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