1 /* 2 * Copyright (c) 2001, 2011, 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_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "memory/barrierSet.hpp" 31 #include "memory/cardTableModRefBS.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/graphKit.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/locknode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/parse.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "runtime/deoptimization.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 43 //----------------------------GraphKit----------------------------------------- 44 // Main utility constructor. 45 GraphKit::GraphKit(JVMState* jvms) 46 : Phase(Phase::Parser), 47 _env(C->env()), 48 _gvn(*C->initial_gvn()) 49 { 50 _exceptions = jvms->map()->next_exception(); 51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 52 set_jvms(jvms); 53 } 54 55 // Private constructor for parser. 56 GraphKit::GraphKit() 57 : Phase(Phase::Parser), 58 _env(C->env()), 59 _gvn(*C->initial_gvn()) 60 { 61 _exceptions = NULL; 62 set_map(NULL); 63 debug_only(_sp = -99); 64 debug_only(set_bci(-99)); 65 } 66 67 68 69 //---------------------------clean_stack--------------------------------------- 70 // Clear away rubbish from the stack area of the JVM state. 71 // This destroys any arguments that may be waiting on the stack. 72 void GraphKit::clean_stack(int from_sp) { 73 SafePointNode* map = this->map(); 74 JVMState* jvms = this->jvms(); 75 int stk_size = jvms->stk_size(); 76 int stkoff = jvms->stkoff(); 77 Node* top = this->top(); 78 for (int i = from_sp; i < stk_size; i++) { 79 if (map->in(stkoff + i) != top) { 80 map->set_req(stkoff + i, top); 81 } 82 } 83 } 84 85 86 //--------------------------------sync_jvms----------------------------------- 87 // Make sure our current jvms agrees with our parse state. 88 JVMState* GraphKit::sync_jvms() const { 89 JVMState* jvms = this->jvms(); 90 jvms->set_bci(bci()); // Record the new bci in the JVMState 91 jvms->set_sp(sp()); // Record the new sp in the JVMState 92 assert(jvms_in_sync(), "jvms is now in sync"); 93 return jvms; 94 } 95 96 #ifdef ASSERT 97 bool GraphKit::jvms_in_sync() const { 98 Parse* parse = is_Parse(); 99 if (parse == NULL) { 100 if (bci() != jvms()->bci()) return false; 101 if (sp() != (int)jvms()->sp()) return false; 102 return true; 103 } 104 if (jvms()->method() != parse->method()) return false; 105 if (jvms()->bci() != parse->bci()) return false; 106 int jvms_sp = jvms()->sp(); 107 if (jvms_sp != parse->sp()) return false; 108 int jvms_depth = jvms()->depth(); 109 if (jvms_depth != parse->depth()) return false; 110 return true; 111 } 112 113 // Local helper checks for special internal merge points 114 // used to accumulate and merge exception states. 115 // They are marked by the region's in(0) edge being the map itself. 116 // Such merge points must never "escape" into the parser at large, 117 // until they have been handed to gvn.transform. 118 static bool is_hidden_merge(Node* reg) { 119 if (reg == NULL) return false; 120 if (reg->is_Phi()) { 121 reg = reg->in(0); 122 if (reg == NULL) return false; 123 } 124 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 125 } 126 127 void GraphKit::verify_map() const { 128 if (map() == NULL) return; // null map is OK 129 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 130 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 131 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 132 } 133 134 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 135 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 136 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 137 } 138 #endif 139 140 //---------------------------stop_and_kill_map--------------------------------- 141 // Set _map to NULL, signalling a stop to further bytecode execution. 142 // First smash the current map's control to a constant, to mark it dead. 143 void GraphKit::stop_and_kill_map() { 144 SafePointNode* dead_map = stop(); 145 if (dead_map != NULL) { 146 dead_map->disconnect_inputs(NULL); // Mark the map as killed. 147 assert(dead_map->is_killed(), "must be so marked"); 148 } 149 } 150 151 152 //--------------------------------stopped-------------------------------------- 153 // Tell if _map is NULL, or control is top. 154 bool GraphKit::stopped() { 155 if (map() == NULL) return true; 156 else if (control() == top()) return true; 157 else return false; 158 } 159 160 161 //-----------------------------has_ex_handler---------------------------------- 162 // Tell if this method or any caller method has exception handlers. 163 bool GraphKit::has_ex_handler() { 164 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 165 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 166 return true; 167 } 168 } 169 return false; 170 } 171 172 //------------------------------save_ex_oop------------------------------------ 173 // Save an exception without blowing stack contents or other JVM state. 174 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 175 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 176 ex_map->add_req(ex_oop); 177 debug_only(verify_exception_state(ex_map)); 178 } 179 180 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 181 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 182 Node* ex_oop = ex_map->in(ex_map->req()-1); 183 if (clear_it) ex_map->del_req(ex_map->req()-1); 184 return ex_oop; 185 } 186 187 //-----------------------------saved_ex_oop------------------------------------ 188 // Recover a saved exception from its map. 189 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 190 return common_saved_ex_oop(ex_map, false); 191 } 192 193 //--------------------------clear_saved_ex_oop--------------------------------- 194 // Erase a previously saved exception from its map. 195 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 196 return common_saved_ex_oop(ex_map, true); 197 } 198 199 #ifdef ASSERT 200 //---------------------------has_saved_ex_oop---------------------------------- 201 // Erase a previously saved exception from its map. 202 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 203 return ex_map->req() == ex_map->jvms()->endoff()+1; 204 } 205 #endif 206 207 //-------------------------make_exception_state-------------------------------- 208 // Turn the current JVM state into an exception state, appending the ex_oop. 209 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 210 sync_jvms(); 211 SafePointNode* ex_map = stop(); // do not manipulate this map any more 212 set_saved_ex_oop(ex_map, ex_oop); 213 return ex_map; 214 } 215 216 217 //--------------------------add_exception_state-------------------------------- 218 // Add an exception to my list of exceptions. 219 void GraphKit::add_exception_state(SafePointNode* ex_map) { 220 if (ex_map == NULL || ex_map->control() == top()) { 221 return; 222 } 223 #ifdef ASSERT 224 verify_exception_state(ex_map); 225 if (has_exceptions()) { 226 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 227 } 228 #endif 229 230 // If there is already an exception of exactly this type, merge with it. 231 // In particular, null-checks and other low-level exceptions common up here. 232 Node* ex_oop = saved_ex_oop(ex_map); 233 const Type* ex_type = _gvn.type(ex_oop); 234 if (ex_oop == top()) { 235 // No action needed. 236 return; 237 } 238 assert(ex_type->isa_instptr(), "exception must be an instance"); 239 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 240 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 241 // We check sp also because call bytecodes can generate exceptions 242 // both before and after arguments are popped! 243 if (ex_type2 == ex_type 244 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 245 combine_exception_states(ex_map, e2); 246 return; 247 } 248 } 249 250 // No pre-existing exception of the same type. Chain it on the list. 251 push_exception_state(ex_map); 252 } 253 254 //-----------------------add_exception_states_from----------------------------- 255 void GraphKit::add_exception_states_from(JVMState* jvms) { 256 SafePointNode* ex_map = jvms->map()->next_exception(); 257 if (ex_map != NULL) { 258 jvms->map()->set_next_exception(NULL); 259 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 260 next_map = ex_map->next_exception(); 261 ex_map->set_next_exception(NULL); 262 add_exception_state(ex_map); 263 } 264 } 265 } 266 267 //-----------------------transfer_exceptions_into_jvms------------------------- 268 JVMState* GraphKit::transfer_exceptions_into_jvms() { 269 if (map() == NULL) { 270 // We need a JVMS to carry the exceptions, but the map has gone away. 271 // Create a scratch JVMS, cloned from any of the exception states... 272 if (has_exceptions()) { 273 _map = _exceptions; 274 _map = clone_map(); 275 _map->set_next_exception(NULL); 276 clear_saved_ex_oop(_map); 277 debug_only(verify_map()); 278 } else { 279 // ...or created from scratch 280 JVMState* jvms = new (C) JVMState(_method, NULL); 281 jvms->set_bci(_bci); 282 jvms->set_sp(_sp); 283 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms)); 284 set_jvms(jvms); 285 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 286 set_all_memory(top()); 287 while (map()->req() < jvms->endoff()) map()->add_req(top()); 288 } 289 // (This is a kludge, in case you didn't notice.) 290 set_control(top()); 291 } 292 JVMState* jvms = sync_jvms(); 293 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 294 jvms->map()->set_next_exception(_exceptions); 295 _exceptions = NULL; // done with this set of exceptions 296 return jvms; 297 } 298 299 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 300 assert(is_hidden_merge(dstphi), "must be a special merge node"); 301 assert(is_hidden_merge(srcphi), "must be a special merge node"); 302 uint limit = srcphi->req(); 303 for (uint i = PhiNode::Input; i < limit; i++) { 304 dstphi->add_req(srcphi->in(i)); 305 } 306 } 307 static inline void add_one_req(Node* dstphi, Node* src) { 308 assert(is_hidden_merge(dstphi), "must be a special merge node"); 309 assert(!is_hidden_merge(src), "must not be a special merge node"); 310 dstphi->add_req(src); 311 } 312 313 //-----------------------combine_exception_states------------------------------ 314 // This helper function combines exception states by building phis on a 315 // specially marked state-merging region. These regions and phis are 316 // untransformed, and can build up gradually. The region is marked by 317 // having a control input of its exception map, rather than NULL. Such 318 // regions do not appear except in this function, and in use_exception_state. 319 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 320 if (failing()) return; // dying anyway... 321 JVMState* ex_jvms = ex_map->_jvms; 322 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 323 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 324 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 325 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 326 assert(ex_map->req() == phi_map->req(), "matching maps"); 327 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 328 Node* hidden_merge_mark = root(); 329 Node* region = phi_map->control(); 330 MergeMemNode* phi_mem = phi_map->merged_memory(); 331 MergeMemNode* ex_mem = ex_map->merged_memory(); 332 if (region->in(0) != hidden_merge_mark) { 333 // The control input is not (yet) a specially-marked region in phi_map. 334 // Make it so, and build some phis. 335 region = new (C, 2) RegionNode(2); 336 _gvn.set_type(region, Type::CONTROL); 337 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 338 region->init_req(1, phi_map->control()); 339 phi_map->set_control(region); 340 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 341 record_for_igvn(io_phi); 342 _gvn.set_type(io_phi, Type::ABIO); 343 phi_map->set_i_o(io_phi); 344 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 345 Node* m = mms.memory(); 346 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 347 record_for_igvn(m_phi); 348 _gvn.set_type(m_phi, Type::MEMORY); 349 mms.set_memory(m_phi); 350 } 351 } 352 353 // Either or both of phi_map and ex_map might already be converted into phis. 354 Node* ex_control = ex_map->control(); 355 // if there is special marking on ex_map also, we add multiple edges from src 356 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 357 // how wide was the destination phi_map, originally? 358 uint orig_width = region->req(); 359 360 if (add_multiple) { 361 add_n_reqs(region, ex_control); 362 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 363 } else { 364 // ex_map has no merges, so we just add single edges everywhere 365 add_one_req(region, ex_control); 366 add_one_req(phi_map->i_o(), ex_map->i_o()); 367 } 368 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 369 if (mms.is_empty()) { 370 // get a copy of the base memory, and patch some inputs into it 371 const TypePtr* adr_type = mms.adr_type(C); 372 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 373 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 374 mms.set_memory(phi); 375 // Prepare to append interesting stuff onto the newly sliced phi: 376 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 377 } 378 // Append stuff from ex_map: 379 if (add_multiple) { 380 add_n_reqs(mms.memory(), mms.memory2()); 381 } else { 382 add_one_req(mms.memory(), mms.memory2()); 383 } 384 } 385 uint limit = ex_map->req(); 386 for (uint i = TypeFunc::Parms; i < limit; i++) { 387 // Skip everything in the JVMS after tos. (The ex_oop follows.) 388 if (i == tos) i = ex_jvms->monoff(); 389 Node* src = ex_map->in(i); 390 Node* dst = phi_map->in(i); 391 if (src != dst) { 392 PhiNode* phi; 393 if (dst->in(0) != region) { 394 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 395 record_for_igvn(phi); 396 _gvn.set_type(phi, phi->type()); 397 phi_map->set_req(i, dst); 398 // Prepare to append interesting stuff onto the new phi: 399 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 400 } else { 401 assert(dst->is_Phi(), "nobody else uses a hidden region"); 402 phi = (PhiNode*)dst; 403 } 404 if (add_multiple && src->in(0) == ex_control) { 405 // Both are phis. 406 add_n_reqs(dst, src); 407 } else { 408 while (dst->req() < region->req()) add_one_req(dst, src); 409 } 410 const Type* srctype = _gvn.type(src); 411 if (phi->type() != srctype) { 412 const Type* dsttype = phi->type()->meet(srctype); 413 if (phi->type() != dsttype) { 414 phi->set_type(dsttype); 415 _gvn.set_type(phi, dsttype); 416 } 417 } 418 } 419 } 420 } 421 422 //--------------------------use_exception_state-------------------------------- 423 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 424 if (failing()) { stop(); return top(); } 425 Node* region = phi_map->control(); 426 Node* hidden_merge_mark = root(); 427 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 428 Node* ex_oop = clear_saved_ex_oop(phi_map); 429 if (region->in(0) == hidden_merge_mark) { 430 // Special marking for internal ex-states. Process the phis now. 431 region->set_req(0, region); // now it's an ordinary region 432 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 433 // Note: Setting the jvms also sets the bci and sp. 434 set_control(_gvn.transform(region)); 435 uint tos = jvms()->stkoff() + sp(); 436 for (uint i = 1; i < tos; i++) { 437 Node* x = phi_map->in(i); 438 if (x->in(0) == region) { 439 assert(x->is_Phi(), "expected a special phi"); 440 phi_map->set_req(i, _gvn.transform(x)); 441 } 442 } 443 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 444 Node* x = mms.memory(); 445 if (x->in(0) == region) { 446 assert(x->is_Phi(), "nobody else uses a hidden region"); 447 mms.set_memory(_gvn.transform(x)); 448 } 449 } 450 if (ex_oop->in(0) == region) { 451 assert(ex_oop->is_Phi(), "expected a special phi"); 452 ex_oop = _gvn.transform(ex_oop); 453 } 454 } else { 455 set_jvms(phi_map->jvms()); 456 } 457 458 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 459 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 460 return ex_oop; 461 } 462 463 //---------------------------------java_bc------------------------------------- 464 Bytecodes::Code GraphKit::java_bc() const { 465 ciMethod* method = this->method(); 466 int bci = this->bci(); 467 if (method != NULL && bci != InvocationEntryBci) 468 return method->java_code_at_bci(bci); 469 else 470 return Bytecodes::_illegal; 471 } 472 473 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 474 bool must_throw) { 475 // if the exception capability is set, then we will generate code 476 // to check the JavaThread.should_post_on_exceptions flag to see 477 // if we actually need to report exception events (for this 478 // thread). If we don't need to report exception events, we will 479 // take the normal fast path provided by add_exception_events. If 480 // exception event reporting is enabled for this thread, we will 481 // take the uncommon_trap in the BuildCutout below. 482 483 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 484 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode()); 485 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 486 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false); 487 488 // Test the should_post_on_exceptions_flag vs. 0 489 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) ); 490 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) ); 491 492 // Branch to slow_path if should_post_on_exceptions_flag was true 493 { BuildCutout unless(this, tst, PROB_MAX); 494 // Do not try anything fancy if we're notifying the VM on every throw. 495 // Cf. case Bytecodes::_athrow in parse2.cpp. 496 uncommon_trap(reason, Deoptimization::Action_none, 497 (ciKlass*)NULL, (char*)NULL, must_throw); 498 } 499 500 } 501 502 //------------------------------builtin_throw---------------------------------- 503 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 504 bool must_throw = true; 505 506 if (env()->jvmti_can_post_on_exceptions()) { 507 // check if we must post exception events, take uncommon trap if so 508 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 509 // here if should_post_on_exceptions is false 510 // continue on with the normal codegen 511 } 512 513 // If this particular condition has not yet happened at this 514 // bytecode, then use the uncommon trap mechanism, and allow for 515 // a future recompilation if several traps occur here. 516 // If the throw is hot, try to use a more complicated inline mechanism 517 // which keeps execution inside the compiled code. 518 bool treat_throw_as_hot = false; 519 ciMethodData* md = method()->method_data(); 520 521 if (ProfileTraps) { 522 if (too_many_traps(reason)) { 523 treat_throw_as_hot = true; 524 } 525 // (If there is no MDO at all, assume it is early in 526 // execution, and that any deopts are part of the 527 // startup transient, and don't need to be remembered.) 528 529 // Also, if there is a local exception handler, treat all throws 530 // as hot if there has been at least one in this method. 531 if (C->trap_count(reason) != 0 532 && method()->method_data()->trap_count(reason) != 0 533 && has_ex_handler()) { 534 treat_throw_as_hot = true; 535 } 536 } 537 538 // If this throw happens frequently, an uncommon trap might cause 539 // a performance pothole. If there is a local exception handler, 540 // and if this particular bytecode appears to be deoptimizing often, 541 // let us handle the throw inline, with a preconstructed instance. 542 // Note: If the deopt count has blown up, the uncommon trap 543 // runtime is going to flush this nmethod, not matter what. 544 if (treat_throw_as_hot 545 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 546 // If the throw is local, we use a pre-existing instance and 547 // punt on the backtrace. This would lead to a missing backtrace 548 // (a repeat of 4292742) if the backtrace object is ever asked 549 // for its backtrace. 550 // Fixing this remaining case of 4292742 requires some flavor of 551 // escape analysis. Leave that for the future. 552 ciInstance* ex_obj = NULL; 553 switch (reason) { 554 case Deoptimization::Reason_null_check: 555 ex_obj = env()->NullPointerException_instance(); 556 break; 557 case Deoptimization::Reason_div0_check: 558 ex_obj = env()->ArithmeticException_instance(); 559 break; 560 case Deoptimization::Reason_range_check: 561 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 562 break; 563 case Deoptimization::Reason_class_check: 564 if (java_bc() == Bytecodes::_aastore) { 565 ex_obj = env()->ArrayStoreException_instance(); 566 } else { 567 ex_obj = env()->ClassCastException_instance(); 568 } 569 break; 570 } 571 if (failing()) { stop(); return; } // exception allocation might fail 572 if (ex_obj != NULL) { 573 // Cheat with a preallocated exception object. 574 if (C->log() != NULL) 575 C->log()->elem("hot_throw preallocated='1' reason='%s'", 576 Deoptimization::trap_reason_name(reason)); 577 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 578 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 579 580 // Clear the detail message of the preallocated exception object. 581 // Weblogic sometimes mutates the detail message of exceptions 582 // using reflection. 583 int offset = java_lang_Throwable::get_detailMessage_offset(); 584 const TypePtr* adr_typ = ex_con->add_offset(offset); 585 586 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 587 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 588 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT); 589 590 add_exception_state(make_exception_state(ex_node)); 591 return; 592 } 593 } 594 595 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 596 // It won't be much cheaper than bailing to the interp., since we'll 597 // have to pass up all the debug-info, and the runtime will have to 598 // create the stack trace. 599 600 // Usual case: Bail to interpreter. 601 // Reserve the right to recompile if we haven't seen anything yet. 602 603 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 604 if (treat_throw_as_hot 605 && (method()->method_data()->trap_recompiled_at(bci()) 606 || C->too_many_traps(reason))) { 607 // We cannot afford to take more traps here. Suffer in the interpreter. 608 if (C->log() != NULL) 609 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 610 Deoptimization::trap_reason_name(reason), 611 C->trap_count(reason)); 612 action = Deoptimization::Action_none; 613 } 614 615 // "must_throw" prunes the JVM state to include only the stack, if there 616 // are no local exception handlers. This should cut down on register 617 // allocation time and code size, by drastically reducing the number 618 // of in-edges on the call to the uncommon trap. 619 620 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 621 } 622 623 624 //----------------------------PreserveJVMState--------------------------------- 625 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 626 debug_only(kit->verify_map()); 627 _kit = kit; 628 _map = kit->map(); // preserve the map 629 _sp = kit->sp(); 630 kit->set_map(clone_map ? kit->clone_map() : NULL); 631 #ifdef ASSERT 632 _bci = kit->bci(); 633 Parse* parser = kit->is_Parse(); 634 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 635 _block = block; 636 #endif 637 } 638 PreserveJVMState::~PreserveJVMState() { 639 GraphKit* kit = _kit; 640 #ifdef ASSERT 641 assert(kit->bci() == _bci, "bci must not shift"); 642 Parse* parser = kit->is_Parse(); 643 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 644 assert(block == _block, "block must not shift"); 645 #endif 646 kit->set_map(_map); 647 kit->set_sp(_sp); 648 } 649 650 651 //-----------------------------BuildCutout------------------------------------- 652 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 653 : PreserveJVMState(kit) 654 { 655 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 656 SafePointNode* outer_map = _map; // preserved map is caller's 657 SafePointNode* inner_map = kit->map(); 658 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 659 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) )); 660 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) )); 661 } 662 BuildCutout::~BuildCutout() { 663 GraphKit* kit = _kit; 664 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 665 } 666 667 //---------------------------PreserveReexecuteState---------------------------- 668 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 669 assert(!kit->stopped(), "must call stopped() before"); 670 _kit = kit; 671 _sp = kit->sp(); 672 _reexecute = kit->jvms()->_reexecute; 673 } 674 PreserveReexecuteState::~PreserveReexecuteState() { 675 if (_kit->stopped()) return; 676 _kit->jvms()->_reexecute = _reexecute; 677 _kit->set_sp(_sp); 678 } 679 680 //------------------------------clone_map-------------------------------------- 681 // Implementation of PreserveJVMState 682 // 683 // Only clone_map(...) here. If this function is only used in the 684 // PreserveJVMState class we may want to get rid of this extra 685 // function eventually and do it all there. 686 687 SafePointNode* GraphKit::clone_map() { 688 if (map() == NULL) return NULL; 689 690 // Clone the memory edge first 691 Node* mem = MergeMemNode::make(C, map()->memory()); 692 gvn().set_type_bottom(mem); 693 694 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 695 JVMState* jvms = this->jvms(); 696 JVMState* clonejvms = jvms->clone_shallow(C); 697 clonemap->set_memory(mem); 698 clonemap->set_jvms(clonejvms); 699 clonejvms->set_map(clonemap); 700 record_for_igvn(clonemap); 701 gvn().set_type_bottom(clonemap); 702 return clonemap; 703 } 704 705 706 //-----------------------------set_map_clone----------------------------------- 707 void GraphKit::set_map_clone(SafePointNode* m) { 708 _map = m; 709 _map = clone_map(); 710 _map->set_next_exception(NULL); 711 debug_only(verify_map()); 712 } 713 714 715 //----------------------------kill_dead_locals--------------------------------- 716 // Detect any locals which are known to be dead, and force them to top. 717 void GraphKit::kill_dead_locals() { 718 // Consult the liveness information for the locals. If any 719 // of them are unused, then they can be replaced by top(). This 720 // should help register allocation time and cut down on the size 721 // of the deoptimization information. 722 723 // This call is made from many of the bytecode handling 724 // subroutines called from the Big Switch in do_one_bytecode. 725 // Every bytecode which might include a slow path is responsible 726 // for killing its dead locals. The more consistent we 727 // are about killing deads, the fewer useless phis will be 728 // constructed for them at various merge points. 729 730 // bci can be -1 (InvocationEntryBci). We return the entry 731 // liveness for the method. 732 733 if (method() == NULL || method()->code_size() == 0) { 734 // We are building a graph for a call to a native method. 735 // All locals are live. 736 return; 737 } 738 739 ResourceMark rm; 740 741 // Consult the liveness information for the locals. If any 742 // of them are unused, then they can be replaced by top(). This 743 // should help register allocation time and cut down on the size 744 // of the deoptimization information. 745 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 746 747 int len = (int)live_locals.size(); 748 assert(len <= jvms()->loc_size(), "too many live locals"); 749 for (int local = 0; local < len; local++) { 750 if (!live_locals.at(local)) { 751 set_local(local, top()); 752 } 753 } 754 } 755 756 #ifdef ASSERT 757 //-------------------------dead_locals_are_killed------------------------------ 758 // Return true if all dead locals are set to top in the map. 759 // Used to assert "clean" debug info at various points. 760 bool GraphKit::dead_locals_are_killed() { 761 if (method() == NULL || method()->code_size() == 0) { 762 // No locals need to be dead, so all is as it should be. 763 return true; 764 } 765 766 // Make sure somebody called kill_dead_locals upstream. 767 ResourceMark rm; 768 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 769 if (jvms->loc_size() == 0) continue; // no locals to consult 770 SafePointNode* map = jvms->map(); 771 ciMethod* method = jvms->method(); 772 int bci = jvms->bci(); 773 if (jvms == this->jvms()) { 774 bci = this->bci(); // it might not yet be synched 775 } 776 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 777 int len = (int)live_locals.size(); 778 if (!live_locals.is_valid() || len == 0) 779 // This method is trivial, or is poisoned by a breakpoint. 780 return true; 781 assert(len == jvms->loc_size(), "live map consistent with locals map"); 782 for (int local = 0; local < len; local++) { 783 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 784 if (PrintMiscellaneous && (Verbose || WizardMode)) { 785 tty->print_cr("Zombie local %d: ", local); 786 jvms->dump(); 787 } 788 return false; 789 } 790 } 791 } 792 return true; 793 } 794 795 #endif //ASSERT 796 797 // Helper function for enforcing certain bytecodes to reexecute if 798 // deoptimization happens 799 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 800 ciMethod* cur_method = jvms->method(); 801 int cur_bci = jvms->bci(); 802 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 803 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 804 return Interpreter::bytecode_should_reexecute(code) || 805 is_anewarray && code == Bytecodes::_multianewarray; 806 // Reexecute _multianewarray bytecode which was replaced with 807 // sequence of [a]newarray. See Parse::do_multianewarray(). 808 // 809 // Note: interpreter should not have it set since this optimization 810 // is limited by dimensions and guarded by flag so in some cases 811 // multianewarray() runtime calls will be generated and 812 // the bytecode should not be reexecutes (stack will not be reset). 813 } else 814 return false; 815 } 816 817 // Helper function for adding JVMState and debug information to node 818 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 819 // Add the safepoint edges to the call (or other safepoint). 820 821 // Make sure dead locals are set to top. This 822 // should help register allocation time and cut down on the size 823 // of the deoptimization information. 824 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 825 826 // Walk the inline list to fill in the correct set of JVMState's 827 // Also fill in the associated edges for each JVMState. 828 829 JVMState* youngest_jvms = sync_jvms(); 830 831 // If we are guaranteed to throw, we can prune everything but the 832 // input to the current bytecode. 833 bool can_prune_locals = false; 834 uint stack_slots_not_pruned = 0; 835 int inputs = 0, depth = 0; 836 if (must_throw) { 837 assert(method() == youngest_jvms->method(), "sanity"); 838 if (compute_stack_effects(inputs, depth)) { 839 can_prune_locals = true; 840 stack_slots_not_pruned = inputs; 841 } 842 } 843 844 if (env()->jvmti_can_access_local_variables()) { 845 // At any safepoint, this method can get breakpointed, which would 846 // then require an immediate deoptimization. 847 can_prune_locals = false; // do not prune locals 848 stack_slots_not_pruned = 0; 849 } 850 851 // do not scribble on the input jvms 852 JVMState* out_jvms = youngest_jvms->clone_deep(C); 853 call->set_jvms(out_jvms); // Start jvms list for call node 854 855 // For a known set of bytecodes, the interpreter should reexecute them if 856 // deoptimization happens. We set the reexecute state for them here 857 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 858 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 859 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 860 } 861 862 // Presize the call: 863 debug_only(uint non_debug_edges = call->req()); 864 call->add_req_batch(top(), youngest_jvms->debug_depth()); 865 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 866 867 // Set up edges so that the call looks like this: 868 // Call [state:] ctl io mem fptr retadr 869 // [parms:] parm0 ... parmN 870 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 871 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 872 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 873 // Note that caller debug info precedes callee debug info. 874 875 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 876 uint debug_ptr = call->req(); 877 878 // Loop over the map input edges associated with jvms, add them 879 // to the call node, & reset all offsets to match call node array. 880 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 881 uint debug_end = debug_ptr; 882 uint debug_start = debug_ptr - in_jvms->debug_size(); 883 debug_ptr = debug_start; // back up the ptr 884 885 uint p = debug_start; // walks forward in [debug_start, debug_end) 886 uint j, k, l; 887 SafePointNode* in_map = in_jvms->map(); 888 out_jvms->set_map(call); 889 890 if (can_prune_locals) { 891 assert(in_jvms->method() == out_jvms->method(), "sanity"); 892 // If the current throw can reach an exception handler in this JVMS, 893 // then we must keep everything live that can reach that handler. 894 // As a quick and dirty approximation, we look for any handlers at all. 895 if (in_jvms->method()->has_exception_handlers()) { 896 can_prune_locals = false; 897 } 898 } 899 900 // Add the Locals 901 k = in_jvms->locoff(); 902 l = in_jvms->loc_size(); 903 out_jvms->set_locoff(p); 904 if (!can_prune_locals) { 905 for (j = 0; j < l; j++) 906 call->set_req(p++, in_map->in(k+j)); 907 } else { 908 p += l; // already set to top above by add_req_batch 909 } 910 911 // Add the Expression Stack 912 k = in_jvms->stkoff(); 913 l = in_jvms->sp(); 914 out_jvms->set_stkoff(p); 915 if (!can_prune_locals) { 916 for (j = 0; j < l; j++) 917 call->set_req(p++, in_map->in(k+j)); 918 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 919 // Divide stack into {S0,...,S1}, where S0 is set to top. 920 uint s1 = stack_slots_not_pruned; 921 stack_slots_not_pruned = 0; // for next iteration 922 if (s1 > l) s1 = l; 923 uint s0 = l - s1; 924 p += s0; // skip the tops preinstalled by add_req_batch 925 for (j = s0; j < l; j++) 926 call->set_req(p++, in_map->in(k+j)); 927 } else { 928 p += l; // already set to top above by add_req_batch 929 } 930 931 // Add the Monitors 932 k = in_jvms->monoff(); 933 l = in_jvms->mon_size(); 934 out_jvms->set_monoff(p); 935 for (j = 0; j < l; j++) 936 call->set_req(p++, in_map->in(k+j)); 937 938 // Copy any scalar object fields. 939 k = in_jvms->scloff(); 940 l = in_jvms->scl_size(); 941 out_jvms->set_scloff(p); 942 for (j = 0; j < l; j++) 943 call->set_req(p++, in_map->in(k+j)); 944 945 // Finish the new jvms. 946 out_jvms->set_endoff(p); 947 948 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 949 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 950 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 951 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 952 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 953 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 954 955 // Update the two tail pointers in parallel. 956 out_jvms = out_jvms->caller(); 957 in_jvms = in_jvms->caller(); 958 } 959 960 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 961 962 // Test the correctness of JVMState::debug_xxx accessors: 963 assert(call->jvms()->debug_start() == non_debug_edges, ""); 964 assert(call->jvms()->debug_end() == call->req(), ""); 965 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 966 } 967 968 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 969 Bytecodes::Code code = java_bc(); 970 if (code == Bytecodes::_wide) { 971 code = method()->java_code_at_bci(bci() + 1); 972 } 973 974 BasicType rtype = T_ILLEGAL; 975 int rsize = 0; 976 977 if (code != Bytecodes::_illegal) { 978 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 979 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 980 if (rtype < T_CONFLICT) 981 rsize = type2size[rtype]; 982 } 983 984 switch (code) { 985 case Bytecodes::_illegal: 986 return false; 987 988 case Bytecodes::_ldc: 989 case Bytecodes::_ldc_w: 990 case Bytecodes::_ldc2_w: 991 inputs = 0; 992 break; 993 994 case Bytecodes::_dup: inputs = 1; break; 995 case Bytecodes::_dup_x1: inputs = 2; break; 996 case Bytecodes::_dup_x2: inputs = 3; break; 997 case Bytecodes::_dup2: inputs = 2; break; 998 case Bytecodes::_dup2_x1: inputs = 3; break; 999 case Bytecodes::_dup2_x2: inputs = 4; break; 1000 case Bytecodes::_swap: inputs = 2; break; 1001 case Bytecodes::_arraylength: inputs = 1; break; 1002 1003 case Bytecodes::_getstatic: 1004 case Bytecodes::_putstatic: 1005 case Bytecodes::_getfield: 1006 case Bytecodes::_putfield: 1007 { 1008 bool is_get = (depth >= 0), is_static = (depth & 1); 1009 bool ignore; 1010 ciBytecodeStream iter(method()); 1011 iter.reset_to_bci(bci()); 1012 iter.next(); 1013 ciField* field = iter.get_field(ignore); 1014 int size = field->type()->size(); 1015 inputs = (is_static ? 0 : 1); 1016 if (is_get) { 1017 depth = size - inputs; 1018 } else { 1019 inputs += size; // putxxx pops the value from the stack 1020 depth = - inputs; 1021 } 1022 } 1023 break; 1024 1025 case Bytecodes::_invokevirtual: 1026 case Bytecodes::_invokespecial: 1027 case Bytecodes::_invokestatic: 1028 case Bytecodes::_invokedynamic: 1029 case Bytecodes::_invokeinterface: 1030 { 1031 bool ignore; 1032 ciBytecodeStream iter(method()); 1033 iter.reset_to_bci(bci()); 1034 iter.next(); 1035 ciMethod* method = iter.get_method(ignore); 1036 // (Do not use ciMethod::arg_size(), because 1037 // it might be an unloaded method, which doesn't 1038 // know whether it is static or not.) 1039 inputs = method->invoke_arg_size(code); 1040 int size = method->return_type()->size(); 1041 depth = size - inputs; 1042 } 1043 break; 1044 1045 case Bytecodes::_multianewarray: 1046 { 1047 ciBytecodeStream iter(method()); 1048 iter.reset_to_bci(bci()); 1049 iter.next(); 1050 inputs = iter.get_dimensions(); 1051 assert(rsize == 1, ""); 1052 depth = rsize - inputs; 1053 } 1054 break; 1055 1056 case Bytecodes::_ireturn: 1057 case Bytecodes::_lreturn: 1058 case Bytecodes::_freturn: 1059 case Bytecodes::_dreturn: 1060 case Bytecodes::_areturn: 1061 assert(rsize = -depth, ""); 1062 inputs = rsize; 1063 break; 1064 1065 case Bytecodes::_jsr: 1066 case Bytecodes::_jsr_w: 1067 inputs = 0; 1068 depth = 1; // S.B. depth=1, not zero 1069 break; 1070 1071 default: 1072 // bytecode produces a typed result 1073 inputs = rsize - depth; 1074 assert(inputs >= 0, ""); 1075 break; 1076 } 1077 1078 #ifdef ASSERT 1079 // spot check 1080 int outputs = depth + inputs; 1081 assert(outputs >= 0, "sanity"); 1082 switch (code) { 1083 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1084 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1085 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1086 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1087 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1088 } 1089 #endif //ASSERT 1090 1091 return true; 1092 } 1093 1094 1095 1096 //------------------------------basic_plus_adr--------------------------------- 1097 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1098 // short-circuit a common case 1099 if (offset == intcon(0)) return ptr; 1100 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); 1101 } 1102 1103 Node* GraphKit::ConvI2L(Node* offset) { 1104 // short-circuit a common case 1105 jint offset_con = find_int_con(offset, Type::OffsetBot); 1106 if (offset_con != Type::OffsetBot) { 1107 return longcon((long) offset_con); 1108 } 1109 return _gvn.transform( new (C, 2) ConvI2LNode(offset)); 1110 } 1111 Node* GraphKit::ConvL2I(Node* offset) { 1112 // short-circuit a common case 1113 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1114 if (offset_con != (jlong)Type::OffsetBot) { 1115 return intcon((int) offset_con); 1116 } 1117 return _gvn.transform( new (C, 2) ConvL2INode(offset)); 1118 } 1119 1120 //-------------------------load_object_klass----------------------------------- 1121 Node* GraphKit::load_object_klass(Node* obj) { 1122 // Special-case a fresh allocation to avoid building nodes: 1123 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1124 if (akls != NULL) return akls; 1125 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1126 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1127 } 1128 1129 //-------------------------load_array_length----------------------------------- 1130 Node* GraphKit::load_array_length(Node* array) { 1131 // Special-case a fresh allocation to avoid building nodes: 1132 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1133 Node *alen; 1134 if (alloc == NULL) { 1135 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1136 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1137 } else { 1138 alen = alloc->Ideal_length(); 1139 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1140 if (ccast != alen) { 1141 alen = _gvn.transform(ccast); 1142 } 1143 } 1144 return alen; 1145 } 1146 1147 //------------------------------do_null_check---------------------------------- 1148 // Helper function to do a NULL pointer check. Returned value is 1149 // the incoming address with NULL casted away. You are allowed to use the 1150 // not-null value only if you are control dependent on the test. 1151 extern int explicit_null_checks_inserted, 1152 explicit_null_checks_elided; 1153 Node* GraphKit::null_check_common(Node* value, BasicType type, 1154 // optional arguments for variations: 1155 bool assert_null, 1156 Node* *null_control) { 1157 assert(!assert_null || null_control == NULL, "not both at once"); 1158 if (stopped()) return top(); 1159 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1160 // For some performance testing, we may wish to suppress null checking. 1161 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1162 return value; 1163 } 1164 explicit_null_checks_inserted++; 1165 1166 // Construct NULL check 1167 Node *chk = NULL; 1168 switch(type) { 1169 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1170 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; 1171 case T_ARRAY : // fall through 1172 type = T_OBJECT; // simplify further tests 1173 case T_OBJECT : { 1174 const Type *t = _gvn.type( value ); 1175 1176 const TypeOopPtr* tp = t->isa_oopptr(); 1177 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1178 // Only for do_null_check, not any of its siblings: 1179 && !assert_null && null_control == NULL) { 1180 // Usually, any field access or invocation on an unloaded oop type 1181 // will simply fail to link, since the statically linked class is 1182 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1183 // the static class is loaded but the sharper oop type is not. 1184 // Rather than checking for this obscure case in lots of places, 1185 // we simply observe that a null check on an unloaded class 1186 // will always be followed by a nonsense operation, so we 1187 // can just issue the uncommon trap here. 1188 // Our access to the unloaded class will only be correct 1189 // after it has been loaded and initialized, which requires 1190 // a trip through the interpreter. 1191 #ifndef PRODUCT 1192 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1193 #endif 1194 uncommon_trap(Deoptimization::Reason_unloaded, 1195 Deoptimization::Action_reinterpret, 1196 tp->klass(), "!loaded"); 1197 return top(); 1198 } 1199 1200 if (assert_null) { 1201 // See if the type is contained in NULL_PTR. 1202 // If so, then the value is already null. 1203 if (t->higher_equal(TypePtr::NULL_PTR)) { 1204 explicit_null_checks_elided++; 1205 return value; // Elided null assert quickly! 1206 } 1207 } else { 1208 // See if mixing in the NULL pointer changes type. 1209 // If so, then the NULL pointer was not allowed in the original 1210 // type. In other words, "value" was not-null. 1211 if (t->meet(TypePtr::NULL_PTR) != t) { 1212 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1213 explicit_null_checks_elided++; 1214 return value; // Elided null check quickly! 1215 } 1216 } 1217 chk = new (C, 3) CmpPNode( value, null() ); 1218 break; 1219 } 1220 1221 default : ShouldNotReachHere(); 1222 } 1223 assert(chk != NULL, "sanity check"); 1224 chk = _gvn.transform(chk); 1225 1226 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1227 BoolNode *btst = new (C, 2) BoolNode( chk, btest); 1228 Node *tst = _gvn.transform( btst ); 1229 1230 //----------- 1231 // if peephole optimizations occurred, a prior test existed. 1232 // If a prior test existed, maybe it dominates as we can avoid this test. 1233 if (tst != btst && type == T_OBJECT) { 1234 // At this point we want to scan up the CFG to see if we can 1235 // find an identical test (and so avoid this test altogether). 1236 Node *cfg = control(); 1237 int depth = 0; 1238 while( depth < 16 ) { // Limit search depth for speed 1239 if( cfg->Opcode() == Op_IfTrue && 1240 cfg->in(0)->in(1) == tst ) { 1241 // Found prior test. Use "cast_not_null" to construct an identical 1242 // CastPP (and hence hash to) as already exists for the prior test. 1243 // Return that casted value. 1244 if (assert_null) { 1245 replace_in_map(value, null()); 1246 return null(); // do not issue the redundant test 1247 } 1248 Node *oldcontrol = control(); 1249 set_control(cfg); 1250 Node *res = cast_not_null(value); 1251 set_control(oldcontrol); 1252 explicit_null_checks_elided++; 1253 return res; 1254 } 1255 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1256 if (cfg == NULL) break; // Quit at region nodes 1257 depth++; 1258 } 1259 } 1260 1261 //----------- 1262 // Branch to failure if null 1263 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1264 Deoptimization::DeoptReason reason; 1265 if (assert_null) 1266 reason = Deoptimization::Reason_null_assert; 1267 else if (type == T_OBJECT) 1268 reason = Deoptimization::Reason_null_check; 1269 else 1270 reason = Deoptimization::Reason_div0_check; 1271 1272 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1273 // ciMethodData::has_trap_at will return a conservative -1 if any 1274 // must-be-null assertion has failed. This could cause performance 1275 // problems for a method after its first do_null_assert failure. 1276 // Consider using 'Reason_class_check' instead? 1277 1278 // To cause an implicit null check, we set the not-null probability 1279 // to the maximum (PROB_MAX). For an explicit check the probability 1280 // is set to a smaller value. 1281 if (null_control != NULL || too_many_traps(reason)) { 1282 // probability is less likely 1283 ok_prob = PROB_LIKELY_MAG(3); 1284 } else if (!assert_null && 1285 (ImplicitNullCheckThreshold > 0) && 1286 method() != NULL && 1287 (method()->method_data()->trap_count(reason) 1288 >= (uint)ImplicitNullCheckThreshold)) { 1289 ok_prob = PROB_LIKELY_MAG(3); 1290 } 1291 1292 if (null_control != NULL) { 1293 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1294 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); 1295 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); 1296 if (null_true == top()) 1297 explicit_null_checks_elided++; 1298 (*null_control) = null_true; 1299 } else { 1300 BuildCutout unless(this, tst, ok_prob); 1301 // Check for optimizer eliding test at parse time 1302 if (stopped()) { 1303 // Failure not possible; do not bother making uncommon trap. 1304 explicit_null_checks_elided++; 1305 } else if (assert_null) { 1306 uncommon_trap(reason, 1307 Deoptimization::Action_make_not_entrant, 1308 NULL, "assert_null"); 1309 } else { 1310 replace_in_map(value, zerocon(type)); 1311 builtin_throw(reason); 1312 } 1313 } 1314 1315 // Must throw exception, fall-thru not possible? 1316 if (stopped()) { 1317 return top(); // No result 1318 } 1319 1320 if (assert_null) { 1321 // Cast obj to null on this path. 1322 replace_in_map(value, zerocon(type)); 1323 return zerocon(type); 1324 } 1325 1326 // Cast obj to not-null on this path, if there is no null_control. 1327 // (If there is a null_control, a non-null value may come back to haunt us.) 1328 if (type == T_OBJECT) { 1329 Node* cast = cast_not_null(value, false); 1330 if (null_control == NULL || (*null_control) == top()) 1331 replace_in_map(value, cast); 1332 value = cast; 1333 } 1334 1335 return value; 1336 } 1337 1338 1339 //------------------------------cast_not_null---------------------------------- 1340 // Cast obj to not-null on this path 1341 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1342 const Type *t = _gvn.type(obj); 1343 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1344 // Object is already not-null? 1345 if( t == t_not_null ) return obj; 1346 1347 Node *cast = new (C, 2) CastPPNode(obj,t_not_null); 1348 cast->init_req(0, control()); 1349 cast = _gvn.transform( cast ); 1350 1351 // Scan for instances of 'obj' in the current JVM mapping. 1352 // These instances are known to be not-null after the test. 1353 if (do_replace_in_map) 1354 replace_in_map(obj, cast); 1355 1356 return cast; // Return casted value 1357 } 1358 1359 1360 //--------------------------replace_in_map------------------------------------- 1361 void GraphKit::replace_in_map(Node* old, Node* neww) { 1362 this->map()->replace_edge(old, neww); 1363 1364 // Note: This operation potentially replaces any edge 1365 // on the map. This includes locals, stack, and monitors 1366 // of the current (innermost) JVM state. 1367 1368 // We can consider replacing in caller maps. 1369 // The idea would be that an inlined function's null checks 1370 // can be shared with the entire inlining tree. 1371 // The expense of doing this is that the PreserveJVMState class 1372 // would have to preserve caller states too, with a deep copy. 1373 } 1374 1375 1376 1377 //============================================================================= 1378 //--------------------------------memory--------------------------------------- 1379 Node* GraphKit::memory(uint alias_idx) { 1380 MergeMemNode* mem = merged_memory(); 1381 Node* p = mem->memory_at(alias_idx); 1382 _gvn.set_type(p, Type::MEMORY); // must be mapped 1383 return p; 1384 } 1385 1386 //-----------------------------reset_memory------------------------------------ 1387 Node* GraphKit::reset_memory() { 1388 Node* mem = map()->memory(); 1389 // do not use this node for any more parsing! 1390 debug_only( map()->set_memory((Node*)NULL) ); 1391 return _gvn.transform( mem ); 1392 } 1393 1394 //------------------------------set_all_memory--------------------------------- 1395 void GraphKit::set_all_memory(Node* newmem) { 1396 Node* mergemem = MergeMemNode::make(C, newmem); 1397 gvn().set_type_bottom(mergemem); 1398 map()->set_memory(mergemem); 1399 } 1400 1401 //------------------------------set_all_memory_call---------------------------- 1402 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1403 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1404 set_all_memory(newmem); 1405 } 1406 1407 //============================================================================= 1408 // 1409 // parser factory methods for MemNodes 1410 // 1411 // These are layered on top of the factory methods in LoadNode and StoreNode, 1412 // and integrate with the parser's memory state and _gvn engine. 1413 // 1414 1415 // factory methods in "int adr_idx" 1416 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1417 int adr_idx, 1418 bool require_atomic_access) { 1419 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1420 const TypePtr* adr_type = NULL; // debug-mode-only argument 1421 debug_only(adr_type = C->get_adr_type(adr_idx)); 1422 Node* mem = memory(adr_idx); 1423 Node* ld; 1424 if (require_atomic_access && bt == T_LONG) { 1425 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); 1426 } else { 1427 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); 1428 } 1429 return _gvn.transform(ld); 1430 } 1431 1432 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1433 int adr_idx, 1434 bool require_atomic_access) { 1435 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1436 const TypePtr* adr_type = NULL; 1437 debug_only(adr_type = C->get_adr_type(adr_idx)); 1438 Node *mem = memory(adr_idx); 1439 Node* st; 1440 if (require_atomic_access && bt == T_LONG) { 1441 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); 1442 } else { 1443 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); 1444 } 1445 st = _gvn.transform(st); 1446 set_memory(st, adr_idx); 1447 // Back-to-back stores can only remove intermediate store with DU info 1448 // so push on worklist for optimizer. 1449 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1450 record_for_igvn(st); 1451 1452 return st; 1453 } 1454 1455 1456 void GraphKit::pre_barrier(bool do_load, 1457 Node* ctl, 1458 Node* obj, 1459 Node* adr, 1460 uint adr_idx, 1461 Node* val, 1462 const TypeOopPtr* val_type, 1463 Node* pre_val, 1464 BasicType bt) { 1465 1466 BarrierSet* bs = Universe::heap()->barrier_set(); 1467 set_control(ctl); 1468 switch (bs->kind()) { 1469 case BarrierSet::G1SATBCT: 1470 case BarrierSet::G1SATBCTLogging: 1471 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1472 break; 1473 1474 case BarrierSet::CardTableModRef: 1475 case BarrierSet::CardTableExtension: 1476 case BarrierSet::ModRef: 1477 break; 1478 1479 case BarrierSet::Other: 1480 default : 1481 ShouldNotReachHere(); 1482 1483 } 1484 } 1485 1486 void GraphKit::post_barrier(Node* ctl, 1487 Node* store, 1488 Node* obj, 1489 Node* adr, 1490 uint adr_idx, 1491 Node* val, 1492 BasicType bt, 1493 bool use_precise) { 1494 BarrierSet* bs = Universe::heap()->barrier_set(); 1495 set_control(ctl); 1496 switch (bs->kind()) { 1497 case BarrierSet::G1SATBCT: 1498 case BarrierSet::G1SATBCTLogging: 1499 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1500 break; 1501 1502 case BarrierSet::CardTableModRef: 1503 case BarrierSet::CardTableExtension: 1504 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1505 break; 1506 1507 case BarrierSet::ModRef: 1508 break; 1509 1510 case BarrierSet::Other: 1511 default : 1512 ShouldNotReachHere(); 1513 1514 } 1515 } 1516 1517 Node* GraphKit::store_oop(Node* ctl, 1518 Node* obj, 1519 Node* adr, 1520 const TypePtr* adr_type, 1521 Node* val, 1522 const TypeOopPtr* val_type, 1523 BasicType bt, 1524 bool use_precise) { 1525 1526 set_control(ctl); 1527 if (stopped()) return top(); // Dead path ? 1528 1529 assert(bt == T_OBJECT, "sanity"); 1530 assert(val != NULL, "not dead path"); 1531 uint adr_idx = C->get_alias_index(adr_type); 1532 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1533 1534 pre_barrier(true /* do_load */, 1535 control(), obj, adr, adr_idx, val, val_type, 1536 NULL /* pre_val */, 1537 bt); 1538 1539 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1540 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1541 return store; 1542 } 1543 1544 // Could be an array or object we don't know at compile time (unsafe ref.) 1545 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1546 Node* obj, // containing obj 1547 Node* adr, // actual adress to store val at 1548 const TypePtr* adr_type, 1549 Node* val, 1550 BasicType bt) { 1551 Compile::AliasType* at = C->alias_type(adr_type); 1552 const TypeOopPtr* val_type = NULL; 1553 if (adr_type->isa_instptr()) { 1554 if (at->field() != NULL) { 1555 // known field. This code is a copy of the do_put_xxx logic. 1556 ciField* field = at->field(); 1557 if (!field->type()->is_loaded()) { 1558 val_type = TypeInstPtr::BOTTOM; 1559 } else { 1560 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1561 } 1562 } 1563 } else if (adr_type->isa_aryptr()) { 1564 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1565 } 1566 if (val_type == NULL) { 1567 val_type = TypeInstPtr::BOTTOM; 1568 } 1569 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); 1570 } 1571 1572 1573 //-------------------------array_element_address------------------------- 1574 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1575 const TypeInt* sizetype) { 1576 uint shift = exact_log2(type2aelembytes(elembt)); 1577 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1578 1579 // short-circuit a common case (saves lots of confusing waste motion) 1580 jint idx_con = find_int_con(idx, -1); 1581 if (idx_con >= 0) { 1582 intptr_t offset = header + ((intptr_t)idx_con << shift); 1583 return basic_plus_adr(ary, offset); 1584 } 1585 1586 // must be correct type for alignment purposes 1587 Node* base = basic_plus_adr(ary, header); 1588 #ifdef _LP64 1589 // The scaled index operand to AddP must be a clean 64-bit value. 1590 // Java allows a 32-bit int to be incremented to a negative 1591 // value, which appears in a 64-bit register as a large 1592 // positive number. Using that large positive number as an 1593 // operand in pointer arithmetic has bad consequences. 1594 // On the other hand, 32-bit overflow is rare, and the possibility 1595 // can often be excluded, if we annotate the ConvI2L node with 1596 // a type assertion that its value is known to be a small positive 1597 // number. (The prior range check has ensured this.) 1598 // This assertion is used by ConvI2LNode::Ideal. 1599 int index_max = max_jint - 1; // array size is max_jint, index is one less 1600 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1601 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1602 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); 1603 #endif 1604 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); 1605 return basic_plus_adr(ary, base, scale); 1606 } 1607 1608 //-------------------------load_array_element------------------------- 1609 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1610 const Type* elemtype = arytype->elem(); 1611 BasicType elembt = elemtype->array_element_basic_type(); 1612 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1613 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1614 return ld; 1615 } 1616 1617 //-------------------------set_arguments_for_java_call------------------------- 1618 // Arguments (pre-popped from the stack) are taken from the JVMS. 1619 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1620 // Add the call arguments: 1621 uint nargs = call->method()->arg_size(); 1622 for (uint i = 0; i < nargs; i++) { 1623 Node* arg = argument(i); 1624 call->init_req(i + TypeFunc::Parms, arg); 1625 } 1626 } 1627 1628 //---------------------------set_edges_for_java_call--------------------------- 1629 // Connect a newly created call into the current JVMS. 1630 // A return value node (if any) is returned from set_edges_for_java_call. 1631 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1632 1633 // Add the predefined inputs: 1634 call->init_req( TypeFunc::Control, control() ); 1635 call->init_req( TypeFunc::I_O , i_o() ); 1636 call->init_req( TypeFunc::Memory , reset_memory() ); 1637 call->init_req( TypeFunc::FramePtr, frameptr() ); 1638 call->init_req( TypeFunc::ReturnAdr, top() ); 1639 1640 add_safepoint_edges(call, must_throw); 1641 1642 Node* xcall = _gvn.transform(call); 1643 1644 if (xcall == top()) { 1645 set_control(top()); 1646 return; 1647 } 1648 assert(xcall == call, "call identity is stable"); 1649 1650 // Re-use the current map to produce the result. 1651 1652 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); 1653 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1654 set_all_memory_call(xcall, separate_io_proj); 1655 1656 //return xcall; // no need, caller already has it 1657 } 1658 1659 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1660 if (stopped()) return top(); // maybe the call folded up? 1661 1662 // Capture the return value, if any. 1663 Node* ret; 1664 if (call->method() == NULL || 1665 call->method()->return_type()->basic_type() == T_VOID) 1666 ret = top(); 1667 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 1668 1669 // Note: Since any out-of-line call can produce an exception, 1670 // we always insert an I_O projection from the call into the result. 1671 1672 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1673 1674 if (separate_io_proj) { 1675 // The caller requested separate projections be used by the fall 1676 // through and exceptional paths, so replace the projections for 1677 // the fall through path. 1678 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) )); 1679 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) )); 1680 } 1681 return ret; 1682 } 1683 1684 //--------------------set_predefined_input_for_runtime_call-------------------- 1685 // Reading and setting the memory state is way conservative here. 1686 // The real problem is that I am not doing real Type analysis on memory, 1687 // so I cannot distinguish card mark stores from other stores. Across a GC 1688 // point the Store Barrier and the card mark memory has to agree. I cannot 1689 // have a card mark store and its barrier split across the GC point from 1690 // either above or below. Here I get that to happen by reading ALL of memory. 1691 // A better answer would be to separate out card marks from other memory. 1692 // For now, return the input memory state, so that it can be reused 1693 // after the call, if this call has restricted memory effects. 1694 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1695 // Set fixed predefined input arguments 1696 Node* memory = reset_memory(); 1697 call->init_req( TypeFunc::Control, control() ); 1698 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1699 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1700 call->init_req( TypeFunc::FramePtr, frameptr() ); 1701 call->init_req( TypeFunc::ReturnAdr, top() ); 1702 return memory; 1703 } 1704 1705 //-------------------set_predefined_output_for_runtime_call-------------------- 1706 // Set control and memory (not i_o) from the call. 1707 // If keep_mem is not NULL, use it for the output state, 1708 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1709 // If hook_mem is NULL, this call produces no memory effects at all. 1710 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1711 // then only that memory slice is taken from the call. 1712 // In the last case, we must put an appropriate memory barrier before 1713 // the call, so as to create the correct anti-dependencies on loads 1714 // preceding the call. 1715 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1716 Node* keep_mem, 1717 const TypePtr* hook_mem) { 1718 // no i/o 1719 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); 1720 if (keep_mem) { 1721 // First clone the existing memory state 1722 set_all_memory(keep_mem); 1723 if (hook_mem != NULL) { 1724 // Make memory for the call 1725 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1726 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1727 // We also use hook_mem to extract specific effects from arraycopy stubs. 1728 set_memory(mem, hook_mem); 1729 } 1730 // ...else the call has NO memory effects. 1731 1732 // Make sure the call advertises its memory effects precisely. 1733 // This lets us build accurate anti-dependences in gcm.cpp. 1734 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1735 "call node must be constructed correctly"); 1736 } else { 1737 assert(hook_mem == NULL, ""); 1738 // This is not a "slow path" call; all memory comes from the call. 1739 set_all_memory_call(call); 1740 } 1741 } 1742 1743 1744 // Replace the call with the current state of the kit. 1745 void GraphKit::replace_call(CallNode* call, Node* result) { 1746 JVMState* ejvms = NULL; 1747 if (has_exceptions()) { 1748 ejvms = transfer_exceptions_into_jvms(); 1749 } 1750 1751 SafePointNode* final_state = stop(); 1752 1753 // Find all the needed outputs of this call 1754 CallProjections callprojs; 1755 call->extract_projections(&callprojs, true); 1756 1757 // Replace all the old call edges with the edges from the inlining result 1758 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control)); 1759 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory)); 1760 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O)); 1761 Node* final_mem = final_state->in(TypeFunc::Memory); 1762 1763 // Replace the result with the new result if it exists and is used 1764 if (callprojs.resproj != NULL && result != NULL) { 1765 C->gvn_replace_by(callprojs.resproj, result); 1766 } 1767 1768 if (ejvms == NULL) { 1769 // No exception edges to simply kill off those paths 1770 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1771 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1772 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1773 1774 // Replace the old exception object with top 1775 if (callprojs.exobj != NULL) { 1776 C->gvn_replace_by(callprojs.exobj, C->top()); 1777 } 1778 } else { 1779 GraphKit ekit(ejvms); 1780 1781 // Load my combined exception state into the kit, with all phis transformed: 1782 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1783 1784 Node* ex_oop = ekit.use_exception_state(ex_map); 1785 1786 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1787 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1788 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1789 1790 // Replace the old exception object with the newly created one 1791 if (callprojs.exobj != NULL) { 1792 C->gvn_replace_by(callprojs.exobj, ex_oop); 1793 } 1794 } 1795 1796 // Disconnect the call from the graph 1797 call->disconnect_inputs(NULL); 1798 C->gvn_replace_by(call, C->top()); 1799 1800 // Clean up any MergeMems that feed other MergeMems since the 1801 // optimizer doesn't like that. 1802 if (final_mem->is_MergeMem()) { 1803 Node_List wl; 1804 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1805 Node* m = i.get(); 1806 if (m->is_MergeMem() && !wl.contains(m)) { 1807 wl.push(m); 1808 } 1809 } 1810 while (wl.size() > 0) { 1811 _gvn.transform(wl.pop()); 1812 } 1813 } 1814 } 1815 1816 1817 //------------------------------increment_counter------------------------------ 1818 // for statistics: increment a VM counter by 1 1819 1820 void GraphKit::increment_counter(address counter_addr) { 1821 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1822 increment_counter(adr1); 1823 } 1824 1825 void GraphKit::increment_counter(Node* counter_addr) { 1826 int adr_type = Compile::AliasIdxRaw; 1827 Node* ctrl = control(); 1828 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type); 1829 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 1830 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type ); 1831 } 1832 1833 1834 //------------------------------uncommon_trap---------------------------------- 1835 // Bail out to the interpreter in mid-method. Implemented by calling the 1836 // uncommon_trap blob. This helper function inserts a runtime call with the 1837 // right debug info. 1838 void GraphKit::uncommon_trap(int trap_request, 1839 ciKlass* klass, const char* comment, 1840 bool must_throw, 1841 bool keep_exact_action) { 1842 if (failing()) stop(); 1843 if (stopped()) return; // trap reachable? 1844 1845 // Note: If ProfileTraps is true, and if a deopt. actually 1846 // occurs here, the runtime will make sure an MDO exists. There is 1847 // no need to call method()->ensure_method_data() at this point. 1848 1849 #ifdef ASSERT 1850 if (!must_throw) { 1851 // Make sure the stack has at least enough depth to execute 1852 // the current bytecode. 1853 int inputs, ignore; 1854 if (compute_stack_effects(inputs, ignore)) { 1855 assert(sp() >= inputs, "must have enough JVMS stack to execute"); 1856 // It is a frequent error in library_call.cpp to issue an 1857 // uncommon trap with the _sp value already popped. 1858 } 1859 } 1860 #endif 1861 1862 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1863 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1864 1865 switch (action) { 1866 case Deoptimization::Action_maybe_recompile: 1867 case Deoptimization::Action_reinterpret: 1868 // Temporary fix for 6529811 to allow virtual calls to be sure they 1869 // get the chance to go from mono->bi->mega 1870 if (!keep_exact_action && 1871 Deoptimization::trap_request_index(trap_request) < 0 && 1872 too_many_recompiles(reason)) { 1873 // This BCI is causing too many recompilations. 1874 action = Deoptimization::Action_none; 1875 trap_request = Deoptimization::make_trap_request(reason, action); 1876 } else { 1877 C->set_trap_can_recompile(true); 1878 } 1879 break; 1880 case Deoptimization::Action_make_not_entrant: 1881 C->set_trap_can_recompile(true); 1882 break; 1883 #ifdef ASSERT 1884 case Deoptimization::Action_none: 1885 case Deoptimization::Action_make_not_compilable: 1886 break; 1887 default: 1888 assert(false, "bad action"); 1889 #endif 1890 } 1891 1892 if (TraceOptoParse) { 1893 char buf[100]; 1894 tty->print_cr("Uncommon trap %s at bci:%d", 1895 Deoptimization::format_trap_request(buf, sizeof(buf), 1896 trap_request), bci()); 1897 } 1898 1899 CompileLog* log = C->log(); 1900 if (log != NULL) { 1901 int kid = (klass == NULL)? -1: log->identify(klass); 1902 log->begin_elem("uncommon_trap bci='%d'", bci()); 1903 char buf[100]; 1904 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1905 trap_request)); 1906 if (kid >= 0) log->print(" klass='%d'", kid); 1907 if (comment != NULL) log->print(" comment='%s'", comment); 1908 log->end_elem(); 1909 } 1910 1911 // Make sure any guarding test views this path as very unlikely 1912 Node *i0 = control()->in(0); 1913 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1914 IfNode *iff = i0->as_If(); 1915 float f = iff->_prob; // Get prob 1916 if (control()->Opcode() == Op_IfTrue) { 1917 if (f > PROB_UNLIKELY_MAG(4)) 1918 iff->_prob = PROB_MIN; 1919 } else { 1920 if (f < PROB_LIKELY_MAG(4)) 1921 iff->_prob = PROB_MAX; 1922 } 1923 } 1924 1925 // Clear out dead values from the debug info. 1926 kill_dead_locals(); 1927 1928 // Now insert the uncommon trap subroutine call 1929 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 1930 const TypePtr* no_memory_effects = NULL; 1931 // Pass the index of the class to be loaded 1932 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1933 (must_throw ? RC_MUST_THROW : 0), 1934 OptoRuntime::uncommon_trap_Type(), 1935 call_addr, "uncommon_trap", no_memory_effects, 1936 intcon(trap_request)); 1937 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 1938 "must extract request correctly from the graph"); 1939 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 1940 1941 call->set_req(TypeFunc::ReturnAdr, returnadr()); 1942 // The debug info is the only real input to this call. 1943 1944 // Halt-and-catch fire here. The above call should never return! 1945 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); 1946 _gvn.set_type_bottom(halt); 1947 root()->add_req(halt); 1948 1949 stop_and_kill_map(); 1950 } 1951 1952 1953 //--------------------------just_allocated_object------------------------------ 1954 // Report the object that was just allocated. 1955 // It must be the case that there are no intervening safepoints. 1956 // We use this to determine if an object is so "fresh" that 1957 // it does not require card marks. 1958 Node* GraphKit::just_allocated_object(Node* current_control) { 1959 if (C->recent_alloc_ctl() == current_control) 1960 return C->recent_alloc_obj(); 1961 return NULL; 1962 } 1963 1964 1965 void GraphKit::round_double_arguments(ciMethod* dest_method) { 1966 // (Note: TypeFunc::make has a cache that makes this fast.) 1967 const TypeFunc* tf = TypeFunc::make(dest_method); 1968 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 1969 for (int j = 0; j < nargs; j++) { 1970 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 1971 if( targ->basic_type() == T_DOUBLE ) { 1972 // If any parameters are doubles, they must be rounded before 1973 // the call, dstore_rounding does gvn.transform 1974 Node *arg = argument(j); 1975 arg = dstore_rounding(arg); 1976 set_argument(j, arg); 1977 } 1978 } 1979 } 1980 1981 void GraphKit::round_double_result(ciMethod* dest_method) { 1982 // A non-strict method may return a double value which has an extended 1983 // exponent, but this must not be visible in a caller which is 'strict' 1984 // If a strict caller invokes a non-strict callee, round a double result 1985 1986 BasicType result_type = dest_method->return_type()->basic_type(); 1987 assert( method() != NULL, "must have caller context"); 1988 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 1989 // Destination method's return value is on top of stack 1990 // dstore_rounding() does gvn.transform 1991 Node *result = pop_pair(); 1992 result = dstore_rounding(result); 1993 push_pair(result); 1994 } 1995 } 1996 1997 // rounding for strict float precision conformance 1998 Node* GraphKit::precision_rounding(Node* n) { 1999 return UseStrictFP && _method->flags().is_strict() 2000 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2001 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) 2002 : n; 2003 } 2004 2005 // rounding for strict double precision conformance 2006 Node* GraphKit::dprecision_rounding(Node *n) { 2007 return UseStrictFP && _method->flags().is_strict() 2008 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2009 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 2010 : n; 2011 } 2012 2013 // rounding for non-strict double stores 2014 Node* GraphKit::dstore_rounding(Node* n) { 2015 return Matcher::strict_fp_requires_explicit_rounding 2016 && UseSSE <= 1 2017 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 2018 : n; 2019 } 2020 2021 //============================================================================= 2022 // Generate a fast path/slow path idiom. Graph looks like: 2023 // [foo] indicates that 'foo' is a parameter 2024 // 2025 // [in] NULL 2026 // \ / 2027 // CmpP 2028 // Bool ne 2029 // If 2030 // / \ 2031 // True False-<2> 2032 // / | 2033 // / cast_not_null 2034 // Load | | ^ 2035 // [fast_test] | | 2036 // gvn to opt_test | | 2037 // / \ | <1> 2038 // True False | 2039 // | \\ | 2040 // [slow_call] \[fast_result] 2041 // Ctl Val \ \ 2042 // | \ \ 2043 // Catch <1> \ \ 2044 // / \ ^ \ \ 2045 // Ex No_Ex | \ \ 2046 // | \ \ | \ <2> \ 2047 // ... \ [slow_res] | | \ [null_result] 2048 // \ \--+--+--- | | 2049 // \ | / \ | / 2050 // --------Region Phi 2051 // 2052 //============================================================================= 2053 // Code is structured as a series of driver functions all called 'do_XXX' that 2054 // call a set of helper functions. Helper functions first, then drivers. 2055 2056 //------------------------------null_check_oop--------------------------------- 2057 // Null check oop. Set null-path control into Region in slot 3. 2058 // Make a cast-not-nullness use the other not-null control. Return cast. 2059 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2060 bool never_see_null) { 2061 // Initial NULL check taken path 2062 (*null_control) = top(); 2063 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2064 2065 // Generate uncommon_trap: 2066 if (never_see_null && (*null_control) != top()) { 2067 // If we see an unexpected null at a check-cast we record it and force a 2068 // recompile; the offending check-cast will be compiled to handle NULLs. 2069 // If we see more than one offending BCI, then all checkcasts in the 2070 // method will be compiled to handle NULLs. 2071 PreserveJVMState pjvms(this); 2072 set_control(*null_control); 2073 replace_in_map(value, null()); 2074 uncommon_trap(Deoptimization::Reason_null_check, 2075 Deoptimization::Action_make_not_entrant); 2076 (*null_control) = top(); // NULL path is dead 2077 } 2078 2079 // Cast away null-ness on the result 2080 return cast; 2081 } 2082 2083 //------------------------------opt_iff---------------------------------------- 2084 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2085 // Return slow-path control. 2086 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2087 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2088 2089 // Fast path taken; set region slot 2 2090 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); 2091 region->init_req(2,fast_taken); // Capture fast-control 2092 2093 // Fast path not-taken, i.e. slow path 2094 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); 2095 return slow_taken; 2096 } 2097 2098 //-----------------------------make_runtime_call------------------------------- 2099 Node* GraphKit::make_runtime_call(int flags, 2100 const TypeFunc* call_type, address call_addr, 2101 const char* call_name, 2102 const TypePtr* adr_type, 2103 // The following parms are all optional. 2104 // The first NULL ends the list. 2105 Node* parm0, Node* parm1, 2106 Node* parm2, Node* parm3, 2107 Node* parm4, Node* parm5, 2108 Node* parm6, Node* parm7) { 2109 // Slow-path call 2110 int size = call_type->domain()->cnt(); 2111 bool is_leaf = !(flags & RC_NO_LEAF); 2112 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2113 if (call_name == NULL) { 2114 assert(!is_leaf, "must supply name for leaf"); 2115 call_name = OptoRuntime::stub_name(call_addr); 2116 } 2117 CallNode* call; 2118 if (!is_leaf) { 2119 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, 2120 bci(), adr_type); 2121 } else if (flags & RC_NO_FP) { 2122 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2123 } else { 2124 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); 2125 } 2126 2127 // The following is similar to set_edges_for_java_call, 2128 // except that the memory effects of the call are restricted to AliasIdxRaw. 2129 2130 // Slow path call has no side-effects, uses few values 2131 bool wide_in = !(flags & RC_NARROW_MEM); 2132 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2133 2134 Node* prev_mem = NULL; 2135 if (wide_in) { 2136 prev_mem = set_predefined_input_for_runtime_call(call); 2137 } else { 2138 assert(!wide_out, "narrow in => narrow out"); 2139 Node* narrow_mem = memory(adr_type); 2140 prev_mem = reset_memory(); 2141 map()->set_memory(narrow_mem); 2142 set_predefined_input_for_runtime_call(call); 2143 } 2144 2145 // Hook each parm in order. Stop looking at the first NULL. 2146 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2147 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2148 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2149 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2150 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2151 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2152 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2153 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2154 /* close each nested if ===> */ } } } } } } } } 2155 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2156 2157 if (!is_leaf) { 2158 // Non-leaves can block and take safepoints: 2159 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2160 } 2161 // Non-leaves can throw exceptions: 2162 if (has_io) { 2163 call->set_req(TypeFunc::I_O, i_o()); 2164 } 2165 2166 if (flags & RC_UNCOMMON) { 2167 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2168 // (An "if" probability corresponds roughly to an unconditional count. 2169 // Sort of.) 2170 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2171 } 2172 2173 Node* c = _gvn.transform(call); 2174 assert(c == call, "cannot disappear"); 2175 2176 if (wide_out) { 2177 // Slow path call has full side-effects. 2178 set_predefined_output_for_runtime_call(call); 2179 } else { 2180 // Slow path call has few side-effects, and/or sets few values. 2181 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2182 } 2183 2184 if (has_io) { 2185 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); 2186 } 2187 return call; 2188 2189 } 2190 2191 //------------------------------merge_memory----------------------------------- 2192 // Merge memory from one path into the current memory state. 2193 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2194 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2195 Node* old_slice = mms.force_memory(); 2196 Node* new_slice = mms.memory2(); 2197 if (old_slice != new_slice) { 2198 PhiNode* phi; 2199 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2200 phi = new_slice->as_Phi(); 2201 #ifdef ASSERT 2202 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2203 old_slice = old_slice->in(new_path); 2204 // Caller is responsible for ensuring that any pre-existing 2205 // phis are already aware of old memory. 2206 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2207 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2208 #endif 2209 mms.set_memory(phi); 2210 } else { 2211 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2212 _gvn.set_type(phi, Type::MEMORY); 2213 phi->set_req(new_path, new_slice); 2214 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2215 } 2216 } 2217 } 2218 } 2219 2220 //------------------------------make_slow_call_ex------------------------------ 2221 // Make the exception handler hookups for the slow call 2222 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2223 if (stopped()) return; 2224 2225 // Make a catch node with just two handlers: fall-through and catch-all 2226 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2227 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); 2228 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2229 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2230 2231 { PreserveJVMState pjvms(this); 2232 set_control(excp); 2233 set_i_o(i_o); 2234 2235 if (excp != top()) { 2236 // Create an exception state also. 2237 // Use an exact type if the caller has specified a specific exception. 2238 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2239 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); 2240 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2241 } 2242 } 2243 2244 // Get the no-exception control from the CatchNode. 2245 set_control(norm); 2246 } 2247 2248 2249 //-------------------------------gen_subtype_check----------------------------- 2250 // Generate a subtyping check. Takes as input the subtype and supertype. 2251 // Returns 2 values: sets the default control() to the true path and returns 2252 // the false path. Only reads invariant memory; sets no (visible) memory. 2253 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2254 // but that's not exposed to the optimizer. This call also doesn't take in an 2255 // Object; if you wish to check an Object you need to load the Object's class 2256 // prior to coming here. 2257 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2258 // Fast check for identical types, perhaps identical constants. 2259 // The types can even be identical non-constants, in cases 2260 // involving Array.newInstance, Object.clone, etc. 2261 if (subklass == superklass) 2262 return top(); // false path is dead; no test needed. 2263 2264 if (_gvn.type(superklass)->singleton()) { 2265 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2266 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2267 2268 // In the common case of an exact superklass, try to fold up the 2269 // test before generating code. You may ask, why not just generate 2270 // the code and then let it fold up? The answer is that the generated 2271 // code will necessarily include null checks, which do not always 2272 // completely fold away. If they are also needless, then they turn 2273 // into a performance loss. Example: 2274 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2275 // Here, the type of 'fa' is often exact, so the store check 2276 // of fa[1]=x will fold up, without testing the nullness of x. 2277 switch (static_subtype_check(superk, subk)) { 2278 case SSC_always_false: 2279 { 2280 Node* always_fail = control(); 2281 set_control(top()); 2282 return always_fail; 2283 } 2284 case SSC_always_true: 2285 return top(); 2286 case SSC_easy_test: 2287 { 2288 // Just do a direct pointer compare and be done. 2289 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); 2290 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2291 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2292 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); 2293 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2294 } 2295 case SSC_full_test: 2296 break; 2297 default: 2298 ShouldNotReachHere(); 2299 } 2300 } 2301 2302 // %%% Possible further optimization: Even if the superklass is not exact, 2303 // if the subklass is the unique subtype of the superklass, the check 2304 // will always succeed. We could leave a dependency behind to ensure this. 2305 2306 // First load the super-klass's check-offset 2307 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2308 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2309 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2310 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2311 2312 // Load from the sub-klass's super-class display list, or a 1-word cache of 2313 // the secondary superclass list, or a failing value with a sentinel offset 2314 // if the super-klass is an interface or exceptionally deep in the Java 2315 // hierarchy and we have to scan the secondary superclass list the hard way. 2316 // Worst-case type is a little odd: NULL is allowed as a result (usually 2317 // klass loads can never produce a NULL). 2318 Node *chk_off_X = ConvI2X(chk_off); 2319 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); 2320 // For some types like interfaces the following loadKlass is from a 1-word 2321 // cache which is mutable so can't use immutable memory. Other 2322 // types load from the super-class display table which is immutable. 2323 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2324 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2325 2326 // Compile speed common case: ARE a subtype and we canNOT fail 2327 if( superklass == nkls ) 2328 return top(); // false path is dead; no test needed. 2329 2330 // See if we get an immediate positive hit. Happens roughly 83% of the 2331 // time. Test to see if the value loaded just previously from the subklass 2332 // is exactly the superklass. 2333 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); 2334 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); 2335 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2336 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); 2337 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); 2338 2339 // Compile speed common case: Check for being deterministic right now. If 2340 // chk_off is a constant and not equal to cacheoff then we are NOT a 2341 // subklass. In this case we need exactly the 1 test above and we can 2342 // return those results immediately. 2343 if (!might_be_cache) { 2344 Node* not_subtype_ctrl = control(); 2345 set_control(iftrue1); // We need exactly the 1 test above 2346 return not_subtype_ctrl; 2347 } 2348 2349 // Gather the various success & failures here 2350 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); 2351 record_for_igvn(r_ok_subtype); 2352 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); 2353 record_for_igvn(r_not_subtype); 2354 2355 r_ok_subtype->init_req(1, iftrue1); 2356 2357 // Check for immediate negative hit. Happens roughly 11% of the time (which 2358 // is roughly 63% of the remaining cases). Test to see if the loaded 2359 // check-offset points into the subklass display list or the 1-element 2360 // cache. If it points to the display (and NOT the cache) and the display 2361 // missed then it's not a subtype. 2362 Node *cacheoff = _gvn.intcon(cacheoff_con); 2363 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); 2364 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); 2365 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2366 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); 2367 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); 2368 2369 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2370 // No performance impact (too rare) but allows sharing of secondary arrays 2371 // which has some footprint reduction. 2372 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); 2373 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); 2374 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2375 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); 2376 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); 2377 2378 // -- Roads not taken here: -- 2379 // We could also have chosen to perform the self-check at the beginning 2380 // of this code sequence, as the assembler does. This would not pay off 2381 // the same way, since the optimizer, unlike the assembler, can perform 2382 // static type analysis to fold away many successful self-checks. 2383 // Non-foldable self checks work better here in second position, because 2384 // the initial primary superclass check subsumes a self-check for most 2385 // types. An exception would be a secondary type like array-of-interface, 2386 // which does not appear in its own primary supertype display. 2387 // Finally, we could have chosen to move the self-check into the 2388 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2389 // dependent manner. But it is worthwhile to have the check here, 2390 // where it can be perhaps be optimized. The cost in code space is 2391 // small (register compare, branch). 2392 2393 // Now do a linear scan of the secondary super-klass array. Again, no real 2394 // performance impact (too rare) but it's gotta be done. 2395 // Since the code is rarely used, there is no penalty for moving it 2396 // out of line, and it can only improve I-cache density. 2397 // The decision to inline or out-of-line this final check is platform 2398 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2399 Node* psc = _gvn.transform( 2400 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2401 2402 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); 2403 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); 2404 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2405 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); 2406 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); 2407 2408 // Return false path; set default control to true path. 2409 set_control( _gvn.transform(r_ok_subtype) ); 2410 return _gvn.transform(r_not_subtype); 2411 } 2412 2413 //----------------------------static_subtype_check----------------------------- 2414 // Shortcut important common cases when superklass is exact: 2415 // (0) superklass is java.lang.Object (can occur in reflective code) 2416 // (1) subklass is already limited to a subtype of superklass => always ok 2417 // (2) subklass does not overlap with superklass => always fail 2418 // (3) superklass has NO subtypes and we can check with a simple compare. 2419 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2420 if (StressReflectiveCode) { 2421 return SSC_full_test; // Let caller generate the general case. 2422 } 2423 2424 if (superk == env()->Object_klass()) { 2425 return SSC_always_true; // (0) this test cannot fail 2426 } 2427 2428 ciType* superelem = superk; 2429 if (superelem->is_array_klass()) 2430 superelem = superelem->as_array_klass()->base_element_type(); 2431 2432 if (!subk->is_interface()) { // cannot trust static interface types yet 2433 if (subk->is_subtype_of(superk)) { 2434 return SSC_always_true; // (1) false path dead; no dynamic test needed 2435 } 2436 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2437 !superk->is_subtype_of(subk)) { 2438 return SSC_always_false; 2439 } 2440 } 2441 2442 // If casting to an instance klass, it must have no subtypes 2443 if (superk->is_interface()) { 2444 // Cannot trust interfaces yet. 2445 // %%% S.B. superk->nof_implementors() == 1 2446 } else if (superelem->is_instance_klass()) { 2447 ciInstanceKlass* ik = superelem->as_instance_klass(); 2448 if (!ik->has_subklass() && !ik->is_interface()) { 2449 if (!ik->is_final()) { 2450 // Add a dependency if there is a chance of a later subclass. 2451 C->dependencies()->assert_leaf_type(ik); 2452 } 2453 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2454 } 2455 } else { 2456 // A primitive array type has no subtypes. 2457 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2458 } 2459 2460 return SSC_full_test; 2461 } 2462 2463 // Profile-driven exact type check: 2464 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2465 float prob, 2466 Node* *casted_receiver) { 2467 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2468 Node* recv_klass = load_object_klass(receiver); 2469 Node* want_klass = makecon(tklass); 2470 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); 2471 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2472 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2473 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); 2474 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2475 2476 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2477 assert(recv_xtype->klass_is_exact(), ""); 2478 2479 // Subsume downstream occurrences of receiver with a cast to 2480 // recv_xtype, since now we know what the type will be. 2481 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); 2482 (*casted_receiver) = _gvn.transform(cast); 2483 // (User must make the replace_in_map call.) 2484 2485 return fail; 2486 } 2487 2488 2489 //------------------------------seems_never_null------------------------------- 2490 // Use null_seen information if it is available from the profile. 2491 // If we see an unexpected null at a type check we record it and force a 2492 // recompile; the offending check will be recompiled to handle NULLs. 2493 // If we see several offending BCIs, then all checks in the 2494 // method will be recompiled. 2495 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2496 if (UncommonNullCast // Cutout for this technique 2497 && obj != null() // And not the -Xcomp stupid case? 2498 && !too_many_traps(Deoptimization::Reason_null_check) 2499 ) { 2500 if (data == NULL) 2501 // Edge case: no mature data. Be optimistic here. 2502 return true; 2503 // If the profile has not seen a null, assume it won't happen. 2504 assert(java_bc() == Bytecodes::_checkcast || 2505 java_bc() == Bytecodes::_instanceof || 2506 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2507 return !data->as_BitData()->null_seen(); 2508 } 2509 return false; 2510 } 2511 2512 //------------------------maybe_cast_profiled_receiver------------------------- 2513 // If the profile has seen exactly one type, narrow to exactly that type. 2514 // Subsequent type checks will always fold up. 2515 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2516 ciProfileData* data, 2517 ciKlass* require_klass) { 2518 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2519 if (data == NULL) return NULL; 2520 2521 // Make sure we haven't already deoptimized from this tactic. 2522 if (too_many_traps(Deoptimization::Reason_class_check)) 2523 return NULL; 2524 2525 // (No, this isn't a call, but it's enough like a virtual call 2526 // to use the same ciMethod accessor to get the profile info...) 2527 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2528 if (profile.count() >= 0 && // no cast failures here 2529 profile.has_receiver(0) && 2530 profile.morphism() == 1) { 2531 ciKlass* exact_kls = profile.receiver(0); 2532 if (require_klass == NULL || 2533 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2534 // If we narrow the type to match what the type profile sees, 2535 // we can then remove the rest of the cast. 2536 // This is a win, even if the exact_kls is very specific, 2537 // because downstream operations, such as method calls, 2538 // will often benefit from the sharper type. 2539 Node* exact_obj = not_null_obj; // will get updated in place... 2540 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2541 &exact_obj); 2542 { PreserveJVMState pjvms(this); 2543 set_control(slow_ctl); 2544 uncommon_trap(Deoptimization::Reason_class_check, 2545 Deoptimization::Action_maybe_recompile); 2546 } 2547 replace_in_map(not_null_obj, exact_obj); 2548 return exact_obj; 2549 } 2550 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2551 } 2552 2553 return NULL; 2554 } 2555 2556 2557 //-------------------------------gen_instanceof-------------------------------- 2558 // Generate an instance-of idiom. Used by both the instance-of bytecode 2559 // and the reflective instance-of call. 2560 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) { 2561 kill_dead_locals(); // Benefit all the uncommon traps 2562 assert( !stopped(), "dead parse path should be checked in callers" ); 2563 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2564 "must check for not-null not-dead klass in callers"); 2565 2566 // Make the merge point 2567 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2568 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2569 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2570 C->set_has_split_ifs(true); // Has chance for split-if optimization 2571 2572 ciProfileData* data = NULL; 2573 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2574 data = method()->method_data()->bci_to_data(bci()); 2575 } 2576 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2577 && seems_never_null(obj, data)); 2578 2579 // Null check; get casted pointer; set region slot 3 2580 Node* null_ctl = top(); 2581 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2582 2583 // If not_null_obj is dead, only null-path is taken 2584 if (stopped()) { // Doing instance-of on a NULL? 2585 set_control(null_ctl); 2586 return intcon(0); 2587 } 2588 region->init_req(_null_path, null_ctl); 2589 phi ->init_req(_null_path, intcon(0)); // Set null path value 2590 if (null_ctl == top()) { 2591 // Do this eagerly, so that pattern matches like is_diamond_phi 2592 // will work even during parsing. 2593 assert(_null_path == PATH_LIMIT-1, "delete last"); 2594 region->del_req(_null_path); 2595 phi ->del_req(_null_path); 2596 } 2597 2598 if (ProfileDynamicTypes && data != NULL) { 2599 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL); 2600 if (stopped()) { // Profile disagrees with this path. 2601 set_control(null_ctl); // Null is the only remaining possibility. 2602 return intcon(0); 2603 } 2604 if (cast_obj != NULL) 2605 not_null_obj = cast_obj; 2606 } 2607 2608 // Load the object's klass 2609 Node* obj_klass = load_object_klass(not_null_obj); 2610 2611 // Generate the subtype check 2612 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2613 2614 // Plug in the success path to the general merge in slot 1. 2615 region->init_req(_obj_path, control()); 2616 phi ->init_req(_obj_path, intcon(1)); 2617 2618 // Plug in the failing path to the general merge in slot 2. 2619 region->init_req(_fail_path, not_subtype_ctrl); 2620 phi ->init_req(_fail_path, intcon(0)); 2621 2622 // Return final merged results 2623 set_control( _gvn.transform(region) ); 2624 record_for_igvn(region); 2625 return _gvn.transform(phi); 2626 } 2627 2628 //-------------------------------gen_checkcast--------------------------------- 2629 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2630 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2631 // uncommon-trap paths work. Adjust stack after this call. 2632 // If failure_control is supplied and not null, it is filled in with 2633 // the control edge for the cast failure. Otherwise, an appropriate 2634 // uncommon trap or exception is thrown. 2635 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2636 Node* *failure_control) { 2637 kill_dead_locals(); // Benefit all the uncommon traps 2638 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2639 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2640 2641 // Fast cutout: Check the case that the cast is vacuously true. 2642 // This detects the common cases where the test will short-circuit 2643 // away completely. We do this before we perform the null check, 2644 // because if the test is going to turn into zero code, we don't 2645 // want a residual null check left around. (Causes a slowdown, 2646 // for example, in some objArray manipulations, such as a[i]=a[j].) 2647 if (tk->singleton()) { 2648 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2649 if (objtp != NULL && objtp->klass() != NULL) { 2650 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2651 case SSC_always_true: 2652 return obj; 2653 case SSC_always_false: 2654 // It needs a null check because a null will *pass* the cast check. 2655 // A non-null value will always produce an exception. 2656 return do_null_assert(obj, T_OBJECT); 2657 } 2658 } 2659 } 2660 2661 ciProfileData* data = NULL; 2662 if (failure_control == NULL) { // use MDO in regular case only 2663 assert(java_bc() == Bytecodes::_aastore || 2664 java_bc() == Bytecodes::_checkcast, 2665 "interpreter profiles type checks only for these BCs"); 2666 data = method()->method_data()->bci_to_data(bci()); 2667 } 2668 2669 // Make the merge point 2670 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2671 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2672 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); 2673 C->set_has_split_ifs(true); // Has chance for split-if optimization 2674 2675 // Use null-cast information if it is available 2676 bool never_see_null = ((failure_control == NULL) // regular case only 2677 && seems_never_null(obj, data)); 2678 2679 // Null check; get casted pointer; set region slot 3 2680 Node* null_ctl = top(); 2681 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2682 2683 // If not_null_obj is dead, only null-path is taken 2684 if (stopped()) { // Doing instance-of on a NULL? 2685 set_control(null_ctl); 2686 return null(); 2687 } 2688 region->init_req(_null_path, null_ctl); 2689 phi ->init_req(_null_path, null()); // Set null path value 2690 if (null_ctl == top()) { 2691 // Do this eagerly, so that pattern matches like is_diamond_phi 2692 // will work even during parsing. 2693 assert(_null_path == PATH_LIMIT-1, "delete last"); 2694 region->del_req(_null_path); 2695 phi ->del_req(_null_path); 2696 } 2697 2698 Node* cast_obj = NULL; 2699 if (data != NULL && 2700 // Counter has never been decremented (due to cast failure). 2701 // ...This is a reasonable thing to expect. It is true of 2702 // all casts inserted by javac to implement generic types. 2703 data->as_CounterData()->count() >= 0) { 2704 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass()); 2705 if (cast_obj != NULL) { 2706 if (failure_control != NULL) // failure is now impossible 2707 (*failure_control) = top(); 2708 // adjust the type of the phi to the exact klass: 2709 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR)); 2710 } 2711 } 2712 2713 if (cast_obj == NULL) { 2714 // Load the object's klass 2715 Node* obj_klass = load_object_klass(not_null_obj); 2716 2717 // Generate the subtype check 2718 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2719 2720 // Plug in success path into the merge 2721 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), 2722 not_null_obj, toop)); 2723 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2724 if (failure_control == NULL) { 2725 if (not_subtype_ctrl != top()) { // If failure is possible 2726 PreserveJVMState pjvms(this); 2727 set_control(not_subtype_ctrl); 2728 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2729 } 2730 } else { 2731 (*failure_control) = not_subtype_ctrl; 2732 } 2733 } 2734 2735 region->init_req(_obj_path, control()); 2736 phi ->init_req(_obj_path, cast_obj); 2737 2738 // A merge of NULL or Casted-NotNull obj 2739 Node* res = _gvn.transform(phi); 2740 2741 // Note I do NOT always 'replace_in_map(obj,result)' here. 2742 // if( tk->klass()->can_be_primary_super() ) 2743 // This means that if I successfully store an Object into an array-of-String 2744 // I 'forget' that the Object is really now known to be a String. I have to 2745 // do this because we don't have true union types for interfaces - if I store 2746 // a Baz into an array-of-Interface and then tell the optimizer it's an 2747 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2748 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2749 // replace_in_map( obj, res ); 2750 2751 // Return final merged results 2752 set_control( _gvn.transform(region) ); 2753 record_for_igvn(region); 2754 return res; 2755 } 2756 2757 //------------------------------next_monitor----------------------------------- 2758 // What number should be given to the next monitor? 2759 int GraphKit::next_monitor() { 2760 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2761 int next = current + C->sync_stack_slots(); 2762 // Keep the toplevel high water mark current: 2763 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2764 return current; 2765 } 2766 2767 //------------------------------insert_mem_bar--------------------------------- 2768 // Memory barrier to avoid floating things around 2769 // The membar serves as a pinch point between both control and all memory slices. 2770 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2771 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2772 mb->init_req(TypeFunc::Control, control()); 2773 mb->init_req(TypeFunc::Memory, reset_memory()); 2774 Node* membar = _gvn.transform(mb); 2775 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); 2776 set_all_memory_call(membar); 2777 return membar; 2778 } 2779 2780 //-------------------------insert_mem_bar_volatile---------------------------- 2781 // Memory barrier to avoid floating things around 2782 // The membar serves as a pinch point between both control and memory(alias_idx). 2783 // If you want to make a pinch point on all memory slices, do not use this 2784 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2785 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2786 // When Parse::do_put_xxx updates a volatile field, it appends a series 2787 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2788 // The first membar is on the same memory slice as the field store opcode. 2789 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2790 // All the other membars (for other volatile slices, including AliasIdxBot, 2791 // which stands for all unknown volatile slices) are control-dependent 2792 // on the first membar. This prevents later volatile loads or stores 2793 // from sliding up past the just-emitted store. 2794 2795 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2796 mb->set_req(TypeFunc::Control,control()); 2797 if (alias_idx == Compile::AliasIdxBot) { 2798 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2799 } else { 2800 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2801 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2802 } 2803 Node* membar = _gvn.transform(mb); 2804 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); 2805 if (alias_idx == Compile::AliasIdxBot) { 2806 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); 2807 } else { 2808 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2809 } 2810 return membar; 2811 } 2812 2813 //------------------------------shared_lock------------------------------------ 2814 // Emit locking code. 2815 FastLockNode* GraphKit::shared_lock(Node* obj) { 2816 // bci is either a monitorenter bc or InvocationEntryBci 2817 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2818 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2819 2820 if( !GenerateSynchronizationCode ) 2821 return NULL; // Not locking things? 2822 if (stopped()) // Dead monitor? 2823 return NULL; 2824 2825 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2826 2827 // Box the stack location 2828 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); 2829 Node* mem = reset_memory(); 2830 2831 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); 2832 if (PrintPreciseBiasedLockingStatistics) { 2833 // Create the counters for this fast lock. 2834 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2835 } 2836 // Add monitor to debug info for the slow path. If we block inside the 2837 // slow path and de-opt, we need the monitor hanging around 2838 map()->push_monitor( flock ); 2839 2840 const TypeFunc *tf = LockNode::lock_type(); 2841 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); 2842 2843 lock->init_req( TypeFunc::Control, control() ); 2844 lock->init_req( TypeFunc::Memory , mem ); 2845 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2846 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2847 lock->init_req( TypeFunc::ReturnAdr, top() ); 2848 2849 lock->init_req(TypeFunc::Parms + 0, obj); 2850 lock->init_req(TypeFunc::Parms + 1, box); 2851 lock->init_req(TypeFunc::Parms + 2, flock); 2852 add_safepoint_edges(lock); 2853 2854 lock = _gvn.transform( lock )->as_Lock(); 2855 2856 // lock has no side-effects, sets few values 2857 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2858 2859 insert_mem_bar(Op_MemBarAcquireLock); 2860 2861 // Add this to the worklist so that the lock can be eliminated 2862 record_for_igvn(lock); 2863 2864 #ifndef PRODUCT 2865 if (PrintLockStatistics) { 2866 // Update the counter for this lock. Don't bother using an atomic 2867 // operation since we don't require absolute accuracy. 2868 lock->create_lock_counter(map()->jvms()); 2869 increment_counter(lock->counter()->addr()); 2870 } 2871 #endif 2872 2873 return flock; 2874 } 2875 2876 2877 //------------------------------shared_unlock---------------------------------- 2878 // Emit unlocking code. 2879 void GraphKit::shared_unlock(Node* box, Node* obj) { 2880 // bci is either a monitorenter bc or InvocationEntryBci 2881 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2882 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2883 2884 if( !GenerateSynchronizationCode ) 2885 return; 2886 if (stopped()) { // Dead monitor? 2887 map()->pop_monitor(); // Kill monitor from debug info 2888 return; 2889 } 2890 2891 // Memory barrier to avoid floating things down past the locked region 2892 insert_mem_bar(Op_MemBarReleaseLock); 2893 2894 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2895 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); 2896 uint raw_idx = Compile::AliasIdxRaw; 2897 unlock->init_req( TypeFunc::Control, control() ); 2898 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2899 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2900 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2901 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2902 2903 unlock->init_req(TypeFunc::Parms + 0, obj); 2904 unlock->init_req(TypeFunc::Parms + 1, box); 2905 unlock = _gvn.transform(unlock)->as_Unlock(); 2906 2907 Node* mem = reset_memory(); 2908 2909 // unlock has no side-effects, sets few values 2910 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 2911 2912 // Kill monitor from debug info 2913 map()->pop_monitor( ); 2914 } 2915 2916 //-------------------------------get_layout_helper----------------------------- 2917 // If the given klass is a constant or known to be an array, 2918 // fetch the constant layout helper value into constant_value 2919 // and return (Node*)NULL. Otherwise, load the non-constant 2920 // layout helper value, and return the node which represents it. 2921 // This two-faced routine is useful because allocation sites 2922 // almost always feature constant types. 2923 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 2924 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 2925 if (!StressReflectiveCode && inst_klass != NULL) { 2926 ciKlass* klass = inst_klass->klass(); 2927 bool xklass = inst_klass->klass_is_exact(); 2928 if (xklass || klass->is_array_klass()) { 2929 jint lhelper = klass->layout_helper(); 2930 if (lhelper != Klass::_lh_neutral_value) { 2931 constant_value = lhelper; 2932 return (Node*) NULL; 2933 } 2934 } 2935 } 2936 constant_value = Klass::_lh_neutral_value; // put in a known value 2937 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 2938 return make_load(NULL, lhp, TypeInt::INT, T_INT); 2939 } 2940 2941 // We just put in an allocate/initialize with a big raw-memory effect. 2942 // Hook selected additional alias categories on the initialization. 2943 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 2944 MergeMemNode* init_in_merge, 2945 Node* init_out_raw) { 2946 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 2947 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 2948 2949 Node* prevmem = kit.memory(alias_idx); 2950 init_in_merge->set_memory_at(alias_idx, prevmem); 2951 kit.set_memory(init_out_raw, alias_idx); 2952 } 2953 2954 //---------------------------set_output_for_allocation------------------------- 2955 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 2956 const TypeOopPtr* oop_type) { 2957 int rawidx = Compile::AliasIdxRaw; 2958 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 2959 add_safepoint_edges(alloc); 2960 Node* allocx = _gvn.transform(alloc); 2961 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); 2962 // create memory projection for i_o 2963 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 2964 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); 2965 2966 // create a memory projection as for the normal control path 2967 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); 2968 set_memory(malloc, rawidx); 2969 2970 // a normal slow-call doesn't change i_o, but an allocation does 2971 // we create a separate i_o projection for the normal control path 2972 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); 2973 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); 2974 2975 // put in an initialization barrier 2976 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 2977 rawoop)->as_Initialize(); 2978 assert(alloc->initialization() == init, "2-way macro link must work"); 2979 assert(init ->allocation() == alloc, "2-way macro link must work"); 2980 { 2981 // Extract memory strands which may participate in the new object's 2982 // initialization, and source them from the new InitializeNode. 2983 // This will allow us to observe initializations when they occur, 2984 // and link them properly (as a group) to the InitializeNode. 2985 assert(init->in(InitializeNode::Memory) == malloc, ""); 2986 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 2987 init->set_req(InitializeNode::Memory, minit_in); 2988 record_for_igvn(minit_in); // fold it up later, if possible 2989 Node* minit_out = memory(rawidx); 2990 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 2991 if (oop_type->isa_aryptr()) { 2992 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 2993 int elemidx = C->get_alias_index(telemref); 2994 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 2995 } else if (oop_type->isa_instptr()) { 2996 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 2997 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 2998 ciField* field = ik->nonstatic_field_at(i); 2999 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3000 continue; // do not bother to track really large numbers of fields 3001 // Find (or create) the alias category for this field: 3002 int fieldidx = C->alias_type(field)->index(); 3003 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3004 } 3005 } 3006 } 3007 3008 // Cast raw oop to the real thing... 3009 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); 3010 javaoop = _gvn.transform(javaoop); 3011 C->set_recent_alloc(control(), javaoop); 3012 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3013 3014 #ifdef ASSERT 3015 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3016 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3017 "Ideal_allocation works"); 3018 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3019 "Ideal_allocation works"); 3020 if (alloc->is_AllocateArray()) { 3021 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3022 "Ideal_allocation works"); 3023 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3024 "Ideal_allocation works"); 3025 } else { 3026 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3027 } 3028 } 3029 #endif //ASSERT 3030 3031 return javaoop; 3032 } 3033 3034 //---------------------------new_instance-------------------------------------- 3035 // This routine takes a klass_node which may be constant (for a static type) 3036 // or may be non-constant (for reflective code). It will work equally well 3037 // for either, and the graph will fold nicely if the optimizer later reduces 3038 // the type to a constant. 3039 // The optional arguments are for specialized use by intrinsics: 3040 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3041 // - If 'return_size_val', report the the total object size to the caller. 3042 Node* GraphKit::new_instance(Node* klass_node, 3043 Node* extra_slow_test, 3044 Node* *return_size_val) { 3045 // Compute size in doublewords 3046 // The size is always an integral number of doublewords, represented 3047 // as a positive bytewise size stored in the klass's layout_helper. 3048 // The layout_helper also encodes (in a low bit) the need for a slow path. 3049 jint layout_con = Klass::_lh_neutral_value; 3050 Node* layout_val = get_layout_helper(klass_node, layout_con); 3051 int layout_is_con = (layout_val == NULL); 3052 3053 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3054 // Generate the initial go-slow test. It's either ALWAYS (return a 3055 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3056 // case) a computed value derived from the layout_helper. 3057 Node* initial_slow_test = NULL; 3058 if (layout_is_con) { 3059 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3060 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3061 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3062 3063 } else { // reflective case 3064 // This reflective path is used by Unsafe.allocateInstance. 3065 // (It may be stress-tested by specifying StressReflectiveCode.) 3066 // Basically, we want to get into the VM is there's an illegal argument. 3067 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3068 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); 3069 if (extra_slow_test != intcon(0)) { 3070 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); 3071 } 3072 // (Macro-expander will further convert this to a Bool, if necessary.) 3073 } 3074 3075 // Find the size in bytes. This is easy; it's the layout_helper. 3076 // The size value must be valid even if the slow path is taken. 3077 Node* size = NULL; 3078 if (layout_is_con) { 3079 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3080 } else { // reflective case 3081 // This reflective path is used by clone and Unsafe.allocateInstance. 3082 size = ConvI2X(layout_val); 3083 3084 // Clear the low bits to extract layout_helper_size_in_bytes: 3085 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3086 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3087 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); 3088 } 3089 if (return_size_val != NULL) { 3090 (*return_size_val) = size; 3091 } 3092 3093 // This is a precise notnull oop of the klass. 3094 // (Actually, it need not be precise if this is a reflective allocation.) 3095 // It's what we cast the result to. 3096 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3097 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3098 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3099 3100 // Now generate allocation code 3101 3102 // The entire memory state is needed for slow path of the allocation 3103 // since GC and deoptimization can happened. 3104 Node *mem = reset_memory(); 3105 set_all_memory(mem); // Create new memory state 3106 3107 AllocateNode* alloc 3108 = new (C, AllocateNode::ParmLimit) 3109 AllocateNode(C, AllocateNode::alloc_type(), 3110 control(), mem, i_o(), 3111 size, klass_node, 3112 initial_slow_test); 3113 3114 return set_output_for_allocation(alloc, oop_type); 3115 } 3116 3117 //-------------------------------new_array------------------------------------- 3118 // helper for both newarray and anewarray 3119 // The 'length' parameter is (obviously) the length of the array. 3120 // See comments on new_instance for the meaning of the other arguments. 3121 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3122 Node* length, // number of array elements 3123 int nargs, // number of arguments to push back for uncommon trap 3124 Node* *return_size_val) { 3125 jint layout_con = Klass::_lh_neutral_value; 3126 Node* layout_val = get_layout_helper(klass_node, layout_con); 3127 int layout_is_con = (layout_val == NULL); 3128 3129 if (!layout_is_con && !StressReflectiveCode && 3130 !too_many_traps(Deoptimization::Reason_class_check)) { 3131 // This is a reflective array creation site. 3132 // Optimistically assume that it is a subtype of Object[], 3133 // so that we can fold up all the address arithmetic. 3134 layout_con = Klass::array_layout_helper(T_OBJECT); 3135 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); 3136 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); 3137 { BuildCutout unless(this, bol_lh, PROB_MAX); 3138 _sp += nargs; 3139 uncommon_trap(Deoptimization::Reason_class_check, 3140 Deoptimization::Action_maybe_recompile); 3141 } 3142 layout_val = NULL; 3143 layout_is_con = true; 3144 } 3145 3146 // Generate the initial go-slow test. Make sure we do not overflow 3147 // if length is huge (near 2Gig) or negative! We do not need 3148 // exact double-words here, just a close approximation of needed 3149 // double-words. We can't add any offset or rounding bits, lest we 3150 // take a size -1 of bytes and make it positive. Use an unsigned 3151 // compare, so negative sizes look hugely positive. 3152 int fast_size_limit = FastAllocateSizeLimit; 3153 if (layout_is_con) { 3154 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3155 // Increase the size limit if we have exact knowledge of array type. 3156 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3157 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3158 } 3159 3160 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); 3161 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3162 if (initial_slow_test->is_Bool()) { 3163 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3164 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3165 } 3166 3167 // --- Size Computation --- 3168 // array_size = round_to_heap(array_header + (length << elem_shift)); 3169 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3170 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3171 // The rounding mask is strength-reduced, if possible. 3172 int round_mask = MinObjAlignmentInBytes - 1; 3173 Node* header_size = NULL; 3174 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3175 // (T_BYTE has the weakest alignment and size restrictions...) 3176 if (layout_is_con) { 3177 int hsize = Klass::layout_helper_header_size(layout_con); 3178 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3179 BasicType etype = Klass::layout_helper_element_type(layout_con); 3180 if ((round_mask & ~right_n_bits(eshift)) == 0) 3181 round_mask = 0; // strength-reduce it if it goes away completely 3182 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3183 assert(header_size_min <= hsize, "generic minimum is smallest"); 3184 header_size_min = hsize; 3185 header_size = intcon(hsize + round_mask); 3186 } else { 3187 Node* hss = intcon(Klass::_lh_header_size_shift); 3188 Node* hsm = intcon(Klass::_lh_header_size_mask); 3189 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); 3190 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); 3191 Node* mask = intcon(round_mask); 3192 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); 3193 } 3194 3195 Node* elem_shift = NULL; 3196 if (layout_is_con) { 3197 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3198 if (eshift != 0) 3199 elem_shift = intcon(eshift); 3200 } else { 3201 // There is no need to mask or shift this value. 3202 // The semantics of LShiftINode include an implicit mask to 0x1F. 3203 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3204 elem_shift = layout_val; 3205 } 3206 3207 // Transition to native address size for all offset calculations: 3208 Node* lengthx = ConvI2X(length); 3209 Node* headerx = ConvI2X(header_size); 3210 #ifdef _LP64 3211 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3212 if (tllen != NULL && tllen->_lo < 0) { 3213 // Add a manual constraint to a positive range. Cf. array_element_address. 3214 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3215 if (size_max > tllen->_hi) size_max = tllen->_hi; 3216 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3217 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); 3218 } 3219 } 3220 #endif 3221 3222 // Combine header size (plus rounding) and body size. Then round down. 3223 // This computation cannot overflow, because it is used only in two 3224 // places, one where the length is sharply limited, and the other 3225 // after a successful allocation. 3226 Node* abody = lengthx; 3227 if (elem_shift != NULL) 3228 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); 3229 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); 3230 if (round_mask != 0) { 3231 Node* mask = MakeConX(~round_mask); 3232 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); 3233 } 3234 // else if round_mask == 0, the size computation is self-rounding 3235 3236 if (return_size_val != NULL) { 3237 // This is the size 3238 (*return_size_val) = size; 3239 } 3240 3241 // Now generate allocation code 3242 3243 // The entire memory state is needed for slow path of the allocation 3244 // since GC and deoptimization can happened. 3245 Node *mem = reset_memory(); 3246 set_all_memory(mem); // Create new memory state 3247 3248 // Create the AllocateArrayNode and its result projections 3249 AllocateArrayNode* alloc 3250 = new (C, AllocateArrayNode::ParmLimit) 3251 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3252 control(), mem, i_o(), 3253 size, klass_node, 3254 initial_slow_test, 3255 length); 3256 3257 // Cast to correct type. Note that the klass_node may be constant or not, 3258 // and in the latter case the actual array type will be inexact also. 3259 // (This happens via a non-constant argument to inline_native_newArray.) 3260 // In any case, the value of klass_node provides the desired array type. 3261 const TypeInt* length_type = _gvn.find_int_type(length); 3262 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3263 if (ary_type->isa_aryptr() && length_type != NULL) { 3264 // Try to get a better type than POS for the size 3265 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3266 } 3267 3268 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3269 3270 // Cast length on remaining path to be as narrow as possible 3271 if (map()->find_edge(length) >= 0) { 3272 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3273 if (ccast != length) { 3274 _gvn.set_type_bottom(ccast); 3275 record_for_igvn(ccast); 3276 replace_in_map(length, ccast); 3277 } 3278 } 3279 3280 return javaoop; 3281 } 3282 3283 // The following "Ideal_foo" functions are placed here because they recognize 3284 // the graph shapes created by the functions immediately above. 3285 3286 //---------------------------Ideal_allocation---------------------------------- 3287 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3288 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3289 if (ptr == NULL) { // reduce dumb test in callers 3290 return NULL; 3291 } 3292 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3293 ptr = ptr->in(1); 3294 if (ptr == NULL) return NULL; 3295 } 3296 if (ptr->is_Proj()) { 3297 Node* allo = ptr->in(0); 3298 if (allo != NULL && allo->is_Allocate()) { 3299 return allo->as_Allocate(); 3300 } 3301 } 3302 // Report failure to match. 3303 return NULL; 3304 } 3305 3306 // Fancy version which also strips off an offset (and reports it to caller). 3307 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3308 intptr_t& offset) { 3309 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3310 if (base == NULL) return NULL; 3311 return Ideal_allocation(base, phase); 3312 } 3313 3314 // Trace Initialize <- Proj[Parm] <- Allocate 3315 AllocateNode* InitializeNode::allocation() { 3316 Node* rawoop = in(InitializeNode::RawAddress); 3317 if (rawoop->is_Proj()) { 3318 Node* alloc = rawoop->in(0); 3319 if (alloc->is_Allocate()) { 3320 return alloc->as_Allocate(); 3321 } 3322 } 3323 return NULL; 3324 } 3325 3326 // Trace Allocate -> Proj[Parm] -> Initialize 3327 InitializeNode* AllocateNode::initialization() { 3328 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3329 if (rawoop == NULL) return NULL; 3330 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3331 Node* init = rawoop->fast_out(i); 3332 if (init->is_Initialize()) { 3333 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3334 return init->as_Initialize(); 3335 } 3336 } 3337 return NULL; 3338 } 3339 3340 // Trace Allocate -> Proj[Parm] -> MemBarStoreStore 3341 MemBarStoreStoreNode* AllocateNode::storestore() { 3342 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3343 if (rawoop == NULL) return NULL; 3344 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3345 Node* storestore = rawoop->fast_out(i); 3346 if (storestore->is_MemBarStoreStore()) { 3347 return storestore->as_MemBarStoreStore(); 3348 } 3349 } 3350 return NULL; 3351 } 3352 3353 //----------------------------- loop predicates --------------------------- 3354 3355 //------------------------------add_predicate_impl---------------------------- 3356 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3357 // Too many traps seen? 3358 if (too_many_traps(reason)) { 3359 #ifdef ASSERT 3360 if (TraceLoopPredicate) { 3361 int tc = C->trap_count(reason); 3362 tty->print("too many traps=%s tcount=%d in ", 3363 Deoptimization::trap_reason_name(reason), tc); 3364 method()->print(); // which method has too many predicate traps 3365 tty->cr(); 3366 } 3367 #endif 3368 // We cannot afford to take more traps here, 3369 // do not generate predicate. 3370 return; 3371 } 3372 3373 Node *cont = _gvn.intcon(1); 3374 Node* opq = _gvn.transform(new (C, 2) Opaque1Node(C, cont)); 3375 Node *bol = _gvn.transform(new (C, 2) Conv2BNode(opq)); 3376 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3377 Node* iffalse = _gvn.transform(new (C, 1) IfFalseNode(iff)); 3378 C->add_predicate_opaq(opq); 3379 { 3380 PreserveJVMState pjvms(this); 3381 set_control(iffalse); 3382 _sp += nargs; 3383 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3384 } 3385 Node* iftrue = _gvn.transform(new (C, 1) IfTrueNode(iff)); 3386 set_control(iftrue); 3387 } 3388 3389 //------------------------------add_predicate--------------------------------- 3390 void GraphKit::add_predicate(int nargs) { 3391 if (UseLoopPredicate) { 3392 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3393 } 3394 // loop's limit check predicate should be near the loop. 3395 if (LoopLimitCheck) { 3396 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3397 } 3398 } 3399 3400 //----------------------------- store barriers ---------------------------- 3401 #define __ ideal. 3402 3403 void GraphKit::sync_kit(IdealKit& ideal) { 3404 set_all_memory(__ merged_memory()); 3405 set_i_o(__ i_o()); 3406 set_control(__ ctrl()); 3407 } 3408 3409 void GraphKit::final_sync(IdealKit& ideal) { 3410 // Final sync IdealKit and graphKit. 3411 __ drain_delay_transform(); 3412 sync_kit(ideal); 3413 } 3414 3415 // vanilla/CMS post barrier 3416 // Insert a write-barrier store. This is to let generational GC work; we have 3417 // to flag all oop-stores before the next GC point. 3418 void GraphKit::write_barrier_post(Node* oop_store, 3419 Node* obj, 3420 Node* adr, 3421 uint adr_idx, 3422 Node* val, 3423 bool use_precise) { 3424 // No store check needed if we're storing a NULL or an old object 3425 // (latter case is probably a string constant). The concurrent 3426 // mark sweep garbage collector, however, needs to have all nonNull 3427 // oop updates flagged via card-marks. 3428 if (val != NULL && val->is_Con()) { 3429 // must be either an oop or NULL 3430 const Type* t = val->bottom_type(); 3431 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3432 // stores of null never (?) need barriers 3433 return; 3434 ciObject* con = t->is_oopptr()->const_oop(); 3435 if (con != NULL 3436 && con->is_perm() 3437 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 3438 // no store barrier needed, because no old-to-new ref created 3439 return; 3440 } 3441 3442 if (use_ReduceInitialCardMarks() 3443 && obj == just_allocated_object(control())) { 3444 // We can skip marks on a freshly-allocated object in Eden. 3445 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3446 // That routine informs GC to take appropriate compensating steps, 3447 // upon a slow-path allocation, so as to make this card-mark 3448 // elision safe. 3449 return; 3450 } 3451 3452 if (!use_precise) { 3453 // All card marks for a (non-array) instance are in one place: 3454 adr = obj; 3455 } 3456 // (Else it's an array (or unknown), and we want more precise card marks.) 3457 assert(adr != NULL, ""); 3458 3459 IdealKit ideal(this, true); 3460 3461 // Convert the pointer to an int prior to doing math on it 3462 Node* cast = __ CastPX(__ ctrl(), adr); 3463 3464 // Divide by card size 3465 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3466 "Only one we handle so far."); 3467 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3468 3469 // Combine card table base and card offset 3470 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3471 3472 // Get the alias_index for raw card-mark memory 3473 int adr_type = Compile::AliasIdxRaw; 3474 Node* zero = __ ConI(0); // Dirty card value 3475 BasicType bt = T_BYTE; 3476 3477 if (UseCondCardMark) { 3478 // The classic GC reference write barrier is typically implemented 3479 // as a store into the global card mark table. Unfortunately 3480 // unconditional stores can result in false sharing and excessive 3481 // coherence traffic as well as false transactional aborts. 3482 // UseCondCardMark enables MP "polite" conditional card mark 3483 // stores. In theory we could relax the load from ctrl() to 3484 // no_ctrl, but that doesn't buy much latitude. 3485 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3486 __ if_then(card_val, BoolTest::ne, zero); 3487 } 3488 3489 // Smash zero into card 3490 if( !UseConcMarkSweepGC ) { 3491 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3492 } else { 3493 // Specialized path for CM store barrier 3494 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3495 } 3496 3497 if (UseCondCardMark) { 3498 __ end_if(); 3499 } 3500 3501 // Final sync IdealKit and GraphKit. 3502 final_sync(ideal); 3503 } 3504 3505 // G1 pre/post barriers 3506 void GraphKit::g1_write_barrier_pre(bool do_load, 3507 Node* obj, 3508 Node* adr, 3509 uint alias_idx, 3510 Node* val, 3511 const TypeOopPtr* val_type, 3512 Node* pre_val, 3513 BasicType bt) { 3514 3515 // Some sanity checks 3516 // Note: val is unused in this routine. 3517 3518 if (do_load) { 3519 // We need to generate the load of the previous value 3520 assert(obj != NULL, "must have a base"); 3521 assert(adr != NULL, "where are loading from?"); 3522 assert(pre_val == NULL, "loaded already?"); 3523 assert(val_type != NULL, "need a type"); 3524 } else { 3525 // In this case both val_type and alias_idx are unused. 3526 assert(pre_val != NULL, "must be loaded already"); 3527 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3528 } 3529 assert(bt == T_OBJECT, "or we shouldn't be here"); 3530 3531 IdealKit ideal(this, true); 3532 3533 Node* tls = __ thread(); // ThreadLocalStorage 3534 3535 Node* no_ctrl = NULL; 3536 Node* no_base = __ top(); 3537 Node* zero = __ ConI(0); 3538 3539 float likely = PROB_LIKELY(0.999); 3540 float unlikely = PROB_UNLIKELY(0.999); 3541 3542 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3543 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3544 3545 // Offsets into the thread 3546 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3547 PtrQueue::byte_offset_of_active()); 3548 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3549 PtrQueue::byte_offset_of_index()); 3550 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3551 PtrQueue::byte_offset_of_buf()); 3552 3553 // Now the actual pointers into the thread 3554 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3555 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3556 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3557 3558 // Now some of the values 3559 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3560 3561 // if (!marking) 3562 __ if_then(marking, BoolTest::ne, zero); { 3563 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3564 3565 if (do_load) { 3566 // load original value 3567 // alias_idx correct?? 3568 pre_val = __ load(no_ctrl, adr, val_type, bt, alias_idx); 3569 } 3570 3571 // if (pre_val != NULL) 3572 __ if_then(pre_val, BoolTest::ne, null()); { 3573 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3574 3575 // is the queue for this thread full? 3576 __ if_then(index, BoolTest::ne, zero, likely); { 3577 3578 // decrement the index 3579 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3580 Node* next_indexX = next_index; 3581 #ifdef _LP64 3582 // We could refine the type for what it's worth 3583 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3584 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3585 #endif 3586 3587 // Now get the buffer location we will log the previous value into and store it 3588 Node *log_addr = __ AddP(no_base, buffer, next_indexX); 3589 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw); 3590 // update the index 3591 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3592 3593 } __ else_(); { 3594 3595 // logging buffer is full, call the runtime 3596 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3597 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3598 } __ end_if(); // (!index) 3599 } __ end_if(); // (pre_val != NULL) 3600 } __ end_if(); // (!marking) 3601 3602 // Final sync IdealKit and GraphKit. 3603 final_sync(ideal); 3604 } 3605 3606 // 3607 // Update the card table and add card address to the queue 3608 // 3609 void GraphKit::g1_mark_card(IdealKit& ideal, 3610 Node* card_adr, 3611 Node* oop_store, 3612 uint oop_alias_idx, 3613 Node* index, 3614 Node* index_adr, 3615 Node* buffer, 3616 const TypeFunc* tf) { 3617 3618 Node* zero = __ ConI(0); 3619 Node* no_base = __ top(); 3620 BasicType card_bt = T_BYTE; 3621 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3622 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3623 3624 // Now do the queue work 3625 __ if_then(index, BoolTest::ne, zero); { 3626 3627 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3628 Node* next_indexX = next_index; 3629 #ifdef _LP64 3630 // We could refine the type for what it's worth 3631 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3632 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3633 #endif // _LP64 3634 Node* log_addr = __ AddP(no_base, buffer, next_indexX); 3635 3636 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3637 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3638 3639 } __ else_(); { 3640 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3641 } __ end_if(); 3642 3643 } 3644 3645 void GraphKit::g1_write_barrier_post(Node* oop_store, 3646 Node* obj, 3647 Node* adr, 3648 uint alias_idx, 3649 Node* val, 3650 BasicType bt, 3651 bool use_precise) { 3652 // If we are writing a NULL then we need no post barrier 3653 3654 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3655 // Must be NULL 3656 const Type* t = val->bottom_type(); 3657 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3658 // No post barrier if writing NULLx 3659 return; 3660 } 3661 3662 if (!use_precise) { 3663 // All card marks for a (non-array) instance are in one place: 3664 adr = obj; 3665 } 3666 // (Else it's an array (or unknown), and we want more precise card marks.) 3667 assert(adr != NULL, ""); 3668 3669 IdealKit ideal(this, true); 3670 3671 Node* tls = __ thread(); // ThreadLocalStorage 3672 3673 Node* no_base = __ top(); 3674 float likely = PROB_LIKELY(0.999); 3675 float unlikely = PROB_UNLIKELY(0.999); 3676 Node* zero = __ ConI(0); 3677 Node* zeroX = __ ConX(0); 3678 3679 // Get the alias_index for raw card-mark memory 3680 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3681 3682 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3683 3684 // Offsets into the thread 3685 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3686 PtrQueue::byte_offset_of_index()); 3687 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3688 PtrQueue::byte_offset_of_buf()); 3689 3690 // Pointers into the thread 3691 3692 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3693 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3694 3695 // Now some values 3696 // Use ctrl to avoid hoisting these values past a safepoint, which could 3697 // potentially reset these fields in the JavaThread. 3698 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3699 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3700 3701 // Convert the store obj pointer to an int prior to doing math on it 3702 // Must use ctrl to prevent "integerized oop" existing across safepoint 3703 Node* cast = __ CastPX(__ ctrl(), adr); 3704 3705 // Divide pointer by card size 3706 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3707 3708 // Combine card table base and card offset 3709 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3710 3711 // If we know the value being stored does it cross regions? 3712 3713 if (val != NULL) { 3714 // Does the store cause us to cross regions? 3715 3716 // Should be able to do an unsigned compare of region_size instead of 3717 // and extra shift. Do we have an unsigned compare?? 3718 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3719 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3720 3721 // if (xor_res == 0) same region so skip 3722 __ if_then(xor_res, BoolTest::ne, zeroX); { 3723 3724 // No barrier if we are storing a NULL 3725 __ if_then(val, BoolTest::ne, null(), unlikely); { 3726 3727 // Ok must mark the card if not already dirty 3728 3729 // load the original value of the card 3730 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3731 3732 __ if_then(card_val, BoolTest::ne, zero); { 3733 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3734 } __ end_if(); 3735 } __ end_if(); 3736 } __ end_if(); 3737 } else { 3738 // Object.clone() instrinsic uses this path. 3739 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3740 } 3741 3742 // Final sync IdealKit and GraphKit. 3743 final_sync(ideal); 3744 } 3745 #undef __