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/g1CardTable.hpp" 53 #include "gc/g1/g1ThreadLocalData.hpp" 54 #include "gc/g1/heapRegion.hpp" 55 #endif // INCLUDE_ALL_GCS 56 57 //----------------------------GraphKit----------------------------------------- 58 // Main utility constructor. 59 GraphKit::GraphKit(JVMState* jvms) 60 : Phase(Phase::Parser), 61 _env(C->env()), 62 _gvn(*C->initial_gvn()) 63 { 64 _exceptions = jvms->map()->next_exception(); 65 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 66 set_jvms(jvms); 67 } 68 69 // Private constructor for parser. 70 GraphKit::GraphKit() 71 : Phase(Phase::Parser), 72 _env(C->env()), 73 _gvn(*C->initial_gvn()) 74 { 75 _exceptions = NULL; 76 set_map(NULL); 77 debug_only(_sp = -99); 78 debug_only(set_bci(-99)); 79 } 80 81 82 83 //---------------------------clean_stack--------------------------------------- 84 // Clear away rubbish from the stack area of the JVM state. 85 // This destroys any arguments that may be waiting on the stack. 86 void GraphKit::clean_stack(int from_sp) { 87 SafePointNode* map = this->map(); 88 JVMState* jvms = this->jvms(); 89 int stk_size = jvms->stk_size(); 90 int stkoff = jvms->stkoff(); 91 Node* top = this->top(); 92 for (int i = from_sp; i < stk_size; i++) { 93 if (map->in(stkoff + i) != top) { 94 map->set_req(stkoff + i, top); 95 } 96 } 97 } 98 99 100 //--------------------------------sync_jvms----------------------------------- 101 // Make sure our current jvms agrees with our parse state. 102 JVMState* GraphKit::sync_jvms() const { 103 JVMState* jvms = this->jvms(); 104 jvms->set_bci(bci()); // Record the new bci in the JVMState 105 jvms->set_sp(sp()); // Record the new sp in the JVMState 106 assert(jvms_in_sync(), "jvms is now in sync"); 107 return jvms; 108 } 109 110 //--------------------------------sync_jvms_for_reexecute--------------------- 111 // Make sure our current jvms agrees with our parse state. This version 112 // uses the reexecute_sp for reexecuting bytecodes. 113 JVMState* GraphKit::sync_jvms_for_reexecute() { 114 JVMState* jvms = this->jvms(); 115 jvms->set_bci(bci()); // Record the new bci in the JVMState 116 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 117 return jvms; 118 } 119 120 #ifdef ASSERT 121 bool GraphKit::jvms_in_sync() const { 122 Parse* parse = is_Parse(); 123 if (parse == NULL) { 124 if (bci() != jvms()->bci()) return false; 125 if (sp() != (int)jvms()->sp()) return false; 126 return true; 127 } 128 if (jvms()->method() != parse->method()) return false; 129 if (jvms()->bci() != parse->bci()) return false; 130 int jvms_sp = jvms()->sp(); 131 if (jvms_sp != parse->sp()) return false; 132 int jvms_depth = jvms()->depth(); 133 if (jvms_depth != parse->depth()) return false; 134 return true; 135 } 136 137 // Local helper checks for special internal merge points 138 // used to accumulate and merge exception states. 139 // They are marked by the region's in(0) edge being the map itself. 140 // Such merge points must never "escape" into the parser at large, 141 // until they have been handed to gvn.transform. 142 static bool is_hidden_merge(Node* reg) { 143 if (reg == NULL) return false; 144 if (reg->is_Phi()) { 145 reg = reg->in(0); 146 if (reg == NULL) return false; 147 } 148 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 149 } 150 151 void GraphKit::verify_map() const { 152 if (map() == NULL) return; // null map is OK 153 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 154 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 155 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 156 } 157 158 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 159 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 160 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 161 } 162 #endif 163 164 //---------------------------stop_and_kill_map--------------------------------- 165 // Set _map to NULL, signalling a stop to further bytecode execution. 166 // First smash the current map's control to a constant, to mark it dead. 167 void GraphKit::stop_and_kill_map() { 168 SafePointNode* dead_map = stop(); 169 if (dead_map != NULL) { 170 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 171 assert(dead_map->is_killed(), "must be so marked"); 172 } 173 } 174 175 176 //--------------------------------stopped-------------------------------------- 177 // Tell if _map is NULL, or control is top. 178 bool GraphKit::stopped() { 179 if (map() == NULL) return true; 180 else if (control() == top()) return true; 181 else return false; 182 } 183 184 185 //-----------------------------has_ex_handler---------------------------------- 186 // Tell if this method or any caller method has exception handlers. 187 bool GraphKit::has_ex_handler() { 188 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 189 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 190 return true; 191 } 192 } 193 return false; 194 } 195 196 //------------------------------save_ex_oop------------------------------------ 197 // Save an exception without blowing stack contents or other JVM state. 198 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 199 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 200 ex_map->add_req(ex_oop); 201 debug_only(verify_exception_state(ex_map)); 202 } 203 204 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 205 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 206 Node* ex_oop = ex_map->in(ex_map->req()-1); 207 if (clear_it) ex_map->del_req(ex_map->req()-1); 208 return ex_oop; 209 } 210 211 //-----------------------------saved_ex_oop------------------------------------ 212 // Recover a saved exception from its map. 213 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 214 return common_saved_ex_oop(ex_map, false); 215 } 216 217 //--------------------------clear_saved_ex_oop--------------------------------- 218 // Erase a previously saved exception from its map. 219 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 220 return common_saved_ex_oop(ex_map, true); 221 } 222 223 #ifdef ASSERT 224 //---------------------------has_saved_ex_oop---------------------------------- 225 // Erase a previously saved exception from its map. 226 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 227 return ex_map->req() == ex_map->jvms()->endoff()+1; 228 } 229 #endif 230 231 //-------------------------make_exception_state-------------------------------- 232 // Turn the current JVM state into an exception state, appending the ex_oop. 233 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 234 sync_jvms(); 235 SafePointNode* ex_map = stop(); // do not manipulate this map any more 236 set_saved_ex_oop(ex_map, ex_oop); 237 return ex_map; 238 } 239 240 241 //--------------------------add_exception_state-------------------------------- 242 // Add an exception to my list of exceptions. 243 void GraphKit::add_exception_state(SafePointNode* ex_map) { 244 if (ex_map == NULL || ex_map->control() == top()) { 245 return; 246 } 247 #ifdef ASSERT 248 verify_exception_state(ex_map); 249 if (has_exceptions()) { 250 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 251 } 252 #endif 253 254 // If there is already an exception of exactly this type, merge with it. 255 // In particular, null-checks and other low-level exceptions common up here. 256 Node* ex_oop = saved_ex_oop(ex_map); 257 const Type* ex_type = _gvn.type(ex_oop); 258 if (ex_oop == top()) { 259 // No action needed. 260 return; 261 } 262 assert(ex_type->isa_instptr(), "exception must be an instance"); 263 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 264 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 265 // We check sp also because call bytecodes can generate exceptions 266 // both before and after arguments are popped! 267 if (ex_type2 == ex_type 268 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 269 combine_exception_states(ex_map, e2); 270 return; 271 } 272 } 273 274 // No pre-existing exception of the same type. Chain it on the list. 275 push_exception_state(ex_map); 276 } 277 278 //-----------------------add_exception_states_from----------------------------- 279 void GraphKit::add_exception_states_from(JVMState* jvms) { 280 SafePointNode* ex_map = jvms->map()->next_exception(); 281 if (ex_map != NULL) { 282 jvms->map()->set_next_exception(NULL); 283 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 284 next_map = ex_map->next_exception(); 285 ex_map->set_next_exception(NULL); 286 add_exception_state(ex_map); 287 } 288 } 289 } 290 291 //-----------------------transfer_exceptions_into_jvms------------------------- 292 JVMState* GraphKit::transfer_exceptions_into_jvms() { 293 if (map() == NULL) { 294 // We need a JVMS to carry the exceptions, but the map has gone away. 295 // Create a scratch JVMS, cloned from any of the exception states... 296 if (has_exceptions()) { 297 _map = _exceptions; 298 _map = clone_map(); 299 _map->set_next_exception(NULL); 300 clear_saved_ex_oop(_map); 301 debug_only(verify_map()); 302 } else { 303 // ...or created from scratch 304 JVMState* jvms = new (C) JVMState(_method, NULL); 305 jvms->set_bci(_bci); 306 jvms->set_sp(_sp); 307 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 308 set_jvms(jvms); 309 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 310 set_all_memory(top()); 311 while (map()->req() < jvms->endoff()) map()->add_req(top()); 312 } 313 // (This is a kludge, in case you didn't notice.) 314 set_control(top()); 315 } 316 JVMState* jvms = sync_jvms(); 317 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 318 jvms->map()->set_next_exception(_exceptions); 319 _exceptions = NULL; // done with this set of exceptions 320 return jvms; 321 } 322 323 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 324 assert(is_hidden_merge(dstphi), "must be a special merge node"); 325 assert(is_hidden_merge(srcphi), "must be a special merge node"); 326 uint limit = srcphi->req(); 327 for (uint i = PhiNode::Input; i < limit; i++) { 328 dstphi->add_req(srcphi->in(i)); 329 } 330 } 331 static inline void add_one_req(Node* dstphi, Node* src) { 332 assert(is_hidden_merge(dstphi), "must be a special merge node"); 333 assert(!is_hidden_merge(src), "must not be a special merge node"); 334 dstphi->add_req(src); 335 } 336 337 //-----------------------combine_exception_states------------------------------ 338 // This helper function combines exception states by building phis on a 339 // specially marked state-merging region. These regions and phis are 340 // untransformed, and can build up gradually. The region is marked by 341 // having a control input of its exception map, rather than NULL. Such 342 // regions do not appear except in this function, and in use_exception_state. 343 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 344 if (failing()) return; // dying anyway... 345 JVMState* ex_jvms = ex_map->_jvms; 346 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 347 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 348 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 349 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 350 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 351 assert(ex_map->req() == phi_map->req(), "matching maps"); 352 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 353 Node* hidden_merge_mark = root(); 354 Node* region = phi_map->control(); 355 MergeMemNode* phi_mem = phi_map->merged_memory(); 356 MergeMemNode* ex_mem = ex_map->merged_memory(); 357 if (region->in(0) != hidden_merge_mark) { 358 // The control input is not (yet) a specially-marked region in phi_map. 359 // Make it so, and build some phis. 360 region = new RegionNode(2); 361 _gvn.set_type(region, Type::CONTROL); 362 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 363 region->init_req(1, phi_map->control()); 364 phi_map->set_control(region); 365 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 366 record_for_igvn(io_phi); 367 _gvn.set_type(io_phi, Type::ABIO); 368 phi_map->set_i_o(io_phi); 369 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 370 Node* m = mms.memory(); 371 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 372 record_for_igvn(m_phi); 373 _gvn.set_type(m_phi, Type::MEMORY); 374 mms.set_memory(m_phi); 375 } 376 } 377 378 // Either or both of phi_map and ex_map might already be converted into phis. 379 Node* ex_control = ex_map->control(); 380 // if there is special marking on ex_map also, we add multiple edges from src 381 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 382 // how wide was the destination phi_map, originally? 383 uint orig_width = region->req(); 384 385 if (add_multiple) { 386 add_n_reqs(region, ex_control); 387 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 388 } else { 389 // ex_map has no merges, so we just add single edges everywhere 390 add_one_req(region, ex_control); 391 add_one_req(phi_map->i_o(), ex_map->i_o()); 392 } 393 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 394 if (mms.is_empty()) { 395 // get a copy of the base memory, and patch some inputs into it 396 const TypePtr* adr_type = mms.adr_type(C); 397 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 398 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 399 mms.set_memory(phi); 400 // Prepare to append interesting stuff onto the newly sliced phi: 401 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 402 } 403 // Append stuff from ex_map: 404 if (add_multiple) { 405 add_n_reqs(mms.memory(), mms.memory2()); 406 } else { 407 add_one_req(mms.memory(), mms.memory2()); 408 } 409 } 410 uint limit = ex_map->req(); 411 for (uint i = TypeFunc::Parms; i < limit; i++) { 412 // Skip everything in the JVMS after tos. (The ex_oop follows.) 413 if (i == tos) i = ex_jvms->monoff(); 414 Node* src = ex_map->in(i); 415 Node* dst = phi_map->in(i); 416 if (src != dst) { 417 PhiNode* phi; 418 if (dst->in(0) != region) { 419 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 420 record_for_igvn(phi); 421 _gvn.set_type(phi, phi->type()); 422 phi_map->set_req(i, dst); 423 // Prepare to append interesting stuff onto the new phi: 424 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 425 } else { 426 assert(dst->is_Phi(), "nobody else uses a hidden region"); 427 phi = dst->as_Phi(); 428 } 429 if (add_multiple && src->in(0) == ex_control) { 430 // Both are phis. 431 add_n_reqs(dst, src); 432 } else { 433 while (dst->req() < region->req()) add_one_req(dst, src); 434 } 435 const Type* srctype = _gvn.type(src); 436 if (phi->type() != srctype) { 437 const Type* dsttype = phi->type()->meet_speculative(srctype); 438 if (phi->type() != dsttype) { 439 phi->set_type(dsttype); 440 _gvn.set_type(phi, dsttype); 441 } 442 } 443 } 444 } 445 phi_map->merge_replaced_nodes_with(ex_map); 446 } 447 448 //--------------------------use_exception_state-------------------------------- 449 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 450 if (failing()) { stop(); return top(); } 451 Node* region = phi_map->control(); 452 Node* hidden_merge_mark = root(); 453 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 454 Node* ex_oop = clear_saved_ex_oop(phi_map); 455 if (region->in(0) == hidden_merge_mark) { 456 // Special marking for internal ex-states. Process the phis now. 457 region->set_req(0, region); // now it's an ordinary region 458 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 459 // Note: Setting the jvms also sets the bci and sp. 460 set_control(_gvn.transform(region)); 461 uint tos = jvms()->stkoff() + sp(); 462 for (uint i = 1; i < tos; i++) { 463 Node* x = phi_map->in(i); 464 if (x->in(0) == region) { 465 assert(x->is_Phi(), "expected a special phi"); 466 phi_map->set_req(i, _gvn.transform(x)); 467 } 468 } 469 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 470 Node* x = mms.memory(); 471 if (x->in(0) == region) { 472 assert(x->is_Phi(), "nobody else uses a hidden region"); 473 mms.set_memory(_gvn.transform(x)); 474 } 475 } 476 if (ex_oop->in(0) == region) { 477 assert(ex_oop->is_Phi(), "expected a special phi"); 478 ex_oop = _gvn.transform(ex_oop); 479 } 480 } else { 481 set_jvms(phi_map->jvms()); 482 } 483 484 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 485 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 486 return ex_oop; 487 } 488 489 //---------------------------------java_bc------------------------------------- 490 Bytecodes::Code GraphKit::java_bc() const { 491 ciMethod* method = this->method(); 492 int bci = this->bci(); 493 if (method != NULL && bci != InvocationEntryBci) 494 return method->java_code_at_bci(bci); 495 else 496 return Bytecodes::_illegal; 497 } 498 499 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 500 bool must_throw) { 501 // if the exception capability is set, then we will generate code 502 // to check the JavaThread.should_post_on_exceptions flag to see 503 // if we actually need to report exception events (for this 504 // thread). If we don't need to report exception events, we will 505 // take the normal fast path provided by add_exception_events. If 506 // exception event reporting is enabled for this thread, we will 507 // take the uncommon_trap in the BuildCutout below. 508 509 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 510 Node* jthread = _gvn.transform(new ThreadLocalNode()); 511 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 512 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 513 514 // Test the should_post_on_exceptions_flag vs. 0 515 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 516 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 517 518 // Branch to slow_path if should_post_on_exceptions_flag was true 519 { BuildCutout unless(this, tst, PROB_MAX); 520 // Do not try anything fancy if we're notifying the VM on every throw. 521 // Cf. case Bytecodes::_athrow in parse2.cpp. 522 uncommon_trap(reason, Deoptimization::Action_none, 523 (ciKlass*)NULL, (char*)NULL, must_throw); 524 } 525 526 } 527 528 //------------------------------builtin_throw---------------------------------- 529 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 530 bool must_throw = true; 531 532 if (env()->jvmti_can_post_on_exceptions()) { 533 // check if we must post exception events, take uncommon trap if so 534 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 535 // here if should_post_on_exceptions is false 536 // continue on with the normal codegen 537 } 538 539 // If this particular condition has not yet happened at this 540 // bytecode, then use the uncommon trap mechanism, and allow for 541 // a future recompilation if several traps occur here. 542 // If the throw is hot, try to use a more complicated inline mechanism 543 // which keeps execution inside the compiled code. 544 bool treat_throw_as_hot = false; 545 ciMethodData* md = method()->method_data(); 546 547 if (ProfileTraps) { 548 if (too_many_traps(reason)) { 549 treat_throw_as_hot = true; 550 } 551 // (If there is no MDO at all, assume it is early in 552 // execution, and that any deopts are part of the 553 // startup transient, and don't need to be remembered.) 554 555 // Also, if there is a local exception handler, treat all throws 556 // as hot if there has been at least one in this method. 557 if (C->trap_count(reason) != 0 558 && method()->method_data()->trap_count(reason) != 0 559 && has_ex_handler()) { 560 treat_throw_as_hot = true; 561 } 562 } 563 564 // If this throw happens frequently, an uncommon trap might cause 565 // a performance pothole. If there is a local exception handler, 566 // and if this particular bytecode appears to be deoptimizing often, 567 // let us handle the throw inline, with a preconstructed instance. 568 // Note: If the deopt count has blown up, the uncommon trap 569 // runtime is going to flush this nmethod, not matter what. 570 if (treat_throw_as_hot 571 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 572 // If the throw is local, we use a pre-existing instance and 573 // punt on the backtrace. This would lead to a missing backtrace 574 // (a repeat of 4292742) if the backtrace object is ever asked 575 // for its backtrace. 576 // Fixing this remaining case of 4292742 requires some flavor of 577 // escape analysis. Leave that for the future. 578 ciInstance* ex_obj = NULL; 579 switch (reason) { 580 case Deoptimization::Reason_null_check: 581 ex_obj = env()->NullPointerException_instance(); 582 break; 583 case Deoptimization::Reason_div0_check: 584 ex_obj = env()->ArithmeticException_instance(); 585 break; 586 case Deoptimization::Reason_range_check: 587 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 588 break; 589 case Deoptimization::Reason_class_check: 590 if (java_bc() == Bytecodes::_aastore) { 591 ex_obj = env()->ArrayStoreException_instance(); 592 } else { 593 ex_obj = env()->ClassCastException_instance(); 594 } 595 break; 596 default: 597 break; 598 } 599 if (failing()) { stop(); return; } // exception allocation might fail 600 if (ex_obj != NULL) { 601 // Cheat with a preallocated exception object. 602 if (C->log() != NULL) 603 C->log()->elem("hot_throw preallocated='1' reason='%s'", 604 Deoptimization::trap_reason_name(reason)); 605 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 606 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 607 608 // Clear the detail message of the preallocated exception object. 609 // Weblogic sometimes mutates the detail message of exceptions 610 // using reflection. 611 int offset = java_lang_Throwable::get_detailMessage_offset(); 612 const TypePtr* adr_typ = ex_con->add_offset(offset); 613 614 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 615 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 616 // Conservatively release stores of object references. 617 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 618 619 add_exception_state(make_exception_state(ex_node)); 620 return; 621 } 622 } 623 624 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 625 // It won't be much cheaper than bailing to the interp., since we'll 626 // have to pass up all the debug-info, and the runtime will have to 627 // create the stack trace. 628 629 // Usual case: Bail to interpreter. 630 // Reserve the right to recompile if we haven't seen anything yet. 631 632 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 633 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 634 if (treat_throw_as_hot 635 && (method()->method_data()->trap_recompiled_at(bci(), m) 636 || C->too_many_traps(reason))) { 637 // We cannot afford to take more traps here. Suffer in the interpreter. 638 if (C->log() != NULL) 639 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 640 Deoptimization::trap_reason_name(reason), 641 C->trap_count(reason)); 642 action = Deoptimization::Action_none; 643 } 644 645 // "must_throw" prunes the JVM state to include only the stack, if there 646 // are no local exception handlers. This should cut down on register 647 // allocation time and code size, by drastically reducing the number 648 // of in-edges on the call to the uncommon trap. 649 650 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 651 } 652 653 654 //----------------------------PreserveJVMState--------------------------------- 655 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 656 debug_only(kit->verify_map()); 657 _kit = kit; 658 _map = kit->map(); // preserve the map 659 _sp = kit->sp(); 660 kit->set_map(clone_map ? kit->clone_map() : NULL); 661 #ifdef ASSERT 662 _bci = kit->bci(); 663 Parse* parser = kit->is_Parse(); 664 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 665 _block = block; 666 #endif 667 } 668 PreserveJVMState::~PreserveJVMState() { 669 GraphKit* kit = _kit; 670 #ifdef ASSERT 671 assert(kit->bci() == _bci, "bci must not shift"); 672 Parse* parser = kit->is_Parse(); 673 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 674 assert(block == _block, "block must not shift"); 675 #endif 676 kit->set_map(_map); 677 kit->set_sp(_sp); 678 } 679 680 681 //-----------------------------BuildCutout------------------------------------- 682 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 683 : PreserveJVMState(kit) 684 { 685 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 686 SafePointNode* outer_map = _map; // preserved map is caller's 687 SafePointNode* inner_map = kit->map(); 688 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 689 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 690 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 691 } 692 BuildCutout::~BuildCutout() { 693 GraphKit* kit = _kit; 694 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 695 } 696 697 //---------------------------PreserveReexecuteState---------------------------- 698 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 699 assert(!kit->stopped(), "must call stopped() before"); 700 _kit = kit; 701 _sp = kit->sp(); 702 _reexecute = kit->jvms()->_reexecute; 703 } 704 PreserveReexecuteState::~PreserveReexecuteState() { 705 if (_kit->stopped()) return; 706 _kit->jvms()->_reexecute = _reexecute; 707 _kit->set_sp(_sp); 708 } 709 710 //------------------------------clone_map-------------------------------------- 711 // Implementation of PreserveJVMState 712 // 713 // Only clone_map(...) here. If this function is only used in the 714 // PreserveJVMState class we may want to get rid of this extra 715 // function eventually and do it all there. 716 717 SafePointNode* GraphKit::clone_map() { 718 if (map() == NULL) return NULL; 719 720 // Clone the memory edge first 721 Node* mem = MergeMemNode::make(map()->memory()); 722 gvn().set_type_bottom(mem); 723 724 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 725 JVMState* jvms = this->jvms(); 726 JVMState* clonejvms = jvms->clone_shallow(C); 727 clonemap->set_memory(mem); 728 clonemap->set_jvms(clonejvms); 729 clonejvms->set_map(clonemap); 730 record_for_igvn(clonemap); 731 gvn().set_type_bottom(clonemap); 732 return clonemap; 733 } 734 735 736 //-----------------------------set_map_clone----------------------------------- 737 void GraphKit::set_map_clone(SafePointNode* m) { 738 _map = m; 739 _map = clone_map(); 740 _map->set_next_exception(NULL); 741 debug_only(verify_map()); 742 } 743 744 745 //----------------------------kill_dead_locals--------------------------------- 746 // Detect any locals which are known to be dead, and force them to top. 747 void GraphKit::kill_dead_locals() { 748 // Consult the liveness information for the locals. If any 749 // of them are unused, then they can be replaced by top(). This 750 // should help register allocation time and cut down on the size 751 // of the deoptimization information. 752 753 // This call is made from many of the bytecode handling 754 // subroutines called from the Big Switch in do_one_bytecode. 755 // Every bytecode which might include a slow path is responsible 756 // for killing its dead locals. The more consistent we 757 // are about killing deads, the fewer useless phis will be 758 // constructed for them at various merge points. 759 760 // bci can be -1 (InvocationEntryBci). We return the entry 761 // liveness for the method. 762 763 if (method() == NULL || method()->code_size() == 0) { 764 // We are building a graph for a call to a native method. 765 // All locals are live. 766 return; 767 } 768 769 ResourceMark rm; 770 771 // Consult the liveness information for the locals. If any 772 // of them are unused, then they can be replaced by top(). This 773 // should help register allocation time and cut down on the size 774 // of the deoptimization information. 775 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 776 777 int len = (int)live_locals.size(); 778 assert(len <= jvms()->loc_size(), "too many live locals"); 779 for (int local = 0; local < len; local++) { 780 if (!live_locals.at(local)) { 781 set_local(local, top()); 782 } 783 } 784 } 785 786 #ifdef ASSERT 787 //-------------------------dead_locals_are_killed------------------------------ 788 // Return true if all dead locals are set to top in the map. 789 // Used to assert "clean" debug info at various points. 790 bool GraphKit::dead_locals_are_killed() { 791 if (method() == NULL || method()->code_size() == 0) { 792 // No locals need to be dead, so all is as it should be. 793 return true; 794 } 795 796 // Make sure somebody called kill_dead_locals upstream. 797 ResourceMark rm; 798 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 799 if (jvms->loc_size() == 0) continue; // no locals to consult 800 SafePointNode* map = jvms->map(); 801 ciMethod* method = jvms->method(); 802 int bci = jvms->bci(); 803 if (jvms == this->jvms()) { 804 bci = this->bci(); // it might not yet be synched 805 } 806 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 807 int len = (int)live_locals.size(); 808 if (!live_locals.is_valid() || len == 0) 809 // This method is trivial, or is poisoned by a breakpoint. 810 return true; 811 assert(len == jvms->loc_size(), "live map consistent with locals map"); 812 for (int local = 0; local < len; local++) { 813 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 814 if (PrintMiscellaneous && (Verbose || WizardMode)) { 815 tty->print_cr("Zombie local %d: ", local); 816 jvms->dump(); 817 } 818 return false; 819 } 820 } 821 } 822 return true; 823 } 824 825 #endif //ASSERT 826 827 // Helper function for enforcing certain bytecodes to reexecute if 828 // deoptimization happens 829 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 830 ciMethod* cur_method = jvms->method(); 831 int cur_bci = jvms->bci(); 832 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 833 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 834 return Interpreter::bytecode_should_reexecute(code) || 835 (is_anewarray && code == Bytecodes::_multianewarray); 836 // Reexecute _multianewarray bytecode which was replaced with 837 // sequence of [a]newarray. See Parse::do_multianewarray(). 838 // 839 // Note: interpreter should not have it set since this optimization 840 // is limited by dimensions and guarded by flag so in some cases 841 // multianewarray() runtime calls will be generated and 842 // the bytecode should not be reexecutes (stack will not be reset). 843 } else 844 return false; 845 } 846 847 // Helper function for adding JVMState and debug information to node 848 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 849 // Add the safepoint edges to the call (or other safepoint). 850 851 // Make sure dead locals are set to top. This 852 // should help register allocation time and cut down on the size 853 // of the deoptimization information. 854 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 855 856 // Walk the inline list to fill in the correct set of JVMState's 857 // Also fill in the associated edges for each JVMState. 858 859 // If the bytecode needs to be reexecuted we need to put 860 // the arguments back on the stack. 861 const bool should_reexecute = jvms()->should_reexecute(); 862 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 863 864 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 865 // undefined if the bci is different. This is normal for Parse but it 866 // should not happen for LibraryCallKit because only one bci is processed. 867 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 868 "in LibraryCallKit the reexecute bit should not change"); 869 870 // If we are guaranteed to throw, we can prune everything but the 871 // input to the current bytecode. 872 bool can_prune_locals = false; 873 uint stack_slots_not_pruned = 0; 874 int inputs = 0, depth = 0; 875 if (must_throw) { 876 assert(method() == youngest_jvms->method(), "sanity"); 877 if (compute_stack_effects(inputs, depth)) { 878 can_prune_locals = true; 879 stack_slots_not_pruned = inputs; 880 } 881 } 882 883 if (env()->should_retain_local_variables()) { 884 // At any safepoint, this method can get breakpointed, which would 885 // then require an immediate deoptimization. 886 can_prune_locals = false; // do not prune locals 887 stack_slots_not_pruned = 0; 888 } 889 890 // do not scribble on the input jvms 891 JVMState* out_jvms = youngest_jvms->clone_deep(C); 892 call->set_jvms(out_jvms); // Start jvms list for call node 893 894 // For a known set of bytecodes, the interpreter should reexecute them if 895 // deoptimization happens. We set the reexecute state for them here 896 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 897 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 898 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 899 } 900 901 // Presize the call: 902 DEBUG_ONLY(uint non_debug_edges = call->req()); 903 call->add_req_batch(top(), youngest_jvms->debug_depth()); 904 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 905 906 // Set up edges so that the call looks like this: 907 // Call [state:] ctl io mem fptr retadr 908 // [parms:] parm0 ... parmN 909 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 910 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 911 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 912 // Note that caller debug info precedes callee debug info. 913 914 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 915 uint debug_ptr = call->req(); 916 917 // Loop over the map input edges associated with jvms, add them 918 // to the call node, & reset all offsets to match call node array. 919 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 920 uint debug_end = debug_ptr; 921 uint debug_start = debug_ptr - in_jvms->debug_size(); 922 debug_ptr = debug_start; // back up the ptr 923 924 uint p = debug_start; // walks forward in [debug_start, debug_end) 925 uint j, k, l; 926 SafePointNode* in_map = in_jvms->map(); 927 out_jvms->set_map(call); 928 929 if (can_prune_locals) { 930 assert(in_jvms->method() == out_jvms->method(), "sanity"); 931 // If the current throw can reach an exception handler in this JVMS, 932 // then we must keep everything live that can reach that handler. 933 // As a quick and dirty approximation, we look for any handlers at all. 934 if (in_jvms->method()->has_exception_handlers()) { 935 can_prune_locals = false; 936 } 937 } 938 939 // Add the Locals 940 k = in_jvms->locoff(); 941 l = in_jvms->loc_size(); 942 out_jvms->set_locoff(p); 943 if (!can_prune_locals) { 944 for (j = 0; j < l; j++) 945 call->set_req(p++, in_map->in(k+j)); 946 } else { 947 p += l; // already set to top above by add_req_batch 948 } 949 950 // Add the Expression Stack 951 k = in_jvms->stkoff(); 952 l = in_jvms->sp(); 953 out_jvms->set_stkoff(p); 954 if (!can_prune_locals) { 955 for (j = 0; j < l; j++) 956 call->set_req(p++, in_map->in(k+j)); 957 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 958 // Divide stack into {S0,...,S1}, where S0 is set to top. 959 uint s1 = stack_slots_not_pruned; 960 stack_slots_not_pruned = 0; // for next iteration 961 if (s1 > l) s1 = l; 962 uint s0 = l - s1; 963 p += s0; // skip the tops preinstalled by add_req_batch 964 for (j = s0; j < l; j++) 965 call->set_req(p++, in_map->in(k+j)); 966 } else { 967 p += l; // already set to top above by add_req_batch 968 } 969 970 // Add the Monitors 971 k = in_jvms->monoff(); 972 l = in_jvms->mon_size(); 973 out_jvms->set_monoff(p); 974 for (j = 0; j < l; j++) 975 call->set_req(p++, in_map->in(k+j)); 976 977 // Copy any scalar object fields. 978 k = in_jvms->scloff(); 979 l = in_jvms->scl_size(); 980 out_jvms->set_scloff(p); 981 for (j = 0; j < l; j++) 982 call->set_req(p++, in_map->in(k+j)); 983 984 // Finish the new jvms. 985 out_jvms->set_endoff(p); 986 987 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 988 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 989 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 990 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 991 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 992 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 993 994 // Update the two tail pointers in parallel. 995 out_jvms = out_jvms->caller(); 996 in_jvms = in_jvms->caller(); 997 } 998 999 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1000 1001 // Test the correctness of JVMState::debug_xxx accessors: 1002 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1003 assert(call->jvms()->debug_end() == call->req(), ""); 1004 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1005 } 1006 1007 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1008 Bytecodes::Code code = java_bc(); 1009 if (code == Bytecodes::_wide) { 1010 code = method()->java_code_at_bci(bci() + 1); 1011 } 1012 1013 BasicType rtype = T_ILLEGAL; 1014 int rsize = 0; 1015 1016 if (code != Bytecodes::_illegal) { 1017 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1018 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1019 if (rtype < T_CONFLICT) 1020 rsize = type2size[rtype]; 1021 } 1022 1023 switch (code) { 1024 case Bytecodes::_illegal: 1025 return false; 1026 1027 case Bytecodes::_ldc: 1028 case Bytecodes::_ldc_w: 1029 case Bytecodes::_ldc2_w: 1030 inputs = 0; 1031 break; 1032 1033 case Bytecodes::_dup: inputs = 1; break; 1034 case Bytecodes::_dup_x1: inputs = 2; break; 1035 case Bytecodes::_dup_x2: inputs = 3; break; 1036 case Bytecodes::_dup2: inputs = 2; break; 1037 case Bytecodes::_dup2_x1: inputs = 3; break; 1038 case Bytecodes::_dup2_x2: inputs = 4; break; 1039 case Bytecodes::_swap: inputs = 2; break; 1040 case Bytecodes::_arraylength: inputs = 1; break; 1041 1042 case Bytecodes::_getstatic: 1043 case Bytecodes::_putstatic: 1044 case Bytecodes::_getfield: 1045 case Bytecodes::_putfield: 1046 { 1047 bool ignored_will_link; 1048 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1049 int size = field->type()->size(); 1050 bool is_get = (depth >= 0), is_static = (depth & 1); 1051 inputs = (is_static ? 0 : 1); 1052 if (is_get) { 1053 depth = size - inputs; 1054 } else { 1055 inputs += size; // putxxx pops the value from the stack 1056 depth = - inputs; 1057 } 1058 } 1059 break; 1060 1061 case Bytecodes::_invokevirtual: 1062 case Bytecodes::_invokespecial: 1063 case Bytecodes::_invokestatic: 1064 case Bytecodes::_invokedynamic: 1065 case Bytecodes::_invokeinterface: 1066 { 1067 bool ignored_will_link; 1068 ciSignature* declared_signature = NULL; 1069 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1070 assert(declared_signature != NULL, "cannot be null"); 1071 inputs = declared_signature->arg_size_for_bc(code); 1072 int size = declared_signature->return_type()->size(); 1073 depth = size - inputs; 1074 } 1075 break; 1076 1077 case Bytecodes::_multianewarray: 1078 { 1079 ciBytecodeStream iter(method()); 1080 iter.reset_to_bci(bci()); 1081 iter.next(); 1082 inputs = iter.get_dimensions(); 1083 assert(rsize == 1, ""); 1084 depth = rsize - inputs; 1085 } 1086 break; 1087 1088 case Bytecodes::_ireturn: 1089 case Bytecodes::_lreturn: 1090 case Bytecodes::_freturn: 1091 case Bytecodes::_dreturn: 1092 case Bytecodes::_areturn: 1093 assert(rsize == -depth, ""); 1094 inputs = rsize; 1095 break; 1096 1097 case Bytecodes::_jsr: 1098 case Bytecodes::_jsr_w: 1099 inputs = 0; 1100 depth = 1; // S.B. depth=1, not zero 1101 break; 1102 1103 default: 1104 // bytecode produces a typed result 1105 inputs = rsize - depth; 1106 assert(inputs >= 0, ""); 1107 break; 1108 } 1109 1110 #ifdef ASSERT 1111 // spot check 1112 int outputs = depth + inputs; 1113 assert(outputs >= 0, "sanity"); 1114 switch (code) { 1115 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1116 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1117 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1118 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1119 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1120 default: break; 1121 } 1122 #endif //ASSERT 1123 1124 return true; 1125 } 1126 1127 1128 1129 //------------------------------basic_plus_adr--------------------------------- 1130 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1131 // short-circuit a common case 1132 if (offset == intcon(0)) return ptr; 1133 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1134 } 1135 1136 Node* GraphKit::ConvI2L(Node* offset) { 1137 // short-circuit a common case 1138 jint offset_con = find_int_con(offset, Type::OffsetBot); 1139 if (offset_con != Type::OffsetBot) { 1140 return longcon((jlong) offset_con); 1141 } 1142 return _gvn.transform( new ConvI2LNode(offset)); 1143 } 1144 1145 Node* GraphKit::ConvI2UL(Node* offset) { 1146 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1147 if (offset_con != (juint) Type::OffsetBot) { 1148 return longcon((julong) offset_con); 1149 } 1150 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1151 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1152 return _gvn.transform( new AndLNode(conv, mask) ); 1153 } 1154 1155 Node* GraphKit::ConvL2I(Node* offset) { 1156 // short-circuit a common case 1157 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1158 if (offset_con != (jlong)Type::OffsetBot) { 1159 return intcon((int) offset_con); 1160 } 1161 return _gvn.transform( new ConvL2INode(offset)); 1162 } 1163 1164 //-------------------------load_object_klass----------------------------------- 1165 Node* GraphKit::load_object_klass(Node* obj) { 1166 // Special-case a fresh allocation to avoid building nodes: 1167 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1168 if (akls != NULL) return akls; 1169 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1170 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1171 } 1172 1173 //-------------------------load_array_length----------------------------------- 1174 Node* GraphKit::load_array_length(Node* array) { 1175 // Special-case a fresh allocation to avoid building nodes: 1176 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1177 Node *alen; 1178 if (alloc == NULL) { 1179 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1180 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1181 } else { 1182 alen = alloc->Ideal_length(); 1183 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1184 if (ccast != alen) { 1185 alen = _gvn.transform(ccast); 1186 } 1187 } 1188 return alen; 1189 } 1190 1191 //------------------------------do_null_check---------------------------------- 1192 // Helper function to do a NULL pointer check. Returned value is 1193 // the incoming address with NULL casted away. You are allowed to use the 1194 // not-null value only if you are control dependent on the test. 1195 #ifndef PRODUCT 1196 extern int explicit_null_checks_inserted, 1197 explicit_null_checks_elided; 1198 #endif 1199 Node* GraphKit::null_check_common(Node* value, BasicType type, 1200 // optional arguments for variations: 1201 bool assert_null, 1202 Node* *null_control, 1203 bool speculative) { 1204 assert(!assert_null || null_control == NULL, "not both at once"); 1205 if (stopped()) return top(); 1206 NOT_PRODUCT(explicit_null_checks_inserted++); 1207 1208 // Construct NULL check 1209 Node *chk = NULL; 1210 switch(type) { 1211 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1212 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1213 case T_ARRAY : // fall through 1214 type = T_OBJECT; // simplify further tests 1215 case T_OBJECT : { 1216 const Type *t = _gvn.type( value ); 1217 1218 const TypeOopPtr* tp = t->isa_oopptr(); 1219 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1220 // Only for do_null_check, not any of its siblings: 1221 && !assert_null && null_control == NULL) { 1222 // Usually, any field access or invocation on an unloaded oop type 1223 // will simply fail to link, since the statically linked class is 1224 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1225 // the static class is loaded but the sharper oop type is not. 1226 // Rather than checking for this obscure case in lots of places, 1227 // we simply observe that a null check on an unloaded class 1228 // will always be followed by a nonsense operation, so we 1229 // can just issue the uncommon trap here. 1230 // Our access to the unloaded class will only be correct 1231 // after it has been loaded and initialized, which requires 1232 // a trip through the interpreter. 1233 #ifndef PRODUCT 1234 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1235 #endif 1236 uncommon_trap(Deoptimization::Reason_unloaded, 1237 Deoptimization::Action_reinterpret, 1238 tp->klass(), "!loaded"); 1239 return top(); 1240 } 1241 1242 if (assert_null) { 1243 // See if the type is contained in NULL_PTR. 1244 // If so, then the value is already null. 1245 if (t->higher_equal(TypePtr::NULL_PTR)) { 1246 NOT_PRODUCT(explicit_null_checks_elided++); 1247 return value; // Elided null assert quickly! 1248 } 1249 } else { 1250 // See if mixing in the NULL pointer changes type. 1251 // If so, then the NULL pointer was not allowed in the original 1252 // type. In other words, "value" was not-null. 1253 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1254 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1255 NOT_PRODUCT(explicit_null_checks_elided++); 1256 return value; // Elided null check quickly! 1257 } 1258 } 1259 chk = new CmpPNode( value, null() ); 1260 break; 1261 } 1262 1263 default: 1264 fatal("unexpected type: %s", type2name(type)); 1265 } 1266 assert(chk != NULL, "sanity check"); 1267 chk = _gvn.transform(chk); 1268 1269 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1270 BoolNode *btst = new BoolNode( chk, btest); 1271 Node *tst = _gvn.transform( btst ); 1272 1273 //----------- 1274 // if peephole optimizations occurred, a prior test existed. 1275 // If a prior test existed, maybe it dominates as we can avoid this test. 1276 if (tst != btst && type == T_OBJECT) { 1277 // At this point we want to scan up the CFG to see if we can 1278 // find an identical test (and so avoid this test altogether). 1279 Node *cfg = control(); 1280 int depth = 0; 1281 while( depth < 16 ) { // Limit search depth for speed 1282 if( cfg->Opcode() == Op_IfTrue && 1283 cfg->in(0)->in(1) == tst ) { 1284 // Found prior test. Use "cast_not_null" to construct an identical 1285 // CastPP (and hence hash to) as already exists for the prior test. 1286 // Return that casted value. 1287 if (assert_null) { 1288 replace_in_map(value, null()); 1289 return null(); // do not issue the redundant test 1290 } 1291 Node *oldcontrol = control(); 1292 set_control(cfg); 1293 Node *res = cast_not_null(value); 1294 set_control(oldcontrol); 1295 NOT_PRODUCT(explicit_null_checks_elided++); 1296 return res; 1297 } 1298 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1299 if (cfg == NULL) break; // Quit at region nodes 1300 depth++; 1301 } 1302 } 1303 1304 //----------- 1305 // Branch to failure if null 1306 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1307 Deoptimization::DeoptReason reason; 1308 if (assert_null) { 1309 reason = Deoptimization::reason_null_assert(speculative); 1310 } else if (type == T_OBJECT) { 1311 reason = Deoptimization::reason_null_check(speculative); 1312 } else { 1313 reason = Deoptimization::Reason_div0_check; 1314 } 1315 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1316 // ciMethodData::has_trap_at will return a conservative -1 if any 1317 // must-be-null assertion has failed. This could cause performance 1318 // problems for a method after its first do_null_assert failure. 1319 // Consider using 'Reason_class_check' instead? 1320 1321 // To cause an implicit null check, we set the not-null probability 1322 // to the maximum (PROB_MAX). For an explicit check the probability 1323 // is set to a smaller value. 1324 if (null_control != NULL || too_many_traps(reason)) { 1325 // probability is less likely 1326 ok_prob = PROB_LIKELY_MAG(3); 1327 } else if (!assert_null && 1328 (ImplicitNullCheckThreshold > 0) && 1329 method() != NULL && 1330 (method()->method_data()->trap_count(reason) 1331 >= (uint)ImplicitNullCheckThreshold)) { 1332 ok_prob = PROB_LIKELY_MAG(3); 1333 } 1334 1335 if (null_control != NULL) { 1336 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1337 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1338 set_control( _gvn.transform( new IfTrueNode(iff))); 1339 #ifndef PRODUCT 1340 if (null_true == top()) { 1341 explicit_null_checks_elided++; 1342 } 1343 #endif 1344 (*null_control) = null_true; 1345 } else { 1346 BuildCutout unless(this, tst, ok_prob); 1347 // Check for optimizer eliding test at parse time 1348 if (stopped()) { 1349 // Failure not possible; do not bother making uncommon trap. 1350 NOT_PRODUCT(explicit_null_checks_elided++); 1351 } else if (assert_null) { 1352 uncommon_trap(reason, 1353 Deoptimization::Action_make_not_entrant, 1354 NULL, "assert_null"); 1355 } else { 1356 replace_in_map(value, zerocon(type)); 1357 builtin_throw(reason); 1358 } 1359 } 1360 1361 // Must throw exception, fall-thru not possible? 1362 if (stopped()) { 1363 return top(); // No result 1364 } 1365 1366 if (assert_null) { 1367 // Cast obj to null on this path. 1368 replace_in_map(value, zerocon(type)); 1369 return zerocon(type); 1370 } 1371 1372 // Cast obj to not-null on this path, if there is no null_control. 1373 // (If there is a null_control, a non-null value may come back to haunt us.) 1374 if (type == T_OBJECT) { 1375 Node* cast = cast_not_null(value, false); 1376 if (null_control == NULL || (*null_control) == top()) 1377 replace_in_map(value, cast); 1378 value = cast; 1379 } 1380 1381 return value; 1382 } 1383 1384 1385 //------------------------------cast_not_null---------------------------------- 1386 // Cast obj to not-null on this path 1387 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1388 const Type *t = _gvn.type(obj); 1389 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1390 // Object is already not-null? 1391 if( t == t_not_null ) return obj; 1392 1393 Node *cast = new CastPPNode(obj,t_not_null); 1394 cast->init_req(0, control()); 1395 cast = _gvn.transform( cast ); 1396 1397 // Scan for instances of 'obj' in the current JVM mapping. 1398 // These instances are known to be not-null after the test. 1399 if (do_replace_in_map) 1400 replace_in_map(obj, cast); 1401 1402 return cast; // Return casted value 1403 } 1404 1405 // Sometimes in intrinsics, we implicitly know an object is not null 1406 // (there's no actual null check) so we can cast it to not null. In 1407 // the course of optimizations, the input to the cast can become null. 1408 // In that case that data path will die and we need the control path 1409 // to become dead as well to keep the graph consistent. So we have to 1410 // add a check for null for which one branch can't be taken. It uses 1411 // an Opaque4 node that will cause the check to be removed after loop 1412 // opts so the test goes away and the compiled code doesn't execute a 1413 // useless check. 1414 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1415 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1416 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1417 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1418 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1419 _gvn.set_type(iff, iff->Value(&_gvn)); 1420 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1421 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1422 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1423 C->root()->add_req(halt); 1424 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1425 set_control(if_t); 1426 return cast_not_null(value, do_replace_in_map); 1427 } 1428 1429 1430 //--------------------------replace_in_map------------------------------------- 1431 void GraphKit::replace_in_map(Node* old, Node* neww) { 1432 if (old == neww) { 1433 return; 1434 } 1435 1436 map()->replace_edge(old, neww); 1437 1438 // Note: This operation potentially replaces any edge 1439 // on the map. This includes locals, stack, and monitors 1440 // of the current (innermost) JVM state. 1441 1442 // don't let inconsistent types from profiling escape this 1443 // method 1444 1445 const Type* told = _gvn.type(old); 1446 const Type* tnew = _gvn.type(neww); 1447 1448 if (!tnew->higher_equal(told)) { 1449 return; 1450 } 1451 1452 map()->record_replaced_node(old, neww); 1453 } 1454 1455 1456 //============================================================================= 1457 //--------------------------------memory--------------------------------------- 1458 Node* GraphKit::memory(uint alias_idx) { 1459 MergeMemNode* mem = merged_memory(); 1460 Node* p = mem->memory_at(alias_idx); 1461 _gvn.set_type(p, Type::MEMORY); // must be mapped 1462 return p; 1463 } 1464 1465 //-----------------------------reset_memory------------------------------------ 1466 Node* GraphKit::reset_memory() { 1467 Node* mem = map()->memory(); 1468 // do not use this node for any more parsing! 1469 debug_only( map()->set_memory((Node*)NULL) ); 1470 return _gvn.transform( mem ); 1471 } 1472 1473 //------------------------------set_all_memory--------------------------------- 1474 void GraphKit::set_all_memory(Node* newmem) { 1475 Node* mergemem = MergeMemNode::make(newmem); 1476 gvn().set_type_bottom(mergemem); 1477 map()->set_memory(mergemem); 1478 } 1479 1480 //------------------------------set_all_memory_call---------------------------- 1481 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1482 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1483 set_all_memory(newmem); 1484 } 1485 1486 //============================================================================= 1487 // 1488 // parser factory methods for MemNodes 1489 // 1490 // These are layered on top of the factory methods in LoadNode and StoreNode, 1491 // and integrate with the parser's memory state and _gvn engine. 1492 // 1493 1494 // factory methods in "int adr_idx" 1495 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1496 int adr_idx, 1497 MemNode::MemOrd mo, 1498 LoadNode::ControlDependency control_dependency, 1499 bool require_atomic_access, 1500 bool unaligned, 1501 bool mismatched) { 1502 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1503 const TypePtr* adr_type = NULL; // debug-mode-only argument 1504 debug_only(adr_type = C->get_adr_type(adr_idx)); 1505 Node* mem = memory(adr_idx); 1506 Node* ld; 1507 if (require_atomic_access && bt == T_LONG) { 1508 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1509 } else if (require_atomic_access && bt == T_DOUBLE) { 1510 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1511 } else { 1512 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1513 } 1514 ld = _gvn.transform(ld); 1515 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1516 // Improve graph before escape analysis and boxing elimination. 1517 record_for_igvn(ld); 1518 } 1519 return ld; 1520 } 1521 1522 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1523 int adr_idx, 1524 MemNode::MemOrd mo, 1525 bool require_atomic_access, 1526 bool unaligned, 1527 bool mismatched) { 1528 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1529 const TypePtr* adr_type = NULL; 1530 debug_only(adr_type = C->get_adr_type(adr_idx)); 1531 Node *mem = memory(adr_idx); 1532 Node* st; 1533 if (require_atomic_access && bt == T_LONG) { 1534 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1535 } else if (require_atomic_access && bt == T_DOUBLE) { 1536 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1537 } else { 1538 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1539 } 1540 if (unaligned) { 1541 st->as_Store()->set_unaligned_access(); 1542 } 1543 if (mismatched) { 1544 st->as_Store()->set_mismatched_access(); 1545 } 1546 st = _gvn.transform(st); 1547 set_memory(st, adr_idx); 1548 // Back-to-back stores can only remove intermediate store with DU info 1549 // so push on worklist for optimizer. 1550 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1551 record_for_igvn(st); 1552 1553 return st; 1554 } 1555 1556 1557 void GraphKit::pre_barrier(bool do_load, 1558 Node* ctl, 1559 Node* obj, 1560 Node* adr, 1561 uint adr_idx, 1562 Node* val, 1563 const TypeOopPtr* val_type, 1564 Node* pre_val, 1565 BasicType bt) { 1566 1567 BarrierSet* bs = BarrierSet::barrier_set(); 1568 set_control(ctl); 1569 switch (bs->kind()) { 1570 case BarrierSet::G1BarrierSet: 1571 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1572 break; 1573 1574 case BarrierSet::CardTableBarrierSet: 1575 break; 1576 1577 default : 1578 ShouldNotReachHere(); 1579 1580 } 1581 } 1582 1583 bool GraphKit::can_move_pre_barrier() const { 1584 BarrierSet* bs = BarrierSet::barrier_set(); 1585 switch (bs->kind()) { 1586 case BarrierSet::G1BarrierSet: 1587 return true; // Can move it if no safepoint 1588 1589 case BarrierSet::CardTableBarrierSet: 1590 return true; // There is no pre-barrier 1591 1592 default : 1593 ShouldNotReachHere(); 1594 } 1595 return false; 1596 } 1597 1598 void GraphKit::post_barrier(Node* ctl, 1599 Node* store, 1600 Node* obj, 1601 Node* adr, 1602 uint adr_idx, 1603 Node* val, 1604 BasicType bt, 1605 bool use_precise) { 1606 BarrierSet* bs = BarrierSet::barrier_set(); 1607 set_control(ctl); 1608 switch (bs->kind()) { 1609 case BarrierSet::G1BarrierSet: 1610 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1611 break; 1612 1613 case BarrierSet::CardTableBarrierSet: 1614 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1615 break; 1616 1617 default : 1618 ShouldNotReachHere(); 1619 1620 } 1621 } 1622 1623 Node* GraphKit::store_oop(Node* ctl, 1624 Node* obj, 1625 Node* adr, 1626 const TypePtr* adr_type, 1627 Node* val, 1628 const TypeOopPtr* val_type, 1629 BasicType bt, 1630 bool use_precise, 1631 MemNode::MemOrd mo, 1632 bool mismatched) { 1633 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1634 // could be delayed during Parse (for example, in adjust_map_after_if()). 1635 // Execute transformation here to avoid barrier generation in such case. 1636 if (_gvn.type(val) == TypePtr::NULL_PTR) 1637 val = _gvn.makecon(TypePtr::NULL_PTR); 1638 1639 set_control(ctl); 1640 if (stopped()) return top(); // Dead path ? 1641 1642 assert(bt == T_OBJECT, "sanity"); 1643 assert(val != NULL, "not dead path"); 1644 uint adr_idx = C->get_alias_index(adr_type); 1645 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1646 1647 pre_barrier(true /* do_load */, 1648 control(), obj, adr, adr_idx, val, val_type, 1649 NULL /* pre_val */, 1650 bt); 1651 1652 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1653 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1654 return store; 1655 } 1656 1657 // Could be an array or object we don't know at compile time (unsafe ref.) 1658 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1659 Node* obj, // containing obj 1660 Node* adr, // actual adress to store val at 1661 const TypePtr* adr_type, 1662 Node* val, 1663 BasicType bt, 1664 MemNode::MemOrd mo, 1665 bool mismatched) { 1666 Compile::AliasType* at = C->alias_type(adr_type); 1667 const TypeOopPtr* val_type = NULL; 1668 if (adr_type->isa_instptr()) { 1669 if (at->field() != NULL) { 1670 // known field. This code is a copy of the do_put_xxx logic. 1671 ciField* field = at->field(); 1672 if (!field->type()->is_loaded()) { 1673 val_type = TypeInstPtr::BOTTOM; 1674 } else { 1675 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1676 } 1677 } 1678 } else if (adr_type->isa_aryptr()) { 1679 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1680 } 1681 if (val_type == NULL) { 1682 val_type = TypeInstPtr::BOTTOM; 1683 } 1684 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1685 } 1686 1687 1688 //-------------------------array_element_address------------------------- 1689 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1690 const TypeInt* sizetype, Node* ctrl) { 1691 uint shift = exact_log2(type2aelembytes(elembt)); 1692 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1693 1694 // short-circuit a common case (saves lots of confusing waste motion) 1695 jint idx_con = find_int_con(idx, -1); 1696 if (idx_con >= 0) { 1697 intptr_t offset = header + ((intptr_t)idx_con << shift); 1698 return basic_plus_adr(ary, offset); 1699 } 1700 1701 // must be correct type for alignment purposes 1702 Node* base = basic_plus_adr(ary, header); 1703 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1704 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1705 return basic_plus_adr(ary, base, scale); 1706 } 1707 1708 //-------------------------load_array_element------------------------- 1709 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1710 const Type* elemtype = arytype->elem(); 1711 BasicType elembt = elemtype->array_element_basic_type(); 1712 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1713 if (elembt == T_NARROWOOP) { 1714 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1715 } 1716 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1717 return ld; 1718 } 1719 1720 //-------------------------set_arguments_for_java_call------------------------- 1721 // Arguments (pre-popped from the stack) are taken from the JVMS. 1722 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1723 // Add the call arguments: 1724 uint nargs = call->method()->arg_size(); 1725 for (uint i = 0; i < nargs; i++) { 1726 Node* arg = argument(i); 1727 call->init_req(i + TypeFunc::Parms, arg); 1728 } 1729 } 1730 1731 //---------------------------set_edges_for_java_call--------------------------- 1732 // Connect a newly created call into the current JVMS. 1733 // A return value node (if any) is returned from set_edges_for_java_call. 1734 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1735 1736 // Add the predefined inputs: 1737 call->init_req( TypeFunc::Control, control() ); 1738 call->init_req( TypeFunc::I_O , i_o() ); 1739 call->init_req( TypeFunc::Memory , reset_memory() ); 1740 call->init_req( TypeFunc::FramePtr, frameptr() ); 1741 call->init_req( TypeFunc::ReturnAdr, top() ); 1742 1743 add_safepoint_edges(call, must_throw); 1744 1745 Node* xcall = _gvn.transform(call); 1746 1747 if (xcall == top()) { 1748 set_control(top()); 1749 return; 1750 } 1751 assert(xcall == call, "call identity is stable"); 1752 1753 // Re-use the current map to produce the result. 1754 1755 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1756 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1757 set_all_memory_call(xcall, separate_io_proj); 1758 1759 //return xcall; // no need, caller already has it 1760 } 1761 1762 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1763 if (stopped()) return top(); // maybe the call folded up? 1764 1765 // Capture the return value, if any. 1766 Node* ret; 1767 if (call->method() == NULL || 1768 call->method()->return_type()->basic_type() == T_VOID) 1769 ret = top(); 1770 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1771 1772 // Note: Since any out-of-line call can produce an exception, 1773 // we always insert an I_O projection from the call into the result. 1774 1775 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1776 1777 if (separate_io_proj) { 1778 // The caller requested separate projections be used by the fall 1779 // through and exceptional paths, so replace the projections for 1780 // the fall through path. 1781 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1782 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1783 } 1784 return ret; 1785 } 1786 1787 //--------------------set_predefined_input_for_runtime_call-------------------- 1788 // Reading and setting the memory state is way conservative here. 1789 // The real problem is that I am not doing real Type analysis on memory, 1790 // so I cannot distinguish card mark stores from other stores. Across a GC 1791 // point the Store Barrier and the card mark memory has to agree. I cannot 1792 // have a card mark store and its barrier split across the GC point from 1793 // either above or below. Here I get that to happen by reading ALL of memory. 1794 // A better answer would be to separate out card marks from other memory. 1795 // For now, return the input memory state, so that it can be reused 1796 // after the call, if this call has restricted memory effects. 1797 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1798 // Set fixed predefined input arguments 1799 Node* memory = reset_memory(); 1800 call->init_req( TypeFunc::Control, control() ); 1801 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1802 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1803 call->init_req( TypeFunc::FramePtr, frameptr() ); 1804 call->init_req( TypeFunc::ReturnAdr, top() ); 1805 return memory; 1806 } 1807 1808 //-------------------set_predefined_output_for_runtime_call-------------------- 1809 // Set control and memory (not i_o) from the call. 1810 // If keep_mem is not NULL, use it for the output state, 1811 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1812 // If hook_mem is NULL, this call produces no memory effects at all. 1813 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1814 // then only that memory slice is taken from the call. 1815 // In the last case, we must put an appropriate memory barrier before 1816 // the call, so as to create the correct anti-dependencies on loads 1817 // preceding the call. 1818 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1819 Node* keep_mem, 1820 const TypePtr* hook_mem) { 1821 // no i/o 1822 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1823 if (keep_mem) { 1824 // First clone the existing memory state 1825 set_all_memory(keep_mem); 1826 if (hook_mem != NULL) { 1827 // Make memory for the call 1828 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1829 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1830 // We also use hook_mem to extract specific effects from arraycopy stubs. 1831 set_memory(mem, hook_mem); 1832 } 1833 // ...else the call has NO memory effects. 1834 1835 // Make sure the call advertises its memory effects precisely. 1836 // This lets us build accurate anti-dependences in gcm.cpp. 1837 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1838 "call node must be constructed correctly"); 1839 } else { 1840 assert(hook_mem == NULL, ""); 1841 // This is not a "slow path" call; all memory comes from the call. 1842 set_all_memory_call(call); 1843 } 1844 } 1845 1846 1847 // Replace the call with the current state of the kit. 1848 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1849 JVMState* ejvms = NULL; 1850 if (has_exceptions()) { 1851 ejvms = transfer_exceptions_into_jvms(); 1852 } 1853 1854 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1855 ReplacedNodes replaced_nodes_exception; 1856 Node* ex_ctl = top(); 1857 1858 SafePointNode* final_state = stop(); 1859 1860 // Find all the needed outputs of this call 1861 CallProjections callprojs; 1862 call->extract_projections(&callprojs, true); 1863 1864 Node* init_mem = call->in(TypeFunc::Memory); 1865 Node* final_mem = final_state->in(TypeFunc::Memory); 1866 Node* final_ctl = final_state->in(TypeFunc::Control); 1867 Node* final_io = final_state->in(TypeFunc::I_O); 1868 1869 // Replace all the old call edges with the edges from the inlining result 1870 if (callprojs.fallthrough_catchproj != NULL) { 1871 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1872 } 1873 if (callprojs.fallthrough_memproj != NULL) { 1874 if (final_mem->is_MergeMem()) { 1875 // Parser's exits MergeMem was not transformed but may be optimized 1876 final_mem = _gvn.transform(final_mem); 1877 } 1878 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1879 } 1880 if (callprojs.fallthrough_ioproj != NULL) { 1881 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1882 } 1883 1884 // Replace the result with the new result if it exists and is used 1885 if (callprojs.resproj != NULL && result != NULL) { 1886 C->gvn_replace_by(callprojs.resproj, result); 1887 } 1888 1889 if (ejvms == NULL) { 1890 // No exception edges to simply kill off those paths 1891 if (callprojs.catchall_catchproj != NULL) { 1892 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1893 } 1894 if (callprojs.catchall_memproj != NULL) { 1895 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1896 } 1897 if (callprojs.catchall_ioproj != NULL) { 1898 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1899 } 1900 // Replace the old exception object with top 1901 if (callprojs.exobj != NULL) { 1902 C->gvn_replace_by(callprojs.exobj, C->top()); 1903 } 1904 } else { 1905 GraphKit ekit(ejvms); 1906 1907 // Load my combined exception state into the kit, with all phis transformed: 1908 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1909 replaced_nodes_exception = ex_map->replaced_nodes(); 1910 1911 Node* ex_oop = ekit.use_exception_state(ex_map); 1912 1913 if (callprojs.catchall_catchproj != NULL) { 1914 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1915 ex_ctl = ekit.control(); 1916 } 1917 if (callprojs.catchall_memproj != NULL) { 1918 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1919 } 1920 if (callprojs.catchall_ioproj != NULL) { 1921 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1922 } 1923 1924 // Replace the old exception object with the newly created one 1925 if (callprojs.exobj != NULL) { 1926 C->gvn_replace_by(callprojs.exobj, ex_oop); 1927 } 1928 } 1929 1930 // Disconnect the call from the graph 1931 call->disconnect_inputs(NULL, C); 1932 C->gvn_replace_by(call, C->top()); 1933 1934 // Clean up any MergeMems that feed other MergeMems since the 1935 // optimizer doesn't like that. 1936 if (final_mem->is_MergeMem()) { 1937 Node_List wl; 1938 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1939 Node* m = i.get(); 1940 if (m->is_MergeMem() && !wl.contains(m)) { 1941 wl.push(m); 1942 } 1943 } 1944 while (wl.size() > 0) { 1945 _gvn.transform(wl.pop()); 1946 } 1947 } 1948 1949 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1950 replaced_nodes.apply(C, final_ctl); 1951 } 1952 if (!ex_ctl->is_top() && do_replaced_nodes) { 1953 replaced_nodes_exception.apply(C, ex_ctl); 1954 } 1955 } 1956 1957 1958 //------------------------------increment_counter------------------------------ 1959 // for statistics: increment a VM counter by 1 1960 1961 void GraphKit::increment_counter(address counter_addr) { 1962 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1963 increment_counter(adr1); 1964 } 1965 1966 void GraphKit::increment_counter(Node* counter_addr) { 1967 int adr_type = Compile::AliasIdxRaw; 1968 Node* ctrl = control(); 1969 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1970 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1971 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1972 } 1973 1974 1975 //------------------------------uncommon_trap---------------------------------- 1976 // Bail out to the interpreter in mid-method. Implemented by calling the 1977 // uncommon_trap blob. This helper function inserts a runtime call with the 1978 // right debug info. 1979 void GraphKit::uncommon_trap(int trap_request, 1980 ciKlass* klass, const char* comment, 1981 bool must_throw, 1982 bool keep_exact_action) { 1983 if (failing()) stop(); 1984 if (stopped()) return; // trap reachable? 1985 1986 // Note: If ProfileTraps is true, and if a deopt. actually 1987 // occurs here, the runtime will make sure an MDO exists. There is 1988 // no need to call method()->ensure_method_data() at this point. 1989 1990 // Set the stack pointer to the right value for reexecution: 1991 set_sp(reexecute_sp()); 1992 1993 #ifdef ASSERT 1994 if (!must_throw) { 1995 // Make sure the stack has at least enough depth to execute 1996 // the current bytecode. 1997 int inputs, ignored_depth; 1998 if (compute_stack_effects(inputs, ignored_depth)) { 1999 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2000 Bytecodes::name(java_bc()), sp(), inputs); 2001 } 2002 } 2003 #endif 2004 2005 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2006 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2007 2008 switch (action) { 2009 case Deoptimization::Action_maybe_recompile: 2010 case Deoptimization::Action_reinterpret: 2011 // Temporary fix for 6529811 to allow virtual calls to be sure they 2012 // get the chance to go from mono->bi->mega 2013 if (!keep_exact_action && 2014 Deoptimization::trap_request_index(trap_request) < 0 && 2015 too_many_recompiles(reason)) { 2016 // This BCI is causing too many recompilations. 2017 if (C->log() != NULL) { 2018 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2019 Deoptimization::trap_reason_name(reason), 2020 Deoptimization::trap_action_name(action)); 2021 } 2022 action = Deoptimization::Action_none; 2023 trap_request = Deoptimization::make_trap_request(reason, action); 2024 } else { 2025 C->set_trap_can_recompile(true); 2026 } 2027 break; 2028 case Deoptimization::Action_make_not_entrant: 2029 C->set_trap_can_recompile(true); 2030 break; 2031 case Deoptimization::Action_none: 2032 case Deoptimization::Action_make_not_compilable: 2033 break; 2034 default: 2035 #ifdef ASSERT 2036 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2037 #endif 2038 break; 2039 } 2040 2041 if (TraceOptoParse) { 2042 char buf[100]; 2043 tty->print_cr("Uncommon trap %s at bci:%d", 2044 Deoptimization::format_trap_request(buf, sizeof(buf), 2045 trap_request), bci()); 2046 } 2047 2048 CompileLog* log = C->log(); 2049 if (log != NULL) { 2050 int kid = (klass == NULL)? -1: log->identify(klass); 2051 log->begin_elem("uncommon_trap bci='%d'", bci()); 2052 char buf[100]; 2053 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2054 trap_request)); 2055 if (kid >= 0) log->print(" klass='%d'", kid); 2056 if (comment != NULL) log->print(" comment='%s'", comment); 2057 log->end_elem(); 2058 } 2059 2060 // Make sure any guarding test views this path as very unlikely 2061 Node *i0 = control()->in(0); 2062 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2063 IfNode *iff = i0->as_If(); 2064 float f = iff->_prob; // Get prob 2065 if (control()->Opcode() == Op_IfTrue) { 2066 if (f > PROB_UNLIKELY_MAG(4)) 2067 iff->_prob = PROB_MIN; 2068 } else { 2069 if (f < PROB_LIKELY_MAG(4)) 2070 iff->_prob = PROB_MAX; 2071 } 2072 } 2073 2074 // Clear out dead values from the debug info. 2075 kill_dead_locals(); 2076 2077 // Now insert the uncommon trap subroutine call 2078 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2079 const TypePtr* no_memory_effects = NULL; 2080 // Pass the index of the class to be loaded 2081 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2082 (must_throw ? RC_MUST_THROW : 0), 2083 OptoRuntime::uncommon_trap_Type(), 2084 call_addr, "uncommon_trap", no_memory_effects, 2085 intcon(trap_request)); 2086 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2087 "must extract request correctly from the graph"); 2088 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2089 2090 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2091 // The debug info is the only real input to this call. 2092 2093 // Halt-and-catch fire here. The above call should never return! 2094 HaltNode* halt = new HaltNode(control(), frameptr()); 2095 _gvn.set_type_bottom(halt); 2096 root()->add_req(halt); 2097 2098 stop_and_kill_map(); 2099 } 2100 2101 2102 //--------------------------just_allocated_object------------------------------ 2103 // Report the object that was just allocated. 2104 // It must be the case that there are no intervening safepoints. 2105 // We use this to determine if an object is so "fresh" that 2106 // it does not require card marks. 2107 Node* GraphKit::just_allocated_object(Node* current_control) { 2108 if (C->recent_alloc_ctl() == current_control) 2109 return C->recent_alloc_obj(); 2110 return NULL; 2111 } 2112 2113 2114 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2115 // (Note: TypeFunc::make has a cache that makes this fast.) 2116 const TypeFunc* tf = TypeFunc::make(dest_method); 2117 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2118 for (int j = 0; j < nargs; j++) { 2119 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2120 if( targ->basic_type() == T_DOUBLE ) { 2121 // If any parameters are doubles, they must be rounded before 2122 // the call, dstore_rounding does gvn.transform 2123 Node *arg = argument(j); 2124 arg = dstore_rounding(arg); 2125 set_argument(j, arg); 2126 } 2127 } 2128 } 2129 2130 /** 2131 * Record profiling data exact_kls for Node n with the type system so 2132 * that it can propagate it (speculation) 2133 * 2134 * @param n node that the type applies to 2135 * @param exact_kls type from profiling 2136 * @param maybe_null did profiling see null? 2137 * 2138 * @return node with improved type 2139 */ 2140 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2141 const Type* current_type = _gvn.type(n); 2142 assert(UseTypeSpeculation, "type speculation must be on"); 2143 2144 const TypePtr* speculative = current_type->speculative(); 2145 2146 // Should the klass from the profile be recorded in the speculative type? 2147 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2148 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2149 const TypeOopPtr* xtype = tklass->as_instance_type(); 2150 assert(xtype->klass_is_exact(), "Should be exact"); 2151 // Any reason to believe n is not null (from this profiling or a previous one)? 2152 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2153 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2154 // record the new speculative type's depth 2155 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2156 speculative = speculative->with_inline_depth(jvms()->depth()); 2157 } else if (current_type->would_improve_ptr(ptr_kind)) { 2158 // Profiling report that null was never seen so we can change the 2159 // speculative type to non null ptr. 2160 if (ptr_kind == ProfileAlwaysNull) { 2161 speculative = TypePtr::NULL_PTR; 2162 } else { 2163 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2164 const TypePtr* ptr = TypePtr::NOTNULL; 2165 if (speculative != NULL) { 2166 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2167 } else { 2168 speculative = ptr; 2169 } 2170 } 2171 } 2172 2173 if (speculative != current_type->speculative()) { 2174 // Build a type with a speculative type (what we think we know 2175 // about the type but will need a guard when we use it) 2176 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2177 // We're changing the type, we need a new CheckCast node to carry 2178 // the new type. The new type depends on the control: what 2179 // profiling tells us is only valid from here as far as we can 2180 // tell. 2181 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2182 cast = _gvn.transform(cast); 2183 replace_in_map(n, cast); 2184 n = cast; 2185 } 2186 2187 return n; 2188 } 2189 2190 /** 2191 * Record profiling data from receiver profiling at an invoke with the 2192 * type system so that it can propagate it (speculation) 2193 * 2194 * @param n receiver node 2195 * 2196 * @return node with improved type 2197 */ 2198 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2199 if (!UseTypeSpeculation) { 2200 return n; 2201 } 2202 ciKlass* exact_kls = profile_has_unique_klass(); 2203 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2204 if ((java_bc() == Bytecodes::_checkcast || 2205 java_bc() == Bytecodes::_instanceof || 2206 java_bc() == Bytecodes::_aastore) && 2207 method()->method_data()->is_mature()) { 2208 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2209 if (data != NULL) { 2210 if (!data->as_BitData()->null_seen()) { 2211 ptr_kind = ProfileNeverNull; 2212 } else { 2213 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2214 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2215 uint i = 0; 2216 for (; i < call->row_limit(); i++) { 2217 ciKlass* receiver = call->receiver(i); 2218 if (receiver != NULL) { 2219 break; 2220 } 2221 } 2222 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2223 } 2224 } 2225 } 2226 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2227 } 2228 2229 /** 2230 * Record profiling data from argument profiling at an invoke with the 2231 * type system so that it can propagate it (speculation) 2232 * 2233 * @param dest_method target method for the call 2234 * @param bc what invoke bytecode is this? 2235 */ 2236 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2237 if (!UseTypeSpeculation) { 2238 return; 2239 } 2240 const TypeFunc* tf = TypeFunc::make(dest_method); 2241 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2242 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2243 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2244 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2245 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2246 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2247 ciKlass* better_type = NULL; 2248 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2249 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2250 } 2251 i++; 2252 } 2253 } 2254 } 2255 2256 /** 2257 * Record profiling data from parameter profiling at an invoke with 2258 * the type system so that it can propagate it (speculation) 2259 */ 2260 void GraphKit::record_profiled_parameters_for_speculation() { 2261 if (!UseTypeSpeculation) { 2262 return; 2263 } 2264 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2265 if (_gvn.type(local(i))->isa_oopptr()) { 2266 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2267 ciKlass* better_type = NULL; 2268 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2269 record_profile_for_speculation(local(i), better_type, ptr_kind); 2270 } 2271 j++; 2272 } 2273 } 2274 } 2275 2276 /** 2277 * Record profiling data from return value profiling at an invoke with 2278 * the type system so that it can propagate it (speculation) 2279 */ 2280 void GraphKit::record_profiled_return_for_speculation() { 2281 if (!UseTypeSpeculation) { 2282 return; 2283 } 2284 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2285 ciKlass* better_type = NULL; 2286 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2287 // If profiling reports a single type for the return value, 2288 // feed it to the type system so it can propagate it as a 2289 // speculative type 2290 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2291 } 2292 } 2293 2294 void GraphKit::round_double_result(ciMethod* dest_method) { 2295 // A non-strict method may return a double value which has an extended 2296 // exponent, but this must not be visible in a caller which is 'strict' 2297 // If a strict caller invokes a non-strict callee, round a double result 2298 2299 BasicType result_type = dest_method->return_type()->basic_type(); 2300 assert( method() != NULL, "must have caller context"); 2301 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2302 // Destination method's return value is on top of stack 2303 // dstore_rounding() does gvn.transform 2304 Node *result = pop_pair(); 2305 result = dstore_rounding(result); 2306 push_pair(result); 2307 } 2308 } 2309 2310 // rounding for strict float precision conformance 2311 Node* GraphKit::precision_rounding(Node* n) { 2312 return UseStrictFP && _method->flags().is_strict() 2313 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2314 ? _gvn.transform( new RoundFloatNode(0, n) ) 2315 : n; 2316 } 2317 2318 // rounding for strict double precision conformance 2319 Node* GraphKit::dprecision_rounding(Node *n) { 2320 return UseStrictFP && _method->flags().is_strict() 2321 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2322 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2323 : n; 2324 } 2325 2326 // rounding for non-strict double stores 2327 Node* GraphKit::dstore_rounding(Node* n) { 2328 return Matcher::strict_fp_requires_explicit_rounding 2329 && UseSSE <= 1 2330 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2331 : n; 2332 } 2333 2334 //============================================================================= 2335 // Generate a fast path/slow path idiom. Graph looks like: 2336 // [foo] indicates that 'foo' is a parameter 2337 // 2338 // [in] NULL 2339 // \ / 2340 // CmpP 2341 // Bool ne 2342 // If 2343 // / \ 2344 // True False-<2> 2345 // / | 2346 // / cast_not_null 2347 // Load | | ^ 2348 // [fast_test] | | 2349 // gvn to opt_test | | 2350 // / \ | <1> 2351 // True False | 2352 // | \\ | 2353 // [slow_call] \[fast_result] 2354 // Ctl Val \ \ 2355 // | \ \ 2356 // Catch <1> \ \ 2357 // / \ ^ \ \ 2358 // Ex No_Ex | \ \ 2359 // | \ \ | \ <2> \ 2360 // ... \ [slow_res] | | \ [null_result] 2361 // \ \--+--+--- | | 2362 // \ | / \ | / 2363 // --------Region Phi 2364 // 2365 //============================================================================= 2366 // Code is structured as a series of driver functions all called 'do_XXX' that 2367 // call a set of helper functions. Helper functions first, then drivers. 2368 2369 //------------------------------null_check_oop--------------------------------- 2370 // Null check oop. Set null-path control into Region in slot 3. 2371 // Make a cast-not-nullness use the other not-null control. Return cast. 2372 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2373 bool never_see_null, 2374 bool safe_for_replace, 2375 bool speculative) { 2376 // Initial NULL check taken path 2377 (*null_control) = top(); 2378 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2379 2380 // Generate uncommon_trap: 2381 if (never_see_null && (*null_control) != top()) { 2382 // If we see an unexpected null at a check-cast we record it and force a 2383 // recompile; the offending check-cast will be compiled to handle NULLs. 2384 // If we see more than one offending BCI, then all checkcasts in the 2385 // method will be compiled to handle NULLs. 2386 PreserveJVMState pjvms(this); 2387 set_control(*null_control); 2388 replace_in_map(value, null()); 2389 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2390 uncommon_trap(reason, 2391 Deoptimization::Action_make_not_entrant); 2392 (*null_control) = top(); // NULL path is dead 2393 } 2394 if ((*null_control) == top() && safe_for_replace) { 2395 replace_in_map(value, cast); 2396 } 2397 2398 // Cast away null-ness on the result 2399 return cast; 2400 } 2401 2402 //------------------------------opt_iff---------------------------------------- 2403 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2404 // Return slow-path control. 2405 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2406 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2407 2408 // Fast path taken; set region slot 2 2409 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2410 region->init_req(2,fast_taken); // Capture fast-control 2411 2412 // Fast path not-taken, i.e. slow path 2413 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2414 return slow_taken; 2415 } 2416 2417 //-----------------------------make_runtime_call------------------------------- 2418 Node* GraphKit::make_runtime_call(int flags, 2419 const TypeFunc* call_type, address call_addr, 2420 const char* call_name, 2421 const TypePtr* adr_type, 2422 // The following parms are all optional. 2423 // The first NULL ends the list. 2424 Node* parm0, Node* parm1, 2425 Node* parm2, Node* parm3, 2426 Node* parm4, Node* parm5, 2427 Node* parm6, Node* parm7) { 2428 // Slow-path call 2429 bool is_leaf = !(flags & RC_NO_LEAF); 2430 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2431 if (call_name == NULL) { 2432 assert(!is_leaf, "must supply name for leaf"); 2433 call_name = OptoRuntime::stub_name(call_addr); 2434 } 2435 CallNode* call; 2436 if (!is_leaf) { 2437 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2438 bci(), adr_type); 2439 } else if (flags & RC_NO_FP) { 2440 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2441 } else { 2442 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2443 } 2444 2445 // The following is similar to set_edges_for_java_call, 2446 // except that the memory effects of the call are restricted to AliasIdxRaw. 2447 2448 // Slow path call has no side-effects, uses few values 2449 bool wide_in = !(flags & RC_NARROW_MEM); 2450 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2451 2452 Node* prev_mem = NULL; 2453 if (wide_in) { 2454 prev_mem = set_predefined_input_for_runtime_call(call); 2455 } else { 2456 assert(!wide_out, "narrow in => narrow out"); 2457 Node* narrow_mem = memory(adr_type); 2458 prev_mem = reset_memory(); 2459 map()->set_memory(narrow_mem); 2460 set_predefined_input_for_runtime_call(call); 2461 } 2462 2463 // Hook each parm in order. Stop looking at the first NULL. 2464 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2465 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2466 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2467 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2468 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2469 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2470 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2471 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2472 /* close each nested if ===> */ } } } } } } } } 2473 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2474 2475 if (!is_leaf) { 2476 // Non-leaves can block and take safepoints: 2477 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2478 } 2479 // Non-leaves can throw exceptions: 2480 if (has_io) { 2481 call->set_req(TypeFunc::I_O, i_o()); 2482 } 2483 2484 if (flags & RC_UNCOMMON) { 2485 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2486 // (An "if" probability corresponds roughly to an unconditional count. 2487 // Sort of.) 2488 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2489 } 2490 2491 Node* c = _gvn.transform(call); 2492 assert(c == call, "cannot disappear"); 2493 2494 if (wide_out) { 2495 // Slow path call has full side-effects. 2496 set_predefined_output_for_runtime_call(call); 2497 } else { 2498 // Slow path call has few side-effects, and/or sets few values. 2499 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2500 } 2501 2502 if (has_io) { 2503 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2504 } 2505 return call; 2506 2507 } 2508 2509 //------------------------------merge_memory----------------------------------- 2510 // Merge memory from one path into the current memory state. 2511 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2512 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2513 Node* old_slice = mms.force_memory(); 2514 Node* new_slice = mms.memory2(); 2515 if (old_slice != new_slice) { 2516 PhiNode* phi; 2517 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2518 if (mms.is_empty()) { 2519 // clone base memory Phi's inputs for this memory slice 2520 assert(old_slice == mms.base_memory(), "sanity"); 2521 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2522 _gvn.set_type(phi, Type::MEMORY); 2523 for (uint i = 1; i < phi->req(); i++) { 2524 phi->init_req(i, old_slice->in(i)); 2525 } 2526 } else { 2527 phi = old_slice->as_Phi(); // Phi was generated already 2528 } 2529 } else { 2530 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2531 _gvn.set_type(phi, Type::MEMORY); 2532 } 2533 phi->set_req(new_path, new_slice); 2534 mms.set_memory(phi); 2535 } 2536 } 2537 } 2538 2539 //------------------------------make_slow_call_ex------------------------------ 2540 // Make the exception handler hookups for the slow call 2541 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2542 if (stopped()) return; 2543 2544 // Make a catch node with just two handlers: fall-through and catch-all 2545 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2546 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2547 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2548 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2549 2550 { PreserveJVMState pjvms(this); 2551 set_control(excp); 2552 set_i_o(i_o); 2553 2554 if (excp != top()) { 2555 if (deoptimize) { 2556 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2557 uncommon_trap(Deoptimization::Reason_unhandled, 2558 Deoptimization::Action_none); 2559 } else { 2560 // Create an exception state also. 2561 // Use an exact type if the caller has specified a specific exception. 2562 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2563 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2564 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2565 } 2566 } 2567 } 2568 2569 // Get the no-exception control from the CatchNode. 2570 set_control(norm); 2571 } 2572 2573 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2574 Node* cmp = NULL; 2575 switch(bt) { 2576 case T_INT: cmp = new CmpINode(in1, in2); break; 2577 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2578 default: fatal("unexpected comparison type %s", type2name(bt)); 2579 } 2580 gvn->transform(cmp); 2581 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2582 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2583 gvn->transform(iff); 2584 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2585 return iff; 2586 } 2587 2588 2589 //-------------------------------gen_subtype_check----------------------------- 2590 // Generate a subtyping check. Takes as input the subtype and supertype. 2591 // Returns 2 values: sets the default control() to the true path and returns 2592 // the false path. Only reads invariant memory; sets no (visible) memory. 2593 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2594 // but that's not exposed to the optimizer. This call also doesn't take in an 2595 // Object; if you wish to check an Object you need to load the Object's class 2596 // prior to coming here. 2597 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2598 Compile* C = gvn->C; 2599 2600 if ((*ctrl)->is_top()) { 2601 return C->top(); 2602 } 2603 2604 // Fast check for identical types, perhaps identical constants. 2605 // The types can even be identical non-constants, in cases 2606 // involving Array.newInstance, Object.clone, etc. 2607 if (subklass == superklass) 2608 return C->top(); // false path is dead; no test needed. 2609 2610 if (gvn->type(superklass)->singleton()) { 2611 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2612 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2613 2614 // In the common case of an exact superklass, try to fold up the 2615 // test before generating code. You may ask, why not just generate 2616 // the code and then let it fold up? The answer is that the generated 2617 // code will necessarily include null checks, which do not always 2618 // completely fold away. If they are also needless, then they turn 2619 // into a performance loss. Example: 2620 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2621 // Here, the type of 'fa' is often exact, so the store check 2622 // of fa[1]=x will fold up, without testing the nullness of x. 2623 switch (C->static_subtype_check(superk, subk)) { 2624 case Compile::SSC_always_false: 2625 { 2626 Node* always_fail = *ctrl; 2627 *ctrl = gvn->C->top(); 2628 return always_fail; 2629 } 2630 case Compile::SSC_always_true: 2631 return C->top(); 2632 case Compile::SSC_easy_test: 2633 { 2634 // Just do a direct pointer compare and be done. 2635 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2636 *ctrl = gvn->transform(new IfTrueNode(iff)); 2637 return gvn->transform(new IfFalseNode(iff)); 2638 } 2639 case Compile::SSC_full_test: 2640 break; 2641 default: 2642 ShouldNotReachHere(); 2643 } 2644 } 2645 2646 // %%% Possible further optimization: Even if the superklass is not exact, 2647 // if the subklass is the unique subtype of the superklass, the check 2648 // will always succeed. We could leave a dependency behind to ensure this. 2649 2650 // First load the super-klass's check-offset 2651 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2652 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2653 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2654 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2655 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2656 2657 // Load from the sub-klass's super-class display list, or a 1-word cache of 2658 // the secondary superclass list, or a failing value with a sentinel offset 2659 // if the super-klass is an interface or exceptionally deep in the Java 2660 // hierarchy and we have to scan the secondary superclass list the hard way. 2661 // Worst-case type is a little odd: NULL is allowed as a result (usually 2662 // klass loads can never produce a NULL). 2663 Node *chk_off_X = chk_off; 2664 #ifdef _LP64 2665 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2666 #endif 2667 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2668 // For some types like interfaces the following loadKlass is from a 1-word 2669 // cache which is mutable so can't use immutable memory. Other 2670 // types load from the super-class display table which is immutable. 2671 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2672 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2673 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2674 2675 // Compile speed common case: ARE a subtype and we canNOT fail 2676 if( superklass == nkls ) 2677 return C->top(); // false path is dead; no test needed. 2678 2679 // See if we get an immediate positive hit. Happens roughly 83% of the 2680 // time. Test to see if the value loaded just previously from the subklass 2681 // is exactly the superklass. 2682 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2683 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2684 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2685 2686 // Compile speed common case: Check for being deterministic right now. If 2687 // chk_off is a constant and not equal to cacheoff then we are NOT a 2688 // subklass. In this case we need exactly the 1 test above and we can 2689 // return those results immediately. 2690 if (!might_be_cache) { 2691 Node* not_subtype_ctrl = *ctrl; 2692 *ctrl = iftrue1; // We need exactly the 1 test above 2693 return not_subtype_ctrl; 2694 } 2695 2696 // Gather the various success & failures here 2697 RegionNode *r_ok_subtype = new RegionNode(4); 2698 gvn->record_for_igvn(r_ok_subtype); 2699 RegionNode *r_not_subtype = new RegionNode(3); 2700 gvn->record_for_igvn(r_not_subtype); 2701 2702 r_ok_subtype->init_req(1, iftrue1); 2703 2704 // Check for immediate negative hit. Happens roughly 11% of the time (which 2705 // is roughly 63% of the remaining cases). Test to see if the loaded 2706 // check-offset points into the subklass display list or the 1-element 2707 // cache. If it points to the display (and NOT the cache) and the display 2708 // missed then it's not a subtype. 2709 Node *cacheoff = gvn->intcon(cacheoff_con); 2710 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2711 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2712 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2713 2714 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2715 // No performance impact (too rare) but allows sharing of secondary arrays 2716 // which has some footprint reduction. 2717 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2718 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2719 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2720 2721 // -- Roads not taken here: -- 2722 // We could also have chosen to perform the self-check at the beginning 2723 // of this code sequence, as the assembler does. This would not pay off 2724 // the same way, since the optimizer, unlike the assembler, can perform 2725 // static type analysis to fold away many successful self-checks. 2726 // Non-foldable self checks work better here in second position, because 2727 // the initial primary superclass check subsumes a self-check for most 2728 // types. An exception would be a secondary type like array-of-interface, 2729 // which does not appear in its own primary supertype display. 2730 // Finally, we could have chosen to move the self-check into the 2731 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2732 // dependent manner. But it is worthwhile to have the check here, 2733 // where it can be perhaps be optimized. The cost in code space is 2734 // small (register compare, branch). 2735 2736 // Now do a linear scan of the secondary super-klass array. Again, no real 2737 // performance impact (too rare) but it's gotta be done. 2738 // Since the code is rarely used, there is no penalty for moving it 2739 // out of line, and it can only improve I-cache density. 2740 // The decision to inline or out-of-line this final check is platform 2741 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2742 Node* psc = gvn->transform( 2743 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2744 2745 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2746 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2747 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2748 2749 // Return false path; set default control to true path. 2750 *ctrl = gvn->transform(r_ok_subtype); 2751 return gvn->transform(r_not_subtype); 2752 } 2753 2754 // Profile-driven exact type check: 2755 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2756 float prob, 2757 Node* *casted_receiver) { 2758 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2759 Node* recv_klass = load_object_klass(receiver); 2760 Node* want_klass = makecon(tklass); 2761 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2762 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2763 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2764 set_control( _gvn.transform( new IfTrueNode (iff) )); 2765 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2766 2767 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2768 assert(recv_xtype->klass_is_exact(), ""); 2769 2770 // Subsume downstream occurrences of receiver with a cast to 2771 // recv_xtype, since now we know what the type will be. 2772 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2773 (*casted_receiver) = _gvn.transform(cast); 2774 // (User must make the replace_in_map call.) 2775 2776 return fail; 2777 } 2778 2779 2780 //------------------------------seems_never_null------------------------------- 2781 // Use null_seen information if it is available from the profile. 2782 // If we see an unexpected null at a type check we record it and force a 2783 // recompile; the offending check will be recompiled to handle NULLs. 2784 // If we see several offending BCIs, then all checks in the 2785 // method will be recompiled. 2786 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2787 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2788 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2789 if (UncommonNullCast // Cutout for this technique 2790 && obj != null() // And not the -Xcomp stupid case? 2791 && !too_many_traps(reason) 2792 ) { 2793 if (speculating) { 2794 return true; 2795 } 2796 if (data == NULL) 2797 // Edge case: no mature data. Be optimistic here. 2798 return true; 2799 // If the profile has not seen a null, assume it won't happen. 2800 assert(java_bc() == Bytecodes::_checkcast || 2801 java_bc() == Bytecodes::_instanceof || 2802 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2803 return !data->as_BitData()->null_seen(); 2804 } 2805 speculating = false; 2806 return false; 2807 } 2808 2809 //------------------------maybe_cast_profiled_receiver------------------------- 2810 // If the profile has seen exactly one type, narrow to exactly that type. 2811 // Subsequent type checks will always fold up. 2812 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2813 ciKlass* require_klass, 2814 ciKlass* spec_klass, 2815 bool safe_for_replace) { 2816 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2817 2818 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2819 2820 // Make sure we haven't already deoptimized from this tactic. 2821 if (too_many_traps(reason) || too_many_recompiles(reason)) 2822 return NULL; 2823 2824 // (No, this isn't a call, but it's enough like a virtual call 2825 // to use the same ciMethod accessor to get the profile info...) 2826 // If we have a speculative type use it instead of profiling (which 2827 // may not help us) 2828 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2829 if (exact_kls != NULL) {// no cast failures here 2830 if (require_klass == NULL || 2831 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2832 // If we narrow the type to match what the type profile sees or 2833 // the speculative type, we can then remove the rest of the 2834 // cast. 2835 // This is a win, even if the exact_kls is very specific, 2836 // because downstream operations, such as method calls, 2837 // will often benefit from the sharper type. 2838 Node* exact_obj = not_null_obj; // will get updated in place... 2839 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2840 &exact_obj); 2841 { PreserveJVMState pjvms(this); 2842 set_control(slow_ctl); 2843 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2844 } 2845 if (safe_for_replace) { 2846 replace_in_map(not_null_obj, exact_obj); 2847 } 2848 return exact_obj; 2849 } 2850 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2851 } 2852 2853 return NULL; 2854 } 2855 2856 /** 2857 * Cast obj to type and emit guard unless we had too many traps here 2858 * already 2859 * 2860 * @param obj node being casted 2861 * @param type type to cast the node to 2862 * @param not_null true if we know node cannot be null 2863 */ 2864 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2865 ciKlass* type, 2866 bool not_null) { 2867 if (stopped()) { 2868 return obj; 2869 } 2870 2871 // type == NULL if profiling tells us this object is always null 2872 if (type != NULL) { 2873 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2874 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2875 2876 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2877 !too_many_traps(class_reason) && 2878 !too_many_recompiles(class_reason)) { 2879 Node* not_null_obj = NULL; 2880 // not_null is true if we know the object is not null and 2881 // there's no need for a null check 2882 if (!not_null) { 2883 Node* null_ctl = top(); 2884 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2885 assert(null_ctl->is_top(), "no null control here"); 2886 } else { 2887 not_null_obj = obj; 2888 } 2889 2890 Node* exact_obj = not_null_obj; 2891 ciKlass* exact_kls = type; 2892 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2893 &exact_obj); 2894 { 2895 PreserveJVMState pjvms(this); 2896 set_control(slow_ctl); 2897 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2898 } 2899 replace_in_map(not_null_obj, exact_obj); 2900 obj = exact_obj; 2901 } 2902 } else { 2903 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2904 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2905 Node* exact_obj = null_assert(obj); 2906 replace_in_map(obj, exact_obj); 2907 obj = exact_obj; 2908 } 2909 } 2910 return obj; 2911 } 2912 2913 //-------------------------------gen_instanceof-------------------------------- 2914 // Generate an instance-of idiom. Used by both the instance-of bytecode 2915 // and the reflective instance-of call. 2916 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2917 kill_dead_locals(); // Benefit all the uncommon traps 2918 assert( !stopped(), "dead parse path should be checked in callers" ); 2919 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2920 "must check for not-null not-dead klass in callers"); 2921 2922 // Make the merge point 2923 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2924 RegionNode* region = new RegionNode(PATH_LIMIT); 2925 Node* phi = new PhiNode(region, TypeInt::BOOL); 2926 C->set_has_split_ifs(true); // Has chance for split-if optimization 2927 2928 ciProfileData* data = NULL; 2929 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2930 data = method()->method_data()->bci_to_data(bci()); 2931 } 2932 bool speculative_not_null = false; 2933 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2934 && seems_never_null(obj, data, speculative_not_null)); 2935 2936 // Null check; get casted pointer; set region slot 3 2937 Node* null_ctl = top(); 2938 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2939 2940 // If not_null_obj is dead, only null-path is taken 2941 if (stopped()) { // Doing instance-of on a NULL? 2942 set_control(null_ctl); 2943 return intcon(0); 2944 } 2945 region->init_req(_null_path, null_ctl); 2946 phi ->init_req(_null_path, intcon(0)); // Set null path value 2947 if (null_ctl == top()) { 2948 // Do this eagerly, so that pattern matches like is_diamond_phi 2949 // will work even during parsing. 2950 assert(_null_path == PATH_LIMIT-1, "delete last"); 2951 region->del_req(_null_path); 2952 phi ->del_req(_null_path); 2953 } 2954 2955 // Do we know the type check always succeed? 2956 bool known_statically = false; 2957 if (_gvn.type(superklass)->singleton()) { 2958 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2959 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2960 if (subk != NULL && subk->is_loaded()) { 2961 int static_res = C->static_subtype_check(superk, subk); 2962 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2963 } 2964 } 2965 2966 if (!known_statically) { 2967 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2968 // We may not have profiling here or it may not help us. If we 2969 // have a speculative type use it to perform an exact cast. 2970 ciKlass* spec_obj_type = obj_type->speculative_type(); 2971 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2972 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2973 if (stopped()) { // Profile disagrees with this path. 2974 set_control(null_ctl); // Null is the only remaining possibility. 2975 return intcon(0); 2976 } 2977 if (cast_obj != NULL) { 2978 not_null_obj = cast_obj; 2979 } 2980 } 2981 } 2982 2983 // Load the object's klass 2984 Node* obj_klass = load_object_klass(not_null_obj); 2985 2986 // Generate the subtype check 2987 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2988 2989 // Plug in the success path to the general merge in slot 1. 2990 region->init_req(_obj_path, control()); 2991 phi ->init_req(_obj_path, intcon(1)); 2992 2993 // Plug in the failing path to the general merge in slot 2. 2994 region->init_req(_fail_path, not_subtype_ctrl); 2995 phi ->init_req(_fail_path, intcon(0)); 2996 2997 // Return final merged results 2998 set_control( _gvn.transform(region) ); 2999 record_for_igvn(region); 3000 3001 // If we know the type check always succeeds then we don't use the 3002 // profiling data at this bytecode. Don't lose it, feed it to the 3003 // type system as a speculative type. 3004 if (safe_for_replace) { 3005 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3006 replace_in_map(obj, casted_obj); 3007 } 3008 3009 return _gvn.transform(phi); 3010 } 3011 3012 //-------------------------------gen_checkcast--------------------------------- 3013 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3014 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3015 // uncommon-trap paths work. Adjust stack after this call. 3016 // If failure_control is supplied and not null, it is filled in with 3017 // the control edge for the cast failure. Otherwise, an appropriate 3018 // uncommon trap or exception is thrown. 3019 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3020 Node* *failure_control) { 3021 kill_dead_locals(); // Benefit all the uncommon traps 3022 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3023 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3024 3025 // Fast cutout: Check the case that the cast is vacuously true. 3026 // This detects the common cases where the test will short-circuit 3027 // away completely. We do this before we perform the null check, 3028 // because if the test is going to turn into zero code, we don't 3029 // want a residual null check left around. (Causes a slowdown, 3030 // for example, in some objArray manipulations, such as a[i]=a[j].) 3031 if (tk->singleton()) { 3032 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3033 if (objtp != NULL && objtp->klass() != NULL) { 3034 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3035 case Compile::SSC_always_true: 3036 // If we know the type check always succeed then we don't use 3037 // the profiling data at this bytecode. Don't lose it, feed it 3038 // to the type system as a speculative type. 3039 return record_profiled_receiver_for_speculation(obj); 3040 case Compile::SSC_always_false: 3041 // It needs a null check because a null will *pass* the cast check. 3042 // A non-null value will always produce an exception. 3043 return null_assert(obj); 3044 } 3045 } 3046 } 3047 3048 ciProfileData* data = NULL; 3049 bool safe_for_replace = false; 3050 if (failure_control == NULL) { // use MDO in regular case only 3051 assert(java_bc() == Bytecodes::_aastore || 3052 java_bc() == Bytecodes::_checkcast, 3053 "interpreter profiles type checks only for these BCs"); 3054 data = method()->method_data()->bci_to_data(bci()); 3055 safe_for_replace = true; 3056 } 3057 3058 // Make the merge point 3059 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3060 RegionNode* region = new RegionNode(PATH_LIMIT); 3061 Node* phi = new PhiNode(region, toop); 3062 C->set_has_split_ifs(true); // Has chance for split-if optimization 3063 3064 // Use null-cast information if it is available 3065 bool speculative_not_null = false; 3066 bool never_see_null = ((failure_control == NULL) // regular case only 3067 && seems_never_null(obj, data, speculative_not_null)); 3068 3069 // Null check; get casted pointer; set region slot 3 3070 Node* null_ctl = top(); 3071 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3072 3073 // If not_null_obj is dead, only null-path is taken 3074 if (stopped()) { // Doing instance-of on a NULL? 3075 set_control(null_ctl); 3076 return null(); 3077 } 3078 region->init_req(_null_path, null_ctl); 3079 phi ->init_req(_null_path, null()); // Set null path value 3080 if (null_ctl == top()) { 3081 // Do this eagerly, so that pattern matches like is_diamond_phi 3082 // will work even during parsing. 3083 assert(_null_path == PATH_LIMIT-1, "delete last"); 3084 region->del_req(_null_path); 3085 phi ->del_req(_null_path); 3086 } 3087 3088 Node* cast_obj = NULL; 3089 if (tk->klass_is_exact()) { 3090 // The following optimization tries to statically cast the speculative type of the object 3091 // (for example obtained during profiling) to the type of the superklass and then do a 3092 // dynamic check that the type of the object is what we expect. To work correctly 3093 // for checkcast and aastore the type of superklass should be exact. 3094 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3095 // We may not have profiling here or it may not help us. If we have 3096 // a speculative type use it to perform an exact cast. 3097 ciKlass* spec_obj_type = obj_type->speculative_type(); 3098 if (spec_obj_type != NULL || data != NULL) { 3099 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3100 if (cast_obj != NULL) { 3101 if (failure_control != NULL) // failure is now impossible 3102 (*failure_control) = top(); 3103 // adjust the type of the phi to the exact klass: 3104 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3105 } 3106 } 3107 } 3108 3109 if (cast_obj == NULL) { 3110 // Load the object's klass 3111 Node* obj_klass = load_object_klass(not_null_obj); 3112 3113 // Generate the subtype check 3114 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3115 3116 // Plug in success path into the merge 3117 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3118 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3119 if (failure_control == NULL) { 3120 if (not_subtype_ctrl != top()) { // If failure is possible 3121 PreserveJVMState pjvms(this); 3122 set_control(not_subtype_ctrl); 3123 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3124 } 3125 } else { 3126 (*failure_control) = not_subtype_ctrl; 3127 } 3128 } 3129 3130 region->init_req(_obj_path, control()); 3131 phi ->init_req(_obj_path, cast_obj); 3132 3133 // A merge of NULL or Casted-NotNull obj 3134 Node* res = _gvn.transform(phi); 3135 3136 // Note I do NOT always 'replace_in_map(obj,result)' here. 3137 // if( tk->klass()->can_be_primary_super() ) 3138 // This means that if I successfully store an Object into an array-of-String 3139 // I 'forget' that the Object is really now known to be a String. I have to 3140 // do this because we don't have true union types for interfaces - if I store 3141 // a Baz into an array-of-Interface and then tell the optimizer it's an 3142 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3143 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3144 // replace_in_map( obj, res ); 3145 3146 // Return final merged results 3147 set_control( _gvn.transform(region) ); 3148 record_for_igvn(region); 3149 3150 return record_profiled_receiver_for_speculation(res); 3151 } 3152 3153 //------------------------------next_monitor----------------------------------- 3154 // What number should be given to the next monitor? 3155 int GraphKit::next_monitor() { 3156 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3157 int next = current + C->sync_stack_slots(); 3158 // Keep the toplevel high water mark current: 3159 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3160 return current; 3161 } 3162 3163 //------------------------------insert_mem_bar--------------------------------- 3164 // Memory barrier to avoid floating things around 3165 // The membar serves as a pinch point between both control and all memory slices. 3166 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3167 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3168 mb->init_req(TypeFunc::Control, control()); 3169 mb->init_req(TypeFunc::Memory, reset_memory()); 3170 Node* membar = _gvn.transform(mb); 3171 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3172 set_all_memory_call(membar); 3173 return membar; 3174 } 3175 3176 //-------------------------insert_mem_bar_volatile---------------------------- 3177 // Memory barrier to avoid floating things around 3178 // The membar serves as a pinch point between both control and memory(alias_idx). 3179 // If you want to make a pinch point on all memory slices, do not use this 3180 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3181 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3182 // When Parse::do_put_xxx updates a volatile field, it appends a series 3183 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3184 // The first membar is on the same memory slice as the field store opcode. 3185 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3186 // All the other membars (for other volatile slices, including AliasIdxBot, 3187 // which stands for all unknown volatile slices) are control-dependent 3188 // on the first membar. This prevents later volatile loads or stores 3189 // from sliding up past the just-emitted store. 3190 3191 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3192 mb->set_req(TypeFunc::Control,control()); 3193 if (alias_idx == Compile::AliasIdxBot) { 3194 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3195 } else { 3196 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3197 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3198 } 3199 Node* membar = _gvn.transform(mb); 3200 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3201 if (alias_idx == Compile::AliasIdxBot) { 3202 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3203 } else { 3204 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3205 } 3206 return membar; 3207 } 3208 3209 //------------------------------shared_lock------------------------------------ 3210 // Emit locking code. 3211 FastLockNode* GraphKit::shared_lock(Node* obj) { 3212 // bci is either a monitorenter bc or InvocationEntryBci 3213 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3214 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3215 3216 if( !GenerateSynchronizationCode ) 3217 return NULL; // Not locking things? 3218 if (stopped()) // Dead monitor? 3219 return NULL; 3220 3221 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3222 3223 // Box the stack location 3224 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3225 Node* mem = reset_memory(); 3226 3227 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3228 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3229 // Create the counters for this fast lock. 3230 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3231 } 3232 3233 // Create the rtm counters for this fast lock if needed. 3234 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3235 3236 // Add monitor to debug info for the slow path. If we block inside the 3237 // slow path and de-opt, we need the monitor hanging around 3238 map()->push_monitor( flock ); 3239 3240 const TypeFunc *tf = LockNode::lock_type(); 3241 LockNode *lock = new LockNode(C, tf); 3242 3243 lock->init_req( TypeFunc::Control, control() ); 3244 lock->init_req( TypeFunc::Memory , mem ); 3245 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3246 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3247 lock->init_req( TypeFunc::ReturnAdr, top() ); 3248 3249 lock->init_req(TypeFunc::Parms + 0, obj); 3250 lock->init_req(TypeFunc::Parms + 1, box); 3251 lock->init_req(TypeFunc::Parms + 2, flock); 3252 add_safepoint_edges(lock); 3253 3254 lock = _gvn.transform( lock )->as_Lock(); 3255 3256 // lock has no side-effects, sets few values 3257 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3258 3259 insert_mem_bar(Op_MemBarAcquireLock); 3260 3261 // Add this to the worklist so that the lock can be eliminated 3262 record_for_igvn(lock); 3263 3264 #ifndef PRODUCT 3265 if (PrintLockStatistics) { 3266 // Update the counter for this lock. Don't bother using an atomic 3267 // operation since we don't require absolute accuracy. 3268 lock->create_lock_counter(map()->jvms()); 3269 increment_counter(lock->counter()->addr()); 3270 } 3271 #endif 3272 3273 return flock; 3274 } 3275 3276 3277 //------------------------------shared_unlock---------------------------------- 3278 // Emit unlocking code. 3279 void GraphKit::shared_unlock(Node* box, Node* obj) { 3280 // bci is either a monitorenter bc or InvocationEntryBci 3281 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3282 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3283 3284 if( !GenerateSynchronizationCode ) 3285 return; 3286 if (stopped()) { // Dead monitor? 3287 map()->pop_monitor(); // Kill monitor from debug info 3288 return; 3289 } 3290 3291 // Memory barrier to avoid floating things down past the locked region 3292 insert_mem_bar(Op_MemBarReleaseLock); 3293 3294 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3295 UnlockNode *unlock = new UnlockNode(C, tf); 3296 #ifdef ASSERT 3297 unlock->set_dbg_jvms(sync_jvms()); 3298 #endif 3299 uint raw_idx = Compile::AliasIdxRaw; 3300 unlock->init_req( TypeFunc::Control, control() ); 3301 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3302 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3303 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3304 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3305 3306 unlock->init_req(TypeFunc::Parms + 0, obj); 3307 unlock->init_req(TypeFunc::Parms + 1, box); 3308 unlock = _gvn.transform(unlock)->as_Unlock(); 3309 3310 Node* mem = reset_memory(); 3311 3312 // unlock has no side-effects, sets few values 3313 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3314 3315 // Kill monitor from debug info 3316 map()->pop_monitor( ); 3317 } 3318 3319 //-------------------------------get_layout_helper----------------------------- 3320 // If the given klass is a constant or known to be an array, 3321 // fetch the constant layout helper value into constant_value 3322 // and return (Node*)NULL. Otherwise, load the non-constant 3323 // layout helper value, and return the node which represents it. 3324 // This two-faced routine is useful because allocation sites 3325 // almost always feature constant types. 3326 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3327 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3328 if (!StressReflectiveCode && inst_klass != NULL) { 3329 ciKlass* klass = inst_klass->klass(); 3330 bool xklass = inst_klass->klass_is_exact(); 3331 if (xklass || klass->is_array_klass()) { 3332 jint lhelper = klass->layout_helper(); 3333 if (lhelper != Klass::_lh_neutral_value) { 3334 constant_value = lhelper; 3335 return (Node*) NULL; 3336 } 3337 } 3338 } 3339 constant_value = Klass::_lh_neutral_value; // put in a known value 3340 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3341 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3342 } 3343 3344 // We just put in an allocate/initialize with a big raw-memory effect. 3345 // Hook selected additional alias categories on the initialization. 3346 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3347 MergeMemNode* init_in_merge, 3348 Node* init_out_raw) { 3349 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3350 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3351 3352 Node* prevmem = kit.memory(alias_idx); 3353 init_in_merge->set_memory_at(alias_idx, prevmem); 3354 kit.set_memory(init_out_raw, alias_idx); 3355 } 3356 3357 //---------------------------set_output_for_allocation------------------------- 3358 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3359 const TypeOopPtr* oop_type, 3360 bool deoptimize_on_exception) { 3361 int rawidx = Compile::AliasIdxRaw; 3362 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3363 add_safepoint_edges(alloc); 3364 Node* allocx = _gvn.transform(alloc); 3365 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3366 // create memory projection for i_o 3367 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3368 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3369 3370 // create a memory projection as for the normal control path 3371 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3372 set_memory(malloc, rawidx); 3373 3374 // a normal slow-call doesn't change i_o, but an allocation does 3375 // we create a separate i_o projection for the normal control path 3376 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3377 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3378 3379 // put in an initialization barrier 3380 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3381 rawoop)->as_Initialize(); 3382 assert(alloc->initialization() == init, "2-way macro link must work"); 3383 assert(init ->allocation() == alloc, "2-way macro link must work"); 3384 { 3385 // Extract memory strands which may participate in the new object's 3386 // initialization, and source them from the new InitializeNode. 3387 // This will allow us to observe initializations when they occur, 3388 // and link them properly (as a group) to the InitializeNode. 3389 assert(init->in(InitializeNode::Memory) == malloc, ""); 3390 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3391 init->set_req(InitializeNode::Memory, minit_in); 3392 record_for_igvn(minit_in); // fold it up later, if possible 3393 Node* minit_out = memory(rawidx); 3394 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3395 if (oop_type->isa_aryptr()) { 3396 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3397 int elemidx = C->get_alias_index(telemref); 3398 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3399 } else if (oop_type->isa_instptr()) { 3400 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3401 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3402 ciField* field = ik->nonstatic_field_at(i); 3403 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3404 continue; // do not bother to track really large numbers of fields 3405 // Find (or create) the alias category for this field: 3406 int fieldidx = C->alias_type(field)->index(); 3407 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3408 } 3409 } 3410 } 3411 3412 // Cast raw oop to the real thing... 3413 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3414 javaoop = _gvn.transform(javaoop); 3415 C->set_recent_alloc(control(), javaoop); 3416 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3417 3418 #ifdef ASSERT 3419 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3420 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3421 "Ideal_allocation works"); 3422 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3423 "Ideal_allocation works"); 3424 if (alloc->is_AllocateArray()) { 3425 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3426 "Ideal_allocation works"); 3427 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3428 "Ideal_allocation works"); 3429 } else { 3430 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3431 } 3432 } 3433 #endif //ASSERT 3434 3435 return javaoop; 3436 } 3437 3438 //---------------------------new_instance-------------------------------------- 3439 // This routine takes a klass_node which may be constant (for a static type) 3440 // or may be non-constant (for reflective code). It will work equally well 3441 // for either, and the graph will fold nicely if the optimizer later reduces 3442 // the type to a constant. 3443 // The optional arguments are for specialized use by intrinsics: 3444 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3445 // - If 'return_size_val', report the the total object size to the caller. 3446 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3447 Node* GraphKit::new_instance(Node* klass_node, 3448 Node* extra_slow_test, 3449 Node* *return_size_val, 3450 bool deoptimize_on_exception) { 3451 // Compute size in doublewords 3452 // The size is always an integral number of doublewords, represented 3453 // as a positive bytewise size stored in the klass's layout_helper. 3454 // The layout_helper also encodes (in a low bit) the need for a slow path. 3455 jint layout_con = Klass::_lh_neutral_value; 3456 Node* layout_val = get_layout_helper(klass_node, layout_con); 3457 int layout_is_con = (layout_val == NULL); 3458 3459 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3460 // Generate the initial go-slow test. It's either ALWAYS (return a 3461 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3462 // case) a computed value derived from the layout_helper. 3463 Node* initial_slow_test = NULL; 3464 if (layout_is_con) { 3465 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3466 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3467 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3468 } else { // reflective case 3469 // This reflective path is used by Unsafe.allocateInstance. 3470 // (It may be stress-tested by specifying StressReflectiveCode.) 3471 // Basically, we want to get into the VM is there's an illegal argument. 3472 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3473 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3474 if (extra_slow_test != intcon(0)) { 3475 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3476 } 3477 // (Macro-expander will further convert this to a Bool, if necessary.) 3478 } 3479 3480 // Find the size in bytes. This is easy; it's the layout_helper. 3481 // The size value must be valid even if the slow path is taken. 3482 Node* size = NULL; 3483 if (layout_is_con) { 3484 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3485 } else { // reflective case 3486 // This reflective path is used by clone and Unsafe.allocateInstance. 3487 size = ConvI2X(layout_val); 3488 3489 // Clear the low bits to extract layout_helper_size_in_bytes: 3490 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3491 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3492 size = _gvn.transform( new AndXNode(size, mask) ); 3493 } 3494 if (return_size_val != NULL) { 3495 (*return_size_val) = size; 3496 } 3497 3498 // This is a precise notnull oop of the klass. 3499 // (Actually, it need not be precise if this is a reflective allocation.) 3500 // It's what we cast the result to. 3501 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3502 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3503 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3504 3505 // Now generate allocation code 3506 3507 // The entire memory state is needed for slow path of the allocation 3508 // since GC and deoptimization can happened. 3509 Node *mem = reset_memory(); 3510 set_all_memory(mem); // Create new memory state 3511 3512 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3513 control(), mem, i_o(), 3514 size, klass_node, 3515 initial_slow_test); 3516 3517 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3518 } 3519 3520 //-------------------------------new_array------------------------------------- 3521 // helper for both newarray and anewarray 3522 // The 'length' parameter is (obviously) the length of the array. 3523 // See comments on new_instance for the meaning of the other arguments. 3524 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3525 Node* length, // number of array elements 3526 int nargs, // number of arguments to push back for uncommon trap 3527 Node* *return_size_val, 3528 bool deoptimize_on_exception) { 3529 jint layout_con = Klass::_lh_neutral_value; 3530 Node* layout_val = get_layout_helper(klass_node, layout_con); 3531 int layout_is_con = (layout_val == NULL); 3532 3533 if (!layout_is_con && !StressReflectiveCode && 3534 !too_many_traps(Deoptimization::Reason_class_check)) { 3535 // This is a reflective array creation site. 3536 // Optimistically assume that it is a subtype of Object[], 3537 // so that we can fold up all the address arithmetic. 3538 layout_con = Klass::array_layout_helper(T_OBJECT); 3539 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3540 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3541 { BuildCutout unless(this, bol_lh, PROB_MAX); 3542 inc_sp(nargs); 3543 uncommon_trap(Deoptimization::Reason_class_check, 3544 Deoptimization::Action_maybe_recompile); 3545 } 3546 layout_val = NULL; 3547 layout_is_con = true; 3548 } 3549 3550 // Generate the initial go-slow test. Make sure we do not overflow 3551 // if length is huge (near 2Gig) or negative! We do not need 3552 // exact double-words here, just a close approximation of needed 3553 // double-words. We can't add any offset or rounding bits, lest we 3554 // take a size -1 of bytes and make it positive. Use an unsigned 3555 // compare, so negative sizes look hugely positive. 3556 int fast_size_limit = FastAllocateSizeLimit; 3557 if (layout_is_con) { 3558 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3559 // Increase the size limit if we have exact knowledge of array type. 3560 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3561 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3562 } 3563 3564 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3565 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3566 3567 // --- Size Computation --- 3568 // array_size = round_to_heap(array_header + (length << elem_shift)); 3569 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3570 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3571 // The rounding mask is strength-reduced, if possible. 3572 int round_mask = MinObjAlignmentInBytes - 1; 3573 Node* header_size = NULL; 3574 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3575 // (T_BYTE has the weakest alignment and size restrictions...) 3576 if (layout_is_con) { 3577 int hsize = Klass::layout_helper_header_size(layout_con); 3578 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3579 BasicType etype = Klass::layout_helper_element_type(layout_con); 3580 if ((round_mask & ~right_n_bits(eshift)) == 0) 3581 round_mask = 0; // strength-reduce it if it goes away completely 3582 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3583 assert(header_size_min <= hsize, "generic minimum is smallest"); 3584 header_size_min = hsize; 3585 header_size = intcon(hsize + round_mask); 3586 } else { 3587 Node* hss = intcon(Klass::_lh_header_size_shift); 3588 Node* hsm = intcon(Klass::_lh_header_size_mask); 3589 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3590 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3591 Node* mask = intcon(round_mask); 3592 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3593 } 3594 3595 Node* elem_shift = NULL; 3596 if (layout_is_con) { 3597 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3598 if (eshift != 0) 3599 elem_shift = intcon(eshift); 3600 } else { 3601 // There is no need to mask or shift this value. 3602 // The semantics of LShiftINode include an implicit mask to 0x1F. 3603 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3604 elem_shift = layout_val; 3605 } 3606 3607 // Transition to native address size for all offset calculations: 3608 Node* lengthx = ConvI2X(length); 3609 Node* headerx = ConvI2X(header_size); 3610 #ifdef _LP64 3611 { const TypeInt* tilen = _gvn.find_int_type(length); 3612 if (tilen != NULL && tilen->_lo < 0) { 3613 // Add a manual constraint to a positive range. Cf. array_element_address. 3614 jint size_max = fast_size_limit; 3615 if (size_max > tilen->_hi) size_max = tilen->_hi; 3616 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3617 3618 // Only do a narrow I2L conversion if the range check passed. 3619 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3620 _gvn.transform(iff); 3621 RegionNode* region = new RegionNode(3); 3622 _gvn.set_type(region, Type::CONTROL); 3623 lengthx = new PhiNode(region, TypeLong::LONG); 3624 _gvn.set_type(lengthx, TypeLong::LONG); 3625 3626 // Range check passed. Use ConvI2L node with narrow type. 3627 Node* passed = IfFalse(iff); 3628 region->init_req(1, passed); 3629 // Make I2L conversion control dependent to prevent it from 3630 // floating above the range check during loop optimizations. 3631 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3632 3633 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3634 region->init_req(2, IfTrue(iff)); 3635 lengthx->init_req(2, ConvI2X(length)); 3636 3637 set_control(region); 3638 record_for_igvn(region); 3639 record_for_igvn(lengthx); 3640 } 3641 } 3642 #endif 3643 3644 // Combine header size (plus rounding) and body size. Then round down. 3645 // This computation cannot overflow, because it is used only in two 3646 // places, one where the length is sharply limited, and the other 3647 // after a successful allocation. 3648 Node* abody = lengthx; 3649 if (elem_shift != NULL) 3650 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3651 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3652 if (round_mask != 0) { 3653 Node* mask = MakeConX(~round_mask); 3654 size = _gvn.transform( new AndXNode(size, mask) ); 3655 } 3656 // else if round_mask == 0, the size computation is self-rounding 3657 3658 if (return_size_val != NULL) { 3659 // This is the size 3660 (*return_size_val) = size; 3661 } 3662 3663 // Now generate allocation code 3664 3665 // The entire memory state is needed for slow path of the allocation 3666 // since GC and deoptimization can happened. 3667 Node *mem = reset_memory(); 3668 set_all_memory(mem); // Create new memory state 3669 3670 if (initial_slow_test->is_Bool()) { 3671 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3672 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3673 } 3674 3675 // Create the AllocateArrayNode and its result projections 3676 AllocateArrayNode* alloc 3677 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3678 control(), mem, i_o(), 3679 size, klass_node, 3680 initial_slow_test, 3681 length); 3682 3683 // Cast to correct type. Note that the klass_node may be constant or not, 3684 // and in the latter case the actual array type will be inexact also. 3685 // (This happens via a non-constant argument to inline_native_newArray.) 3686 // In any case, the value of klass_node provides the desired array type. 3687 const TypeInt* length_type = _gvn.find_int_type(length); 3688 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3689 if (ary_type->isa_aryptr() && length_type != NULL) { 3690 // Try to get a better type than POS for the size 3691 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3692 } 3693 3694 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3695 3696 // Cast length on remaining path to be as narrow as possible 3697 if (map()->find_edge(length) >= 0) { 3698 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3699 if (ccast != length) { 3700 _gvn.set_type_bottom(ccast); 3701 record_for_igvn(ccast); 3702 replace_in_map(length, ccast); 3703 } 3704 } 3705 3706 return javaoop; 3707 } 3708 3709 // The following "Ideal_foo" functions are placed here because they recognize 3710 // the graph shapes created by the functions immediately above. 3711 3712 //---------------------------Ideal_allocation---------------------------------- 3713 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3714 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3715 if (ptr == NULL) { // reduce dumb test in callers 3716 return NULL; 3717 } 3718 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3719 ptr = ptr->in(1); 3720 if (ptr == NULL) return NULL; 3721 } 3722 // Return NULL for allocations with several casts: 3723 // j.l.reflect.Array.newInstance(jobject, jint) 3724 // Object.clone() 3725 // to keep more precise type from last cast. 3726 if (ptr->is_Proj()) { 3727 Node* allo = ptr->in(0); 3728 if (allo != NULL && allo->is_Allocate()) { 3729 return allo->as_Allocate(); 3730 } 3731 } 3732 // Report failure to match. 3733 return NULL; 3734 } 3735 3736 // Fancy version which also strips off an offset (and reports it to caller). 3737 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3738 intptr_t& offset) { 3739 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3740 if (base == NULL) return NULL; 3741 return Ideal_allocation(base, phase); 3742 } 3743 3744 // Trace Initialize <- Proj[Parm] <- Allocate 3745 AllocateNode* InitializeNode::allocation() { 3746 Node* rawoop = in(InitializeNode::RawAddress); 3747 if (rawoop->is_Proj()) { 3748 Node* alloc = rawoop->in(0); 3749 if (alloc->is_Allocate()) { 3750 return alloc->as_Allocate(); 3751 } 3752 } 3753 return NULL; 3754 } 3755 3756 // Trace Allocate -> Proj[Parm] -> Initialize 3757 InitializeNode* AllocateNode::initialization() { 3758 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3759 if (rawoop == NULL) return NULL; 3760 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3761 Node* init = rawoop->fast_out(i); 3762 if (init->is_Initialize()) { 3763 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3764 return init->as_Initialize(); 3765 } 3766 } 3767 return NULL; 3768 } 3769 3770 //----------------------------- loop predicates --------------------------- 3771 3772 //------------------------------add_predicate_impl---------------------------- 3773 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3774 // Too many traps seen? 3775 if (too_many_traps(reason)) { 3776 #ifdef ASSERT 3777 if (TraceLoopPredicate) { 3778 int tc = C->trap_count(reason); 3779 tty->print("too many traps=%s tcount=%d in ", 3780 Deoptimization::trap_reason_name(reason), tc); 3781 method()->print(); // which method has too many predicate traps 3782 tty->cr(); 3783 } 3784 #endif 3785 // We cannot afford to take more traps here, 3786 // do not generate predicate. 3787 return; 3788 } 3789 3790 Node *cont = _gvn.intcon(1); 3791 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3792 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3793 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3794 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3795 C->add_predicate_opaq(opq); 3796 { 3797 PreserveJVMState pjvms(this); 3798 set_control(iffalse); 3799 inc_sp(nargs); 3800 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3801 } 3802 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3803 set_control(iftrue); 3804 } 3805 3806 //------------------------------add_predicate--------------------------------- 3807 void GraphKit::add_predicate(int nargs) { 3808 if (UseLoopPredicate) { 3809 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3810 } 3811 // loop's limit check predicate should be near the loop. 3812 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3813 } 3814 3815 //----------------------------- store barriers ---------------------------- 3816 #define __ ideal. 3817 3818 bool GraphKit::use_ReduceInitialCardMarks() { 3819 BarrierSet *bs = BarrierSet::barrier_set(); 3820 return bs->is_a(BarrierSet::CardTableBarrierSet) 3821 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers() 3822 && ReduceInitialCardMarks; 3823 } 3824 3825 void GraphKit::sync_kit(IdealKit& ideal) { 3826 set_all_memory(__ merged_memory()); 3827 set_i_o(__ i_o()); 3828 set_control(__ ctrl()); 3829 } 3830 3831 void GraphKit::final_sync(IdealKit& ideal) { 3832 // Final sync IdealKit and graphKit. 3833 sync_kit(ideal); 3834 } 3835 3836 Node* GraphKit::byte_map_base_node() { 3837 // Get base of card map 3838 jbyte* card_table_base = ci_card_table_address(); 3839 if (card_table_base != NULL) { 3840 return makecon(TypeRawPtr::make((address)card_table_base)); 3841 } else { 3842 return null(); 3843 } 3844 } 3845 3846 // vanilla/CMS post barrier 3847 // Insert a write-barrier store. This is to let generational GC work; we have 3848 // to flag all oop-stores before the next GC point. 3849 void GraphKit::write_barrier_post(Node* oop_store, 3850 Node* obj, 3851 Node* adr, 3852 uint adr_idx, 3853 Node* val, 3854 bool use_precise) { 3855 // No store check needed if we're storing a NULL or an old object 3856 // (latter case is probably a string constant). The concurrent 3857 // mark sweep garbage collector, however, needs to have all nonNull 3858 // oop updates flagged via card-marks. 3859 if (val != NULL && val->is_Con()) { 3860 // must be either an oop or NULL 3861 const Type* t = val->bottom_type(); 3862 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3863 // stores of null never (?) need barriers 3864 return; 3865 } 3866 3867 if (use_ReduceInitialCardMarks() 3868 && obj == just_allocated_object(control())) { 3869 // We can skip marks on a freshly-allocated object in Eden. 3870 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 3871 // That routine informs GC to take appropriate compensating steps, 3872 // upon a slow-path allocation, so as to make this card-mark 3873 // elision safe. 3874 return; 3875 } 3876 3877 if (!use_precise) { 3878 // All card marks for a (non-array) instance are in one place: 3879 adr = obj; 3880 } 3881 // (Else it's an array (or unknown), and we want more precise card marks.) 3882 assert(adr != NULL, ""); 3883 3884 IdealKit ideal(this, true); 3885 3886 // Convert the pointer to an int prior to doing math on it 3887 Node* cast = __ CastPX(__ ctrl(), adr); 3888 3889 // Divide by card size 3890 assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet), 3891 "Only one we handle so far."); 3892 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 3893 3894 // Combine card table base and card offset 3895 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3896 3897 // Get the alias_index for raw card-mark memory 3898 int adr_type = Compile::AliasIdxRaw; 3899 Node* zero = __ ConI(0); // Dirty card value 3900 BasicType bt = T_BYTE; 3901 3902 if (UseConcMarkSweepGC && UseCondCardMark) { 3903 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3904 __ sync_kit(this); 3905 } 3906 3907 if (UseCondCardMark) { 3908 // The classic GC reference write barrier is typically implemented 3909 // as a store into the global card mark table. Unfortunately 3910 // unconditional stores can result in false sharing and excessive 3911 // coherence traffic as well as false transactional aborts. 3912 // UseCondCardMark enables MP "polite" conditional card mark 3913 // stores. In theory we could relax the load from ctrl() to 3914 // no_ctrl, but that doesn't buy much latitude. 3915 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3916 __ if_then(card_val, BoolTest::ne, zero); 3917 } 3918 3919 // Smash zero into card 3920 if( !UseConcMarkSweepGC ) { 3921 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3922 } else { 3923 // Specialized path for CM store barrier 3924 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3925 } 3926 3927 if (UseCondCardMark) { 3928 __ end_if(); 3929 } 3930 3931 // Final sync IdealKit and GraphKit. 3932 final_sync(ideal); 3933 } 3934 /* 3935 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3936 * required by SATB to make sure all objects live at the start of the 3937 * marking are kept alive, all reference updates need to any previous 3938 * reference stored before writing. 3939 * 3940 * If the previous value is NULL there is no need to save the old value. 3941 * References that are NULL are filtered during runtime by the barrier 3942 * code to avoid unnecessary queuing. 3943 * 3944 * However in the case of newly allocated objects it might be possible to 3945 * prove that the reference about to be overwritten is NULL during compile 3946 * time and avoid adding the barrier code completely. 3947 * 3948 * The compiler needs to determine that the object in which a field is about 3949 * to be written is newly allocated, and that no prior store to the same field 3950 * has happened since the allocation. 3951 * 3952 * Returns true if the pre-barrier can be removed 3953 */ 3954 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3955 BasicType bt, uint adr_idx) { 3956 intptr_t offset = 0; 3957 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3958 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3959 3960 if (offset == Type::OffsetBot) { 3961 return false; // cannot unalias unless there are precise offsets 3962 } 3963 3964 if (alloc == NULL) { 3965 return false; // No allocation found 3966 } 3967 3968 intptr_t size_in_bytes = type2aelembytes(bt); 3969 3970 Node* mem = memory(adr_idx); // start searching here... 3971 3972 for (int cnt = 0; cnt < 50; cnt++) { 3973 3974 if (mem->is_Store()) { 3975 3976 Node* st_adr = mem->in(MemNode::Address); 3977 intptr_t st_offset = 0; 3978 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3979 3980 if (st_base == NULL) { 3981 break; // inscrutable pointer 3982 } 3983 3984 // Break we have found a store with same base and offset as ours so break 3985 if (st_base == base && st_offset == offset) { 3986 break; 3987 } 3988 3989 if (st_offset != offset && st_offset != Type::OffsetBot) { 3990 const int MAX_STORE = BytesPerLong; 3991 if (st_offset >= offset + size_in_bytes || 3992 st_offset <= offset - MAX_STORE || 3993 st_offset <= offset - mem->as_Store()->memory_size()) { 3994 // Success: The offsets are provably independent. 3995 // (You may ask, why not just test st_offset != offset and be done? 3996 // The answer is that stores of different sizes can co-exist 3997 // in the same sequence of RawMem effects. We sometimes initialize 3998 // a whole 'tile' of array elements with a single jint or jlong.) 3999 mem = mem->in(MemNode::Memory); 4000 continue; // advance through independent store memory 4001 } 4002 } 4003 4004 if (st_base != base 4005 && MemNode::detect_ptr_independence(base, alloc, st_base, 4006 AllocateNode::Ideal_allocation(st_base, phase), 4007 phase)) { 4008 // Success: The bases are provably independent. 4009 mem = mem->in(MemNode::Memory); 4010 continue; // advance through independent store memory 4011 } 4012 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4013 4014 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4015 AllocateNode* st_alloc = st_init->allocation(); 4016 4017 // Make sure that we are looking at the same allocation site. 4018 // The alloc variable is guaranteed to not be null here from earlier check. 4019 if (alloc == st_alloc) { 4020 // Check that the initialization is storing NULL so that no previous store 4021 // has been moved up and directly write a reference 4022 Node* captured_store = st_init->find_captured_store(offset, 4023 type2aelembytes(T_OBJECT), 4024 phase); 4025 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4026 return true; 4027 } 4028 } 4029 } 4030 4031 // Unless there is an explicit 'continue', we must bail out here, 4032 // because 'mem' is an inscrutable memory state (e.g., a call). 4033 break; 4034 } 4035 4036 return false; 4037 } 4038 4039 // G1 pre/post barriers 4040 void GraphKit::g1_write_barrier_pre(bool do_load, 4041 Node* obj, 4042 Node* adr, 4043 uint alias_idx, 4044 Node* val, 4045 const TypeOopPtr* val_type, 4046 Node* pre_val, 4047 BasicType bt) { 4048 4049 // Some sanity checks 4050 // Note: val is unused in this routine. 4051 4052 if (do_load) { 4053 // We need to generate the load of the previous value 4054 assert(obj != NULL, "must have a base"); 4055 assert(adr != NULL, "where are loading from?"); 4056 assert(pre_val == NULL, "loaded already?"); 4057 assert(val_type != NULL, "need a type"); 4058 4059 if (use_ReduceInitialCardMarks() 4060 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4061 return; 4062 } 4063 4064 } else { 4065 // In this case both val_type and alias_idx are unused. 4066 assert(pre_val != NULL, "must be loaded already"); 4067 // Nothing to be done if pre_val is null. 4068 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4069 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4070 } 4071 assert(bt == T_OBJECT, "or we shouldn't be here"); 4072 4073 IdealKit ideal(this, true); 4074 4075 Node* tls = __ thread(); // ThreadLocalStorage 4076 4077 Node* no_ctrl = NULL; 4078 Node* no_base = __ top(); 4079 Node* zero = __ ConI(0); 4080 Node* zeroX = __ ConX(0); 4081 4082 float likely = PROB_LIKELY(0.999); 4083 float unlikely = PROB_UNLIKELY(0.999); 4084 4085 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4086 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4087 4088 // Offsets into the thread 4089 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 4090 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset()); 4091 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()); 4092 4093 // Now the actual pointers into the thread 4094 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4095 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4096 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4097 4098 // Now some of the values 4099 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4100 4101 // if (!marking) 4102 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4103 BasicType index_bt = TypeX_X->basic_type(); 4104 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4105 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4106 4107 if (do_load) { 4108 // load original value 4109 // alias_idx correct?? 4110 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4111 } 4112 4113 // if (pre_val != NULL) 4114 __ if_then(pre_val, BoolTest::ne, null()); { 4115 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4116 4117 // is the queue for this thread full? 4118 __ if_then(index, BoolTest::ne, zeroX, likely); { 4119 4120 // decrement the index 4121 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4122 4123 // Now get the buffer location we will log the previous value into and store it 4124 Node *log_addr = __ AddP(no_base, buffer, next_index); 4125 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4126 // update the index 4127 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4128 4129 } __ else_(); { 4130 4131 // logging buffer is full, call the runtime 4132 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4133 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4134 } __ end_if(); // (!index) 4135 } __ end_if(); // (pre_val != NULL) 4136 } __ end_if(); // (!marking) 4137 4138 // Final sync IdealKit and GraphKit. 4139 final_sync(ideal); 4140 } 4141 4142 /* 4143 * G1 similar to any GC with a Young Generation requires a way to keep track of 4144 * references from Old Generation to Young Generation to make sure all live 4145 * objects are found. G1 also requires to keep track of object references 4146 * between different regions to enable evacuation of old regions, which is done 4147 * as part of mixed collections. References are tracked in remembered sets and 4148 * is continuously updated as reference are written to with the help of the 4149 * post-barrier. 4150 * 4151 * To reduce the number of updates to the remembered set the post-barrier 4152 * filters updates to fields in objects located in the Young Generation, 4153 * the same region as the reference, when the NULL is being written or 4154 * if the card is already marked as dirty by an earlier write. 4155 * 4156 * Under certain circumstances it is possible to avoid generating the 4157 * post-barrier completely if it is possible during compile time to prove 4158 * the object is newly allocated and that no safepoint exists between the 4159 * allocation and the store. 4160 * 4161 * In the case of slow allocation the allocation code must handle the barrier 4162 * as part of the allocation in the case the allocated object is not located 4163 * in the nursery, this would happen for humongous objects. This is similar to 4164 * how CMS is required to handle this case, see the comments for the method 4165 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier. 4166 * A deferred card mark is required for these objects and handled in the above 4167 * mentioned methods. 4168 * 4169 * Returns true if the post barrier can be removed 4170 */ 4171 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4172 Node* adr) { 4173 intptr_t offset = 0; 4174 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4175 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4176 4177 if (offset == Type::OffsetBot) { 4178 return false; // cannot unalias unless there are precise offsets 4179 } 4180 4181 if (alloc == NULL) { 4182 return false; // No allocation found 4183 } 4184 4185 // Start search from Store node 4186 Node* mem = store->in(MemNode::Control); 4187 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4188 4189 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4190 AllocateNode* st_alloc = st_init->allocation(); 4191 4192 // Make sure we are looking at the same allocation 4193 if (alloc == st_alloc) { 4194 return true; 4195 } 4196 } 4197 4198 return false; 4199 } 4200 4201 // 4202 // Update the card table and add card address to the queue 4203 // 4204 void GraphKit::g1_mark_card(IdealKit& ideal, 4205 Node* card_adr, 4206 Node* oop_store, 4207 uint oop_alias_idx, 4208 Node* index, 4209 Node* index_adr, 4210 Node* buffer, 4211 const TypeFunc* tf) { 4212 4213 Node* zero = __ ConI(0); 4214 Node* zeroX = __ ConX(0); 4215 Node* no_base = __ top(); 4216 BasicType card_bt = T_BYTE; 4217 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4218 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4219 4220 // Now do the queue work 4221 __ if_then(index, BoolTest::ne, zeroX); { 4222 4223 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4224 Node* log_addr = __ AddP(no_base, buffer, next_index); 4225 4226 // Order, see storeCM. 4227 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4228 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4229 4230 } __ else_(); { 4231 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4232 } __ end_if(); 4233 4234 } 4235 4236 void GraphKit::g1_write_barrier_post(Node* oop_store, 4237 Node* obj, 4238 Node* adr, 4239 uint alias_idx, 4240 Node* val, 4241 BasicType bt, 4242 bool use_precise) { 4243 // If we are writing a NULL then we need no post barrier 4244 4245 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4246 // Must be NULL 4247 const Type* t = val->bottom_type(); 4248 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4249 // No post barrier if writing NULLx 4250 return; 4251 } 4252 4253 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4254 // We can skip marks on a freshly-allocated object in Eden. 4255 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 4256 // That routine informs GC to take appropriate compensating steps, 4257 // upon a slow-path allocation, so as to make this card-mark 4258 // elision safe. 4259 return; 4260 } 4261 4262 if (use_ReduceInitialCardMarks() 4263 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4264 return; 4265 } 4266 4267 if (!use_precise) { 4268 // All card marks for a (non-array) instance are in one place: 4269 adr = obj; 4270 } 4271 // (Else it's an array (or unknown), and we want more precise card marks.) 4272 assert(adr != NULL, ""); 4273 4274 IdealKit ideal(this, true); 4275 4276 Node* tls = __ thread(); // ThreadLocalStorage 4277 4278 Node* no_base = __ top(); 4279 float likely = PROB_LIKELY(0.999); 4280 float unlikely = PROB_UNLIKELY(0.999); 4281 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val()); 4282 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val()); 4283 Node* zeroX = __ ConX(0); 4284 4285 // Get the alias_index for raw card-mark memory 4286 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4287 4288 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4289 4290 // Offsets into the thread 4291 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset()); 4292 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset()); 4293 4294 // Pointers into the thread 4295 4296 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4297 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4298 4299 // Now some values 4300 // Use ctrl to avoid hoisting these values past a safepoint, which could 4301 // potentially reset these fields in the JavaThread. 4302 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4303 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4304 4305 // Convert the store obj pointer to an int prior to doing math on it 4306 // Must use ctrl to prevent "integerized oop" existing across safepoint 4307 Node* cast = __ CastPX(__ ctrl(), adr); 4308 4309 // Divide pointer by card size 4310 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 4311 4312 // Combine card table base and card offset 4313 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4314 4315 // If we know the value being stored does it cross regions? 4316 4317 if (val != NULL) { 4318 // Does the store cause us to cross regions? 4319 4320 // Should be able to do an unsigned compare of region_size instead of 4321 // and extra shift. Do we have an unsigned compare?? 4322 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4323 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4324 4325 // if (xor_res == 0) same region so skip 4326 __ if_then(xor_res, BoolTest::ne, zeroX); { 4327 4328 // No barrier if we are storing a NULL 4329 __ if_then(val, BoolTest::ne, null(), unlikely); { 4330 4331 // Ok must mark the card if not already dirty 4332 4333 // load the original value of the card 4334 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4335 4336 __ if_then(card_val, BoolTest::ne, young_card); { 4337 sync_kit(ideal); 4338 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4339 insert_mem_bar(Op_MemBarVolatile, oop_store); 4340 __ sync_kit(this); 4341 4342 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4343 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4344 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4345 } __ end_if(); 4346 } __ end_if(); 4347 } __ end_if(); 4348 } __ end_if(); 4349 } else { 4350 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4351 // We don't need a barrier here if the destination is a newly allocated object 4352 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4353 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()). 4354 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4355 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4356 __ if_then(card_val, BoolTest::ne, young_card); { 4357 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4358 } __ end_if(); 4359 } 4360 4361 // Final sync IdealKit and GraphKit. 4362 final_sync(ideal); 4363 } 4364 #undef __ 4365 4366 4367 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4368 Node* len = load_array_length(load_String_value(ctrl, str)); 4369 Node* coder = load_String_coder(ctrl, str); 4370 // Divide length by 2 if coder is UTF16 4371 return _gvn.transform(new RShiftINode(len, coder)); 4372 } 4373 4374 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4375 int value_offset = java_lang_String::value_offset_in_bytes(); 4376 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4377 false, NULL, 0); 4378 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4379 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4380 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4381 ciTypeArrayKlass::make(T_BYTE), true, 0); 4382 int value_field_idx = C->get_alias_index(value_field_type); 4383 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4384 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4385 // String.value field is known to be @Stable. 4386 if (UseImplicitStableValues) { 4387 load = cast_array_to_stable(load, value_type); 4388 } 4389 return load; 4390 } 4391 4392 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4393 if (!CompactStrings) { 4394 return intcon(java_lang_String::CODER_UTF16); 4395 } 4396 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4397 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4398 false, NULL, 0); 4399 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4400 int coder_field_idx = C->get_alias_index(coder_field_type); 4401 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4402 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4403 } 4404 4405 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4406 int value_offset = java_lang_String::value_offset_in_bytes(); 4407 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4408 false, NULL, 0); 4409 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4410 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4411 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4412 } 4413 4414 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4415 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4416 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4417 false, NULL, 0); 4418 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4419 int coder_field_idx = C->get_alias_index(coder_field_type); 4420 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4421 value, T_BYTE, coder_field_idx, MemNode::unordered); 4422 } 4423 4424 // Capture src and dst memory state with a MergeMemNode 4425 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4426 if (src_type == dst_type) { 4427 // Types are equal, we don't need a MergeMemNode 4428 return memory(src_type); 4429 } 4430 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4431 record_for_igvn(merge); // fold it up later, if possible 4432 int src_idx = C->get_alias_index(src_type); 4433 int dst_idx = C->get_alias_index(dst_type); 4434 merge->set_memory_at(src_idx, memory(src_idx)); 4435 merge->set_memory_at(dst_idx, memory(dst_idx)); 4436 return merge; 4437 } 4438 4439 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4440 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4441 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4442 // If input and output memory types differ, capture both states to preserve 4443 // the dependency between preceding and subsequent loads/stores. 4444 // For example, the following program: 4445 // StoreB 4446 // compress_string 4447 // LoadB 4448 // has this memory graph (use->def): 4449 // LoadB -> compress_string -> CharMem 4450 // ... -> StoreB -> ByteMem 4451 // The intrinsic hides the dependency between LoadB and StoreB, causing 4452 // the load to read from memory not containing the result of the StoreB. 4453 // The correct memory graph should look like this: 4454 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4455 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4456 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4457 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4458 set_memory(res_mem, TypeAryPtr::BYTES); 4459 return str; 4460 } 4461 4462 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4463 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4464 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4465 // Capture src and dst memory (see comment in 'compress_string'). 4466 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4467 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4468 set_memory(_gvn.transform(str), dst_type); 4469 } 4470 4471 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4472 /** 4473 * int i_char = start; 4474 * for (int i_byte = 0; i_byte < count; i_byte++) { 4475 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4476 * } 4477 */ 4478 add_predicate(); 4479 RegionNode* head = new RegionNode(3); 4480 head->init_req(1, control()); 4481 gvn().set_type(head, Type::CONTROL); 4482 record_for_igvn(head); 4483 4484 Node* i_byte = new PhiNode(head, TypeInt::INT); 4485 i_byte->init_req(1, intcon(0)); 4486 gvn().set_type(i_byte, TypeInt::INT); 4487 record_for_igvn(i_byte); 4488 4489 Node* i_char = new PhiNode(head, TypeInt::INT); 4490 i_char->init_req(1, start); 4491 gvn().set_type(i_char, TypeInt::INT); 4492 record_for_igvn(i_char); 4493 4494 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4495 gvn().set_type(mem, Type::MEMORY); 4496 record_for_igvn(mem); 4497 set_control(head); 4498 set_memory(mem, TypeAryPtr::BYTES); 4499 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4500 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4501 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4502 false, false, true /* mismatched */); 4503 4504 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4505 head->init_req(2, IfTrue(iff)); 4506 mem->init_req(2, st); 4507 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4508 i_char->init_req(2, AddI(i_char, intcon(2))); 4509 4510 set_control(IfFalse(iff)); 4511 set_memory(st, TypeAryPtr::BYTES); 4512 } 4513 4514 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4515 if (!field->is_constant()) { 4516 return NULL; // Field not marked as constant. 4517 } 4518 ciInstance* holder = NULL; 4519 if (!field->is_static()) { 4520 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4521 if (const_oop != NULL && const_oop->is_instance()) { 4522 holder = const_oop->as_instance(); 4523 } 4524 } 4525 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4526 /*is_unsigned_load=*/false); 4527 if (con_type != NULL) { 4528 return makecon(con_type); 4529 } 4530 return NULL; 4531 } 4532 4533 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4534 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4535 // assumption of CCP analysis. 4536 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4537 }