1 /* 2 * Copyright (c) 2001, 2012, 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 //--------------------------------sync_jvms_for_reexecute--------------------- 97 // Make sure our current jvms agrees with our parse state. This version 98 // uses the reexecute_sp for reexecuting bytecodes. 99 JVMState* GraphKit::sync_jvms_for_reexecute() { 100 JVMState* jvms = this->jvms(); 101 jvms->set_bci(bci()); // Record the new bci in the JVMState 102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 103 return jvms; 104 } 105 106 #ifdef ASSERT 107 bool GraphKit::jvms_in_sync() const { 108 Parse* parse = is_Parse(); 109 if (parse == NULL) { 110 if (bci() != jvms()->bci()) return false; 111 if (sp() != (int)jvms()->sp()) return false; 112 return true; 113 } 114 if (jvms()->method() != parse->method()) return false; 115 if (jvms()->bci() != parse->bci()) return false; 116 int jvms_sp = jvms()->sp(); 117 if (jvms_sp != parse->sp()) return false; 118 int jvms_depth = jvms()->depth(); 119 if (jvms_depth != parse->depth()) return false; 120 return true; 121 } 122 123 // Local helper checks for special internal merge points 124 // used to accumulate and merge exception states. 125 // They are marked by the region's in(0) edge being the map itself. 126 // Such merge points must never "escape" into the parser at large, 127 // until they have been handed to gvn.transform. 128 static bool is_hidden_merge(Node* reg) { 129 if (reg == NULL) return false; 130 if (reg->is_Phi()) { 131 reg = reg->in(0); 132 if (reg == NULL) return false; 133 } 134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 135 } 136 137 void GraphKit::verify_map() const { 138 if (map() == NULL) return; // null map is OK 139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 142 } 143 144 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 145 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 147 } 148 #endif 149 150 //---------------------------stop_and_kill_map--------------------------------- 151 // Set _map to NULL, signalling a stop to further bytecode execution. 152 // First smash the current map's control to a constant, to mark it dead. 153 void GraphKit::stop_and_kill_map() { 154 SafePointNode* dead_map = stop(); 155 if (dead_map != NULL) { 156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 157 assert(dead_map->is_killed(), "must be so marked"); 158 } 159 } 160 161 162 //--------------------------------stopped-------------------------------------- 163 // Tell if _map is NULL, or control is top. 164 bool GraphKit::stopped() { 165 if (map() == NULL) return true; 166 else if (control() == top()) return true; 167 else return false; 168 } 169 170 171 //-----------------------------has_ex_handler---------------------------------- 172 // Tell if this method or any caller method has exception handlers. 173 bool GraphKit::has_ex_handler() { 174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 176 return true; 177 } 178 } 179 return false; 180 } 181 182 //------------------------------save_ex_oop------------------------------------ 183 // Save an exception without blowing stack contents or other JVM state. 184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 186 ex_map->add_req(ex_oop); 187 debug_only(verify_exception_state(ex_map)); 188 } 189 190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 192 Node* ex_oop = ex_map->in(ex_map->req()-1); 193 if (clear_it) ex_map->del_req(ex_map->req()-1); 194 return ex_oop; 195 } 196 197 //-----------------------------saved_ex_oop------------------------------------ 198 // Recover a saved exception from its map. 199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 200 return common_saved_ex_oop(ex_map, false); 201 } 202 203 //--------------------------clear_saved_ex_oop--------------------------------- 204 // Erase a previously saved exception from its map. 205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 206 return common_saved_ex_oop(ex_map, true); 207 } 208 209 #ifdef ASSERT 210 //---------------------------has_saved_ex_oop---------------------------------- 211 // Erase a previously saved exception from its map. 212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 213 return ex_map->req() == ex_map->jvms()->endoff()+1; 214 } 215 #endif 216 217 //-------------------------make_exception_state-------------------------------- 218 // Turn the current JVM state into an exception state, appending the ex_oop. 219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 220 sync_jvms(); 221 SafePointNode* ex_map = stop(); // do not manipulate this map any more 222 set_saved_ex_oop(ex_map, ex_oop); 223 return ex_map; 224 } 225 226 227 //--------------------------add_exception_state-------------------------------- 228 // Add an exception to my list of exceptions. 229 void GraphKit::add_exception_state(SafePointNode* ex_map) { 230 if (ex_map == NULL || ex_map->control() == top()) { 231 return; 232 } 233 #ifdef ASSERT 234 verify_exception_state(ex_map); 235 if (has_exceptions()) { 236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 237 } 238 #endif 239 240 // If there is already an exception of exactly this type, merge with it. 241 // In particular, null-checks and other low-level exceptions common up here. 242 Node* ex_oop = saved_ex_oop(ex_map); 243 const Type* ex_type = _gvn.type(ex_oop); 244 if (ex_oop == top()) { 245 // No action needed. 246 return; 247 } 248 assert(ex_type->isa_instptr(), "exception must be an instance"); 249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 251 // We check sp also because call bytecodes can generate exceptions 252 // both before and after arguments are popped! 253 if (ex_type2 == ex_type 254 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 255 combine_exception_states(ex_map, e2); 256 return; 257 } 258 } 259 260 // No pre-existing exception of the same type. Chain it on the list. 261 push_exception_state(ex_map); 262 } 263 264 //-----------------------add_exception_states_from----------------------------- 265 void GraphKit::add_exception_states_from(JVMState* jvms) { 266 SafePointNode* ex_map = jvms->map()->next_exception(); 267 if (ex_map != NULL) { 268 jvms->map()->set_next_exception(NULL); 269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 270 next_map = ex_map->next_exception(); 271 ex_map->set_next_exception(NULL); 272 add_exception_state(ex_map); 273 } 274 } 275 } 276 277 //-----------------------transfer_exceptions_into_jvms------------------------- 278 JVMState* GraphKit::transfer_exceptions_into_jvms() { 279 if (map() == NULL) { 280 // We need a JVMS to carry the exceptions, but the map has gone away. 281 // Create a scratch JVMS, cloned from any of the exception states... 282 if (has_exceptions()) { 283 _map = _exceptions; 284 _map = clone_map(); 285 _map->set_next_exception(NULL); 286 clear_saved_ex_oop(_map); 287 debug_only(verify_map()); 288 } else { 289 // ...or created from scratch 290 JVMState* jvms = new (C) JVMState(_method, NULL); 291 jvms->set_bci(_bci); 292 jvms->set_sp(_sp); 293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); 294 set_jvms(jvms); 295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 296 set_all_memory(top()); 297 while (map()->req() < jvms->endoff()) map()->add_req(top()); 298 } 299 // (This is a kludge, in case you didn't notice.) 300 set_control(top()); 301 } 302 JVMState* jvms = sync_jvms(); 303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 304 jvms->map()->set_next_exception(_exceptions); 305 _exceptions = NULL; // done with this set of exceptions 306 return jvms; 307 } 308 309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 310 assert(is_hidden_merge(dstphi), "must be a special merge node"); 311 assert(is_hidden_merge(srcphi), "must be a special merge node"); 312 uint limit = srcphi->req(); 313 for (uint i = PhiNode::Input; i < limit; i++) { 314 dstphi->add_req(srcphi->in(i)); 315 } 316 } 317 static inline void add_one_req(Node* dstphi, Node* src) { 318 assert(is_hidden_merge(dstphi), "must be a special merge node"); 319 assert(!is_hidden_merge(src), "must not be a special merge node"); 320 dstphi->add_req(src); 321 } 322 323 //-----------------------combine_exception_states------------------------------ 324 // This helper function combines exception states by building phis on a 325 // specially marked state-merging region. These regions and phis are 326 // untransformed, and can build up gradually. The region is marked by 327 // having a control input of its exception map, rather than NULL. Such 328 // regions do not appear except in this function, and in use_exception_state. 329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 330 if (failing()) return; // dying anyway... 331 JVMState* ex_jvms = ex_map->_jvms; 332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 336 assert(ex_map->req() == phi_map->req(), "matching maps"); 337 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 338 Node* hidden_merge_mark = root(); 339 Node* region = phi_map->control(); 340 MergeMemNode* phi_mem = phi_map->merged_memory(); 341 MergeMemNode* ex_mem = ex_map->merged_memory(); 342 if (region->in(0) != hidden_merge_mark) { 343 // The control input is not (yet) a specially-marked region in phi_map. 344 // Make it so, and build some phis. 345 region = new (C) RegionNode(2); 346 _gvn.set_type(region, Type::CONTROL); 347 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 348 region->init_req(1, phi_map->control()); 349 phi_map->set_control(region); 350 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 351 record_for_igvn(io_phi); 352 _gvn.set_type(io_phi, Type::ABIO); 353 phi_map->set_i_o(io_phi); 354 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 355 Node* m = mms.memory(); 356 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 357 record_for_igvn(m_phi); 358 _gvn.set_type(m_phi, Type::MEMORY); 359 mms.set_memory(m_phi); 360 } 361 } 362 363 // Either or both of phi_map and ex_map might already be converted into phis. 364 Node* ex_control = ex_map->control(); 365 // if there is special marking on ex_map also, we add multiple edges from src 366 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 367 // how wide was the destination phi_map, originally? 368 uint orig_width = region->req(); 369 370 if (add_multiple) { 371 add_n_reqs(region, ex_control); 372 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 373 } else { 374 // ex_map has no merges, so we just add single edges everywhere 375 add_one_req(region, ex_control); 376 add_one_req(phi_map->i_o(), ex_map->i_o()); 377 } 378 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 379 if (mms.is_empty()) { 380 // get a copy of the base memory, and patch some inputs into it 381 const TypePtr* adr_type = mms.adr_type(C); 382 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 383 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 384 mms.set_memory(phi); 385 // Prepare to append interesting stuff onto the newly sliced phi: 386 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 387 } 388 // Append stuff from ex_map: 389 if (add_multiple) { 390 add_n_reqs(mms.memory(), mms.memory2()); 391 } else { 392 add_one_req(mms.memory(), mms.memory2()); 393 } 394 } 395 uint limit = ex_map->req(); 396 for (uint i = TypeFunc::Parms; i < limit; i++) { 397 // Skip everything in the JVMS after tos. (The ex_oop follows.) 398 if (i == tos) i = ex_jvms->monoff(); 399 Node* src = ex_map->in(i); 400 Node* dst = phi_map->in(i); 401 if (src != dst) { 402 PhiNode* phi; 403 if (dst->in(0) != region) { 404 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 405 record_for_igvn(phi); 406 _gvn.set_type(phi, phi->type()); 407 phi_map->set_req(i, dst); 408 // Prepare to append interesting stuff onto the new phi: 409 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 410 } else { 411 assert(dst->is_Phi(), "nobody else uses a hidden region"); 412 phi = (PhiNode*)dst; 413 } 414 if (add_multiple && src->in(0) == ex_control) { 415 // Both are phis. 416 add_n_reqs(dst, src); 417 } else { 418 while (dst->req() < region->req()) add_one_req(dst, src); 419 } 420 const Type* srctype = _gvn.type(src); 421 if (phi->type() != srctype) { 422 const Type* dsttype = phi->type()->meet(srctype); 423 if (phi->type() != dsttype) { 424 phi->set_type(dsttype); 425 _gvn.set_type(phi, dsttype); 426 } 427 } 428 } 429 } 430 phi_map->merge_replaced_nodes_with(ex_map); 431 } 432 433 //--------------------------use_exception_state-------------------------------- 434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 435 if (failing()) { stop(); return top(); } 436 Node* region = phi_map->control(); 437 Node* hidden_merge_mark = root(); 438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 439 Node* ex_oop = clear_saved_ex_oop(phi_map); 440 if (region->in(0) == hidden_merge_mark) { 441 // Special marking for internal ex-states. Process the phis now. 442 region->set_req(0, region); // now it's an ordinary region 443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 444 // Note: Setting the jvms also sets the bci and sp. 445 set_control(_gvn.transform(region)); 446 uint tos = jvms()->stkoff() + sp(); 447 for (uint i = 1; i < tos; i++) { 448 Node* x = phi_map->in(i); 449 if (x->in(0) == region) { 450 assert(x->is_Phi(), "expected a special phi"); 451 phi_map->set_req(i, _gvn.transform(x)); 452 } 453 } 454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 455 Node* x = mms.memory(); 456 if (x->in(0) == region) { 457 assert(x->is_Phi(), "nobody else uses a hidden region"); 458 mms.set_memory(_gvn.transform(x)); 459 } 460 } 461 if (ex_oop->in(0) == region) { 462 assert(ex_oop->is_Phi(), "expected a special phi"); 463 ex_oop = _gvn.transform(ex_oop); 464 } 465 } else { 466 set_jvms(phi_map->jvms()); 467 } 468 469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 471 return ex_oop; 472 } 473 474 //---------------------------------java_bc------------------------------------- 475 Bytecodes::Code GraphKit::java_bc() const { 476 ciMethod* method = this->method(); 477 int bci = this->bci(); 478 if (method != NULL && bci != InvocationEntryBci) 479 return method->java_code_at_bci(bci); 480 else 481 return Bytecodes::_illegal; 482 } 483 484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 485 bool must_throw) { 486 // if the exception capability is set, then we will generate code 487 // to check the JavaThread.should_post_on_exceptions flag to see 488 // if we actually need to report exception events (for this 489 // thread). If we don't need to report exception events, we will 490 // take the normal fast path provided by add_exception_events. If 491 // exception event reporting is enabled for this thread, we will 492 // take the uncommon_trap in the BuildCutout below. 493 494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false); 498 499 // Test the should_post_on_exceptions_flag vs. 0 500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 502 503 // Branch to slow_path if should_post_on_exceptions_flag was true 504 { BuildCutout unless(this, tst, PROB_MAX); 505 // Do not try anything fancy if we're notifying the VM on every throw. 506 // Cf. case Bytecodes::_athrow in parse2.cpp. 507 uncommon_trap(reason, Deoptimization::Action_none, 508 (ciKlass*)NULL, (char*)NULL, must_throw); 509 } 510 511 } 512 513 //------------------------------builtin_throw---------------------------------- 514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 515 bool must_throw = true; 516 517 if (env()->jvmti_can_post_on_exceptions()) { 518 // check if we must post exception events, take uncommon trap if so 519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 520 // here if should_post_on_exceptions is false 521 // continue on with the normal codegen 522 } 523 524 // If this particular condition has not yet happened at this 525 // bytecode, then use the uncommon trap mechanism, and allow for 526 // a future recompilation if several traps occur here. 527 // If the throw is hot, try to use a more complicated inline mechanism 528 // which keeps execution inside the compiled code. 529 bool treat_throw_as_hot = false; 530 ciMethodData* md = method()->method_data(); 531 532 if (ProfileTraps) { 533 if (too_many_traps(reason)) { 534 treat_throw_as_hot = true; 535 } 536 // (If there is no MDO at all, assume it is early in 537 // execution, and that any deopts are part of the 538 // startup transient, and don't need to be remembered.) 539 540 // Also, if there is a local exception handler, treat all throws 541 // as hot if there has been at least one in this method. 542 if (C->trap_count(reason) != 0 543 && method()->method_data()->trap_count(reason) != 0 544 && has_ex_handler()) { 545 treat_throw_as_hot = true; 546 } 547 } 548 549 // If this throw happens frequently, an uncommon trap might cause 550 // a performance pothole. If there is a local exception handler, 551 // and if this particular bytecode appears to be deoptimizing often, 552 // let us handle the throw inline, with a preconstructed instance. 553 // Note: If the deopt count has blown up, the uncommon trap 554 // runtime is going to flush this nmethod, not matter what. 555 if (treat_throw_as_hot 556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 557 // If the throw is local, we use a pre-existing instance and 558 // punt on the backtrace. This would lead to a missing backtrace 559 // (a repeat of 4292742) if the backtrace object is ever asked 560 // for its backtrace. 561 // Fixing this remaining case of 4292742 requires some flavor of 562 // escape analysis. Leave that for the future. 563 ciInstance* ex_obj = NULL; 564 switch (reason) { 565 case Deoptimization::Reason_null_check: 566 ex_obj = env()->NullPointerException_instance(); 567 break; 568 case Deoptimization::Reason_div0_check: 569 ex_obj = env()->ArithmeticException_instance(); 570 break; 571 case Deoptimization::Reason_range_check: 572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 573 break; 574 case Deoptimization::Reason_class_check: 575 if (java_bc() == Bytecodes::_aastore) { 576 ex_obj = env()->ArrayStoreException_instance(); 577 } else { 578 ex_obj = env()->ClassCastException_instance(); 579 } 580 break; 581 } 582 if (failing()) { stop(); return; } // exception allocation might fail 583 if (ex_obj != NULL) { 584 // Cheat with a preallocated exception object. 585 if (C->log() != NULL) 586 C->log()->elem("hot_throw preallocated='1' reason='%s'", 587 Deoptimization::trap_reason_name(reason)); 588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 590 591 // Clear the detail message of the preallocated exception object. 592 // Weblogic sometimes mutates the detail message of exceptions 593 // using reflection. 594 int offset = java_lang_Throwable::get_detailMessage_offset(); 595 const TypePtr* adr_typ = ex_con->add_offset(offset); 596 597 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 599 // Conservatively release stores of object references. 600 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, StoreNode::release); 601 602 add_exception_state(make_exception_state(ex_node)); 603 return; 604 } 605 } 606 607 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 608 // It won't be much cheaper than bailing to the interp., since we'll 609 // have to pass up all the debug-info, and the runtime will have to 610 // create the stack trace. 611 612 // Usual case: Bail to interpreter. 613 // Reserve the right to recompile if we haven't seen anything yet. 614 615 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 616 if (treat_throw_as_hot 617 && (method()->method_data()->trap_recompiled_at(bci()) 618 || C->too_many_traps(reason))) { 619 // We cannot afford to take more traps here. Suffer in the interpreter. 620 if (C->log() != NULL) 621 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 622 Deoptimization::trap_reason_name(reason), 623 C->trap_count(reason)); 624 action = Deoptimization::Action_none; 625 } 626 627 // "must_throw" prunes the JVM state to include only the stack, if there 628 // are no local exception handlers. This should cut down on register 629 // allocation time and code size, by drastically reducing the number 630 // of in-edges on the call to the uncommon trap. 631 632 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 633 } 634 635 636 //----------------------------PreserveJVMState--------------------------------- 637 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 638 debug_only(kit->verify_map()); 639 _kit = kit; 640 _map = kit->map(); // preserve the map 641 _sp = kit->sp(); 642 kit->set_map(clone_map ? kit->clone_map() : NULL); 643 #ifdef ASSERT 644 _bci = kit->bci(); 645 Parse* parser = kit->is_Parse(); 646 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 647 _block = block; 648 #endif 649 } 650 PreserveJVMState::~PreserveJVMState() { 651 GraphKit* kit = _kit; 652 #ifdef ASSERT 653 assert(kit->bci() == _bci, "bci must not shift"); 654 Parse* parser = kit->is_Parse(); 655 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 656 assert(block == _block, "block must not shift"); 657 #endif 658 kit->set_map(_map); 659 kit->set_sp(_sp); 660 } 661 662 663 //-----------------------------BuildCutout------------------------------------- 664 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 665 : PreserveJVMState(kit) 666 { 667 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 668 SafePointNode* outer_map = _map; // preserved map is caller's 669 SafePointNode* inner_map = kit->map(); 670 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 671 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 672 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 673 } 674 BuildCutout::~BuildCutout() { 675 GraphKit* kit = _kit; 676 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 677 } 678 679 //---------------------------PreserveReexecuteState---------------------------- 680 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 681 assert(!kit->stopped(), "must call stopped() before"); 682 _kit = kit; 683 _sp = kit->sp(); 684 _reexecute = kit->jvms()->_reexecute; 685 } 686 PreserveReexecuteState::~PreserveReexecuteState() { 687 if (_kit->stopped()) return; 688 _kit->jvms()->_reexecute = _reexecute; 689 _kit->set_sp(_sp); 690 } 691 692 //------------------------------clone_map-------------------------------------- 693 // Implementation of PreserveJVMState 694 // 695 // Only clone_map(...) here. If this function is only used in the 696 // PreserveJVMState class we may want to get rid of this extra 697 // function eventually and do it all there. 698 699 SafePointNode* GraphKit::clone_map() { 700 if (map() == NULL) return NULL; 701 702 // Clone the memory edge first 703 Node* mem = MergeMemNode::make(C, map()->memory()); 704 gvn().set_type_bottom(mem); 705 706 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 707 JVMState* jvms = this->jvms(); 708 JVMState* clonejvms = jvms->clone_shallow(C); 709 clonemap->set_memory(mem); 710 clonemap->set_jvms(clonejvms); 711 clonejvms->set_map(clonemap); 712 record_for_igvn(clonemap); 713 gvn().set_type_bottom(clonemap); 714 return clonemap; 715 } 716 717 718 //-----------------------------set_map_clone----------------------------------- 719 void GraphKit::set_map_clone(SafePointNode* m) { 720 _map = m; 721 _map = clone_map(); 722 _map->set_next_exception(NULL); 723 debug_only(verify_map()); 724 } 725 726 727 //----------------------------kill_dead_locals--------------------------------- 728 // Detect any locals which are known to be dead, and force them to top. 729 void GraphKit::kill_dead_locals() { 730 // Consult the liveness information for the locals. If any 731 // of them are unused, then they can be replaced by top(). This 732 // should help register allocation time and cut down on the size 733 // of the deoptimization information. 734 735 // This call is made from many of the bytecode handling 736 // subroutines called from the Big Switch in do_one_bytecode. 737 // Every bytecode which might include a slow path is responsible 738 // for killing its dead locals. The more consistent we 739 // are about killing deads, the fewer useless phis will be 740 // constructed for them at various merge points. 741 742 // bci can be -1 (InvocationEntryBci). We return the entry 743 // liveness for the method. 744 745 if (method() == NULL || method()->code_size() == 0) { 746 // We are building a graph for a call to a native method. 747 // All locals are live. 748 return; 749 } 750 751 ResourceMark rm; 752 753 // Consult the liveness information for the locals. If any 754 // of them are unused, then they can be replaced by top(). This 755 // should help register allocation time and cut down on the size 756 // of the deoptimization information. 757 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 758 759 int len = (int)live_locals.size(); 760 assert(len <= jvms()->loc_size(), "too many live locals"); 761 for (int local = 0; local < len; local++) { 762 if (!live_locals.at(local)) { 763 set_local(local, top()); 764 } 765 } 766 } 767 768 #ifdef ASSERT 769 //-------------------------dead_locals_are_killed------------------------------ 770 // Return true if all dead locals are set to top in the map. 771 // Used to assert "clean" debug info at various points. 772 bool GraphKit::dead_locals_are_killed() { 773 if (method() == NULL || method()->code_size() == 0) { 774 // No locals need to be dead, so all is as it should be. 775 return true; 776 } 777 778 // Make sure somebody called kill_dead_locals upstream. 779 ResourceMark rm; 780 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 781 if (jvms->loc_size() == 0) continue; // no locals to consult 782 SafePointNode* map = jvms->map(); 783 ciMethod* method = jvms->method(); 784 int bci = jvms->bci(); 785 if (jvms == this->jvms()) { 786 bci = this->bci(); // it might not yet be synched 787 } 788 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 789 int len = (int)live_locals.size(); 790 if (!live_locals.is_valid() || len == 0) 791 // This method is trivial, or is poisoned by a breakpoint. 792 return true; 793 assert(len == jvms->loc_size(), "live map consistent with locals map"); 794 for (int local = 0; local < len; local++) { 795 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 796 if (PrintMiscellaneous && (Verbose || WizardMode)) { 797 tty->print_cr("Zombie local %d: ", local); 798 jvms->dump(); 799 } 800 return false; 801 } 802 } 803 } 804 return true; 805 } 806 807 #endif //ASSERT 808 809 // Helper function for enforcing certain bytecodes to reexecute if 810 // deoptimization happens 811 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 812 ciMethod* cur_method = jvms->method(); 813 int cur_bci = jvms->bci(); 814 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 815 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 816 return Interpreter::bytecode_should_reexecute(code) || 817 is_anewarray && code == Bytecodes::_multianewarray; 818 // Reexecute _multianewarray bytecode which was replaced with 819 // sequence of [a]newarray. See Parse::do_multianewarray(). 820 // 821 // Note: interpreter should not have it set since this optimization 822 // is limited by dimensions and guarded by flag so in some cases 823 // multianewarray() runtime calls will be generated and 824 // the bytecode should not be reexecutes (stack will not be reset). 825 } else 826 return false; 827 } 828 829 // Helper function for adding JVMState and debug information to node 830 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 831 // Add the safepoint edges to the call (or other safepoint). 832 833 // Make sure dead locals are set to top. This 834 // should help register allocation time and cut down on the size 835 // of the deoptimization information. 836 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 837 838 // Walk the inline list to fill in the correct set of JVMState's 839 // Also fill in the associated edges for each JVMState. 840 841 // If the bytecode needs to be reexecuted we need to put 842 // the arguments back on the stack. 843 const bool should_reexecute = jvms()->should_reexecute(); 844 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 845 846 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 847 // undefined if the bci is different. This is normal for Parse but it 848 // should not happen for LibraryCallKit because only one bci is processed. 849 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 850 "in LibraryCallKit the reexecute bit should not change"); 851 852 // If we are guaranteed to throw, we can prune everything but the 853 // input to the current bytecode. 854 bool can_prune_locals = false; 855 uint stack_slots_not_pruned = 0; 856 int inputs = 0, depth = 0; 857 if (must_throw) { 858 assert(method() == youngest_jvms->method(), "sanity"); 859 if (compute_stack_effects(inputs, depth)) { 860 can_prune_locals = true; 861 stack_slots_not_pruned = inputs; 862 } 863 } 864 865 if (env()->jvmti_can_access_local_variables()) { 866 // At any safepoint, this method can get breakpointed, which would 867 // then require an immediate deoptimization. 868 can_prune_locals = false; // do not prune locals 869 stack_slots_not_pruned = 0; 870 } 871 872 // do not scribble on the input jvms 873 JVMState* out_jvms = youngest_jvms->clone_deep(C); 874 call->set_jvms(out_jvms); // Start jvms list for call node 875 876 // For a known set of bytecodes, the interpreter should reexecute them if 877 // deoptimization happens. We set the reexecute state for them here 878 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 879 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 880 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 881 } 882 883 // Presize the call: 884 DEBUG_ONLY(uint non_debug_edges = call->req()); 885 call->add_req_batch(top(), youngest_jvms->debug_depth()); 886 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 887 888 // Set up edges so that the call looks like this: 889 // Call [state:] ctl io mem fptr retadr 890 // [parms:] parm0 ... parmN 891 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 892 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 893 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 894 // Note that caller debug info precedes callee debug info. 895 896 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 897 uint debug_ptr = call->req(); 898 899 // Loop over the map input edges associated with jvms, add them 900 // to the call node, & reset all offsets to match call node array. 901 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 902 uint debug_end = debug_ptr; 903 uint debug_start = debug_ptr - in_jvms->debug_size(); 904 debug_ptr = debug_start; // back up the ptr 905 906 uint p = debug_start; // walks forward in [debug_start, debug_end) 907 uint j, k, l; 908 SafePointNode* in_map = in_jvms->map(); 909 out_jvms->set_map(call); 910 911 if (can_prune_locals) { 912 assert(in_jvms->method() == out_jvms->method(), "sanity"); 913 // If the current throw can reach an exception handler in this JVMS, 914 // then we must keep everything live that can reach that handler. 915 // As a quick and dirty approximation, we look for any handlers at all. 916 if (in_jvms->method()->has_exception_handlers()) { 917 can_prune_locals = false; 918 } 919 } 920 921 // Add the Locals 922 k = in_jvms->locoff(); 923 l = in_jvms->loc_size(); 924 out_jvms->set_locoff(p); 925 if (!can_prune_locals) { 926 for (j = 0; j < l; j++) 927 call->set_req(p++, in_map->in(k+j)); 928 } else { 929 p += l; // already set to top above by add_req_batch 930 } 931 932 // Add the Expression Stack 933 k = in_jvms->stkoff(); 934 l = in_jvms->sp(); 935 out_jvms->set_stkoff(p); 936 if (!can_prune_locals) { 937 for (j = 0; j < l; j++) 938 call->set_req(p++, in_map->in(k+j)); 939 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 940 // Divide stack into {S0,...,S1}, where S0 is set to top. 941 uint s1 = stack_slots_not_pruned; 942 stack_slots_not_pruned = 0; // for next iteration 943 if (s1 > l) s1 = l; 944 uint s0 = l - s1; 945 p += s0; // skip the tops preinstalled by add_req_batch 946 for (j = s0; j < l; j++) 947 call->set_req(p++, in_map->in(k+j)); 948 } else { 949 p += l; // already set to top above by add_req_batch 950 } 951 952 // Add the Monitors 953 k = in_jvms->monoff(); 954 l = in_jvms->mon_size(); 955 out_jvms->set_monoff(p); 956 for (j = 0; j < l; j++) 957 call->set_req(p++, in_map->in(k+j)); 958 959 // Copy any scalar object fields. 960 k = in_jvms->scloff(); 961 l = in_jvms->scl_size(); 962 out_jvms->set_scloff(p); 963 for (j = 0; j < l; j++) 964 call->set_req(p++, in_map->in(k+j)); 965 966 // Finish the new jvms. 967 out_jvms->set_endoff(p); 968 969 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 970 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 971 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 972 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 973 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 974 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 975 976 // Update the two tail pointers in parallel. 977 out_jvms = out_jvms->caller(); 978 in_jvms = in_jvms->caller(); 979 } 980 981 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 982 983 // Test the correctness of JVMState::debug_xxx accessors: 984 assert(call->jvms()->debug_start() == non_debug_edges, ""); 985 assert(call->jvms()->debug_end() == call->req(), ""); 986 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 987 } 988 989 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 990 Bytecodes::Code code = java_bc(); 991 if (code == Bytecodes::_wide) { 992 code = method()->java_code_at_bci(bci() + 1); 993 } 994 995 BasicType rtype = T_ILLEGAL; 996 int rsize = 0; 997 998 if (code != Bytecodes::_illegal) { 999 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1000 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1001 if (rtype < T_CONFLICT) 1002 rsize = type2size[rtype]; 1003 } 1004 1005 switch (code) { 1006 case Bytecodes::_illegal: 1007 return false; 1008 1009 case Bytecodes::_ldc: 1010 case Bytecodes::_ldc_w: 1011 case Bytecodes::_ldc2_w: 1012 inputs = 0; 1013 break; 1014 1015 case Bytecodes::_dup: inputs = 1; break; 1016 case Bytecodes::_dup_x1: inputs = 2; break; 1017 case Bytecodes::_dup_x2: inputs = 3; break; 1018 case Bytecodes::_dup2: inputs = 2; break; 1019 case Bytecodes::_dup2_x1: inputs = 3; break; 1020 case Bytecodes::_dup2_x2: inputs = 4; break; 1021 case Bytecodes::_swap: inputs = 2; break; 1022 case Bytecodes::_arraylength: inputs = 1; break; 1023 1024 case Bytecodes::_getstatic: 1025 case Bytecodes::_putstatic: 1026 case Bytecodes::_getfield: 1027 case Bytecodes::_putfield: 1028 { 1029 bool ignored_will_link; 1030 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1031 int size = field->type()->size(); 1032 bool is_get = (depth >= 0), is_static = (depth & 1); 1033 inputs = (is_static ? 0 : 1); 1034 if (is_get) { 1035 depth = size - inputs; 1036 } else { 1037 inputs += size; // putxxx pops the value from the stack 1038 depth = - inputs; 1039 } 1040 } 1041 break; 1042 1043 case Bytecodes::_invokevirtual: 1044 case Bytecodes::_invokespecial: 1045 case Bytecodes::_invokestatic: 1046 case Bytecodes::_invokedynamic: 1047 case Bytecodes::_invokeinterface: 1048 { 1049 bool ignored_will_link; 1050 ciSignature* declared_signature = NULL; 1051 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1052 assert(declared_signature != NULL, "cannot be null"); 1053 inputs = declared_signature->arg_size_for_bc(code); 1054 int size = declared_signature->return_type()->size(); 1055 depth = size - inputs; 1056 } 1057 break; 1058 1059 case Bytecodes::_multianewarray: 1060 { 1061 ciBytecodeStream iter(method()); 1062 iter.reset_to_bci(bci()); 1063 iter.next(); 1064 inputs = iter.get_dimensions(); 1065 assert(rsize == 1, ""); 1066 depth = rsize - inputs; 1067 } 1068 break; 1069 1070 case Bytecodes::_ireturn: 1071 case Bytecodes::_lreturn: 1072 case Bytecodes::_freturn: 1073 case Bytecodes::_dreturn: 1074 case Bytecodes::_areturn: 1075 assert(rsize = -depth, ""); 1076 inputs = rsize; 1077 break; 1078 1079 case Bytecodes::_jsr: 1080 case Bytecodes::_jsr_w: 1081 inputs = 0; 1082 depth = 1; // S.B. depth=1, not zero 1083 break; 1084 1085 default: 1086 // bytecode produces a typed result 1087 inputs = rsize - depth; 1088 assert(inputs >= 0, ""); 1089 break; 1090 } 1091 1092 #ifdef ASSERT 1093 // spot check 1094 int outputs = depth + inputs; 1095 assert(outputs >= 0, "sanity"); 1096 switch (code) { 1097 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1098 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1099 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1100 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1101 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1102 } 1103 #endif //ASSERT 1104 1105 return true; 1106 } 1107 1108 1109 1110 //------------------------------basic_plus_adr--------------------------------- 1111 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1112 // short-circuit a common case 1113 if (offset == intcon(0)) return ptr; 1114 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1115 } 1116 1117 Node* GraphKit::ConvI2L(Node* offset) { 1118 // short-circuit a common case 1119 jint offset_con = find_int_con(offset, Type::OffsetBot); 1120 if (offset_con != Type::OffsetBot) { 1121 return longcon((jlong) offset_con); 1122 } 1123 return _gvn.transform( new (C) ConvI2LNode(offset)); 1124 } 1125 1126 Node* GraphKit::ConvI2UL(Node* offset) { 1127 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1128 if (offset_con != (juint) Type::OffsetBot) { 1129 return longcon((julong) offset_con); 1130 } 1131 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset)); 1132 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) ); 1133 return _gvn.transform( new (C) AndLNode(conv, mask) ); 1134 } 1135 1136 Node* GraphKit::ConvL2I(Node* offset) { 1137 // short-circuit a common case 1138 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1139 if (offset_con != (jlong)Type::OffsetBot) { 1140 return intcon((int) offset_con); 1141 } 1142 return _gvn.transform( new (C) ConvL2INode(offset)); 1143 } 1144 1145 //-------------------------load_object_klass----------------------------------- 1146 Node* GraphKit::load_object_klass(Node* obj) { 1147 // Special-case a fresh allocation to avoid building nodes: 1148 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1149 if (akls != NULL) return akls; 1150 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1151 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1152 } 1153 1154 //-------------------------load_array_length----------------------------------- 1155 Node* GraphKit::load_array_length(Node* array) { 1156 // Special-case a fresh allocation to avoid building nodes: 1157 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1158 Node *alen; 1159 if (alloc == NULL) { 1160 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1161 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1162 } else { 1163 alen = alloc->Ideal_length(); 1164 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1165 if (ccast != alen) { 1166 alen = _gvn.transform(ccast); 1167 } 1168 } 1169 return alen; 1170 } 1171 1172 //------------------------------do_null_check---------------------------------- 1173 // Helper function to do a NULL pointer check. Returned value is 1174 // the incoming address with NULL casted away. You are allowed to use the 1175 // not-null value only if you are control dependent on the test. 1176 extern int explicit_null_checks_inserted, 1177 explicit_null_checks_elided; 1178 Node* GraphKit::null_check_common(Node* value, BasicType type, 1179 // optional arguments for variations: 1180 bool assert_null, 1181 Node* *null_control) { 1182 assert(!assert_null || null_control == NULL, "not both at once"); 1183 if (stopped()) return top(); 1184 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1185 // For some performance testing, we may wish to suppress null checking. 1186 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1187 return value; 1188 } 1189 explicit_null_checks_inserted++; 1190 1191 // Construct NULL check 1192 Node *chk = NULL; 1193 switch(type) { 1194 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1195 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1196 case T_ARRAY : // fall through 1197 type = T_OBJECT; // simplify further tests 1198 case T_OBJECT : { 1199 const Type *t = _gvn.type( value ); 1200 1201 const TypeOopPtr* tp = t->isa_oopptr(); 1202 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1203 // Only for do_null_check, not any of its siblings: 1204 && !assert_null && null_control == NULL) { 1205 // Usually, any field access or invocation on an unloaded oop type 1206 // will simply fail to link, since the statically linked class is 1207 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1208 // the static class is loaded but the sharper oop type is not. 1209 // Rather than checking for this obscure case in lots of places, 1210 // we simply observe that a null check on an unloaded class 1211 // will always be followed by a nonsense operation, so we 1212 // can just issue the uncommon trap here. 1213 // Our access to the unloaded class will only be correct 1214 // after it has been loaded and initialized, which requires 1215 // a trip through the interpreter. 1216 #ifndef PRODUCT 1217 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1218 #endif 1219 uncommon_trap(Deoptimization::Reason_unloaded, 1220 Deoptimization::Action_reinterpret, 1221 tp->klass(), "!loaded"); 1222 return top(); 1223 } 1224 1225 if (assert_null) { 1226 // See if the type is contained in NULL_PTR. 1227 // If so, then the value is already null. 1228 if (t->higher_equal(TypePtr::NULL_PTR)) { 1229 explicit_null_checks_elided++; 1230 return value; // Elided null assert quickly! 1231 } 1232 } else { 1233 // See if mixing in the NULL pointer changes type. 1234 // If so, then the NULL pointer was not allowed in the original 1235 // type. In other words, "value" was not-null. 1236 if (t->meet(TypePtr::NULL_PTR) != t) { 1237 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1238 explicit_null_checks_elided++; 1239 return value; // Elided null check quickly! 1240 } 1241 } 1242 chk = new (C) CmpPNode( value, null() ); 1243 break; 1244 } 1245 1246 default: 1247 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1248 } 1249 assert(chk != NULL, "sanity check"); 1250 chk = _gvn.transform(chk); 1251 1252 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1253 BoolNode *btst = new (C) BoolNode( chk, btest); 1254 Node *tst = _gvn.transform( btst ); 1255 1256 //----------- 1257 // if peephole optimizations occurred, a prior test existed. 1258 // If a prior test existed, maybe it dominates as we can avoid this test. 1259 if (tst != btst && type == T_OBJECT) { 1260 // At this point we want to scan up the CFG to see if we can 1261 // find an identical test (and so avoid this test altogether). 1262 Node *cfg = control(); 1263 int depth = 0; 1264 while( depth < 16 ) { // Limit search depth for speed 1265 if( cfg->Opcode() == Op_IfTrue && 1266 cfg->in(0)->in(1) == tst ) { 1267 // Found prior test. Use "cast_not_null" to construct an identical 1268 // CastPP (and hence hash to) as already exists for the prior test. 1269 // Return that casted value. 1270 if (assert_null) { 1271 replace_in_map(value, null()); 1272 return null(); // do not issue the redundant test 1273 } 1274 Node *oldcontrol = control(); 1275 set_control(cfg); 1276 Node *res = cast_not_null(value); 1277 set_control(oldcontrol); 1278 explicit_null_checks_elided++; 1279 return res; 1280 } 1281 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1282 if (cfg == NULL) break; // Quit at region nodes 1283 depth++; 1284 } 1285 } 1286 1287 //----------- 1288 // Branch to failure if null 1289 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1290 Deoptimization::DeoptReason reason; 1291 if (assert_null) 1292 reason = Deoptimization::Reason_null_assert; 1293 else if (type == T_OBJECT) 1294 reason = Deoptimization::Reason_null_check; 1295 else 1296 reason = Deoptimization::Reason_div0_check; 1297 1298 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1299 // ciMethodData::has_trap_at will return a conservative -1 if any 1300 // must-be-null assertion has failed. This could cause performance 1301 // problems for a method after its first do_null_assert failure. 1302 // Consider using 'Reason_class_check' instead? 1303 1304 // To cause an implicit null check, we set the not-null probability 1305 // to the maximum (PROB_MAX). For an explicit check the probability 1306 // is set to a smaller value. 1307 if (null_control != NULL || too_many_traps(reason)) { 1308 // probability is less likely 1309 ok_prob = PROB_LIKELY_MAG(3); 1310 } else if (!assert_null && 1311 (ImplicitNullCheckThreshold > 0) && 1312 method() != NULL && 1313 (method()->method_data()->trap_count(reason) 1314 >= (uint)ImplicitNullCheckThreshold)) { 1315 ok_prob = PROB_LIKELY_MAG(3); 1316 } 1317 1318 if (null_control != NULL) { 1319 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1320 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1321 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1322 if (null_true == top()) 1323 explicit_null_checks_elided++; 1324 (*null_control) = null_true; 1325 } else { 1326 BuildCutout unless(this, tst, ok_prob); 1327 // Check for optimizer eliding test at parse time 1328 if (stopped()) { 1329 // Failure not possible; do not bother making uncommon trap. 1330 explicit_null_checks_elided++; 1331 } else if (assert_null) { 1332 uncommon_trap(reason, 1333 Deoptimization::Action_make_not_entrant, 1334 NULL, "assert_null"); 1335 } else { 1336 replace_in_map(value, zerocon(type)); 1337 builtin_throw(reason); 1338 } 1339 } 1340 1341 // Must throw exception, fall-thru not possible? 1342 if (stopped()) { 1343 return top(); // No result 1344 } 1345 1346 if (assert_null) { 1347 // Cast obj to null on this path. 1348 replace_in_map(value, zerocon(type)); 1349 return zerocon(type); 1350 } 1351 1352 // Cast obj to not-null on this path, if there is no null_control. 1353 // (If there is a null_control, a non-null value may come back to haunt us.) 1354 if (type == T_OBJECT) { 1355 Node* cast = cast_not_null(value, false); 1356 if (null_control == NULL || (*null_control) == top()) 1357 replace_in_map(value, cast); 1358 value = cast; 1359 } 1360 1361 return value; 1362 } 1363 1364 1365 //------------------------------cast_not_null---------------------------------- 1366 // Cast obj to not-null on this path 1367 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1368 const Type *t = _gvn.type(obj); 1369 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1370 // Object is already not-null? 1371 if( t == t_not_null ) return obj; 1372 1373 Node *cast = new (C) CastPPNode(obj,t_not_null); 1374 cast->init_req(0, control()); 1375 cast = _gvn.transform( cast ); 1376 1377 // Scan for instances of 'obj' in the current JVM mapping. 1378 // These instances are known to be not-null after the test. 1379 if (do_replace_in_map) 1380 replace_in_map(obj, cast); 1381 1382 return cast; // Return casted value 1383 } 1384 1385 1386 //--------------------------replace_in_map------------------------------------- 1387 void GraphKit::replace_in_map(Node* old, Node* neww) { 1388 if (old == neww) { 1389 return; 1390 } 1391 1392 map()->replace_edge(old, neww); 1393 1394 // Note: This operation potentially replaces any edge 1395 // on the map. This includes locals, stack, and monitors 1396 // of the current (innermost) JVM state. 1397 1398 // don't let inconsistent types from profiling escape this 1399 // method 1400 1401 const Type* told = _gvn.type(old); 1402 const Type* tnew = _gvn.type(neww); 1403 1404 if (!tnew->higher_equal(told)) { 1405 return; 1406 } 1407 1408 map()->record_replaced_node(old, neww); 1409 } 1410 1411 1412 //============================================================================= 1413 //--------------------------------memory--------------------------------------- 1414 Node* GraphKit::memory(uint alias_idx) { 1415 MergeMemNode* mem = merged_memory(); 1416 Node* p = mem->memory_at(alias_idx); 1417 _gvn.set_type(p, Type::MEMORY); // must be mapped 1418 return p; 1419 } 1420 1421 //-----------------------------reset_memory------------------------------------ 1422 Node* GraphKit::reset_memory() { 1423 Node* mem = map()->memory(); 1424 // do not use this node for any more parsing! 1425 debug_only( map()->set_memory((Node*)NULL) ); 1426 return _gvn.transform( mem ); 1427 } 1428 1429 //------------------------------set_all_memory--------------------------------- 1430 void GraphKit::set_all_memory(Node* newmem) { 1431 Node* mergemem = MergeMemNode::make(C, newmem); 1432 gvn().set_type_bottom(mergemem); 1433 map()->set_memory(mergemem); 1434 } 1435 1436 //------------------------------set_all_memory_call---------------------------- 1437 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1438 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1439 set_all_memory(newmem); 1440 } 1441 1442 //============================================================================= 1443 // 1444 // parser factory methods for MemNodes 1445 // 1446 // These are layered on top of the factory methods in LoadNode and StoreNode, 1447 // and integrate with the parser's memory state and _gvn engine. 1448 // 1449 1450 // factory methods in "int adr_idx" 1451 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1452 int adr_idx, 1453 bool require_atomic_access, LoadNode::Sem sem) { 1454 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1455 const TypePtr* adr_type = NULL; // debug-mode-only argument 1456 debug_only(adr_type = C->get_adr_type(adr_idx)); 1457 Node* mem = memory(adr_idx); 1458 Node* ld; 1459 if (require_atomic_access && bt == T_LONG) { 1460 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, sem); 1461 } else { 1462 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, sem); 1463 } 1464 return _gvn.transform(ld); 1465 } 1466 1467 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1468 int adr_idx, 1469 bool require_atomic_access, 1470 StoreNode::Sem sem) { 1471 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1472 const TypePtr* adr_type = NULL; 1473 debug_only(adr_type = C->get_adr_type(adr_idx)); 1474 Node *mem = memory(adr_idx); 1475 Node* st; 1476 if (require_atomic_access && bt == T_LONG) { 1477 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, sem); 1478 } else { 1479 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, sem); 1480 } 1481 st = _gvn.transform(st); 1482 set_memory(st, adr_idx); 1483 // Back-to-back stores can only remove intermediate store with DU info 1484 // so push on worklist for optimizer. 1485 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1486 record_for_igvn(st); 1487 1488 return st; 1489 } 1490 1491 1492 void GraphKit::pre_barrier(bool do_load, 1493 Node* ctl, 1494 Node* obj, 1495 Node* adr, 1496 uint adr_idx, 1497 Node* val, 1498 const TypeOopPtr* val_type, 1499 Node* pre_val, 1500 BasicType bt) { 1501 1502 BarrierSet* bs = Universe::heap()->barrier_set(); 1503 set_control(ctl); 1504 switch (bs->kind()) { 1505 case BarrierSet::G1SATBCT: 1506 case BarrierSet::G1SATBCTLogging: 1507 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1508 break; 1509 1510 case BarrierSet::CardTableModRef: 1511 case BarrierSet::CardTableExtension: 1512 case BarrierSet::ModRef: 1513 break; 1514 1515 case BarrierSet::Other: 1516 default : 1517 ShouldNotReachHere(); 1518 1519 } 1520 } 1521 1522 void GraphKit::post_barrier(Node* ctl, 1523 Node* store, 1524 Node* obj, 1525 Node* adr, 1526 uint adr_idx, 1527 Node* val, 1528 BasicType bt, 1529 bool use_precise) { 1530 BarrierSet* bs = Universe::heap()->barrier_set(); 1531 set_control(ctl); 1532 switch (bs->kind()) { 1533 case BarrierSet::G1SATBCT: 1534 case BarrierSet::G1SATBCTLogging: 1535 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1536 break; 1537 1538 case BarrierSet::CardTableModRef: 1539 case BarrierSet::CardTableExtension: 1540 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1541 break; 1542 1543 case BarrierSet::ModRef: 1544 break; 1545 1546 case BarrierSet::Other: 1547 default : 1548 ShouldNotReachHere(); 1549 1550 } 1551 } 1552 1553 Node* GraphKit::store_oop(Node* ctl, 1554 Node* obj, 1555 Node* adr, 1556 const TypePtr* adr_type, 1557 Node* val, 1558 const TypeOopPtr* val_type, 1559 BasicType bt, 1560 bool use_precise, 1561 StoreNode::Sem sem) { 1562 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1563 // could be delayed during Parse (for example, in adjust_map_after_if()). 1564 // Execute transformation here to avoid barrier generation in such case. 1565 if (_gvn.type(val) == TypePtr::NULL_PTR) 1566 val = _gvn.makecon(TypePtr::NULL_PTR); 1567 1568 set_control(ctl); 1569 if (stopped()) return top(); // Dead path ? 1570 1571 assert(bt == T_OBJECT, "sanity"); 1572 assert(val != NULL, "not dead path"); 1573 uint adr_idx = C->get_alias_index(adr_type); 1574 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1575 1576 pre_barrier(true /* do_load */, 1577 control(), obj, adr, adr_idx, val, val_type, 1578 NULL /* pre_val */, 1579 bt); 1580 1581 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, false, sem); 1582 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1583 return store; 1584 } 1585 1586 // Could be an array or object we don't know at compile time (unsafe ref.) 1587 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1588 Node* obj, // containing obj 1589 Node* adr, // actual adress to store val at 1590 const TypePtr* adr_type, 1591 Node* val, 1592 BasicType bt, 1593 StoreNode::Sem sem) { 1594 Compile::AliasType* at = C->alias_type(adr_type); 1595 const TypeOopPtr* val_type = NULL; 1596 if (adr_type->isa_instptr()) { 1597 if (at->field() != NULL) { 1598 // known field. This code is a copy of the do_put_xxx logic. 1599 ciField* field = at->field(); 1600 if (!field->type()->is_loaded()) { 1601 val_type = TypeInstPtr::BOTTOM; 1602 } else { 1603 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1604 } 1605 } 1606 } else if (adr_type->isa_aryptr()) { 1607 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1608 } 1609 if (val_type == NULL) { 1610 val_type = TypeInstPtr::BOTTOM; 1611 } 1612 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, sem); 1613 } 1614 1615 1616 //-------------------------array_element_address------------------------- 1617 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1618 const TypeInt* sizetype) { 1619 uint shift = exact_log2(type2aelembytes(elembt)); 1620 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1621 1622 // short-circuit a common case (saves lots of confusing waste motion) 1623 jint idx_con = find_int_con(idx, -1); 1624 if (idx_con >= 0) { 1625 intptr_t offset = header + ((intptr_t)idx_con << shift); 1626 return basic_plus_adr(ary, offset); 1627 } 1628 1629 // must be correct type for alignment purposes 1630 Node* base = basic_plus_adr(ary, header); 1631 #ifdef _LP64 1632 // The scaled index operand to AddP must be a clean 64-bit value. 1633 // Java allows a 32-bit int to be incremented to a negative 1634 // value, which appears in a 64-bit register as a large 1635 // positive number. Using that large positive number as an 1636 // operand in pointer arithmetic has bad consequences. 1637 // On the other hand, 32-bit overflow is rare, and the possibility 1638 // can often be excluded, if we annotate the ConvI2L node with 1639 // a type assertion that its value is known to be a small positive 1640 // number. (The prior range check has ensured this.) 1641 // This assertion is used by ConvI2LNode::Ideal. 1642 int index_max = max_jint - 1; // array size is max_jint, index is one less 1643 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1644 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1645 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) ); 1646 #endif 1647 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1648 return basic_plus_adr(ary, base, scale); 1649 } 1650 1651 //-------------------------load_array_element------------------------- 1652 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1653 const Type* elemtype = arytype->elem(); 1654 BasicType elembt = elemtype->array_element_basic_type(); 1655 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1656 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1657 return ld; 1658 } 1659 1660 //-------------------------set_arguments_for_java_call------------------------- 1661 // Arguments (pre-popped from the stack) are taken from the JVMS. 1662 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1663 // Add the call arguments: 1664 uint nargs = call->method()->arg_size(); 1665 for (uint i = 0; i < nargs; i++) { 1666 Node* arg = argument(i); 1667 call->init_req(i + TypeFunc::Parms, arg); 1668 } 1669 } 1670 1671 //---------------------------set_edges_for_java_call--------------------------- 1672 // Connect a newly created call into the current JVMS. 1673 // A return value node (if any) is returned from set_edges_for_java_call. 1674 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1675 1676 // Add the predefined inputs: 1677 call->init_req( TypeFunc::Control, control() ); 1678 call->init_req( TypeFunc::I_O , i_o() ); 1679 call->init_req( TypeFunc::Memory , reset_memory() ); 1680 call->init_req( TypeFunc::FramePtr, frameptr() ); 1681 call->init_req( TypeFunc::ReturnAdr, top() ); 1682 1683 add_safepoint_edges(call, must_throw); 1684 1685 Node* xcall = _gvn.transform(call); 1686 1687 if (xcall == top()) { 1688 set_control(top()); 1689 return; 1690 } 1691 assert(xcall == call, "call identity is stable"); 1692 1693 // Re-use the current map to produce the result. 1694 1695 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1696 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1697 set_all_memory_call(xcall, separate_io_proj); 1698 1699 //return xcall; // no need, caller already has it 1700 } 1701 1702 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1703 if (stopped()) return top(); // maybe the call folded up? 1704 1705 // Capture the return value, if any. 1706 Node* ret; 1707 if (call->method() == NULL || 1708 call->method()->return_type()->basic_type() == T_VOID) 1709 ret = top(); 1710 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1711 1712 // Note: Since any out-of-line call can produce an exception, 1713 // we always insert an I_O projection from the call into the result. 1714 1715 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1716 1717 if (separate_io_proj) { 1718 // The caller requested separate projections be used by the fall 1719 // through and exceptional paths, so replace the projections for 1720 // the fall through path. 1721 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1722 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1723 } 1724 return ret; 1725 } 1726 1727 //--------------------set_predefined_input_for_runtime_call-------------------- 1728 // Reading and setting the memory state is way conservative here. 1729 // The real problem is that I am not doing real Type analysis on memory, 1730 // so I cannot distinguish card mark stores from other stores. Across a GC 1731 // point the Store Barrier and the card mark memory has to agree. I cannot 1732 // have a card mark store and its barrier split across the GC point from 1733 // either above or below. Here I get that to happen by reading ALL of memory. 1734 // A better answer would be to separate out card marks from other memory. 1735 // For now, return the input memory state, so that it can be reused 1736 // after the call, if this call has restricted memory effects. 1737 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1738 // Set fixed predefined input arguments 1739 Node* memory = reset_memory(); 1740 call->init_req( TypeFunc::Control, control() ); 1741 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1742 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1743 call->init_req( TypeFunc::FramePtr, frameptr() ); 1744 call->init_req( TypeFunc::ReturnAdr, top() ); 1745 return memory; 1746 } 1747 1748 //-------------------set_predefined_output_for_runtime_call-------------------- 1749 // Set control and memory (not i_o) from the call. 1750 // If keep_mem is not NULL, use it for the output state, 1751 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1752 // If hook_mem is NULL, this call produces no memory effects at all. 1753 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1754 // then only that memory slice is taken from the call. 1755 // In the last case, we must put an appropriate memory barrier before 1756 // the call, so as to create the correct anti-dependencies on loads 1757 // preceding the call. 1758 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1759 Node* keep_mem, 1760 const TypePtr* hook_mem) { 1761 // no i/o 1762 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1763 if (keep_mem) { 1764 // First clone the existing memory state 1765 set_all_memory(keep_mem); 1766 if (hook_mem != NULL) { 1767 // Make memory for the call 1768 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1769 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1770 // We also use hook_mem to extract specific effects from arraycopy stubs. 1771 set_memory(mem, hook_mem); 1772 } 1773 // ...else the call has NO memory effects. 1774 1775 // Make sure the call advertises its memory effects precisely. 1776 // This lets us build accurate anti-dependences in gcm.cpp. 1777 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1778 "call node must be constructed correctly"); 1779 } else { 1780 assert(hook_mem == NULL, ""); 1781 // This is not a "slow path" call; all memory comes from the call. 1782 set_all_memory_call(call); 1783 } 1784 } 1785 1786 1787 // Replace the call with the current state of the kit. 1788 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1789 JVMState* ejvms = NULL; 1790 if (has_exceptions()) { 1791 ejvms = transfer_exceptions_into_jvms(); 1792 } 1793 1794 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1795 ReplacedNodes replaced_nodes_exception; 1796 Node* ex_ctl = top(); 1797 1798 SafePointNode* final_state = stop(); 1799 1800 // Find all the needed outputs of this call 1801 CallProjections callprojs; 1802 call->extract_projections(&callprojs, true); 1803 1804 Node* init_mem = call->in(TypeFunc::Memory); 1805 Node* final_mem = final_state->in(TypeFunc::Memory); 1806 Node* final_ctl = final_state->in(TypeFunc::Control); 1807 Node* final_io = final_state->in(TypeFunc::I_O); 1808 1809 // Replace all the old call edges with the edges from the inlining result 1810 if (callprojs.fallthrough_catchproj != NULL) { 1811 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1812 } 1813 if (callprojs.fallthrough_memproj != NULL) { 1814 if (final_mem->is_MergeMem()) { 1815 // Parser's exits MergeMem was not transformed but may be optimized 1816 final_mem = _gvn.transform(final_mem); 1817 } 1818 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1819 } 1820 if (callprojs.fallthrough_ioproj != NULL) { 1821 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1822 } 1823 1824 // Replace the result with the new result if it exists and is used 1825 if (callprojs.resproj != NULL && result != NULL) { 1826 C->gvn_replace_by(callprojs.resproj, result); 1827 } 1828 1829 if (ejvms == NULL) { 1830 // No exception edges to simply kill off those paths 1831 if (callprojs.catchall_catchproj != NULL) { 1832 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1833 } 1834 if (callprojs.catchall_memproj != NULL) { 1835 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1836 } 1837 if (callprojs.catchall_ioproj != NULL) { 1838 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1839 } 1840 // Replace the old exception object with top 1841 if (callprojs.exobj != NULL) { 1842 C->gvn_replace_by(callprojs.exobj, C->top()); 1843 } 1844 } else { 1845 GraphKit ekit(ejvms); 1846 1847 // Load my combined exception state into the kit, with all phis transformed: 1848 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1849 replaced_nodes_exception = ex_map->replaced_nodes(); 1850 1851 Node* ex_oop = ekit.use_exception_state(ex_map); 1852 1853 if (callprojs.catchall_catchproj != NULL) { 1854 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1855 ex_ctl = ekit.control(); 1856 } 1857 if (callprojs.catchall_memproj != NULL) { 1858 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1859 } 1860 if (callprojs.catchall_ioproj != NULL) { 1861 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1862 } 1863 1864 // Replace the old exception object with the newly created one 1865 if (callprojs.exobj != NULL) { 1866 C->gvn_replace_by(callprojs.exobj, ex_oop); 1867 } 1868 } 1869 1870 // Disconnect the call from the graph 1871 call->disconnect_inputs(NULL, C); 1872 C->gvn_replace_by(call, C->top()); 1873 1874 // Clean up any MergeMems that feed other MergeMems since the 1875 // optimizer doesn't like that. 1876 if (final_mem->is_MergeMem()) { 1877 Node_List wl; 1878 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1879 Node* m = i.get(); 1880 if (m->is_MergeMem() && !wl.contains(m)) { 1881 wl.push(m); 1882 } 1883 } 1884 while (wl.size() > 0) { 1885 _gvn.transform(wl.pop()); 1886 } 1887 } 1888 1889 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1890 replaced_nodes.apply(C, final_ctl); 1891 } 1892 if (!ex_ctl->is_top() && do_replaced_nodes) { 1893 replaced_nodes_exception.apply(C, ex_ctl); 1894 } 1895 } 1896 1897 1898 //------------------------------increment_counter------------------------------ 1899 // for statistics: increment a VM counter by 1 1900 1901 void GraphKit::increment_counter(address counter_addr) { 1902 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1903 increment_counter(adr1); 1904 } 1905 1906 void GraphKit::increment_counter(Node* counter_addr) { 1907 int adr_type = Compile::AliasIdxRaw; 1908 Node* ctrl = control(); 1909 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type); 1910 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1911 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type ); 1912 } 1913 1914 1915 //------------------------------uncommon_trap---------------------------------- 1916 // Bail out to the interpreter in mid-method. Implemented by calling the 1917 // uncommon_trap blob. This helper function inserts a runtime call with the 1918 // right debug info. 1919 void GraphKit::uncommon_trap(int trap_request, 1920 ciKlass* klass, const char* comment, 1921 bool must_throw, 1922 bool keep_exact_action) { 1923 if (failing()) stop(); 1924 if (stopped()) return; // trap reachable? 1925 1926 // Note: If ProfileTraps is true, and if a deopt. actually 1927 // occurs here, the runtime will make sure an MDO exists. There is 1928 // no need to call method()->ensure_method_data() at this point. 1929 1930 // Set the stack pointer to the right value for reexecution: 1931 set_sp(reexecute_sp()); 1932 1933 #ifdef ASSERT 1934 if (!must_throw) { 1935 // Make sure the stack has at least enough depth to execute 1936 // the current bytecode. 1937 int inputs, ignored_depth; 1938 if (compute_stack_effects(inputs, ignored_depth)) { 1939 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1940 Bytecodes::name(java_bc()), sp(), inputs)); 1941 } 1942 } 1943 #endif 1944 1945 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1946 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1947 1948 switch (action) { 1949 case Deoptimization::Action_maybe_recompile: 1950 case Deoptimization::Action_reinterpret: 1951 // Temporary fix for 6529811 to allow virtual calls to be sure they 1952 // get the chance to go from mono->bi->mega 1953 if (!keep_exact_action && 1954 Deoptimization::trap_request_index(trap_request) < 0 && 1955 too_many_recompiles(reason)) { 1956 // This BCI is causing too many recompilations. 1957 action = Deoptimization::Action_none; 1958 trap_request = Deoptimization::make_trap_request(reason, action); 1959 } else { 1960 C->set_trap_can_recompile(true); 1961 } 1962 break; 1963 case Deoptimization::Action_make_not_entrant: 1964 C->set_trap_can_recompile(true); 1965 break; 1966 #ifdef ASSERT 1967 case Deoptimization::Action_none: 1968 case Deoptimization::Action_make_not_compilable: 1969 break; 1970 default: 1971 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 1972 break; 1973 #endif 1974 } 1975 1976 if (TraceOptoParse) { 1977 char buf[100]; 1978 tty->print_cr("Uncommon trap %s at bci:%d", 1979 Deoptimization::format_trap_request(buf, sizeof(buf), 1980 trap_request), bci()); 1981 } 1982 1983 CompileLog* log = C->log(); 1984 if (log != NULL) { 1985 int kid = (klass == NULL)? -1: log->identify(klass); 1986 log->begin_elem("uncommon_trap bci='%d'", bci()); 1987 char buf[100]; 1988 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1989 trap_request)); 1990 if (kid >= 0) log->print(" klass='%d'", kid); 1991 if (comment != NULL) log->print(" comment='%s'", comment); 1992 log->end_elem(); 1993 } 1994 1995 // Make sure any guarding test views this path as very unlikely 1996 Node *i0 = control()->in(0); 1997 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1998 IfNode *iff = i0->as_If(); 1999 float f = iff->_prob; // Get prob 2000 if (control()->Opcode() == Op_IfTrue) { 2001 if (f > PROB_UNLIKELY_MAG(4)) 2002 iff->_prob = PROB_MIN; 2003 } else { 2004 if (f < PROB_LIKELY_MAG(4)) 2005 iff->_prob = PROB_MAX; 2006 } 2007 } 2008 2009 // Clear out dead values from the debug info. 2010 kill_dead_locals(); 2011 2012 // Now insert the uncommon trap subroutine call 2013 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2014 const TypePtr* no_memory_effects = NULL; 2015 // Pass the index of the class to be loaded 2016 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2017 (must_throw ? RC_MUST_THROW : 0), 2018 OptoRuntime::uncommon_trap_Type(), 2019 call_addr, "uncommon_trap", no_memory_effects, 2020 intcon(trap_request)); 2021 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2022 "must extract request correctly from the graph"); 2023 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2024 2025 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2026 // The debug info is the only real input to this call. 2027 2028 // Halt-and-catch fire here. The above call should never return! 2029 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2030 _gvn.set_type_bottom(halt); 2031 root()->add_req(halt); 2032 2033 stop_and_kill_map(); 2034 } 2035 2036 2037 //--------------------------just_allocated_object------------------------------ 2038 // Report the object that was just allocated. 2039 // It must be the case that there are no intervening safepoints. 2040 // We use this to determine if an object is so "fresh" that 2041 // it does not require card marks. 2042 Node* GraphKit::just_allocated_object(Node* current_control) { 2043 if (C->recent_alloc_ctl() == current_control) 2044 return C->recent_alloc_obj(); 2045 return NULL; 2046 } 2047 2048 2049 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2050 // (Note: TypeFunc::make has a cache that makes this fast.) 2051 const TypeFunc* tf = TypeFunc::make(dest_method); 2052 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2053 for (int j = 0; j < nargs; j++) { 2054 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2055 if( targ->basic_type() == T_DOUBLE ) { 2056 // If any parameters are doubles, they must be rounded before 2057 // the call, dstore_rounding does gvn.transform 2058 Node *arg = argument(j); 2059 arg = dstore_rounding(arg); 2060 set_argument(j, arg); 2061 } 2062 } 2063 } 2064 2065 void GraphKit::round_double_result(ciMethod* dest_method) { 2066 // A non-strict method may return a double value which has an extended 2067 // exponent, but this must not be visible in a caller which is 'strict' 2068 // If a strict caller invokes a non-strict callee, round a double result 2069 2070 BasicType result_type = dest_method->return_type()->basic_type(); 2071 assert( method() != NULL, "must have caller context"); 2072 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2073 // Destination method's return value is on top of stack 2074 // dstore_rounding() does gvn.transform 2075 Node *result = pop_pair(); 2076 result = dstore_rounding(result); 2077 push_pair(result); 2078 } 2079 } 2080 2081 // rounding for strict float precision conformance 2082 Node* GraphKit::precision_rounding(Node* n) { 2083 return UseStrictFP && _method->flags().is_strict() 2084 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2085 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2086 : n; 2087 } 2088 2089 // rounding for strict double precision conformance 2090 Node* GraphKit::dprecision_rounding(Node *n) { 2091 return UseStrictFP && _method->flags().is_strict() 2092 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2093 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2094 : n; 2095 } 2096 2097 // rounding for non-strict double stores 2098 Node* GraphKit::dstore_rounding(Node* n) { 2099 return Matcher::strict_fp_requires_explicit_rounding 2100 && UseSSE <= 1 2101 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2102 : n; 2103 } 2104 2105 //============================================================================= 2106 // Generate a fast path/slow path idiom. Graph looks like: 2107 // [foo] indicates that 'foo' is a parameter 2108 // 2109 // [in] NULL 2110 // \ / 2111 // CmpP 2112 // Bool ne 2113 // If 2114 // / \ 2115 // True False-<2> 2116 // / | 2117 // / cast_not_null 2118 // Load | | ^ 2119 // [fast_test] | | 2120 // gvn to opt_test | | 2121 // / \ | <1> 2122 // True False | 2123 // | \\ | 2124 // [slow_call] \[fast_result] 2125 // Ctl Val \ \ 2126 // | \ \ 2127 // Catch <1> \ \ 2128 // / \ ^ \ \ 2129 // Ex No_Ex | \ \ 2130 // | \ \ | \ <2> \ 2131 // ... \ [slow_res] | | \ [null_result] 2132 // \ \--+--+--- | | 2133 // \ | / \ | / 2134 // --------Region Phi 2135 // 2136 //============================================================================= 2137 // Code is structured as a series of driver functions all called 'do_XXX' that 2138 // call a set of helper functions. Helper functions first, then drivers. 2139 2140 //------------------------------null_check_oop--------------------------------- 2141 // Null check oop. Set null-path control into Region in slot 3. 2142 // Make a cast-not-nullness use the other not-null control. Return cast. 2143 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2144 bool never_see_null) { 2145 // Initial NULL check taken path 2146 (*null_control) = top(); 2147 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2148 2149 // Generate uncommon_trap: 2150 if (never_see_null && (*null_control) != top()) { 2151 // If we see an unexpected null at a check-cast we record it and force a 2152 // recompile; the offending check-cast will be compiled to handle NULLs. 2153 // If we see more than one offending BCI, then all checkcasts in the 2154 // method will be compiled to handle NULLs. 2155 PreserveJVMState pjvms(this); 2156 set_control(*null_control); 2157 replace_in_map(value, null()); 2158 uncommon_trap(Deoptimization::Reason_null_check, 2159 Deoptimization::Action_make_not_entrant); 2160 (*null_control) = top(); // NULL path is dead 2161 } 2162 2163 // Cast away null-ness on the result 2164 return cast; 2165 } 2166 2167 //------------------------------opt_iff---------------------------------------- 2168 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2169 // Return slow-path control. 2170 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2171 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2172 2173 // Fast path taken; set region slot 2 2174 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2175 region->init_req(2,fast_taken); // Capture fast-control 2176 2177 // Fast path not-taken, i.e. slow path 2178 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2179 return slow_taken; 2180 } 2181 2182 //-----------------------------make_runtime_call------------------------------- 2183 Node* GraphKit::make_runtime_call(int flags, 2184 const TypeFunc* call_type, address call_addr, 2185 const char* call_name, 2186 const TypePtr* adr_type, 2187 // The following parms are all optional. 2188 // The first NULL ends the list. 2189 Node* parm0, Node* parm1, 2190 Node* parm2, Node* parm3, 2191 Node* parm4, Node* parm5, 2192 Node* parm6, Node* parm7) { 2193 // Slow-path call 2194 bool is_leaf = !(flags & RC_NO_LEAF); 2195 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2196 if (call_name == NULL) { 2197 assert(!is_leaf, "must supply name for leaf"); 2198 call_name = OptoRuntime::stub_name(call_addr); 2199 } 2200 CallNode* call; 2201 if (!is_leaf) { 2202 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2203 bci(), adr_type); 2204 } else if (flags & RC_NO_FP) { 2205 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2206 } else { 2207 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2208 } 2209 2210 // The following is similar to set_edges_for_java_call, 2211 // except that the memory effects of the call are restricted to AliasIdxRaw. 2212 2213 // Slow path call has no side-effects, uses few values 2214 bool wide_in = !(flags & RC_NARROW_MEM); 2215 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2216 2217 Node* prev_mem = NULL; 2218 if (wide_in) { 2219 prev_mem = set_predefined_input_for_runtime_call(call); 2220 } else { 2221 assert(!wide_out, "narrow in => narrow out"); 2222 Node* narrow_mem = memory(adr_type); 2223 prev_mem = reset_memory(); 2224 map()->set_memory(narrow_mem); 2225 set_predefined_input_for_runtime_call(call); 2226 } 2227 2228 // Hook each parm in order. Stop looking at the first NULL. 2229 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2230 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2231 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2232 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2233 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2234 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2235 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2236 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2237 /* close each nested if ===> */ } } } } } } } } 2238 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2239 2240 if (!is_leaf) { 2241 // Non-leaves can block and take safepoints: 2242 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2243 } 2244 // Non-leaves can throw exceptions: 2245 if (has_io) { 2246 call->set_req(TypeFunc::I_O, i_o()); 2247 } 2248 2249 if (flags & RC_UNCOMMON) { 2250 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2251 // (An "if" probability corresponds roughly to an unconditional count. 2252 // Sort of.) 2253 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2254 } 2255 2256 Node* c = _gvn.transform(call); 2257 assert(c == call, "cannot disappear"); 2258 2259 if (wide_out) { 2260 // Slow path call has full side-effects. 2261 set_predefined_output_for_runtime_call(call); 2262 } else { 2263 // Slow path call has few side-effects, and/or sets few values. 2264 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2265 } 2266 2267 if (has_io) { 2268 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2269 } 2270 return call; 2271 2272 } 2273 2274 //------------------------------merge_memory----------------------------------- 2275 // Merge memory from one path into the current memory state. 2276 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2277 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2278 Node* old_slice = mms.force_memory(); 2279 Node* new_slice = mms.memory2(); 2280 if (old_slice != new_slice) { 2281 PhiNode* phi; 2282 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2283 phi = new_slice->as_Phi(); 2284 #ifdef ASSERT 2285 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2286 old_slice = old_slice->in(new_path); 2287 // Caller is responsible for ensuring that any pre-existing 2288 // phis are already aware of old memory. 2289 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2290 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2291 #endif 2292 mms.set_memory(phi); 2293 } else { 2294 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2295 _gvn.set_type(phi, Type::MEMORY); 2296 phi->set_req(new_path, new_slice); 2297 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2298 } 2299 } 2300 } 2301 } 2302 2303 //------------------------------make_slow_call_ex------------------------------ 2304 // Make the exception handler hookups for the slow call 2305 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2306 if (stopped()) return; 2307 2308 // Make a catch node with just two handlers: fall-through and catch-all 2309 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2310 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2311 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2312 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2313 2314 { PreserveJVMState pjvms(this); 2315 set_control(excp); 2316 set_i_o(i_o); 2317 2318 if (excp != top()) { 2319 // Create an exception state also. 2320 // Use an exact type if the caller has specified a specific exception. 2321 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2322 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2323 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2324 } 2325 } 2326 2327 // Get the no-exception control from the CatchNode. 2328 set_control(norm); 2329 } 2330 2331 2332 //-------------------------------gen_subtype_check----------------------------- 2333 // Generate a subtyping check. Takes as input the subtype and supertype. 2334 // Returns 2 values: sets the default control() to the true path and returns 2335 // the false path. Only reads invariant memory; sets no (visible) memory. 2336 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2337 // but that's not exposed to the optimizer. This call also doesn't take in an 2338 // Object; if you wish to check an Object you need to load the Object's class 2339 // prior to coming here. 2340 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2341 // Fast check for identical types, perhaps identical constants. 2342 // The types can even be identical non-constants, in cases 2343 // involving Array.newInstance, Object.clone, etc. 2344 if (subklass == superklass) 2345 return top(); // false path is dead; no test needed. 2346 2347 if (_gvn.type(superklass)->singleton()) { 2348 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2349 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2350 2351 // In the common case of an exact superklass, try to fold up the 2352 // test before generating code. You may ask, why not just generate 2353 // the code and then let it fold up? The answer is that the generated 2354 // code will necessarily include null checks, which do not always 2355 // completely fold away. If they are also needless, then they turn 2356 // into a performance loss. Example: 2357 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2358 // Here, the type of 'fa' is often exact, so the store check 2359 // of fa[1]=x will fold up, without testing the nullness of x. 2360 switch (static_subtype_check(superk, subk)) { 2361 case SSC_always_false: 2362 { 2363 Node* always_fail = control(); 2364 set_control(top()); 2365 return always_fail; 2366 } 2367 case SSC_always_true: 2368 return top(); 2369 case SSC_easy_test: 2370 { 2371 // Just do a direct pointer compare and be done. 2372 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2373 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2374 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2375 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2376 return _gvn.transform( new(C) IfFalseNode(iff) ); 2377 } 2378 case SSC_full_test: 2379 break; 2380 default: 2381 ShouldNotReachHere(); 2382 } 2383 } 2384 2385 // %%% Possible further optimization: Even if the superklass is not exact, 2386 // if the subklass is the unique subtype of the superklass, the check 2387 // will always succeed. We could leave a dependency behind to ensure this. 2388 2389 // First load the super-klass's check-offset 2390 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2391 Node *chk_off = _gvn.transform( new (C) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2392 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2393 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2394 2395 // Load from the sub-klass's super-class display list, or a 1-word cache of 2396 // the secondary superclass list, or a failing value with a sentinel offset 2397 // if the super-klass is an interface or exceptionally deep in the Java 2398 // hierarchy and we have to scan the secondary superclass list the hard way. 2399 // Worst-case type is a little odd: NULL is allowed as a result (usually 2400 // klass loads can never produce a NULL). 2401 Node *chk_off_X = ConvI2X(chk_off); 2402 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2403 // For some types like interfaces the following loadKlass is from a 1-word 2404 // cache which is mutable so can't use immutable memory. Other 2405 // types load from the super-class display table which is immutable. 2406 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2407 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2408 2409 // Compile speed common case: ARE a subtype and we canNOT fail 2410 if( superklass == nkls ) 2411 return top(); // false path is dead; no test needed. 2412 2413 // See if we get an immediate positive hit. Happens roughly 83% of the 2414 // time. Test to see if the value loaded just previously from the subklass 2415 // is exactly the superklass. 2416 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2417 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2418 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2419 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2420 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2421 2422 // Compile speed common case: Check for being deterministic right now. If 2423 // chk_off is a constant and not equal to cacheoff then we are NOT a 2424 // subklass. In this case we need exactly the 1 test above and we can 2425 // return those results immediately. 2426 if (!might_be_cache) { 2427 Node* not_subtype_ctrl = control(); 2428 set_control(iftrue1); // We need exactly the 1 test above 2429 return not_subtype_ctrl; 2430 } 2431 2432 // Gather the various success & failures here 2433 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2434 record_for_igvn(r_ok_subtype); 2435 RegionNode *r_not_subtype = new (C) RegionNode(3); 2436 record_for_igvn(r_not_subtype); 2437 2438 r_ok_subtype->init_req(1, iftrue1); 2439 2440 // Check for immediate negative hit. Happens roughly 11% of the time (which 2441 // is roughly 63% of the remaining cases). Test to see if the loaded 2442 // check-offset points into the subklass display list or the 1-element 2443 // cache. If it points to the display (and NOT the cache) and the display 2444 // missed then it's not a subtype. 2445 Node *cacheoff = _gvn.intcon(cacheoff_con); 2446 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2447 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2448 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2449 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2450 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2451 2452 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2453 // No performance impact (too rare) but allows sharing of secondary arrays 2454 // which has some footprint reduction. 2455 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2456 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2457 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2458 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2459 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2460 2461 // -- Roads not taken here: -- 2462 // We could also have chosen to perform the self-check at the beginning 2463 // of this code sequence, as the assembler does. This would not pay off 2464 // the same way, since the optimizer, unlike the assembler, can perform 2465 // static type analysis to fold away many successful self-checks. 2466 // Non-foldable self checks work better here in second position, because 2467 // the initial primary superclass check subsumes a self-check for most 2468 // types. An exception would be a secondary type like array-of-interface, 2469 // which does not appear in its own primary supertype display. 2470 // Finally, we could have chosen to move the self-check into the 2471 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2472 // dependent manner. But it is worthwhile to have the check here, 2473 // where it can be perhaps be optimized. The cost in code space is 2474 // small (register compare, branch). 2475 2476 // Now do a linear scan of the secondary super-klass array. Again, no real 2477 // performance impact (too rare) but it's gotta be done. 2478 // Since the code is rarely used, there is no penalty for moving it 2479 // out of line, and it can only improve I-cache density. 2480 // The decision to inline or out-of-line this final check is platform 2481 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2482 Node* psc = _gvn.transform( 2483 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2484 2485 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2486 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2487 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2488 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2489 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2490 2491 // Return false path; set default control to true path. 2492 set_control( _gvn.transform(r_ok_subtype) ); 2493 return _gvn.transform(r_not_subtype); 2494 } 2495 2496 //----------------------------static_subtype_check----------------------------- 2497 // Shortcut important common cases when superklass is exact: 2498 // (0) superklass is java.lang.Object (can occur in reflective code) 2499 // (1) subklass is already limited to a subtype of superklass => always ok 2500 // (2) subklass does not overlap with superklass => always fail 2501 // (3) superklass has NO subtypes and we can check with a simple compare. 2502 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2503 if (StressReflectiveCode) { 2504 return SSC_full_test; // Let caller generate the general case. 2505 } 2506 2507 if (superk == env()->Object_klass()) { 2508 return SSC_always_true; // (0) this test cannot fail 2509 } 2510 2511 ciType* superelem = superk; 2512 if (superelem->is_array_klass()) 2513 superelem = superelem->as_array_klass()->base_element_type(); 2514 2515 if (!subk->is_interface()) { // cannot trust static interface types yet 2516 if (subk->is_subtype_of(superk)) { 2517 return SSC_always_true; // (1) false path dead; no dynamic test needed 2518 } 2519 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2520 !superk->is_subtype_of(subk)) { 2521 return SSC_always_false; 2522 } 2523 } 2524 2525 // If casting to an instance klass, it must have no subtypes 2526 if (superk->is_interface()) { 2527 // Cannot trust interfaces yet. 2528 // %%% S.B. superk->nof_implementors() == 1 2529 } else if (superelem->is_instance_klass()) { 2530 ciInstanceKlass* ik = superelem->as_instance_klass(); 2531 if (!ik->has_subklass() && !ik->is_interface()) { 2532 if (!ik->is_final()) { 2533 // Add a dependency if there is a chance of a later subclass. 2534 C->dependencies()->assert_leaf_type(ik); 2535 } 2536 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2537 } 2538 } else { 2539 // A primitive array type has no subtypes. 2540 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2541 } 2542 2543 return SSC_full_test; 2544 } 2545 2546 // Profile-driven exact type check: 2547 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2548 float prob, 2549 Node* *casted_receiver) { 2550 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2551 Node* recv_klass = load_object_klass(receiver); 2552 Node* want_klass = makecon(tklass); 2553 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2554 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2555 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2556 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2557 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2558 2559 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2560 assert(recv_xtype->klass_is_exact(), ""); 2561 2562 // Subsume downstream occurrences of receiver with a cast to 2563 // recv_xtype, since now we know what the type will be. 2564 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2565 (*casted_receiver) = _gvn.transform(cast); 2566 // (User must make the replace_in_map call.) 2567 2568 return fail; 2569 } 2570 2571 2572 //------------------------------seems_never_null------------------------------- 2573 // Use null_seen information if it is available from the profile. 2574 // If we see an unexpected null at a type check we record it and force a 2575 // recompile; the offending check will be recompiled to handle NULLs. 2576 // If we see several offending BCIs, then all checks in the 2577 // method will be recompiled. 2578 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2579 if (UncommonNullCast // Cutout for this technique 2580 && obj != null() // And not the -Xcomp stupid case? 2581 && !too_many_traps(Deoptimization::Reason_null_check) 2582 ) { 2583 if (data == NULL) 2584 // Edge case: no mature data. Be optimistic here. 2585 return true; 2586 // If the profile has not seen a null, assume it won't happen. 2587 assert(java_bc() == Bytecodes::_checkcast || 2588 java_bc() == Bytecodes::_instanceof || 2589 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2590 return !data->as_BitData()->null_seen(); 2591 } 2592 return false; 2593 } 2594 2595 //------------------------maybe_cast_profiled_receiver------------------------- 2596 // If the profile has seen exactly one type, narrow to exactly that type. 2597 // Subsequent type checks will always fold up. 2598 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2599 ciProfileData* data, 2600 ciKlass* require_klass) { 2601 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2602 if (data == NULL) return NULL; 2603 2604 // Make sure we haven't already deoptimized from this tactic. 2605 if (too_many_traps(Deoptimization::Reason_class_check)) 2606 return NULL; 2607 2608 // (No, this isn't a call, but it's enough like a virtual call 2609 // to use the same ciMethod accessor to get the profile info...) 2610 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2611 if (profile.count() >= 0 && // no cast failures here 2612 profile.has_receiver(0) && 2613 profile.morphism() == 1) { 2614 ciKlass* exact_kls = profile.receiver(0); 2615 if (require_klass == NULL || 2616 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2617 // If we narrow the type to match what the type profile sees, 2618 // we can then remove the rest of the cast. 2619 // This is a win, even if the exact_kls is very specific, 2620 // because downstream operations, such as method calls, 2621 // will often benefit from the sharper type. 2622 Node* exact_obj = not_null_obj; // will get updated in place... 2623 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2624 &exact_obj); 2625 { PreserveJVMState pjvms(this); 2626 set_control(slow_ctl); 2627 uncommon_trap(Deoptimization::Reason_class_check, 2628 Deoptimization::Action_maybe_recompile); 2629 } 2630 replace_in_map(not_null_obj, exact_obj); 2631 return exact_obj; 2632 } 2633 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2634 } 2635 2636 return NULL; 2637 } 2638 2639 2640 //-------------------------------gen_instanceof-------------------------------- 2641 // Generate an instance-of idiom. Used by both the instance-of bytecode 2642 // and the reflective instance-of call. 2643 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) { 2644 kill_dead_locals(); // Benefit all the uncommon traps 2645 assert( !stopped(), "dead parse path should be checked in callers" ); 2646 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2647 "must check for not-null not-dead klass in callers"); 2648 2649 // Make the merge point 2650 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2651 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2652 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2653 C->set_has_split_ifs(true); // Has chance for split-if optimization 2654 2655 ciProfileData* data = NULL; 2656 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2657 data = method()->method_data()->bci_to_data(bci()); 2658 } 2659 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2660 && seems_never_null(obj, data)); 2661 2662 // Null check; get casted pointer; set region slot 3 2663 Node* null_ctl = top(); 2664 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2665 2666 // If not_null_obj is dead, only null-path is taken 2667 if (stopped()) { // Doing instance-of on a NULL? 2668 set_control(null_ctl); 2669 return intcon(0); 2670 } 2671 region->init_req(_null_path, null_ctl); 2672 phi ->init_req(_null_path, intcon(0)); // Set null path value 2673 if (null_ctl == top()) { 2674 // Do this eagerly, so that pattern matches like is_diamond_phi 2675 // will work even during parsing. 2676 assert(_null_path == PATH_LIMIT-1, "delete last"); 2677 region->del_req(_null_path); 2678 phi ->del_req(_null_path); 2679 } 2680 2681 if (ProfileDynamicTypes && data != NULL) { 2682 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL); 2683 if (stopped()) { // Profile disagrees with this path. 2684 set_control(null_ctl); // Null is the only remaining possibility. 2685 return intcon(0); 2686 } 2687 if (cast_obj != NULL) 2688 not_null_obj = cast_obj; 2689 } 2690 2691 // Load the object's klass 2692 Node* obj_klass = load_object_klass(not_null_obj); 2693 2694 // Generate the subtype check 2695 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2696 2697 // Plug in the success path to the general merge in slot 1. 2698 region->init_req(_obj_path, control()); 2699 phi ->init_req(_obj_path, intcon(1)); 2700 2701 // Plug in the failing path to the general merge in slot 2. 2702 region->init_req(_fail_path, not_subtype_ctrl); 2703 phi ->init_req(_fail_path, intcon(0)); 2704 2705 // Return final merged results 2706 set_control( _gvn.transform(region) ); 2707 record_for_igvn(region); 2708 return _gvn.transform(phi); 2709 } 2710 2711 //-------------------------------gen_checkcast--------------------------------- 2712 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2713 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2714 // uncommon-trap paths work. Adjust stack after this call. 2715 // If failure_control is supplied and not null, it is filled in with 2716 // the control edge for the cast failure. Otherwise, an appropriate 2717 // uncommon trap or exception is thrown. 2718 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2719 Node* *failure_control) { 2720 kill_dead_locals(); // Benefit all the uncommon traps 2721 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2722 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2723 2724 // Fast cutout: Check the case that the cast is vacuously true. 2725 // This detects the common cases where the test will short-circuit 2726 // away completely. We do this before we perform the null check, 2727 // because if the test is going to turn into zero code, we don't 2728 // want a residual null check left around. (Causes a slowdown, 2729 // for example, in some objArray manipulations, such as a[i]=a[j].) 2730 if (tk->singleton()) { 2731 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2732 if (objtp != NULL && objtp->klass() != NULL) { 2733 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2734 case SSC_always_true: 2735 return obj; 2736 case SSC_always_false: 2737 // It needs a null check because a null will *pass* the cast check. 2738 // A non-null value will always produce an exception. 2739 return null_assert(obj); 2740 } 2741 } 2742 } 2743 2744 ciProfileData* data = NULL; 2745 if (failure_control == NULL) { // use MDO in regular case only 2746 assert(java_bc() == Bytecodes::_aastore || 2747 java_bc() == Bytecodes::_checkcast, 2748 "interpreter profiles type checks only for these BCs"); 2749 data = method()->method_data()->bci_to_data(bci()); 2750 } 2751 2752 // Make the merge point 2753 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2754 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 2755 Node* phi = new (C) PhiNode(region, toop); 2756 C->set_has_split_ifs(true); // Has chance for split-if optimization 2757 2758 // Use null-cast information if it is available 2759 bool never_see_null = ((failure_control == NULL) // regular case only 2760 && seems_never_null(obj, data)); 2761 2762 // Null check; get casted pointer; set region slot 3 2763 Node* null_ctl = top(); 2764 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2765 2766 // If not_null_obj is dead, only null-path is taken 2767 if (stopped()) { // Doing instance-of on a NULL? 2768 set_control(null_ctl); 2769 return null(); 2770 } 2771 region->init_req(_null_path, null_ctl); 2772 phi ->init_req(_null_path, null()); // Set null path value 2773 if (null_ctl == top()) { 2774 // Do this eagerly, so that pattern matches like is_diamond_phi 2775 // will work even during parsing. 2776 assert(_null_path == PATH_LIMIT-1, "delete last"); 2777 region->del_req(_null_path); 2778 phi ->del_req(_null_path); 2779 } 2780 2781 Node* cast_obj = NULL; 2782 if (tk->klass_is_exact()) { 2783 // The following optimization tries to statically cast the speculative type of the object 2784 // (for example obtained during profiling) to the type of the superklass and then do a 2785 // dynamic check that the type of the object is what we expect. To work correctly 2786 // for checkcast and aastore the type of superklass should be exact. 2787 if (data != NULL && 2788 // Counter has never been decremented (due to cast failure). 2789 // ...This is a reasonable thing to expect. It is true of 2790 // all casts inserted by javac to implement generic types. 2791 data->as_CounterData()->count() >= 0) { 2792 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass()); 2793 if (cast_obj != NULL) { 2794 if (failure_control != NULL) // failure is now impossible 2795 (*failure_control) = top(); 2796 // adjust the type of the phi to the exact klass: 2797 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR)); 2798 } 2799 } 2800 } 2801 2802 if (cast_obj == NULL) { 2803 // Load the object's klass 2804 Node* obj_klass = load_object_klass(not_null_obj); 2805 2806 // Generate the subtype check 2807 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2808 2809 // Plug in success path into the merge 2810 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 2811 not_null_obj, toop)); 2812 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2813 if (failure_control == NULL) { 2814 if (not_subtype_ctrl != top()) { // If failure is possible 2815 PreserveJVMState pjvms(this); 2816 set_control(not_subtype_ctrl); 2817 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2818 } 2819 } else { 2820 (*failure_control) = not_subtype_ctrl; 2821 } 2822 } 2823 2824 region->init_req(_obj_path, control()); 2825 phi ->init_req(_obj_path, cast_obj); 2826 2827 // A merge of NULL or Casted-NotNull obj 2828 Node* res = _gvn.transform(phi); 2829 2830 // Note I do NOT always 'replace_in_map(obj,result)' here. 2831 // if( tk->klass()->can_be_primary_super() ) 2832 // This means that if I successfully store an Object into an array-of-String 2833 // I 'forget' that the Object is really now known to be a String. I have to 2834 // do this because we don't have true union types for interfaces - if I store 2835 // a Baz into an array-of-Interface and then tell the optimizer it's an 2836 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2837 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2838 // replace_in_map( obj, res ); 2839 2840 // Return final merged results 2841 set_control( _gvn.transform(region) ); 2842 record_for_igvn(region); 2843 return res; 2844 } 2845 2846 //------------------------------next_monitor----------------------------------- 2847 // What number should be given to the next monitor? 2848 int GraphKit::next_monitor() { 2849 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2850 int next = current + C->sync_stack_slots(); 2851 // Keep the toplevel high water mark current: 2852 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2853 return current; 2854 } 2855 2856 //------------------------------insert_mem_bar--------------------------------- 2857 // Memory barrier to avoid floating things around 2858 // The membar serves as a pinch point between both control and all memory slices. 2859 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2860 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2861 mb->init_req(TypeFunc::Control, control()); 2862 mb->init_req(TypeFunc::Memory, reset_memory()); 2863 Node* membar = _gvn.transform(mb); 2864 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 2865 set_all_memory_call(membar); 2866 return membar; 2867 } 2868 2869 //-------------------------insert_mem_bar_volatile---------------------------- 2870 // Memory barrier to avoid floating things around 2871 // The membar serves as a pinch point between both control and memory(alias_idx). 2872 // If you want to make a pinch point on all memory slices, do not use this 2873 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2874 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2875 // When Parse::do_put_xxx updates a volatile field, it appends a series 2876 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2877 // The first membar is on the same memory slice as the field store opcode. 2878 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2879 // All the other membars (for other volatile slices, including AliasIdxBot, 2880 // which stands for all unknown volatile slices) are control-dependent 2881 // on the first membar. This prevents later volatile loads or stores 2882 // from sliding up past the just-emitted store. 2883 2884 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2885 mb->set_req(TypeFunc::Control,control()); 2886 if (alias_idx == Compile::AliasIdxBot) { 2887 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2888 } else { 2889 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2890 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2891 } 2892 Node* membar = _gvn.transform(mb); 2893 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 2894 if (alias_idx == Compile::AliasIdxBot) { 2895 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 2896 } else { 2897 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2898 } 2899 return membar; 2900 } 2901 2902 //------------------------------shared_lock------------------------------------ 2903 // Emit locking code. 2904 FastLockNode* GraphKit::shared_lock(Node* obj) { 2905 // bci is either a monitorenter bc or InvocationEntryBci 2906 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2907 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2908 2909 if( !GenerateSynchronizationCode ) 2910 return NULL; // Not locking things? 2911 if (stopped()) // Dead monitor? 2912 return NULL; 2913 2914 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2915 2916 // Box the stack location 2917 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 2918 Node* mem = reset_memory(); 2919 2920 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 2921 if (PrintPreciseBiasedLockingStatistics) { 2922 // Create the counters for this fast lock. 2923 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2924 } 2925 // Add monitor to debug info for the slow path. If we block inside the 2926 // slow path and de-opt, we need the monitor hanging around 2927 map()->push_monitor( flock ); 2928 2929 const TypeFunc *tf = LockNode::lock_type(); 2930 LockNode *lock = new (C) LockNode(C, tf); 2931 2932 lock->init_req( TypeFunc::Control, control() ); 2933 lock->init_req( TypeFunc::Memory , mem ); 2934 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2935 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2936 lock->init_req( TypeFunc::ReturnAdr, top() ); 2937 2938 lock->init_req(TypeFunc::Parms + 0, obj); 2939 lock->init_req(TypeFunc::Parms + 1, box); 2940 lock->init_req(TypeFunc::Parms + 2, flock); 2941 add_safepoint_edges(lock); 2942 2943 lock = _gvn.transform( lock )->as_Lock(); 2944 2945 // lock has no side-effects, sets few values 2946 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2947 2948 insert_mem_bar(Op_MemBarAcquireLock); 2949 2950 // Add this to the worklist so that the lock can be eliminated 2951 record_for_igvn(lock); 2952 2953 #ifndef PRODUCT 2954 if (PrintLockStatistics) { 2955 // Update the counter for this lock. Don't bother using an atomic 2956 // operation since we don't require absolute accuracy. 2957 lock->create_lock_counter(map()->jvms()); 2958 increment_counter(lock->counter()->addr()); 2959 } 2960 #endif 2961 2962 return flock; 2963 } 2964 2965 2966 //------------------------------shared_unlock---------------------------------- 2967 // Emit unlocking code. 2968 void GraphKit::shared_unlock(Node* box, Node* obj) { 2969 // bci is either a monitorenter bc or InvocationEntryBci 2970 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2971 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2972 2973 if( !GenerateSynchronizationCode ) 2974 return; 2975 if (stopped()) { // Dead monitor? 2976 map()->pop_monitor(); // Kill monitor from debug info 2977 return; 2978 } 2979 2980 // Memory barrier to avoid floating things down past the locked region 2981 insert_mem_bar(Op_MemBarReleaseLock); 2982 2983 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2984 UnlockNode *unlock = new (C) UnlockNode(C, tf); 2985 uint raw_idx = Compile::AliasIdxRaw; 2986 unlock->init_req( TypeFunc::Control, control() ); 2987 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2988 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2989 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2990 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2991 2992 unlock->init_req(TypeFunc::Parms + 0, obj); 2993 unlock->init_req(TypeFunc::Parms + 1, box); 2994 unlock = _gvn.transform(unlock)->as_Unlock(); 2995 2996 Node* mem = reset_memory(); 2997 2998 // unlock has no side-effects, sets few values 2999 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3000 3001 // Kill monitor from debug info 3002 map()->pop_monitor( ); 3003 } 3004 3005 //-------------------------------get_layout_helper----------------------------- 3006 // If the given klass is a constant or known to be an array, 3007 // fetch the constant layout helper value into constant_value 3008 // and return (Node*)NULL. Otherwise, load the non-constant 3009 // layout helper value, and return the node which represents it. 3010 // This two-faced routine is useful because allocation sites 3011 // almost always feature constant types. 3012 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3013 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3014 if (!StressReflectiveCode && inst_klass != NULL) { 3015 ciKlass* klass = inst_klass->klass(); 3016 bool xklass = inst_klass->klass_is_exact(); 3017 if (xklass || klass->is_array_klass()) { 3018 jint lhelper = klass->layout_helper(); 3019 if (lhelper != Klass::_lh_neutral_value) { 3020 constant_value = lhelper; 3021 return (Node*) NULL; 3022 } 3023 } 3024 } 3025 constant_value = Klass::_lh_neutral_value; // put in a known value 3026 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3027 return make_load(NULL, lhp, TypeInt::INT, T_INT); 3028 } 3029 3030 // We just put in an allocate/initialize with a big raw-memory effect. 3031 // Hook selected additional alias categories on the initialization. 3032 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3033 MergeMemNode* init_in_merge, 3034 Node* init_out_raw) { 3035 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3036 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3037 3038 Node* prevmem = kit.memory(alias_idx); 3039 init_in_merge->set_memory_at(alias_idx, prevmem); 3040 kit.set_memory(init_out_raw, alias_idx); 3041 } 3042 3043 //---------------------------set_output_for_allocation------------------------- 3044 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3045 const TypeOopPtr* oop_type) { 3046 int rawidx = Compile::AliasIdxRaw; 3047 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3048 add_safepoint_edges(alloc); 3049 Node* allocx = _gvn.transform(alloc); 3050 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3051 // create memory projection for i_o 3052 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3053 make_slow_call_ex(allocx, env()->Throwable_klass(), true); 3054 3055 // create a memory projection as for the normal control path 3056 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3057 set_memory(malloc, rawidx); 3058 3059 // a normal slow-call doesn't change i_o, but an allocation does 3060 // we create a separate i_o projection for the normal control path 3061 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3062 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3063 3064 // put in an initialization barrier 3065 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3066 rawoop)->as_Initialize(); 3067 assert(alloc->initialization() == init, "2-way macro link must work"); 3068 assert(init ->allocation() == alloc, "2-way macro link must work"); 3069 { 3070 // Extract memory strands which may participate in the new object's 3071 // initialization, and source them from the new InitializeNode. 3072 // This will allow us to observe initializations when they occur, 3073 // and link them properly (as a group) to the InitializeNode. 3074 assert(init->in(InitializeNode::Memory) == malloc, ""); 3075 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3076 init->set_req(InitializeNode::Memory, minit_in); 3077 record_for_igvn(minit_in); // fold it up later, if possible 3078 Node* minit_out = memory(rawidx); 3079 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3080 if (oop_type->isa_aryptr()) { 3081 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3082 int elemidx = C->get_alias_index(telemref); 3083 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3084 } else if (oop_type->isa_instptr()) { 3085 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3086 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3087 ciField* field = ik->nonstatic_field_at(i); 3088 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3089 continue; // do not bother to track really large numbers of fields 3090 // Find (or create) the alias category for this field: 3091 int fieldidx = C->alias_type(field)->index(); 3092 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3093 } 3094 } 3095 } 3096 3097 // Cast raw oop to the real thing... 3098 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3099 javaoop = _gvn.transform(javaoop); 3100 C->set_recent_alloc(control(), javaoop); 3101 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3102 3103 #ifdef ASSERT 3104 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3105 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3106 "Ideal_allocation works"); 3107 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3108 "Ideal_allocation works"); 3109 if (alloc->is_AllocateArray()) { 3110 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3111 "Ideal_allocation works"); 3112 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3113 "Ideal_allocation works"); 3114 } else { 3115 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3116 } 3117 } 3118 #endif //ASSERT 3119 3120 return javaoop; 3121 } 3122 3123 //---------------------------new_instance-------------------------------------- 3124 // This routine takes a klass_node which may be constant (for a static type) 3125 // or may be non-constant (for reflective code). It will work equally well 3126 // for either, and the graph will fold nicely if the optimizer later reduces 3127 // the type to a constant. 3128 // The optional arguments are for specialized use by intrinsics: 3129 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3130 // - If 'return_size_val', report the the total object size to the caller. 3131 Node* GraphKit::new_instance(Node* klass_node, 3132 Node* extra_slow_test, 3133 Node* *return_size_val) { 3134 // Compute size in doublewords 3135 // The size is always an integral number of doublewords, represented 3136 // as a positive bytewise size stored in the klass's layout_helper. 3137 // The layout_helper also encodes (in a low bit) the need for a slow path. 3138 jint layout_con = Klass::_lh_neutral_value; 3139 Node* layout_val = get_layout_helper(klass_node, layout_con); 3140 int layout_is_con = (layout_val == NULL); 3141 3142 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3143 // Generate the initial go-slow test. It's either ALWAYS (return a 3144 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3145 // case) a computed value derived from the layout_helper. 3146 Node* initial_slow_test = NULL; 3147 if (layout_is_con) { 3148 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3149 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3150 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3151 3152 } else { // reflective case 3153 // This reflective path is used by Unsafe.allocateInstance. 3154 // (It may be stress-tested by specifying StressReflectiveCode.) 3155 // Basically, we want to get into the VM is there's an illegal argument. 3156 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3157 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3158 if (extra_slow_test != intcon(0)) { 3159 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3160 } 3161 // (Macro-expander will further convert this to a Bool, if necessary.) 3162 } 3163 3164 // Find the size in bytes. This is easy; it's the layout_helper. 3165 // The size value must be valid even if the slow path is taken. 3166 Node* size = NULL; 3167 if (layout_is_con) { 3168 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3169 } else { // reflective case 3170 // This reflective path is used by clone and Unsafe.allocateInstance. 3171 size = ConvI2X(layout_val); 3172 3173 // Clear the low bits to extract layout_helper_size_in_bytes: 3174 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3175 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3176 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3177 } 3178 if (return_size_val != NULL) { 3179 (*return_size_val) = size; 3180 } 3181 3182 // This is a precise notnull oop of the klass. 3183 // (Actually, it need not be precise if this is a reflective allocation.) 3184 // It's what we cast the result to. 3185 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3186 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3187 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3188 3189 // Now generate allocation code 3190 3191 // The entire memory state is needed for slow path of the allocation 3192 // since GC and deoptimization can happened. 3193 Node *mem = reset_memory(); 3194 set_all_memory(mem); // Create new memory state 3195 3196 AllocateNode* alloc 3197 = new (C) AllocateNode(C, AllocateNode::alloc_type(), 3198 control(), mem, i_o(), 3199 size, klass_node, 3200 initial_slow_test); 3201 3202 return set_output_for_allocation(alloc, oop_type); 3203 } 3204 3205 //-------------------------------new_array------------------------------------- 3206 // helper for both newarray and anewarray 3207 // The 'length' parameter is (obviously) the length of the array. 3208 // See comments on new_instance for the meaning of the other arguments. 3209 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3210 Node* length, // number of array elements 3211 int nargs, // number of arguments to push back for uncommon trap 3212 Node* *return_size_val) { 3213 jint layout_con = Klass::_lh_neutral_value; 3214 Node* layout_val = get_layout_helper(klass_node, layout_con); 3215 int layout_is_con = (layout_val == NULL); 3216 3217 if (!layout_is_con && !StressReflectiveCode && 3218 !too_many_traps(Deoptimization::Reason_class_check)) { 3219 // This is a reflective array creation site. 3220 // Optimistically assume that it is a subtype of Object[], 3221 // so that we can fold up all the address arithmetic. 3222 layout_con = Klass::array_layout_helper(T_OBJECT); 3223 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3224 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3225 { BuildCutout unless(this, bol_lh, PROB_MAX); 3226 inc_sp(nargs); 3227 uncommon_trap(Deoptimization::Reason_class_check, 3228 Deoptimization::Action_maybe_recompile); 3229 } 3230 layout_val = NULL; 3231 layout_is_con = true; 3232 } 3233 3234 // Generate the initial go-slow test. Make sure we do not overflow 3235 // if length is huge (near 2Gig) or negative! We do not need 3236 // exact double-words here, just a close approximation of needed 3237 // double-words. We can't add any offset or rounding bits, lest we 3238 // take a size -1 of bytes and make it positive. Use an unsigned 3239 // compare, so negative sizes look hugely positive. 3240 int fast_size_limit = FastAllocateSizeLimit; 3241 if (layout_is_con) { 3242 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3243 // Increase the size limit if we have exact knowledge of array type. 3244 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3245 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3246 } 3247 3248 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3249 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3250 if (initial_slow_test->is_Bool()) { 3251 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3252 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3253 } 3254 3255 // --- Size Computation --- 3256 // array_size = round_to_heap(array_header + (length << elem_shift)); 3257 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3258 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3259 // The rounding mask is strength-reduced, if possible. 3260 int round_mask = MinObjAlignmentInBytes - 1; 3261 Node* header_size = NULL; 3262 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3263 // (T_BYTE has the weakest alignment and size restrictions...) 3264 if (layout_is_con) { 3265 int hsize = Klass::layout_helper_header_size(layout_con); 3266 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3267 BasicType etype = Klass::layout_helper_element_type(layout_con); 3268 if ((round_mask & ~right_n_bits(eshift)) == 0) 3269 round_mask = 0; // strength-reduce it if it goes away completely 3270 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3271 assert(header_size_min <= hsize, "generic minimum is smallest"); 3272 header_size_min = hsize; 3273 header_size = intcon(hsize + round_mask); 3274 } else { 3275 Node* hss = intcon(Klass::_lh_header_size_shift); 3276 Node* hsm = intcon(Klass::_lh_header_size_mask); 3277 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3278 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3279 Node* mask = intcon(round_mask); 3280 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3281 } 3282 3283 Node* elem_shift = NULL; 3284 if (layout_is_con) { 3285 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3286 if (eshift != 0) 3287 elem_shift = intcon(eshift); 3288 } else { 3289 // There is no need to mask or shift this value. 3290 // The semantics of LShiftINode include an implicit mask to 0x1F. 3291 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3292 elem_shift = layout_val; 3293 } 3294 3295 // Transition to native address size for all offset calculations: 3296 Node* lengthx = ConvI2X(length); 3297 Node* headerx = ConvI2X(header_size); 3298 #ifdef _LP64 3299 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3300 if (tllen != NULL && tllen->_lo < 0) { 3301 // Add a manual constraint to a positive range. Cf. array_element_address. 3302 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3303 if (size_max > tllen->_hi) size_max = tllen->_hi; 3304 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3305 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon)); 3306 } 3307 } 3308 #endif 3309 3310 // Combine header size (plus rounding) and body size. Then round down. 3311 // This computation cannot overflow, because it is used only in two 3312 // places, one where the length is sharply limited, and the other 3313 // after a successful allocation. 3314 Node* abody = lengthx; 3315 if (elem_shift != NULL) 3316 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3317 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3318 if (round_mask != 0) { 3319 Node* mask = MakeConX(~round_mask); 3320 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3321 } 3322 // else if round_mask == 0, the size computation is self-rounding 3323 3324 if (return_size_val != NULL) { 3325 // This is the size 3326 (*return_size_val) = size; 3327 } 3328 3329 // Now generate allocation code 3330 3331 // The entire memory state is needed for slow path of the allocation 3332 // since GC and deoptimization can happened. 3333 Node *mem = reset_memory(); 3334 set_all_memory(mem); // Create new memory state 3335 3336 // Create the AllocateArrayNode and its result projections 3337 AllocateArrayNode* alloc 3338 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3339 control(), mem, i_o(), 3340 size, klass_node, 3341 initial_slow_test, 3342 length); 3343 3344 // Cast to correct type. Note that the klass_node may be constant or not, 3345 // and in the latter case the actual array type will be inexact also. 3346 // (This happens via a non-constant argument to inline_native_newArray.) 3347 // In any case, the value of klass_node provides the desired array type. 3348 const TypeInt* length_type = _gvn.find_int_type(length); 3349 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3350 if (ary_type->isa_aryptr() && length_type != NULL) { 3351 // Try to get a better type than POS for the size 3352 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3353 } 3354 3355 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3356 3357 // Cast length on remaining path to be as narrow as possible 3358 if (map()->find_edge(length) >= 0) { 3359 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3360 if (ccast != length) { 3361 _gvn.set_type_bottom(ccast); 3362 record_for_igvn(ccast); 3363 replace_in_map(length, ccast); 3364 } 3365 } 3366 3367 return javaoop; 3368 } 3369 3370 // The following "Ideal_foo" functions are placed here because they recognize 3371 // the graph shapes created by the functions immediately above. 3372 3373 //---------------------------Ideal_allocation---------------------------------- 3374 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3375 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3376 if (ptr == NULL) { // reduce dumb test in callers 3377 return NULL; 3378 } 3379 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3380 ptr = ptr->in(1); 3381 if (ptr == NULL) return NULL; 3382 } 3383 if (ptr->is_Proj()) { 3384 Node* allo = ptr->in(0); 3385 if (allo != NULL && allo->is_Allocate()) { 3386 return allo->as_Allocate(); 3387 } 3388 } 3389 // Report failure to match. 3390 return NULL; 3391 } 3392 3393 // Fancy version which also strips off an offset (and reports it to caller). 3394 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3395 intptr_t& offset) { 3396 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3397 if (base == NULL) return NULL; 3398 return Ideal_allocation(base, phase); 3399 } 3400 3401 // Trace Initialize <- Proj[Parm] <- Allocate 3402 AllocateNode* InitializeNode::allocation() { 3403 Node* rawoop = in(InitializeNode::RawAddress); 3404 if (rawoop->is_Proj()) { 3405 Node* alloc = rawoop->in(0); 3406 if (alloc->is_Allocate()) { 3407 return alloc->as_Allocate(); 3408 } 3409 } 3410 return NULL; 3411 } 3412 3413 // Trace Allocate -> Proj[Parm] -> Initialize 3414 InitializeNode* AllocateNode::initialization() { 3415 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3416 if (rawoop == NULL) return NULL; 3417 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3418 Node* init = rawoop->fast_out(i); 3419 if (init->is_Initialize()) { 3420 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3421 return init->as_Initialize(); 3422 } 3423 } 3424 return NULL; 3425 } 3426 3427 // Trace Allocate -> Proj[Parm] -> MemBarStoreStore 3428 MemBarStoreStoreNode* AllocateNode::storestore() { 3429 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3430 if (rawoop == NULL) return NULL; 3431 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3432 Node* storestore = rawoop->fast_out(i); 3433 if (storestore->is_MemBarStoreStore()) { 3434 return storestore->as_MemBarStoreStore(); 3435 } 3436 } 3437 return NULL; 3438 } 3439 3440 //----------------------------- loop predicates --------------------------- 3441 3442 //------------------------------add_predicate_impl---------------------------- 3443 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3444 // Too many traps seen? 3445 if (too_many_traps(reason)) { 3446 #ifdef ASSERT 3447 if (TraceLoopPredicate) { 3448 int tc = C->trap_count(reason); 3449 tty->print("too many traps=%s tcount=%d in ", 3450 Deoptimization::trap_reason_name(reason), tc); 3451 method()->print(); // which method has too many predicate traps 3452 tty->cr(); 3453 } 3454 #endif 3455 // We cannot afford to take more traps here, 3456 // do not generate predicate. 3457 return; 3458 } 3459 3460 Node *cont = _gvn.intcon(1); 3461 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3462 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3463 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3464 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3465 C->add_predicate_opaq(opq); 3466 { 3467 PreserveJVMState pjvms(this); 3468 set_control(iffalse); 3469 inc_sp(nargs); 3470 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3471 } 3472 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3473 set_control(iftrue); 3474 } 3475 3476 //------------------------------add_predicate--------------------------------- 3477 void GraphKit::add_predicate(int nargs) { 3478 if (UseLoopPredicate) { 3479 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3480 } 3481 // loop's limit check predicate should be near the loop. 3482 if (LoopLimitCheck) { 3483 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3484 } 3485 } 3486 3487 //----------------------------- store barriers ---------------------------- 3488 #define __ ideal. 3489 3490 void GraphKit::sync_kit(IdealKit& ideal) { 3491 set_all_memory(__ merged_memory()); 3492 set_i_o(__ i_o()); 3493 set_control(__ ctrl()); 3494 } 3495 3496 void GraphKit::final_sync(IdealKit& ideal) { 3497 // Final sync IdealKit and graphKit. 3498 sync_kit(ideal); 3499 } 3500 3501 // vanilla/CMS post barrier 3502 // Insert a write-barrier store. This is to let generational GC work; we have 3503 // to flag all oop-stores before the next GC point. 3504 void GraphKit::write_barrier_post(Node* oop_store, 3505 Node* obj, 3506 Node* adr, 3507 uint adr_idx, 3508 Node* val, 3509 bool use_precise) { 3510 // No store check needed if we're storing a NULL or an old object 3511 // (latter case is probably a string constant). The concurrent 3512 // mark sweep garbage collector, however, needs to have all nonNull 3513 // oop updates flagged via card-marks. 3514 if (val != NULL && val->is_Con()) { 3515 // must be either an oop or NULL 3516 const Type* t = val->bottom_type(); 3517 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3518 // stores of null never (?) need barriers 3519 return; 3520 ciObject* con = t->is_oopptr()->const_oop(); 3521 if (con != NULL 3522 && con->is_perm() 3523 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 3524 // no store barrier needed, because no old-to-new ref created 3525 return; 3526 } 3527 3528 if (use_ReduceInitialCardMarks() 3529 && obj == just_allocated_object(control())) { 3530 // We can skip marks on a freshly-allocated object in Eden. 3531 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3532 // That routine informs GC to take appropriate compensating steps, 3533 // upon a slow-path allocation, so as to make this card-mark 3534 // elision safe. 3535 return; 3536 } 3537 3538 if (!use_precise) { 3539 // All card marks for a (non-array) instance are in one place: 3540 adr = obj; 3541 } 3542 // (Else it's an array (or unknown), and we want more precise card marks.) 3543 assert(adr != NULL, ""); 3544 3545 IdealKit ideal(this, true); 3546 3547 // Convert the pointer to an int prior to doing math on it 3548 Node* cast = __ CastPX(__ ctrl(), adr); 3549 3550 // Divide by card size 3551 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3552 "Only one we handle so far."); 3553 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3554 3555 // Combine card table base and card offset 3556 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3557 3558 // Get the alias_index for raw card-mark memory 3559 int adr_type = Compile::AliasIdxRaw; 3560 Node* zero = __ ConI(0); // Dirty card value 3561 BasicType bt = T_BYTE; 3562 3563 if (UseCondCardMark) { 3564 // The classic GC reference write barrier is typically implemented 3565 // as a store into the global card mark table. Unfortunately 3566 // unconditional stores can result in false sharing and excessive 3567 // coherence traffic as well as false transactional aborts. 3568 // UseCondCardMark enables MP "polite" conditional card mark 3569 // stores. In theory we could relax the load from ctrl() to 3570 // no_ctrl, but that doesn't buy much latitude. 3571 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3572 __ if_then(card_val, BoolTest::ne, zero); 3573 } 3574 3575 // Smash zero into card 3576 if( !UseConcMarkSweepGC ) { 3577 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3578 } else { 3579 // Specialized path for CM store barrier 3580 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3581 } 3582 3583 if (UseCondCardMark) { 3584 __ end_if(); 3585 } 3586 3587 // Final sync IdealKit and GraphKit. 3588 final_sync(ideal); 3589 } 3590 3591 // G1 pre/post barriers 3592 void GraphKit::g1_write_barrier_pre(bool do_load, 3593 Node* obj, 3594 Node* adr, 3595 uint alias_idx, 3596 Node* val, 3597 const TypeOopPtr* val_type, 3598 Node* pre_val, 3599 BasicType bt) { 3600 3601 // Some sanity checks 3602 // Note: val is unused in this routine. 3603 3604 if (do_load) { 3605 // We need to generate the load of the previous value 3606 assert(obj != NULL, "must have a base"); 3607 assert(adr != NULL, "where are loading from?"); 3608 assert(pre_val == NULL, "loaded already?"); 3609 assert(val_type != NULL, "need a type"); 3610 } else { 3611 // In this case both val_type and alias_idx are unused. 3612 assert(pre_val != NULL, "must be loaded already"); 3613 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3614 } 3615 assert(bt == T_OBJECT, "or we shouldn't be here"); 3616 3617 IdealKit ideal(this, true); 3618 3619 Node* tls = __ thread(); // ThreadLocalStorage 3620 3621 Node* no_ctrl = NULL; 3622 Node* no_base = __ top(); 3623 Node* zero = __ ConI(0); 3624 Node* zeroX = __ ConX(0); 3625 3626 float likely = PROB_LIKELY(0.999); 3627 float unlikely = PROB_UNLIKELY(0.999); 3628 3629 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3630 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3631 3632 // Offsets into the thread 3633 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3634 PtrQueue::byte_offset_of_active()); 3635 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3636 PtrQueue::byte_offset_of_index()); 3637 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3638 PtrQueue::byte_offset_of_buf()); 3639 3640 // Now the actual pointers into the thread 3641 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3642 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3643 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3644 3645 // Now some of the values 3646 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3647 3648 // if (!marking) 3649 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3650 BasicType index_bt = TypeX_X->basic_type(); 3651 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3652 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3653 3654 if (do_load) { 3655 // load original value 3656 // alias_idx correct?? 3657 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3658 } 3659 3660 // if (pre_val != NULL) 3661 __ if_then(pre_val, BoolTest::ne, null()); { 3662 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3663 3664 // is the queue for this thread full? 3665 __ if_then(index, BoolTest::ne, zeroX, likely); { 3666 3667 // decrement the index 3668 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3669 3670 // Now get the buffer location we will log the previous value into and store it 3671 Node *log_addr = __ AddP(no_base, buffer, next_index); 3672 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, StoreNode::release); 3673 // update the index 3674 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw); 3675 3676 } __ else_(); { 3677 3678 // logging buffer is full, call the runtime 3679 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3680 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3681 } __ end_if(); // (!index) 3682 } __ end_if(); // (pre_val != NULL) 3683 } __ end_if(); // (!marking) 3684 3685 // Final sync IdealKit and GraphKit. 3686 final_sync(ideal); 3687 } 3688 3689 // 3690 // Update the card table and add card address to the queue 3691 // 3692 void GraphKit::g1_mark_card(IdealKit& ideal, 3693 Node* card_adr, 3694 Node* oop_store, 3695 uint oop_alias_idx, 3696 Node* index, 3697 Node* index_adr, 3698 Node* buffer, 3699 const TypeFunc* tf) { 3700 3701 Node* zero = __ ConI(0); 3702 Node* zeroX = __ ConX(0); 3703 Node* no_base = __ top(); 3704 BasicType card_bt = T_BYTE; 3705 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3706 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3707 3708 // Now do the queue work 3709 __ if_then(index, BoolTest::ne, zeroX); { 3710 3711 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3712 Node* log_addr = __ AddP(no_base, buffer, next_index); 3713 3714 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, StoreNode::release); 3715 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, StoreNode::release); 3716 3717 } __ else_(); { 3718 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3719 } __ end_if(); 3720 3721 } 3722 3723 void GraphKit::g1_write_barrier_post(Node* oop_store, 3724 Node* obj, 3725 Node* adr, 3726 uint alias_idx, 3727 Node* val, 3728 BasicType bt, 3729 bool use_precise) { 3730 // If we are writing a NULL then we need no post barrier 3731 3732 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3733 // Must be NULL 3734 const Type* t = val->bottom_type(); 3735 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3736 // No post barrier if writing NULLx 3737 return; 3738 } 3739 3740 if (!use_precise) { 3741 // All card marks for a (non-array) instance are in one place: 3742 adr = obj; 3743 } 3744 // (Else it's an array (or unknown), and we want more precise card marks.) 3745 assert(adr != NULL, ""); 3746 3747 IdealKit ideal(this, true); 3748 3749 Node* tls = __ thread(); // ThreadLocalStorage 3750 3751 Node* no_base = __ top(); 3752 float likely = PROB_LIKELY(0.999); 3753 float unlikely = PROB_UNLIKELY(0.999); 3754 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 3755 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 3756 Node* zeroX = __ ConX(0); 3757 3758 // Get the alias_index for raw card-mark memory 3759 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3760 3761 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3762 3763 // Offsets into the thread 3764 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3765 PtrQueue::byte_offset_of_index()); 3766 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3767 PtrQueue::byte_offset_of_buf()); 3768 3769 // Pointers into the thread 3770 3771 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3772 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3773 3774 // Now some values 3775 // Use ctrl to avoid hoisting these values past a safepoint, which could 3776 // potentially reset these fields in the JavaThread. 3777 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 3778 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3779 3780 // Convert the store obj pointer to an int prior to doing math on it 3781 // Must use ctrl to prevent "integerized oop" existing across safepoint 3782 Node* cast = __ CastPX(__ ctrl(), adr); 3783 3784 // Divide pointer by card size 3785 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3786 3787 // Combine card table base and card offset 3788 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3789 3790 // If we know the value being stored does it cross regions? 3791 3792 if (val != NULL) { 3793 // Does the store cause us to cross regions? 3794 3795 // Should be able to do an unsigned compare of region_size instead of 3796 // and extra shift. Do we have an unsigned compare?? 3797 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3798 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3799 3800 // if (xor_res == 0) same region so skip 3801 __ if_then(xor_res, BoolTest::ne, zeroX); { 3802 3803 // No barrier if we are storing a NULL 3804 __ if_then(val, BoolTest::ne, null(), unlikely); { 3805 3806 // Ok must mark the card if not already dirty 3807 3808 // load the original value of the card 3809 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3810 3811 __ if_then(card_val, BoolTest::ne, young_card); { 3812 sync_kit(ideal); 3813 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 3814 insert_mem_bar(Op_MemBarVolatile, oop_store); 3815 __ sync_kit(this); 3816 3817 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3818 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 3819 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3820 } __ end_if(); 3821 } __ end_if(); 3822 } __ end_if(); 3823 } __ end_if(); 3824 } else { 3825 // Object.clone() instrinsic uses this path. 3826 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3827 } 3828 3829 // Final sync IdealKit and GraphKit. 3830 final_sync(ideal); 3831 } 3832 #undef __ 3833 3834 3835 3836 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 3837 if (java_lang_String::has_offset_field()) { 3838 int offset_offset = java_lang_String::offset_offset_in_bytes(); 3839 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3840 false, NULL, 0); 3841 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 3842 int offset_field_idx = C->get_alias_index(offset_field_type); 3843 return make_load(ctrl, 3844 basic_plus_adr(str, str, offset_offset), 3845 TypeInt::INT, T_INT, offset_field_idx); 3846 } else { 3847 return intcon(0); 3848 } 3849 } 3850 3851 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 3852 if (java_lang_String::has_count_field()) { 3853 int count_offset = java_lang_String::count_offset_in_bytes(); 3854 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3855 false, NULL, 0); 3856 const TypePtr* count_field_type = string_type->add_offset(count_offset); 3857 int count_field_idx = C->get_alias_index(count_field_type); 3858 return make_load(ctrl, 3859 basic_plus_adr(str, str, count_offset), 3860 TypeInt::INT, T_INT, count_field_idx); 3861 } else { 3862 return load_array_length(load_String_value(ctrl, str)); 3863 } 3864 } 3865 3866 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 3867 int value_offset = java_lang_String::value_offset_in_bytes(); 3868 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3869 false, NULL, 0); 3870 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3871 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 3872 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 3873 ciTypeArrayKlass::make(T_CHAR), true, 0); 3874 int value_field_idx = C->get_alias_index(value_field_type); 3875 return make_load(ctrl, basic_plus_adr(str, str, value_offset), 3876 value_type, T_OBJECT, value_field_idx); 3877 } 3878 3879 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 3880 int offset_offset = java_lang_String::offset_offset_in_bytes(); 3881 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3882 false, NULL, 0); 3883 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 3884 int offset_field_idx = C->get_alias_index(offset_field_type); 3885 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 3886 value, T_INT, offset_field_idx); 3887 } 3888 3889 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 3890 int value_offset = java_lang_String::value_offset_in_bytes(); 3891 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3892 false, NULL, 0); 3893 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3894 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 3895 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 3896 ciTypeArrayKlass::make(T_CHAR), true, 0); 3897 int value_field_idx = C->get_alias_index(value_field_type); 3898 store_to_memory(ctrl, basic_plus_adr(str, value_offset), 3899 value, T_OBJECT, value_field_idx); 3900 } 3901 3902 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 3903 int count_offset = java_lang_String::count_offset_in_bytes(); 3904 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3905 false, NULL, 0); 3906 const TypePtr* count_field_type = string_type->add_offset(count_offset); 3907 int count_field_idx = C->get_alias_index(count_field_type); 3908 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 3909 value, T_INT, count_field_idx); 3910 }