is similar to set_edges_for_java_call, 2444 // except that the memory effects of the call are restricted to AliasIdxRaw. 2445 2446 // Slow path call has no side-effects, uses few values 2447 bool wide_in = !(flags & RC_NARROW_MEM); 2448 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2449 2450 Node* prev_mem = NULL; 2451 if (wide_in) { 2452 prev_mem = set_predefined_input_for_runtime_call(call); 2453 } else { 2454 assert(!wide_out, "narrow in => narrow out"); 2455 Node* narrow_mem = memory(adr_type); 2456 prev_mem = reset_memory(); 2457 map()->set_memory(narrow_mem); 2458 set_predefined_input_for_runtime_call(call); 2459 } 2460 2461 // Hook each parm in order. Stop looking at the first NULL. 2462 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2463 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2464 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2465 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2466 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2467 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2468 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2469 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2470 /* close each nested if ===> */ } } } } } } } } 2471 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2472 2473 if (!is_leaf) { 2474 // Non-leaves can block and take safepoints: 2475 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2476 } 2477 // Non-leaves can throw exceptions: 2478 if (has_io) { 2479 call->set_req(TypeFunc::I_O, i_o()); 2480 } 2481 2482 if (flags & RC_UNCOMMON) { 2483 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2484 // (An "if" probability corresponds roughly to an unconditional count. 2485 // Sort of.) 2486 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2487 } 2488 2489 Node* c = _gvn.transform(call); 2490 assert(c == call, "cannot disappear"); 2491 2492 if (wide_out) { 2493 // Slow path call has full side-effects. 2494 set_predefined_output_for_runtime_call(call); 2495 } else { 2496 // Slow path call has few side-effects, and/or sets few values. 2497 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2498 } 2499 2500 if (has_io) { 2501 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2502 } 2503 return call; 2504 2505 } 2506 2507 //------------------------------merge_memory----------------------------------- 2508 // Merge memory from one path into the current memory state. 2509 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2510 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2511 Node* old_slice = mms.force_memory(); 2512 Node* new_slice = mms.memory2(); 2513 if (old_slice != new_slice) { 2514 PhiNode* phi; 2515 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2516 if (mms.is_empty()) { 2517 // clone base memory Phi's inputs for this memory slice 2518 assert(old_slice == mms.base_memory(), "sanity"); 2519 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2520 _gvn.set_type(phi, Type::MEMORY); 2521 for (uint i = 1; i < phi->req(); i++) { 2522 phi->init_req(i, old_slice->in(i)); 2523 } 2524 } else { 2525 phi = old_slice->as_Phi(); // Phi was generated already 2526 } 2527 } else { 2528 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2529 _gvn.set_type(phi, Type::MEMORY); 2530 } 2531 phi->set_req(new_path, new_slice); 2532 mms.set_memory(phi); 2533 } 2534 } 2535 } 2536 2537 //------------------------------make_slow_call_ex------------------------------ 2538 // Make the exception handler hookups for the slow call 2539 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2540 if (stopped()) return; 2541 2542 // Make a catch node with just two handlers: fall-through and catch-all 2543 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2544 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2545 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2546 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2547 2548 { PreserveJVMState pjvms(this); 2549 set_control(excp); 2550 set_i_o(i_o); 2551 2552 if (excp != top()) { 2553 if (deoptimize) { 2554 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2555 uncommon_trap(Deoptimization::Reason_unhandled, 2556 Deoptimization::Action_none); 2557 } else { 2558 // Create an exception state also. 2559 // Use an exact type if the caller has specified a specific exception. 2560 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2561 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2562 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2563 } 2564 } 2565 } 2566 2567 // Get the no-exception control from the CatchNode. 2568 set_control(norm); 2569 } 2570 2571 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2572 Node* cmp = NULL; 2573 switch(bt) { 2574 case T_INT: cmp = new CmpINode(in1, in2); break; 2575 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2576 default: fatal("unexpected comparison type %s", type2name(bt)); 2577 } 2578 gvn->transform(cmp); 2579 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2580 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2581 gvn->transform(iff); 2582 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2583 return iff; 2584 } 2585 2586 2587 //-------------------------------gen_subtype_check----------------------------- 2588 // Generate a subtyping check. Takes as input the subtype and supertype. 2589 // Returns 2 values: sets the default control() to the true path and returns 2590 // the false path. Only reads invariant memory; sets no (visible) memory. 2591 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2592 // but that's not exposed to the optimizer. This call also doesn't take in an 2593 // Object; if you wish to check an Object you need to load the Object's class 2594 // prior to coming here. 2595 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2596 Compile* C = gvn->C; 2597 2598 if ((*ctrl)->is_top()) { 2599 return C->top(); 2600 } 2601 2602 // Fast check for identical types, perhaps identical constants. 2603 // The types can even be identical non-constants, in cases 2604 // involving Array.newInstance, Object.clone, etc. 2605 if (subklass == superklass) 2606 return C->top(); // false path is dead; no test needed. 2607 2608 if (gvn->type(superklass)->singleton()) { 2609 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2610 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2611 2612 // In the common case of an exact superklass, try to fold up the 2613 // test before generating code. You may ask, why not just generate 2614 // the code and then let it fold up? The answer is that the generated 2615 // code will necessarily include null checks, which do not always 2616 // completely fold away. If they are also needless, then they turn 2617 // into a performance loss. Example: 2618 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2619 // Here, the type of 'fa' is often exact, so the store check 2620 // of fa[1]=x will fold up, without testing the nullness of x. 2621 switch (C->static_subtype_check(superk, subk)) { 2622 case Compile::SSC_always_false: 2623 { 2624 Node* always_fail = *ctrl; 2625 *ctrl = gvn->C->top(); 2626 return always_fail; 2627 } 2628 case Compile::SSC_always_true: 2629 return C->top(); 2630 case Compile::SSC_easy_test: 2631 { 2632 // Just do a direct pointer compare and be done. 2633 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2634 *ctrl = gvn->transform(new IfTrueNode(iff)); 2635 return gvn->transform(new IfFalseNode(iff)); 2636 } 2637 case Compile::SSC_full_test: 2638 break; 2639 default: 2640 ShouldNotReachHere(); 2641 } 2642 } 2643 2644 // %%% Possible further optimization: Even if the superklass is not exact, 2645 // if the subklass is the unique subtype of the superklass, the check 2646 // will always succeed. We could leave a dependency behind to ensure this. 2647 2648 // First load the super-klass's check-offset 2649 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2650 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2651 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2652 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2653 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2654 2655 // Load from the sub-klass's super-class display list, or a 1-word cache of 2656 // the secondary superclass list, or a failing value with a sentinel offset 2657 // if the super-klass is an interface or exceptionally deep in the Java 2658 // hierarchy and we have to scan the secondary superclass list the hard way. 2659 // Worst-case type is a little odd: NULL is allowed as a result (usually 2660 // klass loads can never produce a NULL). 2661 Node *chk_off_X = chk_off; 2662 #ifdef _LP64 2663 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2664 #endif 2665 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2666 // For some types like interfaces the following loadKlass is from a 1-word 2667 // cache which is mutable so can't use immutable memory. Other 2668 // types load from the super-class display table which is immutable. 2669 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2670 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2671 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2672 2673 // Compile speed common case: ARE a subtype and we canNOT fail 2674 if( superklass == nkls ) 2675 return C->top(); // false path is dead; no test needed. 2676 2677 // See if we get an immediate positive hit. Happens roughly 83% of the 2678 // time. Test to see if the value loaded just previously from the subklass 2679 // is exactly the superklass. 2680 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2681 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2682 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2683 2684 // Compile speed common case: Check for being deterministic right now. If 2685 // chk_off is a constant and not equal to cacheoff then we are NOT a 2686 // subklass. In this case we need exactly the 1 test above and we can 2687 // return those results immediately. 2688 if (!might_be_cache) { 2689 Node* not_subtype_ctrl = *ctrl; 2690 *ctrl = iftrue1; // We need exactly the 1 test above 2691 return not_subtype_ctrl; 2692 } 2693 2694 // Gather the various success & failures here 2695 RegionNode *r_ok_subtype = new RegionNode(4); 2696 gvn->record_for_igvn(r_ok_subtype); 2697 RegionNode *r_not_subtype = new RegionNode(3); 2698 gvn->record_for_igvn(r_not_subtype); 2699 2700 r_ok_subtype->init_req(1, iftrue1); 2701 2702 // Check for immediate negative hit. Happens roughly 11% of the time (which 2703 // is roughly 63% of the remaining cases). Test to see if the loaded 2704 // check-offset points into the subklass display list or the 1-element 2705 // cache. If it points to the display (and NOT the cache) and the display 2706 // missed then it's not a subtype. 2707 Node *cacheoff = gvn->intcon(cacheoff_con); 2708 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2709 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2710 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2711 2712 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2713 // No performance impact (too rare) but allows sharing of secondary arrays 2714 // which has some footprint reduction. 2715 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2716 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2717 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2718 2719 // -- Roads not taken here: -- 2720 // We could also have chosen to perform the self-check at the beginning 2721 // of this code sequence, as the assembler does. This would not pay off 2722 // the same way, since the optimizer, unlike the assembler, can perform 2723 // static type analysis to fold away many successful self-checks. 2724 // Non-foldable self checks work better here in second position, because 2725 // the initial primary superclass check subsumes a self-check for most 2726 // types. An exception would be a secondary type like array-of-interface, 2727 // which does not appear in its own primary supertype display. 2728 // Finally, we could have chosen to move the self-check into the 2729 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2730 // dependent manner. But it is worthwhile to have the check here, 2731 // where it can be perhaps be optimized. The cost in code space is 2732 // small (register compare, branch). 2733 2734 // Now do a linear scan of the secondary super-klass array. Again, no real 2735 // performance impact (too rare) but it's gotta be done. 2736 // Since the code is rarely used, there is no penalty for moving it 2737 // out of line, and it can only improve I-cache density. 2738 // The decision to inline or out-of-line this final check is platform 2739 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2740 Node* psc = gvn->transform( 2741 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2742 2743 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2744 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2745 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2746 2747 // Return false path; set default control to true path. 2748 *ctrl = gvn->transform(r_ok_subtype); 2749 return gvn->transform(r_not_subtype); 2750 } 2751 2752 // Profile-driven exact type check: 2753 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2754 float prob, 2755 Node* *casted_receiver) { 2756 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2757 Node* recv_klass = load_object_klass(receiver); 2758 Node* want_klass = makecon(tklass); 2759 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2760 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2761 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2762 set_control( _gvn.transform( new IfTrueNode (iff) )); 2763 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2764 2765 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2766 assert(recv_xtype->klass_is_exact(), ""); 2767 2768 // Subsume downstream occurrences of receiver with a cast to 2769 // recv_xtype, since now we know what the type will be. 2770 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2771 (*casted_receiver) = _gvn.transform(cast); 2772 // (User must make the replace_in_map call.) 2773 2774 return fail; 2775 } 2776 2777 2778 //------------------------------seems_never_null------------------------------- 2779 // Use null_seen information if it is available from the profile. 2780 // If we see an unexpected null at a type check we record it and force a 2781 // recompile; the offending check will be recompiled to handle NULLs. 2782 // If we see several offending BCIs, then all checks in the 2783 // method will be recompiled. 2784 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2785 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2786 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2787 if (UncommonNullCast // Cutout for this technique 2788 && obj != null() // And not the -Xcomp stupid case? 2789 && !too_many_traps(reason) 2790 ) { 2791 if (speculating) { 2792 return true; 2793 } 2794 if (data == NULL) 2795 // Edge case: no mature data. Be optimistic here. 2796 return true; 2797 // If the profile has not seen a null, assume it won't happen. 2798 assert(java_bc() == Bytecodes::_checkcast || 2799 java_bc() == Bytecodes::_instanceof || 2800 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2801 return !data->as_BitData()->null_seen(); 2802 } 2803 speculating = false; 2804 return false; 2805 } 2806 2807 //------------------------maybe_cast_profiled_receiver------------------------- 2808 // If the profile has seen exactly one type, narrow to exactly that type. 2809 // Subsequent type checks will always fold up. 2810 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2811 ciKlass* require_klass, 2812 ciKlass* spec_klass, 2813 bool safe_for_replace) { 2814 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2815 2816 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2817 2818 // Make sure we haven't already deoptimized from this tactic. 2819 if (too_many_traps(reason) || too_many_recompiles(reason)) 2820 return NULL; 2821 2822 // (No, this isn't a call, but it's enough like a virtual call 2823 // to use the same ciMethod accessor to get the profile info...) 2824 // If we have a speculative type use it instead of profiling (which 2825 // may not help us) 2826 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2827 if (exact_kls != NULL) {// no cast failures here 2828 if (require_klass == NULL || 2829 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2830 // If we narrow the type to match what the type profile sees or 2831 // the speculative type, we can then remove the rest of the 2832 // cast. 2833 // This is a win, even if the exact_kls is very specific, 2834 // because downstream operations, such as method calls, 2835 // will often benefit from the sharper type. 2836 Node* exact_obj = not_null_obj; // will get updated in place... 2837 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2838 &exact_obj); 2839 { PreserveJVMState pjvms(this); 2840 set_control(slow_ctl); 2841 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2842 } 2843 if (safe_for_replace) { 2844 replace_in_map(not_null_obj, exact_obj); 2845 } 2846 return exact_obj; 2847 } 2848 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2849 } 2850 2851 return NULL; 2852 } 2853 2854 /** 2855 * Cast obj to type and emit guard unless we had too many traps here 2856 * already 2857 * 2858 * @param obj node being casted 2859 * @param type type to cast the node to 2860 * @param not_null true if we know node cannot be null 2861 */ 2862 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2863 ciKlass* type, 2864 bool not_null) { 2865 if (stopped()) { 2866 return obj; 2867 } 2868 2869 // type == NULL if profiling tells us this object is always null 2870 if (type != NULL) { 2871 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2872 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2873 2874 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2875 !too_many_traps(class_reason) && 2876 !too_many_recompiles(class_reason)) { 2877 Node* not_null_obj = NULL; 2878 // not_null is true if we know the object is not null and 2879 // there's no need for a null check 2880 if (!not_null) { 2881 Node* null_ctl = top(); 2882 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2883 assert(null_ctl->is_top(), "no null control here"); 2884 } else { 2885 not_null_obj = obj; 2886 } 2887 2888 Node* exact_obj = not_null_obj; 2889 ciKlass* exact_kls = type; 2890 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2891 &exact_obj); 2892 { 2893 PreserveJVMState pjvms(this); 2894 set_control(slow_ctl); 2895 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2896 } 2897 replace_in_map(not_null_obj, exact_obj); 2898 obj = exact_obj; 2899 } 2900 } else { 2901 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2902 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2903 Node* exact_obj = null_assert(obj); 2904 replace_in_map(obj, exact_obj); 2905 obj = exact_obj; 2906 } 2907 } 2908 return obj; 2909 } 2910 2911 //-------------------------------gen_instanceof-------------------------------- 2912 // Generate an instance-of idiom. Used by both the instance-of bytecode 2913 // and the reflective instance-of call. 2914 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2915 kill_dead_locals(); // Benefit all the uncommon traps 2916 assert( !stopped(), "dead parse path should be checked in callers" ); 2917 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2918 "must check for not-null not-dead klass in callers"); 2919 2920 // Make the merge point 2921 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2922 RegionNode* region = new RegionNode(PATH_LIMIT); 2923 Node* phi = new PhiNode(region, TypeInt::BOOL); 2924 C->set_has_split_ifs(true); // Has chance for split-if optimization 2925 2926 ciProfileData* data = NULL; 2927 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2928 data = method()->method_data()->bci_to_data(bci()); 2929 } 2930 bool speculative_not_null = false; 2931 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2932 && seems_never_null(obj, data, speculative_not_null)); 2933 2934 // Null check; get casted pointer; set region slot 3 2935 Node* null_ctl = top(); 2936 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2937 2938 // If not_null_obj is dead, only null-path is taken 2939 if (stopped()) { // Doing instance-of on a NULL? 2940 set_control(null_ctl); 2941 return intcon(0); 2942 } 2943 region->init_req(_null_path, null_ctl); 2944 phi ->init_req(_null_path, intcon(0)); // Set null path value 2945 if (null_ctl == top()) { 2946 // Do this eagerly, so that pattern matches like is_diamond_phi 2947 // will work even during parsing. 2948 assert(_null_path == PATH_LIMIT-1, "delete last"); 2949 region->del_req(_null_path); 2950 phi ->del_req(_null_path); 2951 } 2952 2953 // Do we know the type check always succeed? 2954 bool known_statically = false; 2955 if (_gvn.type(superklass)->singleton()) { 2956 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2957 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2958 if (subk != NULL && subk->is_loaded()) { 2959 int static_res = C->static_subtype_check(superk, subk); 2960 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2961 } 2962 } 2963 2964 if (!known_statically) { 2965 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2966 // We may not have profiling here or it may not help us. If we 2967 // have a speculative type use it to perform an exact cast. 2968 ciKlass* spec_obj_type = obj_type->speculative_type(); 2969 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2970 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2971 if (stopped()) { // Profile disagrees with this path. 2972 set_control(null_ctl); // Null is the only remaining possibility. 2973 return intcon(0); 2974 } 2975 if (cast_obj != NULL) { 2976 not_null_obj = cast_obj; 2977 } 2978 } 2979 } 2980 2981 // Load the object's klass 2982 Node* obj_klass = load_object_klass(not_null_obj); 2983 2984 // Generate the subtype check 2985 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2986 2987 // Plug in the success path to the general merge in slot 1. 2988 region->init_req(_obj_path, control()); 2989 phi ->init_req(_obj_path, intcon(1)); 2990 2991 // Plug in the failing path to the general merge in slot 2. 2992 region->init_req(_fail_path, not_subtype_ctrl); 2993 phi ->init_req(_fail_path, intcon(0)); 2994 2995 // Return final merged results 2996 set_control( _gvn.transform(region) ); 2997 record_for_igvn(region); 2998 2999 // If we know the type check always succeeds then we don't use the 3000 // profiling data at this bytecode. Don't lose it, feed it to the 3001 // type system as a speculative type. 3002 if (safe_for_replace) { 3003 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3004 replace_in_map(obj, casted_obj); 3005 } 3006 3007 return _gvn.transform(phi); 3008 } 3009 3010 //-------------------------------gen_checkcast--------------------------------- 3011 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3012 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3013 // uncommon-trap paths work. Adjust stack after this call. 3014 // If failure_control is supplied and not null, it is filled in with 3015 // the control edge for the cast failure. Otherwise, an appropriate 3016 // uncommon trap or exception is thrown. 3017 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3018 Node* *failure_control) { 3019 kill_dead_locals(); // Benefit all the uncommon traps 3020 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3021 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3022 3023 // Fast cutout: Check the case that the cast is vacuously true. 3024 // This detects the common cases where the test will short-circuit 3025 // away completely. We do this before we perform the null check, 3026 // because if the test is going to turn into zero code, we don't 3027 // want a residual null check left around. (Causes a slowdown, 3028 // for example, in some objArray manipulations, such as a[i]=a[j].) 3029 if (tk->singleton()) { 3030 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3031 if (objtp != NULL && objtp->klass() != NULL) { 3032 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3033 case Compile::SSC_always_true: 3034 // If we know the type check always succeed then we don't use 3035 // the profiling data at this bytecode. Don't lose it, feed it 3036 // to the type system as a speculative type. 3037 return record_profiled_receiver_for_speculation(obj); 3038 case Compile::SSC_always_false: 3039 // It needs a null check because a null will *pass* the cast check. 3040 // A non-null value will always produce an exception. 3041 return null_assert(obj); 3042 } 3043 } 3044 } 3045 3046 ciProfileData* data = NULL; 3047 bool safe_for_replace = false; 3048 if (failure_control == NULL) { // use MDO in regular case only 3049 assert(java_bc() == Bytecodes::_aastore || 3050 java_bc() == Bytecodes::_checkcast, 3051 "interpreter profiles type checks only for these BCs"); 3052 data = method()->method_data()->bci_to_data(bci()); 3053 safe_for_replace = true; 3054 } 3055 3056 // Make the merge point 3057 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3058 RegionNode* region = new RegionNode(PATH_LIMIT); 3059 Node* phi = new PhiNode(region, toop); 3060 C->set_has_split_ifs(true); // Has chance for split-if optimization 3061 3062 // Use null-cast information if it is available 3063 bool speculative_not_null = false; 3064 bool never_see_null = ((failure_control == NULL) // regular case only 3065 && seems_never_null(obj, data, speculative_not_null)); 3066 3067 // Null check; get casted pointer; set region slot 3 3068 Node* null_ctl = top(); 3069 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3070 3071 // If not_null_obj is dead, only null-path is taken 3072 if (stopped()) { // Doing instance-of on a NULL? 3073 set_control(null_ctl); 3074 return null(); 3075 } 3076 region->init_req(_null_path, null_ctl); 3077 phi ->init_req(_null_path, null()); // Set null path value 3078 if (null_ctl == top()) { 3079 // Do this eagerly, so that pattern matches like is_diamond_phi 3080 // will work even during parsing. 3081 assert(_null_path == PATH_LIMIT-1, "delete last"); 3082 region->del_req(_null_path); 3083 phi ->del_req(_null_path); 3084 } 3085 3086 Node* cast_obj = NULL; 3087 if (tk->klass_is_exact()) { 3088 // The following optimization tries to statically cast the speculative type of the object 3089 // (for example obtained during profiling) to the type of the superklass and then do a 3090 // dynamic check that the type of the object is what we expect. To work correctly 3091 // for checkcast and aastore the type of superklass should be exact. 3092 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3093 // We may not have profiling here or it may not help us. If we have 3094 // a speculative type use it to perform an exact cast. 3095 ciKlass* spec_obj_type = obj_type->speculative_type(); 3096 if (spec_obj_type != NULL || data != NULL) { 3097 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3098 if (cast_obj != NULL) { 3099 if (failure_control != NULL) // failure is now impossible 3100 (*failure_control) = top(); 3101 // adjust the type of the phi to the exact klass: 3102 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3103 } 3104 } 3105 } 3106 3107 if (cast_obj == NULL) { 3108 // Load the object's klass 3109 Node* obj_klass = load_object_klass(not_null_obj); 3110 3111 // Generate the subtype check 3112 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3113 3114 // Plug in success path into the merge 3115 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3116 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3117 if (failure_control == NULL) { 3118 if (not_subtype_ctrl != top()) { // If failure is possible 3119 PreserveJVMState pjvms(this); 3120 set_control(not_subtype_ctrl); 3121 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3122 } 3123 } else { 3124 (*failure_control) = not_subtype_ctrl; 3125 } 3126 } 3127 3128 region->init_req(_obj_path, control()); 3129 phi ->init_req(_obj_path, cast_obj); 3130 3131 // A merge of NULL or Casted-NotNull obj 3132 Node* res = _gvn.transform(phi); 3133 3134 // Note I do NOT always 'replace_in_map(obj,result)' here. 3135 // if( tk->klass()->can_be_primary_super() ) 3136 // This means that if I successfully store an Object into an array-of-String 3137 // I 'forget' that the Object is really now known to be a String. I have to 3138 // do this because we don't have true union types for interfaces - if I store 3139 // a Baz into an array-of-Interface and then tell the optimizer it's an 3140 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3141 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3142 // replace_in_map( obj, res ); 3143 3144 // Return final merged results 3145 set_control( _gvn.transform(region) ); 3146 record_for_igvn(region); 3147 3148 return record_profiled_receiver_for_speculation(res); 3149 } 3150 3151 //------------------------------next_monitor----------------------------------- 3152 // What number should be given to the next monitor? 3153 int GraphKit::next_monitor() { 3154 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3155 int next = current + C->sync_stack_slots(); 3156 // Keep the toplevel high water mark current: 3157 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3158 return current; 3159 } 3160 3161 //------------------------------insert_mem_bar--------------------------------- 3162 // Memory barrier to avoid floating things around 3163 // The membar serves as a pinch point between both control and all memory slices. 3164 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3165 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3166 mb->init_req(TypeFunc::Control, control()); 3167 mb->init_req(TypeFunc::Memory, reset_memory()); 3168 Node* membar = _gvn.transform(mb); 3169 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3170 set_all_memory_call(membar); 3171 return membar; 3172 } 3173 3174 //-------------------------insert_mem_bar_volatile---------------------------- 3175 // Memory barrier to avoid floating things around 3176 // The membar serves as a pinch point between both control and memory(alias_idx). 3177 // If you want to make a pinch point on all memory slices, do not use this 3178 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3179 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3180 // When Parse::do_put_xxx updates a volatile field, it appends a series 3181 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3182 // The first membar is on the same memory slice as the field store opcode. 3183 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3184 // All the other membars (for other volatile slices, including AliasIdxBot, 3185 // which stands for all unknown volatile slices) are control-dependent 3186 // on the first membar. This prevents later volatile loads or stores 3187 // from sliding up past the just-emitted store. 3188 3189 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3190 mb->set_req(TypeFunc::Control,control()); 3191 if (alias_idx == Compile::AliasIdxBot) { 3192 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3193 } else { 3194 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3195 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3196 } 3197 Node* membar = _gvn.transform(mb); 3198 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3199 if (alias_idx == Compile::AliasIdxBot) { 3200 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3201 } else { 3202 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3203 } 3204 return membar; 3205 } 3206 3207 //------------------------------shared_lock------------------------------------ 3208 // Emit locking code. 3209 FastLockNode* GraphKit::shared_lock(Node* obj) { 3210 // bci is either a monitorenter bc or InvocationEntryBci 3211 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3212 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3213 3214 if( !GenerateSynchronizationCode ) 3215 return NULL; // Not locking things? 3216 if (stopped()) // Dead monitor? 3217 return NULL; 3218 3219 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3220 3221 // Box the stack location 3222 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3223 Node* mem = reset_memory(); 3224 3225 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3226 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3227 // Create the counters for this fast lock. 3228 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3229 } 3230 3231 // Create the rtm counters for this fast lock if needed. 3232 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3233 3234 // Add monitor to debug info for the slow path. If we block inside the 3235 // slow path and de-opt, we need the monitor hanging around 3236 map()->push_monitor( flock ); 3237 3238 const TypeFunc *tf = LockNode::lock_type(); 3239 LockNode *lock = new LockNode(C, tf); 3240 3241 lock->init_req( TypeFunc::Control, control() ); 3242 lock->init_req( TypeFunc::Memory , mem ); 3243 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3244 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3245 lock->init_req( TypeFunc::ReturnAdr, top() ); 3246 3247 lock->init_req(TypeFunc::Parms + 0, obj); 3248 lock->init_req(TypeFunc::Parms + 1, box); 3249 lock->init_req(TypeFunc::Parms + 2, flock); 3250 add_safepoint_edges(lock); 3251 3252 lock = _gvn.transform( lock )->as_Lock(); 3253 3254 // lock has no side-effects, sets few values 3255 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3256 3257 insert_mem_bar(Op_MemBarAcquireLock); 3258 3259 // Add this to the worklist so that the lock can be eliminated 3260 record_for_igvn(lock); 3261 3262 #ifndef PRODUCT 3263 if (PrintLockStatistics) { 3264 // Update the counter for this lock. Don't bother using an atomic 3265 // operation since we don't require absolute accuracy. 3266 lock->create_lock_counter(map()->jvms()); 3267 increment_counter(lock->counter()->addr()); 3268 } 3269 #endif 3270 3271 return flock; 3272 } 3273 3274 3275 //------------------------------shared_unlock---------------------------------- 3276 // Emit unlocking code. 3277 void GraphKit::shared_unlock(Node* box, Node* obj) { 3278 // bci is either a monitorenter bc or InvocationEntryBci 3279 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3280 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3281 3282 if( !GenerateSynchronizationCode ) 3283 return; 3284 if (stopped()) { // Dead monitor? 3285 map()->pop_monitor(); // Kill monitor from debug info 3286 return; 3287 } 3288 3289 // Memory barrier to avoid floating things down past the locked region 3290 insert_mem_bar(Op_MemBarReleaseLock); 3291 3292 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3293 UnlockNode *unlock = new UnlockNode(C, tf); 3294 #ifdef ASSERT 3295 unlock->set_dbg_jvms(sync_jvms()); 3296 #endif 3297 uint raw_idx = Compile::AliasIdxRaw; 3298 unlock->init_req( TypeFunc::Control, control() ); 3299 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3300 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3301 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3302 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3303 3304 unlock->init_req(TypeFunc::Parms + 0, obj); 3305 unlock->init_req(TypeFunc::Parms + 1, box); 3306 unlock = _gvn.transform(unlock)->as_Unlock(); 3307 3308 Node* mem = reset_memory(); 3309 3310 // unlock has no side-effects, sets few values 3311 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3312 3313 // Kill monitor from debug info 3314 map()->pop_monitor( ); 3315 } 3316 3317 //-------------------------------get_layout_helper----------------------------- 3318 // If the given klass is a constant or known to be an array, 3319 // fetch the constant layout helper value into constant_value 3320 // and return (Node*)NULL. Otherwise, load the non-constant 3321 // layout helper value, and return the node which represents it. 3322 // This two-faced routine is useful because allocation sites 3323 // almost always feature constant types. 3324 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3325 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3326 if (!StressReflectiveCode && inst_klass != NULL) { 3327 ciKlass* klass = inst_klass->klass(); 3328 bool xklass = inst_klass->klass_is_exact(); 3329 if (xklass || klass->is_array_klass()) { 3330 jint lhelper = klass->layout_helper(); 3331 if (lhelper != Klass::_lh_neutral_value) { 3332 constant_value = lhelper; 3333 return (Node*) NULL; 3334 } 3335 } 3336 } 3337 constant_value = Klass::_lh_neutral_value; // put in a known value 3338 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3339 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3340 } 3341 3342 // We just put in an allocate/initialize with a big raw-memory effect. 3343 // Hook selected additional alias categories on the initialization. 3344 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3345 MergeMemNode* init_in_merge, 3346 Node* init_out_raw) { 3347 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3348 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3349 3350 Node* prevmem = kit.memory(alias_idx); 3351 init_in_merge->set_memory_at(alias_idx, prevmem); 3352 kit.set_memory(init_out_raw, alias_idx); 3353 } 3354 3355 //---------------------------set_output_for_allocation------------------------- 3356 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3357 const TypeOopPtr* oop_type, 3358 bool deoptimize_on_exception) { 3359 int rawidx = Compile::AliasIdxRaw; 3360 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3361 add_safepoint_edges(alloc); 3362 Node* allocx = _gvn.transform(alloc); 3363 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3364 // create memory projection for i_o 3365 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3366 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3367 3368 // create a memory projection as for the normal control path 3369 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3370 set_memory(malloc, rawidx); 3371 3372 // a normal slow-call doesn't change i_o, but an allocation does 3373 // we create a separate i_o projection for the normal control path 3374 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3375 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3376 3377 // put in an initialization barrier 3378 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3379 rawoop)->as_Initialize(); 3380 assert(alloc->initialization() == init, "2-way macro link must work"); 3381 assert(init ->allocation() == alloc, "2-way macro link must work"); 3382 { 3383 // Extract memory strands which may participate in the new object's 3384 // initialization, and source them from the new InitializeNode. 3385 // This will allow us to observe initializations when they occur, 3386 // and link them properly (as a group) to the InitializeNode. 3387 assert(init->in(InitializeNode::Memory) == malloc, ""); 3388 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3389 init->set_req(InitializeNode::Memory, minit_in); 3390 record_for_igvn(minit_in); // fold it up later, if possible 3391 Node* minit_out = memory(rawidx); 3392 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3393 if (oop_type->isa_aryptr()) { 3394 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3395 int elemidx = C->get_alias_index(telemref); 3396 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3397 } else if (oop_type->isa_instptr()) { 3398 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3399 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3400 ciField* field = ik->nonstatic_field_at(i); 3401 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3402 continue; // do not bother to track really large numbers of fields 3403 // Find (or create) the alias category for this field: 3404 int fieldidx = C->alias_type(field)->index(); 3405 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3406 } 3407 } 3408 } 3409 3410 // Cast raw oop to the real thing... 3411 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3412 javaoop = _gvn.transform(javaoop); 3413 C->set_recent_alloc(control(), javaoop); 3414 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3415 3416 #ifdef ASSERT 3417 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3418 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3419 "Ideal_allocation works"); 3420 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3421 "Ideal_allocation works"); 3422 if (alloc->is_AllocateArray()) { 3423 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3424 "Ideal_allocation works"); 3425 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3426 "Ideal_allocation works"); 3427 } else { 3428 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3429 } 3430 } 3431 #endif //ASSERT 3432 3433 return javaoop; 3434 } 3435 3436 //---------------------------new_instance-------------------------------------- 3437 // This routine takes a klass_node which may be constant (for a static type) 3438 // or may be non-constant (for reflective code). It will work equally well 3439 // for either, and the graph will fold nicely if the optimizer later reduces 3440 // the type to a constant. 3441 // The optional arguments are for specialized use by intrinsics: 3442 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3443 // - If 'return_size_val', report the the total object size to the caller. 3444 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3445 Node* GraphKit::new_instance(Node* klass_node, 3446 Node* extra_slow_test, 3447 Node* *return_size_val, 3448 bool deoptimize_on_exception) { 3449 // Compute size in doublewords 3450 // The size is always an integral number of doublewords, represented 3451 // as a positive bytewise size stored in the klass's layout_helper. 3452 // The layout_helper also encodes (in a low bit) the need for a slow path. 3453 jint layout_con = Klass::_lh_neutral_value; 3454 Node* layout_val = get_layout_helper(klass_node, layout_con); 3455 int layout_is_con = (layout_val == NULL); 3456 3457 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3458 // Generate the initial go-slow test. It's either ALWAYS (return a 3459 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3460 // case) a computed value derived from the layout_helper. 3461 Node* initial_slow_test = NULL; 3462 if (layout_is_con) { 3463 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3464 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3465 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3466 } else { // reflective case 3467 // This reflective path is used by Unsafe.allocateInstance. 3468 // (It may be stress-tested by specifying StressReflectiveCode.) 3469 // Basically, we want to get into the VM is there's an illegal argument. 3470 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3471 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3472 if (extra_slow_test != intcon(0)) { 3473 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3474 } 3475 // (Macro-expander will further convert this to a Bool, if necessary.) 3476 } 3477 3478 // Find the size in bytes. This is easy; it's the layout_helper. 3479 // The size value must be valid even if the slow path is taken. 3480 Node* size = NULL; 3481 if (layout_is_con) { 3482 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3483 } else { // reflective case 3484 // This reflective path is used by clone and Unsafe.allocateInstance. 3485 size = ConvI2X(layout_val); 3486 3487 // Clear the low bits to extract layout_helper_size_in_bytes: 3488 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3489 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3490 size = _gvn.transform( new AndXNode(size, mask) ); 3491 } 3492 if (return_size_val != NULL) { 3493 (*return_size_val) = size; 3494 } 3495 3496 // This is a precise notnull oop of the klass. 3497 // (Actually, it need not be precise if this is a reflective allocation.) 3498 // It's what we cast the result to. 3499 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3500 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3501 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3502 3503 // Now generate allocation code 3504 3505 // The entire memory state is needed for slow path of the allocation 3506 // since GC and deoptimization can happened. 3507 Node *mem = reset_memory(); 3508 set_all_memory(mem); // Create new memory state 3509 3510 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3511 control(), mem, i_o(), 3512 size, klass_node, 3513 initial_slow_test); 3514 3515 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3516 } 3517 3518 //-------------------------------new_array------------------------------------- 3519 // helper for both newarray and anewarray 3520 // The 'length' parameter is (obviously) the length of the array. 3521 // See comments on new_instance for the meaning of the other arguments. 3522 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3523 Node* length, // number of array elements 3524 int nargs, // number of arguments to push back for uncommon trap 3525 Node* *return_size_val, 3526 bool deoptimize_on_exception) { 3527 jint layout_con = Klass::_lh_neutral_value; 3528 Node* layout_val = get_layout_helper(klass_node, layout_con); 3529 int layout_is_con = (layout_val == NULL); 3530 3531 if (!layout_is_con && !StressReflectiveCode && 3532 !too_many_traps(Deoptimization::Reason_class_check)) { 3533 // This is a reflective array creation site. 3534 // Optimistically assume that it is a subtype of Object[], 3535 // so that we can fold up all the address arithmetic. 3536 layout_con = Klass::array_layout_helper(T_OBJECT); 3537 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3538 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3539 { BuildCutout unless(this, bol_lh, PROB_MAX); 3540 inc_sp(nargs); 3541 uncommon_trap(Deoptimization::Reason_class_check, 3542 Deoptimization::Action_maybe_recompile); 3543 } 3544 layout_val = NULL; 3545 layout_is_con = true; 3546 } 3547 3548 // Generate the initial go-slow test. Make sure we do not overflow 3549 // if length is huge (near 2Gig) or negative! We do not need 3550 // exact double-words here, just a close approximation of needed 3551 // double-words. We can't add any offset or rounding bits, lest we 3552 // take a size -1 of bytes and make it positive. Use an unsigned 3553 // compare, so negative sizes look hugely positive. 3554 int fast_size_limit = FastAllocateSizeLimit; 3555 if (layout_is_con) { 3556 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3557 // Increase the size limit if we have exact knowledge of array type. 3558 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3559 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3560 } 3561 3562 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3563 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3564 3565 // --- Size Computation --- 3566 // array_size = round_to_heap(array_header + (length << elem_shift)); 3567 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3568 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3569 // The rounding mask is strength-reduced, if possible. 3570 int round_mask = MinObjAlignmentInBytes - 1; 3571 Node* header_size = NULL; 3572 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3573 // (T_BYTE has the weakest alignment and size restrictions...) 3574 if (layout_is_con) { 3575 int hsize = Klass::layout_helper_header_size(layout_con); 3576 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3577 BasicType etype = Klass::layout_helper_element_type(layout_con); 3578 if ((round_mask & ~right_n_bits(eshift)) == 0) 3579 round_mask = 0; // strength-reduce it if it goes away completely 3580 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3581 assert(header_size_min <= hsize, "generic minimum is smallest"); 3582 header_size_min = hsize; 3583 header_size = intcon(hsize + round_mask); 3584 } else { 3585 Node* hss = intcon(Klass::_lh_header_size_shift); 3586 Node* hsm = intcon(Klass::_lh_header_size_mask); 3587 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3588 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3589 Node* mask = intcon(round_mask); 3590 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3591 } 3592 3593 Node* elem_shift = NULL; 3594 if (layout_is_con) { 3595 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3596 if (eshift != 0) 3597 elem_shift = intcon(eshift); 3598 } else { 3599 // There is no need to mask or shift this value. 3600 // The semantics of LShiftINode include an implicit mask to 0x1F. 3601 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3602 elem_shift = layout_val; 3603 } 3604 3605 // Transition to native address size for all offset calculations: 3606 Node* lengthx = ConvI2X(length); 3607 Node* headerx = ConvI2X(header_size); 3608 #ifdef _LP64 3609 { const TypeInt* tilen = _gvn.find_int_type(length); 3610 if (tilen != NULL && tilen->_lo < 0) { 3611 // Add a manual constraint to a positive range. Cf. array_element_address. 3612 jint size_max = fast_size_limit; 3613 if (size_max > tilen->_hi) size_max = tilen->_hi; 3614 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3615 3616 // Only do a narrow I2L conversion if the range check passed. 3617 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3618 _gvn.transform(iff); 3619 RegionNode* region = new RegionNode(3); 3620 _gvn.set_type(region, Type::CONTROL); 3621 lengthx = new PhiNode(region, TypeLong::LONG); 3622 _gvn.set_type(lengthx, TypeLong::LONG); 3623 3624 // Range check passed. Use ConvI2L node with narrow type. 3625 Node* passed = IfFalse(iff); 3626 region->init_req(1, passed); 3627 // Make I2L conversion control dependent to prevent it from 3628 // floating above the range check during loop optimizations. 3629 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3630 3631 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3632 region->init_req(2, IfTrue(iff)); 3633 lengthx->init_req(2, ConvI2X(length)); 3634 3635 set_control(region); 3636 record_for_igvn(region); 3637 record_for_igvn(lengthx); 3638 } 3639 } 3640 #endif 3641 3642 // Combine header size (plus rounding) and body size. Then round down. 3643 // This computation cannot overflow, because it is used only in two 3644 // places, one where the length is sharply limited, and the other 3645 // after a successful allocation. 3646 Node* abody = lengthx; 3647 if (elem_shift != NULL) 3648 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3649 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3650 if (round_mask != 0) { 3651 Node* mask = MakeConX(~round_mask); 3652 size = _gvn.transform( new AndXNode(size, mask) ); 3653 } 3654 // else if round_mask == 0, the size computation is self-rounding 3655 3656 if (return_size_val != NULL) { 3657 // This is the size 3658 (*return_size_val) = size; 3659 } 3660 3661 // Now generate allocation code 3662 3663 // The entire memory state is needed for slow path of the allocation 3664 // since GC and deoptimization can happened. 3665 Node *mem = reset_memory(); 3666 set_all_memory(mem); // Create new memory state 3667 3668 if (initial_slow_test->is_Bool()) { 3669 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3670 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3671 } 3672 3673 // Create the AllocateArrayNode and its result projections 3674 AllocateArrayNode* alloc 3675 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3676 control(), mem, i_o(), 3677 size, klass_node, 3678 initial_slow_test, 3679 length); 3680 3681 // Cast to correct type. Note that the klass_node may be constant or not, 3682 // and in the latter case the actual array type will be inexact also. 3683 // (This happens via a non-constant argument to inline_native_newArray.) 3684 // In any case, the value of klass_node provides the desired array type. 3685 const TypeInt* length_type = _gvn.find_int_type(length); 3686 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3687 if (ary_type->isa_aryptr() && length_type != NULL) { 3688 // Try to get a better type than POS for the size 3689 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3690 } 3691 3692 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3693 3694 // Cast length on remaining path to be as narrow as possible 3695 if (map()->find_edge(length) >= 0) { 3696 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3697 if (ccast != length) { 3698 _gvn.set_type_bottom(ccast); 3699 record_for_igvn(ccast); 3700 replace_in_map(length, ccast); 3701 } 3702 } 3703 3704 return javaoop; 3705 } 3706 3707 // The following "Ideal_foo" functions are placed here because they recognize 3708 // the graph shapes created by the functions immediately above. 3709 3710 //---------------------------Ideal_allocation---------------------------------- 3711 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3712 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3713 if (ptr == NULL) { // reduce dumb test in callers 3714 return NULL; 3715 } 3716 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3717 ptr = ptr->in(1); 3718 if (ptr == NULL) return NULL; 3719 } 3720 // Return NULL for allocations with several casts: 3721 // j.l.reflect.Array.newInstance(jobject, jint) 3722 // Object.clone() 3723 // to keep more precise type from last cast. 3724 if (ptr->is_Proj()) { 3725 Node* allo = ptr->in(0); 3726 if (allo != NULL && allo->is_Allocate()) { 3727 return allo->as_Allocate(); 3728 } 3729 } 3730 // Report failure to match. 3731 return NULL; 3732 } 3733 3734 // Fancy version which also strips off an offset (and reports it to caller). 3735 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3736 intptr_t& offset) { 3737 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3738 if (base == NULL) return NULL; 3739 return Ideal_allocation(base, phase); 3740 } 3741 3742 // Trace Initialize <- Proj[Parm] <- Allocate 3743 AllocateNode* InitializeNode::allocation() { 3744 Node* rawoop = in(InitializeNode::RawAddress); 3745 if (rawoop->is_Proj()) { 3746 Node* alloc = rawoop->in(0); 3747 if (alloc->is_Allocate()) { 3748 return alloc->as_Allocate(); 3749 } 3750 } 3751 return NULL; 3752 } 3753 3754 // Trace Allocate -> Proj[Parm] -> Initialize 3755 InitializeNode* AllocateNode::initialization() { 3756 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3757 if (rawoop == NULL) return NULL; 3758 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3759 Node* init = rawoop->fast_out(i); 3760 if (init->is_Initialize()) { 3761 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3762 return init->as_Initialize(); 3763 } 3764 } 3765 return NULL; 3766 } 3767 3768 //----------------------------- loop predicates --------------------------- 3769 3770 //------------------------------add_predicate_impl---------------------------- 3771 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3772 // Too many traps seen? 3773 if (too_many_traps(reason)) { 3774 #ifdef ASSERT 3775 if (TraceLoopPredicate) { 3776 int tc = C->trap_count(reason); 3777 tty->print("too many traps=%s tcount=%d in ", 3778 Deoptimization::trap_reason_name(reason), tc); 3779 method()->print(); // which method has too many predicate traps 3780 tty->cr(); 3781 } 3782 #endif 3783 // We cannot afford to take more traps here, 3784 // do not generate predicate. 3785 return; 3786 } 3787 3788 Node *cont = _gvn.intcon(1); 3789 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3790 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3791 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3792 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3793 C->add_predicate_opaq(opq); 3794 { 3795 PreserveJVMState pjvms(this); 3796 set_control(iffalse); 3797 inc_sp(nargs); 3798 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3799 } 3800 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3801 set_control(iftrue); 3802 } 3803 3804 //------------------------------add_predicate--------------------------------- 3805 void GraphKit::add_predicate(int nargs) { 3806 if (UseLoopPredicate) { 3807 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3808 } 3809 // loop's limit check predicate should be near the loop. 3810 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3811 } 3812 3813 //----------------------------- store barriers ---------------------------- 3814 #define __ ideal. 3815 3816 bool GraphKit::use_ReduceInitialCardMarks() { 3817 BarrierSet *bs = BarrierSet::barrier_set(); 3818 return bs->is_a(BarrierSet::CardTableBarrierSet) 3819 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers() 3820 && ReduceInitialCardMarks; 3821 } 3822 3823 void GraphKit::sync_kit(IdealKit& ideal) { 3824 set_all_memory(__ merged_memory()); 3825 set_i_o(__ i_o()); 3826 set_control(__ ctrl()); 3827 } 3828 3829 void GraphKit::final_sync(IdealKit& ideal) { 3830 // Final sync IdealKit and graphKit. 3831 sync_kit(ideal); 3832 } 3833 3834 Node* GraphKit::byte_map_base_node() { 3835 // Get base of card map 3836 jbyte* card_table_base = ci_card_table_address(); 3837 if (card_table_base != NULL) { 3838 return makecon(TypeRawPtr::make((address)card_table_base)); 3839 } else { 3840 return null(); 3841 } 3842 } 3843 3844 // vanilla/CMS post barrier 3845 // Insert a write-barrier store. This is to let generational GC work; we have 3846 // to flag all oop-stores before the next GC point. 3847 void GraphKit::write_barrier_post(Node* oop_store, 3848 Node* obj, 3849 Node* adr, 3850 uint adr_idx, 3851 Node* val, 3852 bool use_precise) { 3853 // No store check needed if we're storing a NULL or an old object 3854 // (latter case is probably a string constant). The concurrent 3855 // mark sweep garbage collector, however, needs to have all nonNull 3856 // oop updates flagged via card-marks. 3857 if (val != NULL && val->is_Con()) { 3858 // must be either an oop or NULL 3859 const Type* t = val->bottom_type(); 3860 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3861 // stores of null never (?) need barriers 3862 return; 3863 } 3864 3865 if (use_ReduceInitialCardMarks() 3866 && obj == just_allocated_object(control())) { 3867 // We can skip marks on a freshly-allocated object in Eden. 3868 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 3869 // That routine informs GC to take appropriate compensating steps, 3870 // upon a slow-path allocation, so as to make this card-mark 3871 // elision safe. 3872 return; 3873 } 3874 3875 if (!use_precise) { 3876 // All card marks for a (non-array) instance are in one place: 3877 adr = obj; 3878 } 3879 // (Else it's an array (or unknown), and we want more precise card marks.) 3880 assert(adr != NULL, ""); 3881 3882 IdealKit ideal(this, true); 3883 3884 // Convert the pointer to an int prior to doing math on it 3885 Node* cast = __ CastPX(__ ctrl(), adr); 3886 3887 // Divide by card size 3888 assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet), 3889 "Only one we handle so far."); 3890 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 3891 3892 // Combine card table base and card offset 3893 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3894 3895 // Get the alias_index for raw card-mark memory 3896 int adr_type = Compile::AliasIdxRaw; 3897 Node* zero = __ ConI(0); // Dirty card value 3898 BasicType bt = T_BYTE; 3899 3900 if (UseConcMarkSweepGC && UseCondCardMark) { 3901 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3902 __ sync_kit(this); 3903 } 3904 3905 if (UseCondCardMark) { 3906 // The classic GC reference write barrier is typically implemented 3907 // as a store into the global card mark table. Unfortunately 3908 // unconditional stores can result in false sharing and excessive 3909 // coherence traffic as well as false transactional aborts. 3910 // UseCondCardMark enables MP "polite" conditional card mark 3911 // stores. In theory we could relax the load from ctrl() to 3912 // no_ctrl, but that doesn't buy much latitude. 3913 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3914 __ if_then(card_val, BoolTest::ne, zero); 3915 } 3916 3917 // Smash zero into card 3918 if( !UseConcMarkSweepGC ) { 3919 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3920 } else { 3921 // Specialized path for CM store barrier 3922 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3923 } 3924 3925 if (UseCondCardMark) { 3926 __ end_if(); 3927 } 3928 3929 // Final sync IdealKit and GraphKit. 3930 final_sync(ideal); 3931 } 3932 /* 3933 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3934 * required by SATB to make sure all objects live at the start of the 3935 * marking are kept alive, all reference updates need to any previous 3936 * reference stored before writing. 3937 * 3938 * If the previous value is NULL there is no need to save the old value. 3939 * References that are NULL are filtered during runtime by the barrier 3940 * code to avoid unnecessary queuing. 3941 * 3942 * However in the case of newly allocated objects it might be possible to 3943 * prove that the reference about to be overwritten is NULL during compile 3944 * time and avoid adding the barrier code completely. 3945 * 3946 * The compiler needs to determine that the object in which a field is about 3947 * to be written is newly allocated, and that no prior store to the same field 3948 * has happened since the allocation. 3949 * 3950 * Returns true if the pre-barrier can be removed 3951 */ 3952 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3953 BasicType bt, uint adr_idx) { 3954 intptr_t offset = 0; 3955 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3956 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3957 3958 if (offset == Type::OffsetBot) { 3959 return false; // cannot unalias unless there are precise offsets 3960 } 3961 3962 if (alloc == NULL) { 3963 return false; // No allocation found 3964 } 3965 3966 intptr_t size_in_bytes = type2aelembytes(bt); 3967 3968 Node* mem = memory(adr_idx); // start searching here... 3969 3970 for (int cnt = 0; cnt < 50; cnt++) { 3971 3972 if (mem->is_Store()) { 3973 3974 Node* st_adr = mem->in(MemNode::Address); 3975 intptr_t st_offset = 0; 3976 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3977 3978 if (st_base == NULL) { 3979 break; // inscrutable pointer 3980 } 3981 3982 // Break we have found a store with same base and offset as ours so break 3983 if (st_base == base && st_offset == offset) { 3984 break; 3985 } 3986 3987 if (st_offset != offset && st_offset != Type::OffsetBot) { 3988 const int MAX_STORE = BytesPerLong; 3989 if (st_offset >= offset + size_in_bytes || 3990 st_offset <= offset - MAX_STORE || 3991 st_offset <= offset - mem->as_Store()->memory_size()) { 3992 // Success: The offsets are provably independent. 3993 // (You may ask, why not just test st_offset != offset and be done? 3994 // The answer is that stores of different sizes can co-exist 3995 // in the same sequence of RawMem effects. We sometimes initialize 3996 // a whole 'tile' of array elements with a single jint or jlong.) 3997 mem = mem->in(MemNode::Memory); 3998 continue; // advance through independent store memory 3999 } 4000 } 4001 4002 if (st_base != base 4003 && MemNode::detect_ptr_independence(base, alloc, st_base, 4004 AllocateNode::Ideal_allocation(st_base, phase), 4005 phase)) { 4006 // Success: The bases are provably independent. 4007 mem = mem->in(MemNode::Memory); 4008 continue; // advance through independent store memory 4009 } 4010 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4011 4012 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4013 AllocateNode* st_alloc = st_init->allocation(); 4014 4015 // Make sure that we are looking at the same allocation site. 4016 // The alloc variable is guaranteed to not be null here from earlier check. 4017 if (alloc == st_alloc) { 4018 // Check that the initialization is storing NULL so that no previous store 4019 // has been moved up and directly write a reference 4020 Node* captured_store = st_init->find_captured_store(offset, 4021 type2aelembytes(T_OBJECT), 4022 phase); 4023 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4024 return true; 4025 } 4026 } 4027 } 4028 4029 // Unless there is an explicit 'continue', we must bail out here, 4030 // because 'mem' is an inscrutable memory state (e.g., a call). 4031 break; 4032 } 4033 4034 return false; 4035 } 4036 4037 // G1 pre/post barriers 4038 void GraphKit::g1_write_barrier_pre(bool do_load, 4039 Node* obj, 4040 Node* adr, 4041 uint alias_idx, 4042 Node* val, 4043 const TypeOopPtr* val_type, 4044 Node* pre_val, 4045 BasicType bt) { 4046 4047 // Some sanity checks 4048 // Note: val is unused in this routine. 4049 4050 if (do_load) { 4051 // We need to generate the load of the previous value 4052 assert(obj != NULL, "must have a base"); 4053 assert(adr != NULL, "where are loading from?"); 4054 assert(pre_val == NULL, "loaded already?"); 4055 assert(val_type != NULL, "need a type"); 4056 4057 if (use_ReduceInitialCardMarks() 4058 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4059 return; 4060 } 4061 4062 } else { 4063 // In this case both val_type and alias_idx are unused. 4064 assert(pre_val != NULL, "must be loaded already"); 4065 // Nothing to be done if pre_val is null. 4066 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4067 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4068 } 4069 assert(bt == T_OBJECT, "or we shouldn't be here"); 4070 4071 IdealKit ideal(this, true); 4072 4073 Node* tls = __ thread(); // ThreadLocalStorage 4074 4075 Node* no_ctrl = NULL; 4076 Node* no_base = __ top(); 4077 Node* zero = __ ConI(0); 4078 Node* zeroX = __ ConX(0); 4079 4080 float likely = PROB_LIKELY(0.999); 4081 float unlikely = PROB_UNLIKELY(0.999); 4082 4083 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4084 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4085 4086 // Offsets into the thread 4087 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 4088 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset()); 4089 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()); 4090 4091 // Now the actual pointers into the thread 4092 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4093 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4094 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4095 4096 // Now some of the values 4097 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4098 4099 // if (!marking) 4100 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4101 BasicType index_bt = TypeX_X->basic_type(); 4102 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4103 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4104 4105 if (do_load) { 4106 // load original value 4107 // alias_idx correct?? 4108 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4109 } 4110 4111 // if (pre_val != NULL) 4112 __ if_then(pre_val, BoolTest::ne, null()); { 4113 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4114 4115 // is the queue for this thread full? 4116 __ if_then(index, BoolTest::ne, zeroX, likely); { 4117 4118 // decrement the index 4119 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4120 4121 // Now get the buffer location we will log the previous value into and store it 4122 Node *log_addr = __ AddP(no_base, buffer, next_index); 4123 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4124 // update the index 4125 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4126 4127 } __ else_(); { 4128 4129 // logging buffer is full, call the runtime 4130 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4131 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4132 } __ end_if(); // (!index) 4133 } __ end_if(); // (pre_val != NULL) 4134 } __ end_if(); // (!marking) 4135 4136 // Final sync IdealKit and GraphKit. 4137 final_sync(ideal); 4138 } 4139 4140 /* 4141 * G1 similar to any GC with a Young Generation requires a way to keep track of 4142 * references from Old Generation to Young Generation to make sure all live 4143 * objects are found. G1 also requires to keep track of object references 4144 * between different regions to enable evacuation of old regions, which is done 4145 * as part of mixed collections. References are tracked in remembered sets and 4146 * is continuously updated as reference are written to with the help of the 4147 * post-barrier. 4148 * 4149 * To reduce the number of updates to the remembered set the post-barrier 4150 * filters updates to fields in objects located in the Young Generation, 4151 * the same region as the reference, when the NULL is being written or 4152 * if the card is already marked as dirty by an earlier write. 4153 * 4154 * Under certain circumstances it is possible to avoid generating the 4155 * post-barrier completely if it is possible during compile time to prove 4156 * the object is newly allocated and that no safepoint exists between the 4157 * allocation and the store. 4158 * 4159 * In the case of slow allocation the allocation code must handle the barrier 4160 * as part of the allocation in the case the allocated object is not located 4161 * in the nursery, this would happen for humongous objects. This is similar to 4162 * how CMS is required to handle this case, see the comments for the method 4163 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier. 4164 * A deferred card mark is required for these objects and handled in the above 4165 * mentioned methods. 4166 * 4167 * Returns true if the post barrier can be removed 4168 */ 4169 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4170 Node* adr) { 4171 intptr_t offset = 0; 4172 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4173 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4174 4175 if (offset == Type::OffsetBot) { 4176 return false; // cannot unalias unless there are precise offsets 4177 } 4178 4179 if (alloc == NULL) { 4180 return false; // No allocation found 4181 } 4182 4183 // Start search from Store node 4184 Node* mem = store->in(MemNode::Control); 4185 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4186 4187 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4188 AllocateNode* st_alloc = st_init->allocation(); 4189 4190 // Make sure we are looking at the same allocation 4191 if (alloc == st_alloc) { 4192 return true; 4193 } 4194 } 4195 4196 return false; 4197 } 4198 4199 // 4200 // Update the card table and add card address to the queue 4201 // 4202 void GraphKit::g1_mark_card(IdealKit& ideal, 4203 Node* card_adr, 4204 Node* oop_store, 4205 uint oop_alias_idx, 4206 Node* index, 4207 Node* index_adr, 4208 Node* buffer, 4209 const TypeFunc* tf) { 4210 4211 Node* zero = __ ConI(0); 4212 Node* zeroX = __ ConX(0); 4213 Node* no_base = __ top(); 4214 BasicType card_bt = T_BYTE; 4215 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4216 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4217 4218 // Now do the queue work 4219 __ if_then(index, BoolTest::ne, zeroX); { 4220 4221 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4222 Node* log_addr = __ AddP(no_base, buffer, next_index); 4223 4224 // Order, see storeCM. 4225 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4226 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4227 4228 } __ else_(); { 4229 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4230 } __ end_if(); 4231 4232 } 4233 4234 void GraphKit::g1_write_barrier_post(Node* oop_store, 4235 Node* obj, 4236 Node* adr, 4237 uint alias_idx, 4238 Node* val, 4239 BasicType bt, 4240 bool use_precise) { 4241 // If we are writing a NULL then we need no post barrier 4242 4243 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4244 // Must be NULL 4245 const Type* t = val->bottom_type(); 4246 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4247 // No post barrier if writing NULLx 4248 return; 4249 } 4250 4251 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4252 // We can skip marks on a freshly-allocated object in Eden. 4253 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 4254 // That routine informs GC to take appropriate compensating steps, 4255 // upon a slow-path allocation, so as to make this card-mark 4256 // elision safe. 4257 return; 4258 } 4259 4260 if (use_ReduceInitialCardMarks() 4261 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4262 return; 4263 } 4264 4265 if (!use_precise) { 4266 // All card marks for a (non-array) instance are in one place: 4267 adr = obj; 4268 } 4269 // (Else it's an array (or unknown), and we want more precise card marks.) 4270 assert(adr != NULL, ""); 4271 4272 IdealKit ideal(this, true); 4273 4274 Node* tls = __ thread(); // ThreadLocalStorage 4275 4276 Node* no_base = __ top(); 4277 float likely = PROB_LIKELY(0.999); 4278 float unlikely = PROB_UNLIKELY(0.999); 4279 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val()); 4280 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val()); 4281 Node* zeroX = __ ConX(0); 4282 4283 // Get the alias_index for raw card-mark memory 4284 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4285 4286 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4287 4288 // Offsets into the thread 4289 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset()); 4290 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset()); 4291 4292 // Pointers into the thread 4293 4294 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4295 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4296 4297 // Now some values 4298 // Use ctrl to avoid hoisting these values past a safepoint, which could 4299 // potentially reset these fields in the JavaThread. 4300 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4301 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4302 4303 // Convert the store obj pointer to an int prior to doing math on it 4304 // Must use ctrl to prevent "integerized oop" existing across safepoint 4305 Node* cast = __ CastPX(__ ctrl(), adr); 4306 4307 // Divide pointer by card size 4308 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 4309 4310 // Combine card table base and card offset 4311 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4312 4313 // If we know the value being stored does it cross regions? 4314 4315 if (val != NULL) { 4316 // Does the store cause us to cross regions? 4317 4318 // Should be able to do an unsigned compare of region_size instead of 4319 // and extra shift. Do we have an unsigned compare?? 4320 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4321 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4322 4323 // if (xor_res == 0) same region so skip 4324 __ if_then(xor_res, BoolTest::ne, zeroX); { 4325 4326 // No barrier if we are storing a NULL 4327 __ if_then(val, BoolTest::ne, null(), unlikely); { 4328 4329 // Ok must mark the card if not already dirty 4330 4331 // load the original value of the card 4332 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4333 4334 __ if_then(card_val, BoolTest::ne, young_card); { 4335 sync_kit(ideal); 4336 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4337 insert_mem_bar(Op_MemBarVolatile, oop_store); 4338 __ sync_kit(this); 4339 4340 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4341 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4342 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4343 } __ end_if(); 4344 } __ end_if(); 4345 } __ end_if(); 4346 } __ end_if(); 4347 } else { 4348 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4349 // We don't need a barrier here if the destination is a newly allocated object 4350 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4351 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()). 4352 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4353 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4354 __ if_then(card_val, BoolTest::ne, young_card); { 4355 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4356 } __ end_if(); 4357 } 4358 4359 // Final sync IdealKit and GraphKit. 4360 final_sync(ideal); 4361 } 4362 #undef __ 4363 4364 4365 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4366 Node* len = load_array_length(load_String_value(ctrl, str)); 4367 Node* coder = load_String_coder(ctrl, str); 4368 // Divide length by 2 if coder is UTF16 4369 return _gvn.transform(new RShiftINode(len, coder)); 4370 } 4371 4372 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4373 int value_offset = java_lang_String::value_offset_in_bytes(); 4374 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4375 false, NULL, 0); 4376 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4377 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4378 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4379 ciTypeArrayKlass::make(T_BYTE), true, 0); 4380 int value_field_idx = C->get_alias_index(value_field_type); 4381 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4382 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4383 // String.value field is known to be @Stable. 4384 if (UseImplicitStableValues) { 4385 load = cast_array_to_stable(load, value_type); 4386 } 4387 return load; 4388 } 4389 4390 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4391 if (!CompactStrings) { 4392 return intcon(java_lang_String::CODER_UTF16); 4393 } 4394 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4395 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4396 false, NULL, 0); 4397 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4398 int coder_field_idx = C->get_alias_index(coder_field_type); 4399 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4400 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4401 } 4402 4403 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4404 int value_offset = java_lang_String::value_offset_in_bytes(); 4405 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4406 false, NULL, 0); 4407 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4408 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4409 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4410 } 4411 4412 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4413 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4414 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4415 false, NULL, 0); 4416 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4417 int coder_field_idx = C->get_alias_index(coder_field_type); 4418 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4419 value, T_BYTE, coder_field_idx, MemNode::unordered); 4420 } 4421 4422 // Capture src and dst memory state with a MergeMemNode 4423 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4424 if (src_type == dst_type) { 4425 // Types are equal, we don't need a MergeMemNode 4426 return memory(src_type); 4427 } 4428 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4429 record_for_igvn(merge); // fold it up later, if possible 4430 int src_idx = C->get_alias_index(src_type); 4431 int dst_idx = C->get_alias_index(dst_type); 4432 merge->set_memory_at(src_idx, memory(src_idx)); 4433 merge->set_memory_at(dst_idx, memory(dst_idx)); 4434 return merge; 4435 } 4436 4437 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4438 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4439 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4440 // If input and output memory types differ, capture both states to preserve 4441 // the dependency between preceding and subsequent loads/stores. 4442 // For example, the following program: 4443 // StoreB 4444 // compress_string 4445 // LoadB 4446 // has this memory graph (use->def): 4447 // LoadB -> compress_string -> CharMem 4448 // ... -> StoreB -> ByteMem 4449 // The intrinsic hides the dependency between LoadB and StoreB, causing 4450 // the load to read from memory not containing the result of the StoreB. 4451 // The correct memory graph should look like this: 4452 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4453 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4454 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4455 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4456 set_memory(res_mem, TypeAryPtr::BYTES); 4457 return str; 4458 } 4459 4460 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4461 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4462 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4463 // Capture src and dst memory (see comment in 'compress_string'). 4464 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4465 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4466 set_memory(_gvn.transform(str), dst_type); 4467 } 4468 4469 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4470 /** 4471 * int i_char = start; 4472 * for (int i_byte = 0; i_byte < count; i_byte++) { 4473 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4474 * } 4475 */ 4476 add_predicate(); 4477 RegionNode* head = new RegionNode(3); 4478 head->init_req(1, control()); 4479 gvn().set_type(head, Type::CONTROL); 4480 record_for_igvn(head); 4481 4482 Node* i_byte = new PhiNode(head, TypeInt::INT); 4483 i_byte->init_req(1, intcon(0)); 4484 gvn().set_type(i_byte, TypeInt::INT); 4485 record_for_igvn(i_byte); 4486 4487 Node* i_char = new PhiNode(head, TypeInt::INT); 4488 i_char->init_req(1, start); 4489 gvn().set_type(i_char, TypeInt::INT); 4490 record_for_igvn(i_char); 4491 4492 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4493 gvn().set_type(mem, Type::MEMORY); 4494 record_for_igvn(mem); 4495 set_control(head); 4496 set_memory(mem, TypeAryPtr::BYTES); 4497 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4498 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4499 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4500 false, false, true /* mismatched */); 4501 4502 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4503 head->init_req(2, IfTrue(iff)); 4504 mem->init_req(2, st); 4505 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4506 i_char->init_req(2, AddI(i_char, intcon(2))); 4507 4508 set_control(IfFalse(iff)); 4509 set_memory(st, TypeAryPtr::BYTES); 4510 } 4511 4512 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4513 if (!field->is_constant()) { 4514 return NULL; // Field not marked as constant. 4515 } 4516 ciInstance* holder = NULL; 4517 if (!field->is_static()) { 4518 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4519 if (const_oop != NULL && const_oop->is_instance()) { 4520 holder = const_oop->as_instance(); 4521 } 4522 } 4523 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4524 /*is_unsigned_load=*/false); 4525 if (con_type != NULL) { 4526 return makecon(con_type); 4527 } 4528 return NULL; 4529 } 4530 4531 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4532 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4533 // assumption of CCP analysis. 4534 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4535 }