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 52 //----------------------------GraphKit----------------------------------------- 53 // Main utility constructor. 54 GraphKit::GraphKit(JVMState* jvms) 55 : Phase(Phase::Parser), 56 _env(C->env()), 57 _gvn(*C->initial_gvn()) 58 { 59 _exceptions = jvms->map()->next_exception(); 60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 61 set_jvms(jvms); 62 } 63 64 // Private constructor for parser. 65 GraphKit::GraphKit() 66 : Phase(Phase::Parser), 67 _env(C->env()), 68 _gvn(*C->initial_gvn()) 69 { 70 _exceptions = NULL; 71 set_map(NULL); 72 debug_only(_sp = -99); 73 debug_only(set_bci(-99)); 74 } 75 76 77 78 //---------------------------clean_stack--------------------------------------- 79 // Clear away rubbish from the stack area of the JVM state. 80 // This destroys any arguments that may be waiting on the stack. 81 void GraphKit::clean_stack(int from_sp) { 82 SafePointNode* map = this->map(); 83 JVMState* jvms = this->jvms(); 84 int stk_size = jvms->stk_size(); 85 int stkoff = jvms->stkoff(); 86 Node* top = this->top(); 87 for (int i = from_sp; i < stk_size; i++) { 88 if (map->in(stkoff + i) != top) { 89 map->set_req(stkoff + i, top); 90 } 91 } 92 } 93 94 95 //--------------------------------sync_jvms----------------------------------- 96 // Make sure our current jvms agrees with our parse state. 97 JVMState* GraphKit::sync_jvms() const { 98 JVMState* jvms = this->jvms(); 99 jvms->set_bci(bci()); // Record the new bci in the JVMState 100 jvms->set_sp(sp()); // Record the new sp in the JVMState 101 assert(jvms_in_sync(), "jvms is now in sync"); 102 return jvms; 103 } 104 105 //--------------------------------sync_jvms_for_reexecute--------------------- 106 // Make sure our current jvms agrees with our parse state. This version 107 // uses the reexecute_sp for reexecuting bytecodes. 108 JVMState* GraphKit::sync_jvms_for_reexecute() { 109 JVMState* jvms = this->jvms(); 110 jvms->set_bci(bci()); // Record the new bci in the JVMState 111 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 112 return jvms; 113 } 114 115 #ifdef ASSERT 116 bool GraphKit::jvms_in_sync() const { 117 Parse* parse = is_Parse(); 118 if (parse == NULL) { 119 if (bci() != jvms()->bci()) return false; 120 if (sp() != (int)jvms()->sp()) return false; 121 return true; 122 } 123 if (jvms()->method() != parse->method()) return false; 124 if (jvms()->bci() != parse->bci()) return false; 125 int jvms_sp = jvms()->sp(); 126 if (jvms_sp != parse->sp()) return false; 127 int jvms_depth = jvms()->depth(); 128 if (jvms_depth != parse->depth()) return false; 129 return true; 130 } 131 132 // Local helper checks for special internal merge points 133 // used to accumulate and merge exception states. 134 // They are marked by the region's in(0) edge being the map itself. 135 // Such merge points must never "escape" into the parser at large, 136 // until they have been handed to gvn.transform. 137 static bool is_hidden_merge(Node* reg) { 138 if (reg == NULL) return false; 139 if (reg->is_Phi()) { 140 reg = reg->in(0); 141 if (reg == NULL) return false; 142 } 143 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 144 } 145 146 void GraphKit::verify_map() const { 147 if (map() == NULL) return; // null map is OK 148 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 149 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 150 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 151 } 152 153 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 154 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 155 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 156 } 157 #endif 158 159 //---------------------------stop_and_kill_map--------------------------------- 160 // Set _map to NULL, signalling a stop to further bytecode execution. 161 // First smash the current map's control to a constant, to mark it dead. 162 void GraphKit::stop_and_kill_map() { 163 SafePointNode* dead_map = stop(); 164 if (dead_map != NULL) { 165 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 166 assert(dead_map->is_killed(), "must be so marked"); 167 } 168 } 169 170 171 //--------------------------------stopped-------------------------------------- 172 // Tell if _map is NULL, or control is top. 173 bool GraphKit::stopped() { 174 if (map() == NULL) return true; 175 else if (control() == top()) return true; 176 else return false; 177 } 178 179 180 //-----------------------------has_ex_handler---------------------------------- 181 // Tell if this method or any caller method has exception handlers. 182 bool GraphKit::has_ex_handler() { 183 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 184 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 185 return true; 186 } 187 } 188 return false; 189 } 190 191 //------------------------------save_ex_oop------------------------------------ 192 // Save an exception without blowing stack contents or other JVM state. 193 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 194 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 195 ex_map->add_req(ex_oop); 196 debug_only(verify_exception_state(ex_map)); 197 } 198 199 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 200 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 201 Node* ex_oop = ex_map->in(ex_map->req()-1); 202 if (clear_it) ex_map->del_req(ex_map->req()-1); 203 return ex_oop; 204 } 205 206 //-----------------------------saved_ex_oop------------------------------------ 207 // Recover a saved exception from its map. 208 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 209 return common_saved_ex_oop(ex_map, false); 210 } 211 212 //--------------------------clear_saved_ex_oop--------------------------------- 213 // Erase a previously saved exception from its map. 214 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 215 return common_saved_ex_oop(ex_map, true); 216 } 217 218 #ifdef ASSERT 219 //---------------------------has_saved_ex_oop---------------------------------- 220 // Erase a previously saved exception from its map. 221 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 222 return ex_map->req() == ex_map->jvms()->endoff()+1; 223 } 224 #endif 225 226 //-------------------------make_exception_state-------------------------------- 227 // Turn the current JVM state into an exception state, appending the ex_oop. 228 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 229 sync_jvms(); 230 SafePointNode* ex_map = stop(); // do not manipulate this map any more 231 set_saved_ex_oop(ex_map, ex_oop); 232 return ex_map; 233 } 234 235 236 //--------------------------add_exception_state-------------------------------- 237 // Add an exception to my list of exceptions. 238 void GraphKit::add_exception_state(SafePointNode* ex_map) { 239 if (ex_map == NULL || ex_map->control() == top()) { 240 return; 241 } 242 #ifdef ASSERT 243 verify_exception_state(ex_map); 244 if (has_exceptions()) { 245 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 246 } 247 #endif 248 249 // If there is already an exception of exactly this type, merge with it. 250 // In particular, null-checks and other low-level exceptions common up here. 251 Node* ex_oop = saved_ex_oop(ex_map); 252 const Type* ex_type = _gvn.type(ex_oop); 253 if (ex_oop == top()) { 254 // No action needed. 255 return; 256 } 257 assert(ex_type->isa_instptr(), "exception must be an instance"); 258 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 259 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 260 // We check sp also because call bytecodes can generate exceptions 261 // both before and after arguments are popped! 262 if (ex_type2 == ex_type 263 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 264 combine_exception_states(ex_map, e2); 265 return; 266 } 267 } 268 269 // No pre-existing exception of the same type. Chain it on the list. 270 push_exception_state(ex_map); 271 } 272 273 //-----------------------add_exception_states_from----------------------------- 274 void GraphKit::add_exception_states_from(JVMState* jvms) { 275 SafePointNode* ex_map = jvms->map()->next_exception(); 276 if (ex_map != NULL) { 277 jvms->map()->set_next_exception(NULL); 278 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 279 next_map = ex_map->next_exception(); 280 ex_map->set_next_exception(NULL); 281 add_exception_state(ex_map); 282 } 283 } 284 } 285 286 //-----------------------transfer_exceptions_into_jvms------------------------- 287 JVMState* GraphKit::transfer_exceptions_into_jvms() { 288 if (map() == NULL) { 289 // We need a JVMS to carry the exceptions, but the map has gone away. 290 // Create a scratch JVMS, cloned from any of the exception states... 291 if (has_exceptions()) { 292 _map = _exceptions; 293 _map = clone_map(); 294 _map->set_next_exception(NULL); 295 clear_saved_ex_oop(_map); 296 debug_only(verify_map()); 297 } else { 298 // ...or created from scratch 299 JVMState* jvms = new (C) JVMState(_method, NULL); 300 jvms->set_bci(_bci); 301 jvms->set_sp(_sp); 302 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 303 set_jvms(jvms); 304 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 305 set_all_memory(top()); 306 while (map()->req() < jvms->endoff()) map()->add_req(top()); 307 } 308 // (This is a kludge, in case you didn't notice.) 309 set_control(top()); 310 } 311 JVMState* jvms = sync_jvms(); 312 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 313 jvms->map()->set_next_exception(_exceptions); 314 _exceptions = NULL; // done with this set of exceptions 315 return jvms; 316 } 317 318 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 319 assert(is_hidden_merge(dstphi), "must be a special merge node"); 320 assert(is_hidden_merge(srcphi), "must be a special merge node"); 321 uint limit = srcphi->req(); 322 for (uint i = PhiNode::Input; i < limit; i++) { 323 dstphi->add_req(srcphi->in(i)); 324 } 325 } 326 static inline void add_one_req(Node* dstphi, Node* src) { 327 assert(is_hidden_merge(dstphi), "must be a special merge node"); 328 assert(!is_hidden_merge(src), "must not be a special merge node"); 329 dstphi->add_req(src); 330 } 331 332 //-----------------------combine_exception_states------------------------------ 333 // This helper function combines exception states by building phis on a 334 // specially marked state-merging region. These regions and phis are 335 // untransformed, and can build up gradually. The region is marked by 336 // having a control input of its exception map, rather than NULL. Such 337 // regions do not appear except in this function, and in use_exception_state. 338 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 339 if (failing()) return; // dying anyway... 340 JVMState* ex_jvms = ex_map->_jvms; 341 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 342 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 343 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 344 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 345 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 346 assert(ex_map->req() == phi_map->req(), "matching maps"); 347 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 348 Node* hidden_merge_mark = root(); 349 Node* region = phi_map->control(); 350 MergeMemNode* phi_mem = phi_map->merged_memory(); 351 MergeMemNode* ex_mem = ex_map->merged_memory(); 352 if (region->in(0) != hidden_merge_mark) { 353 // The control input is not (yet) a specially-marked region in phi_map. 354 // Make it so, and build some phis. 355 region = new RegionNode(2); 356 _gvn.set_type(region, Type::CONTROL); 357 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 358 region->init_req(1, phi_map->control()); 359 phi_map->set_control(region); 360 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 361 record_for_igvn(io_phi); 362 _gvn.set_type(io_phi, Type::ABIO); 363 phi_map->set_i_o(io_phi); 364 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 365 Node* m = mms.memory(); 366 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 367 record_for_igvn(m_phi); 368 _gvn.set_type(m_phi, Type::MEMORY); 369 mms.set_memory(m_phi); 370 } 371 } 372 373 // Either or both of phi_map and ex_map might already be converted into phis. 374 Node* ex_control = ex_map->control(); 375 // if there is special marking on ex_map also, we add multiple edges from src 376 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 377 // how wide was the destination phi_map, originally? 378 uint orig_width = region->req(); 379 380 if (add_multiple) { 381 add_n_reqs(region, ex_control); 382 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 383 } else { 384 // ex_map has no merges, so we just add single edges everywhere 385 add_one_req(region, ex_control); 386 add_one_req(phi_map->i_o(), ex_map->i_o()); 387 } 388 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 389 if (mms.is_empty()) { 390 // get a copy of the base memory, and patch some inputs into it 391 const TypePtr* adr_type = mms.adr_type(C); 392 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 393 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 394 mms.set_memory(phi); 395 // Prepare to append interesting stuff onto the newly sliced phi: 396 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 397 } 398 // Append stuff from ex_map: 399 if (add_multiple) { 400 add_n_reqs(mms.memory(), mms.memory2()); 401 } else { 402 add_one_req(mms.memory(), mms.memory2()); 403 } 404 } 405 uint limit = ex_map->req(); 406 for (uint i = TypeFunc::Parms; i < limit; i++) { 407 // Skip everything in the JVMS after tos. (The ex_oop follows.) 408 if (i == tos) i = ex_jvms->monoff(); 409 Node* src = ex_map->in(i); 410 Node* dst = phi_map->in(i); 411 if (src != dst) { 412 PhiNode* phi; 413 if (dst->in(0) != region) { 414 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 415 record_for_igvn(phi); 416 _gvn.set_type(phi, phi->type()); 417 phi_map->set_req(i, dst); 418 // Prepare to append interesting stuff onto the new phi: 419 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 420 } else { 421 assert(dst->is_Phi(), "nobody else uses a hidden region"); 422 phi = dst->as_Phi(); 423 } 424 if (add_multiple && src->in(0) == ex_control) { 425 // Both are phis. 426 add_n_reqs(dst, src); 427 } else { 428 while (dst->req() < region->req()) add_one_req(dst, src); 429 } 430 const Type* srctype = _gvn.type(src); 431 if (phi->type() != srctype) { 432 const Type* dsttype = phi->type()->meet_speculative(srctype); 433 if (phi->type() != dsttype) { 434 phi->set_type(dsttype); 435 _gvn.set_type(phi, dsttype); 436 } 437 } 438 } 439 } 440 phi_map->merge_replaced_nodes_with(ex_map); 441 } 442 443 //--------------------------use_exception_state-------------------------------- 444 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 445 if (failing()) { stop(); return top(); } 446 Node* region = phi_map->control(); 447 Node* hidden_merge_mark = root(); 448 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 449 Node* ex_oop = clear_saved_ex_oop(phi_map); 450 if (region->in(0) == hidden_merge_mark) { 451 // Special marking for internal ex-states. Process the phis now. 452 region->set_req(0, region); // now it's an ordinary region 453 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 454 // Note: Setting the jvms also sets the bci and sp. 455 set_control(_gvn.transform(region)); 456 uint tos = jvms()->stkoff() + sp(); 457 for (uint i = 1; i < tos; i++) { 458 Node* x = phi_map->in(i); 459 if (x->in(0) == region) { 460 assert(x->is_Phi(), "expected a special phi"); 461 phi_map->set_req(i, _gvn.transform(x)); 462 } 463 } 464 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 465 Node* x = mms.memory(); 466 if (x->in(0) == region) { 467 assert(x->is_Phi(), "nobody else uses a hidden region"); 468 mms.set_memory(_gvn.transform(x)); 469 } 470 } 471 if (ex_oop->in(0) == region) { 472 assert(ex_oop->is_Phi(), "expected a special phi"); 473 ex_oop = _gvn.transform(ex_oop); 474 } 475 } else { 476 set_jvms(phi_map->jvms()); 477 } 478 479 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 480 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 481 return ex_oop; 482 } 483 484 //---------------------------------java_bc------------------------------------- 485 Bytecodes::Code GraphKit::java_bc() const { 486 ciMethod* method = this->method(); 487 int bci = this->bci(); 488 if (method != NULL && bci != InvocationEntryBci) 489 return method->java_code_at_bci(bci); 490 else 491 return Bytecodes::_illegal; 492 } 493 494 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 495 bool must_throw) { 496 // if the exception capability is set, then we will generate code 497 // to check the JavaThread.should_post_on_exceptions flag to see 498 // if we actually need to report exception events (for this 499 // thread). If we don't need to report exception events, we will 500 // take the normal fast path provided by add_exception_events. If 501 // exception event reporting is enabled for this thread, we will 502 // take the uncommon_trap in the BuildCutout below. 503 504 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 505 Node* jthread = _gvn.transform(new ThreadLocalNode()); 506 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 507 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 508 509 // Test the should_post_on_exceptions_flag vs. 0 510 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 511 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 512 513 // Branch to slow_path if should_post_on_exceptions_flag was true 514 { BuildCutout unless(this, tst, PROB_MAX); 515 // Do not try anything fancy if we're notifying the VM on every throw. 516 // Cf. case Bytecodes::_athrow in parse2.cpp. 517 uncommon_trap(reason, Deoptimization::Action_none, 518 (ciKlass*)NULL, (char*)NULL, must_throw); 519 } 520 521 } 522 523 //------------------------------builtin_throw---------------------------------- 524 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 525 bool must_throw = true; 526 527 if (env()->jvmti_can_post_on_exceptions()) { 528 // check if we must post exception events, take uncommon trap if so 529 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 530 // here if should_post_on_exceptions is false 531 // continue on with the normal codegen 532 } 533 534 // If this particular condition has not yet happened at this 535 // bytecode, then use the uncommon trap mechanism, and allow for 536 // a future recompilation if several traps occur here. 537 // If the throw is hot, try to use a more complicated inline mechanism 538 // which keeps execution inside the compiled code. 539 bool treat_throw_as_hot = false; 540 ciMethodData* md = method()->method_data(); 541 542 if (ProfileTraps) { 543 if (too_many_traps(reason)) { 544 treat_throw_as_hot = true; 545 } 546 // (If there is no MDO at all, assume it is early in 547 // execution, and that any deopts are part of the 548 // startup transient, and don't need to be remembered.) 549 550 // Also, if there is a local exception handler, treat all throws 551 // as hot if there has been at least one in this method. 552 if (C->trap_count(reason) != 0 553 && method()->method_data()->trap_count(reason) != 0 554 && has_ex_handler()) { 555 treat_throw_as_hot = true; 556 } 557 } 558 559 // If this throw happens frequently, an uncommon trap might cause 560 // a performance pothole. If there is a local exception handler, 561 // and if this particular bytecode appears to be deoptimizing often, 562 // let us handle the throw inline, with a preconstructed instance. 563 // Note: If the deopt count has blown up, the uncommon trap 564 // runtime is going to flush this nmethod, not matter what. 565 if (treat_throw_as_hot 566 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 567 // If the throw is local, we use a pre-existing instance and 568 // punt on the backtrace. This would lead to a missing backtrace 569 // (a repeat of 4292742) if the backtrace object is ever asked 570 // for its backtrace. 571 // Fixing this remaining case of 4292742 requires some flavor of 572 // escape analysis. Leave that for the future. 573 ciInstance* ex_obj = NULL; 574 switch (reason) { 575 case Deoptimization::Reason_null_check: 576 ex_obj = env()->NullPointerException_instance(); 577 break; 578 case Deoptimization::Reason_div0_check: 579 ex_obj = env()->ArithmeticException_instance(); 580 break; 581 case Deoptimization::Reason_range_check: 582 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 583 break; 584 case Deoptimization::Reason_class_check: 585 if (java_bc() == Bytecodes::_aastore) { 586 ex_obj = env()->ArrayStoreException_instance(); 587 } else { 588 ex_obj = env()->ClassCastException_instance(); 589 } 590 break; 591 default: 592 break; 593 } 594 if (failing()) { stop(); return; } // exception allocation might fail 595 if (ex_obj != NULL) { 596 // Cheat with a preallocated exception object. 597 if (C->log() != NULL) 598 C->log()->elem("hot_throw preallocated='1' reason='%s'", 599 Deoptimization::trap_reason_name(reason)); 600 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 601 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 602 603 // Clear the detail message of the preallocated exception object. 604 // Weblogic sometimes mutates the detail message of exceptions 605 // using reflection. 606 int offset = java_lang_Throwable::get_detailMessage_offset(); 607 const TypePtr* adr_typ = ex_con->add_offset(offset); 608 609 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 610 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 611 // Conservatively release stores of object references. 612 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 613 614 add_exception_state(make_exception_state(ex_node)); 615 return; 616 } 617 } 618 619 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 620 // It won't be much cheaper than bailing to the interp., since we'll 621 // have to pass up all the debug-info, and the runtime will have to 622 // create the stack trace. 623 624 // Usual case: Bail to interpreter. 625 // Reserve the right to recompile if we haven't seen anything yet. 626 627 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 628 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 629 if (treat_throw_as_hot 630 && (method()->method_data()->trap_recompiled_at(bci(), m) 631 || C->too_many_traps(reason))) { 632 // We cannot afford to take more traps here. Suffer in the interpreter. 633 if (C->log() != NULL) 634 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 635 Deoptimization::trap_reason_name(reason), 636 C->trap_count(reason)); 637 action = Deoptimization::Action_none; 638 } 639 640 // "must_throw" prunes the JVM state to include only the stack, if there 641 // are no local exception handlers. This should cut down on register 642 // allocation time and code size, by drastically reducing the number 643 // of in-edges on the call to the uncommon trap. 644 645 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 646 } 647 648 649 //----------------------------PreserveJVMState--------------------------------- 650 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 651 debug_only(kit->verify_map()); 652 _kit = kit; 653 _map = kit->map(); // preserve the map 654 _sp = kit->sp(); 655 kit->set_map(clone_map ? kit->clone_map() : NULL); 656 #ifdef ASSERT 657 _bci = kit->bci(); 658 Parse* parser = kit->is_Parse(); 659 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 660 _block = block; 661 #endif 662 } 663 PreserveJVMState::~PreserveJVMState() { 664 GraphKit* kit = _kit; 665 #ifdef ASSERT 666 assert(kit->bci() == _bci, "bci must not shift"); 667 Parse* parser = kit->is_Parse(); 668 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 669 assert(block == _block, "block must not shift"); 670 #endif 671 kit->set_map(_map); 672 kit->set_sp(_sp); 673 } 674 675 676 //-----------------------------BuildCutout------------------------------------- 677 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 678 : PreserveJVMState(kit) 679 { 680 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 681 SafePointNode* outer_map = _map; // preserved map is caller's 682 SafePointNode* inner_map = kit->map(); 683 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 684 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 685 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 686 } 687 BuildCutout::~BuildCutout() { 688 GraphKit* kit = _kit; 689 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 690 } 691 692 //---------------------------PreserveReexecuteState---------------------------- 693 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 694 assert(!kit->stopped(), "must call stopped() before"); 695 _kit = kit; 696 _sp = kit->sp(); 697 _reexecute = kit->jvms()->_reexecute; 698 } 699 PreserveReexecuteState::~PreserveReexecuteState() { 700 if (_kit->stopped()) return; 701 _kit->jvms()->_reexecute = _reexecute; 702 _kit->set_sp(_sp); 703 } 704 705 //------------------------------clone_map-------------------------------------- 706 // Implementation of PreserveJVMState 707 // 708 // Only clone_map(...) here. If this function is only used in the 709 // PreserveJVMState class we may want to get rid of this extra 710 // function eventually and do it all there. 711 712 SafePointNode* GraphKit::clone_map() { 713 if (map() == NULL) return NULL; 714 715 // Clone the memory edge first 716 Node* mem = MergeMemNode::make(map()->memory()); 717 gvn().set_type_bottom(mem); 718 719 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 720 JVMState* jvms = this->jvms(); 721 JVMState* clonejvms = jvms->clone_shallow(C); 722 clonemap->set_memory(mem); 723 clonemap->set_jvms(clonejvms); 724 clonejvms->set_map(clonemap); 725 record_for_igvn(clonemap); 726 gvn().set_type_bottom(clonemap); 727 return clonemap; 728 } 729 730 731 //-----------------------------set_map_clone----------------------------------- 732 void GraphKit::set_map_clone(SafePointNode* m) { 733 _map = m; 734 _map = clone_map(); 735 _map->set_next_exception(NULL); 736 debug_only(verify_map()); 737 } 738 739 740 //----------------------------kill_dead_locals--------------------------------- 741 // Detect any locals which are known to be dead, and force them to top. 742 void GraphKit::kill_dead_locals() { 743 // Consult the liveness information for the locals. If any 744 // of them are unused, then they can be replaced by top(). This 745 // should help register allocation time and cut down on the size 746 // of the deoptimization information. 747 748 // This call is made from many of the bytecode handling 749 // subroutines called from the Big Switch in do_one_bytecode. 750 // Every bytecode which might include a slow path is responsible 751 // for killing its dead locals. The more consistent we 752 // are about killing deads, the fewer useless phis will be 753 // constructed for them at various merge points. 754 755 // bci can be -1 (InvocationEntryBci). We return the entry 756 // liveness for the method. 757 758 if (method() == NULL || method()->code_size() == 0) { 759 // We are building a graph for a call to a native method. 760 // All locals are live. 761 return; 762 } 763 764 ResourceMark rm; 765 766 // Consult the liveness information for the locals. If any 767 // of them are unused, then they can be replaced by top(). This 768 // should help register allocation time and cut down on the size 769 // of the deoptimization information. 770 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 771 772 int len = (int)live_locals.size(); 773 assert(len <= jvms()->loc_size(), "too many live locals"); 774 for (int local = 0; local < len; local++) { 775 if (!live_locals.at(local)) { 776 set_local(local, top()); 777 } 778 } 779 } 780 781 #ifdef ASSERT 782 //-------------------------dead_locals_are_killed------------------------------ 783 // Return true if all dead locals are set to top in the map. 784 // Used to assert "clean" debug info at various points. 785 bool GraphKit::dead_locals_are_killed() { 786 if (method() == NULL || method()->code_size() == 0) { 787 // No locals need to be dead, so all is as it should be. 788 return true; 789 } 790 791 // Make sure somebody called kill_dead_locals upstream. 792 ResourceMark rm; 793 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 794 if (jvms->loc_size() == 0) continue; // no locals to consult 795 SafePointNode* map = jvms->map(); 796 ciMethod* method = jvms->method(); 797 int bci = jvms->bci(); 798 if (jvms == this->jvms()) { 799 bci = this->bci(); // it might not yet be synched 800 } 801 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 802 int len = (int)live_locals.size(); 803 if (!live_locals.is_valid() || len == 0) 804 // This method is trivial, or is poisoned by a breakpoint. 805 return true; 806 assert(len == jvms->loc_size(), "live map consistent with locals map"); 807 for (int local = 0; local < len; local++) { 808 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 809 if (PrintMiscellaneous && (Verbose || WizardMode)) { 810 tty->print_cr("Zombie local %d: ", local); 811 jvms->dump(); 812 } 813 return false; 814 } 815 } 816 } 817 return true; 818 } 819 820 #endif //ASSERT 821 822 // Helper function for enforcing certain bytecodes to reexecute if 823 // deoptimization happens 824 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 825 ciMethod* cur_method = jvms->method(); 826 int cur_bci = jvms->bci(); 827 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 828 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 829 return Interpreter::bytecode_should_reexecute(code) || 830 (is_anewarray && code == Bytecodes::_multianewarray); 831 // Reexecute _multianewarray bytecode which was replaced with 832 // sequence of [a]newarray. See Parse::do_multianewarray(). 833 // 834 // Note: interpreter should not have it set since this optimization 835 // is limited by dimensions and guarded by flag so in some cases 836 // multianewarray() runtime calls will be generated and 837 // the bytecode should not be reexecutes (stack will not be reset). 838 } else 839 return false; 840 } 841 842 // Helper function for adding JVMState and debug information to node 843 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 844 // Add the safepoint edges to the call (or other safepoint). 845 846 // Make sure dead locals are set to top. This 847 // should help register allocation time and cut down on the size 848 // of the deoptimization information. 849 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 850 851 // Walk the inline list to fill in the correct set of JVMState's 852 // Also fill in the associated edges for each JVMState. 853 854 // If the bytecode needs to be reexecuted we need to put 855 // the arguments back on the stack. 856 const bool should_reexecute = jvms()->should_reexecute(); 857 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 858 859 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 860 // undefined if the bci is different. This is normal for Parse but it 861 // should not happen for LibraryCallKit because only one bci is processed. 862 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 863 "in LibraryCallKit the reexecute bit should not change"); 864 865 // If we are guaranteed to throw, we can prune everything but the 866 // input to the current bytecode. 867 bool can_prune_locals = false; 868 uint stack_slots_not_pruned = 0; 869 int inputs = 0, depth = 0; 870 if (must_throw) { 871 assert(method() == youngest_jvms->method(), "sanity"); 872 if (compute_stack_effects(inputs, depth)) { 873 can_prune_locals = true; 874 stack_slots_not_pruned = inputs; 875 } 876 } 877 878 if (env()->should_retain_local_variables()) { 879 // At any safepoint, this method can get breakpointed, which would 880 // then require an immediate deoptimization. 881 can_prune_locals = false; // do not prune locals 882 stack_slots_not_pruned = 0; 883 } 884 885 // do not scribble on the input jvms 886 JVMState* out_jvms = youngest_jvms->clone_deep(C); 887 call->set_jvms(out_jvms); // Start jvms list for call node 888 889 // For a known set of bytecodes, the interpreter should reexecute them if 890 // deoptimization happens. We set the reexecute state for them here 891 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 892 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 893 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 894 } 895 896 // Presize the call: 897 DEBUG_ONLY(uint non_debug_edges = call->req()); 898 call->add_req_batch(top(), youngest_jvms->debug_depth()); 899 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 900 901 // Set up edges so that the call looks like this: 902 // Call [state:] ctl io mem fptr retadr 903 // [parms:] parm0 ... parmN 904 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 905 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 906 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 907 // Note that caller debug info precedes callee debug info. 908 909 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 910 uint debug_ptr = call->req(); 911 912 // Loop over the map input edges associated with jvms, add them 913 // to the call node, & reset all offsets to match call node array. 914 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 915 uint debug_end = debug_ptr; 916 uint debug_start = debug_ptr - in_jvms->debug_size(); 917 debug_ptr = debug_start; // back up the ptr 918 919 uint p = debug_start; // walks forward in [debug_start, debug_end) 920 uint j, k, l; 921 SafePointNode* in_map = in_jvms->map(); 922 out_jvms->set_map(call); 923 924 if (can_prune_locals) { 925 assert(in_jvms->method() == out_jvms->method(), "sanity"); 926 // If the current throw can reach an exception handler in this JVMS, 927 // then we must keep everything live that can reach that handler. 928 // As a quick and dirty approximation, we look for any handlers at all. 929 if (in_jvms->method()->has_exception_handlers()) { 930 can_prune_locals = false; 931 } 932 } 933 934 // Add the Locals 935 k = in_jvms->locoff(); 936 l = in_jvms->loc_size(); 937 out_jvms->set_locoff(p); 938 if (!can_prune_locals) { 939 for (j = 0; j < l; j++) 940 call->set_req(p++, in_map->in(k+j)); 941 } else { 942 p += l; // already set to top above by add_req_batch 943 } 944 945 // Add the Expression Stack 946 k = in_jvms->stkoff(); 947 l = in_jvms->sp(); 948 out_jvms->set_stkoff(p); 949 if (!can_prune_locals) { 950 for (j = 0; j < l; j++) 951 call->set_req(p++, in_map->in(k+j)); 952 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 953 // Divide stack into {S0,...,S1}, where S0 is set to top. 954 uint s1 = stack_slots_not_pruned; 955 stack_slots_not_pruned = 0; // for next iteration 956 if (s1 > l) s1 = l; 957 uint s0 = l - s1; 958 p += s0; // skip the tops preinstalled by add_req_batch 959 for (j = s0; j < l; j++) 960 call->set_req(p++, in_map->in(k+j)); 961 } else { 962 p += l; // already set to top above by add_req_batch 963 } 964 965 // Add the Monitors 966 k = in_jvms->monoff(); 967 l = in_jvms->mon_size(); 968 out_jvms->set_monoff(p); 969 for (j = 0; j < l; j++) 970 call->set_req(p++, in_map->in(k+j)); 971 972 // Copy any scalar object fields. 973 k = in_jvms->scloff(); 974 l = in_jvms->scl_size(); 975 out_jvms->set_scloff(p); 976 for (j = 0; j < l; j++) 977 call->set_req(p++, in_map->in(k+j)); 978 979 // Finish the new jvms. 980 out_jvms->set_endoff(p); 981 982 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 983 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 984 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 985 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 986 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 987 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 988 989 // Update the two tail pointers in parallel. 990 out_jvms = out_jvms->caller(); 991 in_jvms = in_jvms->caller(); 992 } 993 994 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 995 996 // Test the correctness of JVMState::debug_xxx accessors: 997 assert(call->jvms()->debug_start() == non_debug_edges, ""); 998 assert(call->jvms()->debug_end() == call->req(), ""); 999 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1000 } 1001 1002 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1003 Bytecodes::Code code = java_bc(); 1004 if (code == Bytecodes::_wide) { 1005 code = method()->java_code_at_bci(bci() + 1); 1006 } 1007 1008 BasicType rtype = T_ILLEGAL; 1009 int rsize = 0; 1010 1011 if (code != Bytecodes::_illegal) { 1012 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1013 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1014 if (rtype < T_CONFLICT) 1015 rsize = type2size[rtype]; 1016 } 1017 1018 switch (code) { 1019 case Bytecodes::_illegal: 1020 return false; 1021 1022 case Bytecodes::_ldc: 1023 case Bytecodes::_ldc_w: 1024 case Bytecodes::_ldc2_w: 1025 inputs = 0; 1026 break; 1027 1028 case Bytecodes::_dup: inputs = 1; break; 1029 case Bytecodes::_dup_x1: inputs = 2; break; 1030 case Bytecodes::_dup_x2: inputs = 3; break; 1031 case Bytecodes::_dup2: inputs = 2; break; 1032 case Bytecodes::_dup2_x1: inputs = 3; break; 1033 case Bytecodes::_dup2_x2: inputs = 4; break; 1034 case Bytecodes::_swap: inputs = 2; break; 1035 case Bytecodes::_arraylength: inputs = 1; break; 1036 1037 case Bytecodes::_getstatic: 1038 case Bytecodes::_putstatic: 1039 case Bytecodes::_getfield: 1040 case Bytecodes::_putfield: 1041 { 1042 bool ignored_will_link; 1043 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1044 int size = field->type()->size(); 1045 bool is_get = (depth >= 0), is_static = (depth & 1); 1046 inputs = (is_static ? 0 : 1); 1047 if (is_get) { 1048 depth = size - inputs; 1049 } else { 1050 inputs += size; // putxxx pops the value from the stack 1051 depth = - inputs; 1052 } 1053 } 1054 break; 1055 1056 case Bytecodes::_invokevirtual: 1057 case Bytecodes::_invokespecial: 1058 case Bytecodes::_invokestatic: 1059 case Bytecodes::_invokedynamic: 1060 case Bytecodes::_invokeinterface: 1061 { 1062 bool ignored_will_link; 1063 ciSignature* declared_signature = NULL; 1064 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1065 assert(declared_signature != NULL, "cannot be null"); 1066 inputs = declared_signature->arg_size_for_bc(code); 1067 int size = declared_signature->return_type()->size(); 1068 depth = size - inputs; 1069 } 1070 break; 1071 1072 case Bytecodes::_multianewarray: 1073 { 1074 ciBytecodeStream iter(method()); 1075 iter.reset_to_bci(bci()); 1076 iter.next(); 1077 inputs = iter.get_dimensions(); 1078 assert(rsize == 1, ""); 1079 depth = rsize - inputs; 1080 } 1081 break; 1082 1083 case Bytecodes::_ireturn: 1084 case Bytecodes::_lreturn: 1085 case Bytecodes::_freturn: 1086 case Bytecodes::_dreturn: 1087 case Bytecodes::_areturn: 1088 assert(rsize == -depth, ""); 1089 inputs = rsize; 1090 break; 1091 1092 case Bytecodes::_jsr: 1093 case Bytecodes::_jsr_w: 1094 inputs = 0; 1095 depth = 1; // S.B. depth=1, not zero 1096 break; 1097 1098 default: 1099 // bytecode produces a typed result 1100 inputs = rsize - depth; 1101 assert(inputs >= 0, ""); 1102 break; 1103 } 1104 1105 #ifdef ASSERT 1106 // spot check 1107 int outputs = depth + inputs; 1108 assert(outputs >= 0, "sanity"); 1109 switch (code) { 1110 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1111 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1112 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1113 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1114 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1115 default: break; 1116 } 1117 #endif //ASSERT 1118 1119 return true; 1120 } 1121 1122 1123 1124 //------------------------------basic_plus_adr--------------------------------- 1125 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1126 // short-circuit a common case 1127 if (offset == intcon(0)) return ptr; 1128 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1129 } 1130 1131 Node* GraphKit::ConvI2L(Node* offset) { 1132 // short-circuit a common case 1133 jint offset_con = find_int_con(offset, Type::OffsetBot); 1134 if (offset_con != Type::OffsetBot) { 1135 return longcon((jlong) offset_con); 1136 } 1137 return _gvn.transform( new ConvI2LNode(offset)); 1138 } 1139 1140 Node* GraphKit::ConvI2UL(Node* offset) { 1141 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1142 if (offset_con != (juint) Type::OffsetBot) { 1143 return longcon((julong) offset_con); 1144 } 1145 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1146 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1147 return _gvn.transform( new AndLNode(conv, mask) ); 1148 } 1149 1150 Node* GraphKit::ConvL2I(Node* offset) { 1151 // short-circuit a common case 1152 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1153 if (offset_con != (jlong)Type::OffsetBot) { 1154 return intcon((int) offset_con); 1155 } 1156 return _gvn.transform( new ConvL2INode(offset)); 1157 } 1158 1159 //-------------------------load_object_klass----------------------------------- 1160 Node* GraphKit::load_object_klass(Node* obj) { 1161 // Special-case a fresh allocation to avoid building nodes: 1162 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1163 if (akls != NULL) return akls; 1164 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1165 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1166 } 1167 1168 //-------------------------load_array_length----------------------------------- 1169 Node* GraphKit::load_array_length(Node* array) { 1170 // Special-case a fresh allocation to avoid building nodes: 1171 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1172 Node *alen; 1173 if (alloc == NULL) { 1174 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1175 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1176 } else { 1177 alen = alloc->Ideal_length(); 1178 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1179 if (ccast != alen) { 1180 alen = _gvn.transform(ccast); 1181 } 1182 } 1183 return alen; 1184 } 1185 1186 //------------------------------do_null_check---------------------------------- 1187 // Helper function to do a NULL pointer check. Returned value is 1188 // the incoming address with NULL casted away. You are allowed to use the 1189 // not-null value only if you are control dependent on the test. 1190 #ifndef PRODUCT 1191 extern int explicit_null_checks_inserted, 1192 explicit_null_checks_elided; 1193 #endif 1194 Node* GraphKit::null_check_common(Node* value, BasicType type, 1195 // optional arguments for variations: 1196 bool assert_null, 1197 Node* *null_control, 1198 bool speculative) { 1199 assert(!assert_null || null_control == NULL, "not both at once"); 1200 if (stopped()) return top(); 1201 NOT_PRODUCT(explicit_null_checks_inserted++); 1202 1203 // Construct NULL check 1204 Node *chk = NULL; 1205 switch(type) { 1206 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1207 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1208 case T_ARRAY : // fall through 1209 type = T_OBJECT; // simplify further tests 1210 case T_OBJECT : { 1211 const Type *t = _gvn.type( value ); 1212 1213 const TypeOopPtr* tp = t->isa_oopptr(); 1214 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1215 // Only for do_null_check, not any of its siblings: 1216 && !assert_null && null_control == NULL) { 1217 // Usually, any field access or invocation on an unloaded oop type 1218 // will simply fail to link, since the statically linked class is 1219 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1220 // the static class is loaded but the sharper oop type is not. 1221 // Rather than checking for this obscure case in lots of places, 1222 // we simply observe that a null check on an unloaded class 1223 // will always be followed by a nonsense operation, so we 1224 // can just issue the uncommon trap here. 1225 // Our access to the unloaded class will only be correct 1226 // after it has been loaded and initialized, which requires 1227 // a trip through the interpreter. 1228 #ifndef PRODUCT 1229 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1230 #endif 1231 uncommon_trap(Deoptimization::Reason_unloaded, 1232 Deoptimization::Action_reinterpret, 1233 tp->klass(), "!loaded"); 1234 return top(); 1235 } 1236 1237 if (assert_null) { 1238 // See if the type is contained in NULL_PTR. 1239 // If so, then the value is already null. 1240 if (t->higher_equal(TypePtr::NULL_PTR)) { 1241 NOT_PRODUCT(explicit_null_checks_elided++); 1242 return value; // Elided null assert quickly! 1243 } 1244 } else { 1245 // See if mixing in the NULL pointer changes type. 1246 // If so, then the NULL pointer was not allowed in the original 1247 // type. In other words, "value" was not-null. 1248 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1249 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1250 NOT_PRODUCT(explicit_null_checks_elided++); 1251 return value; // Elided null check quickly! 1252 } 1253 } 1254 chk = new CmpPNode( value, null() ); 1255 break; 1256 } 1257 1258 default: 1259 fatal("unexpected type: %s", type2name(type)); 1260 } 1261 assert(chk != NULL, "sanity check"); 1262 chk = _gvn.transform(chk); 1263 1264 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1265 BoolNode *btst = new BoolNode( chk, btest); 1266 Node *tst = _gvn.transform( btst ); 1267 1268 //----------- 1269 // if peephole optimizations occurred, a prior test existed. 1270 // If a prior test existed, maybe it dominates as we can avoid this test. 1271 if (tst != btst && type == T_OBJECT) { 1272 // At this point we want to scan up the CFG to see if we can 1273 // find an identical test (and so avoid this test altogether). 1274 Node *cfg = control(); 1275 int depth = 0; 1276 while( depth < 16 ) { // Limit search depth for speed 1277 if( cfg->Opcode() == Op_IfTrue && 1278 cfg->in(0)->in(1) == tst ) { 1279 // Found prior test. Use "cast_not_null" to construct an identical 1280 // CastPP (and hence hash to) as already exists for the prior test. 1281 // Return that casted value. 1282 if (assert_null) { 1283 replace_in_map(value, null()); 1284 return null(); // do not issue the redundant test 1285 } 1286 Node *oldcontrol = control(); 1287 set_control(cfg); 1288 Node *res = cast_not_null(value); 1289 set_control(oldcontrol); 1290 NOT_PRODUCT(explicit_null_checks_elided++); 1291 return res; 1292 } 1293 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1294 if (cfg == NULL) break; // Quit at region nodes 1295 depth++; 1296 } 1297 } 1298 1299 //----------- 1300 // Branch to failure if null 1301 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1302 Deoptimization::DeoptReason reason; 1303 if (assert_null) { 1304 reason = Deoptimization::reason_null_assert(speculative); 1305 } else if (type == T_OBJECT) { 1306 reason = Deoptimization::reason_null_check(speculative); 1307 } else { 1308 reason = Deoptimization::Reason_div0_check; 1309 } 1310 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1311 // ciMethodData::has_trap_at will return a conservative -1 if any 1312 // must-be-null assertion has failed. This could cause performance 1313 // problems for a method after its first do_null_assert failure. 1314 // Consider using 'Reason_class_check' instead? 1315 1316 // To cause an implicit null check, we set the not-null probability 1317 // to the maximum (PROB_MAX). For an explicit check the probability 1318 // is set to a smaller value. 1319 if (null_control != NULL || too_many_traps(reason)) { 1320 // probability is less likely 1321 ok_prob = PROB_LIKELY_MAG(3); 1322 } else if (!assert_null && 1323 (ImplicitNullCheckThreshold > 0) && 1324 method() != NULL && 1325 (method()->method_data()->trap_count(reason) 1326 >= (uint)ImplicitNullCheckThreshold)) { 1327 ok_prob = PROB_LIKELY_MAG(3); 1328 } 1329 1330 if (null_control != NULL) { 1331 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1332 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1333 set_control( _gvn.transform( new IfTrueNode(iff))); 1334 #ifndef PRODUCT 1335 if (null_true == top()) { 1336 explicit_null_checks_elided++; 1337 } 1338 #endif 1339 (*null_control) = null_true; 1340 } else { 1341 BuildCutout unless(this, tst, ok_prob); 1342 // Check for optimizer eliding test at parse time 1343 if (stopped()) { 1344 // Failure not possible; do not bother making uncommon trap. 1345 NOT_PRODUCT(explicit_null_checks_elided++); 1346 } else if (assert_null) { 1347 uncommon_trap(reason, 1348 Deoptimization::Action_make_not_entrant, 1349 NULL, "assert_null"); 1350 } else { 1351 replace_in_map(value, zerocon(type)); 1352 builtin_throw(reason); 1353 } 1354 } 1355 1356 // Must throw exception, fall-thru not possible? 1357 if (stopped()) { 1358 return top(); // No result 1359 } 1360 1361 if (assert_null) { 1362 // Cast obj to null on this path. 1363 replace_in_map(value, zerocon(type)); 1364 return zerocon(type); 1365 } 1366 1367 // Cast obj to not-null on this path, if there is no null_control. 1368 // (If there is a null_control, a non-null value may come back to haunt us.) 1369 if (type == T_OBJECT) { 1370 Node* cast = cast_not_null(value, false); 1371 if (null_control == NULL || (*null_control) == top()) 1372 replace_in_map(value, cast); 1373 value = cast; 1374 } 1375 1376 return value; 1377 } 1378 1379 1380 //------------------------------cast_not_null---------------------------------- 1381 // Cast obj to not-null on this path 1382 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1383 const Type *t = _gvn.type(obj); 1384 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1385 // Object is already not-null? 1386 if( t == t_not_null ) return obj; 1387 1388 Node *cast = new CastPPNode(obj,t_not_null); 1389 cast->init_req(0, control()); 1390 cast = _gvn.transform( cast ); 1391 1392 // Scan for instances of 'obj' in the current JVM mapping. 1393 // These instances are known to be not-null after the test. 1394 if (do_replace_in_map) 1395 replace_in_map(obj, cast); 1396 1397 return cast; // Return casted value 1398 } 1399 1400 // Sometimes in intrinsics, we implicitly know an object is not null 1401 // (there's no actual null check) so we can cast it to not null. In 1402 // the course of optimizations, the input to the cast can become null. 1403 // In that case that data path will die and we need the control path 1404 // to become dead as well to keep the graph consistent. So we have to 1405 // add a check for null for which one branch can't be taken. It uses 1406 // an Opaque4 node that will cause the check to be removed after loop 1407 // opts so the test goes away and the compiled code doesn't execute a 1408 // useless check. 1409 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1410 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1411 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1412 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1413 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1414 _gvn.set_type(iff, iff->Value(&_gvn)); 1415 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1416 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1417 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1418 C->root()->add_req(halt); 1419 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1420 set_control(if_t); 1421 return cast_not_null(value, do_replace_in_map); 1422 } 1423 1424 1425 //--------------------------replace_in_map------------------------------------- 1426 void GraphKit::replace_in_map(Node* old, Node* neww) { 1427 if (old == neww) { 1428 return; 1429 } 1430 1431 map()->replace_edge(old, neww); 1432 1433 // Note: This operation potentially replaces any edge 1434 // on the map. This includes locals, stack, and monitors 1435 // of the current (innermost) JVM state. 1436 1437 // don't let inconsistent types from profiling escape this 1438 // method 1439 1440 const Type* told = _gvn.type(old); 1441 const Type* tnew = _gvn.type(neww); 1442 1443 if (!tnew->higher_equal(told)) { 1444 return; 1445 } 1446 1447 map()->record_replaced_node(old, neww); 1448 } 1449 1450 1451 //============================================================================= 1452 //--------------------------------memory--------------------------------------- 1453 Node* GraphKit::memory(uint alias_idx) { 1454 MergeMemNode* mem = merged_memory(); 1455 Node* p = mem->memory_at(alias_idx); 1456 _gvn.set_type(p, Type::MEMORY); // must be mapped 1457 return p; 1458 } 1459 1460 //-----------------------------reset_memory------------------------------------ 1461 Node* GraphKit::reset_memory() { 1462 Node* mem = map()->memory(); 1463 // do not use this node for any more parsing! 1464 debug_only( map()->set_memory((Node*)NULL) ); 1465 return _gvn.transform( mem ); 1466 } 1467 1468 //------------------------------set_all_memory--------------------------------- 1469 void GraphKit::set_all_memory(Node* newmem) { 1470 Node* mergemem = MergeMemNode::make(newmem); 1471 gvn().set_type_bottom(mergemem); 1472 map()->set_memory(mergemem); 1473 } 1474 1475 //------------------------------set_all_memory_call---------------------------- 1476 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1477 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1478 set_all_memory(newmem); 1479 } 1480 1481 //============================================================================= 1482 // 1483 // parser factory methods for MemNodes 1484 // 1485 // These are layered on top of the factory methods in LoadNode and StoreNode, 1486 // and integrate with the parser's memory state and _gvn engine. 1487 // 1488 1489 // factory methods in "int adr_idx" 1490 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1491 int adr_idx, 1492 MemNode::MemOrd mo, 1493 LoadNode::ControlDependency control_dependency, 1494 bool require_atomic_access, 1495 bool unaligned, 1496 bool mismatched) { 1497 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1498 const TypePtr* adr_type = NULL; // debug-mode-only argument 1499 debug_only(adr_type = C->get_adr_type(adr_idx)); 1500 Node* mem = memory(adr_idx); 1501 Node* ld; 1502 if (require_atomic_access && bt == T_LONG) { 1503 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1504 } else if (require_atomic_access && bt == T_DOUBLE) { 1505 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1506 } else { 1507 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1508 } 1509 ld = _gvn.transform(ld); 1510 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1511 // Improve graph before escape analysis and boxing elimination. 1512 record_for_igvn(ld); 1513 } 1514 return ld; 1515 } 1516 1517 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1518 int adr_idx, 1519 MemNode::MemOrd mo, 1520 bool require_atomic_access, 1521 bool unaligned, 1522 bool mismatched) { 1523 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1524 const TypePtr* adr_type = NULL; 1525 debug_only(adr_type = C->get_adr_type(adr_idx)); 1526 Node *mem = memory(adr_idx); 1527 Node* st; 1528 if (require_atomic_access && bt == T_LONG) { 1529 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1530 } else if (require_atomic_access && bt == T_DOUBLE) { 1531 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1532 } else { 1533 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1534 } 1535 if (unaligned) { 1536 st->as_Store()->set_unaligned_access(); 1537 } 1538 if (mismatched) { 1539 st->as_Store()->set_mismatched_access(); 1540 } 1541 st = _gvn.transform(st); 1542 set_memory(st, adr_idx); 1543 // Back-to-back stores can only remove intermediate store with DU info 1544 // so push on worklist for optimizer. 1545 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1546 record_for_igvn(st); 1547 1548 return st; 1549 } 1550 1551 1552 void GraphKit::pre_barrier(bool do_load, 1553 Node* ctl, 1554 Node* obj, 1555 Node* adr, 1556 uint adr_idx, 1557 Node* val, 1558 const TypeOopPtr* val_type, 1559 Node* pre_val, 1560 BasicType bt) { 1561 1562 BarrierSet* bs = Universe::heap()->barrier_set(); 1563 set_control(ctl); 1564 switch (bs->kind()) { 1565 case BarrierSet::G1BarrierSet: 1566 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1567 break; 1568 1569 case BarrierSet::CardTableBarrierSet: 1570 break; 1571 1572 default : 1573 ShouldNotReachHere(); 1574 1575 } 1576 } 1577 1578 bool GraphKit::can_move_pre_barrier() const { 1579 BarrierSet* bs = Universe::heap()->barrier_set(); 1580 switch (bs->kind()) { 1581 case BarrierSet::G1BarrierSet: 1582 return true; // Can move it if no safepoint 1583 1584 case BarrierSet::CardTableBarrierSet: 1585 return true; // There is no pre-barrier 1586 1587 default : 1588 ShouldNotReachHere(); 1589 } 1590 return false; 1591 } 1592 1593 void GraphKit::post_barrier(Node* ctl, 1594 Node* store, 1595 Node* obj, 1596 Node* adr, 1597 uint adr_idx, 1598 Node* val, 1599 BasicType bt, 1600 bool use_precise) { 1601 BarrierSet* bs = Universe::heap()->barrier_set(); 1602 set_control(ctl); 1603 switch (bs->kind()) { 1604 case BarrierSet::G1BarrierSet: 1605 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1606 break; 1607 1608 case BarrierSet::CardTableBarrierSet: 1609 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1610 break; 1611 1612 default : 1613 ShouldNotReachHere(); 1614 1615 } 1616 } 1617 1618 Node* GraphKit::store_oop(Node* ctl, 1619 Node* obj, 1620 Node* adr, 1621 const TypePtr* adr_type, 1622 Node* val, 1623 const TypeOopPtr* val_type, 1624 BasicType bt, 1625 bool use_precise, 1626 MemNode::MemOrd mo, 1627 bool mismatched) { 1628 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1629 // could be delayed during Parse (for example, in adjust_map_after_if()). 1630 // Execute transformation here to avoid barrier generation in such case. 1631 if (_gvn.type(val) == TypePtr::NULL_PTR) 1632 val = _gvn.makecon(TypePtr::NULL_PTR); 1633 1634 set_control(ctl); 1635 if (stopped()) return top(); // Dead path ? 1636 1637 assert(bt == T_OBJECT, "sanity"); 1638 assert(val != NULL, "not dead path"); 1639 uint adr_idx = C->get_alias_index(adr_type); 1640 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1641 1642 pre_barrier(true /* do_load */, 1643 control(), obj, adr, adr_idx, val, val_type, 1644 NULL /* pre_val */, 1645 bt); 1646 1647 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1648 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1649 return store; 1650 } 1651 1652 // Could be an array or object we don't know at compile time (unsafe ref.) 1653 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1654 Node* obj, // containing obj 1655 Node* adr, // actual adress to store val at 1656 const TypePtr* adr_type, 1657 Node* val, 1658 BasicType bt, 1659 MemNode::MemOrd mo, 1660 bool mismatched) { 1661 Compile::AliasType* at = C->alias_type(adr_type); 1662 const TypeOopPtr* val_type = NULL; 1663 if (adr_type->isa_instptr()) { 1664 if (at->field() != NULL) { 1665 // known field. This code is a copy of the do_put_xxx logic. 1666 ciField* field = at->field(); 1667 if (!field->type()->is_loaded()) { 1668 val_type = TypeInstPtr::BOTTOM; 1669 } else { 1670 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1671 } 1672 } 1673 } else if (adr_type->isa_aryptr()) { 1674 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1675 } 1676 if (val_type == NULL) { 1677 val_type = TypeInstPtr::BOTTOM; 1678 } 1679 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1680 } 1681 1682 1683 //-------------------------array_element_address------------------------- 1684 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1685 const TypeInt* sizetype, Node* ctrl) { 1686 uint shift = exact_log2(type2aelembytes(elembt)); 1687 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1688 1689 // short-circuit a common case (saves lots of confusing waste motion) 1690 jint idx_con = find_int_con(idx, -1); 1691 if (idx_con >= 0) { 1692 intptr_t offset = header + ((intptr_t)idx_con << shift); 1693 return basic_plus_adr(ary, offset); 1694 } 1695 1696 // must be correct type for alignment purposes 1697 Node* base = basic_plus_adr(ary, header); 1698 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1699 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1700 return basic_plus_adr(ary, base, scale); 1701 } 1702 1703 //-------------------------load_array_element------------------------- 1704 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1705 const Type* elemtype = arytype->elem(); 1706 BasicType elembt = elemtype->array_element_basic_type(); 1707 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1708 if (elembt == T_NARROWOOP) { 1709 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1710 } 1711 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1712 return ld; 1713 } 1714 1715 //-------------------------set_arguments_for_java_call------------------------- 1716 // Arguments (pre-popped from the stack) are taken from the JVMS. 1717 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1718 // Add the call arguments: 1719 uint nargs = call->method()->arg_size(); 1720 for (uint i = 0; i < nargs; i++) { 1721 Node* arg = argument(i); 1722 call->init_req(i + TypeFunc::Parms, arg); 1723 } 1724 } 1725 1726 //---------------------------set_edges_for_java_call--------------------------- 1727 // Connect a newly created call into the current JVMS. 1728 // A return value node (if any) is returned from set_edges_for_java_call. 1729 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1730 1731 // Add the predefined inputs: 1732 call->init_req( TypeFunc::Control, control() ); 1733 call->init_req( TypeFunc::I_O , i_o() ); 1734 call->init_req( TypeFunc::Memory , reset_memory() ); 1735 call->init_req( TypeFunc::FramePtr, frameptr() ); 1736 call->init_req( TypeFunc::ReturnAdr, top() ); 1737 1738 add_safepoint_edges(call, must_throw); 1739 1740 Node* xcall = _gvn.transform(call); 1741 1742 if (xcall == top()) { 1743 set_control(top()); 1744 return; 1745 } 1746 assert(xcall == call, "call identity is stable"); 1747 1748 // Re-use the current map to produce the result. 1749 1750 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1751 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1752 set_all_memory_call(xcall, separate_io_proj); 1753 1754 //return xcall; // no need, caller already has it 1755 } 1756 1757 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1758 if (stopped()) return top(); // maybe the call folded up? 1759 1760 // Capture the return value, if any. 1761 Node* ret; 1762 if (call->method() == NULL || 1763 call->method()->return_type()->basic_type() == T_VOID) 1764 ret = top(); 1765 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1766 1767 // Note: Since any out-of-line call can produce an exception, 1768 // we always insert an I_O projection from the call into the result. 1769 1770 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1771 1772 if (separate_io_proj) { 1773 // The caller requested separate projections be used by the fall 1774 // through and exceptional paths, so replace the projections for 1775 // the fall through path. 1776 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1777 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1778 } 1779 return ret; 1780 } 1781 1782 //--------------------set_predefined_input_for_runtime_call-------------------- 1783 // Reading and setting the memory state is way conservative here. 1784 // The real problem is that I am not doing real Type analysis on memory, 1785 // so I cannot distinguish card mark stores from other stores. Across a GC 1786 // point the Store Barrier and the card mark memory has to agree. I cannot 1787 // have a card mark store and its barrier split across the GC point from 1788 // either above or below. Here I get that to happen by reading ALL of memory. 1789 // A better answer would be to separate out card marks from other memory. 1790 // For now, return the input memory state, so that it can be reused 1791 // after the call, if this call has restricted memory effects. 1792 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1793 // Set fixed predefined input arguments 1794 Node* memory = reset_memory(); 1795 call->init_req( TypeFunc::Control, control() ); 1796 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1797 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1798 call->init_req( TypeFunc::FramePtr, frameptr() ); 1799 call->init_req( TypeFunc::ReturnAdr, top() ); 1800 return memory; 1801 } 1802 1803 //-------------------set_predefined_output_for_runtime_call-------------------- 1804 // Set control and memory (not i_o) from the call. 1805 // If keep_mem is not NULL, use it for the output state, 1806 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1807 // If hook_mem is NULL, this call produces no memory effects at all. 1808 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1809 // then only that memory slice is taken from the call. 1810 // In the last case, we must put an appropriate memory barrier before 1811 // the call, so as to create the correct anti-dependencies on loads 1812 // preceding the call. 1813 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1814 Node* keep_mem, 1815 const TypePtr* hook_mem) { 1816 // no i/o 1817 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1818 if (keep_mem) { 1819 // First clone the existing memory state 1820 set_all_memory(keep_mem); 1821 if (hook_mem != NULL) { 1822 // Make memory for the call 1823 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1824 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1825 // We also use hook_mem to extract specific effects from arraycopy stubs. 1826 set_memory(mem, hook_mem); 1827 } 1828 // ...else the call has NO memory effects. 1829 1830 // Make sure the call advertises its memory effects precisely. 1831 // This lets us build accurate anti-dependences in gcm.cpp. 1832 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1833 "call node must be constructed correctly"); 1834 } else { 1835 assert(hook_mem == NULL, ""); 1836 // This is not a "slow path" call; all memory comes from the call. 1837 set_all_memory_call(call); 1838 } 1839 } 1840 1841 1842 // Replace the call with the current state of the kit. 1843 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1844 JVMState* ejvms = NULL; 1845 if (has_exceptions()) { 1846 ejvms = transfer_exceptions_into_jvms(); 1847 } 1848 1849 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1850 ReplacedNodes replaced_nodes_exception; 1851 Node* ex_ctl = top(); 1852 1853 SafePointNode* final_state = stop(); 1854 1855 // Find all the needed outputs of this call 1856 CallProjections callprojs; 1857 call->extract_projections(&callprojs, true); 1858 1859 Node* init_mem = call->in(TypeFunc::Memory); 1860 Node* final_mem = final_state->in(TypeFunc::Memory); 1861 Node* final_ctl = final_state->in(TypeFunc::Control); 1862 Node* final_io = final_state->in(TypeFunc::I_O); 1863 1864 // Replace all the old call edges with the edges from the inlining result 1865 if (callprojs.fallthrough_catchproj != NULL) { 1866 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1867 } 1868 if (callprojs.fallthrough_memproj != NULL) { 1869 if (final_mem->is_MergeMem()) { 1870 // Parser's exits MergeMem was not transformed but may be optimized 1871 final_mem = _gvn.transform(final_mem); 1872 } 1873 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1874 } 1875 if (callprojs.fallthrough_ioproj != NULL) { 1876 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1877 } 1878 1879 // Replace the result with the new result if it exists and is used 1880 if (callprojs.resproj != NULL && result != NULL) { 1881 C->gvn_replace_by(callprojs.resproj, result); 1882 } 1883 1884 if (ejvms == NULL) { 1885 // No exception edges to simply kill off those paths 1886 if (callprojs.catchall_catchproj != NULL) { 1887 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1888 } 1889 if (callprojs.catchall_memproj != NULL) { 1890 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1891 } 1892 if (callprojs.catchall_ioproj != NULL) { 1893 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1894 } 1895 // Replace the old exception object with top 1896 if (callprojs.exobj != NULL) { 1897 C->gvn_replace_by(callprojs.exobj, C->top()); 1898 } 1899 } else { 1900 GraphKit ekit(ejvms); 1901 1902 // Load my combined exception state into the kit, with all phis transformed: 1903 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1904 replaced_nodes_exception = ex_map->replaced_nodes(); 1905 1906 Node* ex_oop = ekit.use_exception_state(ex_map); 1907 1908 if (callprojs.catchall_catchproj != NULL) { 1909 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1910 ex_ctl = ekit.control(); 1911 } 1912 if (callprojs.catchall_memproj != NULL) { 1913 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1914 } 1915 if (callprojs.catchall_ioproj != NULL) { 1916 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1917 } 1918 1919 // Replace the old exception object with the newly created one 1920 if (callprojs.exobj != NULL) { 1921 C->gvn_replace_by(callprojs.exobj, ex_oop); 1922 } 1923 } 1924 1925 // Disconnect the call from the graph 1926 call->disconnect_inputs(NULL, C); 1927 C->gvn_replace_by(call, C->top()); 1928 1929 // Clean up any MergeMems that feed other MergeMems since the 1930 // optimizer doesn't like that. 1931 if (final_mem->is_MergeMem()) { 1932 Node_List wl; 1933 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1934 Node* m = i.get(); 1935 if (m->is_MergeMem() && !wl.contains(m)) { 1936 wl.push(m); 1937 } 1938 } 1939 while (wl.size() > 0) { 1940 _gvn.transform(wl.pop()); 1941 } 1942 } 1943 1944 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1945 replaced_nodes.apply(C, final_ctl); 1946 } 1947 if (!ex_ctl->is_top() && do_replaced_nodes) { 1948 replaced_nodes_exception.apply(C, ex_ctl); 1949 } 1950 } 1951 1952 1953 //------------------------------increment_counter------------------------------ 1954 // for statistics: increment a VM counter by 1 1955 1956 void GraphKit::increment_counter(address counter_addr) { 1957 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1958 increment_counter(adr1); 1959 } 1960 1961 void GraphKit::increment_counter(Node* counter_addr) { 1962 int adr_type = Compile::AliasIdxRaw; 1963 Node* ctrl = control(); 1964 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1965 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1966 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1967 } 1968 1969 1970 //------------------------------uncommon_trap---------------------------------- 1971 // Bail out to the interpreter in mid-method. Implemented by calling the 1972 // uncommon_trap blob. This helper function inserts a runtime call with the 1973 // right debug info. 1974 void GraphKit::uncommon_trap(int trap_request, 1975 ciKlass* klass, const char* comment, 1976 bool must_throw, 1977 bool keep_exact_action) { 1978 if (failing()) stop(); 1979 if (stopped()) return; // trap reachable? 1980 1981 // Note: If ProfileTraps is true, and if a deopt. actually 1982 // occurs here, the runtime will make sure an MDO exists. There is 1983 // no need to call method()->ensure_method_data() at this point. 1984 1985 // Set the stack pointer to the right value for reexecution: 1986 set_sp(reexecute_sp()); 1987 1988 #ifdef ASSERT 1989 if (!must_throw) { 1990 // Make sure the stack has at least enough depth to execute 1991 // the current bytecode. 1992 int inputs, ignored_depth; 1993 if (compute_stack_effects(inputs, ignored_depth)) { 1994 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1995 Bytecodes::name(java_bc()), sp(), inputs); 1996 } 1997 } 1998 #endif 1999 2000 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2001 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2002 2003 switch (action) { 2004 case Deoptimization::Action_maybe_recompile: 2005 case Deoptimization::Action_reinterpret: 2006 // Temporary fix for 6529811 to allow virtual calls to be sure they 2007 // get the chance to go from mono->bi->mega 2008 if (!keep_exact_action && 2009 Deoptimization::trap_request_index(trap_request) < 0 && 2010 too_many_recompiles(reason)) { 2011 // This BCI is causing too many recompilations. 2012 if (C->log() != NULL) { 2013 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2014 Deoptimization::trap_reason_name(reason), 2015 Deoptimization::trap_action_name(action)); 2016 } 2017 action = Deoptimization::Action_none; 2018 trap_request = Deoptimization::make_trap_request(reason, action); 2019 } else { 2020 C->set_trap_can_recompile(true); 2021 } 2022 break; 2023 case Deoptimization::Action_make_not_entrant: 2024 C->set_trap_can_recompile(true); 2025 break; 2026 case Deoptimization::Action_none: 2027 case Deoptimization::Action_make_not_compilable: 2028 break; 2029 default: 2030 #ifdef ASSERT 2031 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2032 #endif 2033 break; 2034 } 2035 2036 if (TraceOptoParse) { 2037 char buf[100]; 2038 tty->print_cr("Uncommon trap %s at bci:%d", 2039 Deoptimization::format_trap_request(buf, sizeof(buf), 2040 trap_request), bci()); 2041 } 2042 2043 CompileLog* log = C->log(); 2044 if (log != NULL) { 2045 int kid = (klass == NULL)? -1: log->identify(klass); 2046 log->begin_elem("uncommon_trap bci='%d'", bci()); 2047 char buf[100]; 2048 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2049 trap_request)); 2050 if (kid >= 0) log->print(" klass='%d'", kid); 2051 if (comment != NULL) log->print(" comment='%s'", comment); 2052 log->end_elem(); 2053 } 2054 2055 // Make sure any guarding test views this path as very unlikely 2056 Node *i0 = control()->in(0); 2057 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2058 IfNode *iff = i0->as_If(); 2059 float f = iff->_prob; // Get prob 2060 if (control()->Opcode() == Op_IfTrue) { 2061 if (f > PROB_UNLIKELY_MAG(4)) 2062 iff->_prob = PROB_MIN; 2063 } else { 2064 if (f < PROB_LIKELY_MAG(4)) 2065 iff->_prob = PROB_MAX; 2066 } 2067 } 2068 2069 // Clear out dead values from the debug info. 2070 kill_dead_locals(); 2071 2072 // Now insert the uncommon trap subroutine call 2073 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2074 const TypePtr* no_memory_effects = NULL; 2075 // Pass the index of the class to be loaded 2076 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2077 (must_throw ? RC_MUST_THROW : 0), 2078 OptoRuntime::uncommon_trap_Type(), 2079 call_addr, "uncommon_trap", no_memory_effects, 2080 intcon(trap_request)); 2081 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2082 "must extract request correctly from the graph"); 2083 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2084 2085 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2086 // The debug info is the only real input to this call. 2087 2088 // Halt-and-catch fire here. The above call should never return! 2089 HaltNode* halt = new HaltNode(control(), frameptr()); 2090 _gvn.set_type_bottom(halt); 2091 root()->add_req(halt); 2092 2093 stop_and_kill_map(); 2094 } 2095 2096 2097 //--------------------------just_allocated_object------------------------------ 2098 // Report the object that was just allocated. 2099 // It must be the case that there are no intervening safepoints. 2100 // We use this to determine if an object is so "fresh" that 2101 // it does not require card marks. 2102 Node* GraphKit::just_allocated_object(Node* current_control) { 2103 if (C->recent_alloc_ctl() == current_control) 2104 return C->recent_alloc_obj(); 2105 return NULL; 2106 } 2107 2108 2109 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2110 // (Note: TypeFunc::make has a cache that makes this fast.) 2111 const TypeFunc* tf = TypeFunc::make(dest_method); 2112 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2113 for (int j = 0; j < nargs; j++) { 2114 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2115 if( targ->basic_type() == T_DOUBLE ) { 2116 // If any parameters are doubles, they must be rounded before 2117 // the call, dstore_rounding does gvn.transform 2118 Node *arg = argument(j); 2119 arg = dstore_rounding(arg); 2120 set_argument(j, arg); 2121 } 2122 } 2123 } 2124 2125 /** 2126 * Record profiling data exact_kls for Node n with the type system so 2127 * that it can propagate it (speculation) 2128 * 2129 * @param n node that the type applies to 2130 * @param exact_kls type from profiling 2131 * @param maybe_null did profiling see null? 2132 * 2133 * @return node with improved type 2134 */ 2135 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2136 const Type* current_type = _gvn.type(n); 2137 assert(UseTypeSpeculation, "type speculation must be on"); 2138 2139 const TypePtr* speculative = current_type->speculative(); 2140 2141 // Should the klass from the profile be recorded in the speculative type? 2142 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2143 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2144 const TypeOopPtr* xtype = tklass->as_instance_type(); 2145 assert(xtype->klass_is_exact(), "Should be exact"); 2146 // Any reason to believe n is not null (from this profiling or a previous one)? 2147 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2148 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2149 // record the new speculative type's depth 2150 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2151 speculative = speculative->with_inline_depth(jvms()->depth()); 2152 } else if (current_type->would_improve_ptr(ptr_kind)) { 2153 // Profiling report that null was never seen so we can change the 2154 // speculative type to non null ptr. 2155 if (ptr_kind == ProfileAlwaysNull) { 2156 speculative = TypePtr::NULL_PTR; 2157 } else { 2158 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2159 const TypePtr* ptr = TypePtr::NOTNULL; 2160 if (speculative != NULL) { 2161 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2162 } else { 2163 speculative = ptr; 2164 } 2165 } 2166 } 2167 2168 if (speculative != current_type->speculative()) { 2169 // Build a type with a speculative type (what we think we know 2170 // about the type but will need a guard when we use it) 2171 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2172 // We're changing the type, we need a new CheckCast node to carry 2173 // the new type. The new type depends on the control: what 2174 // profiling tells us is only valid from here as far as we can 2175 // tell. 2176 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2177 cast = _gvn.transform(cast); 2178 replace_in_map(n, cast); 2179 n = cast; 2180 } 2181 2182 return n; 2183 } 2184 2185 /** 2186 * Record profiling data from receiver profiling at an invoke with the 2187 * type system so that it can propagate it (speculation) 2188 * 2189 * @param n receiver node 2190 * 2191 * @return node with improved type 2192 */ 2193 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2194 if (!UseTypeSpeculation) { 2195 return n; 2196 } 2197 ciKlass* exact_kls = profile_has_unique_klass(); 2198 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2199 if ((java_bc() == Bytecodes::_checkcast || 2200 java_bc() == Bytecodes::_instanceof || 2201 java_bc() == Bytecodes::_aastore) && 2202 method()->method_data()->is_mature()) { 2203 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2204 if (data != NULL) { 2205 if (!data->as_BitData()->null_seen()) { 2206 ptr_kind = ProfileNeverNull; 2207 } else { 2208 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2209 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2210 uint i = 0; 2211 for (; i < call->row_limit(); i++) { 2212 ciKlass* receiver = call->receiver(i); 2213 if (receiver != NULL) { 2214 break; 2215 } 2216 } 2217 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2218 } 2219 } 2220 } 2221 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2222 } 2223 2224 /** 2225 * Record profiling data from argument profiling at an invoke with the 2226 * type system so that it can propagate it (speculation) 2227 * 2228 * @param dest_method target method for the call 2229 * @param bc what invoke bytecode is this? 2230 */ 2231 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2232 if (!UseTypeSpeculation) { 2233 return; 2234 } 2235 const TypeFunc* tf = TypeFunc::make(dest_method); 2236 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2237 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2238 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2239 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2240 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2241 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2242 ciKlass* better_type = NULL; 2243 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2244 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2245 } 2246 i++; 2247 } 2248 } 2249 } 2250 2251 /** 2252 * Record profiling data from parameter profiling at an invoke with 2253 * the type system so that it can propagate it (speculation) 2254 */ 2255 void GraphKit::record_profiled_parameters_for_speculation() { 2256 if (!UseTypeSpeculation) { 2257 return; 2258 } 2259 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2260 if (_gvn.type(local(i))->isa_oopptr()) { 2261 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2262 ciKlass* better_type = NULL; 2263 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2264 record_profile_for_speculation(local(i), better_type, ptr_kind); 2265 } 2266 j++; 2267 } 2268 } 2269 } 2270 2271 /** 2272 * Record profiling data from return value profiling at an invoke with 2273 * the type system so that it can propagate it (speculation) 2274 */ 2275 void GraphKit::record_profiled_return_for_speculation() { 2276 if (!UseTypeSpeculation) { 2277 return; 2278 } 2279 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2280 ciKlass* better_type = NULL; 2281 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2282 // If profiling reports a single type for the return value, 2283 // feed it to the type system so it can propagate it as a 2284 // speculative type 2285 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2286 } 2287 } 2288 2289 void GraphKit::round_double_result(ciMethod* dest_method) { 2290 // A non-strict method may return a double value which has an extended 2291 // exponent, but this must not be visible in a caller which is 'strict' 2292 // If a strict caller invokes a non-strict callee, round a double result 2293 2294 BasicType result_type = dest_method->return_type()->basic_type(); 2295 assert( method() != NULL, "must have caller context"); 2296 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2297 // Destination method's return value is on top of stack 2298 // dstore_rounding() does gvn.transform 2299 Node *result = pop_pair(); 2300 result = dstore_rounding(result); 2301 push_pair(result); 2302 } 2303 } 2304 2305 // rounding for strict float precision conformance 2306 Node* GraphKit::precision_rounding(Node* n) { 2307 return UseStrictFP && _method->flags().is_strict() 2308 X86_ONLY(&& UseSSE == 0) && Matcher::strict_fp_requires_explicit_rounding 2309 ? _gvn.transform( new RoundFloatNode(0, n) ) 2310 : n; 2311 } 2312 2313 // rounding for strict double precision conformance 2314 Node* GraphKit::dprecision_rounding(Node *n) { 2315 return UseStrictFP && _method->flags().is_strict() 2316 X86_ONLY(&& UseSSE <= 1) && Matcher::strict_fp_requires_explicit_rounding 2317 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2318 : n; 2319 } 2320 2321 // rounding for non-strict double stores 2322 Node* GraphKit::dstore_rounding(Node* n) { 2323 return Matcher::strict_fp_requires_explicit_rounding 2324 X86_ONLY(&& UseSSE <= 1) 2325 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2326 : n; 2327 } 2328 2329 //============================================================================= 2330 // Generate a fast path/slow path idiom. Graph looks like: 2331 // [foo] indicates that 'foo' is a parameter 2332 // 2333 // [in] NULL 2334 // \ / 2335 // CmpP 2336 // Bool ne 2337 // If 2338 // / \ 2339 // True False-<2> 2340 // / | 2341 // / cast_not_null 2342 // Load | | ^ 2343 // [fast_test] | | 2344 // gvn to opt_test | | 2345 // / \ | <1> 2346 // True False | 2347 // | \\ | 2348 // [slow_call] \[fast_result] 2349 // Ctl Val \ \ 2350 // | \ \ 2351 // Catch <1> \ \ 2352 // / \ ^ \ \ 2353 // Ex No_Ex | \ \ 2354 // | \ \ | \ <2> \ 2355 // ... \ [slow_res] | | \ [null_result] 2356 // \ \--+--+--- | | 2357 // \ | / \ | / 2358 // --------Region Phi 2359 // 2360 //============================================================================= 2361 // Code is structured as a series of driver functions all called 'do_XXX' that 2362 // call a set of helper functions. Helper functions first, then drivers. 2363 2364 //------------------------------null_check_oop--------------------------------- 2365 // Null check oop. Set null-path control into Region in slot 3. 2366 // Make a cast-not-nullness use the other not-null control. Return cast. 2367 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2368 bool never_see_null, 2369 bool safe_for_replace, 2370 bool speculative) { 2371 // Initial NULL check taken path 2372 (*null_control) = top(); 2373 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2374 2375 // Generate uncommon_trap: 2376 if (never_see_null && (*null_control) != top()) { 2377 // If we see an unexpected null at a check-cast we record it and force a 2378 // recompile; the offending check-cast will be compiled to handle NULLs. 2379 // If we see more than one offending BCI, then all checkcasts in the 2380 // method will be compiled to handle NULLs. 2381 PreserveJVMState pjvms(this); 2382 set_control(*null_control); 2383 replace_in_map(value, null()); 2384 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2385 uncommon_trap(reason, 2386 Deoptimization::Action_make_not_entrant); 2387 (*null_control) = top(); // NULL path is dead 2388 } 2389 if ((*null_control) == top() && safe_for_replace) { 2390 replace_in_map(value, cast); 2391 } 2392 2393 // Cast away null-ness on the result 2394 return cast; 2395 } 2396 2397 //------------------------------opt_iff---------------------------------------- 2398 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2399 // Return slow-path control. 2400 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2401 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2402 2403 // Fast path taken; set region slot 2 2404 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2405 region->init_req(2,fast_taken); // Capture fast-control 2406 2407 // Fast path not-taken, i.e. slow path 2408 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2409 return slow_taken; 2410 } 2411 2412 //-----------------------------make_runtime_call------------------------------- 2413 Node* GraphKit::make_runtime_call(int flags, 2414 const TypeFunc* call_type, address call_addr, 2415 const char* call_name, 2416 const TypePtr* adr_type, 2417 // The following parms are all optional. 2418 // The first NULL ends the list. 2419 Node* parm0, Node* parm1, 2420 Node* parm2, Node* parm3, 2421 Node* parm4, Node* parm5, 2422 Node* parm6, Node* parm7) { 2423 // Slow-path call 2424 bool is_leaf = !(flags & RC_NO_LEAF); 2425 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2426 if (call_name == NULL) { 2427 assert(!is_leaf, "must supply name for leaf"); 2428 call_name = OptoRuntime::stub_name(call_addr); 2429 } 2430 CallNode* call; 2431 if (!is_leaf) { 2432 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2433 bci(), adr_type); 2434 } else if (flags & RC_NO_FP) { 2435 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2436 } else { 2437 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2438 } 2439 2440 // The following is similar to set_edges_for_java_call, 2441 // except that the memory effects of the call are restricted to AliasIdxRaw. 2442 2443 // Slow path call has no side-effects, uses few values 2444 bool wide_in = !(flags & RC_NARROW_MEM); 2445 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2446 2447 Node* prev_mem = NULL; 2448 if (wide_in) { 2449 prev_mem = set_predefined_input_for_runtime_call(call); 2450 } else { 2451 assert(!wide_out, "narrow in => narrow out"); 2452 Node* narrow_mem = memory(adr_type); 2453 prev_mem = reset_memory(); 2454 map()->set_memory(narrow_mem); 2455 set_predefined_input_for_runtime_call(call); 2456 } 2457 2458 // Hook each parm in order. Stop looking at the first NULL. 2459 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2460 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2461 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2462 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2463 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2464 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2465 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2466 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2467 /* close each nested if ===> */ } } } } } } } } 2468 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2469 2470 if (!is_leaf) { 2471 // Non-leaves can block and take safepoints: 2472 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2473 } 2474 // Non-leaves can throw exceptions: 2475 if (has_io) { 2476 call->set_req(TypeFunc::I_O, i_o()); 2477 } 2478 2479 if (flags & RC_UNCOMMON) { 2480 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2481 // (An "if" probability corresponds roughly to an unconditional count. 2482 // Sort of.) 2483 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2484 } 2485 2486 Node* c = _gvn.transform(call); 2487 assert(c == call, "cannot disappear"); 2488 2489 if (wide_out) { 2490 // Slow path call has full side-effects. 2491 set_predefined_output_for_runtime_call(call); 2492 } else { 2493 // Slow path call has few side-effects, and/or sets few values. 2494 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2495 } 2496 2497 if (has_io) { 2498 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2499 } 2500 return call; 2501 2502 } 2503 2504 //------------------------------merge_memory----------------------------------- 2505 // Merge memory from one path into the current memory state. 2506 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2507 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2508 Node* old_slice = mms.force_memory(); 2509 Node* new_slice = mms.memory2(); 2510 if (old_slice != new_slice) { 2511 PhiNode* phi; 2512 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2513 if (mms.is_empty()) { 2514 // clone base memory Phi's inputs for this memory slice 2515 assert(old_slice == mms.base_memory(), "sanity"); 2516 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2517 _gvn.set_type(phi, Type::MEMORY); 2518 for (uint i = 1; i < phi->req(); i++) { 2519 phi->init_req(i, old_slice->in(i)); 2520 } 2521 } else { 2522 phi = old_slice->as_Phi(); // Phi was generated already 2523 } 2524 } else { 2525 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2526 _gvn.set_type(phi, Type::MEMORY); 2527 } 2528 phi->set_req(new_path, new_slice); 2529 mms.set_memory(phi); 2530 } 2531 } 2532 } 2533 2534 //------------------------------make_slow_call_ex------------------------------ 2535 // Make the exception handler hookups for the slow call 2536 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2537 if (stopped()) return; 2538 2539 // Make a catch node with just two handlers: fall-through and catch-all 2540 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2541 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2542 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2543 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2544 2545 { PreserveJVMState pjvms(this); 2546 set_control(excp); 2547 set_i_o(i_o); 2548 2549 if (excp != top()) { 2550 if (deoptimize) { 2551 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2552 uncommon_trap(Deoptimization::Reason_unhandled, 2553 Deoptimization::Action_none); 2554 } else { 2555 // Create an exception state also. 2556 // Use an exact type if the caller has specified a specific exception. 2557 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2558 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2559 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2560 } 2561 } 2562 } 2563 2564 // Get the no-exception control from the CatchNode. 2565 set_control(norm); 2566 } 2567 2568 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2569 Node* cmp = NULL; 2570 switch(bt) { 2571 case T_INT: cmp = new CmpINode(in1, in2); break; 2572 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2573 default: fatal("unexpected comparison type %s", type2name(bt)); 2574 } 2575 gvn->transform(cmp); 2576 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2577 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2578 gvn->transform(iff); 2579 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2580 return iff; 2581 } 2582 2583 2584 //-------------------------------gen_subtype_check----------------------------- 2585 // Generate a subtyping check. Takes as input the subtype and supertype. 2586 // Returns 2 values: sets the default control() to the true path and returns 2587 // the false path. Only reads invariant memory; sets no (visible) memory. 2588 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2589 // but that's not exposed to the optimizer. This call also doesn't take in an 2590 // Object; if you wish to check an Object you need to load the Object's class 2591 // prior to coming here. 2592 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2593 Compile* C = gvn->C; 2594 2595 if ((*ctrl)->is_top()) { 2596 return C->top(); 2597 } 2598 2599 // Fast check for identical types, perhaps identical constants. 2600 // The types can even be identical non-constants, in cases 2601 // involving Array.newInstance, Object.clone, etc. 2602 if (subklass == superklass) 2603 return C->top(); // false path is dead; no test needed. 2604 2605 if (gvn->type(superklass)->singleton()) { 2606 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2607 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2608 2609 // In the common case of an exact superklass, try to fold up the 2610 // test before generating code. You may ask, why not just generate 2611 // the code and then let it fold up? The answer is that the generated 2612 // code will necessarily include null checks, which do not always 2613 // completely fold away. If they are also needless, then they turn 2614 // into a performance loss. Example: 2615 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2616 // Here, the type of 'fa' is often exact, so the store check 2617 // of fa[1]=x will fold up, without testing the nullness of x. 2618 switch (C->static_subtype_check(superk, subk)) { 2619 case Compile::SSC_always_false: 2620 { 2621 Node* always_fail = *ctrl; 2622 *ctrl = gvn->C->top(); 2623 return always_fail; 2624 } 2625 case Compile::SSC_always_true: 2626 return C->top(); 2627 case Compile::SSC_easy_test: 2628 { 2629 // Just do a direct pointer compare and be done. 2630 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2631 *ctrl = gvn->transform(new IfTrueNode(iff)); 2632 return gvn->transform(new IfFalseNode(iff)); 2633 } 2634 case Compile::SSC_full_test: 2635 break; 2636 default: 2637 ShouldNotReachHere(); 2638 } 2639 } 2640 2641 // %%% Possible further optimization: Even if the superklass is not exact, 2642 // if the subklass is the unique subtype of the superklass, the check 2643 // will always succeed. We could leave a dependency behind to ensure this. 2644 2645 // First load the super-klass's check-offset 2646 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2647 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2648 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2649 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2650 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2651 2652 // Load from the sub-klass's super-class display list, or a 1-word cache of 2653 // the secondary superclass list, or a failing value with a sentinel offset 2654 // if the super-klass is an interface or exceptionally deep in the Java 2655 // hierarchy and we have to scan the secondary superclass list the hard way. 2656 // Worst-case type is a little odd: NULL is allowed as a result (usually 2657 // klass loads can never produce a NULL). 2658 Node *chk_off_X = chk_off; 2659 #ifdef _LP64 2660 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2661 #endif 2662 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2663 // For some types like interfaces the following loadKlass is from a 1-word 2664 // cache which is mutable so can't use immutable memory. Other 2665 // types load from the super-class display table which is immutable. 2666 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2667 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2668 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2669 2670 // Compile speed common case: ARE a subtype and we canNOT fail 2671 if( superklass == nkls ) 2672 return C->top(); // false path is dead; no test needed. 2673 2674 // See if we get an immediate positive hit. Happens roughly 83% of the 2675 // time. Test to see if the value loaded just previously from the subklass 2676 // is exactly the superklass. 2677 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2678 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2679 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2680 2681 // Compile speed common case: Check for being deterministic right now. If 2682 // chk_off is a constant and not equal to cacheoff then we are NOT a 2683 // subklass. In this case we need exactly the 1 test above and we can 2684 // return those results immediately. 2685 if (!might_be_cache) { 2686 Node* not_subtype_ctrl = *ctrl; 2687 *ctrl = iftrue1; // We need exactly the 1 test above 2688 return not_subtype_ctrl; 2689 } 2690 2691 // Gather the various success & failures here 2692 RegionNode *r_ok_subtype = new RegionNode(4); 2693 gvn->record_for_igvn(r_ok_subtype); 2694 RegionNode *r_not_subtype = new RegionNode(3); 2695 gvn->record_for_igvn(r_not_subtype); 2696 2697 r_ok_subtype->init_req(1, iftrue1); 2698 2699 // Check for immediate negative hit. Happens roughly 11% of the time (which 2700 // is roughly 63% of the remaining cases). Test to see if the loaded 2701 // check-offset points into the subklass display list or the 1-element 2702 // cache. If it points to the display (and NOT the cache) and the display 2703 // missed then it's not a subtype. 2704 Node *cacheoff = gvn->intcon(cacheoff_con); 2705 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2706 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2707 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2708 2709 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2710 // No performance impact (too rare) but allows sharing of secondary arrays 2711 // which has some footprint reduction. 2712 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2713 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2714 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2715 2716 // -- Roads not taken here: -- 2717 // We could also have chosen to perform the self-check at the beginning 2718 // of this code sequence, as the assembler does. This would not pay off 2719 // the same way, since the optimizer, unlike the assembler, can perform 2720 // static type analysis to fold away many successful self-checks. 2721 // Non-foldable self checks work better here in second position, because 2722 // the initial primary superclass check subsumes a self-check for most 2723 // types. An exception would be a secondary type like array-of-interface, 2724 // which does not appear in its own primary supertype display. 2725 // Finally, we could have chosen to move the self-check into the 2726 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2727 // dependent manner. But it is worthwhile to have the check here, 2728 // where it can be perhaps be optimized. The cost in code space is 2729 // small (register compare, branch). 2730 2731 // Now do a linear scan of the secondary super-klass array. Again, no real 2732 // performance impact (too rare) but it's gotta be done. 2733 // Since the code is rarely used, there is no penalty for moving it 2734 // out of line, and it can only improve I-cache density. 2735 // The decision to inline or out-of-line this final check is platform 2736 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2737 Node* psc = gvn->transform( 2738 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2739 2740 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2741 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2742 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2743 2744 // Return false path; set default control to true path. 2745 *ctrl = gvn->transform(r_ok_subtype); 2746 return gvn->transform(r_not_subtype); 2747 } 2748 2749 // Profile-driven exact type check: 2750 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2751 float prob, 2752 Node* *casted_receiver) { 2753 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2754 Node* recv_klass = load_object_klass(receiver); 2755 Node* want_klass = makecon(tklass); 2756 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2757 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2758 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2759 set_control( _gvn.transform( new IfTrueNode (iff) )); 2760 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2761 2762 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2763 assert(recv_xtype->klass_is_exact(), ""); 2764 2765 // Subsume downstream occurrences of receiver with a cast to 2766 // recv_xtype, since now we know what the type will be. 2767 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2768 (*casted_receiver) = _gvn.transform(cast); 2769 // (User must make the replace_in_map call.) 2770 2771 return fail; 2772 } 2773 2774 2775 //------------------------------seems_never_null------------------------------- 2776 // Use null_seen information if it is available from the profile. 2777 // If we see an unexpected null at a type check we record it and force a 2778 // recompile; the offending check will be recompiled to handle NULLs. 2779 // If we see several offending BCIs, then all checks in the 2780 // method will be recompiled. 2781 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2782 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2783 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2784 if (UncommonNullCast // Cutout for this technique 2785 && obj != null() // And not the -Xcomp stupid case? 2786 && !too_many_traps(reason) 2787 ) { 2788 if (speculating) { 2789 return true; 2790 } 2791 if (data == NULL) 2792 // Edge case: no mature data. Be optimistic here. 2793 return true; 2794 // If the profile has not seen a null, assume it won't happen. 2795 assert(java_bc() == Bytecodes::_checkcast || 2796 java_bc() == Bytecodes::_instanceof || 2797 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2798 return !data->as_BitData()->null_seen(); 2799 } 2800 speculating = false; 2801 return false; 2802 } 2803 2804 //------------------------maybe_cast_profiled_receiver------------------------- 2805 // If the profile has seen exactly one type, narrow to exactly that type. 2806 // Subsequent type checks will always fold up. 2807 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2808 ciKlass* require_klass, 2809 ciKlass* spec_klass, 2810 bool safe_for_replace) { 2811 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2812 2813 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2814 2815 // Make sure we haven't already deoptimized from this tactic. 2816 if (too_many_traps(reason) || too_many_recompiles(reason)) 2817 return NULL; 2818 2819 // (No, this isn't a call, but it's enough like a virtual call 2820 // to use the same ciMethod accessor to get the profile info...) 2821 // If we have a speculative type use it instead of profiling (which 2822 // may not help us) 2823 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2824 if (exact_kls != NULL) {// no cast failures here 2825 if (require_klass == NULL || 2826 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2827 // If we narrow the type to match what the type profile sees or 2828 // the speculative type, we can then remove the rest of the 2829 // cast. 2830 // This is a win, even if the exact_kls is very specific, 2831 // because downstream operations, such as method calls, 2832 // will often benefit from the sharper type. 2833 Node* exact_obj = not_null_obj; // will get updated in place... 2834 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2835 &exact_obj); 2836 { PreserveJVMState pjvms(this); 2837 set_control(slow_ctl); 2838 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2839 } 2840 if (safe_for_replace) { 2841 replace_in_map(not_null_obj, exact_obj); 2842 } 2843 return exact_obj; 2844 } 2845 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2846 } 2847 2848 return NULL; 2849 } 2850 2851 /** 2852 * Cast obj to type and emit guard unless we had too many traps here 2853 * already 2854 * 2855 * @param obj node being casted 2856 * @param type type to cast the node to 2857 * @param not_null true if we know node cannot be null 2858 */ 2859 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2860 ciKlass* type, 2861 bool not_null) { 2862 if (stopped()) { 2863 return obj; 2864 } 2865 2866 // type == NULL if profiling tells us this object is always null 2867 if (type != NULL) { 2868 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2869 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2870 2871 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2872 !too_many_traps(class_reason) && 2873 !too_many_recompiles(class_reason)) { 2874 Node* not_null_obj = NULL; 2875 // not_null is true if we know the object is not null and 2876 // there's no need for a null check 2877 if (!not_null) { 2878 Node* null_ctl = top(); 2879 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2880 assert(null_ctl->is_top(), "no null control here"); 2881 } else { 2882 not_null_obj = obj; 2883 } 2884 2885 Node* exact_obj = not_null_obj; 2886 ciKlass* exact_kls = type; 2887 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2888 &exact_obj); 2889 { 2890 PreserveJVMState pjvms(this); 2891 set_control(slow_ctl); 2892 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2893 } 2894 replace_in_map(not_null_obj, exact_obj); 2895 obj = exact_obj; 2896 } 2897 } else { 2898 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2899 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2900 Node* exact_obj = null_assert(obj); 2901 replace_in_map(obj, exact_obj); 2902 obj = exact_obj; 2903 } 2904 } 2905 return obj; 2906 } 2907 2908 //-------------------------------gen_instanceof-------------------------------- 2909 // Generate an instance-of idiom. Used by both the instance-of bytecode 2910 // and the reflective instance-of call. 2911 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2912 kill_dead_locals(); // Benefit all the uncommon traps 2913 assert( !stopped(), "dead parse path should be checked in callers" ); 2914 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2915 "must check for not-null not-dead klass in callers"); 2916 2917 // Make the merge point 2918 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2919 RegionNode* region = new RegionNode(PATH_LIMIT); 2920 Node* phi = new PhiNode(region, TypeInt::BOOL); 2921 C->set_has_split_ifs(true); // Has chance for split-if optimization 2922 2923 ciProfileData* data = NULL; 2924 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2925 data = method()->method_data()->bci_to_data(bci()); 2926 } 2927 bool speculative_not_null = false; 2928 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2929 && seems_never_null(obj, data, speculative_not_null)); 2930 2931 // Null check; get casted pointer; set region slot 3 2932 Node* null_ctl = top(); 2933 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2934 2935 // If not_null_obj is dead, only null-path is taken 2936 if (stopped()) { // Doing instance-of on a NULL? 2937 set_control(null_ctl); 2938 return intcon(0); 2939 } 2940 region->init_req(_null_path, null_ctl); 2941 phi ->init_req(_null_path, intcon(0)); // Set null path value 2942 if (null_ctl == top()) { 2943 // Do this eagerly, so that pattern matches like is_diamond_phi 2944 // will work even during parsing. 2945 assert(_null_path == PATH_LIMIT-1, "delete last"); 2946 region->del_req(_null_path); 2947 phi ->del_req(_null_path); 2948 } 2949 2950 // Do we know the type check always succeed? 2951 bool known_statically = false; 2952 if (_gvn.type(superklass)->singleton()) { 2953 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2954 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2955 if (subk != NULL && subk->is_loaded()) { 2956 int static_res = C->static_subtype_check(superk, subk); 2957 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2958 } 2959 } 2960 2961 if (!known_statically) { 2962 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2963 // We may not have profiling here or it may not help us. If we 2964 // have a speculative type use it to perform an exact cast. 2965 ciKlass* spec_obj_type = obj_type->speculative_type(); 2966 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2967 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2968 if (stopped()) { // Profile disagrees with this path. 2969 set_control(null_ctl); // Null is the only remaining possibility. 2970 return intcon(0); 2971 } 2972 if (cast_obj != NULL) { 2973 not_null_obj = cast_obj; 2974 } 2975 } 2976 } 2977 2978 // Load the object's klass 2979 Node* obj_klass = load_object_klass(not_null_obj); 2980 2981 // Generate the subtype check 2982 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2983 2984 // Plug in the success path to the general merge in slot 1. 2985 region->init_req(_obj_path, control()); 2986 phi ->init_req(_obj_path, intcon(1)); 2987 2988 // Plug in the failing path to the general merge in slot 2. 2989 region->init_req(_fail_path, not_subtype_ctrl); 2990 phi ->init_req(_fail_path, intcon(0)); 2991 2992 // Return final merged results 2993 set_control( _gvn.transform(region) ); 2994 record_for_igvn(region); 2995 2996 // If we know the type check always succeeds then we don't use the 2997 // profiling data at this bytecode. Don't lose it, feed it to the 2998 // type system as a speculative type. 2999 if (safe_for_replace) { 3000 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3001 replace_in_map(obj, casted_obj); 3002 } 3003 3004 return _gvn.transform(phi); 3005 } 3006 3007 //-------------------------------gen_checkcast--------------------------------- 3008 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3009 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3010 // uncommon-trap paths work. Adjust stack after this call. 3011 // If failure_control is supplied and not null, it is filled in with 3012 // the control edge for the cast failure. Otherwise, an appropriate 3013 // uncommon trap or exception is thrown. 3014 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3015 Node* *failure_control) { 3016 kill_dead_locals(); // Benefit all the uncommon traps 3017 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3018 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3019 3020 // Fast cutout: Check the case that the cast is vacuously true. 3021 // This detects the common cases where the test will short-circuit 3022 // away completely. We do this before we perform the null check, 3023 // because if the test is going to turn into zero code, we don't 3024 // want a residual null check left around. (Causes a slowdown, 3025 // for example, in some objArray manipulations, such as a[i]=a[j].) 3026 if (tk->singleton()) { 3027 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3028 if (objtp != NULL && objtp->klass() != NULL) { 3029 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3030 case Compile::SSC_always_true: 3031 // If we know the type check always succeed then we don't use 3032 // the profiling data at this bytecode. Don't lose it, feed it 3033 // to the type system as a speculative type. 3034 return record_profiled_receiver_for_speculation(obj); 3035 case Compile::SSC_always_false: 3036 // It needs a null check because a null will *pass* the cast check. 3037 // A non-null value will always produce an exception. 3038 return null_assert(obj); 3039 } 3040 } 3041 } 3042 3043 ciProfileData* data = NULL; 3044 bool safe_for_replace = false; 3045 if (failure_control == NULL) { // use MDO in regular case only 3046 assert(java_bc() == Bytecodes::_aastore || 3047 java_bc() == Bytecodes::_checkcast, 3048 "interpreter profiles type checks only for these BCs"); 3049 data = method()->method_data()->bci_to_data(bci()); 3050 safe_for_replace = true; 3051 } 3052 3053 // Make the merge point 3054 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3055 RegionNode* region = new RegionNode(PATH_LIMIT); 3056 Node* phi = new PhiNode(region, toop); 3057 C->set_has_split_ifs(true); // Has chance for split-if optimization 3058 3059 // Use null-cast information if it is available 3060 bool speculative_not_null = false; 3061 bool never_see_null = ((failure_control == NULL) // regular case only 3062 && seems_never_null(obj, data, speculative_not_null)); 3063 3064 // Null check; get casted pointer; set region slot 3 3065 Node* null_ctl = top(); 3066 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3067 3068 // If not_null_obj is dead, only null-path is taken 3069 if (stopped()) { // Doing instance-of on a NULL? 3070 set_control(null_ctl); 3071 return null(); 3072 } 3073 region->init_req(_null_path, null_ctl); 3074 phi ->init_req(_null_path, null()); // Set null path value 3075 if (null_ctl == top()) { 3076 // Do this eagerly, so that pattern matches like is_diamond_phi 3077 // will work even during parsing. 3078 assert(_null_path == PATH_LIMIT-1, "delete last"); 3079 region->del_req(_null_path); 3080 phi ->del_req(_null_path); 3081 } 3082 3083 Node* cast_obj = NULL; 3084 if (tk->klass_is_exact()) { 3085 // The following optimization tries to statically cast the speculative type of the object 3086 // (for example obtained during profiling) to the type of the superklass and then do a 3087 // dynamic check that the type of the object is what we expect. To work correctly 3088 // for checkcast and aastore the type of superklass should be exact. 3089 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3090 // We may not have profiling here or it may not help us. If we have 3091 // a speculative type use it to perform an exact cast. 3092 ciKlass* spec_obj_type = obj_type->speculative_type(); 3093 if (spec_obj_type != NULL || data != NULL) { 3094 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3095 if (cast_obj != NULL) { 3096 if (failure_control != NULL) // failure is now impossible 3097 (*failure_control) = top(); 3098 // adjust the type of the phi to the exact klass: 3099 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3100 } 3101 } 3102 } 3103 3104 if (cast_obj == NULL) { 3105 // Load the object's klass 3106 Node* obj_klass = load_object_klass(not_null_obj); 3107 3108 // Generate the subtype check 3109 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3110 3111 // Plug in success path into the merge 3112 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3113 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3114 if (failure_control == NULL) { 3115 if (not_subtype_ctrl != top()) { // If failure is possible 3116 PreserveJVMState pjvms(this); 3117 set_control(not_subtype_ctrl); 3118 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3119 } 3120 } else { 3121 (*failure_control) = not_subtype_ctrl; 3122 } 3123 } 3124 3125 region->init_req(_obj_path, control()); 3126 phi ->init_req(_obj_path, cast_obj); 3127 3128 // A merge of NULL or Casted-NotNull obj 3129 Node* res = _gvn.transform(phi); 3130 3131 // Note I do NOT always 'replace_in_map(obj,result)' here. 3132 // if( tk->klass()->can_be_primary_super() ) 3133 // This means that if I successfully store an Object into an array-of-String 3134 // I 'forget' that the Object is really now known to be a String. I have to 3135 // do this because we don't have true union types for interfaces - if I store 3136 // a Baz into an array-of-Interface and then tell the optimizer it's an 3137 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3138 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3139 // replace_in_map( obj, res ); 3140 3141 // Return final merged results 3142 set_control( _gvn.transform(region) ); 3143 record_for_igvn(region); 3144 3145 return record_profiled_receiver_for_speculation(res); 3146 } 3147 3148 //------------------------------next_monitor----------------------------------- 3149 // What number should be given to the next monitor? 3150 int GraphKit::next_monitor() { 3151 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3152 int next = current + C->sync_stack_slots(); 3153 // Keep the toplevel high water mark current: 3154 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3155 return current; 3156 } 3157 3158 //------------------------------insert_mem_bar--------------------------------- 3159 // Memory barrier to avoid floating things around 3160 // The membar serves as a pinch point between both control and all memory slices. 3161 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3162 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3163 mb->init_req(TypeFunc::Control, control()); 3164 mb->init_req(TypeFunc::Memory, reset_memory()); 3165 Node* membar = _gvn.transform(mb); 3166 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3167 set_all_memory_call(membar); 3168 return membar; 3169 } 3170 3171 //-------------------------insert_mem_bar_volatile---------------------------- 3172 // Memory barrier to avoid floating things around 3173 // The membar serves as a pinch point between both control and memory(alias_idx). 3174 // If you want to make a pinch point on all memory slices, do not use this 3175 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3176 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3177 // When Parse::do_put_xxx updates a volatile field, it appends a series 3178 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3179 // The first membar is on the same memory slice as the field store opcode. 3180 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3181 // All the other membars (for other volatile slices, including AliasIdxBot, 3182 // which stands for all unknown volatile slices) are control-dependent 3183 // on the first membar. This prevents later volatile loads or stores 3184 // from sliding up past the just-emitted store. 3185 3186 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3187 mb->set_req(TypeFunc::Control,control()); 3188 if (alias_idx == Compile::AliasIdxBot) { 3189 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3190 } else { 3191 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3192 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3193 } 3194 Node* membar = _gvn.transform(mb); 3195 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3196 if (alias_idx == Compile::AliasIdxBot) { 3197 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3198 } else { 3199 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3200 } 3201 return membar; 3202 } 3203 3204 //------------------------------shared_lock------------------------------------ 3205 // Emit locking code. 3206 FastLockNode* GraphKit::shared_lock(Node* obj) { 3207 // bci is either a monitorenter bc or InvocationEntryBci 3208 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3209 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3210 3211 if( !GenerateSynchronizationCode ) 3212 return NULL; // Not locking things? 3213 if (stopped()) // Dead monitor? 3214 return NULL; 3215 3216 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3217 3218 // Box the stack location 3219 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3220 Node* mem = reset_memory(); 3221 3222 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3223 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3224 // Create the counters for this fast lock. 3225 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3226 } 3227 3228 // Create the rtm counters for this fast lock if needed. 3229 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3230 3231 // Add monitor to debug info for the slow path. If we block inside the 3232 // slow path and de-opt, we need the monitor hanging around 3233 map()->push_monitor( flock ); 3234 3235 const TypeFunc *tf = LockNode::lock_type(); 3236 LockNode *lock = new LockNode(C, tf); 3237 3238 lock->init_req( TypeFunc::Control, control() ); 3239 lock->init_req( TypeFunc::Memory , mem ); 3240 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3241 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3242 lock->init_req( TypeFunc::ReturnAdr, top() ); 3243 3244 lock->init_req(TypeFunc::Parms + 0, obj); 3245 lock->init_req(TypeFunc::Parms + 1, box); 3246 lock->init_req(TypeFunc::Parms + 2, flock); 3247 add_safepoint_edges(lock); 3248 3249 lock = _gvn.transform( lock )->as_Lock(); 3250 3251 // lock has no side-effects, sets few values 3252 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3253 3254 insert_mem_bar(Op_MemBarAcquireLock); 3255 3256 // Add this to the worklist so that the lock can be eliminated 3257 record_for_igvn(lock); 3258 3259 #ifndef PRODUCT 3260 if (PrintLockStatistics) { 3261 // Update the counter for this lock. Don't bother using an atomic 3262 // operation since we don't require absolute accuracy. 3263 lock->create_lock_counter(map()->jvms()); 3264 increment_counter(lock->counter()->addr()); 3265 } 3266 #endif 3267 3268 return flock; 3269 } 3270 3271 3272 //------------------------------shared_unlock---------------------------------- 3273 // Emit unlocking code. 3274 void GraphKit::shared_unlock(Node* box, Node* obj) { 3275 // bci is either a monitorenter bc or InvocationEntryBci 3276 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3277 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3278 3279 if( !GenerateSynchronizationCode ) 3280 return; 3281 if (stopped()) { // Dead monitor? 3282 map()->pop_monitor(); // Kill monitor from debug info 3283 return; 3284 } 3285 3286 // Memory barrier to avoid floating things down past the locked region 3287 insert_mem_bar(Op_MemBarReleaseLock); 3288 3289 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3290 UnlockNode *unlock = new UnlockNode(C, tf); 3291 #ifdef ASSERT 3292 unlock->set_dbg_jvms(sync_jvms()); 3293 #endif 3294 uint raw_idx = Compile::AliasIdxRaw; 3295 unlock->init_req( TypeFunc::Control, control() ); 3296 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3297 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3298 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3299 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3300 3301 unlock->init_req(TypeFunc::Parms + 0, obj); 3302 unlock->init_req(TypeFunc::Parms + 1, box); 3303 unlock = _gvn.transform(unlock)->as_Unlock(); 3304 3305 Node* mem = reset_memory(); 3306 3307 // unlock has no side-effects, sets few values 3308 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3309 3310 // Kill monitor from debug info 3311 map()->pop_monitor( ); 3312 } 3313 3314 //-------------------------------get_layout_helper----------------------------- 3315 // If the given klass is a constant or known to be an array, 3316 // fetch the constant layout helper value into constant_value 3317 // and return (Node*)NULL. Otherwise, load the non-constant 3318 // layout helper value, and return the node which represents it. 3319 // This two-faced routine is useful because allocation sites 3320 // almost always feature constant types. 3321 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3322 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3323 if (!StressReflectiveCode && inst_klass != NULL) { 3324 ciKlass* klass = inst_klass->klass(); 3325 bool xklass = inst_klass->klass_is_exact(); 3326 if (xklass || klass->is_array_klass()) { 3327 jint lhelper = klass->layout_helper(); 3328 if (lhelper != Klass::_lh_neutral_value) { 3329 constant_value = lhelper; 3330 return (Node*) NULL; 3331 } 3332 } 3333 } 3334 constant_value = Klass::_lh_neutral_value; // put in a known value 3335 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3336 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3337 } 3338 3339 // We just put in an allocate/initialize with a big raw-memory effect. 3340 // Hook selected additional alias categories on the initialization. 3341 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3342 MergeMemNode* init_in_merge, 3343 Node* init_out_raw) { 3344 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3345 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3346 3347 Node* prevmem = kit.memory(alias_idx); 3348 init_in_merge->set_memory_at(alias_idx, prevmem); 3349 kit.set_memory(init_out_raw, alias_idx); 3350 } 3351 3352 //---------------------------set_output_for_allocation------------------------- 3353 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3354 const TypeOopPtr* oop_type, 3355 bool deoptimize_on_exception) { 3356 int rawidx = Compile::AliasIdxRaw; 3357 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3358 add_safepoint_edges(alloc); 3359 Node* allocx = _gvn.transform(alloc); 3360 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3361 // create memory projection for i_o 3362 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3363 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3364 3365 // create a memory projection as for the normal control path 3366 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3367 set_memory(malloc, rawidx); 3368 3369 // a normal slow-call doesn't change i_o, but an allocation does 3370 // we create a separate i_o projection for the normal control path 3371 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3372 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3373 3374 // put in an initialization barrier 3375 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3376 rawoop)->as_Initialize(); 3377 assert(alloc->initialization() == init, "2-way macro link must work"); 3378 assert(init ->allocation() == alloc, "2-way macro link must work"); 3379 { 3380 // Extract memory strands which may participate in the new object's 3381 // initialization, and source them from the new InitializeNode. 3382 // This will allow us to observe initializations when they occur, 3383 // and link them properly (as a group) to the InitializeNode. 3384 assert(init->in(InitializeNode::Memory) == malloc, ""); 3385 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3386 init->set_req(InitializeNode::Memory, minit_in); 3387 record_for_igvn(minit_in); // fold it up later, if possible 3388 Node* minit_out = memory(rawidx); 3389 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3390 if (oop_type->isa_aryptr()) { 3391 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3392 int elemidx = C->get_alias_index(telemref); 3393 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3394 } else if (oop_type->isa_instptr()) { 3395 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3396 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3397 ciField* field = ik->nonstatic_field_at(i); 3398 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3399 continue; // do not bother to track really large numbers of fields 3400 // Find (or create) the alias category for this field: 3401 int fieldidx = C->alias_type(field)->index(); 3402 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3403 } 3404 } 3405 } 3406 3407 // Cast raw oop to the real thing... 3408 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3409 javaoop = _gvn.transform(javaoop); 3410 C->set_recent_alloc(control(), javaoop); 3411 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3412 3413 #ifdef ASSERT 3414 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3415 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3416 "Ideal_allocation works"); 3417 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3418 "Ideal_allocation works"); 3419 if (alloc->is_AllocateArray()) { 3420 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3421 "Ideal_allocation works"); 3422 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3423 "Ideal_allocation works"); 3424 } else { 3425 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3426 } 3427 } 3428 #endif //ASSERT 3429 3430 return javaoop; 3431 } 3432 3433 //---------------------------new_instance-------------------------------------- 3434 // This routine takes a klass_node which may be constant (for a static type) 3435 // or may be non-constant (for reflective code). It will work equally well 3436 // for either, and the graph will fold nicely if the optimizer later reduces 3437 // the type to a constant. 3438 // The optional arguments are for specialized use by intrinsics: 3439 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3440 // - If 'return_size_val', report the the total object size to the caller. 3441 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3442 Node* GraphKit::new_instance(Node* klass_node, 3443 Node* extra_slow_test, 3444 Node* *return_size_val, 3445 bool deoptimize_on_exception) { 3446 // Compute size in doublewords 3447 // The size is always an integral number of doublewords, represented 3448 // as a positive bytewise size stored in the klass's layout_helper. 3449 // The layout_helper also encodes (in a low bit) the need for a slow path. 3450 jint layout_con = Klass::_lh_neutral_value; 3451 Node* layout_val = get_layout_helper(klass_node, layout_con); 3452 int layout_is_con = (layout_val == NULL); 3453 3454 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3455 // Generate the initial go-slow test. It's either ALWAYS (return a 3456 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3457 // case) a computed value derived from the layout_helper. 3458 Node* initial_slow_test = NULL; 3459 if (layout_is_con) { 3460 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3461 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3462 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3463 } else { // reflective case 3464 // This reflective path is used by Unsafe.allocateInstance. 3465 // (It may be stress-tested by specifying StressReflectiveCode.) 3466 // Basically, we want to get into the VM is there's an illegal argument. 3467 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3468 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3469 if (extra_slow_test != intcon(0)) { 3470 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3471 } 3472 // (Macro-expander will further convert this to a Bool, if necessary.) 3473 } 3474 3475 // Find the size in bytes. This is easy; it's the layout_helper. 3476 // The size value must be valid even if the slow path is taken. 3477 Node* size = NULL; 3478 if (layout_is_con) { 3479 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3480 } else { // reflective case 3481 // This reflective path is used by clone and Unsafe.allocateInstance. 3482 size = ConvI2X(layout_val); 3483 3484 // Clear the low bits to extract layout_helper_size_in_bytes: 3485 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3486 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3487 size = _gvn.transform( new AndXNode(size, mask) ); 3488 } 3489 if (return_size_val != NULL) { 3490 (*return_size_val) = size; 3491 } 3492 3493 // This is a precise notnull oop of the klass. 3494 // (Actually, it need not be precise if this is a reflective allocation.) 3495 // It's what we cast the result to. 3496 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3497 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3498 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3499 3500 // Now generate allocation code 3501 3502 // The entire memory state is needed for slow path of the allocation 3503 // since GC and deoptimization can happened. 3504 Node *mem = reset_memory(); 3505 set_all_memory(mem); // Create new memory state 3506 3507 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3508 control(), mem, i_o(), 3509 size, klass_node, 3510 initial_slow_test); 3511 3512 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3513 } 3514 3515 //-------------------------------new_array------------------------------------- 3516 // helper for both newarray and anewarray 3517 // The 'length' parameter is (obviously) the length of the array. 3518 // See comments on new_instance for the meaning of the other arguments. 3519 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3520 Node* length, // number of array elements 3521 int nargs, // number of arguments to push back for uncommon trap 3522 Node* *return_size_val, 3523 bool deoptimize_on_exception) { 3524 jint layout_con = Klass::_lh_neutral_value; 3525 Node* layout_val = get_layout_helper(klass_node, layout_con); 3526 int layout_is_con = (layout_val == NULL); 3527 3528 if (!layout_is_con && !StressReflectiveCode && 3529 !too_many_traps(Deoptimization::Reason_class_check)) { 3530 // This is a reflective array creation site. 3531 // Optimistically assume that it is a subtype of Object[], 3532 // so that we can fold up all the address arithmetic. 3533 layout_con = Klass::array_layout_helper(T_OBJECT); 3534 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3535 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3536 { BuildCutout unless(this, bol_lh, PROB_MAX); 3537 inc_sp(nargs); 3538 uncommon_trap(Deoptimization::Reason_class_check, 3539 Deoptimization::Action_maybe_recompile); 3540 } 3541 layout_val = NULL; 3542 layout_is_con = true; 3543 } 3544 3545 // Generate the initial go-slow test. Make sure we do not overflow 3546 // if length is huge (near 2Gig) or negative! We do not need 3547 // exact double-words here, just a close approximation of needed 3548 // double-words. We can't add any offset or rounding bits, lest we 3549 // take a size -1 of bytes and make it positive. Use an unsigned 3550 // compare, so negative sizes look hugely positive. 3551 int fast_size_limit = FastAllocateSizeLimit; 3552 if (layout_is_con) { 3553 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3554 // Increase the size limit if we have exact knowledge of array type. 3555 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3556 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3557 } 3558 3559 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3560 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3561 3562 // --- Size Computation --- 3563 // array_size = round_to_heap(array_header + (length << elem_shift)); 3564 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3565 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3566 // The rounding mask is strength-reduced, if possible. 3567 int round_mask = MinObjAlignmentInBytes - 1; 3568 Node* header_size = NULL; 3569 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3570 // (T_BYTE has the weakest alignment and size restrictions...) 3571 if (layout_is_con) { 3572 int hsize = Klass::layout_helper_header_size(layout_con); 3573 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3574 BasicType etype = Klass::layout_helper_element_type(layout_con); 3575 if ((round_mask & ~right_n_bits(eshift)) == 0) 3576 round_mask = 0; // strength-reduce it if it goes away completely 3577 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3578 assert(header_size_min <= hsize, "generic minimum is smallest"); 3579 header_size_min = hsize; 3580 header_size = intcon(hsize + round_mask); 3581 } else { 3582 Node* hss = intcon(Klass::_lh_header_size_shift); 3583 Node* hsm = intcon(Klass::_lh_header_size_mask); 3584 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3585 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3586 Node* mask = intcon(round_mask); 3587 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3588 } 3589 3590 Node* elem_shift = NULL; 3591 if (layout_is_con) { 3592 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3593 if (eshift != 0) 3594 elem_shift = intcon(eshift); 3595 } else { 3596 // There is no need to mask or shift this value. 3597 // The semantics of LShiftINode include an implicit mask to 0x1F. 3598 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3599 elem_shift = layout_val; 3600 } 3601 3602 // Transition to native address size for all offset calculations: 3603 Node* lengthx = ConvI2X(length); 3604 Node* headerx = ConvI2X(header_size); 3605 #ifdef _LP64 3606 { const TypeInt* tilen = _gvn.find_int_type(length); 3607 if (tilen != NULL && tilen->_lo < 0) { 3608 // Add a manual constraint to a positive range. Cf. array_element_address. 3609 jint size_max = fast_size_limit; 3610 if (size_max > tilen->_hi) size_max = tilen->_hi; 3611 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3612 3613 // Only do a narrow I2L conversion if the range check passed. 3614 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3615 _gvn.transform(iff); 3616 RegionNode* region = new RegionNode(3); 3617 _gvn.set_type(region, Type::CONTROL); 3618 lengthx = new PhiNode(region, TypeLong::LONG); 3619 _gvn.set_type(lengthx, TypeLong::LONG); 3620 3621 // Range check passed. Use ConvI2L node with narrow type. 3622 Node* passed = IfFalse(iff); 3623 region->init_req(1, passed); 3624 // Make I2L conversion control dependent to prevent it from 3625 // floating above the range check during loop optimizations. 3626 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3627 3628 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3629 region->init_req(2, IfTrue(iff)); 3630 lengthx->init_req(2, ConvI2X(length)); 3631 3632 set_control(region); 3633 record_for_igvn(region); 3634 record_for_igvn(lengthx); 3635 } 3636 } 3637 #endif 3638 3639 // Combine header size (plus rounding) and body size. Then round down. 3640 // This computation cannot overflow, because it is used only in two 3641 // places, one where the length is sharply limited, and the other 3642 // after a successful allocation. 3643 Node* abody = lengthx; 3644 if (elem_shift != NULL) 3645 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3646 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3647 if (round_mask != 0) { 3648 Node* mask = MakeConX(~round_mask); 3649 size = _gvn.transform( new AndXNode(size, mask) ); 3650 } 3651 // else if round_mask == 0, the size computation is self-rounding 3652 3653 if (return_size_val != NULL) { 3654 // This is the size 3655 (*return_size_val) = size; 3656 } 3657 3658 // Now generate allocation code 3659 3660 // The entire memory state is needed for slow path of the allocation 3661 // since GC and deoptimization can happened. 3662 Node *mem = reset_memory(); 3663 set_all_memory(mem); // Create new memory state 3664 3665 if (initial_slow_test->is_Bool()) { 3666 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3667 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3668 } 3669 3670 // Create the AllocateArrayNode and its result projections 3671 AllocateArrayNode* alloc 3672 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3673 control(), mem, i_o(), 3674 size, klass_node, 3675 initial_slow_test, 3676 length); 3677 3678 // Cast to correct type. Note that the klass_node may be constant or not, 3679 // and in the latter case the actual array type will be inexact also. 3680 // (This happens via a non-constant argument to inline_native_newArray.) 3681 // In any case, the value of klass_node provides the desired array type. 3682 const TypeInt* length_type = _gvn.find_int_type(length); 3683 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3684 if (ary_type->isa_aryptr() && length_type != NULL) { 3685 // Try to get a better type than POS for the size 3686 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3687 } 3688 3689 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3690 3691 // Cast length on remaining path to be as narrow as possible 3692 if (map()->find_edge(length) >= 0) { 3693 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3694 if (ccast != length) { 3695 _gvn.set_type_bottom(ccast); 3696 record_for_igvn(ccast); 3697 replace_in_map(length, ccast); 3698 } 3699 } 3700 3701 return javaoop; 3702 } 3703 3704 // The following "Ideal_foo" functions are placed here because they recognize 3705 // the graph shapes created by the functions immediately above. 3706 3707 //---------------------------Ideal_allocation---------------------------------- 3708 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3709 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3710 if (ptr == NULL) { // reduce dumb test in callers 3711 return NULL; 3712 } 3713 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3714 ptr = ptr->in(1); 3715 if (ptr == NULL) return NULL; 3716 } 3717 // Return NULL for allocations with several casts: 3718 // j.l.reflect.Array.newInstance(jobject, jint) 3719 // Object.clone() 3720 // to keep more precise type from last cast. 3721 if (ptr->is_Proj()) { 3722 Node* allo = ptr->in(0); 3723 if (allo != NULL && allo->is_Allocate()) { 3724 return allo->as_Allocate(); 3725 } 3726 } 3727 // Report failure to match. 3728 return NULL; 3729 } 3730 3731 // Fancy version which also strips off an offset (and reports it to caller). 3732 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3733 intptr_t& offset) { 3734 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3735 if (base == NULL) return NULL; 3736 return Ideal_allocation(base, phase); 3737 } 3738 3739 // Trace Initialize <- Proj[Parm] <- Allocate 3740 AllocateNode* InitializeNode::allocation() { 3741 Node* rawoop = in(InitializeNode::RawAddress); 3742 if (rawoop->is_Proj()) { 3743 Node* alloc = rawoop->in(0); 3744 if (alloc->is_Allocate()) { 3745 return alloc->as_Allocate(); 3746 } 3747 } 3748 return NULL; 3749 } 3750 3751 // Trace Allocate -> Proj[Parm] -> Initialize 3752 InitializeNode* AllocateNode::initialization() { 3753 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3754 if (rawoop == NULL) return NULL; 3755 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3756 Node* init = rawoop->fast_out(i); 3757 if (init->is_Initialize()) { 3758 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3759 return init->as_Initialize(); 3760 } 3761 } 3762 return NULL; 3763 } 3764 3765 //----------------------------- loop predicates --------------------------- 3766 3767 //------------------------------add_predicate_impl---------------------------- 3768 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3769 // Too many traps seen? 3770 if (too_many_traps(reason)) { 3771 #ifdef ASSERT 3772 if (TraceLoopPredicate) { 3773 int tc = C->trap_count(reason); 3774 tty->print("too many traps=%s tcount=%d in ", 3775 Deoptimization::trap_reason_name(reason), tc); 3776 method()->print(); // which method has too many predicate traps 3777 tty->cr(); 3778 } 3779 #endif 3780 // We cannot afford to take more traps here, 3781 // do not generate predicate. 3782 return; 3783 } 3784 3785 Node *cont = _gvn.intcon(1); 3786 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3787 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3788 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3789 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3790 C->add_predicate_opaq(opq); 3791 { 3792 PreserveJVMState pjvms(this); 3793 set_control(iffalse); 3794 inc_sp(nargs); 3795 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3796 } 3797 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3798 set_control(iftrue); 3799 } 3800 3801 //------------------------------add_predicate--------------------------------- 3802 void GraphKit::add_predicate(int nargs) { 3803 if (UseLoopPredicate) { 3804 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3805 } 3806 // loop's limit check predicate should be near the loop. 3807 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3808 } 3809 3810 //----------------------------- store barriers ---------------------------- 3811 #define __ ideal. 3812 3813 bool GraphKit::use_ReduceInitialCardMarks() { 3814 BarrierSet *bs = Universe::heap()->barrier_set(); 3815 return bs->is_a(BarrierSet::CardTableBarrierSet) 3816 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers() 3817 && ReduceInitialCardMarks; 3818 } 3819 3820 void GraphKit::sync_kit(IdealKit& ideal) { 3821 set_all_memory(__ merged_memory()); 3822 set_i_o(__ i_o()); 3823 set_control(__ ctrl()); 3824 } 3825 3826 void GraphKit::final_sync(IdealKit& ideal) { 3827 // Final sync IdealKit and graphKit. 3828 sync_kit(ideal); 3829 } 3830 3831 Node* GraphKit::byte_map_base_node() { 3832 // Get base of card map 3833 jbyte* card_table_base = ci_card_table_address(); 3834 if (card_table_base != NULL) { 3835 return makecon(TypeRawPtr::make((address)card_table_base)); 3836 } else { 3837 return null(); 3838 } 3839 } 3840 3841 // vanilla/CMS post barrier 3842 // Insert a write-barrier store. This is to let generational GC work; we have 3843 // to flag all oop-stores before the next GC point. 3844 void GraphKit::write_barrier_post(Node* oop_store, 3845 Node* obj, 3846 Node* adr, 3847 uint adr_idx, 3848 Node* val, 3849 bool use_precise) { 3850 // No store check needed if we're storing a NULL or an old object 3851 // (latter case is probably a string constant). The concurrent 3852 // mark sweep garbage collector, however, needs to have all nonNull 3853 // oop updates flagged via card-marks. 3854 if (val != NULL && val->is_Con()) { 3855 // must be either an oop or NULL 3856 const Type* t = val->bottom_type(); 3857 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3858 // stores of null never (?) need barriers 3859 return; 3860 } 3861 3862 if (use_ReduceInitialCardMarks() 3863 && obj == just_allocated_object(control())) { 3864 // We can skip marks on a freshly-allocated object in Eden. 3865 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 3866 // That routine informs GC to take appropriate compensating steps, 3867 // upon a slow-path allocation, so as to make this card-mark 3868 // elision safe. 3869 return; 3870 } 3871 3872 if (!use_precise) { 3873 // All card marks for a (non-array) instance are in one place: 3874 adr = obj; 3875 } 3876 // (Else it's an array (or unknown), and we want more precise card marks.) 3877 assert(adr != NULL, ""); 3878 3879 IdealKit ideal(this, true); 3880 3881 // Convert the pointer to an int prior to doing math on it 3882 Node* cast = __ CastPX(__ ctrl(), adr); 3883 3884 // Divide by card size 3885 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableBarrierSet), 3886 "Only one we handle so far."); 3887 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 3888 3889 // Combine card table base and card offset 3890 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3891 3892 // Get the alias_index for raw card-mark memory 3893 int adr_type = Compile::AliasIdxRaw; 3894 Node* zero = __ ConI(0); // Dirty card value 3895 BasicType bt = T_BYTE; 3896 3897 if (UseConcMarkSweepGC && UseCondCardMark) { 3898 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3899 __ sync_kit(this); 3900 } 3901 3902 if (UseCondCardMark) { 3903 // The classic GC reference write barrier is typically implemented 3904 // as a store into the global card mark table. Unfortunately 3905 // unconditional stores can result in false sharing and excessive 3906 // coherence traffic as well as false transactional aborts. 3907 // UseCondCardMark enables MP "polite" conditional card mark 3908 // stores. In theory we could relax the load from ctrl() to 3909 // no_ctrl, but that doesn't buy much latitude. 3910 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3911 __ if_then(card_val, BoolTest::ne, zero); 3912 } 3913 3914 // Smash zero into card 3915 if( !UseConcMarkSweepGC ) { 3916 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3917 } else { 3918 // Specialized path for CM store barrier 3919 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3920 } 3921 3922 if (UseCondCardMark) { 3923 __ end_if(); 3924 } 3925 3926 // Final sync IdealKit and GraphKit. 3927 final_sync(ideal); 3928 } 3929 /* 3930 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3931 * required by SATB to make sure all objects live at the start of the 3932 * marking are kept alive, all reference updates need to any previous 3933 * reference stored before writing. 3934 * 3935 * If the previous value is NULL there is no need to save the old value. 3936 * References that are NULL are filtered during runtime by the barrier 3937 * code to avoid unnecessary queuing. 3938 * 3939 * However in the case of newly allocated objects it might be possible to 3940 * prove that the reference about to be overwritten is NULL during compile 3941 * time and avoid adding the barrier code completely. 3942 * 3943 * The compiler needs to determine that the object in which a field is about 3944 * to be written is newly allocated, and that no prior store to the same field 3945 * has happened since the allocation. 3946 * 3947 * Returns true if the pre-barrier can be removed 3948 */ 3949 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3950 BasicType bt, uint adr_idx) { 3951 intptr_t offset = 0; 3952 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3953 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3954 3955 if (offset == Type::OffsetBot) { 3956 return false; // cannot unalias unless there are precise offsets 3957 } 3958 3959 if (alloc == NULL) { 3960 return false; // No allocation found 3961 } 3962 3963 intptr_t size_in_bytes = type2aelembytes(bt); 3964 3965 Node* mem = memory(adr_idx); // start searching here... 3966 3967 for (int cnt = 0; cnt < 50; cnt++) { 3968 3969 if (mem->is_Store()) { 3970 3971 Node* st_adr = mem->in(MemNode::Address); 3972 intptr_t st_offset = 0; 3973 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3974 3975 if (st_base == NULL) { 3976 break; // inscrutable pointer 3977 } 3978 3979 // Break we have found a store with same base and offset as ours so break 3980 if (st_base == base && st_offset == offset) { 3981 break; 3982 } 3983 3984 if (st_offset != offset && st_offset != Type::OffsetBot) { 3985 const int MAX_STORE = BytesPerLong; 3986 if (st_offset >= offset + size_in_bytes || 3987 st_offset <= offset - MAX_STORE || 3988 st_offset <= offset - mem->as_Store()->memory_size()) { 3989 // Success: The offsets are provably independent. 3990 // (You may ask, why not just test st_offset != offset and be done? 3991 // The answer is that stores of different sizes can co-exist 3992 // in the same sequence of RawMem effects. We sometimes initialize 3993 // a whole 'tile' of array elements with a single jint or jlong.) 3994 mem = mem->in(MemNode::Memory); 3995 continue; // advance through independent store memory 3996 } 3997 } 3998 3999 if (st_base != base 4000 && MemNode::detect_ptr_independence(base, alloc, st_base, 4001 AllocateNode::Ideal_allocation(st_base, phase), 4002 phase)) { 4003 // Success: The bases are provably independent. 4004 mem = mem->in(MemNode::Memory); 4005 continue; // advance through independent store memory 4006 } 4007 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4008 4009 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4010 AllocateNode* st_alloc = st_init->allocation(); 4011 4012 // Make sure that we are looking at the same allocation site. 4013 // The alloc variable is guaranteed to not be null here from earlier check. 4014 if (alloc == st_alloc) { 4015 // Check that the initialization is storing NULL so that no previous store 4016 // has been moved up and directly write a reference 4017 Node* captured_store = st_init->find_captured_store(offset, 4018 type2aelembytes(T_OBJECT), 4019 phase); 4020 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4021 return true; 4022 } 4023 } 4024 } 4025 4026 // Unless there is an explicit 'continue', we must bail out here, 4027 // because 'mem' is an inscrutable memory state (e.g., a call). 4028 break; 4029 } 4030 4031 return false; 4032 } 4033 4034 // G1 pre/post barriers 4035 void GraphKit::g1_write_barrier_pre(bool do_load, 4036 Node* obj, 4037 Node* adr, 4038 uint alias_idx, 4039 Node* val, 4040 const TypeOopPtr* val_type, 4041 Node* pre_val, 4042 BasicType bt) { 4043 4044 // Some sanity checks 4045 // Note: val is unused in this routine. 4046 4047 if (do_load) { 4048 // We need to generate the load of the previous value 4049 assert(obj != NULL, "must have a base"); 4050 assert(adr != NULL, "where are loading from?"); 4051 assert(pre_val == NULL, "loaded already?"); 4052 assert(val_type != NULL, "need a type"); 4053 4054 if (use_ReduceInitialCardMarks() 4055 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4056 return; 4057 } 4058 4059 } else { 4060 // In this case both val_type and alias_idx are unused. 4061 assert(pre_val != NULL, "must be loaded already"); 4062 // Nothing to be done if pre_val is null. 4063 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4064 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4065 } 4066 assert(bt == T_OBJECT, "or we shouldn't be here"); 4067 4068 IdealKit ideal(this, true); 4069 4070 Node* tls = __ thread(); // ThreadLocalStorage 4071 4072 Node* no_ctrl = NULL; 4073 Node* no_base = __ top(); 4074 Node* zero = __ ConI(0); 4075 Node* zeroX = __ ConX(0); 4076 4077 float likely = PROB_LIKELY(0.999); 4078 float unlikely = PROB_UNLIKELY(0.999); 4079 4080 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4081 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4082 4083 // Offsets into the thread 4084 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4085 SATBMarkQueue::byte_offset_of_active()); 4086 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4087 SATBMarkQueue::byte_offset_of_index()); 4088 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4089 SATBMarkQueue::byte_offset_of_buf()); 4090 4091 // Now the actual pointers into the thread 4092 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4093 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4094 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4095 4096 // Now some of the values 4097 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4098 4099 // if (!marking) 4100 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4101 BasicType index_bt = TypeX_X->basic_type(); 4102 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4103 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4104 4105 if (do_load) { 4106 // load original value 4107 // alias_idx correct?? 4108 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4109 } 4110 4111 // if (pre_val != NULL) 4112 __ if_then(pre_val, BoolTest::ne, null()); { 4113 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4114 4115 // is the queue for this thread full? 4116 __ if_then(index, BoolTest::ne, zeroX, likely); { 4117 4118 // decrement the index 4119 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4120 4121 // Now get the buffer location we will log the previous value into and store it 4122 Node *log_addr = __ AddP(no_base, buffer, next_index); 4123 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4124 // update the index 4125 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4126 4127 } __ else_(); { 4128 4129 // logging buffer is full, call the runtime 4130 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4131 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4132 } __ end_if(); // (!index) 4133 } __ end_if(); // (pre_val != NULL) 4134 } __ end_if(); // (!marking) 4135 4136 // Final sync IdealKit and GraphKit. 4137 final_sync(ideal); 4138 } 4139 4140 /* 4141 * G1 similar to any GC with a Young Generation requires a way to keep track of 4142 * references from Old Generation to Young Generation to make sure all live 4143 * objects are found. G1 also requires to keep track of object references 4144 * between different regions to enable evacuation of old regions, which is done 4145 * as part of mixed collections. References are tracked in remembered sets and 4146 * is continuously updated as reference are written to with the help of the 4147 * post-barrier. 4148 * 4149 * To reduce the number of updates to the remembered set the post-barrier 4150 * filters updates to fields in objects located in the Young Generation, 4151 * the same region as the reference, when the NULL is being written or 4152 * if the card is already marked as dirty by an earlier write. 4153 * 4154 * Under certain circumstances it is possible to avoid generating the 4155 * post-barrier completely if it is possible during compile time to prove 4156 * the object is newly allocated and that no safepoint exists between the 4157 * allocation and the store. 4158 * 4159 * In the case of slow allocation the allocation code must handle the barrier 4160 * as part of the allocation in the case the allocated object is not located 4161 * in the nursery, this would happen for humongous objects. This is similar to 4162 * how CMS is required to handle this case, see the comments for the method 4163 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier. 4164 * A deferred card mark is required for these objects and handled in the above 4165 * mentioned methods. 4166 * 4167 * Returns true if the post barrier can be removed 4168 */ 4169 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4170 Node* adr) { 4171 intptr_t offset = 0; 4172 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4173 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4174 4175 if (offset == Type::OffsetBot) { 4176 return false; // cannot unalias unless there are precise offsets 4177 } 4178 4179 if (alloc == NULL) { 4180 return false; // No allocation found 4181 } 4182 4183 // Start search from Store node 4184 Node* mem = store->in(MemNode::Control); 4185 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4186 4187 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4188 AllocateNode* st_alloc = st_init->allocation(); 4189 4190 // Make sure we are looking at the same allocation 4191 if (alloc == st_alloc) { 4192 return true; 4193 } 4194 } 4195 4196 return false; 4197 } 4198 4199 // 4200 // Update the card table and add card address to the queue 4201 // 4202 void GraphKit::g1_mark_card(IdealKit& ideal, 4203 Node* card_adr, 4204 Node* oop_store, 4205 uint oop_alias_idx, 4206 Node* index, 4207 Node* index_adr, 4208 Node* buffer, 4209 const TypeFunc* tf) { 4210 4211 Node* zero = __ ConI(0); 4212 Node* zeroX = __ ConX(0); 4213 Node* no_base = __ top(); 4214 BasicType card_bt = T_BYTE; 4215 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4216 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4217 4218 // Now do the queue work 4219 __ if_then(index, BoolTest::ne, zeroX); { 4220 4221 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4222 Node* log_addr = __ AddP(no_base, buffer, next_index); 4223 4224 // Order, see storeCM. 4225 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4226 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4227 4228 } __ else_(); { 4229 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4230 } __ end_if(); 4231 4232 } 4233 4234 void GraphKit::g1_write_barrier_post(Node* oop_store, 4235 Node* obj, 4236 Node* adr, 4237 uint alias_idx, 4238 Node* val, 4239 BasicType bt, 4240 bool use_precise) { 4241 // If we are writing a NULL then we need no post barrier 4242 4243 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4244 // Must be NULL 4245 const Type* t = val->bottom_type(); 4246 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4247 // No post barrier if writing NULLx 4248 return; 4249 } 4250 4251 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4252 // We can skip marks on a freshly-allocated object in Eden. 4253 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 4254 // That routine informs GC to take appropriate compensating steps, 4255 // upon a slow-path allocation, so as to make this card-mark 4256 // elision safe. 4257 return; 4258 } 4259 4260 if (use_ReduceInitialCardMarks() 4261 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4262 return; 4263 } 4264 4265 if (!use_precise) { 4266 // All card marks for a (non-array) instance are in one place: 4267 adr = obj; 4268 } 4269 // (Else it's an array (or unknown), and we want more precise card marks.) 4270 assert(adr != NULL, ""); 4271 4272 IdealKit ideal(this, true); 4273 4274 Node* tls = __ thread(); // ThreadLocalStorage 4275 4276 Node* no_base = __ top(); 4277 float likely = PROB_LIKELY(0.999); 4278 float unlikely = PROB_UNLIKELY(0.999); 4279 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val()); 4280 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val()); 4281 Node* zeroX = __ ConX(0); 4282 4283 // Get the alias_index for raw card-mark memory 4284 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4285 4286 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4287 4288 // Offsets into the thread 4289 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4290 DirtyCardQueue::byte_offset_of_index()); 4291 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4292 DirtyCardQueue::byte_offset_of_buf()); 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 }