1 /* 2 * Copyright (c) 2001, 2015, 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 "gc/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc/g1/heapRegion.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/cardTableModRefBS.hpp" 31 #include "gc/shared/collectedHeap.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/intrinsicnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 47 //----------------------------GraphKit----------------------------------------- 48 // Main utility constructor. 49 GraphKit::GraphKit(JVMState* jvms) 50 : Phase(Phase::Parser), 51 _env(C->env()), 52 _gvn(*C->initial_gvn()) 53 { 54 _exceptions = jvms->map()->next_exception(); 55 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 56 set_jvms(jvms); 57 } 58 59 // Private constructor for parser. 60 GraphKit::GraphKit() 61 : Phase(Phase::Parser), 62 _env(C->env()), 63 _gvn(*C->initial_gvn()) 64 { 65 _exceptions = NULL; 66 set_map(NULL); 67 debug_only(_sp = -99); 68 debug_only(set_bci(-99)); 69 } 70 71 72 73 //---------------------------clean_stack--------------------------------------- 74 // Clear away rubbish from the stack area of the JVM state. 75 // This destroys any arguments that may be waiting on the stack. 76 void GraphKit::clean_stack(int from_sp) { 77 SafePointNode* map = this->map(); 78 JVMState* jvms = this->jvms(); 79 int stk_size = jvms->stk_size(); 80 int stkoff = jvms->stkoff(); 81 Node* top = this->top(); 82 for (int i = from_sp; i < stk_size; i++) { 83 if (map->in(stkoff + i) != top) { 84 map->set_req(stkoff + i, top); 85 } 86 } 87 } 88 89 90 //--------------------------------sync_jvms----------------------------------- 91 // Make sure our current jvms agrees with our parse state. 92 JVMState* GraphKit::sync_jvms() const { 93 JVMState* jvms = this->jvms(); 94 jvms->set_bci(bci()); // Record the new bci in the JVMState 95 jvms->set_sp(sp()); // Record the new sp in the JVMState 96 assert(jvms_in_sync(), "jvms is now in sync"); 97 return jvms; 98 } 99 100 //--------------------------------sync_jvms_for_reexecute--------------------- 101 // Make sure our current jvms agrees with our parse state. This version 102 // uses the reexecute_sp for reexecuting bytecodes. 103 JVMState* GraphKit::sync_jvms_for_reexecute() { 104 JVMState* jvms = this->jvms(); 105 jvms->set_bci(bci()); // Record the new bci in the JVMState 106 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 107 return jvms; 108 } 109 110 #ifdef ASSERT 111 bool GraphKit::jvms_in_sync() const { 112 Parse* parse = is_Parse(); 113 if (parse == NULL) { 114 if (bci() != jvms()->bci()) return false; 115 if (sp() != (int)jvms()->sp()) return false; 116 return true; 117 } 118 if (jvms()->method() != parse->method()) return false; 119 if (jvms()->bci() != parse->bci()) return false; 120 int jvms_sp = jvms()->sp(); 121 if (jvms_sp != parse->sp()) return false; 122 int jvms_depth = jvms()->depth(); 123 if (jvms_depth != parse->depth()) return false; 124 return true; 125 } 126 127 // Local helper checks for special internal merge points 128 // used to accumulate and merge exception states. 129 // They are marked by the region's in(0) edge being the map itself. 130 // Such merge points must never "escape" into the parser at large, 131 // until they have been handed to gvn.transform. 132 static bool is_hidden_merge(Node* reg) { 133 if (reg == NULL) return false; 134 if (reg->is_Phi()) { 135 reg = reg->in(0); 136 if (reg == NULL) return false; 137 } 138 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 139 } 140 141 void GraphKit::verify_map() const { 142 if (map() == NULL) return; // null map is OK 143 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 144 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 145 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 146 } 147 148 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 149 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 150 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 151 } 152 #endif 153 154 //---------------------------stop_and_kill_map--------------------------------- 155 // Set _map to NULL, signalling a stop to further bytecode execution. 156 // First smash the current map's control to a constant, to mark it dead. 157 void GraphKit::stop_and_kill_map() { 158 SafePointNode* dead_map = stop(); 159 if (dead_map != NULL) { 160 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 161 assert(dead_map->is_killed(), "must be so marked"); 162 } 163 } 164 165 166 //--------------------------------stopped-------------------------------------- 167 // Tell if _map is NULL, or control is top. 168 bool GraphKit::stopped() { 169 if (map() == NULL) return true; 170 else if (control() == top()) return true; 171 else return false; 172 } 173 174 175 //-----------------------------has_ex_handler---------------------------------- 176 // Tell if this method or any caller method has exception handlers. 177 bool GraphKit::has_ex_handler() { 178 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 179 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 180 return true; 181 } 182 } 183 return false; 184 } 185 186 //------------------------------save_ex_oop------------------------------------ 187 // Save an exception without blowing stack contents or other JVM state. 188 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 189 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 190 ex_map->add_req(ex_oop); 191 debug_only(verify_exception_state(ex_map)); 192 } 193 194 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 195 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 196 Node* ex_oop = ex_map->in(ex_map->req()-1); 197 if (clear_it) ex_map->del_req(ex_map->req()-1); 198 return ex_oop; 199 } 200 201 //-----------------------------saved_ex_oop------------------------------------ 202 // Recover a saved exception from its map. 203 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 204 return common_saved_ex_oop(ex_map, false); 205 } 206 207 //--------------------------clear_saved_ex_oop--------------------------------- 208 // Erase a previously saved exception from its map. 209 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, true); 211 } 212 213 #ifdef ASSERT 214 //---------------------------has_saved_ex_oop---------------------------------- 215 // Erase a previously saved exception from its map. 216 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 217 return ex_map->req() == ex_map->jvms()->endoff()+1; 218 } 219 #endif 220 221 //-------------------------make_exception_state-------------------------------- 222 // Turn the current JVM state into an exception state, appending the ex_oop. 223 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 224 sync_jvms(); 225 SafePointNode* ex_map = stop(); // do not manipulate this map any more 226 set_saved_ex_oop(ex_map, ex_oop); 227 return ex_map; 228 } 229 230 231 //--------------------------add_exception_state-------------------------------- 232 // Add an exception to my list of exceptions. 233 void GraphKit::add_exception_state(SafePointNode* ex_map) { 234 if (ex_map == NULL || ex_map->control() == top()) { 235 return; 236 } 237 #ifdef ASSERT 238 verify_exception_state(ex_map); 239 if (has_exceptions()) { 240 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 241 } 242 #endif 243 244 // If there is already an exception of exactly this type, merge with it. 245 // In particular, null-checks and other low-level exceptions common up here. 246 Node* ex_oop = saved_ex_oop(ex_map); 247 const Type* ex_type = _gvn.type(ex_oop); 248 if (ex_oop == top()) { 249 // No action needed. 250 return; 251 } 252 assert(ex_type->isa_instptr(), "exception must be an instance"); 253 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 254 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 255 // We check sp also because call bytecodes can generate exceptions 256 // both before and after arguments are popped! 257 if (ex_type2 == ex_type 258 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 259 combine_exception_states(ex_map, e2); 260 return; 261 } 262 } 263 264 // No pre-existing exception of the same type. Chain it on the list. 265 push_exception_state(ex_map); 266 } 267 268 //-----------------------add_exception_states_from----------------------------- 269 void GraphKit::add_exception_states_from(JVMState* jvms) { 270 SafePointNode* ex_map = jvms->map()->next_exception(); 271 if (ex_map != NULL) { 272 jvms->map()->set_next_exception(NULL); 273 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 274 next_map = ex_map->next_exception(); 275 ex_map->set_next_exception(NULL); 276 add_exception_state(ex_map); 277 } 278 } 279 } 280 281 //-----------------------transfer_exceptions_into_jvms------------------------- 282 JVMState* GraphKit::transfer_exceptions_into_jvms() { 283 if (map() == NULL) { 284 // We need a JVMS to carry the exceptions, but the map has gone away. 285 // Create a scratch JVMS, cloned from any of the exception states... 286 if (has_exceptions()) { 287 _map = _exceptions; 288 _map = clone_map(); 289 _map->set_next_exception(NULL); 290 clear_saved_ex_oop(_map); 291 debug_only(verify_map()); 292 } else { 293 // ...or created from scratch 294 JVMState* jvms = new (C) JVMState(_method, NULL); 295 jvms->set_bci(_bci); 296 jvms->set_sp(_sp); 297 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 298 set_jvms(jvms); 299 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 300 set_all_memory(top()); 301 while (map()->req() < jvms->endoff()) map()->add_req(top()); 302 } 303 // (This is a kludge, in case you didn't notice.) 304 set_control(top()); 305 } 306 JVMState* jvms = sync_jvms(); 307 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 308 jvms->map()->set_next_exception(_exceptions); 309 _exceptions = NULL; // done with this set of exceptions 310 return jvms; 311 } 312 313 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 314 assert(is_hidden_merge(dstphi), "must be a special merge node"); 315 assert(is_hidden_merge(srcphi), "must be a special merge node"); 316 uint limit = srcphi->req(); 317 for (uint i = PhiNode::Input; i < limit; i++) { 318 dstphi->add_req(srcphi->in(i)); 319 } 320 } 321 static inline void add_one_req(Node* dstphi, Node* src) { 322 assert(is_hidden_merge(dstphi), "must be a special merge node"); 323 assert(!is_hidden_merge(src), "must not be a special merge node"); 324 dstphi->add_req(src); 325 } 326 327 //-----------------------combine_exception_states------------------------------ 328 // This helper function combines exception states by building phis on a 329 // specially marked state-merging region. These regions and phis are 330 // untransformed, and can build up gradually. The region is marked by 331 // having a control input of its exception map, rather than NULL. Such 332 // regions do not appear except in this function, and in use_exception_state. 333 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 334 if (failing()) return; // dying anyway... 335 JVMState* ex_jvms = ex_map->_jvms; 336 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 337 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 338 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 339 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 340 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 341 assert(ex_map->req() == phi_map->req(), "matching maps"); 342 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 343 Node* hidden_merge_mark = root(); 344 Node* region = phi_map->control(); 345 MergeMemNode* phi_mem = phi_map->merged_memory(); 346 MergeMemNode* ex_mem = ex_map->merged_memory(); 347 if (region->in(0) != hidden_merge_mark) { 348 // The control input is not (yet) a specially-marked region in phi_map. 349 // Make it so, and build some phis. 350 region = new RegionNode(2); 351 _gvn.set_type(region, Type::CONTROL); 352 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 353 region->init_req(1, phi_map->control()); 354 phi_map->set_control(region); 355 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 356 record_for_igvn(io_phi); 357 _gvn.set_type(io_phi, Type::ABIO); 358 phi_map->set_i_o(io_phi); 359 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 360 Node* m = mms.memory(); 361 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 362 record_for_igvn(m_phi); 363 _gvn.set_type(m_phi, Type::MEMORY); 364 mms.set_memory(m_phi); 365 } 366 } 367 368 // Either or both of phi_map and ex_map might already be converted into phis. 369 Node* ex_control = ex_map->control(); 370 // if there is special marking on ex_map also, we add multiple edges from src 371 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 372 // how wide was the destination phi_map, originally? 373 uint orig_width = region->req(); 374 375 if (add_multiple) { 376 add_n_reqs(region, ex_control); 377 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 378 } else { 379 // ex_map has no merges, so we just add single edges everywhere 380 add_one_req(region, ex_control); 381 add_one_req(phi_map->i_o(), ex_map->i_o()); 382 } 383 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 384 if (mms.is_empty()) { 385 // get a copy of the base memory, and patch some inputs into it 386 const TypePtr* adr_type = mms.adr_type(C); 387 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 388 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 389 mms.set_memory(phi); 390 // Prepare to append interesting stuff onto the newly sliced phi: 391 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 392 } 393 // Append stuff from ex_map: 394 if (add_multiple) { 395 add_n_reqs(mms.memory(), mms.memory2()); 396 } else { 397 add_one_req(mms.memory(), mms.memory2()); 398 } 399 } 400 uint limit = ex_map->req(); 401 for (uint i = TypeFunc::Parms; i < limit; i++) { 402 // Skip everything in the JVMS after tos. (The ex_oop follows.) 403 if (i == tos) i = ex_jvms->monoff(); 404 Node* src = ex_map->in(i); 405 Node* dst = phi_map->in(i); 406 if (src != dst) { 407 PhiNode* phi; 408 if (dst->in(0) != region) { 409 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 410 record_for_igvn(phi); 411 _gvn.set_type(phi, phi->type()); 412 phi_map->set_req(i, dst); 413 // Prepare to append interesting stuff onto the new phi: 414 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 415 } else { 416 assert(dst->is_Phi(), "nobody else uses a hidden region"); 417 phi = dst->as_Phi(); 418 } 419 if (add_multiple && src->in(0) == ex_control) { 420 // Both are phis. 421 add_n_reqs(dst, src); 422 } else { 423 while (dst->req() < region->req()) add_one_req(dst, src); 424 } 425 const Type* srctype = _gvn.type(src); 426 if (phi->type() != srctype) { 427 const Type* dsttype = phi->type()->meet_speculative(srctype); 428 if (phi->type() != dsttype) { 429 phi->set_type(dsttype); 430 _gvn.set_type(phi, dsttype); 431 } 432 } 433 } 434 } 435 phi_map->merge_replaced_nodes_with(ex_map); 436 } 437 438 //--------------------------use_exception_state-------------------------------- 439 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 440 if (failing()) { stop(); return top(); } 441 Node* region = phi_map->control(); 442 Node* hidden_merge_mark = root(); 443 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 444 Node* ex_oop = clear_saved_ex_oop(phi_map); 445 if (region->in(0) == hidden_merge_mark) { 446 // Special marking for internal ex-states. Process the phis now. 447 region->set_req(0, region); // now it's an ordinary region 448 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 449 // Note: Setting the jvms also sets the bci and sp. 450 set_control(_gvn.transform(region)); 451 uint tos = jvms()->stkoff() + sp(); 452 for (uint i = 1; i < tos; i++) { 453 Node* x = phi_map->in(i); 454 if (x->in(0) == region) { 455 assert(x->is_Phi(), "expected a special phi"); 456 phi_map->set_req(i, _gvn.transform(x)); 457 } 458 } 459 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 460 Node* x = mms.memory(); 461 if (x->in(0) == region) { 462 assert(x->is_Phi(), "nobody else uses a hidden region"); 463 mms.set_memory(_gvn.transform(x)); 464 } 465 } 466 if (ex_oop->in(0) == region) { 467 assert(ex_oop->is_Phi(), "expected a special phi"); 468 ex_oop = _gvn.transform(ex_oop); 469 } 470 } else { 471 set_jvms(phi_map->jvms()); 472 } 473 474 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 475 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 476 return ex_oop; 477 } 478 479 //---------------------------------java_bc------------------------------------- 480 Bytecodes::Code GraphKit::java_bc() const { 481 ciMethod* method = this->method(); 482 int bci = this->bci(); 483 if (method != NULL && bci != InvocationEntryBci) 484 return method->java_code_at_bci(bci); 485 else 486 return Bytecodes::_illegal; 487 } 488 489 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 490 bool must_throw) { 491 // if the exception capability is set, then we will generate code 492 // to check the JavaThread.should_post_on_exceptions flag to see 493 // if we actually need to report exception events (for this 494 // thread). If we don't need to report exception events, we will 495 // take the normal fast path provided by add_exception_events. If 496 // exception event reporting is enabled for this thread, we will 497 // take the uncommon_trap in the BuildCutout below. 498 499 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 500 Node* jthread = _gvn.transform(new ThreadLocalNode()); 501 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 502 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 503 504 // Test the should_post_on_exceptions_flag vs. 0 505 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 506 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 507 508 // Branch to slow_path if should_post_on_exceptions_flag was true 509 { BuildCutout unless(this, tst, PROB_MAX); 510 // Do not try anything fancy if we're notifying the VM on every throw. 511 // Cf. case Bytecodes::_athrow in parse2.cpp. 512 uncommon_trap(reason, Deoptimization::Action_none, 513 (ciKlass*)NULL, (char*)NULL, must_throw); 514 } 515 516 } 517 518 //------------------------------builtin_throw---------------------------------- 519 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 520 bool must_throw = true; 521 522 if (env()->jvmti_can_post_on_exceptions()) { 523 // check if we must post exception events, take uncommon trap if so 524 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 525 // here if should_post_on_exceptions is false 526 // continue on with the normal codegen 527 } 528 529 // If this particular condition has not yet happened at this 530 // bytecode, then use the uncommon trap mechanism, and allow for 531 // a future recompilation if several traps occur here. 532 // If the throw is hot, try to use a more complicated inline mechanism 533 // which keeps execution inside the compiled code. 534 bool treat_throw_as_hot = false; 535 ciMethodData* md = method()->method_data(); 536 537 if (ProfileTraps) { 538 if (too_many_traps(reason)) { 539 treat_throw_as_hot = true; 540 } 541 // (If there is no MDO at all, assume it is early in 542 // execution, and that any deopts are part of the 543 // startup transient, and don't need to be remembered.) 544 545 // Also, if there is a local exception handler, treat all throws 546 // as hot if there has been at least one in this method. 547 if (C->trap_count(reason) != 0 548 && method()->method_data()->trap_count(reason) != 0 549 && has_ex_handler()) { 550 treat_throw_as_hot = true; 551 } 552 } 553 554 // If this throw happens frequently, an uncommon trap might cause 555 // a performance pothole. If there is a local exception handler, 556 // and if this particular bytecode appears to be deoptimizing often, 557 // let us handle the throw inline, with a preconstructed instance. 558 // Note: If the deopt count has blown up, the uncommon trap 559 // runtime is going to flush this nmethod, not matter what. 560 if (treat_throw_as_hot 561 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 562 // If the throw is local, we use a pre-existing instance and 563 // punt on the backtrace. This would lead to a missing backtrace 564 // (a repeat of 4292742) if the backtrace object is ever asked 565 // for its backtrace. 566 // Fixing this remaining case of 4292742 requires some flavor of 567 // escape analysis. Leave that for the future. 568 ciInstance* ex_obj = NULL; 569 switch (reason) { 570 case Deoptimization::Reason_null_check: 571 ex_obj = env()->NullPointerException_instance(); 572 break; 573 case Deoptimization::Reason_div0_check: 574 ex_obj = env()->ArithmeticException_instance(); 575 break; 576 case Deoptimization::Reason_range_check: 577 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 578 break; 579 case Deoptimization::Reason_class_check: 580 if (java_bc() == Bytecodes::_aastore) { 581 ex_obj = env()->ArrayStoreException_instance(); 582 } else { 583 ex_obj = env()->ClassCastException_instance(); 584 } 585 break; 586 } 587 if (failing()) { stop(); return; } // exception allocation might fail 588 if (ex_obj != NULL) { 589 // Cheat with a preallocated exception object. 590 if (C->log() != NULL) 591 C->log()->elem("hot_throw preallocated='1' reason='%s'", 592 Deoptimization::trap_reason_name(reason)); 593 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 594 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 595 596 // Clear the detail message of the preallocated exception object. 597 // Weblogic sometimes mutates the detail message of exceptions 598 // using reflection. 599 int offset = java_lang_Throwable::get_detailMessage_offset(); 600 const TypePtr* adr_typ = ex_con->add_offset(offset); 601 602 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 603 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 604 // Conservatively release stores of object references. 605 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 606 607 add_exception_state(make_exception_state(ex_node)); 608 return; 609 } 610 } 611 612 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 613 // It won't be much cheaper than bailing to the interp., since we'll 614 // have to pass up all the debug-info, and the runtime will have to 615 // create the stack trace. 616 617 // Usual case: Bail to interpreter. 618 // Reserve the right to recompile if we haven't seen anything yet. 619 620 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 621 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 622 if (treat_throw_as_hot 623 && (method()->method_data()->trap_recompiled_at(bci(), m) 624 || C->too_many_traps(reason))) { 625 // We cannot afford to take more traps here. Suffer in the interpreter. 626 if (C->log() != NULL) 627 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 628 Deoptimization::trap_reason_name(reason), 629 C->trap_count(reason)); 630 action = Deoptimization::Action_none; 631 } 632 633 // "must_throw" prunes the JVM state to include only the stack, if there 634 // are no local exception handlers. This should cut down on register 635 // allocation time and code size, by drastically reducing the number 636 // of in-edges on the call to the uncommon trap. 637 638 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 639 } 640 641 642 //----------------------------PreserveJVMState--------------------------------- 643 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 644 debug_only(kit->verify_map()); 645 _kit = kit; 646 _map = kit->map(); // preserve the map 647 _sp = kit->sp(); 648 kit->set_map(clone_map ? kit->clone_map() : NULL); 649 #ifdef ASSERT 650 _bci = kit->bci(); 651 Parse* parser = kit->is_Parse(); 652 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 653 _block = block; 654 #endif 655 } 656 PreserveJVMState::~PreserveJVMState() { 657 GraphKit* kit = _kit; 658 #ifdef ASSERT 659 assert(kit->bci() == _bci, "bci must not shift"); 660 Parse* parser = kit->is_Parse(); 661 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 662 assert(block == _block, "block must not shift"); 663 #endif 664 kit->set_map(_map); 665 kit->set_sp(_sp); 666 } 667 668 669 //-----------------------------BuildCutout------------------------------------- 670 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 671 : PreserveJVMState(kit) 672 { 673 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 674 SafePointNode* outer_map = _map; // preserved map is caller's 675 SafePointNode* inner_map = kit->map(); 676 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 677 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 678 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 679 } 680 BuildCutout::~BuildCutout() { 681 GraphKit* kit = _kit; 682 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 683 } 684 685 //---------------------------PreserveReexecuteState---------------------------- 686 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 687 assert(!kit->stopped(), "must call stopped() before"); 688 _kit = kit; 689 _sp = kit->sp(); 690 _reexecute = kit->jvms()->_reexecute; 691 } 692 PreserveReexecuteState::~PreserveReexecuteState() { 693 if (_kit->stopped()) return; 694 _kit->jvms()->_reexecute = _reexecute; 695 _kit->set_sp(_sp); 696 } 697 698 //------------------------------clone_map-------------------------------------- 699 // Implementation of PreserveJVMState 700 // 701 // Only clone_map(...) here. If this function is only used in the 702 // PreserveJVMState class we may want to get rid of this extra 703 // function eventually and do it all there. 704 705 SafePointNode* GraphKit::clone_map() { 706 if (map() == NULL) return NULL; 707 708 // Clone the memory edge first 709 Node* mem = MergeMemNode::make(map()->memory()); 710 gvn().set_type_bottom(mem); 711 712 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 713 JVMState* jvms = this->jvms(); 714 JVMState* clonejvms = jvms->clone_shallow(C); 715 clonemap->set_memory(mem); 716 clonemap->set_jvms(clonejvms); 717 clonejvms->set_map(clonemap); 718 record_for_igvn(clonemap); 719 gvn().set_type_bottom(clonemap); 720 return clonemap; 721 } 722 723 724 //-----------------------------set_map_clone----------------------------------- 725 void GraphKit::set_map_clone(SafePointNode* m) { 726 _map = m; 727 _map = clone_map(); 728 _map->set_next_exception(NULL); 729 debug_only(verify_map()); 730 } 731 732 733 //----------------------------kill_dead_locals--------------------------------- 734 // Detect any locals which are known to be dead, and force them to top. 735 void GraphKit::kill_dead_locals() { 736 // Consult the liveness information for the locals. If any 737 // of them are unused, then they can be replaced by top(). This 738 // should help register allocation time and cut down on the size 739 // of the deoptimization information. 740 741 // This call is made from many of the bytecode handling 742 // subroutines called from the Big Switch in do_one_bytecode. 743 // Every bytecode which might include a slow path is responsible 744 // for killing its dead locals. The more consistent we 745 // are about killing deads, the fewer useless phis will be 746 // constructed for them at various merge points. 747 748 // bci can be -1 (InvocationEntryBci). We return the entry 749 // liveness for the method. 750 751 if (method() == NULL || method()->code_size() == 0) { 752 // We are building a graph for a call to a native method. 753 // All locals are live. 754 return; 755 } 756 757 ResourceMark rm; 758 759 // Consult the liveness information for the locals. If any 760 // of them are unused, then they can be replaced by top(). This 761 // should help register allocation time and cut down on the size 762 // of the deoptimization information. 763 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 764 765 int len = (int)live_locals.size(); 766 assert(len <= jvms()->loc_size(), "too many live locals"); 767 for (int local = 0; local < len; local++) { 768 if (!live_locals.at(local)) { 769 set_local(local, top()); 770 } 771 } 772 } 773 774 #ifdef ASSERT 775 //-------------------------dead_locals_are_killed------------------------------ 776 // Return true if all dead locals are set to top in the map. 777 // Used to assert "clean" debug info at various points. 778 bool GraphKit::dead_locals_are_killed() { 779 if (method() == NULL || method()->code_size() == 0) { 780 // No locals need to be dead, so all is as it should be. 781 return true; 782 } 783 784 // Make sure somebody called kill_dead_locals upstream. 785 ResourceMark rm; 786 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 787 if (jvms->loc_size() == 0) continue; // no locals to consult 788 SafePointNode* map = jvms->map(); 789 ciMethod* method = jvms->method(); 790 int bci = jvms->bci(); 791 if (jvms == this->jvms()) { 792 bci = this->bci(); // it might not yet be synched 793 } 794 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 795 int len = (int)live_locals.size(); 796 if (!live_locals.is_valid() || len == 0) 797 // This method is trivial, or is poisoned by a breakpoint. 798 return true; 799 assert(len == jvms->loc_size(), "live map consistent with locals map"); 800 for (int local = 0; local < len; local++) { 801 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 802 if (PrintMiscellaneous && (Verbose || WizardMode)) { 803 tty->print_cr("Zombie local %d: ", local); 804 jvms->dump(); 805 } 806 return false; 807 } 808 } 809 } 810 return true; 811 } 812 813 #endif //ASSERT 814 815 // Helper function for enforcing certain bytecodes to reexecute if 816 // deoptimization happens 817 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 818 ciMethod* cur_method = jvms->method(); 819 int cur_bci = jvms->bci(); 820 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 821 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 822 return Interpreter::bytecode_should_reexecute(code) || 823 is_anewarray && code == Bytecodes::_multianewarray; 824 // Reexecute _multianewarray bytecode which was replaced with 825 // sequence of [a]newarray. See Parse::do_multianewarray(). 826 // 827 // Note: interpreter should not have it set since this optimization 828 // is limited by dimensions and guarded by flag so in some cases 829 // multianewarray() runtime calls will be generated and 830 // the bytecode should not be reexecutes (stack will not be reset). 831 } else 832 return false; 833 } 834 835 // Helper function for adding JVMState and debug information to node 836 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 837 // Add the safepoint edges to the call (or other safepoint). 838 839 // Make sure dead locals are set to top. This 840 // should help register allocation time and cut down on the size 841 // of the deoptimization information. 842 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 843 844 // Walk the inline list to fill in the correct set of JVMState's 845 // Also fill in the associated edges for each JVMState. 846 847 // If the bytecode needs to be reexecuted we need to put 848 // the arguments back on the stack. 849 const bool should_reexecute = jvms()->should_reexecute(); 850 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 851 852 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 853 // undefined if the bci is different. This is normal for Parse but it 854 // should not happen for LibraryCallKit because only one bci is processed. 855 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 856 "in LibraryCallKit the reexecute bit should not change"); 857 858 // If we are guaranteed to throw, we can prune everything but the 859 // input to the current bytecode. 860 bool can_prune_locals = false; 861 uint stack_slots_not_pruned = 0; 862 int inputs = 0, depth = 0; 863 if (must_throw) { 864 assert(method() == youngest_jvms->method(), "sanity"); 865 if (compute_stack_effects(inputs, depth)) { 866 can_prune_locals = true; 867 stack_slots_not_pruned = inputs; 868 } 869 } 870 871 if (env()->should_retain_local_variables()) { 872 // At any safepoint, this method can get breakpointed, which would 873 // then require an immediate deoptimization. 874 can_prune_locals = false; // do not prune locals 875 stack_slots_not_pruned = 0; 876 } 877 878 // do not scribble on the input jvms 879 JVMState* out_jvms = youngest_jvms->clone_deep(C); 880 call->set_jvms(out_jvms); // Start jvms list for call node 881 882 // For a known set of bytecodes, the interpreter should reexecute them if 883 // deoptimization happens. We set the reexecute state for them here 884 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 885 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 886 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 887 } 888 889 // Presize the call: 890 DEBUG_ONLY(uint non_debug_edges = call->req()); 891 call->add_req_batch(top(), youngest_jvms->debug_depth()); 892 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 893 894 // Set up edges so that the call looks like this: 895 // Call [state:] ctl io mem fptr retadr 896 // [parms:] parm0 ... parmN 897 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 898 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 899 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 900 // Note that caller debug info precedes callee debug info. 901 902 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 903 uint debug_ptr = call->req(); 904 905 // Loop over the map input edges associated with jvms, add them 906 // to the call node, & reset all offsets to match call node array. 907 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 908 uint debug_end = debug_ptr; 909 uint debug_start = debug_ptr - in_jvms->debug_size(); 910 debug_ptr = debug_start; // back up the ptr 911 912 uint p = debug_start; // walks forward in [debug_start, debug_end) 913 uint j, k, l; 914 SafePointNode* in_map = in_jvms->map(); 915 out_jvms->set_map(call); 916 917 if (can_prune_locals) { 918 assert(in_jvms->method() == out_jvms->method(), "sanity"); 919 // If the current throw can reach an exception handler in this JVMS, 920 // then we must keep everything live that can reach that handler. 921 // As a quick and dirty approximation, we look for any handlers at all. 922 if (in_jvms->method()->has_exception_handlers()) { 923 can_prune_locals = false; 924 } 925 } 926 927 // Add the Locals 928 k = in_jvms->locoff(); 929 l = in_jvms->loc_size(); 930 out_jvms->set_locoff(p); 931 if (!can_prune_locals) { 932 for (j = 0; j < l; j++) 933 call->set_req(p++, in_map->in(k+j)); 934 } else { 935 p += l; // already set to top above by add_req_batch 936 } 937 938 // Add the Expression Stack 939 k = in_jvms->stkoff(); 940 l = in_jvms->sp(); 941 out_jvms->set_stkoff(p); 942 if (!can_prune_locals) { 943 for (j = 0; j < l; j++) 944 call->set_req(p++, in_map->in(k+j)); 945 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 946 // Divide stack into {S0,...,S1}, where S0 is set to top. 947 uint s1 = stack_slots_not_pruned; 948 stack_slots_not_pruned = 0; // for next iteration 949 if (s1 > l) s1 = l; 950 uint s0 = l - s1; 951 p += s0; // skip the tops preinstalled by add_req_batch 952 for (j = s0; j < l; j++) 953 call->set_req(p++, in_map->in(k+j)); 954 } else { 955 p += l; // already set to top above by add_req_batch 956 } 957 958 // Add the Monitors 959 k = in_jvms->monoff(); 960 l = in_jvms->mon_size(); 961 out_jvms->set_monoff(p); 962 for (j = 0; j < l; j++) 963 call->set_req(p++, in_map->in(k+j)); 964 965 // Copy any scalar object fields. 966 k = in_jvms->scloff(); 967 l = in_jvms->scl_size(); 968 out_jvms->set_scloff(p); 969 for (j = 0; j < l; j++) 970 call->set_req(p++, in_map->in(k+j)); 971 972 // Finish the new jvms. 973 out_jvms->set_endoff(p); 974 975 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 976 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 977 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 978 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 979 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 980 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 981 982 // Update the two tail pointers in parallel. 983 out_jvms = out_jvms->caller(); 984 in_jvms = in_jvms->caller(); 985 } 986 987 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 988 989 // Test the correctness of JVMState::debug_xxx accessors: 990 assert(call->jvms()->debug_start() == non_debug_edges, ""); 991 assert(call->jvms()->debug_end() == call->req(), ""); 992 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 993 } 994 995 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 996 Bytecodes::Code code = java_bc(); 997 if (code == Bytecodes::_wide) { 998 code = method()->java_code_at_bci(bci() + 1); 999 } 1000 1001 BasicType rtype = T_ILLEGAL; 1002 int rsize = 0; 1003 1004 if (code != Bytecodes::_illegal) { 1005 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1006 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1007 if (rtype < T_CONFLICT) 1008 rsize = type2size[rtype]; 1009 } 1010 1011 switch (code) { 1012 case Bytecodes::_illegal: 1013 return false; 1014 1015 case Bytecodes::_ldc: 1016 case Bytecodes::_ldc_w: 1017 case Bytecodes::_ldc2_w: 1018 inputs = 0; 1019 break; 1020 1021 case Bytecodes::_dup: inputs = 1; break; 1022 case Bytecodes::_dup_x1: inputs = 2; break; 1023 case Bytecodes::_dup_x2: inputs = 3; break; 1024 case Bytecodes::_dup2: inputs = 2; break; 1025 case Bytecodes::_dup2_x1: inputs = 3; break; 1026 case Bytecodes::_dup2_x2: inputs = 4; break; 1027 case Bytecodes::_swap: inputs = 2; break; 1028 case Bytecodes::_arraylength: inputs = 1; break; 1029 1030 case Bytecodes::_getstatic: 1031 case Bytecodes::_putstatic: 1032 case Bytecodes::_getfield: 1033 case Bytecodes::_putfield: 1034 { 1035 bool ignored_will_link; 1036 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1037 int size = field->type()->size(); 1038 bool is_get = (depth >= 0), is_static = (depth & 1); 1039 inputs = (is_static ? 0 : 1); 1040 if (is_get) { 1041 depth = size - inputs; 1042 } else { 1043 inputs += size; // putxxx pops the value from the stack 1044 depth = - inputs; 1045 } 1046 } 1047 break; 1048 1049 case Bytecodes::_invokevirtual: 1050 case Bytecodes::_invokespecial: 1051 case Bytecodes::_invokestatic: 1052 case Bytecodes::_invokedynamic: 1053 case Bytecodes::_invokeinterface: 1054 { 1055 bool ignored_will_link; 1056 ciSignature* declared_signature = NULL; 1057 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1058 assert(declared_signature != NULL, "cannot be null"); 1059 inputs = declared_signature->arg_size_for_bc(code); 1060 int size = declared_signature->return_type()->size(); 1061 depth = size - inputs; 1062 } 1063 break; 1064 1065 case Bytecodes::_multianewarray: 1066 { 1067 ciBytecodeStream iter(method()); 1068 iter.reset_to_bci(bci()); 1069 iter.next(); 1070 inputs = iter.get_dimensions(); 1071 assert(rsize == 1, ""); 1072 depth = rsize - inputs; 1073 } 1074 break; 1075 1076 case Bytecodes::_ireturn: 1077 case Bytecodes::_lreturn: 1078 case Bytecodes::_freturn: 1079 case Bytecodes::_dreturn: 1080 case Bytecodes::_areturn: 1081 assert(rsize = -depth, ""); 1082 inputs = rsize; 1083 break; 1084 1085 case Bytecodes::_jsr: 1086 case Bytecodes::_jsr_w: 1087 inputs = 0; 1088 depth = 1; // S.B. depth=1, not zero 1089 break; 1090 1091 default: 1092 // bytecode produces a typed result 1093 inputs = rsize - depth; 1094 assert(inputs >= 0, ""); 1095 break; 1096 } 1097 1098 #ifdef ASSERT 1099 // spot check 1100 int outputs = depth + inputs; 1101 assert(outputs >= 0, "sanity"); 1102 switch (code) { 1103 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1104 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1105 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1106 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1107 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1108 } 1109 #endif //ASSERT 1110 1111 return true; 1112 } 1113 1114 1115 1116 //------------------------------basic_plus_adr--------------------------------- 1117 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1118 // short-circuit a common case 1119 if (offset == intcon(0)) return ptr; 1120 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1121 } 1122 1123 Node* GraphKit::ConvI2L(Node* offset) { 1124 // short-circuit a common case 1125 jint offset_con = find_int_con(offset, Type::OffsetBot); 1126 if (offset_con != Type::OffsetBot) { 1127 return longcon((jlong) offset_con); 1128 } 1129 return _gvn.transform( new ConvI2LNode(offset)); 1130 } 1131 1132 Node* GraphKit::ConvI2UL(Node* offset) { 1133 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1134 if (offset_con != (juint) Type::OffsetBot) { 1135 return longcon((julong) offset_con); 1136 } 1137 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1138 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1139 return _gvn.transform( new AndLNode(conv, mask) ); 1140 } 1141 1142 Node* GraphKit::ConvL2I(Node* offset) { 1143 // short-circuit a common case 1144 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1145 if (offset_con != (jlong)Type::OffsetBot) { 1146 return intcon((int) offset_con); 1147 } 1148 return _gvn.transform( new ConvL2INode(offset)); 1149 } 1150 1151 //-------------------------load_object_klass----------------------------------- 1152 Node* GraphKit::load_object_klass(Node* obj) { 1153 // Special-case a fresh allocation to avoid building nodes: 1154 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1155 if (akls != NULL) return akls; 1156 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1157 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1158 } 1159 1160 //-------------------------load_array_length----------------------------------- 1161 Node* GraphKit::load_array_length(Node* array) { 1162 // Special-case a fresh allocation to avoid building nodes: 1163 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1164 Node *alen; 1165 if (alloc == NULL) { 1166 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1167 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1168 } else { 1169 alen = alloc->Ideal_length(); 1170 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1171 if (ccast != alen) { 1172 alen = _gvn.transform(ccast); 1173 } 1174 } 1175 return alen; 1176 } 1177 1178 //------------------------------do_null_check---------------------------------- 1179 // Helper function to do a NULL pointer check. Returned value is 1180 // the incoming address with NULL casted away. You are allowed to use the 1181 // not-null value only if you are control dependent on the test. 1182 extern int explicit_null_checks_inserted, 1183 explicit_null_checks_elided; 1184 Node* GraphKit::null_check_common(Node* value, BasicType type, 1185 // optional arguments for variations: 1186 bool assert_null, 1187 Node* *null_control, 1188 bool speculative) { 1189 assert(!assert_null || null_control == NULL, "not both at once"); 1190 if (stopped()) return top(); 1191 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1192 // For some performance testing, we may wish to suppress null checking. 1193 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1194 return value; 1195 } 1196 explicit_null_checks_inserted++; 1197 1198 // Construct NULL check 1199 Node *chk = NULL; 1200 switch(type) { 1201 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1202 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1203 case T_ARRAY : // fall through 1204 type = T_OBJECT; // simplify further tests 1205 case T_OBJECT : { 1206 const Type *t = _gvn.type( value ); 1207 1208 const TypeOopPtr* tp = t->isa_oopptr(); 1209 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1210 // Only for do_null_check, not any of its siblings: 1211 && !assert_null && null_control == NULL) { 1212 // Usually, any field access or invocation on an unloaded oop type 1213 // will simply fail to link, since the statically linked class is 1214 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1215 // the static class is loaded but the sharper oop type is not. 1216 // Rather than checking for this obscure case in lots of places, 1217 // we simply observe that a null check on an unloaded class 1218 // will always be followed by a nonsense operation, so we 1219 // can just issue the uncommon trap here. 1220 // Our access to the unloaded class will only be correct 1221 // after it has been loaded and initialized, which requires 1222 // a trip through the interpreter. 1223 #ifndef PRODUCT 1224 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1225 #endif 1226 uncommon_trap(Deoptimization::Reason_unloaded, 1227 Deoptimization::Action_reinterpret, 1228 tp->klass(), "!loaded"); 1229 return top(); 1230 } 1231 1232 if (assert_null) { 1233 // See if the type is contained in NULL_PTR. 1234 // If so, then the value is already null. 1235 if (t->higher_equal(TypePtr::NULL_PTR)) { 1236 explicit_null_checks_elided++; 1237 return value; // Elided null assert quickly! 1238 } 1239 } else { 1240 // See if mixing in the NULL pointer changes type. 1241 // If so, then the NULL pointer was not allowed in the original 1242 // type. In other words, "value" was not-null. 1243 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1244 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1245 explicit_null_checks_elided++; 1246 return value; // Elided null check quickly! 1247 } 1248 } 1249 chk = new CmpPNode( value, null() ); 1250 break; 1251 } 1252 1253 default: 1254 fatal("unexpected type: %s", type2name(type)); 1255 } 1256 assert(chk != NULL, "sanity check"); 1257 chk = _gvn.transform(chk); 1258 1259 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1260 BoolNode *btst = new BoolNode( chk, btest); 1261 Node *tst = _gvn.transform( btst ); 1262 1263 //----------- 1264 // if peephole optimizations occurred, a prior test existed. 1265 // If a prior test existed, maybe it dominates as we can avoid this test. 1266 if (tst != btst && type == T_OBJECT) { 1267 // At this point we want to scan up the CFG to see if we can 1268 // find an identical test (and so avoid this test altogether). 1269 Node *cfg = control(); 1270 int depth = 0; 1271 while( depth < 16 ) { // Limit search depth for speed 1272 if( cfg->Opcode() == Op_IfTrue && 1273 cfg->in(0)->in(1) == tst ) { 1274 // Found prior test. Use "cast_not_null" to construct an identical 1275 // CastPP (and hence hash to) as already exists for the prior test. 1276 // Return that casted value. 1277 if (assert_null) { 1278 replace_in_map(value, null()); 1279 return null(); // do not issue the redundant test 1280 } 1281 Node *oldcontrol = control(); 1282 set_control(cfg); 1283 Node *res = cast_not_null(value); 1284 set_control(oldcontrol); 1285 explicit_null_checks_elided++; 1286 return res; 1287 } 1288 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1289 if (cfg == NULL) break; // Quit at region nodes 1290 depth++; 1291 } 1292 } 1293 1294 //----------- 1295 // Branch to failure if null 1296 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1297 Deoptimization::DeoptReason reason; 1298 if (assert_null) { 1299 reason = Deoptimization::Reason_null_assert; 1300 } else if (type == T_OBJECT) { 1301 reason = Deoptimization::reason_null_check(speculative); 1302 } else { 1303 reason = Deoptimization::Reason_div0_check; 1304 } 1305 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1306 // ciMethodData::has_trap_at will return a conservative -1 if any 1307 // must-be-null assertion has failed. This could cause performance 1308 // problems for a method after its first do_null_assert failure. 1309 // Consider using 'Reason_class_check' instead? 1310 1311 // To cause an implicit null check, we set the not-null probability 1312 // to the maximum (PROB_MAX). For an explicit check the probability 1313 // is set to a smaller value. 1314 if (null_control != NULL || too_many_traps(reason)) { 1315 // probability is less likely 1316 ok_prob = PROB_LIKELY_MAG(3); 1317 } else if (!assert_null && 1318 (ImplicitNullCheckThreshold > 0) && 1319 method() != NULL && 1320 (method()->method_data()->trap_count(reason) 1321 >= (uint)ImplicitNullCheckThreshold)) { 1322 ok_prob = PROB_LIKELY_MAG(3); 1323 } 1324 1325 if (null_control != NULL) { 1326 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1327 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1328 set_control( _gvn.transform( new IfTrueNode(iff))); 1329 if (null_true == top()) 1330 explicit_null_checks_elided++; 1331 (*null_control) = null_true; 1332 } else { 1333 BuildCutout unless(this, tst, ok_prob); 1334 // Check for optimizer eliding test at parse time 1335 if (stopped()) { 1336 // Failure not possible; do not bother making uncommon trap. 1337 explicit_null_checks_elided++; 1338 } else if (assert_null) { 1339 uncommon_trap(reason, 1340 Deoptimization::Action_make_not_entrant, 1341 NULL, "assert_null"); 1342 } else { 1343 replace_in_map(value, zerocon(type)); 1344 builtin_throw(reason); 1345 } 1346 } 1347 1348 // Must throw exception, fall-thru not possible? 1349 if (stopped()) { 1350 return top(); // No result 1351 } 1352 1353 if (assert_null) { 1354 // Cast obj to null on this path. 1355 replace_in_map(value, zerocon(type)); 1356 return zerocon(type); 1357 } 1358 1359 // Cast obj to not-null on this path, if there is no null_control. 1360 // (If there is a null_control, a non-null value may come back to haunt us.) 1361 if (type == T_OBJECT) { 1362 Node* cast = cast_not_null(value, false); 1363 if (null_control == NULL || (*null_control) == top()) 1364 replace_in_map(value, cast); 1365 value = cast; 1366 } 1367 1368 return value; 1369 } 1370 1371 1372 //------------------------------cast_not_null---------------------------------- 1373 // Cast obj to not-null on this path 1374 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1375 const Type *t = _gvn.type(obj); 1376 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1377 // Object is already not-null? 1378 if( t == t_not_null ) return obj; 1379 1380 Node *cast = new CastPPNode(obj,t_not_null); 1381 cast->init_req(0, control()); 1382 cast = _gvn.transform( cast ); 1383 1384 // Scan for instances of 'obj' in the current JVM mapping. 1385 // These instances are known to be not-null after the test. 1386 if (do_replace_in_map) 1387 replace_in_map(obj, cast); 1388 1389 return cast; // Return casted value 1390 } 1391 1392 1393 //--------------------------replace_in_map------------------------------------- 1394 void GraphKit::replace_in_map(Node* old, Node* neww) { 1395 if (old == neww) { 1396 return; 1397 } 1398 1399 map()->replace_edge(old, neww); 1400 1401 // Note: This operation potentially replaces any edge 1402 // on the map. This includes locals, stack, and monitors 1403 // of the current (innermost) JVM state. 1404 1405 // don't let inconsistent types from profiling escape this 1406 // method 1407 1408 const Type* told = _gvn.type(old); 1409 const Type* tnew = _gvn.type(neww); 1410 1411 if (!tnew->higher_equal(told)) { 1412 return; 1413 } 1414 1415 map()->record_replaced_node(old, neww); 1416 } 1417 1418 1419 //============================================================================= 1420 //--------------------------------memory--------------------------------------- 1421 Node* GraphKit::memory(uint alias_idx) { 1422 MergeMemNode* mem = merged_memory(); 1423 Node* p = mem->memory_at(alias_idx); 1424 _gvn.set_type(p, Type::MEMORY); // must be mapped 1425 return p; 1426 } 1427 1428 //-----------------------------reset_memory------------------------------------ 1429 Node* GraphKit::reset_memory() { 1430 Node* mem = map()->memory(); 1431 // do not use this node for any more parsing! 1432 debug_only( map()->set_memory((Node*)NULL) ); 1433 return _gvn.transform( mem ); 1434 } 1435 1436 //------------------------------set_all_memory--------------------------------- 1437 void GraphKit::set_all_memory(Node* newmem) { 1438 Node* mergemem = MergeMemNode::make(newmem); 1439 gvn().set_type_bottom(mergemem); 1440 map()->set_memory(mergemem); 1441 } 1442 1443 //------------------------------set_all_memory_call---------------------------- 1444 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1445 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1446 set_all_memory(newmem); 1447 } 1448 1449 //============================================================================= 1450 // 1451 // parser factory methods for MemNodes 1452 // 1453 // These are layered on top of the factory methods in LoadNode and StoreNode, 1454 // and integrate with the parser's memory state and _gvn engine. 1455 // 1456 1457 // factory methods in "int adr_idx" 1458 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1459 int adr_idx, 1460 MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency, bool require_atomic_access) { 1461 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1462 const TypePtr* adr_type = NULL; // debug-mode-only argument 1463 debug_only(adr_type = C->get_adr_type(adr_idx)); 1464 Node* mem = memory(adr_idx); 1465 Node* ld; 1466 if (require_atomic_access && bt == T_LONG) { 1467 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency); 1468 } else if (require_atomic_access && bt == T_DOUBLE) { 1469 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency); 1470 } else { 1471 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency); 1472 } 1473 ld = _gvn.transform(ld); 1474 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1475 // Improve graph before escape analysis and boxing elimination. 1476 record_for_igvn(ld); 1477 } 1478 return ld; 1479 } 1480 1481 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1482 int adr_idx, 1483 MemNode::MemOrd mo, 1484 bool require_atomic_access) { 1485 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1486 const TypePtr* adr_type = NULL; 1487 debug_only(adr_type = C->get_adr_type(adr_idx)); 1488 Node *mem = memory(adr_idx); 1489 Node* st; 1490 if (require_atomic_access && bt == T_LONG) { 1491 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1492 } else if (require_atomic_access && bt == T_DOUBLE) { 1493 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1494 } else { 1495 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1496 } 1497 st = _gvn.transform(st); 1498 set_memory(st, adr_idx); 1499 // Back-to-back stores can only remove intermediate store with DU info 1500 // so push on worklist for optimizer. 1501 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1502 record_for_igvn(st); 1503 1504 return st; 1505 } 1506 1507 1508 void GraphKit::pre_barrier(bool do_load, 1509 Node* ctl, 1510 Node* obj, 1511 Node* adr, 1512 uint adr_idx, 1513 Node* val, 1514 const TypeOopPtr* val_type, 1515 Node* pre_val, 1516 BasicType bt) { 1517 1518 BarrierSet* bs = Universe::heap()->barrier_set(); 1519 set_control(ctl); 1520 switch (bs->kind()) { 1521 case BarrierSet::G1SATBCTLogging: 1522 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1523 break; 1524 1525 case BarrierSet::CardTableForRS: 1526 case BarrierSet::CardTableExtension: 1527 case BarrierSet::ModRef: 1528 break; 1529 1530 default : 1531 ShouldNotReachHere(); 1532 1533 } 1534 } 1535 1536 bool GraphKit::can_move_pre_barrier() const { 1537 BarrierSet* bs = Universe::heap()->barrier_set(); 1538 switch (bs->kind()) { 1539 case BarrierSet::G1SATBCTLogging: 1540 return true; // Can move it if no safepoint 1541 1542 case BarrierSet::CardTableForRS: 1543 case BarrierSet::CardTableExtension: 1544 case BarrierSet::ModRef: 1545 return true; // There is no pre-barrier 1546 1547 default : 1548 ShouldNotReachHere(); 1549 } 1550 return false; 1551 } 1552 1553 void GraphKit::post_barrier(Node* ctl, 1554 Node* store, 1555 Node* obj, 1556 Node* adr, 1557 uint adr_idx, 1558 Node* val, 1559 BasicType bt, 1560 bool use_precise) { 1561 BarrierSet* bs = Universe::heap()->barrier_set(); 1562 set_control(ctl); 1563 switch (bs->kind()) { 1564 case BarrierSet::G1SATBCTLogging: 1565 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1566 break; 1567 1568 case BarrierSet::CardTableForRS: 1569 case BarrierSet::CardTableExtension: 1570 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1571 break; 1572 1573 case BarrierSet::ModRef: 1574 break; 1575 1576 default : 1577 ShouldNotReachHere(); 1578 1579 } 1580 } 1581 1582 Node* GraphKit::store_oop(Node* ctl, 1583 Node* obj, 1584 Node* adr, 1585 const TypePtr* adr_type, 1586 Node* val, 1587 const TypeOopPtr* val_type, 1588 BasicType bt, 1589 bool use_precise, 1590 MemNode::MemOrd mo) { 1591 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1592 // could be delayed during Parse (for example, in adjust_map_after_if()). 1593 // Execute transformation here to avoid barrier generation in such case. 1594 if (_gvn.type(val) == TypePtr::NULL_PTR) 1595 val = _gvn.makecon(TypePtr::NULL_PTR); 1596 1597 set_control(ctl); 1598 if (stopped()) return top(); // Dead path ? 1599 1600 assert(bt == T_OBJECT, "sanity"); 1601 assert(val != NULL, "not dead path"); 1602 uint adr_idx = C->get_alias_index(adr_type); 1603 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1604 1605 pre_barrier(true /* do_load */, 1606 control(), obj, adr, adr_idx, val, val_type, 1607 NULL /* pre_val */, 1608 bt); 1609 1610 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1611 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1612 return store; 1613 } 1614 1615 // Could be an array or object we don't know at compile time (unsafe ref.) 1616 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1617 Node* obj, // containing obj 1618 Node* adr, // actual adress to store val at 1619 const TypePtr* adr_type, 1620 Node* val, 1621 BasicType bt, 1622 MemNode::MemOrd mo) { 1623 Compile::AliasType* at = C->alias_type(adr_type); 1624 const TypeOopPtr* val_type = NULL; 1625 if (adr_type->isa_instptr()) { 1626 if (at->field() != NULL) { 1627 // known field. This code is a copy of the do_put_xxx logic. 1628 ciField* field = at->field(); 1629 if (!field->type()->is_loaded()) { 1630 val_type = TypeInstPtr::BOTTOM; 1631 } else { 1632 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1633 } 1634 } 1635 } else if (adr_type->isa_aryptr()) { 1636 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1637 } 1638 if (val_type == NULL) { 1639 val_type = TypeInstPtr::BOTTOM; 1640 } 1641 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1642 } 1643 1644 1645 //-------------------------array_element_address------------------------- 1646 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1647 const TypeInt* sizetype) { 1648 uint shift = exact_log2(type2aelembytes(elembt)); 1649 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1650 1651 // short-circuit a common case (saves lots of confusing waste motion) 1652 jint idx_con = find_int_con(idx, -1); 1653 if (idx_con >= 0) { 1654 intptr_t offset = header + ((intptr_t)idx_con << shift); 1655 return basic_plus_adr(ary, offset); 1656 } 1657 1658 // must be correct type for alignment purposes 1659 Node* base = basic_plus_adr(ary, header); 1660 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype); 1661 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1662 return basic_plus_adr(ary, base, scale); 1663 } 1664 1665 //-------------------------load_array_element------------------------- 1666 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1667 const Type* elemtype = arytype->elem(); 1668 BasicType elembt = elemtype->array_element_basic_type(); 1669 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1670 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1671 return ld; 1672 } 1673 1674 //-------------------------set_arguments_for_java_call------------------------- 1675 // Arguments (pre-popped from the stack) are taken from the JVMS. 1676 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1677 // Add the call arguments: 1678 uint nargs = call->method()->arg_size(); 1679 for (uint i = 0; i < nargs; i++) { 1680 Node* arg = argument(i); 1681 call->init_req(i + TypeFunc::Parms, arg); 1682 } 1683 } 1684 1685 //---------------------------set_edges_for_java_call--------------------------- 1686 // Connect a newly created call into the current JVMS. 1687 // A return value node (if any) is returned from set_edges_for_java_call. 1688 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1689 1690 // Add the predefined inputs: 1691 call->init_req( TypeFunc::Control, control() ); 1692 call->init_req( TypeFunc::I_O , i_o() ); 1693 call->init_req( TypeFunc::Memory , reset_memory() ); 1694 call->init_req( TypeFunc::FramePtr, frameptr() ); 1695 call->init_req( TypeFunc::ReturnAdr, top() ); 1696 1697 add_safepoint_edges(call, must_throw); 1698 1699 Node* xcall = _gvn.transform(call); 1700 1701 if (xcall == top()) { 1702 set_control(top()); 1703 return; 1704 } 1705 assert(xcall == call, "call identity is stable"); 1706 1707 // Re-use the current map to produce the result. 1708 1709 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1710 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1711 set_all_memory_call(xcall, separate_io_proj); 1712 1713 //return xcall; // no need, caller already has it 1714 } 1715 1716 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1717 if (stopped()) return top(); // maybe the call folded up? 1718 1719 // Capture the return value, if any. 1720 Node* ret; 1721 if (call->method() == NULL || 1722 call->method()->return_type()->basic_type() == T_VOID) 1723 ret = top(); 1724 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1725 1726 // Note: Since any out-of-line call can produce an exception, 1727 // we always insert an I_O projection from the call into the result. 1728 1729 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1730 1731 if (separate_io_proj) { 1732 // The caller requested separate projections be used by the fall 1733 // through and exceptional paths, so replace the projections for 1734 // the fall through path. 1735 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1736 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1737 } 1738 return ret; 1739 } 1740 1741 //--------------------set_predefined_input_for_runtime_call-------------------- 1742 // Reading and setting the memory state is way conservative here. 1743 // The real problem is that I am not doing real Type analysis on memory, 1744 // so I cannot distinguish card mark stores from other stores. Across a GC 1745 // point the Store Barrier and the card mark memory has to agree. I cannot 1746 // have a card mark store and its barrier split across the GC point from 1747 // either above or below. Here I get that to happen by reading ALL of memory. 1748 // A better answer would be to separate out card marks from other memory. 1749 // For now, return the input memory state, so that it can be reused 1750 // after the call, if this call has restricted memory effects. 1751 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1752 // Set fixed predefined input arguments 1753 Node* memory = reset_memory(); 1754 call->init_req( TypeFunc::Control, control() ); 1755 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1756 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1757 call->init_req( TypeFunc::FramePtr, frameptr() ); 1758 call->init_req( TypeFunc::ReturnAdr, top() ); 1759 return memory; 1760 } 1761 1762 //-------------------set_predefined_output_for_runtime_call-------------------- 1763 // Set control and memory (not i_o) from the call. 1764 // If keep_mem is not NULL, use it for the output state, 1765 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1766 // If hook_mem is NULL, this call produces no memory effects at all. 1767 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1768 // then only that memory slice is taken from the call. 1769 // In the last case, we must put an appropriate memory barrier before 1770 // the call, so as to create the correct anti-dependencies on loads 1771 // preceding the call. 1772 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1773 Node* keep_mem, 1774 const TypePtr* hook_mem) { 1775 // no i/o 1776 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1777 if (keep_mem) { 1778 // First clone the existing memory state 1779 set_all_memory(keep_mem); 1780 if (hook_mem != NULL) { 1781 // Make memory for the call 1782 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1783 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1784 // We also use hook_mem to extract specific effects from arraycopy stubs. 1785 set_memory(mem, hook_mem); 1786 } 1787 // ...else the call has NO memory effects. 1788 1789 // Make sure the call advertises its memory effects precisely. 1790 // This lets us build accurate anti-dependences in gcm.cpp. 1791 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1792 "call node must be constructed correctly"); 1793 } else { 1794 assert(hook_mem == NULL, ""); 1795 // This is not a "slow path" call; all memory comes from the call. 1796 set_all_memory_call(call); 1797 } 1798 } 1799 1800 1801 // Replace the call with the current state of the kit. 1802 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1803 JVMState* ejvms = NULL; 1804 if (has_exceptions()) { 1805 ejvms = transfer_exceptions_into_jvms(); 1806 } 1807 1808 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1809 ReplacedNodes replaced_nodes_exception; 1810 Node* ex_ctl = top(); 1811 1812 SafePointNode* final_state = stop(); 1813 1814 // Find all the needed outputs of this call 1815 CallProjections callprojs; 1816 call->extract_projections(&callprojs, true); 1817 1818 Node* init_mem = call->in(TypeFunc::Memory); 1819 Node* final_mem = final_state->in(TypeFunc::Memory); 1820 Node* final_ctl = final_state->in(TypeFunc::Control); 1821 Node* final_io = final_state->in(TypeFunc::I_O); 1822 1823 // Replace all the old call edges with the edges from the inlining result 1824 if (callprojs.fallthrough_catchproj != NULL) { 1825 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1826 } 1827 if (callprojs.fallthrough_memproj != NULL) { 1828 if (final_mem->is_MergeMem()) { 1829 // Parser's exits MergeMem was not transformed but may be optimized 1830 final_mem = _gvn.transform(final_mem); 1831 } 1832 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1833 } 1834 if (callprojs.fallthrough_ioproj != NULL) { 1835 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1836 } 1837 1838 // Replace the result with the new result if it exists and is used 1839 if (callprojs.resproj != NULL && result != NULL) { 1840 C->gvn_replace_by(callprojs.resproj, result); 1841 } 1842 1843 if (ejvms == NULL) { 1844 // No exception edges to simply kill off those paths 1845 if (callprojs.catchall_catchproj != NULL) { 1846 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1847 } 1848 if (callprojs.catchall_memproj != NULL) { 1849 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1850 } 1851 if (callprojs.catchall_ioproj != NULL) { 1852 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1853 } 1854 // Replace the old exception object with top 1855 if (callprojs.exobj != NULL) { 1856 C->gvn_replace_by(callprojs.exobj, C->top()); 1857 } 1858 } else { 1859 GraphKit ekit(ejvms); 1860 1861 // Load my combined exception state into the kit, with all phis transformed: 1862 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1863 replaced_nodes_exception = ex_map->replaced_nodes(); 1864 1865 Node* ex_oop = ekit.use_exception_state(ex_map); 1866 1867 if (callprojs.catchall_catchproj != NULL) { 1868 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1869 ex_ctl = ekit.control(); 1870 } 1871 if (callprojs.catchall_memproj != NULL) { 1872 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1873 } 1874 if (callprojs.catchall_ioproj != NULL) { 1875 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1876 } 1877 1878 // Replace the old exception object with the newly created one 1879 if (callprojs.exobj != NULL) { 1880 C->gvn_replace_by(callprojs.exobj, ex_oop); 1881 } 1882 } 1883 1884 // Disconnect the call from the graph 1885 call->disconnect_inputs(NULL, C); 1886 C->gvn_replace_by(call, C->top()); 1887 1888 // Clean up any MergeMems that feed other MergeMems since the 1889 // optimizer doesn't like that. 1890 if (final_mem->is_MergeMem()) { 1891 Node_List wl; 1892 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1893 Node* m = i.get(); 1894 if (m->is_MergeMem() && !wl.contains(m)) { 1895 wl.push(m); 1896 } 1897 } 1898 while (wl.size() > 0) { 1899 _gvn.transform(wl.pop()); 1900 } 1901 } 1902 1903 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1904 replaced_nodes.apply(C, final_ctl); 1905 } 1906 if (!ex_ctl->is_top() && do_replaced_nodes) { 1907 replaced_nodes_exception.apply(C, ex_ctl); 1908 } 1909 } 1910 1911 1912 //------------------------------increment_counter------------------------------ 1913 // for statistics: increment a VM counter by 1 1914 1915 void GraphKit::increment_counter(address counter_addr) { 1916 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1917 increment_counter(adr1); 1918 } 1919 1920 void GraphKit::increment_counter(Node* counter_addr) { 1921 int adr_type = Compile::AliasIdxRaw; 1922 Node* ctrl = control(); 1923 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1924 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1925 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1926 } 1927 1928 1929 //------------------------------uncommon_trap---------------------------------- 1930 // Bail out to the interpreter in mid-method. Implemented by calling the 1931 // uncommon_trap blob. This helper function inserts a runtime call with the 1932 // right debug info. 1933 void GraphKit::uncommon_trap(int trap_request, 1934 ciKlass* klass, const char* comment, 1935 bool must_throw, 1936 bool keep_exact_action) { 1937 if (failing()) stop(); 1938 if (stopped()) return; // trap reachable? 1939 1940 // Note: If ProfileTraps is true, and if a deopt. actually 1941 // occurs here, the runtime will make sure an MDO exists. There is 1942 // no need to call method()->ensure_method_data() at this point. 1943 1944 // Set the stack pointer to the right value for reexecution: 1945 set_sp(reexecute_sp()); 1946 1947 #ifdef ASSERT 1948 if (!must_throw) { 1949 // Make sure the stack has at least enough depth to execute 1950 // the current bytecode. 1951 int inputs, ignored_depth; 1952 if (compute_stack_effects(inputs, ignored_depth)) { 1953 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1954 Bytecodes::name(java_bc()), sp(), inputs); 1955 } 1956 } 1957 #endif 1958 1959 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1960 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1961 1962 switch (action) { 1963 case Deoptimization::Action_maybe_recompile: 1964 case Deoptimization::Action_reinterpret: 1965 // Temporary fix for 6529811 to allow virtual calls to be sure they 1966 // get the chance to go from mono->bi->mega 1967 if (!keep_exact_action && 1968 Deoptimization::trap_request_index(trap_request) < 0 && 1969 too_many_recompiles(reason)) { 1970 // This BCI is causing too many recompilations. 1971 if (C->log() != NULL) { 1972 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1973 Deoptimization::trap_reason_name(reason), 1974 Deoptimization::trap_action_name(action)); 1975 } 1976 action = Deoptimization::Action_none; 1977 trap_request = Deoptimization::make_trap_request(reason, action); 1978 } else { 1979 C->set_trap_can_recompile(true); 1980 } 1981 break; 1982 case Deoptimization::Action_make_not_entrant: 1983 C->set_trap_can_recompile(true); 1984 break; 1985 #ifdef ASSERT 1986 case Deoptimization::Action_none: 1987 case Deoptimization::Action_make_not_compilable: 1988 break; 1989 default: 1990 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 1991 break; 1992 #endif 1993 } 1994 1995 if (TraceOptoParse) { 1996 char buf[100]; 1997 tty->print_cr("Uncommon trap %s at bci:%d", 1998 Deoptimization::format_trap_request(buf, sizeof(buf), 1999 trap_request), bci()); 2000 } 2001 2002 CompileLog* log = C->log(); 2003 if (log != NULL) { 2004 int kid = (klass == NULL)? -1: log->identify(klass); 2005 log->begin_elem("uncommon_trap bci='%d'", bci()); 2006 char buf[100]; 2007 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2008 trap_request)); 2009 if (kid >= 0) log->print(" klass='%d'", kid); 2010 if (comment != NULL) log->print(" comment='%s'", comment); 2011 log->end_elem(); 2012 } 2013 2014 // Make sure any guarding test views this path as very unlikely 2015 Node *i0 = control()->in(0); 2016 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2017 IfNode *iff = i0->as_If(); 2018 float f = iff->_prob; // Get prob 2019 if (control()->Opcode() == Op_IfTrue) { 2020 if (f > PROB_UNLIKELY_MAG(4)) 2021 iff->_prob = PROB_MIN; 2022 } else { 2023 if (f < PROB_LIKELY_MAG(4)) 2024 iff->_prob = PROB_MAX; 2025 } 2026 } 2027 2028 // Clear out dead values from the debug info. 2029 kill_dead_locals(); 2030 2031 // Now insert the uncommon trap subroutine call 2032 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2033 const TypePtr* no_memory_effects = NULL; 2034 // Pass the index of the class to be loaded 2035 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2036 (must_throw ? RC_MUST_THROW : 0), 2037 OptoRuntime::uncommon_trap_Type(), 2038 call_addr, "uncommon_trap", no_memory_effects, 2039 intcon(trap_request)); 2040 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2041 "must extract request correctly from the graph"); 2042 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2043 2044 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2045 // The debug info is the only real input to this call. 2046 2047 // Halt-and-catch fire here. The above call should never return! 2048 HaltNode* halt = new HaltNode(control(), frameptr()); 2049 _gvn.set_type_bottom(halt); 2050 root()->add_req(halt); 2051 2052 stop_and_kill_map(); 2053 } 2054 2055 2056 //--------------------------just_allocated_object------------------------------ 2057 // Report the object that was just allocated. 2058 // It must be the case that there are no intervening safepoints. 2059 // We use this to determine if an object is so "fresh" that 2060 // it does not require card marks. 2061 Node* GraphKit::just_allocated_object(Node* current_control) { 2062 if (C->recent_alloc_ctl() == current_control) 2063 return C->recent_alloc_obj(); 2064 return NULL; 2065 } 2066 2067 2068 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2069 // (Note: TypeFunc::make has a cache that makes this fast.) 2070 const TypeFunc* tf = TypeFunc::make(dest_method); 2071 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2072 for (int j = 0; j < nargs; j++) { 2073 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2074 if( targ->basic_type() == T_DOUBLE ) { 2075 // If any parameters are doubles, they must be rounded before 2076 // the call, dstore_rounding does gvn.transform 2077 Node *arg = argument(j); 2078 arg = dstore_rounding(arg); 2079 set_argument(j, arg); 2080 } 2081 } 2082 } 2083 2084 /** 2085 * Record profiling data exact_kls for Node n with the type system so 2086 * that it can propagate it (speculation) 2087 * 2088 * @param n node that the type applies to 2089 * @param exact_kls type from profiling 2090 * @param maybe_null did profiling see null? 2091 * 2092 * @return node with improved type 2093 */ 2094 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2095 const Type* current_type = _gvn.type(n); 2096 assert(UseTypeSpeculation, "type speculation must be on"); 2097 2098 const TypePtr* speculative = current_type->speculative(); 2099 2100 // Should the klass from the profile be recorded in the speculative type? 2101 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2102 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2103 const TypeOopPtr* xtype = tklass->as_instance_type(); 2104 assert(xtype->klass_is_exact(), "Should be exact"); 2105 // Any reason to believe n is not null (from this profiling or a previous one)? 2106 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2107 // record the new speculative type's depth 2108 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2109 speculative = speculative->with_inline_depth(jvms()->depth()); 2110 } else if (current_type->would_improve_ptr(maybe_null)) { 2111 // Profiling report that null was never seen so we can change the 2112 // speculative type to non null ptr. 2113 assert(!maybe_null, "nothing to improve"); 2114 if (speculative == NULL) { 2115 speculative = TypePtr::NOTNULL; 2116 } else { 2117 const TypePtr* ptr = TypePtr::NOTNULL; 2118 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2119 } 2120 } 2121 2122 if (speculative != current_type->speculative()) { 2123 // Build a type with a speculative type (what we think we know 2124 // about the type but will need a guard when we use it) 2125 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2126 // We're changing the type, we need a new CheckCast node to carry 2127 // the new type. The new type depends on the control: what 2128 // profiling tells us is only valid from here as far as we can 2129 // tell. 2130 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2131 cast = _gvn.transform(cast); 2132 replace_in_map(n, cast); 2133 n = cast; 2134 } 2135 2136 return n; 2137 } 2138 2139 /** 2140 * Record profiling data from receiver profiling at an invoke with the 2141 * type system so that it can propagate it (speculation) 2142 * 2143 * @param n receiver node 2144 * 2145 * @return node with improved type 2146 */ 2147 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2148 if (!UseTypeSpeculation) { 2149 return n; 2150 } 2151 ciKlass* exact_kls = profile_has_unique_klass(); 2152 bool maybe_null = true; 2153 if (java_bc() == Bytecodes::_checkcast || 2154 java_bc() == Bytecodes::_instanceof || 2155 java_bc() == Bytecodes::_aastore) { 2156 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2157 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2158 } 2159 return record_profile_for_speculation(n, exact_kls, maybe_null); 2160 return n; 2161 } 2162 2163 /** 2164 * Record profiling data from argument profiling at an invoke with the 2165 * type system so that it can propagate it (speculation) 2166 * 2167 * @param dest_method target method for the call 2168 * @param bc what invoke bytecode is this? 2169 */ 2170 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2171 if (!UseTypeSpeculation) { 2172 return; 2173 } 2174 const TypeFunc* tf = TypeFunc::make(dest_method); 2175 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2176 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2177 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2178 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2179 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2180 bool maybe_null = true; 2181 ciKlass* better_type = NULL; 2182 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2183 record_profile_for_speculation(argument(j), better_type, maybe_null); 2184 } 2185 i++; 2186 } 2187 } 2188 } 2189 2190 /** 2191 * Record profiling data from parameter profiling at an invoke with 2192 * the type system so that it can propagate it (speculation) 2193 */ 2194 void GraphKit::record_profiled_parameters_for_speculation() { 2195 if (!UseTypeSpeculation) { 2196 return; 2197 } 2198 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2199 if (_gvn.type(local(i))->isa_oopptr()) { 2200 bool maybe_null = true; 2201 ciKlass* better_type = NULL; 2202 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2203 record_profile_for_speculation(local(i), better_type, maybe_null); 2204 } 2205 j++; 2206 } 2207 } 2208 } 2209 2210 /** 2211 * Record profiling data from return value profiling at an invoke with 2212 * the type system so that it can propagate it (speculation) 2213 */ 2214 void GraphKit::record_profiled_return_for_speculation() { 2215 if (!UseTypeSpeculation) { 2216 return; 2217 } 2218 bool maybe_null = true; 2219 ciKlass* better_type = NULL; 2220 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2221 // If profiling reports a single type for the return value, 2222 // feed it to the type system so it can propagate it as a 2223 // speculative type 2224 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2225 } 2226 } 2227 2228 void GraphKit::round_double_result(ciMethod* dest_method) { 2229 // A non-strict method may return a double value which has an extended 2230 // exponent, but this must not be visible in a caller which is 'strict' 2231 // If a strict caller invokes a non-strict callee, round a double result 2232 2233 BasicType result_type = dest_method->return_type()->basic_type(); 2234 assert( method() != NULL, "must have caller context"); 2235 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2236 // Destination method's return value is on top of stack 2237 // dstore_rounding() does gvn.transform 2238 Node *result = pop_pair(); 2239 result = dstore_rounding(result); 2240 push_pair(result); 2241 } 2242 } 2243 2244 // rounding for strict float precision conformance 2245 Node* GraphKit::precision_rounding(Node* n) { 2246 return UseStrictFP && _method->flags().is_strict() 2247 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2248 ? _gvn.transform( new RoundFloatNode(0, n) ) 2249 : n; 2250 } 2251 2252 // rounding for strict double precision conformance 2253 Node* GraphKit::dprecision_rounding(Node *n) { 2254 return UseStrictFP && _method->flags().is_strict() 2255 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2256 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2257 : n; 2258 } 2259 2260 // rounding for non-strict double stores 2261 Node* GraphKit::dstore_rounding(Node* n) { 2262 return Matcher::strict_fp_requires_explicit_rounding 2263 && UseSSE <= 1 2264 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2265 : n; 2266 } 2267 2268 //============================================================================= 2269 // Generate a fast path/slow path idiom. Graph looks like: 2270 // [foo] indicates that 'foo' is a parameter 2271 // 2272 // [in] NULL 2273 // \ / 2274 // CmpP 2275 // Bool ne 2276 // If 2277 // / \ 2278 // True False-<2> 2279 // / | 2280 // / cast_not_null 2281 // Load | | ^ 2282 // [fast_test] | | 2283 // gvn to opt_test | | 2284 // / \ | <1> 2285 // True False | 2286 // | \\ | 2287 // [slow_call] \[fast_result] 2288 // Ctl Val \ \ 2289 // | \ \ 2290 // Catch <1> \ \ 2291 // / \ ^ \ \ 2292 // Ex No_Ex | \ \ 2293 // | \ \ | \ <2> \ 2294 // ... \ [slow_res] | | \ [null_result] 2295 // \ \--+--+--- | | 2296 // \ | / \ | / 2297 // --------Region Phi 2298 // 2299 //============================================================================= 2300 // Code is structured as a series of driver functions all called 'do_XXX' that 2301 // call a set of helper functions. Helper functions first, then drivers. 2302 2303 //------------------------------null_check_oop--------------------------------- 2304 // Null check oop. Set null-path control into Region in slot 3. 2305 // Make a cast-not-nullness use the other not-null control. Return cast. 2306 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2307 bool never_see_null, 2308 bool safe_for_replace, 2309 bool speculative) { 2310 // Initial NULL check taken path 2311 (*null_control) = top(); 2312 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2313 2314 // Generate uncommon_trap: 2315 if (never_see_null && (*null_control) != top()) { 2316 // If we see an unexpected null at a check-cast we record it and force a 2317 // recompile; the offending check-cast will be compiled to handle NULLs. 2318 // If we see more than one offending BCI, then all checkcasts in the 2319 // method will be compiled to handle NULLs. 2320 PreserveJVMState pjvms(this); 2321 set_control(*null_control); 2322 replace_in_map(value, null()); 2323 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2324 uncommon_trap(reason, 2325 Deoptimization::Action_make_not_entrant); 2326 (*null_control) = top(); // NULL path is dead 2327 } 2328 if ((*null_control) == top() && safe_for_replace) { 2329 replace_in_map(value, cast); 2330 } 2331 2332 // Cast away null-ness on the result 2333 return cast; 2334 } 2335 2336 //------------------------------opt_iff---------------------------------------- 2337 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2338 // Return slow-path control. 2339 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2340 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2341 2342 // Fast path taken; set region slot 2 2343 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2344 region->init_req(2,fast_taken); // Capture fast-control 2345 2346 // Fast path not-taken, i.e. slow path 2347 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2348 return slow_taken; 2349 } 2350 2351 //-----------------------------make_runtime_call------------------------------- 2352 Node* GraphKit::make_runtime_call(int flags, 2353 const TypeFunc* call_type, address call_addr, 2354 const char* call_name, 2355 const TypePtr* adr_type, 2356 // The following parms are all optional. 2357 // The first NULL ends the list. 2358 Node* parm0, Node* parm1, 2359 Node* parm2, Node* parm3, 2360 Node* parm4, Node* parm5, 2361 Node* parm6, Node* parm7) { 2362 // Slow-path call 2363 bool is_leaf = !(flags & RC_NO_LEAF); 2364 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2365 if (call_name == NULL) { 2366 assert(!is_leaf, "must supply name for leaf"); 2367 call_name = OptoRuntime::stub_name(call_addr); 2368 } 2369 CallNode* call; 2370 if (!is_leaf) { 2371 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2372 bci(), adr_type); 2373 } else if (flags & RC_NO_FP) { 2374 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2375 } else { 2376 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2377 } 2378 2379 // The following is similar to set_edges_for_java_call, 2380 // except that the memory effects of the call are restricted to AliasIdxRaw. 2381 2382 // Slow path call has no side-effects, uses few values 2383 bool wide_in = !(flags & RC_NARROW_MEM); 2384 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2385 2386 Node* prev_mem = NULL; 2387 if (wide_in) { 2388 prev_mem = set_predefined_input_for_runtime_call(call); 2389 } else { 2390 assert(!wide_out, "narrow in => narrow out"); 2391 Node* narrow_mem = memory(adr_type); 2392 prev_mem = reset_memory(); 2393 map()->set_memory(narrow_mem); 2394 set_predefined_input_for_runtime_call(call); 2395 } 2396 2397 // Hook each parm in order. Stop looking at the first NULL. 2398 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2399 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2400 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2401 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2402 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2403 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2404 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2405 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2406 /* close each nested if ===> */ } } } } } } } } 2407 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2408 2409 if (!is_leaf) { 2410 // Non-leaves can block and take safepoints: 2411 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2412 } 2413 // Non-leaves can throw exceptions: 2414 if (has_io) { 2415 call->set_req(TypeFunc::I_O, i_o()); 2416 } 2417 2418 if (flags & RC_UNCOMMON) { 2419 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2420 // (An "if" probability corresponds roughly to an unconditional count. 2421 // Sort of.) 2422 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2423 } 2424 2425 Node* c = _gvn.transform(call); 2426 assert(c == call, "cannot disappear"); 2427 2428 if (wide_out) { 2429 // Slow path call has full side-effects. 2430 set_predefined_output_for_runtime_call(call); 2431 } else { 2432 // Slow path call has few side-effects, and/or sets few values. 2433 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2434 } 2435 2436 if (has_io) { 2437 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2438 } 2439 return call; 2440 2441 } 2442 2443 //------------------------------merge_memory----------------------------------- 2444 // Merge memory from one path into the current memory state. 2445 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2446 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2447 Node* old_slice = mms.force_memory(); 2448 Node* new_slice = mms.memory2(); 2449 if (old_slice != new_slice) { 2450 PhiNode* phi; 2451 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2452 if (mms.is_empty()) { 2453 // clone base memory Phi's inputs for this memory slice 2454 assert(old_slice == mms.base_memory(), "sanity"); 2455 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2456 _gvn.set_type(phi, Type::MEMORY); 2457 for (uint i = 1; i < phi->req(); i++) { 2458 phi->init_req(i, old_slice->in(i)); 2459 } 2460 } else { 2461 phi = old_slice->as_Phi(); // Phi was generated already 2462 } 2463 } else { 2464 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2465 _gvn.set_type(phi, Type::MEMORY); 2466 } 2467 phi->set_req(new_path, new_slice); 2468 mms.set_memory(phi); 2469 } 2470 } 2471 } 2472 2473 //------------------------------make_slow_call_ex------------------------------ 2474 // Make the exception handler hookups for the slow call 2475 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2476 if (stopped()) return; 2477 2478 // Make a catch node with just two handlers: fall-through and catch-all 2479 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2480 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2481 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2482 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2483 2484 { PreserveJVMState pjvms(this); 2485 set_control(excp); 2486 set_i_o(i_o); 2487 2488 if (excp != top()) { 2489 if (deoptimize) { 2490 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2491 uncommon_trap(Deoptimization::Reason_unhandled, 2492 Deoptimization::Action_none); 2493 } else { 2494 // Create an exception state also. 2495 // Use an exact type if the caller has specified a specific exception. 2496 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2497 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2498 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2499 } 2500 } 2501 } 2502 2503 // Get the no-exception control from the CatchNode. 2504 set_control(norm); 2505 } 2506 2507 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2508 Node* cmp = NULL; 2509 switch(bt) { 2510 case T_INT: cmp = new CmpINode(in1, in2); break; 2511 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2512 default: fatal("unexpected comparison type %s", type2name(bt)); 2513 } 2514 gvn->transform(cmp); 2515 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2516 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2517 gvn->transform(iff); 2518 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2519 return iff; 2520 } 2521 2522 2523 //-------------------------------gen_subtype_check----------------------------- 2524 // Generate a subtyping check. Takes as input the subtype and supertype. 2525 // Returns 2 values: sets the default control() to the true path and returns 2526 // the false path. Only reads invariant memory; sets no (visible) memory. 2527 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2528 // but that's not exposed to the optimizer. This call also doesn't take in an 2529 // Object; if you wish to check an Object you need to load the Object's class 2530 // prior to coming here. 2531 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2532 Compile* C = gvn->C; 2533 2534 if ((*ctrl)->is_top()) { 2535 return C->top(); 2536 } 2537 2538 // Fast check for identical types, perhaps identical constants. 2539 // The types can even be identical non-constants, in cases 2540 // involving Array.newInstance, Object.clone, etc. 2541 if (subklass == superklass) 2542 return C->top(); // false path is dead; no test needed. 2543 2544 if (gvn->type(superklass)->singleton()) { 2545 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2546 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2547 2548 // In the common case of an exact superklass, try to fold up the 2549 // test before generating code. You may ask, why not just generate 2550 // the code and then let it fold up? The answer is that the generated 2551 // code will necessarily include null checks, which do not always 2552 // completely fold away. If they are also needless, then they turn 2553 // into a performance loss. Example: 2554 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2555 // Here, the type of 'fa' is often exact, so the store check 2556 // of fa[1]=x will fold up, without testing the nullness of x. 2557 switch (C->static_subtype_check(superk, subk)) { 2558 case Compile::SSC_always_false: 2559 { 2560 Node* always_fail = *ctrl; 2561 *ctrl = gvn->C->top(); 2562 return always_fail; 2563 } 2564 case Compile::SSC_always_true: 2565 return C->top(); 2566 case Compile::SSC_easy_test: 2567 { 2568 // Just do a direct pointer compare and be done. 2569 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2570 *ctrl = gvn->transform(new IfTrueNode(iff)); 2571 return gvn->transform(new IfFalseNode(iff)); 2572 } 2573 case Compile::SSC_full_test: 2574 break; 2575 default: 2576 ShouldNotReachHere(); 2577 } 2578 } 2579 2580 // %%% Possible further optimization: Even if the superklass is not exact, 2581 // if the subklass is the unique subtype of the superklass, the check 2582 // will always succeed. We could leave a dependency behind to ensure this. 2583 2584 // First load the super-klass's check-offset 2585 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2586 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2587 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2588 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2589 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2590 2591 // Load from the sub-klass's super-class display list, or a 1-word cache of 2592 // the secondary superclass list, or a failing value with a sentinel offset 2593 // if the super-klass is an interface or exceptionally deep in the Java 2594 // hierarchy and we have to scan the secondary superclass list the hard way. 2595 // Worst-case type is a little odd: NULL is allowed as a result (usually 2596 // klass loads can never produce a NULL). 2597 Node *chk_off_X = chk_off; 2598 #ifdef _LP64 2599 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2600 #endif 2601 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2602 // For some types like interfaces the following loadKlass is from a 1-word 2603 // cache which is mutable so can't use immutable memory. Other 2604 // types load from the super-class display table which is immutable. 2605 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2606 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2607 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2608 2609 // Compile speed common case: ARE a subtype and we canNOT fail 2610 if( superklass == nkls ) 2611 return C->top(); // false path is dead; no test needed. 2612 2613 // See if we get an immediate positive hit. Happens roughly 83% of the 2614 // time. Test to see if the value loaded just previously from the subklass 2615 // is exactly the superklass. 2616 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2617 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2618 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2619 2620 // Compile speed common case: Check for being deterministic right now. If 2621 // chk_off is a constant and not equal to cacheoff then we are NOT a 2622 // subklass. In this case we need exactly the 1 test above and we can 2623 // return those results immediately. 2624 if (!might_be_cache) { 2625 Node* not_subtype_ctrl = *ctrl; 2626 *ctrl = iftrue1; // We need exactly the 1 test above 2627 return not_subtype_ctrl; 2628 } 2629 2630 // Gather the various success & failures here 2631 RegionNode *r_ok_subtype = new RegionNode(4); 2632 gvn->record_for_igvn(r_ok_subtype); 2633 RegionNode *r_not_subtype = new RegionNode(3); 2634 gvn->record_for_igvn(r_not_subtype); 2635 2636 r_ok_subtype->init_req(1, iftrue1); 2637 2638 // Check for immediate negative hit. Happens roughly 11% of the time (which 2639 // is roughly 63% of the remaining cases). Test to see if the loaded 2640 // check-offset points into the subklass display list or the 1-element 2641 // cache. If it points to the display (and NOT the cache) and the display 2642 // missed then it's not a subtype. 2643 Node *cacheoff = gvn->intcon(cacheoff_con); 2644 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2645 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2646 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2647 2648 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2649 // No performance impact (too rare) but allows sharing of secondary arrays 2650 // which has some footprint reduction. 2651 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2652 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2653 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2654 2655 // -- Roads not taken here: -- 2656 // We could also have chosen to perform the self-check at the beginning 2657 // of this code sequence, as the assembler does. This would not pay off 2658 // the same way, since the optimizer, unlike the assembler, can perform 2659 // static type analysis to fold away many successful self-checks. 2660 // Non-foldable self checks work better here in second position, because 2661 // the initial primary superclass check subsumes a self-check for most 2662 // types. An exception would be a secondary type like array-of-interface, 2663 // which does not appear in its own primary supertype display. 2664 // Finally, we could have chosen to move the self-check into the 2665 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2666 // dependent manner. But it is worthwhile to have the check here, 2667 // where it can be perhaps be optimized. The cost in code space is 2668 // small (register compare, branch). 2669 2670 // Now do a linear scan of the secondary super-klass array. Again, no real 2671 // performance impact (too rare) but it's gotta be done. 2672 // Since the code is rarely used, there is no penalty for moving it 2673 // out of line, and it can only improve I-cache density. 2674 // The decision to inline or out-of-line this final check is platform 2675 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2676 Node* psc = gvn->transform( 2677 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2678 2679 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2680 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2681 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2682 2683 // Return false path; set default control to true path. 2684 *ctrl = gvn->transform(r_ok_subtype); 2685 return gvn->transform(r_not_subtype); 2686 } 2687 2688 // Profile-driven exact type check: 2689 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2690 float prob, 2691 Node* *casted_receiver) { 2692 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2693 Node* recv_klass = load_object_klass(receiver); 2694 Node* want_klass = makecon(tklass); 2695 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2696 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2697 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2698 set_control( _gvn.transform( new IfTrueNode (iff) )); 2699 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2700 2701 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2702 assert(recv_xtype->klass_is_exact(), ""); 2703 2704 // Subsume downstream occurrences of receiver with a cast to 2705 // recv_xtype, since now we know what the type will be. 2706 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2707 (*casted_receiver) = _gvn.transform(cast); 2708 // (User must make the replace_in_map call.) 2709 2710 return fail; 2711 } 2712 2713 2714 //------------------------------seems_never_null------------------------------- 2715 // Use null_seen information if it is available from the profile. 2716 // If we see an unexpected null at a type check we record it and force a 2717 // recompile; the offending check will be recompiled to handle NULLs. 2718 // If we see several offending BCIs, then all checks in the 2719 // method will be recompiled. 2720 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2721 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2722 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2723 if (UncommonNullCast // Cutout for this technique 2724 && obj != null() // And not the -Xcomp stupid case? 2725 && !too_many_traps(reason) 2726 ) { 2727 if (speculating) { 2728 return true; 2729 } 2730 if (data == NULL) 2731 // Edge case: no mature data. Be optimistic here. 2732 return true; 2733 // If the profile has not seen a null, assume it won't happen. 2734 assert(java_bc() == Bytecodes::_checkcast || 2735 java_bc() == Bytecodes::_instanceof || 2736 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2737 return !data->as_BitData()->null_seen(); 2738 } 2739 speculating = false; 2740 return false; 2741 } 2742 2743 //------------------------maybe_cast_profiled_receiver------------------------- 2744 // If the profile has seen exactly one type, narrow to exactly that type. 2745 // Subsequent type checks will always fold up. 2746 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2747 ciKlass* require_klass, 2748 ciKlass* spec_klass, 2749 bool safe_for_replace) { 2750 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2751 2752 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2753 2754 // Make sure we haven't already deoptimized from this tactic. 2755 if (too_many_traps(reason) || too_many_recompiles(reason)) 2756 return NULL; 2757 2758 // (No, this isn't a call, but it's enough like a virtual call 2759 // to use the same ciMethod accessor to get the profile info...) 2760 // If we have a speculative type use it instead of profiling (which 2761 // may not help us) 2762 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2763 if (exact_kls != NULL) {// no cast failures here 2764 if (require_klass == NULL || 2765 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2766 // If we narrow the type to match what the type profile sees or 2767 // the speculative type, we can then remove the rest of the 2768 // cast. 2769 // This is a win, even if the exact_kls is very specific, 2770 // because downstream operations, such as method calls, 2771 // will often benefit from the sharper type. 2772 Node* exact_obj = not_null_obj; // will get updated in place... 2773 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2774 &exact_obj); 2775 { PreserveJVMState pjvms(this); 2776 set_control(slow_ctl); 2777 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2778 } 2779 if (safe_for_replace) { 2780 replace_in_map(not_null_obj, exact_obj); 2781 } 2782 return exact_obj; 2783 } 2784 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2785 } 2786 2787 return NULL; 2788 } 2789 2790 /** 2791 * Cast obj to type and emit guard unless we had too many traps here 2792 * already 2793 * 2794 * @param obj node being casted 2795 * @param type type to cast the node to 2796 * @param not_null true if we know node cannot be null 2797 */ 2798 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2799 ciKlass* type, 2800 bool not_null) { 2801 if (stopped()) { 2802 return obj; 2803 } 2804 2805 // type == NULL if profiling tells us this object is always null 2806 if (type != NULL) { 2807 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2808 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2809 2810 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2811 !too_many_traps(class_reason) && 2812 !too_many_recompiles(class_reason)) { 2813 Node* not_null_obj = NULL; 2814 // not_null is true if we know the object is not null and 2815 // there's no need for a null check 2816 if (!not_null) { 2817 Node* null_ctl = top(); 2818 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2819 assert(null_ctl->is_top(), "no null control here"); 2820 } else { 2821 not_null_obj = obj; 2822 } 2823 2824 Node* exact_obj = not_null_obj; 2825 ciKlass* exact_kls = type; 2826 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2827 &exact_obj); 2828 { 2829 PreserveJVMState pjvms(this); 2830 set_control(slow_ctl); 2831 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2832 } 2833 replace_in_map(not_null_obj, exact_obj); 2834 obj = exact_obj; 2835 } 2836 } else { 2837 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2838 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2839 Node* exact_obj = null_assert(obj); 2840 replace_in_map(obj, exact_obj); 2841 obj = exact_obj; 2842 } 2843 } 2844 return obj; 2845 } 2846 2847 //-------------------------------gen_instanceof-------------------------------- 2848 // Generate an instance-of idiom. Used by both the instance-of bytecode 2849 // and the reflective instance-of call. 2850 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2851 kill_dead_locals(); // Benefit all the uncommon traps 2852 assert( !stopped(), "dead parse path should be checked in callers" ); 2853 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2854 "must check for not-null not-dead klass in callers"); 2855 2856 // Make the merge point 2857 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2858 RegionNode* region = new RegionNode(PATH_LIMIT); 2859 Node* phi = new PhiNode(region, TypeInt::BOOL); 2860 C->set_has_split_ifs(true); // Has chance for split-if optimization 2861 2862 ciProfileData* data = NULL; 2863 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2864 data = method()->method_data()->bci_to_data(bci()); 2865 } 2866 bool speculative_not_null = false; 2867 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2868 && seems_never_null(obj, data, speculative_not_null)); 2869 2870 // Null check; get casted pointer; set region slot 3 2871 Node* null_ctl = top(); 2872 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2873 2874 // If not_null_obj is dead, only null-path is taken 2875 if (stopped()) { // Doing instance-of on a NULL? 2876 set_control(null_ctl); 2877 return intcon(0); 2878 } 2879 region->init_req(_null_path, null_ctl); 2880 phi ->init_req(_null_path, intcon(0)); // Set null path value 2881 if (null_ctl == top()) { 2882 // Do this eagerly, so that pattern matches like is_diamond_phi 2883 // will work even during parsing. 2884 assert(_null_path == PATH_LIMIT-1, "delete last"); 2885 region->del_req(_null_path); 2886 phi ->del_req(_null_path); 2887 } 2888 2889 // Do we know the type check always succeed? 2890 bool known_statically = false; 2891 if (_gvn.type(superklass)->singleton()) { 2892 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2893 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2894 if (subk != NULL && subk->is_loaded()) { 2895 int static_res = C->static_subtype_check(superk, subk); 2896 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2897 } 2898 } 2899 2900 if (known_statically && UseTypeSpeculation) { 2901 // If we know the type check always succeeds then we don't use the 2902 // profiling data at this bytecode. Don't lose it, feed it to the 2903 // type system as a speculative type. 2904 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2905 } else { 2906 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2907 // We may not have profiling here or it may not help us. If we 2908 // have a speculative type use it to perform an exact cast. 2909 ciKlass* spec_obj_type = obj_type->speculative_type(); 2910 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2911 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2912 if (stopped()) { // Profile disagrees with this path. 2913 set_control(null_ctl); // Null is the only remaining possibility. 2914 return intcon(0); 2915 } 2916 if (cast_obj != NULL) { 2917 not_null_obj = cast_obj; 2918 } 2919 } 2920 } 2921 2922 // Load the object's klass 2923 Node* obj_klass = load_object_klass(not_null_obj); 2924 2925 // Generate the subtype check 2926 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2927 2928 // Plug in the success path to the general merge in slot 1. 2929 region->init_req(_obj_path, control()); 2930 phi ->init_req(_obj_path, intcon(1)); 2931 2932 // Plug in the failing path to the general merge in slot 2. 2933 region->init_req(_fail_path, not_subtype_ctrl); 2934 phi ->init_req(_fail_path, intcon(0)); 2935 2936 // Return final merged results 2937 set_control( _gvn.transform(region) ); 2938 record_for_igvn(region); 2939 return _gvn.transform(phi); 2940 } 2941 2942 //-------------------------------gen_checkcast--------------------------------- 2943 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2944 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2945 // uncommon-trap paths work. Adjust stack after this call. 2946 // If failure_control is supplied and not null, it is filled in with 2947 // the control edge for the cast failure. Otherwise, an appropriate 2948 // uncommon trap or exception is thrown. 2949 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2950 Node* *failure_control) { 2951 kill_dead_locals(); // Benefit all the uncommon traps 2952 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2953 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2954 2955 // Fast cutout: Check the case that the cast is vacuously true. 2956 // This detects the common cases where the test will short-circuit 2957 // away completely. We do this before we perform the null check, 2958 // because if the test is going to turn into zero code, we don't 2959 // want a residual null check left around. (Causes a slowdown, 2960 // for example, in some objArray manipulations, such as a[i]=a[j].) 2961 if (tk->singleton()) { 2962 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2963 if (objtp != NULL && objtp->klass() != NULL) { 2964 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2965 case Compile::SSC_always_true: 2966 // If we know the type check always succeed then we don't use 2967 // the profiling data at this bytecode. Don't lose it, feed it 2968 // to the type system as a speculative type. 2969 return record_profiled_receiver_for_speculation(obj); 2970 case Compile::SSC_always_false: 2971 // It needs a null check because a null will *pass* the cast check. 2972 // A non-null value will always produce an exception. 2973 return null_assert(obj); 2974 } 2975 } 2976 } 2977 2978 ciProfileData* data = NULL; 2979 bool safe_for_replace = false; 2980 if (failure_control == NULL) { // use MDO in regular case only 2981 assert(java_bc() == Bytecodes::_aastore || 2982 java_bc() == Bytecodes::_checkcast, 2983 "interpreter profiles type checks only for these BCs"); 2984 data = method()->method_data()->bci_to_data(bci()); 2985 safe_for_replace = true; 2986 } 2987 2988 // Make the merge point 2989 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2990 RegionNode* region = new RegionNode(PATH_LIMIT); 2991 Node* phi = new PhiNode(region, toop); 2992 C->set_has_split_ifs(true); // Has chance for split-if optimization 2993 2994 // Use null-cast information if it is available 2995 bool speculative_not_null = false; 2996 bool never_see_null = ((failure_control == NULL) // regular case only 2997 && seems_never_null(obj, data, speculative_not_null)); 2998 2999 // Null check; get casted pointer; set region slot 3 3000 Node* null_ctl = top(); 3001 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3002 3003 // If not_null_obj is dead, only null-path is taken 3004 if (stopped()) { // Doing instance-of on a NULL? 3005 set_control(null_ctl); 3006 return null(); 3007 } 3008 region->init_req(_null_path, null_ctl); 3009 phi ->init_req(_null_path, null()); // Set null path value 3010 if (null_ctl == top()) { 3011 // Do this eagerly, so that pattern matches like is_diamond_phi 3012 // will work even during parsing. 3013 assert(_null_path == PATH_LIMIT-1, "delete last"); 3014 region->del_req(_null_path); 3015 phi ->del_req(_null_path); 3016 } 3017 3018 Node* cast_obj = NULL; 3019 if (tk->klass_is_exact()) { 3020 // The following optimization tries to statically cast the speculative type of the object 3021 // (for example obtained during profiling) to the type of the superklass and then do a 3022 // dynamic check that the type of the object is what we expect. To work correctly 3023 // for checkcast and aastore the type of superklass should be exact. 3024 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3025 // We may not have profiling here or it may not help us. If we have 3026 // a speculative type use it to perform an exact cast. 3027 ciKlass* spec_obj_type = obj_type->speculative_type(); 3028 if (spec_obj_type != NULL || data != NULL) { 3029 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3030 if (cast_obj != NULL) { 3031 if (failure_control != NULL) // failure is now impossible 3032 (*failure_control) = top(); 3033 // adjust the type of the phi to the exact klass: 3034 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3035 } 3036 } 3037 } 3038 3039 if (cast_obj == NULL) { 3040 // Load the object's klass 3041 Node* obj_klass = load_object_klass(not_null_obj); 3042 3043 // Generate the subtype check 3044 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3045 3046 // Plug in success path into the merge 3047 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3048 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3049 if (failure_control == NULL) { 3050 if (not_subtype_ctrl != top()) { // If failure is possible 3051 PreserveJVMState pjvms(this); 3052 set_control(not_subtype_ctrl); 3053 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3054 } 3055 } else { 3056 (*failure_control) = not_subtype_ctrl; 3057 } 3058 } 3059 3060 region->init_req(_obj_path, control()); 3061 phi ->init_req(_obj_path, cast_obj); 3062 3063 // A merge of NULL or Casted-NotNull obj 3064 Node* res = _gvn.transform(phi); 3065 3066 // Note I do NOT always 'replace_in_map(obj,result)' here. 3067 // if( tk->klass()->can_be_primary_super() ) 3068 // This means that if I successfully store an Object into an array-of-String 3069 // I 'forget' that the Object is really now known to be a String. I have to 3070 // do this because we don't have true union types for interfaces - if I store 3071 // a Baz into an array-of-Interface and then tell the optimizer it's an 3072 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3073 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3074 // replace_in_map( obj, res ); 3075 3076 // Return final merged results 3077 set_control( _gvn.transform(region) ); 3078 record_for_igvn(region); 3079 return res; 3080 } 3081 3082 //------------------------------next_monitor----------------------------------- 3083 // What number should be given to the next monitor? 3084 int GraphKit::next_monitor() { 3085 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3086 int next = current + C->sync_stack_slots(); 3087 // Keep the toplevel high water mark current: 3088 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3089 return current; 3090 } 3091 3092 //------------------------------insert_mem_bar--------------------------------- 3093 // Memory barrier to avoid floating things around 3094 // The membar serves as a pinch point between both control and all memory slices. 3095 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3096 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3097 mb->init_req(TypeFunc::Control, control()); 3098 mb->init_req(TypeFunc::Memory, reset_memory()); 3099 Node* membar = _gvn.transform(mb); 3100 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3101 set_all_memory_call(membar); 3102 return membar; 3103 } 3104 3105 //-------------------------insert_mem_bar_volatile---------------------------- 3106 // Memory barrier to avoid floating things around 3107 // The membar serves as a pinch point between both control and memory(alias_idx). 3108 // If you want to make a pinch point on all memory slices, do not use this 3109 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3110 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3111 // When Parse::do_put_xxx updates a volatile field, it appends a series 3112 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3113 // The first membar is on the same memory slice as the field store opcode. 3114 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3115 // All the other membars (for other volatile slices, including AliasIdxBot, 3116 // which stands for all unknown volatile slices) are control-dependent 3117 // on the first membar. This prevents later volatile loads or stores 3118 // from sliding up past the just-emitted store. 3119 3120 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3121 mb->set_req(TypeFunc::Control,control()); 3122 if (alias_idx == Compile::AliasIdxBot) { 3123 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3124 } else { 3125 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3126 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3127 } 3128 Node* membar = _gvn.transform(mb); 3129 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3130 if (alias_idx == Compile::AliasIdxBot) { 3131 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3132 } else { 3133 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3134 } 3135 return membar; 3136 } 3137 3138 //------------------------------shared_lock------------------------------------ 3139 // Emit locking code. 3140 FastLockNode* GraphKit::shared_lock(Node* obj) { 3141 // bci is either a monitorenter bc or InvocationEntryBci 3142 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3143 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3144 3145 if( !GenerateSynchronizationCode ) 3146 return NULL; // Not locking things? 3147 if (stopped()) // Dead monitor? 3148 return NULL; 3149 3150 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3151 3152 // Box the stack location 3153 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3154 Node* mem = reset_memory(); 3155 3156 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3157 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3158 // Create the counters for this fast lock. 3159 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3160 } 3161 3162 // Create the rtm counters for this fast lock if needed. 3163 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3164 3165 // Add monitor to debug info for the slow path. If we block inside the 3166 // slow path and de-opt, we need the monitor hanging around 3167 map()->push_monitor( flock ); 3168 3169 const TypeFunc *tf = LockNode::lock_type(); 3170 LockNode *lock = new LockNode(C, tf); 3171 3172 lock->init_req( TypeFunc::Control, control() ); 3173 lock->init_req( TypeFunc::Memory , mem ); 3174 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3175 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3176 lock->init_req( TypeFunc::ReturnAdr, top() ); 3177 3178 lock->init_req(TypeFunc::Parms + 0, obj); 3179 lock->init_req(TypeFunc::Parms + 1, box); 3180 lock->init_req(TypeFunc::Parms + 2, flock); 3181 add_safepoint_edges(lock); 3182 3183 lock = _gvn.transform( lock )->as_Lock(); 3184 3185 // lock has no side-effects, sets few values 3186 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3187 3188 insert_mem_bar(Op_MemBarAcquireLock); 3189 3190 // Add this to the worklist so that the lock can be eliminated 3191 record_for_igvn(lock); 3192 3193 #ifndef PRODUCT 3194 if (PrintLockStatistics) { 3195 // Update the counter for this lock. Don't bother using an atomic 3196 // operation since we don't require absolute accuracy. 3197 lock->create_lock_counter(map()->jvms()); 3198 increment_counter(lock->counter()->addr()); 3199 } 3200 #endif 3201 3202 return flock; 3203 } 3204 3205 3206 //------------------------------shared_unlock---------------------------------- 3207 // Emit unlocking code. 3208 void GraphKit::shared_unlock(Node* box, Node* obj) { 3209 // bci is either a monitorenter bc or InvocationEntryBci 3210 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3211 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3212 3213 if( !GenerateSynchronizationCode ) 3214 return; 3215 if (stopped()) { // Dead monitor? 3216 map()->pop_monitor(); // Kill monitor from debug info 3217 return; 3218 } 3219 3220 // Memory barrier to avoid floating things down past the locked region 3221 insert_mem_bar(Op_MemBarReleaseLock); 3222 3223 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3224 UnlockNode *unlock = new UnlockNode(C, tf); 3225 #ifdef ASSERT 3226 unlock->set_dbg_jvms(sync_jvms()); 3227 #endif 3228 uint raw_idx = Compile::AliasIdxRaw; 3229 unlock->init_req( TypeFunc::Control, control() ); 3230 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3231 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3232 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3233 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3234 3235 unlock->init_req(TypeFunc::Parms + 0, obj); 3236 unlock->init_req(TypeFunc::Parms + 1, box); 3237 unlock = _gvn.transform(unlock)->as_Unlock(); 3238 3239 Node* mem = reset_memory(); 3240 3241 // unlock has no side-effects, sets few values 3242 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3243 3244 // Kill monitor from debug info 3245 map()->pop_monitor( ); 3246 } 3247 3248 //-------------------------------get_layout_helper----------------------------- 3249 // If the given klass is a constant or known to be an array, 3250 // fetch the constant layout helper value into constant_value 3251 // and return (Node*)NULL. Otherwise, load the non-constant 3252 // layout helper value, and return the node which represents it. 3253 // This two-faced routine is useful because allocation sites 3254 // almost always feature constant types. 3255 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3256 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3257 if (!StressReflectiveCode && inst_klass != NULL) { 3258 ciKlass* klass = inst_klass->klass(); 3259 bool xklass = inst_klass->klass_is_exact(); 3260 if (xklass || klass->is_array_klass()) { 3261 jint lhelper = klass->layout_helper(); 3262 if (lhelper != Klass::_lh_neutral_value) { 3263 constant_value = lhelper; 3264 return (Node*) NULL; 3265 } 3266 } 3267 } 3268 constant_value = Klass::_lh_neutral_value; // put in a known value 3269 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3270 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3271 } 3272 3273 // We just put in an allocate/initialize with a big raw-memory effect. 3274 // Hook selected additional alias categories on the initialization. 3275 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3276 MergeMemNode* init_in_merge, 3277 Node* init_out_raw) { 3278 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3279 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3280 3281 Node* prevmem = kit.memory(alias_idx); 3282 init_in_merge->set_memory_at(alias_idx, prevmem); 3283 kit.set_memory(init_out_raw, alias_idx); 3284 } 3285 3286 //---------------------------set_output_for_allocation------------------------- 3287 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3288 const TypeOopPtr* oop_type, 3289 bool deoptimize_on_exception) { 3290 int rawidx = Compile::AliasIdxRaw; 3291 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3292 add_safepoint_edges(alloc); 3293 Node* allocx = _gvn.transform(alloc); 3294 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3295 // create memory projection for i_o 3296 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3297 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3298 3299 // create a memory projection as for the normal control path 3300 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3301 set_memory(malloc, rawidx); 3302 3303 // a normal slow-call doesn't change i_o, but an allocation does 3304 // we create a separate i_o projection for the normal control path 3305 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3306 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3307 3308 // put in an initialization barrier 3309 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3310 rawoop)->as_Initialize(); 3311 assert(alloc->initialization() == init, "2-way macro link must work"); 3312 assert(init ->allocation() == alloc, "2-way macro link must work"); 3313 { 3314 // Extract memory strands which may participate in the new object's 3315 // initialization, and source them from the new InitializeNode. 3316 // This will allow us to observe initializations when they occur, 3317 // and link them properly (as a group) to the InitializeNode. 3318 assert(init->in(InitializeNode::Memory) == malloc, ""); 3319 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3320 init->set_req(InitializeNode::Memory, minit_in); 3321 record_for_igvn(minit_in); // fold it up later, if possible 3322 Node* minit_out = memory(rawidx); 3323 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3324 if (oop_type->isa_aryptr()) { 3325 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3326 int elemidx = C->get_alias_index(telemref); 3327 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3328 } else if (oop_type->isa_instptr()) { 3329 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3330 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3331 ciField* field = ik->nonstatic_field_at(i); 3332 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3333 continue; // do not bother to track really large numbers of fields 3334 // Find (or create) the alias category for this field: 3335 int fieldidx = C->alias_type(field)->index(); 3336 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3337 } 3338 } 3339 } 3340 3341 // Cast raw oop to the real thing... 3342 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3343 javaoop = _gvn.transform(javaoop); 3344 C->set_recent_alloc(control(), javaoop); 3345 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3346 3347 #ifdef ASSERT 3348 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3349 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3350 "Ideal_allocation works"); 3351 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3352 "Ideal_allocation works"); 3353 if (alloc->is_AllocateArray()) { 3354 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3355 "Ideal_allocation works"); 3356 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3357 "Ideal_allocation works"); 3358 } else { 3359 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3360 } 3361 } 3362 #endif //ASSERT 3363 3364 return javaoop; 3365 } 3366 3367 //---------------------------new_instance-------------------------------------- 3368 // This routine takes a klass_node which may be constant (for a static type) 3369 // or may be non-constant (for reflective code). It will work equally well 3370 // for either, and the graph will fold nicely if the optimizer later reduces 3371 // the type to a constant. 3372 // The optional arguments are for specialized use by intrinsics: 3373 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3374 // - If 'return_size_val', report the the total object size to the caller. 3375 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3376 Node* GraphKit::new_instance(Node* klass_node, 3377 Node* extra_slow_test, 3378 Node* *return_size_val, 3379 bool deoptimize_on_exception) { 3380 // Compute size in doublewords 3381 // The size is always an integral number of doublewords, represented 3382 // as a positive bytewise size stored in the klass's layout_helper. 3383 // The layout_helper also encodes (in a low bit) the need for a slow path. 3384 jint layout_con = Klass::_lh_neutral_value; 3385 Node* layout_val = get_layout_helper(klass_node, layout_con); 3386 int layout_is_con = (layout_val == NULL); 3387 3388 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3389 // Generate the initial go-slow test. It's either ALWAYS (return a 3390 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3391 // case) a computed value derived from the layout_helper. 3392 Node* initial_slow_test = NULL; 3393 if (layout_is_con) { 3394 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3395 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3396 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3397 3398 } else { // reflective case 3399 // This reflective path is used by Unsafe.allocateInstance. 3400 // (It may be stress-tested by specifying StressReflectiveCode.) 3401 // Basically, we want to get into the VM is there's an illegal argument. 3402 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3403 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3404 if (extra_slow_test != intcon(0)) { 3405 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3406 } 3407 // (Macro-expander will further convert this to a Bool, if necessary.) 3408 } 3409 3410 // Find the size in bytes. This is easy; it's the layout_helper. 3411 // The size value must be valid even if the slow path is taken. 3412 Node* size = NULL; 3413 if (layout_is_con) { 3414 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3415 } else { // reflective case 3416 // This reflective path is used by clone and Unsafe.allocateInstance. 3417 size = ConvI2X(layout_val); 3418 3419 // Clear the low bits to extract layout_helper_size_in_bytes: 3420 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3421 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3422 size = _gvn.transform( new AndXNode(size, mask) ); 3423 } 3424 if (return_size_val != NULL) { 3425 (*return_size_val) = size; 3426 } 3427 3428 // This is a precise notnull oop of the klass. 3429 // (Actually, it need not be precise if this is a reflective allocation.) 3430 // It's what we cast the result to. 3431 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3432 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3433 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3434 3435 // Now generate allocation code 3436 3437 // The entire memory state is needed for slow path of the allocation 3438 // since GC and deoptimization can happened. 3439 Node *mem = reset_memory(); 3440 set_all_memory(mem); // Create new memory state 3441 3442 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3443 control(), mem, i_o(), 3444 size, klass_node, 3445 initial_slow_test); 3446 3447 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3448 } 3449 3450 //-------------------------------new_array------------------------------------- 3451 // helper for both newarray and anewarray 3452 // The 'length' parameter is (obviously) the length of the array. 3453 // See comments on new_instance for the meaning of the other arguments. 3454 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3455 Node* length, // number of array elements 3456 int nargs, // number of arguments to push back for uncommon trap 3457 Node* *return_size_val, 3458 bool deoptimize_on_exception) { 3459 jint layout_con = Klass::_lh_neutral_value; 3460 Node* layout_val = get_layout_helper(klass_node, layout_con); 3461 int layout_is_con = (layout_val == NULL); 3462 3463 if (!layout_is_con && !StressReflectiveCode && 3464 !too_many_traps(Deoptimization::Reason_class_check)) { 3465 // This is a reflective array creation site. 3466 // Optimistically assume that it is a subtype of Object[], 3467 // so that we can fold up all the address arithmetic. 3468 layout_con = Klass::array_layout_helper(T_OBJECT); 3469 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3470 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3471 { BuildCutout unless(this, bol_lh, PROB_MAX); 3472 inc_sp(nargs); 3473 uncommon_trap(Deoptimization::Reason_class_check, 3474 Deoptimization::Action_maybe_recompile); 3475 } 3476 layout_val = NULL; 3477 layout_is_con = true; 3478 } 3479 3480 // Generate the initial go-slow test. Make sure we do not overflow 3481 // if length is huge (near 2Gig) or negative! We do not need 3482 // exact double-words here, just a close approximation of needed 3483 // double-words. We can't add any offset or rounding bits, lest we 3484 // take a size -1 of bytes and make it positive. Use an unsigned 3485 // compare, so negative sizes look hugely positive. 3486 int fast_size_limit = FastAllocateSizeLimit; 3487 if (layout_is_con) { 3488 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3489 // Increase the size limit if we have exact knowledge of array type. 3490 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3491 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3492 } 3493 3494 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3495 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3496 if (initial_slow_test->is_Bool()) { 3497 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3498 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3499 } 3500 3501 // --- Size Computation --- 3502 // array_size = round_to_heap(array_header + (length << elem_shift)); 3503 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3504 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3505 // The rounding mask is strength-reduced, if possible. 3506 int round_mask = MinObjAlignmentInBytes - 1; 3507 Node* header_size = NULL; 3508 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3509 // (T_BYTE has the weakest alignment and size restrictions...) 3510 if (layout_is_con) { 3511 int hsize = Klass::layout_helper_header_size(layout_con); 3512 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3513 BasicType etype = Klass::layout_helper_element_type(layout_con); 3514 if ((round_mask & ~right_n_bits(eshift)) == 0) 3515 round_mask = 0; // strength-reduce it if it goes away completely 3516 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3517 assert(header_size_min <= hsize, "generic minimum is smallest"); 3518 header_size_min = hsize; 3519 header_size = intcon(hsize + round_mask); 3520 } else { 3521 Node* hss = intcon(Klass::_lh_header_size_shift); 3522 Node* hsm = intcon(Klass::_lh_header_size_mask); 3523 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3524 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3525 Node* mask = intcon(round_mask); 3526 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3527 } 3528 3529 Node* elem_shift = NULL; 3530 if (layout_is_con) { 3531 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3532 if (eshift != 0) 3533 elem_shift = intcon(eshift); 3534 } else { 3535 // There is no need to mask or shift this value. 3536 // The semantics of LShiftINode include an implicit mask to 0x1F. 3537 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3538 elem_shift = layout_val; 3539 } 3540 3541 // Transition to native address size for all offset calculations: 3542 Node* lengthx = ConvI2X(length); 3543 Node* headerx = ConvI2X(header_size); 3544 #ifdef _LP64 3545 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3546 if (tllen != NULL && tllen->_lo < 0) { 3547 // Add a manual constraint to a positive range. Cf. array_element_address. 3548 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3549 if (size_max > tllen->_hi) size_max = tllen->_hi; 3550 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3551 lengthx = _gvn.transform( new ConvI2LNode(length, tlcon)); 3552 } 3553 } 3554 #endif 3555 3556 // Combine header size (plus rounding) and body size. Then round down. 3557 // This computation cannot overflow, because it is used only in two 3558 // places, one where the length is sharply limited, and the other 3559 // after a successful allocation. 3560 Node* abody = lengthx; 3561 if (elem_shift != NULL) 3562 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3563 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3564 if (round_mask != 0) { 3565 Node* mask = MakeConX(~round_mask); 3566 size = _gvn.transform( new AndXNode(size, mask) ); 3567 } 3568 // else if round_mask == 0, the size computation is self-rounding 3569 3570 if (return_size_val != NULL) { 3571 // This is the size 3572 (*return_size_val) = size; 3573 } 3574 3575 // Now generate allocation code 3576 3577 // The entire memory state is needed for slow path of the allocation 3578 // since GC and deoptimization can happened. 3579 Node *mem = reset_memory(); 3580 set_all_memory(mem); // Create new memory state 3581 3582 // Create the AllocateArrayNode and its result projections 3583 AllocateArrayNode* alloc 3584 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3585 control(), mem, i_o(), 3586 size, klass_node, 3587 initial_slow_test, 3588 length); 3589 3590 // Cast to correct type. Note that the klass_node may be constant or not, 3591 // and in the latter case the actual array type will be inexact also. 3592 // (This happens via a non-constant argument to inline_native_newArray.) 3593 // In any case, the value of klass_node provides the desired array type. 3594 const TypeInt* length_type = _gvn.find_int_type(length); 3595 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3596 if (ary_type->isa_aryptr() && length_type != NULL) { 3597 // Try to get a better type than POS for the size 3598 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3599 } 3600 3601 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3602 3603 // Cast length on remaining path to be as narrow as possible 3604 if (map()->find_edge(length) >= 0) { 3605 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3606 if (ccast != length) { 3607 _gvn.set_type_bottom(ccast); 3608 record_for_igvn(ccast); 3609 replace_in_map(length, ccast); 3610 } 3611 } 3612 3613 return javaoop; 3614 } 3615 3616 // The following "Ideal_foo" functions are placed here because they recognize 3617 // the graph shapes created by the functions immediately above. 3618 3619 //---------------------------Ideal_allocation---------------------------------- 3620 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3621 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3622 if (ptr == NULL) { // reduce dumb test in callers 3623 return NULL; 3624 } 3625 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3626 ptr = ptr->in(1); 3627 if (ptr == NULL) return NULL; 3628 } 3629 // Return NULL for allocations with several casts: 3630 // j.l.reflect.Array.newInstance(jobject, jint) 3631 // Object.clone() 3632 // to keep more precise type from last cast. 3633 if (ptr->is_Proj()) { 3634 Node* allo = ptr->in(0); 3635 if (allo != NULL && allo->is_Allocate()) { 3636 return allo->as_Allocate(); 3637 } 3638 } 3639 // Report failure to match. 3640 return NULL; 3641 } 3642 3643 // Fancy version which also strips off an offset (and reports it to caller). 3644 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3645 intptr_t& offset) { 3646 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3647 if (base == NULL) return NULL; 3648 return Ideal_allocation(base, phase); 3649 } 3650 3651 // Trace Initialize <- Proj[Parm] <- Allocate 3652 AllocateNode* InitializeNode::allocation() { 3653 Node* rawoop = in(InitializeNode::RawAddress); 3654 if (rawoop->is_Proj()) { 3655 Node* alloc = rawoop->in(0); 3656 if (alloc->is_Allocate()) { 3657 return alloc->as_Allocate(); 3658 } 3659 } 3660 return NULL; 3661 } 3662 3663 // Trace Allocate -> Proj[Parm] -> Initialize 3664 InitializeNode* AllocateNode::initialization() { 3665 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3666 if (rawoop == NULL) return NULL; 3667 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3668 Node* init = rawoop->fast_out(i); 3669 if (init->is_Initialize()) { 3670 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3671 return init->as_Initialize(); 3672 } 3673 } 3674 return NULL; 3675 } 3676 3677 //----------------------------- loop predicates --------------------------- 3678 3679 //------------------------------add_predicate_impl---------------------------- 3680 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3681 // Too many traps seen? 3682 if (too_many_traps(reason)) { 3683 #ifdef ASSERT 3684 if (TraceLoopPredicate) { 3685 int tc = C->trap_count(reason); 3686 tty->print("too many traps=%s tcount=%d in ", 3687 Deoptimization::trap_reason_name(reason), tc); 3688 method()->print(); // which method has too many predicate traps 3689 tty->cr(); 3690 } 3691 #endif 3692 // We cannot afford to take more traps here, 3693 // do not generate predicate. 3694 return; 3695 } 3696 3697 Node *cont = _gvn.intcon(1); 3698 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3699 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3700 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3701 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3702 C->add_predicate_opaq(opq); 3703 { 3704 PreserveJVMState pjvms(this); 3705 set_control(iffalse); 3706 inc_sp(nargs); 3707 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3708 } 3709 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3710 set_control(iftrue); 3711 } 3712 3713 //------------------------------add_predicate--------------------------------- 3714 void GraphKit::add_predicate(int nargs) { 3715 if (UseLoopPredicate) { 3716 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3717 } 3718 // loop's limit check predicate should be near the loop. 3719 if (LoopLimitCheck) { 3720 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3721 } 3722 } 3723 3724 //----------------------------- store barriers ---------------------------- 3725 #define __ ideal. 3726 3727 void GraphKit::sync_kit(IdealKit& ideal) { 3728 set_all_memory(__ merged_memory()); 3729 set_i_o(__ i_o()); 3730 set_control(__ ctrl()); 3731 } 3732 3733 void GraphKit::final_sync(IdealKit& ideal) { 3734 // Final sync IdealKit and graphKit. 3735 sync_kit(ideal); 3736 } 3737 3738 Node* GraphKit::byte_map_base_node() { 3739 // Get base of card map 3740 CardTableModRefBS* ct = 3741 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3742 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3743 if (ct->byte_map_base != NULL) { 3744 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3745 } else { 3746 return null(); 3747 } 3748 } 3749 3750 // vanilla/CMS post barrier 3751 // Insert a write-barrier store. This is to let generational GC work; we have 3752 // to flag all oop-stores before the next GC point. 3753 void GraphKit::write_barrier_post(Node* oop_store, 3754 Node* obj, 3755 Node* adr, 3756 uint adr_idx, 3757 Node* val, 3758 bool use_precise) { 3759 // No store check needed if we're storing a NULL or an old object 3760 // (latter case is probably a string constant). The concurrent 3761 // mark sweep garbage collector, however, needs to have all nonNull 3762 // oop updates flagged via card-marks. 3763 if (val != NULL && val->is_Con()) { 3764 // must be either an oop or NULL 3765 const Type* t = val->bottom_type(); 3766 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3767 // stores of null never (?) need barriers 3768 return; 3769 } 3770 3771 if (use_ReduceInitialCardMarks() 3772 && obj == just_allocated_object(control())) { 3773 // We can skip marks on a freshly-allocated object in Eden. 3774 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3775 // That routine informs GC to take appropriate compensating steps, 3776 // upon a slow-path allocation, so as to make this card-mark 3777 // elision safe. 3778 return; 3779 } 3780 3781 if (!use_precise) { 3782 // All card marks for a (non-array) instance are in one place: 3783 adr = obj; 3784 } 3785 // (Else it's an array (or unknown), and we want more precise card marks.) 3786 assert(adr != NULL, ""); 3787 3788 IdealKit ideal(this, true); 3789 3790 // Convert the pointer to an int prior to doing math on it 3791 Node* cast = __ CastPX(__ ctrl(), adr); 3792 3793 // Divide by card size 3794 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3795 "Only one we handle so far."); 3796 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3797 3798 // Combine card table base and card offset 3799 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3800 3801 // Get the alias_index for raw card-mark memory 3802 int adr_type = Compile::AliasIdxRaw; 3803 Node* zero = __ ConI(0); // Dirty card value 3804 BasicType bt = T_BYTE; 3805 3806 if (UseConcMarkSweepGC && UseCondCardMark) { 3807 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3808 __ sync_kit(this); 3809 } 3810 3811 if (UseCondCardMark) { 3812 // The classic GC reference write barrier is typically implemented 3813 // as a store into the global card mark table. Unfortunately 3814 // unconditional stores can result in false sharing and excessive 3815 // coherence traffic as well as false transactional aborts. 3816 // UseCondCardMark enables MP "polite" conditional card mark 3817 // stores. In theory we could relax the load from ctrl() to 3818 // no_ctrl, but that doesn't buy much latitude. 3819 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3820 __ if_then(card_val, BoolTest::ne, zero); 3821 } 3822 3823 // Smash zero into card 3824 if( !UseConcMarkSweepGC ) { 3825 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3826 } else { 3827 // Specialized path for CM store barrier 3828 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3829 } 3830 3831 if (UseCondCardMark) { 3832 __ end_if(); 3833 } 3834 3835 // Final sync IdealKit and GraphKit. 3836 final_sync(ideal); 3837 } 3838 /* 3839 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3840 * required by SATB to make sure all objects live at the start of the 3841 * marking are kept alive, all reference updates need to any previous 3842 * reference stored before writing. 3843 * 3844 * If the previous value is NULL there is no need to save the old value. 3845 * References that are NULL are filtered during runtime by the barrier 3846 * code to avoid unnecessary queuing. 3847 * 3848 * However in the case of newly allocated objects it might be possible to 3849 * prove that the reference about to be overwritten is NULL during compile 3850 * time and avoid adding the barrier code completely. 3851 * 3852 * The compiler needs to determine that the object in which a field is about 3853 * to be written is newly allocated, and that no prior store to the same field 3854 * has happened since the allocation. 3855 * 3856 * Returns true if the pre-barrier can be removed 3857 */ 3858 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3859 BasicType bt, uint adr_idx) { 3860 intptr_t offset = 0; 3861 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3862 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3863 3864 if (offset == Type::OffsetBot) { 3865 return false; // cannot unalias unless there are precise offsets 3866 } 3867 3868 if (alloc == NULL) { 3869 return false; // No allocation found 3870 } 3871 3872 intptr_t size_in_bytes = type2aelembytes(bt); 3873 3874 Node* mem = memory(adr_idx); // start searching here... 3875 3876 for (int cnt = 0; cnt < 50; cnt++) { 3877 3878 if (mem->is_Store()) { 3879 3880 Node* st_adr = mem->in(MemNode::Address); 3881 intptr_t st_offset = 0; 3882 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3883 3884 if (st_base == NULL) { 3885 break; // inscrutable pointer 3886 } 3887 3888 // Break we have found a store with same base and offset as ours so break 3889 if (st_base == base && st_offset == offset) { 3890 break; 3891 } 3892 3893 if (st_offset != offset && st_offset != Type::OffsetBot) { 3894 const int MAX_STORE = BytesPerLong; 3895 if (st_offset >= offset + size_in_bytes || 3896 st_offset <= offset - MAX_STORE || 3897 st_offset <= offset - mem->as_Store()->memory_size()) { 3898 // Success: The offsets are provably independent. 3899 // (You may ask, why not just test st_offset != offset and be done? 3900 // The answer is that stores of different sizes can co-exist 3901 // in the same sequence of RawMem effects. We sometimes initialize 3902 // a whole 'tile' of array elements with a single jint or jlong.) 3903 mem = mem->in(MemNode::Memory); 3904 continue; // advance through independent store memory 3905 } 3906 } 3907 3908 if (st_base != base 3909 && MemNode::detect_ptr_independence(base, alloc, st_base, 3910 AllocateNode::Ideal_allocation(st_base, phase), 3911 phase)) { 3912 // Success: The bases are provably independent. 3913 mem = mem->in(MemNode::Memory); 3914 continue; // advance through independent store memory 3915 } 3916 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3917 3918 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3919 AllocateNode* st_alloc = st_init->allocation(); 3920 3921 // Make sure that we are looking at the same allocation site. 3922 // The alloc variable is guaranteed to not be null here from earlier check. 3923 if (alloc == st_alloc) { 3924 // Check that the initialization is storing NULL so that no previous store 3925 // has been moved up and directly write a reference 3926 Node* captured_store = st_init->find_captured_store(offset, 3927 type2aelembytes(T_OBJECT), 3928 phase); 3929 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3930 return true; 3931 } 3932 } 3933 } 3934 3935 // Unless there is an explicit 'continue', we must bail out here, 3936 // because 'mem' is an inscrutable memory state (e.g., a call). 3937 break; 3938 } 3939 3940 return false; 3941 } 3942 3943 // G1 pre/post barriers 3944 void GraphKit::g1_write_barrier_pre(bool do_load, 3945 Node* obj, 3946 Node* adr, 3947 uint alias_idx, 3948 Node* val, 3949 const TypeOopPtr* val_type, 3950 Node* pre_val, 3951 BasicType bt) { 3952 3953 // Some sanity checks 3954 // Note: val is unused in this routine. 3955 3956 if (do_load) { 3957 // We need to generate the load of the previous value 3958 assert(obj != NULL, "must have a base"); 3959 assert(adr != NULL, "where are loading from?"); 3960 assert(pre_val == NULL, "loaded already?"); 3961 assert(val_type != NULL, "need a type"); 3962 3963 if (use_ReduceInitialCardMarks() 3964 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 3965 return; 3966 } 3967 3968 } else { 3969 // In this case both val_type and alias_idx are unused. 3970 assert(pre_val != NULL, "must be loaded already"); 3971 // Nothing to be done if pre_val is null. 3972 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3973 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3974 } 3975 assert(bt == T_OBJECT, "or we shouldn't be here"); 3976 3977 IdealKit ideal(this, true); 3978 3979 Node* tls = __ thread(); // ThreadLocalStorage 3980 3981 Node* no_ctrl = NULL; 3982 Node* no_base = __ top(); 3983 Node* zero = __ ConI(0); 3984 Node* zeroX = __ ConX(0); 3985 3986 float likely = PROB_LIKELY(0.999); 3987 float unlikely = PROB_UNLIKELY(0.999); 3988 3989 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3990 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3991 3992 // Offsets into the thread 3993 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3994 PtrQueue::byte_offset_of_active()); 3995 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3996 PtrQueue::byte_offset_of_index()); 3997 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3998 PtrQueue::byte_offset_of_buf()); 3999 4000 // Now the actual pointers into the thread 4001 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4002 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4003 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4004 4005 // Now some of the values 4006 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4007 4008 // if (!marking) 4009 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4010 BasicType index_bt = TypeX_X->basic_type(); 4011 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 4012 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4013 4014 if (do_load) { 4015 // load original value 4016 // alias_idx correct?? 4017 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4018 } 4019 4020 // if (pre_val != NULL) 4021 __ if_then(pre_val, BoolTest::ne, null()); { 4022 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4023 4024 // is the queue for this thread full? 4025 __ if_then(index, BoolTest::ne, zeroX, likely); { 4026 4027 // decrement the index 4028 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4029 4030 // Now get the buffer location we will log the previous value into and store it 4031 Node *log_addr = __ AddP(no_base, buffer, next_index); 4032 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4033 // update the index 4034 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4035 4036 } __ else_(); { 4037 4038 // logging buffer is full, call the runtime 4039 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4040 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4041 } __ end_if(); // (!index) 4042 } __ end_if(); // (pre_val != NULL) 4043 } __ end_if(); // (!marking) 4044 4045 // Final sync IdealKit and GraphKit. 4046 final_sync(ideal); 4047 } 4048 4049 /* 4050 * G1 similar to any GC with a Young Generation requires a way to keep track of 4051 * references from Old Generation to Young Generation to make sure all live 4052 * objects are found. G1 also requires to keep track of object references 4053 * between different regions to enable evacuation of old regions, which is done 4054 * as part of mixed collections. References are tracked in remembered sets and 4055 * is continuously updated as reference are written to with the help of the 4056 * post-barrier. 4057 * 4058 * To reduce the number of updates to the remembered set the post-barrier 4059 * filters updates to fields in objects located in the Young Generation, 4060 * the same region as the reference, when the NULL is being written or 4061 * if the card is already marked as dirty by an earlier write. 4062 * 4063 * Under certain circumstances it is possible to avoid generating the 4064 * post-barrier completely if it is possible during compile time to prove 4065 * the object is newly allocated and that no safepoint exists between the 4066 * allocation and the store. 4067 * 4068 * In the case of slow allocation the allocation code must handle the barrier 4069 * as part of the allocation in the case the allocated object is not located 4070 * in the nursery, this would happen for humongous objects. This is similar to 4071 * how CMS is required to handle this case, see the comments for the method 4072 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4073 * A deferred card mark is required for these objects and handled in the above 4074 * mentioned methods. 4075 * 4076 * Returns true if the post barrier can be removed 4077 */ 4078 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4079 Node* adr) { 4080 intptr_t offset = 0; 4081 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4082 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4083 4084 if (offset == Type::OffsetBot) { 4085 return false; // cannot unalias unless there are precise offsets 4086 } 4087 4088 if (alloc == NULL) { 4089 return false; // No allocation found 4090 } 4091 4092 // Start search from Store node 4093 Node* mem = store->in(MemNode::Control); 4094 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4095 4096 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4097 AllocateNode* st_alloc = st_init->allocation(); 4098 4099 // Make sure we are looking at the same allocation 4100 if (alloc == st_alloc) { 4101 return true; 4102 } 4103 } 4104 4105 return false; 4106 } 4107 4108 // 4109 // Update the card table and add card address to the queue 4110 // 4111 void GraphKit::g1_mark_card(IdealKit& ideal, 4112 Node* card_adr, 4113 Node* oop_store, 4114 uint oop_alias_idx, 4115 Node* index, 4116 Node* index_adr, 4117 Node* buffer, 4118 const TypeFunc* tf) { 4119 4120 Node* zero = __ ConI(0); 4121 Node* zeroX = __ ConX(0); 4122 Node* no_base = __ top(); 4123 BasicType card_bt = T_BYTE; 4124 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4125 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4126 4127 // Now do the queue work 4128 __ if_then(index, BoolTest::ne, zeroX); { 4129 4130 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4131 Node* log_addr = __ AddP(no_base, buffer, next_index); 4132 4133 // Order, see storeCM. 4134 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4135 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4136 4137 } __ else_(); { 4138 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4139 } __ end_if(); 4140 4141 } 4142 4143 void GraphKit::g1_write_barrier_post(Node* oop_store, 4144 Node* obj, 4145 Node* adr, 4146 uint alias_idx, 4147 Node* val, 4148 BasicType bt, 4149 bool use_precise) { 4150 // If we are writing a NULL then we need no post barrier 4151 4152 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4153 // Must be NULL 4154 const Type* t = val->bottom_type(); 4155 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4156 // No post barrier if writing NULLx 4157 return; 4158 } 4159 4160 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4161 // We can skip marks on a freshly-allocated object in Eden. 4162 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4163 // That routine informs GC to take appropriate compensating steps, 4164 // upon a slow-path allocation, so as to make this card-mark 4165 // elision safe. 4166 return; 4167 } 4168 4169 if (use_ReduceInitialCardMarks() 4170 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4171 return; 4172 } 4173 4174 if (!use_precise) { 4175 // All card marks for a (non-array) instance are in one place: 4176 adr = obj; 4177 } 4178 // (Else it's an array (or unknown), and we want more precise card marks.) 4179 assert(adr != NULL, ""); 4180 4181 IdealKit ideal(this, true); 4182 4183 Node* tls = __ thread(); // ThreadLocalStorage 4184 4185 Node* no_base = __ top(); 4186 float likely = PROB_LIKELY(0.999); 4187 float unlikely = PROB_UNLIKELY(0.999); 4188 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4189 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4190 Node* zeroX = __ ConX(0); 4191 4192 // Get the alias_index for raw card-mark memory 4193 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4194 4195 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4196 4197 // Offsets into the thread 4198 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4199 PtrQueue::byte_offset_of_index()); 4200 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4201 PtrQueue::byte_offset_of_buf()); 4202 4203 // Pointers into the thread 4204 4205 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4206 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4207 4208 // Now some values 4209 // Use ctrl to avoid hoisting these values past a safepoint, which could 4210 // potentially reset these fields in the JavaThread. 4211 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4212 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4213 4214 // Convert the store obj pointer to an int prior to doing math on it 4215 // Must use ctrl to prevent "integerized oop" existing across safepoint 4216 Node* cast = __ CastPX(__ ctrl(), adr); 4217 4218 // Divide pointer by card size 4219 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4220 4221 // Combine card table base and card offset 4222 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4223 4224 // If we know the value being stored does it cross regions? 4225 4226 if (val != NULL) { 4227 // Does the store cause us to cross regions? 4228 4229 // Should be able to do an unsigned compare of region_size instead of 4230 // and extra shift. Do we have an unsigned compare?? 4231 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4232 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4233 4234 // if (xor_res == 0) same region so skip 4235 __ if_then(xor_res, BoolTest::ne, zeroX); { 4236 4237 // No barrier if we are storing a NULL 4238 __ if_then(val, BoolTest::ne, null(), unlikely); { 4239 4240 // Ok must mark the card if not already dirty 4241 4242 // load the original value of the card 4243 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4244 4245 __ if_then(card_val, BoolTest::ne, young_card); { 4246 sync_kit(ideal); 4247 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4248 insert_mem_bar(Op_MemBarVolatile, oop_store); 4249 __ sync_kit(this); 4250 4251 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4252 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4253 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4254 } __ end_if(); 4255 } __ end_if(); 4256 } __ end_if(); 4257 } __ end_if(); 4258 } else { 4259 // Object.clone() instrinsic uses this path. 4260 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4261 } 4262 4263 // Final sync IdealKit and GraphKit. 4264 final_sync(ideal); 4265 } 4266 #undef __ 4267 4268 4269 4270 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4271 if (java_lang_String::has_offset_field()) { 4272 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4273 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4274 false, NULL, 0); 4275 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4276 int offset_field_idx = C->get_alias_index(offset_field_type); 4277 return make_load(ctrl, 4278 basic_plus_adr(str, str, offset_offset), 4279 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4280 } else { 4281 return intcon(0); 4282 } 4283 } 4284 4285 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4286 if (java_lang_String::has_count_field()) { 4287 int count_offset = java_lang_String::count_offset_in_bytes(); 4288 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4289 false, NULL, 0); 4290 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4291 int count_field_idx = C->get_alias_index(count_field_type); 4292 return make_load(ctrl, 4293 basic_plus_adr(str, str, count_offset), 4294 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4295 } else { 4296 return load_array_length(load_String_value(ctrl, str)); 4297 } 4298 } 4299 4300 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4301 int value_offset = java_lang_String::value_offset_in_bytes(); 4302 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4303 false, NULL, 0); 4304 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4305 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4306 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4307 ciTypeArrayKlass::make(T_CHAR), true, 0); 4308 int value_field_idx = C->get_alias_index(value_field_type); 4309 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4310 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4311 // String.value field is known to be @Stable. 4312 if (UseImplicitStableValues) { 4313 load = cast_array_to_stable(load, value_type); 4314 } 4315 return load; 4316 } 4317 4318 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4319 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4320 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4321 false, NULL, 0); 4322 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4323 int offset_field_idx = C->get_alias_index(offset_field_type); 4324 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4325 value, T_INT, offset_field_idx, MemNode::unordered); 4326 } 4327 4328 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4329 int value_offset = java_lang_String::value_offset_in_bytes(); 4330 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4331 false, NULL, 0); 4332 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4333 4334 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4335 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4336 } 4337 4338 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4339 int count_offset = java_lang_String::count_offset_in_bytes(); 4340 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4341 false, NULL, 0); 4342 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4343 int count_field_idx = C->get_alias_index(count_field_type); 4344 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4345 value, T_INT, count_field_idx, MemNode::unordered); 4346 } 4347 4348 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4349 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4350 // assumption of CCP analysis. 4351 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4352 }