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::_putfield: 903 { 904 bool ignored_will_link; 905 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 906 int size = field->type()->size(); 907 bool is_get = (depth >= 0), is_static = (depth & 1); 908 inputs = (is_static ? 0 : 1); 909 if (is_get) { 910 depth = size - inputs; 911 } else { 912 inputs += size; // putxxx pops the value from the stack 913 depth = - inputs; 914 } 915 } 916 break; 917 918 case Bytecodes::_invokevirtual: 919 case Bytecodes::_invokespecial: 920 case Bytecodes::_invokestatic: 921 case Bytecodes::_invokedynamic: 922 case Bytecodes::_invokeinterface: 923 { 924 bool ignored_will_link; 925 ciSignature* declared_signature = NULL; 926 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 927 assert(declared_signature != NULL, "cannot be null"); 928 inputs = declared_signature->arg_size_for_bc(code); 929 int size = declared_signature->return_type()->size(); 930 depth = size - inputs; 931 } 932 break; 933 934 case Bytecodes::_multianewarray: 935 { 936 ciBytecodeStream iter(method()); 937 iter.reset_to_bci(bci()); 938 iter.next(); 939 inputs = iter.get_dimensions(); 940 assert(rsize == 1, ""); 941 depth = rsize - inputs; 942 } 943 break; 944 945 case Bytecodes::_vwithfield: { 946 bool ignored_will_link; 947 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 948 int size = field->type()->size(); 949 inputs = size+1; 950 depth = rsize - inputs; 951 break; 952 } 953 954 case Bytecodes::_ireturn: 955 case Bytecodes::_lreturn: 956 case Bytecodes::_freturn: 957 case Bytecodes::_dreturn: 958 case Bytecodes::_areturn: 959 case Bytecodes::_vreturn: 960 assert(rsize == -depth, ""); 961 inputs = rsize; 962 break; 963 964 case Bytecodes::_jsr: 965 case Bytecodes::_jsr_w: 966 inputs = 0; 967 depth = 1; // S.B. depth=1, not zero 968 break; 969 970 default: 971 // bytecode produces a typed result 972 inputs = rsize - depth; 973 assert(inputs >= 0, ""); 974 break; 975 } 976 977 #ifdef ASSERT 978 // spot check 979 int outputs = depth + inputs; 980 assert(outputs >= 0, "sanity"); 981 switch (code) { 982 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 983 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 984 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 985 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 986 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 987 } 988 #endif //ASSERT 989 990 return true; 991 } 992 993 994 995 //------------------------------basic_plus_adr--------------------------------- 996 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 997 // short-circuit a common case 998 if (offset == intcon(0)) return ptr; 999 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1000 } 1001 1002 Node* GraphKit::ConvI2L(Node* offset) { 1003 // short-circuit a common case 1004 jint offset_con = find_int_con(offset, Type::OffsetBot); 1005 if (offset_con != Type::OffsetBot) { 1006 return longcon((jlong) offset_con); 1007 } 1008 return _gvn.transform( new ConvI2LNode(offset)); 1009 } 1010 1011 Node* GraphKit::ConvI2UL(Node* offset) { 1012 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1013 if (offset_con != (juint) Type::OffsetBot) { 1014 return longcon((julong) offset_con); 1015 } 1016 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1017 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1018 return _gvn.transform( new AndLNode(conv, mask) ); 1019 } 1020 1021 Node* GraphKit::ConvL2I(Node* offset) { 1022 // short-circuit a common case 1023 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1024 if (offset_con != (jlong)Type::OffsetBot) { 1025 return intcon((int) offset_con); 1026 } 1027 return _gvn.transform( new ConvL2INode(offset)); 1028 } 1029 1030 //-------------------------load_object_klass----------------------------------- 1031 Node* GraphKit::load_object_klass(Node* obj) { 1032 // Special-case a fresh allocation to avoid building nodes: 1033 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1034 if (akls != NULL) return akls; 1035 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1036 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1037 } 1038 1039 //-------------------------load_array_length----------------------------------- 1040 Node* GraphKit::load_array_length(Node* array) { 1041 // Special-case a fresh allocation to avoid building nodes: 1042 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1043 Node *alen; 1044 if (alloc == NULL) { 1045 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1046 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1047 } else { 1048 alen = alloc->Ideal_length(); 1049 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1050 if (ccast != alen) { 1051 alen = _gvn.transform(ccast); 1052 } 1053 } 1054 return alen; 1055 } 1056 1057 //------------------------------do_null_check---------------------------------- 1058 // Helper function to do a NULL pointer check. Returned value is 1059 // the incoming address with NULL casted away. You are allowed to use the 1060 // not-null value only if you are control dependent on the test. 1061 #ifndef PRODUCT 1062 extern int explicit_null_checks_inserted, 1063 explicit_null_checks_elided; 1064 #endif 1065 Node* GraphKit::null_check_common(Node* value, BasicType type, 1066 // optional arguments for variations: 1067 bool assert_null, 1068 Node* *null_control, 1069 bool speculative) { 1070 assert(!assert_null || null_control == NULL, "not both at once"); 1071 if (stopped()) return top(); 1072 NOT_PRODUCT(explicit_null_checks_inserted++); 1073 1074 // Construct NULL check 1075 Node *chk = NULL; 1076 switch(type) { 1077 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1078 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1079 case T_VALUETYPE : // fall through 1080 case T_ARRAY : // fall through 1081 type = T_OBJECT; // simplify further tests 1082 case T_OBJECT : { 1083 const Type *t = _gvn.type( value ); 1084 1085 const TypeOopPtr* tp = t->isa_oopptr(); 1086 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1087 // Only for do_null_check, not any of its siblings: 1088 && !assert_null && null_control == NULL) { 1089 // Usually, any field access or invocation on an unloaded oop type 1090 // will simply fail to link, since the statically linked class is 1091 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1092 // the static class is loaded but the sharper oop type is not. 1093 // Rather than checking for this obscure case in lots of places, 1094 // we simply observe that a null check on an unloaded class 1095 // will always be followed by a nonsense operation, so we 1096 // can just issue the uncommon trap here. 1097 // Our access to the unloaded class will only be correct 1098 // after it has been loaded and initialized, which requires 1099 // a trip through the interpreter. 1100 #ifndef PRODUCT 1101 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1102 #endif 1103 uncommon_trap(Deoptimization::Reason_unloaded, 1104 Deoptimization::Action_reinterpret, 1105 tp->klass(), "!loaded"); 1106 return top(); 1107 } 1108 1109 if (assert_null) { 1110 // See if the type is contained in NULL_PTR. 1111 // If so, then the value is already null. 1112 if (t->higher_equal(TypePtr::NULL_PTR)) { 1113 NOT_PRODUCT(explicit_null_checks_elided++); 1114 return value; // Elided null assert quickly! 1115 } 1116 } else { 1117 // See if mixing in the NULL pointer changes type. 1118 // If so, then the NULL pointer was not allowed in the original 1119 // type. In other words, "value" was not-null. 1120 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1121 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1122 NOT_PRODUCT(explicit_null_checks_elided++); 1123 return value; // Elided null check quickly! 1124 } 1125 } 1126 chk = new CmpPNode( value, null() ); 1127 break; 1128 } 1129 1130 default: 1131 fatal("unexpected type: %s", type2name(type)); 1132 } 1133 assert(chk != NULL, "sanity check"); 1134 chk = _gvn.transform(chk); 1135 1136 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1137 BoolNode *btst = new BoolNode( chk, btest); 1138 Node *tst = _gvn.transform( btst ); 1139 1140 //----------- 1141 // if peephole optimizations occurred, a prior test existed. 1142 // If a prior test existed, maybe it dominates as we can avoid this test. 1143 if (tst != btst && type == T_OBJECT) { 1144 // At this point we want to scan up the CFG to see if we can 1145 // find an identical test (and so avoid this test altogether). 1146 Node *cfg = control(); 1147 int depth = 0; 1148 while( depth < 16 ) { // Limit search depth for speed 1149 if( cfg->Opcode() == Op_IfTrue && 1150 cfg->in(0)->in(1) == tst ) { 1151 // Found prior test. Use "cast_not_null" to construct an identical 1152 // CastPP (and hence hash to) as already exists for the prior test. 1153 // Return that casted value. 1154 if (assert_null) { 1155 replace_in_map(value, null()); 1156 return null(); // do not issue the redundant test 1157 } 1158 Node *oldcontrol = control(); 1159 set_control(cfg); 1160 Node *res = cast_not_null(value); 1161 set_control(oldcontrol); 1162 NOT_PRODUCT(explicit_null_checks_elided++); 1163 return res; 1164 } 1165 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1166 if (cfg == NULL) break; // Quit at region nodes 1167 depth++; 1168 } 1169 } 1170 1171 //----------- 1172 // Branch to failure if null 1173 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1174 Deoptimization::DeoptReason reason; 1175 if (assert_null) { 1176 reason = Deoptimization::Reason_null_assert; 1177 } else if (type == T_OBJECT) { 1178 reason = Deoptimization::reason_null_check(speculative); 1179 } else { 1180 reason = Deoptimization::Reason_div0_check; 1181 } 1182 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1183 // ciMethodData::has_trap_at will return a conservative -1 if any 1184 // must-be-null assertion has failed. This could cause performance 1185 // problems for a method after its first do_null_assert failure. 1186 // Consider using 'Reason_class_check' instead? 1187 1188 // To cause an implicit null check, we set the not-null probability 1189 // to the maximum (PROB_MAX). For an explicit check the probability 1190 // is set to a smaller value. 1191 if (null_control != NULL || too_many_traps(reason)) { 1192 // probability is less likely 1193 ok_prob = PROB_LIKELY_MAG(3); 1194 } else if (!assert_null && 1195 (ImplicitNullCheckThreshold > 0) && 1196 method() != NULL && 1197 (method()->method_data()->trap_count(reason) 1198 >= (uint)ImplicitNullCheckThreshold)) { 1199 ok_prob = PROB_LIKELY_MAG(3); 1200 } 1201 1202 if (null_control != NULL) { 1203 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1204 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1205 set_control( _gvn.transform( new IfTrueNode(iff))); 1206 #ifndef PRODUCT 1207 if (null_true == top()) { 1208 explicit_null_checks_elided++; 1209 } 1210 #endif 1211 (*null_control) = null_true; 1212 } else { 1213 BuildCutout unless(this, tst, ok_prob); 1214 // Check for optimizer eliding test at parse time 1215 if (stopped()) { 1216 // Failure not possible; do not bother making uncommon trap. 1217 NOT_PRODUCT(explicit_null_checks_elided++); 1218 } else if (assert_null) { 1219 uncommon_trap(reason, 1220 Deoptimization::Action_make_not_entrant, 1221 NULL, "assert_null"); 1222 } else { 1223 replace_in_map(value, zerocon(type)); 1224 builtin_throw(reason); 1225 } 1226 } 1227 1228 // Must throw exception, fall-thru not possible? 1229 if (stopped()) { 1230 return top(); // No result 1231 } 1232 1233 if (assert_null) { 1234 // Cast obj to null on this path. 1235 replace_in_map(value, zerocon(type)); 1236 return zerocon(type); 1237 } 1238 1239 // Cast obj to not-null on this path, if there is no null_control. 1240 // (If there is a null_control, a non-null value may come back to haunt us.) 1241 if (type == T_OBJECT) { 1242 Node* cast = cast_not_null(value, false); 1243 if (null_control == NULL || (*null_control) == top()) 1244 replace_in_map(value, cast); 1245 value = cast; 1246 } 1247 1248 return value; 1249 } 1250 1251 1252 //------------------------------cast_not_null---------------------------------- 1253 // Cast obj to not-null on this path 1254 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1255 const Type *t = _gvn.type(obj); 1256 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1257 // Object is already not-null? 1258 if( t == t_not_null ) return obj; 1259 1260 Node *cast = new CastPPNode(obj,t_not_null); 1261 cast->init_req(0, control()); 1262 cast = _gvn.transform( cast ); 1263 1264 // Scan for instances of 'obj' in the current JVM mapping. 1265 // These instances are known to be not-null after the test. 1266 if (do_replace_in_map) 1267 replace_in_map(obj, cast); 1268 1269 return cast; // Return casted value 1270 } 1271 1272 1273 //--------------------------replace_in_map------------------------------------- 1274 void GraphKit::replace_in_map(Node* old, Node* neww) { 1275 if (old == neww) { 1276 return; 1277 } 1278 1279 map()->replace_edge(old, neww); 1280 1281 // Note: This operation potentially replaces any edge 1282 // on the map. This includes locals, stack, and monitors 1283 // of the current (innermost) JVM state. 1284 1285 // don't let inconsistent types from profiling escape this 1286 // method 1287 1288 const Type* told = _gvn.type(old); 1289 const Type* tnew = _gvn.type(neww); 1290 1291 if (!tnew->higher_equal(told)) { 1292 return; 1293 } 1294 1295 map()->record_replaced_node(old, neww); 1296 } 1297 1298 1299 //============================================================================= 1300 //--------------------------------memory--------------------------------------- 1301 Node* GraphKit::memory(uint alias_idx) { 1302 MergeMemNode* mem = merged_memory(); 1303 Node* p = mem->memory_at(alias_idx); 1304 _gvn.set_type(p, Type::MEMORY); // must be mapped 1305 return p; 1306 } 1307 1308 //-----------------------------reset_memory------------------------------------ 1309 Node* GraphKit::reset_memory() { 1310 Node* mem = map()->memory(); 1311 // do not use this node for any more parsing! 1312 debug_only( map()->set_memory((Node*)NULL) ); 1313 return _gvn.transform( mem ); 1314 } 1315 1316 //------------------------------set_all_memory--------------------------------- 1317 void GraphKit::set_all_memory(Node* newmem) { 1318 Node* mergemem = MergeMemNode::make(newmem); 1319 gvn().set_type_bottom(mergemem); 1320 map()->set_memory(mergemem); 1321 } 1322 1323 //------------------------------set_all_memory_call---------------------------- 1324 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1325 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1326 set_all_memory(newmem); 1327 } 1328 1329 //============================================================================= 1330 // 1331 // parser factory methods for MemNodes 1332 // 1333 // These are layered on top of the factory methods in LoadNode and StoreNode, 1334 // and integrate with the parser's memory state and _gvn engine. 1335 // 1336 1337 // factory methods in "int adr_idx" 1338 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1339 int adr_idx, 1340 MemNode::MemOrd mo, 1341 LoadNode::ControlDependency control_dependency, 1342 bool require_atomic_access, 1343 bool unaligned, 1344 bool mismatched) { 1345 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1346 const TypePtr* adr_type = NULL; // debug-mode-only argument 1347 debug_only(adr_type = C->get_adr_type(adr_idx)); 1348 Node* mem = memory(adr_idx); 1349 Node* ld; 1350 if (require_atomic_access && bt == T_LONG) { 1351 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1352 } else if (require_atomic_access && bt == T_DOUBLE) { 1353 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1354 } else { 1355 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1356 } 1357 ld = _gvn.transform(ld); 1358 1359 if (bt == T_VALUETYPE) { 1360 // Load non-flattened value type from memory. Add a null check and let the 1361 // interpreter take care of initializing the field to the default value type. 1362 Node* null_ctl = top(); 1363 ld = null_check_common(ld, bt, false, &null_ctl, false); 1364 if (null_ctl != top()) { 1365 assert(!adr_type->isa_aryptr(), "value type array must be initialized"); 1366 PreserveJVMState pjvms(this); 1367 set_control(null_ctl); 1368 uncommon_trap(Deoptimization::reason_null_check(false), Deoptimization::Action_maybe_recompile, 1369 t->is_valuetypeptr()->value_type()->value_klass(), "uninitialized non-flattened value type"); 1370 } 1371 ld = ValueTypeNode::make(gvn(), map()->memory(), ld); 1372 } else if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1373 // Improve graph before escape analysis and boxing elimination. 1374 record_for_igvn(ld); 1375 } 1376 return ld; 1377 } 1378 1379 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1380 int adr_idx, 1381 MemNode::MemOrd mo, 1382 bool require_atomic_access, 1383 bool unaligned, 1384 bool mismatched) { 1385 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1386 const TypePtr* adr_type = NULL; 1387 debug_only(adr_type = C->get_adr_type(adr_idx)); 1388 Node *mem = memory(adr_idx); 1389 Node* st; 1390 if (require_atomic_access && bt == T_LONG) { 1391 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1392 } else if (require_atomic_access && bt == T_DOUBLE) { 1393 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1394 } else { 1395 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1396 } 1397 if (unaligned) { 1398 st->as_Store()->set_unaligned_access(); 1399 } 1400 if (mismatched) { 1401 st->as_Store()->set_mismatched_access(); 1402 } 1403 st = _gvn.transform(st); 1404 set_memory(st, adr_idx); 1405 // Back-to-back stores can only remove intermediate store with DU info 1406 // so push on worklist for optimizer. 1407 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1408 record_for_igvn(st); 1409 1410 return st; 1411 } 1412 1413 1414 void GraphKit::pre_barrier(bool do_load, 1415 Node* ctl, 1416 Node* obj, 1417 Node* adr, 1418 uint adr_idx, 1419 Node* val, 1420 const TypeOopPtr* val_type, 1421 Node* pre_val, 1422 BasicType bt) { 1423 1424 BarrierSet* bs = Universe::heap()->barrier_set(); 1425 set_control(ctl); 1426 switch (bs->kind()) { 1427 case BarrierSet::G1SATBCTLogging: 1428 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1429 break; 1430 1431 case BarrierSet::CardTableForRS: 1432 case BarrierSet::CardTableExtension: 1433 case BarrierSet::ModRef: 1434 break; 1435 1436 default : 1437 ShouldNotReachHere(); 1438 1439 } 1440 } 1441 1442 bool GraphKit::can_move_pre_barrier() const { 1443 BarrierSet* bs = Universe::heap()->barrier_set(); 1444 switch (bs->kind()) { 1445 case BarrierSet::G1SATBCTLogging: 1446 return true; // Can move it if no safepoint 1447 1448 case BarrierSet::CardTableForRS: 1449 case BarrierSet::CardTableExtension: 1450 case BarrierSet::ModRef: 1451 return true; // There is no pre-barrier 1452 1453 default : 1454 ShouldNotReachHere(); 1455 } 1456 return false; 1457 } 1458 1459 void GraphKit::post_barrier(Node* ctl, 1460 Node* store, 1461 Node* obj, 1462 Node* adr, 1463 uint adr_idx, 1464 Node* val, 1465 BasicType bt, 1466 bool use_precise) { 1467 BarrierSet* bs = Universe::heap()->barrier_set(); 1468 set_control(ctl); 1469 switch (bs->kind()) { 1470 case BarrierSet::G1SATBCTLogging: 1471 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1472 break; 1473 1474 case BarrierSet::CardTableForRS: 1475 case BarrierSet::CardTableExtension: 1476 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1477 break; 1478 1479 case BarrierSet::ModRef: 1480 break; 1481 1482 default : 1483 ShouldNotReachHere(); 1484 1485 } 1486 } 1487 1488 Node* GraphKit::store_oop(Node* ctl, 1489 Node* obj, 1490 Node* adr, 1491 const TypePtr* adr_type, 1492 Node* val, 1493 const TypeOopPtr* val_type, 1494 BasicType bt, 1495 bool use_precise, 1496 MemNode::MemOrd mo, 1497 bool mismatched) { 1498 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1499 // could be delayed during Parse (for example, in adjust_map_after_if()). 1500 // Execute transformation here to avoid barrier generation in such case. 1501 if (_gvn.type(val) == TypePtr::NULL_PTR) 1502 val = _gvn.makecon(TypePtr::NULL_PTR); 1503 1504 set_control(ctl); 1505 if (stopped()) return top(); // Dead path ? 1506 1507 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity"); 1508 assert(val != NULL, "not dead path"); 1509 uint adr_idx = C->get_alias_index(adr_type); 1510 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1511 1512 if (bt == T_VALUETYPE) { 1513 // Allocate value type and store oop 1514 val = val->as_ValueType()->allocate(this); 1515 } 1516 1517 pre_barrier(true /* do_load */, 1518 control(), obj, adr, adr_idx, val, val_type, 1519 NULL /* pre_val */, 1520 bt); 1521 1522 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1523 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1524 return store; 1525 } 1526 1527 // Could be an array or object we don't know at compile time (unsafe ref.) 1528 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1529 Node* obj, // containing obj 1530 Node* adr, // actual adress to store val at 1531 const TypePtr* adr_type, 1532 Node* val, 1533 BasicType bt, 1534 MemNode::MemOrd mo, 1535 bool mismatched) { 1536 Compile::AliasType* at = C->alias_type(adr_type); 1537 const TypeOopPtr* val_type = NULL; 1538 if (adr_type->isa_instptr()) { 1539 if (at->field() != NULL) { 1540 // known field. This code is a copy of the do_put_xxx logic. 1541 ciField* field = at->field(); 1542 if (!field->type()->is_loaded()) { 1543 val_type = TypeInstPtr::BOTTOM; 1544 } else { 1545 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1546 } 1547 } 1548 } else if (adr_type->isa_aryptr()) { 1549 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1550 } 1551 if (val_type == NULL) { 1552 val_type = TypeInstPtr::BOTTOM; 1553 } 1554 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1555 } 1556 1557 1558 //-------------------------array_element_address------------------------- 1559 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1560 const TypeInt* sizetype, Node* ctrl) { 1561 uint shift = exact_log2(type2aelembytes(elembt)); 1562 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1563 if (arytype_klass->is_value_array_klass()) { 1564 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass(); 1565 shift = vak->log2_element_size(); 1566 } 1567 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1568 1569 // short-circuit a common case (saves lots of confusing waste motion) 1570 jint idx_con = find_int_con(idx, -1); 1571 if (idx_con >= 0) { 1572 intptr_t offset = header + ((intptr_t)idx_con << shift); 1573 return basic_plus_adr(ary, offset); 1574 } 1575 1576 // must be correct type for alignment purposes 1577 Node* base = basic_plus_adr(ary, header); 1578 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1579 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1580 return basic_plus_adr(ary, base, scale); 1581 } 1582 1583 //-------------------------load_array_element------------------------- 1584 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1585 const Type* elemtype = arytype->elem(); 1586 BasicType elembt = elemtype->array_element_basic_type(); 1587 assert(elembt != T_VALUETYPE, "value types are not supported by this method"); 1588 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1589 if (elembt == T_NARROWOOP) { 1590 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1591 } 1592 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1593 return ld; 1594 } 1595 1596 //-------------------------set_arguments_for_java_call------------------------- 1597 // Arguments (pre-popped from the stack) are taken from the JVMS. 1598 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1599 // Add the call arguments: 1600 const TypeTuple* domain = call->tf()->domain_sig(); 1601 uint nargs = domain->cnt(); 1602 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1603 Node* arg = argument(i-TypeFunc::Parms); 1604 if (ValueTypePassFieldsAsArgs) { 1605 if (arg->is_ValueType()) { 1606 ValueTypeNode* vt = arg->as_ValueType(); 1607 if (domain->field_at(i)->is_valuetypeptr()->klass() != C->env()->___Value_klass()) { 1608 // We don't pass value type arguments by reference but instead 1609 // pass each field of the value type 1610 idx += vt->pass_fields(call, idx, *this); 1611 // If a value type argument is passed as fields, attach the Method* to the call site 1612 // to be able to access the extended signature later via attached_method_before_pc(). 1613 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1614 call->set_override_symbolic_info(true); 1615 } else { 1616 arg = arg->as_ValueType()->allocate(this); 1617 call->init_req(idx, arg); 1618 idx++; 1619 } 1620 } else { 1621 call->init_req(idx, arg); 1622 idx++; 1623 } 1624 } else { 1625 if (arg->is_ValueType()) { 1626 // Pass value type argument via oop to callee 1627 arg = arg->as_ValueType()->allocate(this); 1628 } 1629 call->init_req(i, arg); 1630 } 1631 } 1632 } 1633 1634 //---------------------------set_edges_for_java_call--------------------------- 1635 // Connect a newly created call into the current JVMS. 1636 // A return value node (if any) is returned from set_edges_for_java_call. 1637 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1638 1639 // Add the predefined inputs: 1640 call->init_req( TypeFunc::Control, control() ); 1641 call->init_req( TypeFunc::I_O , i_o() ); 1642 call->init_req( TypeFunc::Memory , reset_memory() ); 1643 call->init_req( TypeFunc::FramePtr, frameptr() ); 1644 call->init_req( TypeFunc::ReturnAdr, top() ); 1645 1646 add_safepoint_edges(call, must_throw); 1647 1648 Node* xcall = _gvn.transform(call); 1649 1650 if (xcall == top()) { 1651 set_control(top()); 1652 return; 1653 } 1654 assert(xcall == call, "call identity is stable"); 1655 1656 // Re-use the current map to produce the result. 1657 1658 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1659 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1660 set_all_memory_call(xcall, separate_io_proj); 1661 1662 //return xcall; // no need, caller already has it 1663 } 1664 1665 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1666 if (stopped()) return top(); // maybe the call folded up? 1667 1668 // Capture the return value, if any. 1669 Node* ret; 1670 if (call->method() == NULL || 1671 call->method()->return_type()->basic_type() == T_VOID) 1672 ret = top(); 1673 else { 1674 if (!call->tf()->returns_value_type_as_fields()) { 1675 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1676 } else { 1677 // Return of multiple values (value type fields): we create a 1678 // ValueType node, each field is a projection from the call. 1679 const TypeTuple *range_sig = call->tf()->range_sig(); 1680 const Type* t = range_sig->field_at(TypeFunc::Parms); 1681 assert(t->isa_valuetypeptr(), "only value types for multiple return values"); 1682 ciValueKlass* vk = t->is_valuetypeptr()->value_type()->value_klass(); 1683 ret = ValueTypeNode::make(_gvn, call, vk, TypeFunc::Parms+1, false); 1684 } 1685 } 1686 1687 // Note: Since any out-of-line call can produce an exception, 1688 // we always insert an I_O projection from the call into the result. 1689 1690 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1691 1692 if (separate_io_proj) { 1693 // The caller requested separate projections be used by the fall 1694 // through and exceptional paths, so replace the projections for 1695 // the fall through path. 1696 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1697 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1698 } 1699 return ret; 1700 } 1701 1702 //--------------------set_predefined_input_for_runtime_call-------------------- 1703 // Reading and setting the memory state is way conservative here. 1704 // The real problem is that I am not doing real Type analysis on memory, 1705 // so I cannot distinguish card mark stores from other stores. Across a GC 1706 // point the Store Barrier and the card mark memory has to agree. I cannot 1707 // have a card mark store and its barrier split across the GC point from 1708 // either above or below. Here I get that to happen by reading ALL of memory. 1709 // A better answer would be to separate out card marks from other memory. 1710 // For now, return the input memory state, so that it can be reused 1711 // after the call, if this call has restricted memory effects. 1712 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1713 // Set fixed predefined input arguments 1714 Node* memory = reset_memory(); 1715 call->init_req( TypeFunc::Control, control() ); 1716 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1717 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1718 call->init_req( TypeFunc::FramePtr, frameptr() ); 1719 call->init_req( TypeFunc::ReturnAdr, top() ); 1720 return memory; 1721 } 1722 1723 //-------------------set_predefined_output_for_runtime_call-------------------- 1724 // Set control and memory (not i_o) from the call. 1725 // If keep_mem is not NULL, use it for the output state, 1726 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1727 // If hook_mem is NULL, this call produces no memory effects at all. 1728 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1729 // then only that memory slice is taken from the call. 1730 // In the last case, we must put an appropriate memory barrier before 1731 // the call, so as to create the correct anti-dependencies on loads 1732 // preceding the call. 1733 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1734 Node* keep_mem, 1735 const TypePtr* hook_mem) { 1736 // no i/o 1737 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1738 if (keep_mem) { 1739 // First clone the existing memory state 1740 set_all_memory(keep_mem); 1741 if (hook_mem != NULL) { 1742 // Make memory for the call 1743 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1744 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1745 // We also use hook_mem to extract specific effects from arraycopy stubs. 1746 set_memory(mem, hook_mem); 1747 } 1748 // ...else the call has NO memory effects. 1749 1750 // Make sure the call advertises its memory effects precisely. 1751 // This lets us build accurate anti-dependences in gcm.cpp. 1752 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1753 "call node must be constructed correctly"); 1754 } else { 1755 assert(hook_mem == NULL, ""); 1756 // This is not a "slow path" call; all memory comes from the call. 1757 set_all_memory_call(call); 1758 } 1759 } 1760 1761 1762 // Replace the call with the current state of the kit. 1763 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1764 JVMState* ejvms = NULL; 1765 if (has_exceptions()) { 1766 ejvms = transfer_exceptions_into_jvms(); 1767 } 1768 1769 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1770 ReplacedNodes replaced_nodes_exception; 1771 Node* ex_ctl = top(); 1772 1773 SafePointNode* final_state = stop(); 1774 1775 // Find all the needed outputs of this call 1776 CallProjections callprojs; 1777 call->extract_projections(&callprojs, true); 1778 1779 Node* init_mem = call->in(TypeFunc::Memory); 1780 Node* final_mem = final_state->in(TypeFunc::Memory); 1781 Node* final_ctl = final_state->in(TypeFunc::Control); 1782 Node* final_io = final_state->in(TypeFunc::I_O); 1783 1784 // Replace all the old call edges with the edges from the inlining result 1785 if (callprojs.fallthrough_catchproj != NULL) { 1786 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1787 } 1788 if (callprojs.fallthrough_memproj != NULL) { 1789 if (final_mem->is_MergeMem()) { 1790 // Parser's exits MergeMem was not transformed but may be optimized 1791 final_mem = _gvn.transform(final_mem); 1792 } 1793 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1794 } 1795 if (callprojs.fallthrough_ioproj != NULL) { 1796 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1797 } 1798 1799 // Replace the result with the new result if it exists and is used 1800 if (callprojs.resproj != NULL && result != NULL) { 1801 C->gvn_replace_by(callprojs.resproj, result); 1802 } 1803 1804 if (ejvms == NULL) { 1805 // No exception edges to simply kill off those paths 1806 if (callprojs.catchall_catchproj != NULL) { 1807 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1808 } 1809 if (callprojs.catchall_memproj != NULL) { 1810 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1811 } 1812 if (callprojs.catchall_ioproj != NULL) { 1813 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1814 } 1815 // Replace the old exception object with top 1816 if (callprojs.exobj != NULL) { 1817 C->gvn_replace_by(callprojs.exobj, C->top()); 1818 } 1819 } else { 1820 GraphKit ekit(ejvms); 1821 1822 // Load my combined exception state into the kit, with all phis transformed: 1823 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1824 replaced_nodes_exception = ex_map->replaced_nodes(); 1825 1826 Node* ex_oop = ekit.use_exception_state(ex_map); 1827 1828 if (callprojs.catchall_catchproj != NULL) { 1829 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1830 ex_ctl = ekit.control(); 1831 } 1832 if (callprojs.catchall_memproj != NULL) { 1833 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1834 } 1835 if (callprojs.catchall_ioproj != NULL) { 1836 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1837 } 1838 1839 // Replace the old exception object with the newly created one 1840 if (callprojs.exobj != NULL) { 1841 C->gvn_replace_by(callprojs.exobj, ex_oop); 1842 } 1843 } 1844 1845 // Disconnect the call from the graph 1846 call->disconnect_inputs(NULL, C); 1847 C->gvn_replace_by(call, C->top()); 1848 1849 // Clean up any MergeMems that feed other MergeMems since the 1850 // optimizer doesn't like that. 1851 if (final_mem->is_MergeMem()) { 1852 Node_List wl; 1853 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1854 Node* m = i.get(); 1855 if (m->is_MergeMem() && !wl.contains(m)) { 1856 wl.push(m); 1857 } 1858 } 1859 while (wl.size() > 0) { 1860 _gvn.transform(wl.pop()); 1861 } 1862 } 1863 1864 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1865 replaced_nodes.apply(C, final_ctl); 1866 } 1867 if (!ex_ctl->is_top() && do_replaced_nodes) { 1868 replaced_nodes_exception.apply(C, ex_ctl); 1869 } 1870 } 1871 1872 1873 //------------------------------increment_counter------------------------------ 1874 // for statistics: increment a VM counter by 1 1875 1876 void GraphKit::increment_counter(address counter_addr) { 1877 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1878 increment_counter(adr1); 1879 } 1880 1881 void GraphKit::increment_counter(Node* counter_addr) { 1882 int adr_type = Compile::AliasIdxRaw; 1883 Node* ctrl = control(); 1884 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1885 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1886 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1887 } 1888 1889 1890 //------------------------------uncommon_trap---------------------------------- 1891 // Bail out to the interpreter in mid-method. Implemented by calling the 1892 // uncommon_trap blob. This helper function inserts a runtime call with the 1893 // right debug info. 1894 void GraphKit::uncommon_trap(int trap_request, 1895 ciKlass* klass, const char* comment, 1896 bool must_throw, 1897 bool keep_exact_action) { 1898 if (failing()) stop(); 1899 if (stopped()) return; // trap reachable? 1900 1901 // Note: If ProfileTraps is true, and if a deopt. actually 1902 // occurs here, the runtime will make sure an MDO exists. There is 1903 // no need to call method()->ensure_method_data() at this point. 1904 1905 // Set the stack pointer to the right value for reexecution: 1906 set_sp(reexecute_sp()); 1907 1908 #ifdef ASSERT 1909 if (!must_throw) { 1910 // Make sure the stack has at least enough depth to execute 1911 // the current bytecode. 1912 int inputs, ignored_depth; 1913 if (compute_stack_effects(inputs, ignored_depth)) { 1914 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1915 Bytecodes::name(java_bc()), sp(), inputs); 1916 } 1917 } 1918 #endif 1919 1920 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1921 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1922 1923 switch (action) { 1924 case Deoptimization::Action_maybe_recompile: 1925 case Deoptimization::Action_reinterpret: 1926 // Temporary fix for 6529811 to allow virtual calls to be sure they 1927 // get the chance to go from mono->bi->mega 1928 if (!keep_exact_action && 1929 Deoptimization::trap_request_index(trap_request) < 0 && 1930 too_many_recompiles(reason)) { 1931 // This BCI is causing too many recompilations. 1932 if (C->log() != NULL) { 1933 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1934 Deoptimization::trap_reason_name(reason), 1935 Deoptimization::trap_action_name(action)); 1936 } 1937 action = Deoptimization::Action_none; 1938 trap_request = Deoptimization::make_trap_request(reason, action); 1939 } else { 1940 C->set_trap_can_recompile(true); 1941 } 1942 break; 1943 case Deoptimization::Action_make_not_entrant: 1944 C->set_trap_can_recompile(true); 1945 break; 1946 #ifdef ASSERT 1947 case Deoptimization::Action_none: 1948 case Deoptimization::Action_make_not_compilable: 1949 break; 1950 default: 1951 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 1952 break; 1953 #endif 1954 } 1955 1956 if (TraceOptoParse) { 1957 char buf[100]; 1958 tty->print_cr("Uncommon trap %s at bci:%d", 1959 Deoptimization::format_trap_request(buf, sizeof(buf), 1960 trap_request), bci()); 1961 } 1962 1963 CompileLog* log = C->log(); 1964 if (log != NULL) { 1965 int kid = (klass == NULL)? -1: log->identify(klass); 1966 log->begin_elem("uncommon_trap bci='%d'", bci()); 1967 char buf[100]; 1968 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1969 trap_request)); 1970 if (kid >= 0) log->print(" klass='%d'", kid); 1971 if (comment != NULL) log->print(" comment='%s'", comment); 1972 log->end_elem(); 1973 } 1974 1975 // Make sure any guarding test views this path as very unlikely 1976 Node *i0 = control()->in(0); 1977 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1978 IfNode *iff = i0->as_If(); 1979 float f = iff->_prob; // Get prob 1980 if (control()->Opcode() == Op_IfTrue) { 1981 if (f > PROB_UNLIKELY_MAG(4)) 1982 iff->_prob = PROB_MIN; 1983 } else { 1984 if (f < PROB_LIKELY_MAG(4)) 1985 iff->_prob = PROB_MAX; 1986 } 1987 } 1988 1989 // Clear out dead values from the debug info. 1990 kill_dead_locals(); 1991 1992 // Now insert the uncommon trap subroutine call 1993 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 1994 const TypePtr* no_memory_effects = NULL; 1995 // Pass the index of the class to be loaded 1996 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1997 (must_throw ? RC_MUST_THROW : 0), 1998 OptoRuntime::uncommon_trap_Type(), 1999 call_addr, "uncommon_trap", no_memory_effects, 2000 intcon(trap_request)); 2001 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2002 "must extract request correctly from the graph"); 2003 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2004 2005 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2006 // The debug info is the only real input to this call. 2007 2008 // Halt-and-catch fire here. The above call should never return! 2009 HaltNode* halt = new HaltNode(control(), frameptr()); 2010 _gvn.set_type_bottom(halt); 2011 root()->add_req(halt); 2012 2013 stop_and_kill_map(); 2014 } 2015 2016 2017 //--------------------------just_allocated_object------------------------------ 2018 // Report the object that was just allocated. 2019 // It must be the case that there are no intervening safepoints. 2020 // We use this to determine if an object is so "fresh" that 2021 // it does not require card marks. 2022 Node* GraphKit::just_allocated_object(Node* current_control) { 2023 if (C->recent_alloc_ctl() == current_control) 2024 return C->recent_alloc_obj(); 2025 return NULL; 2026 } 2027 2028 2029 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2030 // (Note: TypeFunc::make has a cache that makes this fast.) 2031 const TypeFunc* tf = TypeFunc::make(dest_method); 2032 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2033 for (int j = 0; j < nargs; j++) { 2034 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2035 if( targ->basic_type() == T_DOUBLE ) { 2036 // If any parameters are doubles, they must be rounded before 2037 // the call, dstore_rounding does gvn.transform 2038 Node *arg = argument(j); 2039 arg = dstore_rounding(arg); 2040 set_argument(j, arg); 2041 } 2042 } 2043 } 2044 2045 /** 2046 * Record profiling data exact_kls for Node n with the type system so 2047 * that it can propagate it (speculation) 2048 * 2049 * @param n node that the type applies to 2050 * @param exact_kls type from profiling 2051 * @param maybe_null did profiling see null? 2052 * 2053 * @return node with improved type 2054 */ 2055 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2056 const Type* current_type = _gvn.type(n); 2057 assert(UseTypeSpeculation, "type speculation must be on"); 2058 2059 const TypePtr* speculative = current_type->speculative(); 2060 2061 // Should the klass from the profile be recorded in the speculative type? 2062 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2063 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2064 const TypeOopPtr* xtype = tklass->as_instance_type(); 2065 assert(xtype->klass_is_exact(), "Should be exact"); 2066 // Any reason to believe n is not null (from this profiling or a previous one)? 2067 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2068 // record the new speculative type's depth 2069 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2070 speculative = speculative->with_inline_depth(jvms()->depth()); 2071 } else if (current_type->would_improve_ptr(maybe_null)) { 2072 // Profiling report that null was never seen so we can change the 2073 // speculative type to non null ptr. 2074 assert(!maybe_null, "nothing to improve"); 2075 if (speculative == NULL) { 2076 speculative = TypePtr::NOTNULL; 2077 } else { 2078 const TypePtr* ptr = TypePtr::NOTNULL; 2079 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2080 } 2081 } 2082 2083 if (speculative != current_type->speculative()) { 2084 // Build a type with a speculative type (what we think we know 2085 // about the type but will need a guard when we use it) 2086 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2087 // We're changing the type, we need a new CheckCast node to carry 2088 // the new type. The new type depends on the control: what 2089 // profiling tells us is only valid from here as far as we can 2090 // tell. 2091 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2092 cast = _gvn.transform(cast); 2093 replace_in_map(n, cast); 2094 n = cast; 2095 } 2096 2097 return n; 2098 } 2099 2100 /** 2101 * Record profiling data from receiver profiling at an invoke with the 2102 * type system so that it can propagate it (speculation) 2103 * 2104 * @param n receiver node 2105 * 2106 * @return node with improved type 2107 */ 2108 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2109 if (!UseTypeSpeculation) { 2110 return n; 2111 } 2112 ciKlass* exact_kls = profile_has_unique_klass(); 2113 bool maybe_null = true; 2114 if (java_bc() == Bytecodes::_checkcast || 2115 java_bc() == Bytecodes::_instanceof || 2116 java_bc() == Bytecodes::_aastore) { 2117 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2118 maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2119 } 2120 return record_profile_for_speculation(n, exact_kls, maybe_null); 2121 } 2122 2123 /** 2124 * Record profiling data from argument profiling at an invoke with the 2125 * type system so that it can propagate it (speculation) 2126 * 2127 * @param dest_method target method for the call 2128 * @param bc what invoke bytecode is this? 2129 */ 2130 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2131 if (!UseTypeSpeculation) { 2132 return; 2133 } 2134 const TypeFunc* tf = TypeFunc::make(dest_method); 2135 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2136 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2137 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2138 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2139 if (targ->isa_oopptr() && !targ->isa_valuetypeptr()) { 2140 bool maybe_null = true; 2141 ciKlass* better_type = NULL; 2142 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2143 record_profile_for_speculation(argument(j), better_type, maybe_null); 2144 } 2145 i++; 2146 } 2147 } 2148 } 2149 2150 /** 2151 * Record profiling data from parameter profiling at an invoke with 2152 * the type system so that it can propagate it (speculation) 2153 */ 2154 void GraphKit::record_profiled_parameters_for_speculation() { 2155 if (!UseTypeSpeculation) { 2156 return; 2157 } 2158 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2159 if (_gvn.type(local(i))->isa_oopptr()) { 2160 bool maybe_null = true; 2161 ciKlass* better_type = NULL; 2162 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2163 record_profile_for_speculation(local(i), better_type, maybe_null); 2164 } 2165 j++; 2166 } 2167 } 2168 } 2169 2170 /** 2171 * Record profiling data from return value profiling at an invoke with 2172 * the type system so that it can propagate it (speculation) 2173 */ 2174 void GraphKit::record_profiled_return_for_speculation() { 2175 if (!UseTypeSpeculation) { 2176 return; 2177 } 2178 bool maybe_null = true; 2179 ciKlass* better_type = NULL; 2180 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2181 // If profiling reports a single type for the return value, 2182 // feed it to the type system so it can propagate it as a 2183 // speculative type 2184 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2185 } 2186 } 2187 2188 void GraphKit::round_double_result(ciMethod* dest_method) { 2189 // A non-strict method may return a double value which has an extended 2190 // exponent, but this must not be visible in a caller which is 'strict' 2191 // If a strict caller invokes a non-strict callee, round a double result 2192 2193 BasicType result_type = dest_method->return_type()->basic_type(); 2194 assert( method() != NULL, "must have caller context"); 2195 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2196 // Destination method's return value is on top of stack 2197 // dstore_rounding() does gvn.transform 2198 Node *result = pop_pair(); 2199 result = dstore_rounding(result); 2200 push_pair(result); 2201 } 2202 } 2203 2204 // rounding for strict float precision conformance 2205 Node* GraphKit::precision_rounding(Node* n) { 2206 return UseStrictFP && _method->flags().is_strict() 2207 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2208 ? _gvn.transform( new RoundFloatNode(0, n) ) 2209 : n; 2210 } 2211 2212 // rounding for strict double precision conformance 2213 Node* GraphKit::dprecision_rounding(Node *n) { 2214 return UseStrictFP && _method->flags().is_strict() 2215 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2216 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2217 : n; 2218 } 2219 2220 // rounding for non-strict double stores 2221 Node* GraphKit::dstore_rounding(Node* n) { 2222 return Matcher::strict_fp_requires_explicit_rounding 2223 && UseSSE <= 1 2224 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2225 : n; 2226 } 2227 2228 //============================================================================= 2229 // Generate a fast path/slow path idiom. Graph looks like: 2230 // [foo] indicates that 'foo' is a parameter 2231 // 2232 // [in] NULL 2233 // \ / 2234 // CmpP 2235 // Bool ne 2236 // If 2237 // / \ 2238 // True False-<2> 2239 // / | 2240 // / cast_not_null 2241 // Load | | ^ 2242 // [fast_test] | | 2243 // gvn to opt_test | | 2244 // / \ | <1> 2245 // True False | 2246 // | \\ | 2247 // [slow_call] \[fast_result] 2248 // Ctl Val \ \ 2249 // | \ \ 2250 // Catch <1> \ \ 2251 // / \ ^ \ \ 2252 // Ex No_Ex | \ \ 2253 // | \ \ | \ <2> \ 2254 // ... \ [slow_res] | | \ [null_result] 2255 // \ \--+--+--- | | 2256 // \ | / \ | / 2257 // --------Region Phi 2258 // 2259 //============================================================================= 2260 // Code is structured as a series of driver functions all called 'do_XXX' that 2261 // call a set of helper functions. Helper functions first, then drivers. 2262 2263 //------------------------------null_check_oop--------------------------------- 2264 // Null check oop. Set null-path control into Region in slot 3. 2265 // Make a cast-not-nullness use the other not-null control. Return cast. 2266 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2267 bool never_see_null, 2268 bool safe_for_replace, 2269 bool speculative) { 2270 // Initial NULL check taken path 2271 (*null_control) = top(); 2272 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2273 2274 // Generate uncommon_trap: 2275 if (never_see_null && (*null_control) != top()) { 2276 // If we see an unexpected null at a check-cast we record it and force a 2277 // recompile; the offending check-cast will be compiled to handle NULLs. 2278 // If we see more than one offending BCI, then all checkcasts in the 2279 // method will be compiled to handle NULLs. 2280 PreserveJVMState pjvms(this); 2281 set_control(*null_control); 2282 replace_in_map(value, null()); 2283 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2284 uncommon_trap(reason, 2285 Deoptimization::Action_make_not_entrant); 2286 (*null_control) = top(); // NULL path is dead 2287 } 2288 if ((*null_control) == top() && safe_for_replace) { 2289 replace_in_map(value, cast); 2290 } 2291 2292 // Cast away null-ness on the result 2293 return cast; 2294 } 2295 2296 //------------------------------opt_iff---------------------------------------- 2297 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2298 // Return slow-path control. 2299 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2300 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2301 2302 // Fast path taken; set region slot 2 2303 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2304 region->init_req(2,fast_taken); // Capture fast-control 2305 2306 // Fast path not-taken, i.e. slow path 2307 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2308 return slow_taken; 2309 } 2310 2311 //-----------------------------make_runtime_call------------------------------- 2312 Node* GraphKit::make_runtime_call(int flags, 2313 const TypeFunc* call_type, address call_addr, 2314 const char* call_name, 2315 const TypePtr* adr_type, 2316 // The following parms are all optional. 2317 // The first NULL ends the list. 2318 Node* parm0, Node* parm1, 2319 Node* parm2, Node* parm3, 2320 Node* parm4, Node* parm5, 2321 Node* parm6, Node* parm7) { 2322 // Slow-path call 2323 bool is_leaf = !(flags & RC_NO_LEAF); 2324 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2325 if (call_name == NULL) { 2326 assert(!is_leaf, "must supply name for leaf"); 2327 call_name = OptoRuntime::stub_name(call_addr); 2328 } 2329 CallNode* call; 2330 if (!is_leaf) { 2331 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2332 bci(), adr_type); 2333 } else if (flags & RC_NO_FP) { 2334 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2335 } else { 2336 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2337 } 2338 2339 // The following is similar to set_edges_for_java_call, 2340 // except that the memory effects of the call are restricted to AliasIdxRaw. 2341 2342 // Slow path call has no side-effects, uses few values 2343 bool wide_in = !(flags & RC_NARROW_MEM); 2344 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2345 2346 Node* prev_mem = NULL; 2347 if (wide_in) { 2348 prev_mem = set_predefined_input_for_runtime_call(call); 2349 } else { 2350 assert(!wide_out, "narrow in => narrow out"); 2351 Node* narrow_mem = memory(adr_type); 2352 prev_mem = reset_memory(); 2353 map()->set_memory(narrow_mem); 2354 set_predefined_input_for_runtime_call(call); 2355 } 2356 2357 // Hook each parm in order. Stop looking at the first NULL. 2358 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2359 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2360 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2361 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2362 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2363 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2364 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2365 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2366 /* close each nested if ===> */ } } } } } } } } 2367 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2368 2369 if (!is_leaf) { 2370 // Non-leaves can block and take safepoints: 2371 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2372 } 2373 // Non-leaves can throw exceptions: 2374 if (has_io) { 2375 call->set_req(TypeFunc::I_O, i_o()); 2376 } 2377 2378 if (flags & RC_UNCOMMON) { 2379 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2380 // (An "if" probability corresponds roughly to an unconditional count. 2381 // Sort of.) 2382 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2383 } 2384 2385 Node* c = _gvn.transform(call); 2386 assert(c == call, "cannot disappear"); 2387 2388 if (wide_out) { 2389 // Slow path call has full side-effects. 2390 set_predefined_output_for_runtime_call(call); 2391 } else { 2392 // Slow path call has few side-effects, and/or sets few values. 2393 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2394 } 2395 2396 if (has_io) { 2397 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2398 } 2399 return call; 2400 2401 } 2402 2403 //------------------------------merge_memory----------------------------------- 2404 // Merge memory from one path into the current memory state. 2405 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2406 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2407 Node* old_slice = mms.force_memory(); 2408 Node* new_slice = mms.memory2(); 2409 if (old_slice != new_slice) { 2410 PhiNode* phi; 2411 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2412 if (mms.is_empty()) { 2413 // clone base memory Phi's inputs for this memory slice 2414 assert(old_slice == mms.base_memory(), "sanity"); 2415 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2416 _gvn.set_type(phi, Type::MEMORY); 2417 for (uint i = 1; i < phi->req(); i++) { 2418 phi->init_req(i, old_slice->in(i)); 2419 } 2420 } else { 2421 phi = old_slice->as_Phi(); // Phi was generated already 2422 } 2423 } else { 2424 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2425 _gvn.set_type(phi, Type::MEMORY); 2426 } 2427 phi->set_req(new_path, new_slice); 2428 mms.set_memory(phi); 2429 } 2430 } 2431 } 2432 2433 //------------------------------make_slow_call_ex------------------------------ 2434 // Make the exception handler hookups for the slow call 2435 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2436 if (stopped()) return; 2437 2438 // Make a catch node with just two handlers: fall-through and catch-all 2439 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2440 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2441 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2442 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2443 2444 { PreserveJVMState pjvms(this); 2445 set_control(excp); 2446 set_i_o(i_o); 2447 2448 if (excp != top()) { 2449 if (deoptimize) { 2450 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2451 uncommon_trap(Deoptimization::Reason_unhandled, 2452 Deoptimization::Action_none); 2453 } else { 2454 // Create an exception state also. 2455 // Use an exact type if the caller has specified a specific exception. 2456 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2457 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2458 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2459 } 2460 } 2461 } 2462 2463 // Get the no-exception control from the CatchNode. 2464 set_control(norm); 2465 } 2466 2467 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2468 Node* cmp = NULL; 2469 switch(bt) { 2470 case T_INT: cmp = new CmpINode(in1, in2); break; 2471 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2472 default: fatal("unexpected comparison type %s", type2name(bt)); 2473 } 2474 gvn->transform(cmp); 2475 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2476 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2477 gvn->transform(iff); 2478 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2479 return iff; 2480 } 2481 2482 2483 //-------------------------------gen_subtype_check----------------------------- 2484 // Generate a subtyping check. Takes as input the subtype and supertype. 2485 // Returns 2 values: sets the default control() to the true path and returns 2486 // the false path. Only reads invariant memory; sets no (visible) memory. 2487 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2488 // but that's not exposed to the optimizer. This call also doesn't take in an 2489 // Object; if you wish to check an Object you need to load the Object's class 2490 // prior to coming here. 2491 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2492 Compile* C = gvn->C; 2493 2494 if ((*ctrl)->is_top()) { 2495 return C->top(); 2496 } 2497 2498 // Fast check for identical types, perhaps identical constants. 2499 // The types can even be identical non-constants, in cases 2500 // involving Array.newInstance, Object.clone, etc. 2501 if (subklass == superklass) 2502 return C->top(); // false path is dead; no test needed. 2503 2504 if (gvn->type(superklass)->singleton()) { 2505 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2506 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2507 2508 // In the common case of an exact superklass, try to fold up the 2509 // test before generating code. You may ask, why not just generate 2510 // the code and then let it fold up? The answer is that the generated 2511 // code will necessarily include null checks, which do not always 2512 // completely fold away. If they are also needless, then they turn 2513 // into a performance loss. Example: 2514 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2515 // Here, the type of 'fa' is often exact, so the store check 2516 // of fa[1]=x will fold up, without testing the nullness of x. 2517 switch (C->static_subtype_check(superk, subk)) { 2518 case Compile::SSC_always_false: 2519 { 2520 Node* always_fail = *ctrl; 2521 *ctrl = gvn->C->top(); 2522 return always_fail; 2523 } 2524 case Compile::SSC_always_true: 2525 return C->top(); 2526 case Compile::SSC_easy_test: 2527 { 2528 // Just do a direct pointer compare and be done. 2529 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2530 *ctrl = gvn->transform(new IfTrueNode(iff)); 2531 return gvn->transform(new IfFalseNode(iff)); 2532 } 2533 case Compile::SSC_full_test: 2534 break; 2535 default: 2536 ShouldNotReachHere(); 2537 } 2538 } 2539 2540 // %%% Possible further optimization: Even if the superklass is not exact, 2541 // if the subklass is the unique subtype of the superklass, the check 2542 // will always succeed. We could leave a dependency behind to ensure this. 2543 2544 // First load the super-klass's check-offset 2545 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2546 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2547 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2548 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2549 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2550 2551 // Load from the sub-klass's super-class display list, or a 1-word cache of 2552 // the secondary superclass list, or a failing value with a sentinel offset 2553 // if the super-klass is an interface or exceptionally deep in the Java 2554 // hierarchy and we have to scan the secondary superclass list the hard way. 2555 // Worst-case type is a little odd: NULL is allowed as a result (usually 2556 // klass loads can never produce a NULL). 2557 Node *chk_off_X = chk_off; 2558 #ifdef _LP64 2559 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2560 #endif 2561 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2562 // For some types like interfaces the following loadKlass is from a 1-word 2563 // cache which is mutable so can't use immutable memory. Other 2564 // types load from the super-class display table which is immutable. 2565 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2566 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2567 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::BOTTOM)); 2568 2569 // Compile speed common case: ARE a subtype and we canNOT fail 2570 if( superklass == nkls ) 2571 return C->top(); // false path is dead; no test needed. 2572 2573 // See if we get an immediate positive hit. Happens roughly 83% of the 2574 // time. Test to see if the value loaded just previously from the subklass 2575 // is exactly the superklass. 2576 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2577 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2578 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2579 2580 // Compile speed common case: Check for being deterministic right now. If 2581 // chk_off is a constant and not equal to cacheoff then we are NOT a 2582 // subklass. In this case we need exactly the 1 test above and we can 2583 // return those results immediately. 2584 if (!might_be_cache) { 2585 Node* not_subtype_ctrl = *ctrl; 2586 *ctrl = iftrue1; // We need exactly the 1 test above 2587 return not_subtype_ctrl; 2588 } 2589 2590 // Gather the various success & failures here 2591 RegionNode *r_ok_subtype = new RegionNode(4); 2592 gvn->record_for_igvn(r_ok_subtype); 2593 RegionNode *r_not_subtype = new RegionNode(3); 2594 gvn->record_for_igvn(r_not_subtype); 2595 2596 r_ok_subtype->init_req(1, iftrue1); 2597 2598 // Check for immediate negative hit. Happens roughly 11% of the time (which 2599 // is roughly 63% of the remaining cases). Test to see if the loaded 2600 // check-offset points into the subklass display list or the 1-element 2601 // cache. If it points to the display (and NOT the cache) and the display 2602 // missed then it's not a subtype. 2603 Node *cacheoff = gvn->intcon(cacheoff_con); 2604 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2605 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2606 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2607 2608 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2609 // No performance impact (too rare) but allows sharing of secondary arrays 2610 // which has some footprint reduction. 2611 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2612 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2613 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2614 2615 // -- Roads not taken here: -- 2616 // We could also have chosen to perform the self-check at the beginning 2617 // of this code sequence, as the assembler does. This would not pay off 2618 // the same way, since the optimizer, unlike the assembler, can perform 2619 // static type analysis to fold away many successful self-checks. 2620 // Non-foldable self checks work better here in second position, because 2621 // the initial primary superclass check subsumes a self-check for most 2622 // types. An exception would be a secondary type like array-of-interface, 2623 // which does not appear in its own primary supertype display. 2624 // Finally, we could have chosen to move the self-check into the 2625 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2626 // dependent manner. But it is worthwhile to have the check here, 2627 // where it can be perhaps be optimized. The cost in code space is 2628 // small (register compare, branch). 2629 2630 // Now do a linear scan of the secondary super-klass array. Again, no real 2631 // performance impact (too rare) but it's gotta be done. 2632 // Since the code is rarely used, there is no penalty for moving it 2633 // out of line, and it can only improve I-cache density. 2634 // The decision to inline or out-of-line this final check is platform 2635 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2636 Node* psc = gvn->transform( 2637 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2638 2639 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2640 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2641 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2642 2643 // Return false path; set default control to true path. 2644 *ctrl = gvn->transform(r_ok_subtype); 2645 return gvn->transform(r_not_subtype); 2646 } 2647 2648 // Profile-driven exact type check: 2649 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2650 float prob, 2651 Node* *casted_receiver) { 2652 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2653 Node* recv_klass = load_object_klass(receiver); 2654 Node* fail = type_check(recv_klass, tklass, prob); 2655 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2656 assert(recv_xtype->klass_is_exact(), ""); 2657 2658 // Subsume downstream occurrences of receiver with a cast to 2659 // recv_xtype, since now we know what the type will be. 2660 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2661 (*casted_receiver) = _gvn.transform(cast); 2662 // (User must make the replace_in_map call.) 2663 2664 return fail; 2665 } 2666 2667 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 2668 float prob) { 2669 //const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2670 Node* want_klass = makecon(tklass); 2671 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass)); 2672 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2673 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2674 set_control( _gvn.transform( new IfTrueNode (iff))); 2675 Node* fail = _gvn.transform( new IfFalseNode(iff)); 2676 return fail; 2677 } 2678 2679 2680 //------------------------------seems_never_null------------------------------- 2681 // Use null_seen information if it is available from the profile. 2682 // If we see an unexpected null at a type check we record it and force a 2683 // recompile; the offending check will be recompiled to handle NULLs. 2684 // If we see several offending BCIs, then all checks in the 2685 // method will be recompiled. 2686 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2687 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2688 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2689 if (UncommonNullCast // Cutout for this technique 2690 && obj != null() // And not the -Xcomp stupid case? 2691 && !too_many_traps(reason) 2692 ) { 2693 if (speculating) { 2694 return true; 2695 } 2696 if (data == NULL) 2697 // Edge case: no mature data. Be optimistic here. 2698 return true; 2699 // If the profile has not seen a null, assume it won't happen. 2700 assert(java_bc() == Bytecodes::_checkcast || 2701 java_bc() == Bytecodes::_instanceof || 2702 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2703 return !data->as_BitData()->null_seen(); 2704 } 2705 speculating = false; 2706 return false; 2707 } 2708 2709 //------------------------maybe_cast_profiled_receiver------------------------- 2710 // If the profile has seen exactly one type, narrow to exactly that type. 2711 // Subsequent type checks will always fold up. 2712 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2713 ciKlass* require_klass, 2714 ciKlass* spec_klass, 2715 bool safe_for_replace) { 2716 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2717 2718 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2719 2720 // Make sure we haven't already deoptimized from this tactic. 2721 if (too_many_traps(reason) || too_many_recompiles(reason)) 2722 return NULL; 2723 2724 // (No, this isn't a call, but it's enough like a virtual call 2725 // to use the same ciMethod accessor to get the profile info...) 2726 // If we have a speculative type use it instead of profiling (which 2727 // may not help us) 2728 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2729 if (exact_kls != NULL) {// no cast failures here 2730 if (require_klass == NULL || 2731 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2732 // If we narrow the type to match what the type profile sees or 2733 // the speculative type, we can then remove the rest of the 2734 // cast. 2735 // This is a win, even if the exact_kls is very specific, 2736 // because downstream operations, such as method calls, 2737 // will often benefit from the sharper type. 2738 Node* exact_obj = not_null_obj; // will get updated in place... 2739 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2740 &exact_obj); 2741 { PreserveJVMState pjvms(this); 2742 set_control(slow_ctl); 2743 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2744 } 2745 if (safe_for_replace) { 2746 replace_in_map(not_null_obj, exact_obj); 2747 } 2748 return exact_obj; 2749 } 2750 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2751 } 2752 2753 return NULL; 2754 } 2755 2756 /** 2757 * Cast obj to type and emit guard unless we had too many traps here 2758 * already 2759 * 2760 * @param obj node being casted 2761 * @param type type to cast the node to 2762 * @param not_null true if we know node cannot be null 2763 */ 2764 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2765 ciKlass* type, 2766 bool not_null) { 2767 if (stopped()) { 2768 return obj; 2769 } 2770 2771 // type == NULL if profiling tells us this object is always null 2772 if (type != NULL) { 2773 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2774 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2775 2776 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2777 !too_many_traps(class_reason) && 2778 !too_many_recompiles(class_reason)) { 2779 Node* not_null_obj = NULL; 2780 // not_null is true if we know the object is not null and 2781 // there's no need for a null check 2782 if (!not_null) { 2783 Node* null_ctl = top(); 2784 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2785 assert(null_ctl->is_top(), "no null control here"); 2786 } else { 2787 not_null_obj = obj; 2788 } 2789 2790 Node* exact_obj = not_null_obj; 2791 ciKlass* exact_kls = type; 2792 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2793 &exact_obj); 2794 { 2795 PreserveJVMState pjvms(this); 2796 set_control(slow_ctl); 2797 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2798 } 2799 replace_in_map(not_null_obj, exact_obj); 2800 obj = exact_obj; 2801 } 2802 } else { 2803 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2804 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2805 Node* exact_obj = null_assert(obj); 2806 replace_in_map(obj, exact_obj); 2807 obj = exact_obj; 2808 } 2809 } 2810 return obj; 2811 } 2812 2813 //-------------------------------gen_instanceof-------------------------------- 2814 // Generate an instance-of idiom. Used by both the instance-of bytecode 2815 // and the reflective instance-of call. 2816 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2817 kill_dead_locals(); // Benefit all the uncommon traps 2818 assert( !stopped(), "dead parse path should be checked in callers" ); 2819 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2820 "must check for not-null not-dead klass in callers"); 2821 2822 // Make the merge point 2823 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2824 RegionNode* region = new RegionNode(PATH_LIMIT); 2825 Node* phi = new PhiNode(region, TypeInt::BOOL); 2826 C->set_has_split_ifs(true); // Has chance for split-if optimization 2827 2828 ciProfileData* data = NULL; 2829 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2830 data = method()->method_data()->bci_to_data(bci()); 2831 } 2832 bool speculative_not_null = false; 2833 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2834 && seems_never_null(obj, data, speculative_not_null)); 2835 2836 // Null check; get casted pointer; set region slot 3 2837 Node* null_ctl = top(); 2838 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2839 2840 // If not_null_obj is dead, only null-path is taken 2841 if (stopped()) { // Doing instance-of on a NULL? 2842 set_control(null_ctl); 2843 return intcon(0); 2844 } 2845 region->init_req(_null_path, null_ctl); 2846 phi ->init_req(_null_path, intcon(0)); // Set null path value 2847 if (null_ctl == top()) { 2848 // Do this eagerly, so that pattern matches like is_diamond_phi 2849 // will work even during parsing. 2850 assert(_null_path == PATH_LIMIT-1, "delete last"); 2851 region->del_req(_null_path); 2852 phi ->del_req(_null_path); 2853 } 2854 2855 // Do we know the type check always succeed? 2856 bool known_statically = false; 2857 if (_gvn.type(superklass)->singleton()) { 2858 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2859 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2860 if (subk != NULL && subk->is_loaded()) { 2861 int static_res = C->static_subtype_check(superk, subk); 2862 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2863 } 2864 } 2865 2866 if (known_statically && UseTypeSpeculation) { 2867 // If we know the type check always succeeds then we don't use the 2868 // profiling data at this bytecode. Don't lose it, feed it to the 2869 // type system as a speculative type. 2870 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2871 } else { 2872 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2873 // We may not have profiling here or it may not help us. If we 2874 // have a speculative type use it to perform an exact cast. 2875 ciKlass* spec_obj_type = obj_type->speculative_type(); 2876 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2877 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2878 if (stopped()) { // Profile disagrees with this path. 2879 set_control(null_ctl); // Null is the only remaining possibility. 2880 return intcon(0); 2881 } 2882 if (cast_obj != NULL) { 2883 not_null_obj = cast_obj; 2884 } 2885 } 2886 } 2887 2888 // Load the object's klass 2889 Node* obj_klass = load_object_klass(not_null_obj); 2890 2891 // Generate the subtype check 2892 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2893 2894 // Plug in the success path to the general merge in slot 1. 2895 region->init_req(_obj_path, control()); 2896 phi ->init_req(_obj_path, intcon(1)); 2897 2898 // Plug in the failing path to the general merge in slot 2. 2899 region->init_req(_fail_path, not_subtype_ctrl); 2900 phi ->init_req(_fail_path, intcon(0)); 2901 2902 // Return final merged results 2903 set_control( _gvn.transform(region) ); 2904 record_for_igvn(region); 2905 return _gvn.transform(phi); 2906 } 2907 2908 //-------------------------------gen_checkcast--------------------------------- 2909 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2910 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2911 // uncommon-trap paths work. Adjust stack after this call. 2912 // If failure_control is supplied and not null, it is filled in with 2913 // the control edge for the cast failure. Otherwise, an appropriate 2914 // uncommon trap or exception is thrown. 2915 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2916 Node* *failure_control) { 2917 kill_dead_locals(); // Benefit all the uncommon traps 2918 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2919 assert(tk->is_loaded(), "must be loaded"); 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 assert(klass != NULL, "klass should not be NULL"); 3227 bool xklass = inst_klass->klass_is_exact(); 3228 if (xklass || klass->is_array_klass()) { 3229 jint lhelper = klass->layout_helper(); 3230 if (lhelper != Klass::_lh_neutral_value) { 3231 constant_value = lhelper; 3232 return (Node*) NULL; 3233 } 3234 } 3235 } 3236 constant_value = Klass::_lh_neutral_value; // put in a known value 3237 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3238 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3239 } 3240 3241 // We just put in an allocate/initialize with a big raw-memory effect. 3242 // Hook selected additional alias categories on the initialization. 3243 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3244 MergeMemNode* init_in_merge, 3245 Node* init_out_raw) { 3246 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3247 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3248 3249 Node* prevmem = kit.memory(alias_idx); 3250 init_in_merge->set_memory_at(alias_idx, prevmem); 3251 kit.set_memory(init_out_raw, alias_idx); 3252 } 3253 3254 //---------------------------set_output_for_allocation------------------------- 3255 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3256 const TypeOopPtr* oop_type, 3257 bool deoptimize_on_exception) { 3258 int rawidx = Compile::AliasIdxRaw; 3259 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3260 add_safepoint_edges(alloc); 3261 Node* allocx = _gvn.transform(alloc); 3262 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3263 // create memory projection for i_o 3264 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3265 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3266 3267 // create a memory projection as for the normal control path 3268 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3269 set_memory(malloc, rawidx); 3270 3271 // a normal slow-call doesn't change i_o, but an allocation does 3272 // we create a separate i_o projection for the normal control path 3273 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3274 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3275 3276 // put in an initialization barrier 3277 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3278 rawoop)->as_Initialize(); 3279 assert(alloc->initialization() == init, "2-way macro link must work"); 3280 assert(init ->allocation() == alloc, "2-way macro link must work"); 3281 { 3282 // Extract memory strands which may participate in the new object's 3283 // initialization, and source them from the new InitializeNode. 3284 // This will allow us to observe initializations when they occur, 3285 // and link them properly (as a group) to the InitializeNode. 3286 assert(init->in(InitializeNode::Memory) == malloc, ""); 3287 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3288 init->set_req(InitializeNode::Memory, minit_in); 3289 record_for_igvn(minit_in); // fold it up later, if possible 3290 Node* minit_out = memory(rawidx); 3291 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3292 if (oop_type->isa_aryptr()) { 3293 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3294 int elemidx = C->get_alias_index(telemref); 3295 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3296 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) { 3297 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3298 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3299 ciField* field = ik->nonstatic_field_at(i); 3300 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3301 continue; // do not bother to track really large numbers of fields 3302 // Find (or create) the alias category for this field: 3303 int fieldidx = C->alias_type(field)->index(); 3304 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3305 } 3306 } 3307 } 3308 3309 // Cast raw oop to the real thing... 3310 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3311 javaoop = _gvn.transform(javaoop); 3312 C->set_recent_alloc(control(), javaoop); 3313 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3314 3315 #ifdef ASSERT 3316 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3317 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3318 "Ideal_allocation works"); 3319 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3320 "Ideal_allocation works"); 3321 if (alloc->is_AllocateArray()) { 3322 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3323 "Ideal_allocation works"); 3324 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3325 "Ideal_allocation works"); 3326 } else { 3327 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3328 } 3329 } 3330 #endif //ASSERT 3331 3332 return javaoop; 3333 } 3334 3335 //---------------------------new_instance-------------------------------------- 3336 // This routine takes a klass_node which may be constant (for a static type) 3337 // or may be non-constant (for reflective code). It will work equally well 3338 // for either, and the graph will fold nicely if the optimizer later reduces 3339 // the type to a constant. 3340 // The optional arguments are for specialized use by intrinsics: 3341 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3342 // - If 'return_size_val', report the the total object size to the caller. 3343 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3344 Node* GraphKit::new_instance(Node* klass_node, 3345 Node* extra_slow_test, 3346 Node* *return_size_val, 3347 bool deoptimize_on_exception, 3348 ValueTypeNode* value_node) { 3349 // Compute size in doublewords 3350 // The size is always an integral number of doublewords, represented 3351 // as a positive bytewise size stored in the klass's layout_helper. 3352 // The layout_helper also encodes (in a low bit) the need for a slow path. 3353 jint layout_con = Klass::_lh_neutral_value; 3354 Node* layout_val = get_layout_helper(klass_node, layout_con); 3355 bool layout_is_con = (layout_val == NULL); 3356 3357 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3358 // Generate the initial go-slow test. It's either ALWAYS (return a 3359 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3360 // case) a computed value derived from the layout_helper. 3361 Node* initial_slow_test = NULL; 3362 if (layout_is_con) { 3363 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3364 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3365 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3366 } else { // reflective case 3367 // This reflective path is used by Unsafe.allocateInstance. 3368 // (It may be stress-tested by specifying StressReflectiveCode.) 3369 // Basically, we want to get into the VM is there's an illegal argument. 3370 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3371 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3372 if (extra_slow_test != intcon(0)) { 3373 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3374 } 3375 // (Macro-expander will further convert this to a Bool, if necessary.) 3376 } 3377 3378 // Find the size in bytes. This is easy; it's the layout_helper. 3379 // The size value must be valid even if the slow path is taken. 3380 Node* size = NULL; 3381 if (layout_is_con) { 3382 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3383 } else { // reflective case 3384 // This reflective path is used by clone and Unsafe.allocateInstance. 3385 size = ConvI2X(layout_val); 3386 3387 // Clear the low bits to extract layout_helper_size_in_bytes: 3388 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3389 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3390 size = _gvn.transform( new AndXNode(size, mask) ); 3391 } 3392 if (return_size_val != NULL) { 3393 (*return_size_val) = size; 3394 } 3395 3396 // This is a precise notnull oop of the klass. 3397 // (Actually, it need not be precise if this is a reflective allocation.) 3398 // It's what we cast the result to. 3399 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3400 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3401 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3402 3403 // Now generate allocation code 3404 3405 // The entire memory state is needed for slow path of the allocation 3406 // since GC and deoptimization can happen. 3407 Node *mem = reset_memory(); 3408 set_all_memory(mem); // Create new memory state 3409 3410 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3411 control(), mem, i_o(), 3412 size, klass_node, 3413 initial_slow_test, value_node); 3414 3415 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3416 } 3417 3418 //-------------------------------new_array------------------------------------- 3419 // helper for newarray and anewarray 3420 // The 'length' parameter is (obviously) the length of the array. 3421 // See comments on new_instance for the meaning of the other arguments. 3422 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3423 Node* length, // number of array elements 3424 int nargs, // number of arguments to push back for uncommon trap 3425 Node* *return_size_val, 3426 bool deoptimize_on_exception) { 3427 jint layout_con = Klass::_lh_neutral_value; 3428 Node* layout_val = get_layout_helper(klass_node, layout_con); 3429 bool layout_is_con = (layout_val == NULL); 3430 3431 if (!layout_is_con && !StressReflectiveCode && 3432 !too_many_traps(Deoptimization::Reason_class_check)) { 3433 // This is a reflective array creation site. 3434 // Optimistically assume that it is a subtype of Object[], 3435 // so that we can fold up all the address arithmetic. 3436 layout_con = Klass::array_layout_helper(T_OBJECT); 3437 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3438 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3439 { BuildCutout unless(this, bol_lh, PROB_MAX); 3440 inc_sp(nargs); 3441 uncommon_trap(Deoptimization::Reason_class_check, 3442 Deoptimization::Action_maybe_recompile); 3443 } 3444 layout_val = NULL; 3445 layout_is_con = true; 3446 } 3447 3448 // Generate the initial go-slow test. Make sure we do not overflow 3449 // if length is huge (near 2Gig) or negative! We do not need 3450 // exact double-words here, just a close approximation of needed 3451 // double-words. We can't add any offset or rounding bits, lest we 3452 // take a size -1 of bytes and make it positive. Use an unsigned 3453 // compare, so negative sizes look hugely positive. 3454 int fast_size_limit = FastAllocateSizeLimit; 3455 if (layout_is_con) { 3456 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3457 // Increase the size limit if we have exact knowledge of array type. 3458 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3459 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3460 } 3461 3462 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3463 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3464 3465 // --- Size Computation --- 3466 // array_size = round_to_heap(array_header + (length << elem_shift)); 3467 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3468 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3469 // The rounding mask is strength-reduced, if possible. 3470 int round_mask = MinObjAlignmentInBytes - 1; 3471 Node* header_size = NULL; 3472 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3473 // (T_BYTE has the weakest alignment and size restrictions...) 3474 if (layout_is_con) { 3475 int hsize = Klass::layout_helper_header_size(layout_con); 3476 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3477 BasicType etype = Klass::layout_helper_element_type(layout_con); 3478 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con); 3479 if ((round_mask & ~right_n_bits(eshift)) == 0) 3480 round_mask = 0; // strength-reduce it if it goes away completely 3481 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3482 assert(header_size_min <= hsize, "generic minimum is smallest"); 3483 header_size_min = hsize; 3484 header_size = intcon(hsize + round_mask); 3485 } else { 3486 Node* hss = intcon(Klass::_lh_header_size_shift); 3487 Node* hsm = intcon(Klass::_lh_header_size_mask); 3488 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3489 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3490 Node* mask = intcon(round_mask); 3491 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3492 } 3493 3494 Node* elem_shift = NULL; 3495 if (layout_is_con) { 3496 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3497 if (eshift != 0) 3498 elem_shift = intcon(eshift); 3499 } else { 3500 // There is no need to mask or shift this value. 3501 // The semantics of LShiftINode include an implicit mask to 0x1F. 3502 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3503 elem_shift = layout_val; 3504 } 3505 3506 // Transition to native address size for all offset calculations: 3507 Node* lengthx = ConvI2X(length); 3508 Node* headerx = ConvI2X(header_size); 3509 #ifdef _LP64 3510 { const TypeInt* tilen = _gvn.find_int_type(length); 3511 if (tilen != NULL && tilen->_lo < 0) { 3512 // Add a manual constraint to a positive range. Cf. array_element_address. 3513 jint size_max = fast_size_limit; 3514 if (size_max > tilen->_hi) size_max = tilen->_hi; 3515 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3516 3517 // Only do a narrow I2L conversion if the range check passed. 3518 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3519 _gvn.transform(iff); 3520 RegionNode* region = new RegionNode(3); 3521 _gvn.set_type(region, Type::CONTROL); 3522 lengthx = new PhiNode(region, TypeLong::LONG); 3523 _gvn.set_type(lengthx, TypeLong::LONG); 3524 3525 // Range check passed. Use ConvI2L node with narrow type. 3526 Node* passed = IfFalse(iff); 3527 region->init_req(1, passed); 3528 // Make I2L conversion control dependent to prevent it from 3529 // floating above the range check during loop optimizations. 3530 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3531 3532 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3533 region->init_req(2, IfTrue(iff)); 3534 lengthx->init_req(2, ConvI2X(length)); 3535 3536 set_control(region); 3537 record_for_igvn(region); 3538 record_for_igvn(lengthx); 3539 } 3540 } 3541 #endif 3542 3543 // Combine header size (plus rounding) and body size. Then round down. 3544 // This computation cannot overflow, because it is used only in two 3545 // places, one where the length is sharply limited, and the other 3546 // after a successful allocation. 3547 Node* abody = lengthx; 3548 if (elem_shift != NULL) 3549 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3550 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3551 if (round_mask != 0) { 3552 Node* mask = MakeConX(~round_mask); 3553 size = _gvn.transform( new AndXNode(size, mask) ); 3554 } 3555 // else if round_mask == 0, the size computation is self-rounding 3556 3557 if (return_size_val != NULL) { 3558 // This is the size 3559 (*return_size_val) = size; 3560 } 3561 3562 // Now generate allocation code 3563 3564 // The entire memory state is needed for slow path of the allocation 3565 // since GC and deoptimization can happen. 3566 Node *mem = reset_memory(); 3567 set_all_memory(mem); // Create new memory state 3568 3569 if (initial_slow_test->is_Bool()) { 3570 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3571 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3572 } 3573 3574 // Create the AllocateArrayNode and its result projections 3575 AllocateArrayNode* alloc 3576 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3577 control(), mem, i_o(), 3578 size, klass_node, 3579 initial_slow_test, 3580 length); 3581 3582 // Cast to correct type. Note that the klass_node may be constant or not, 3583 // and in the latter case the actual array type will be inexact also. 3584 // (This happens via a non-constant argument to inline_native_newArray.) 3585 // In any case, the value of klass_node provides the desired array type. 3586 const TypeInt* length_type = _gvn.find_int_type(length); 3587 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3588 if (ary_type->isa_aryptr() && length_type != NULL) { 3589 // Try to get a better type than POS for the size 3590 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3591 } 3592 3593 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3594 3595 // Cast length on remaining path to be as narrow as possible 3596 if (map()->find_edge(length) >= 0) { 3597 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3598 if (ccast != length) { 3599 _gvn.set_type_bottom(ccast); 3600 record_for_igvn(ccast); 3601 replace_in_map(length, ccast); 3602 } 3603 } 3604 3605 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 3606 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL; 3607 if (elem_klass != NULL && elem_klass->is_valuetype()) { 3608 ciValueKlass* vk = elem_klass->as_value_klass(); 3609 if (!vk->flatten_array()) { 3610 // Non-flattened value type arrays need to be initialized with default value type oops 3611 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs); 3612 InitializeNode* init = alloc->initialization(); 3613 init->set_complete_with_arraycopy(); 3614 } 3615 } 3616 3617 return javaoop; 3618 } 3619 3620 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) { 3621 // Check for zero length 3622 Node* null_ctl = top(); 3623 null_check_common(length, T_INT, false, &null_ctl, false); 3624 if (stopped()) { 3625 set_control(null_ctl); // Always zero 3626 return; 3627 } 3628 3629 // Prepare for merging control and IO 3630 RegionNode* res_ctl = new RegionNode(3); 3631 res_ctl->init_req(1, null_ctl); 3632 gvn().set_type(res_ctl, Type::CONTROL); 3633 record_for_igvn(res_ctl); 3634 Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO); 3635 gvn().set_type(res_io, Type::ABIO); 3636 record_for_igvn(res_io); 3637 3638 // TODO comment 3639 SafePointNode* loop_map = NULL; 3640 { 3641 PreserveJVMState pjvms(this); 3642 // Create default value type and store it to memory 3643 Node* oop = ValueTypeNode::make_default(gvn(), vk); 3644 oop = oop->as_ValueType()->allocate(this); 3645 3646 length = SubI(length, intcon(1)); 3647 add_predicate(nargs); 3648 RegionNode* loop = new RegionNode(3); 3649 loop->init_req(1, control()); 3650 gvn().set_type(loop, Type::CONTROL); 3651 record_for_igvn(loop); 3652 3653 Node* index = new PhiNode(loop, TypeInt::INT); 3654 index->init_req(1, intcon(0)); 3655 gvn().set_type(index, TypeInt::INT); 3656 record_for_igvn(index); 3657 3658 // TODO explain why we need to capture all memory 3659 PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM); 3660 mem->init_req(1, reset_memory()); 3661 gvn().set_type(mem, Type::MEMORY); 3662 record_for_igvn(mem); 3663 set_control(loop); 3664 set_all_memory(mem); 3665 // Initialize array element 3666 Node* adr = array_element_address(array, index, T_OBJECT); 3667 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk); 3668 Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release); 3669 3670 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 3671 loop->init_req(2, IfTrue(iff)); 3672 mem->init_req(2, merged_memory()); 3673 index->init_req(2, AddI(index, intcon(1))); 3674 3675 res_ctl->init_req(2, IfFalse(iff)); 3676 res_io->set_req(2, i_o()); 3677 loop_map = stop(); 3678 } 3679 // Set merged control, IO and memory 3680 set_control(res_ctl); 3681 set_i_o(res_io); 3682 merge_memory(loop_map->merged_memory(), res_ctl, 2); 3683 3684 // Transform new memory Phis. 3685 for (MergeMemStream mms(merged_memory()); mms.next_non_empty();) { 3686 Node* phi = mms.memory(); 3687 if (phi->is_Phi() && phi->in(0) == res_ctl) { 3688 mms.set_memory(gvn().transform(phi)); 3689 } 3690 } 3691 } 3692 3693 // The following "Ideal_foo" functions are placed here because they recognize 3694 // the graph shapes created by the functions immediately above. 3695 3696 //---------------------------Ideal_allocation---------------------------------- 3697 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3698 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3699 if (ptr == NULL) { // reduce dumb test in callers 3700 return NULL; 3701 } 3702 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3703 ptr = ptr->in(1); 3704 if (ptr == NULL) return NULL; 3705 } 3706 // Return NULL for allocations with several casts: 3707 // j.l.reflect.Array.newInstance(jobject, jint) 3708 // Object.clone() 3709 // to keep more precise type from last cast. 3710 if (ptr->is_Proj()) { 3711 Node* allo = ptr->in(0); 3712 if (allo != NULL && allo->is_Allocate()) { 3713 return allo->as_Allocate(); 3714 } 3715 } 3716 // Report failure to match. 3717 return NULL; 3718 } 3719 3720 // Fancy version which also strips off an offset (and reports it to caller). 3721 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3722 intptr_t& offset) { 3723 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3724 if (base == NULL) return NULL; 3725 return Ideal_allocation(base, phase); 3726 } 3727 3728 // Trace Initialize <- Proj[Parm] <- Allocate 3729 AllocateNode* InitializeNode::allocation() { 3730 Node* rawoop = in(InitializeNode::RawAddress); 3731 if (rawoop->is_Proj()) { 3732 Node* alloc = rawoop->in(0); 3733 if (alloc->is_Allocate()) { 3734 return alloc->as_Allocate(); 3735 } 3736 } 3737 return NULL; 3738 } 3739 3740 // Trace Allocate -> Proj[Parm] -> Initialize 3741 InitializeNode* AllocateNode::initialization() { 3742 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3743 if (rawoop == NULL) return NULL; 3744 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3745 Node* init = rawoop->fast_out(i); 3746 if (init->is_Initialize()) { 3747 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3748 return init->as_Initialize(); 3749 } 3750 } 3751 return NULL; 3752 } 3753 3754 //----------------------------- loop predicates --------------------------- 3755 3756 //------------------------------add_predicate_impl---------------------------- 3757 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3758 // Too many traps seen? 3759 if (too_many_traps(reason)) { 3760 #ifdef ASSERT 3761 if (TraceLoopPredicate) { 3762 int tc = C->trap_count(reason); 3763 tty->print("too many traps=%s tcount=%d in ", 3764 Deoptimization::trap_reason_name(reason), tc); 3765 method()->print(); // which method has too many predicate traps 3766 tty->cr(); 3767 } 3768 #endif 3769 // We cannot afford to take more traps here, 3770 // do not generate predicate. 3771 return; 3772 } 3773 3774 Node *cont = _gvn.intcon(1); 3775 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3776 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3777 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3778 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3779 C->add_predicate_opaq(opq); 3780 { 3781 PreserveJVMState pjvms(this); 3782 set_control(iffalse); 3783 inc_sp(nargs); 3784 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3785 } 3786 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3787 set_control(iftrue); 3788 } 3789 3790 //------------------------------add_predicate--------------------------------- 3791 void GraphKit::add_predicate(int nargs) { 3792 if (UseLoopPredicate) { 3793 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3794 } 3795 // loop's limit check predicate should be near the loop. 3796 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3797 } 3798 3799 //----------------------------- store barriers ---------------------------- 3800 #define __ ideal. 3801 3802 void GraphKit::sync_kit(IdealKit& ideal) { 3803 set_all_memory(__ merged_memory()); 3804 set_i_o(__ i_o()); 3805 set_control(__ ctrl()); 3806 } 3807 3808 void GraphKit::final_sync(IdealKit& ideal) { 3809 // Final sync IdealKit and graphKit. 3810 sync_kit(ideal); 3811 } 3812 3813 Node* GraphKit::byte_map_base_node() { 3814 // Get base of card map 3815 CardTableModRefBS* ct = 3816 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3817 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3818 if (ct->byte_map_base != NULL) { 3819 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3820 } else { 3821 return null(); 3822 } 3823 } 3824 3825 // vanilla/CMS post barrier 3826 // Insert a write-barrier store. This is to let generational GC work; we have 3827 // to flag all oop-stores before the next GC point. 3828 void GraphKit::write_barrier_post(Node* oop_store, 3829 Node* obj, 3830 Node* adr, 3831 uint adr_idx, 3832 Node* val, 3833 bool use_precise) { 3834 // No store check needed if we're storing a NULL or an old object 3835 // (latter case is probably a string constant). The concurrent 3836 // mark sweep garbage collector, however, needs to have all nonNull 3837 // oop updates flagged via card-marks. 3838 if (val != NULL && val->is_Con()) { 3839 // must be either an oop or NULL 3840 const Type* t = val->bottom_type(); 3841 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3842 // stores of null never (?) need barriers 3843 return; 3844 } 3845 3846 if (use_ReduceInitialCardMarks() 3847 && obj == just_allocated_object(control())) { 3848 // We can skip marks on a freshly-allocated object in Eden. 3849 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3850 // That routine informs GC to take appropriate compensating steps, 3851 // upon a slow-path allocation, so as to make this card-mark 3852 // elision safe. 3853 return; 3854 } 3855 3856 if (!use_precise) { 3857 // All card marks for a (non-array) instance are in one place: 3858 adr = obj; 3859 } 3860 // (Else it's an array (or unknown), and we want more precise card marks.) 3861 assert(adr != NULL, ""); 3862 3863 IdealKit ideal(this, true); 3864 3865 // Convert the pointer to an int prior to doing math on it 3866 Node* cast = __ CastPX(__ ctrl(), adr); 3867 3868 // Divide by card size 3869 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3870 "Only one we handle so far."); 3871 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3872 3873 // Combine card table base and card offset 3874 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3875 3876 // Get the alias_index for raw card-mark memory 3877 int adr_type = Compile::AliasIdxRaw; 3878 Node* zero = __ ConI(0); // Dirty card value 3879 BasicType bt = T_BYTE; 3880 3881 if (UseConcMarkSweepGC && UseCondCardMark) { 3882 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3883 __ sync_kit(this); 3884 } 3885 3886 if (UseCondCardMark) { 3887 // The classic GC reference write barrier is typically implemented 3888 // as a store into the global card mark table. Unfortunately 3889 // unconditional stores can result in false sharing and excessive 3890 // coherence traffic as well as false transactional aborts. 3891 // UseCondCardMark enables MP "polite" conditional card mark 3892 // stores. In theory we could relax the load from ctrl() to 3893 // no_ctrl, but that doesn't buy much latitude. 3894 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3895 __ if_then(card_val, BoolTest::ne, zero); 3896 } 3897 3898 // Smash zero into card 3899 if( !UseConcMarkSweepGC ) { 3900 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3901 } else { 3902 // Specialized path for CM store barrier 3903 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3904 } 3905 3906 if (UseCondCardMark) { 3907 __ end_if(); 3908 } 3909 3910 // Final sync IdealKit and GraphKit. 3911 final_sync(ideal); 3912 } 3913 /* 3914 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3915 * required by SATB to make sure all objects live at the start of the 3916 * marking are kept alive, all reference updates need to any previous 3917 * reference stored before writing. 3918 * 3919 * If the previous value is NULL there is no need to save the old value. 3920 * References that are NULL are filtered during runtime by the barrier 3921 * code to avoid unnecessary queuing. 3922 * 3923 * However in the case of newly allocated objects it might be possible to 3924 * prove that the reference about to be overwritten is NULL during compile 3925 * time and avoid adding the barrier code completely. 3926 * 3927 * The compiler needs to determine that the object in which a field is about 3928 * to be written is newly allocated, and that no prior store to the same field 3929 * has happened since the allocation. 3930 * 3931 * Returns true if the pre-barrier can be removed 3932 */ 3933 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3934 BasicType bt, uint adr_idx) { 3935 intptr_t offset = 0; 3936 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3937 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3938 3939 if (offset == Type::OffsetBot) { 3940 return false; // cannot unalias unless there are precise offsets 3941 } 3942 3943 if (alloc == NULL) { 3944 return false; // No allocation found 3945 } 3946 3947 intptr_t size_in_bytes = type2aelembytes(bt); 3948 3949 Node* mem = memory(adr_idx); // start searching here... 3950 3951 for (int cnt = 0; cnt < 50; cnt++) { 3952 3953 if (mem->is_Store()) { 3954 3955 Node* st_adr = mem->in(MemNode::Address); 3956 intptr_t st_offset = 0; 3957 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3958 3959 if (st_base == NULL) { 3960 break; // inscrutable pointer 3961 } 3962 3963 // Break we have found a store with same base and offset as ours so break 3964 if (st_base == base && st_offset == offset) { 3965 break; 3966 } 3967 3968 if (st_offset != offset && st_offset != Type::OffsetBot) { 3969 const int MAX_STORE = BytesPerLong; 3970 if (st_offset >= offset + size_in_bytes || 3971 st_offset <= offset - MAX_STORE || 3972 st_offset <= offset - mem->as_Store()->memory_size()) { 3973 // Success: The offsets are provably independent. 3974 // (You may ask, why not just test st_offset != offset and be done? 3975 // The answer is that stores of different sizes can co-exist 3976 // in the same sequence of RawMem effects. We sometimes initialize 3977 // a whole 'tile' of array elements with a single jint or jlong.) 3978 mem = mem->in(MemNode::Memory); 3979 continue; // advance through independent store memory 3980 } 3981 } 3982 3983 if (st_base != base 3984 && MemNode::detect_ptr_independence(base, alloc, st_base, 3985 AllocateNode::Ideal_allocation(st_base, phase), 3986 phase)) { 3987 // Success: The bases are provably independent. 3988 mem = mem->in(MemNode::Memory); 3989 continue; // advance through independent store memory 3990 } 3991 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3992 3993 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3994 AllocateNode* st_alloc = st_init->allocation(); 3995 3996 // Make sure that we are looking at the same allocation site. 3997 // The alloc variable is guaranteed to not be null here from earlier check. 3998 if (alloc == st_alloc) { 3999 // Check that the initialization is storing NULL so that no previous store 4000 // has been moved up and directly write a reference 4001 Node* captured_store = st_init->find_captured_store(offset, 4002 type2aelembytes(T_OBJECT), 4003 phase); 4004 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4005 return true; 4006 } 4007 } 4008 } 4009 4010 // Unless there is an explicit 'continue', we must bail out here, 4011 // because 'mem' is an inscrutable memory state (e.g., a call). 4012 break; 4013 } 4014 4015 return false; 4016 } 4017 4018 // G1 pre/post barriers 4019 void GraphKit::g1_write_barrier_pre(bool do_load, 4020 Node* obj, 4021 Node* adr, 4022 uint alias_idx, 4023 Node* val, 4024 const TypeOopPtr* val_type, 4025 Node* pre_val, 4026 BasicType bt) { 4027 4028 // Some sanity checks 4029 // Note: val is unused in this routine. 4030 4031 if (do_load) { 4032 // We need to generate the load of the previous value 4033 assert(obj != NULL, "must have a base"); 4034 assert(adr != NULL, "where are loading from?"); 4035 assert(pre_val == NULL, "loaded already?"); 4036 assert(val_type != NULL, "need a type"); 4037 4038 if (use_ReduceInitialCardMarks() 4039 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4040 return; 4041 } 4042 4043 } else { 4044 // In this case both val_type and alias_idx are unused. 4045 assert(pre_val != NULL, "must be loaded already"); 4046 // Nothing to be done if pre_val is null. 4047 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4048 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4049 } 4050 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here"); 4051 4052 IdealKit ideal(this, true); 4053 4054 Node* tls = __ thread(); // ThreadLocalStorage 4055 4056 Node* no_ctrl = NULL; 4057 Node* no_base = __ top(); 4058 Node* zero = __ ConI(0); 4059 Node* zeroX = __ ConX(0); 4060 4061 float likely = PROB_LIKELY(0.999); 4062 float unlikely = PROB_UNLIKELY(0.999); 4063 4064 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4065 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4066 4067 // Offsets into the thread 4068 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4069 SATBMarkQueue::byte_offset_of_active()); 4070 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4071 SATBMarkQueue::byte_offset_of_index()); 4072 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4073 SATBMarkQueue::byte_offset_of_buf()); 4074 4075 // Now the actual pointers into the thread 4076 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4077 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4078 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4079 4080 // Now some of the values 4081 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4082 4083 // if (!marking) 4084 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4085 BasicType index_bt = TypeX_X->basic_type(); 4086 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4087 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4088 4089 if (do_load) { 4090 // load original value 4091 // alias_idx correct?? 4092 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4093 } 4094 4095 // if (pre_val != NULL) 4096 __ if_then(pre_val, BoolTest::ne, null()); { 4097 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4098 4099 // is the queue for this thread full? 4100 __ if_then(index, BoolTest::ne, zeroX, likely); { 4101 4102 // decrement the index 4103 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4104 4105 // Now get the buffer location we will log the previous value into and store it 4106 Node *log_addr = __ AddP(no_base, buffer, next_index); 4107 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4108 // update the index 4109 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4110 4111 } __ else_(); { 4112 4113 // logging buffer is full, call the runtime 4114 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4115 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4116 } __ end_if(); // (!index) 4117 } __ end_if(); // (pre_val != NULL) 4118 } __ end_if(); // (!marking) 4119 4120 // Final sync IdealKit and GraphKit. 4121 final_sync(ideal); 4122 } 4123 4124 /* 4125 * G1 similar to any GC with a Young Generation requires a way to keep track of 4126 * references from Old Generation to Young Generation to make sure all live 4127 * objects are found. G1 also requires to keep track of object references 4128 * between different regions to enable evacuation of old regions, which is done 4129 * as part of mixed collections. References are tracked in remembered sets and 4130 * is continuously updated as reference are written to with the help of the 4131 * post-barrier. 4132 * 4133 * To reduce the number of updates to the remembered set the post-barrier 4134 * filters updates to fields in objects located in the Young Generation, 4135 * the same region as the reference, when the NULL is being written or 4136 * if the card is already marked as dirty by an earlier write. 4137 * 4138 * Under certain circumstances it is possible to avoid generating the 4139 * post-barrier completely if it is possible during compile time to prove 4140 * the object is newly allocated and that no safepoint exists between the 4141 * allocation and the store. 4142 * 4143 * In the case of slow allocation the allocation code must handle the barrier 4144 * as part of the allocation in the case the allocated object is not located 4145 * in the nursery, this would happen for humongous objects. This is similar to 4146 * how CMS is required to handle this case, see the comments for the method 4147 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4148 * A deferred card mark is required for these objects and handled in the above 4149 * mentioned methods. 4150 * 4151 * Returns true if the post barrier can be removed 4152 */ 4153 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4154 Node* adr) { 4155 intptr_t offset = 0; 4156 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4157 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4158 4159 if (offset == Type::OffsetBot) { 4160 return false; // cannot unalias unless there are precise offsets 4161 } 4162 4163 if (alloc == NULL) { 4164 return false; // No allocation found 4165 } 4166 4167 // Start search from Store node 4168 Node* mem = store->in(MemNode::Control); 4169 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4170 4171 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4172 AllocateNode* st_alloc = st_init->allocation(); 4173 4174 // Make sure we are looking at the same allocation 4175 if (alloc == st_alloc) { 4176 return true; 4177 } 4178 } 4179 4180 return false; 4181 } 4182 4183 // 4184 // Update the card table and add card address to the queue 4185 // 4186 void GraphKit::g1_mark_card(IdealKit& ideal, 4187 Node* card_adr, 4188 Node* oop_store, 4189 uint oop_alias_idx, 4190 Node* index, 4191 Node* index_adr, 4192 Node* buffer, 4193 const TypeFunc* tf) { 4194 4195 Node* zero = __ ConI(0); 4196 Node* zeroX = __ ConX(0); 4197 Node* no_base = __ top(); 4198 BasicType card_bt = T_BYTE; 4199 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4200 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4201 4202 // Now do the queue work 4203 __ if_then(index, BoolTest::ne, zeroX); { 4204 4205 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4206 Node* log_addr = __ AddP(no_base, buffer, next_index); 4207 4208 // Order, see storeCM. 4209 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4210 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4211 4212 } __ else_(); { 4213 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4214 } __ end_if(); 4215 4216 } 4217 4218 void GraphKit::g1_write_barrier_post(Node* oop_store, 4219 Node* obj, 4220 Node* adr, 4221 uint alias_idx, 4222 Node* val, 4223 BasicType bt, 4224 bool use_precise) { 4225 // If we are writing a NULL then we need no post barrier 4226 4227 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4228 // Must be NULL 4229 const Type* t = val->bottom_type(); 4230 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4231 // No post barrier if writing NULLx 4232 return; 4233 } 4234 4235 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4236 // We can skip marks on a freshly-allocated object in Eden. 4237 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4238 // That routine informs GC to take appropriate compensating steps, 4239 // upon a slow-path allocation, so as to make this card-mark 4240 // elision safe. 4241 return; 4242 } 4243 4244 if (use_ReduceInitialCardMarks() 4245 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4246 return; 4247 } 4248 4249 if (!use_precise) { 4250 // All card marks for a (non-array) instance are in one place: 4251 adr = obj; 4252 } 4253 // (Else it's an array (or unknown), and we want more precise card marks.) 4254 assert(adr != NULL, ""); 4255 4256 IdealKit ideal(this, true); 4257 4258 Node* tls = __ thread(); // ThreadLocalStorage 4259 4260 Node* no_base = __ top(); 4261 float likely = PROB_LIKELY(0.999); 4262 float unlikely = PROB_UNLIKELY(0.999); 4263 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4264 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4265 Node* zeroX = __ ConX(0); 4266 4267 // Get the alias_index for raw card-mark memory 4268 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4269 4270 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4271 4272 // Offsets into the thread 4273 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4274 DirtyCardQueue::byte_offset_of_index()); 4275 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4276 DirtyCardQueue::byte_offset_of_buf()); 4277 4278 // Pointers into the thread 4279 4280 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4281 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4282 4283 // Now some values 4284 // Use ctrl to avoid hoisting these values past a safepoint, which could 4285 // potentially reset these fields in the JavaThread. 4286 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4287 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4288 4289 // Convert the store obj pointer to an int prior to doing math on it 4290 // Must use ctrl to prevent "integerized oop" existing across safepoint 4291 Node* cast = __ CastPX(__ ctrl(), adr); 4292 4293 // Divide pointer by card size 4294 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4295 4296 // Combine card table base and card offset 4297 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4298 4299 // If we know the value being stored does it cross regions? 4300 4301 if (val != NULL) { 4302 // Does the store cause us to cross regions? 4303 4304 // Should be able to do an unsigned compare of region_size instead of 4305 // and extra shift. Do we have an unsigned compare?? 4306 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4307 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4308 4309 // if (xor_res == 0) same region so skip 4310 __ if_then(xor_res, BoolTest::ne, zeroX); { 4311 4312 // No barrier if we are storing a NULL 4313 __ if_then(val, BoolTest::ne, null(), unlikely); { 4314 4315 // Ok must mark the card if not already dirty 4316 4317 // load the original value of the card 4318 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4319 4320 __ if_then(card_val, BoolTest::ne, young_card); { 4321 sync_kit(ideal); 4322 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4323 insert_mem_bar(Op_MemBarVolatile, oop_store); 4324 __ sync_kit(this); 4325 4326 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4327 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4328 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4329 } __ end_if(); 4330 } __ end_if(); 4331 } __ end_if(); 4332 } __ end_if(); 4333 } else { 4334 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4335 // We don't need a barrier here if the destination is a newly allocated object 4336 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4337 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4338 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4339 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4340 __ if_then(card_val, BoolTest::ne, young_card); { 4341 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4342 } __ end_if(); 4343 } 4344 4345 // Final sync IdealKit and GraphKit. 4346 final_sync(ideal); 4347 } 4348 #undef __ 4349 4350 4351 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4352 Node* len = load_array_length(load_String_value(ctrl, str)); 4353 Node* coder = load_String_coder(ctrl, str); 4354 // Divide length by 2 if coder is UTF16 4355 return _gvn.transform(new RShiftINode(len, coder)); 4356 } 4357 4358 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4359 int value_offset = java_lang_String::value_offset_in_bytes(); 4360 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4361 false, NULL, Type::Offset(0)); 4362 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4363 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4364 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4365 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4366 int value_field_idx = C->get_alias_index(value_field_type); 4367 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4368 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4369 // String.value field is known to be @Stable. 4370 if (UseImplicitStableValues) { 4371 load = cast_array_to_stable(load, value_type); 4372 } 4373 return load; 4374 } 4375 4376 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4377 if (!CompactStrings) { 4378 return intcon(java_lang_String::CODER_UTF16); 4379 } 4380 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4381 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4382 false, NULL, Type::Offset(0)); 4383 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4384 int coder_field_idx = C->get_alias_index(coder_field_type); 4385 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4386 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4387 } 4388 4389 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4390 int value_offset = java_lang_String::value_offset_in_bytes(); 4391 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4392 false, NULL, Type::Offset(0)); 4393 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4394 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4395 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4396 } 4397 4398 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4399 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4400 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4401 false, NULL, Type::Offset(0)); 4402 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4403 int coder_field_idx = C->get_alias_index(coder_field_type); 4404 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4405 value, T_BYTE, coder_field_idx, MemNode::unordered); 4406 } 4407 4408 // Capture src and dst memory state with a MergeMemNode 4409 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4410 if (src_type == dst_type) { 4411 // Types are equal, we don't need a MergeMemNode 4412 return memory(src_type); 4413 } 4414 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4415 record_for_igvn(merge); // fold it up later, if possible 4416 int src_idx = C->get_alias_index(src_type); 4417 int dst_idx = C->get_alias_index(dst_type); 4418 merge->set_memory_at(src_idx, memory(src_idx)); 4419 merge->set_memory_at(dst_idx, memory(dst_idx)); 4420 return merge; 4421 } 4422 4423 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4424 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4425 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4426 // If input and output memory types differ, capture both states to preserve 4427 // the dependency between preceding and subsequent loads/stores. 4428 // For example, the following program: 4429 // StoreB 4430 // compress_string 4431 // LoadB 4432 // has this memory graph (use->def): 4433 // LoadB -> compress_string -> CharMem 4434 // ... -> StoreB -> ByteMem 4435 // The intrinsic hides the dependency between LoadB and StoreB, causing 4436 // the load to read from memory not containing the result of the StoreB. 4437 // The correct memory graph should look like this: 4438 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4439 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4440 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4441 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4442 set_memory(res_mem, TypeAryPtr::BYTES); 4443 return str; 4444 } 4445 4446 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4447 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4448 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4449 // Capture src and dst memory (see comment in 'compress_string'). 4450 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4451 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4452 set_memory(_gvn.transform(str), dst_type); 4453 } 4454 4455 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4456 /** 4457 * int i_char = start; 4458 * for (int i_byte = 0; i_byte < count; i_byte++) { 4459 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4460 * } 4461 */ 4462 add_predicate(); 4463 RegionNode* head = new RegionNode(3); 4464 head->init_req(1, control()); 4465 gvn().set_type(head, Type::CONTROL); 4466 record_for_igvn(head); 4467 4468 Node* i_byte = new PhiNode(head, TypeInt::INT); 4469 i_byte->init_req(1, intcon(0)); 4470 gvn().set_type(i_byte, TypeInt::INT); 4471 record_for_igvn(i_byte); 4472 4473 Node* i_char = new PhiNode(head, TypeInt::INT); 4474 i_char->init_req(1, start); 4475 gvn().set_type(i_char, TypeInt::INT); 4476 record_for_igvn(i_char); 4477 4478 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4479 gvn().set_type(mem, Type::MEMORY); 4480 record_for_igvn(mem); 4481 set_control(head); 4482 set_memory(mem, TypeAryPtr::BYTES); 4483 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4484 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4485 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4486 false, false, true /* mismatched */); 4487 4488 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4489 head->init_req(2, IfTrue(iff)); 4490 mem->init_req(2, st); 4491 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4492 i_char->init_req(2, AddI(i_char, intcon(2))); 4493 4494 set_control(IfFalse(iff)); 4495 set_memory(st, TypeAryPtr::BYTES); 4496 } 4497 4498 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4499 if (!field->is_constant()) { 4500 return NULL; // Field not marked as constant. 4501 } 4502 ciInstance* holder = NULL; 4503 if (!field->is_static()) { 4504 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4505 if (const_oop != NULL && const_oop->is_instance()) { 4506 holder = const_oop->as_instance(); 4507 } 4508 } 4509 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4510 /*is_unsigned_load=*/false); 4511 if (con_type != NULL) { 4512 Node* con = makecon(con_type); 4513 if (field->layout_type() == T_VALUETYPE) { 4514 // Load value type from constant oop 4515 con = ValueTypeNode::make(gvn(), map()->memory(), con); 4516 } 4517 return con; 4518 } 4519 return NULL; 4520 } 4521 4522 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4523 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4524 // assumption of CCP analysis. 4525 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4526 }