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