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) { 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 enforcing certain bytecodes to reexecute if 819 // deoptimization happens 820 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 821 ciMethod* cur_method = jvms->method(); 822 int cur_bci = jvms->bci(); 823 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 824 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 825 return Interpreter::bytecode_should_reexecute(code) || 826 is_anewarray && code == Bytecodes::_multianewarray; 827 // Reexecute _multianewarray bytecode which was replaced with 828 // sequence of [a]newarray. See Parse::do_multianewarray(). 829 // 830 // Note: interpreter should not have it set since this optimization 831 // is limited by dimensions and guarded by flag so in some cases 832 // multianewarray() runtime calls will be generated and 833 // the bytecode should not be reexecutes (stack will not be reset). 834 } else 835 return false; 836 } 837 838 // Helper function for adding JVMState and debug information to node 839 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 840 // Add the safepoint edges to the call (or other safepoint). 841 842 // Make sure dead locals are set to top. This 843 // should help register allocation time and cut down on the size 844 // of the deoptimization information. 845 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 846 847 // Walk the inline list to fill in the correct set of JVMState's 848 // Also fill in the associated edges for each JVMState. 849 850 // If the bytecode needs to be reexecuted we need to put 851 // the arguments back on the stack. 852 const bool should_reexecute = jvms()->should_reexecute(); 853 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 854 855 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 856 // undefined if the bci is different. This is normal for Parse but it 857 // should not happen for LibraryCallKit because only one bci is processed. 858 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 859 "in LibraryCallKit the reexecute bit should not change"); 860 861 // If we are guaranteed to throw, we can prune everything but the 862 // input to the current bytecode. 863 bool can_prune_locals = false; 864 uint stack_slots_not_pruned = 0; 865 int inputs = 0, depth = 0; 866 if (must_throw) { 867 assert(method() == youngest_jvms->method(), "sanity"); 868 if (compute_stack_effects(inputs, depth)) { 869 can_prune_locals = true; 870 stack_slots_not_pruned = inputs; 871 } 872 } 873 874 if (env()->should_retain_local_variables()) { 875 // At any safepoint, this method can get breakpointed, which would 876 // then require an immediate deoptimization. 877 can_prune_locals = false; // do not prune locals 878 stack_slots_not_pruned = 0; 879 } 880 881 // do not scribble on the input jvms 882 JVMState* out_jvms = youngest_jvms->clone_deep(C); 883 call->set_jvms(out_jvms); // Start jvms list for call node 884 885 // For a known set of bytecodes, the interpreter should reexecute them if 886 // deoptimization happens. We set the reexecute state for them here 887 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 888 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 889 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 890 } 891 892 // Presize the call: 893 DEBUG_ONLY(uint non_debug_edges = call->req()); 894 call->add_req_batch(top(), youngest_jvms->debug_depth()); 895 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 896 897 // Set up edges so that the call looks like this: 898 // Call [state:] ctl io mem fptr retadr 899 // [parms:] parm0 ... parmN 900 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 901 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 902 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 903 // Note that caller debug info precedes callee debug info. 904 905 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 906 uint debug_ptr = call->req(); 907 908 // Loop over the map input edges associated with jvms, add them 909 // to the call node, & reset all offsets to match call node array. 910 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 911 uint debug_end = debug_ptr; 912 uint debug_start = debug_ptr - in_jvms->debug_size(); 913 debug_ptr = debug_start; // back up the ptr 914 915 uint p = debug_start; // walks forward in [debug_start, debug_end) 916 uint j, k, l; 917 SafePointNode* in_map = in_jvms->map(); 918 out_jvms->set_map(call); 919 920 if (can_prune_locals) { 921 assert(in_jvms->method() == out_jvms->method(), "sanity"); 922 // If the current throw can reach an exception handler in this JVMS, 923 // then we must keep everything live that can reach that handler. 924 // As a quick and dirty approximation, we look for any handlers at all. 925 if (in_jvms->method()->has_exception_handlers()) { 926 can_prune_locals = false; 927 } 928 } 929 930 // Add the Locals 931 k = in_jvms->locoff(); 932 l = in_jvms->loc_size(); 933 out_jvms->set_locoff(p); 934 if (!can_prune_locals) { 935 for (j = 0; j < l; j++) 936 call->set_req(p++, in_map->in(k+j)); 937 } else { 938 p += l; // already set to top above by add_req_batch 939 } 940 941 // Add the Expression Stack 942 k = in_jvms->stkoff(); 943 l = in_jvms->sp(); 944 out_jvms->set_stkoff(p); 945 if (!can_prune_locals) { 946 for (j = 0; j < l; j++) 947 call->set_req(p++, in_map->in(k+j)); 948 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 949 // Divide stack into {S0,...,S1}, where S0 is set to top. 950 uint s1 = stack_slots_not_pruned; 951 stack_slots_not_pruned = 0; // for next iteration 952 if (s1 > l) s1 = l; 953 uint s0 = l - s1; 954 p += s0; // skip the tops preinstalled by add_req_batch 955 for (j = s0; j < l; j++) 956 call->set_req(p++, in_map->in(k+j)); 957 } else { 958 p += l; // already set to top above by add_req_batch 959 } 960 961 // Add the Monitors 962 k = in_jvms->monoff(); 963 l = in_jvms->mon_size(); 964 out_jvms->set_monoff(p); 965 for (j = 0; j < l; j++) 966 call->set_req(p++, in_map->in(k+j)); 967 968 // Copy any scalar object fields. 969 k = in_jvms->scloff(); 970 l = in_jvms->scl_size(); 971 out_jvms->set_scloff(p); 972 for (j = 0; j < l; j++) 973 call->set_req(p++, in_map->in(k+j)); 974 975 // Finish the new jvms. 976 out_jvms->set_endoff(p); 977 978 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 979 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 980 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 981 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 982 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 983 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 984 985 // Update the two tail pointers in parallel. 986 out_jvms = out_jvms->caller(); 987 in_jvms = in_jvms->caller(); 988 } 989 990 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 991 992 // Test the correctness of JVMState::debug_xxx accessors: 993 assert(call->jvms()->debug_start() == non_debug_edges, ""); 994 assert(call->jvms()->debug_end() == call->req(), ""); 995 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 996 } 997 998 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 999 Bytecodes::Code code = java_bc(); 1000 if (code == Bytecodes::_wide) { 1001 code = method()->java_code_at_bci(bci() + 1); 1002 } 1003 1004 BasicType rtype = T_ILLEGAL; 1005 int rsize = 0; 1006 1007 if (code != Bytecodes::_illegal) { 1008 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1009 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1010 if (rtype < T_CONFLICT) 1011 rsize = type2size[rtype]; 1012 } 1013 1014 switch (code) { 1015 case Bytecodes::_illegal: 1016 return false; 1017 1018 case Bytecodes::_ldc: 1019 case Bytecodes::_ldc_w: 1020 case Bytecodes::_ldc2_w: 1021 inputs = 0; 1022 break; 1023 1024 case Bytecodes::_dup: inputs = 1; break; 1025 case Bytecodes::_dup_x1: inputs = 2; break; 1026 case Bytecodes::_dup_x2: inputs = 3; break; 1027 case Bytecodes::_dup2: inputs = 2; break; 1028 case Bytecodes::_dup2_x1: inputs = 3; break; 1029 case Bytecodes::_dup2_x2: inputs = 4; break; 1030 case Bytecodes::_swap: inputs = 2; break; 1031 case Bytecodes::_arraylength: inputs = 1; break; 1032 1033 case Bytecodes::_getstatic: 1034 case Bytecodes::_putstatic: 1035 case Bytecodes::_getfield: 1036 case Bytecodes::_vgetfield: 1037 case Bytecodes::_putfield: 1038 { 1039 bool ignored_will_link; 1040 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1041 int size = field->type()->size(); 1042 bool is_get = (depth >= 0), is_static = (depth & 1); 1043 inputs = (is_static ? 0 : 1); 1044 if (is_get) { 1045 depth = size - inputs; 1046 } else { 1047 inputs += size; // putxxx pops the value from the stack 1048 depth = - inputs; 1049 } 1050 } 1051 break; 1052 1053 case Bytecodes::_invokevirtual: 1054 case Bytecodes::_invokespecial: 1055 case Bytecodes::_invokestatic: 1056 case Bytecodes::_invokedynamic: 1057 case Bytecodes::_invokeinterface: 1058 { 1059 bool ignored_will_link; 1060 ciSignature* declared_signature = NULL; 1061 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1062 assert(declared_signature != NULL, "cannot be null"); 1063 inputs = declared_signature->arg_size_for_bc(code); 1064 int size = declared_signature->return_type()->size(); 1065 depth = size - inputs; 1066 } 1067 break; 1068 1069 case Bytecodes::_multianewarray: 1070 { 1071 ciBytecodeStream iter(method()); 1072 iter.reset_to_bci(bci()); 1073 iter.next(); 1074 inputs = iter.get_dimensions(); 1075 assert(rsize == 1, ""); 1076 depth = rsize - inputs; 1077 } 1078 break; 1079 1080 case Bytecodes::_vwithfield: { 1081 bool ignored_will_link; 1082 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1083 int size = field->type()->size(); 1084 inputs = size+1; 1085 depth = rsize - inputs; 1086 break; 1087 } 1088 1089 case Bytecodes::_ireturn: 1090 case Bytecodes::_lreturn: 1091 case Bytecodes::_freturn: 1092 case Bytecodes::_dreturn: 1093 case Bytecodes::_areturn: 1094 case Bytecodes::_vreturn: 1095 assert(rsize == -depth, ""); 1096 inputs = rsize; 1097 break; 1098 1099 case Bytecodes::_jsr: 1100 case Bytecodes::_jsr_w: 1101 inputs = 0; 1102 depth = 1; // S.B. depth=1, not zero 1103 break; 1104 1105 default: 1106 // bytecode produces a typed result 1107 inputs = rsize - depth; 1108 assert(inputs >= 0, ""); 1109 break; 1110 } 1111 1112 #ifdef ASSERT 1113 // spot check 1114 int outputs = depth + inputs; 1115 assert(outputs >= 0, "sanity"); 1116 switch (code) { 1117 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1118 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1119 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1120 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1121 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1122 } 1123 #endif //ASSERT 1124 1125 return true; 1126 } 1127 1128 1129 1130 //------------------------------basic_plus_adr--------------------------------- 1131 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1132 // short-circuit a common case 1133 if (offset == intcon(0)) return ptr; 1134 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1135 } 1136 1137 Node* GraphKit::ConvI2L(Node* offset) { 1138 // short-circuit a common case 1139 jint offset_con = find_int_con(offset, Type::OffsetBot); 1140 if (offset_con != Type::OffsetBot) { 1141 return longcon((jlong) offset_con); 1142 } 1143 return _gvn.transform( new ConvI2LNode(offset)); 1144 } 1145 1146 Node* GraphKit::ConvI2UL(Node* offset) { 1147 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1148 if (offset_con != (juint) Type::OffsetBot) { 1149 return longcon((julong) offset_con); 1150 } 1151 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1152 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1153 return _gvn.transform( new AndLNode(conv, mask) ); 1154 } 1155 1156 Node* GraphKit::ConvL2I(Node* offset) { 1157 // short-circuit a common case 1158 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1159 if (offset_con != (jlong)Type::OffsetBot) { 1160 return intcon((int) offset_con); 1161 } 1162 return _gvn.transform( new ConvL2INode(offset)); 1163 } 1164 1165 //-------------------------load_object_klass----------------------------------- 1166 Node* GraphKit::load_object_klass(Node* obj) { 1167 // Special-case a fresh allocation to avoid building nodes: 1168 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1169 if (akls != NULL) return akls; 1170 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1171 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1172 } 1173 1174 //-------------------------load_array_length----------------------------------- 1175 Node* GraphKit::load_array_length(Node* array) { 1176 // Special-case a fresh allocation to avoid building nodes: 1177 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1178 Node *alen; 1179 if (alloc == NULL) { 1180 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1181 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1182 } else { 1183 alen = alloc->Ideal_length(); 1184 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1185 if (ccast != alen) { 1186 alen = _gvn.transform(ccast); 1187 } 1188 } 1189 return alen; 1190 } 1191 1192 //------------------------------do_null_check---------------------------------- 1193 // Helper function to do a NULL pointer check. Returned value is 1194 // the incoming address with NULL casted away. You are allowed to use the 1195 // not-null value only if you are control dependent on the test. 1196 #ifndef PRODUCT 1197 extern int explicit_null_checks_inserted, 1198 explicit_null_checks_elided; 1199 #endif 1200 Node* GraphKit::null_check_common(Node* value, BasicType type, 1201 // optional arguments for variations: 1202 bool assert_null, 1203 Node* *null_control, 1204 bool speculative) { 1205 assert(!assert_null || null_control == NULL, "not both at once"); 1206 if (stopped()) return top(); 1207 NOT_PRODUCT(explicit_null_checks_inserted++); 1208 1209 // Construct NULL check 1210 Node *chk = NULL; 1211 switch(type) { 1212 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1213 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1214 case T_ARRAY : // fall through 1215 type = T_OBJECT; // simplify further tests 1216 case T_OBJECT : { 1217 const Type *t = _gvn.type( value ); 1218 1219 const TypeOopPtr* tp = t->isa_oopptr(); 1220 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1221 // Only for do_null_check, not any of its siblings: 1222 && !assert_null && null_control == NULL) { 1223 // Usually, any field access or invocation on an unloaded oop type 1224 // will simply fail to link, since the statically linked class is 1225 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1226 // the static class is loaded but the sharper oop type is not. 1227 // Rather than checking for this obscure case in lots of places, 1228 // we simply observe that a null check on an unloaded class 1229 // will always be followed by a nonsense operation, so we 1230 // can just issue the uncommon trap here. 1231 // Our access to the unloaded class will only be correct 1232 // after it has been loaded and initialized, which requires 1233 // a trip through the interpreter. 1234 #ifndef PRODUCT 1235 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1236 #endif 1237 uncommon_trap(Deoptimization::Reason_unloaded, 1238 Deoptimization::Action_reinterpret, 1239 tp->klass(), "!loaded"); 1240 return top(); 1241 } 1242 1243 if (assert_null) { 1244 // See if the type is contained in NULL_PTR. 1245 // If so, then the value is already null. 1246 if (t->higher_equal(TypePtr::NULL_PTR)) { 1247 NOT_PRODUCT(explicit_null_checks_elided++); 1248 return value; // Elided null assert quickly! 1249 } 1250 } else { 1251 // See if mixing in the NULL pointer changes type. 1252 // If so, then the NULL pointer was not allowed in the original 1253 // type. In other words, "value" was not-null. 1254 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1255 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1256 NOT_PRODUCT(explicit_null_checks_elided++); 1257 return value; // Elided null check quickly! 1258 } 1259 } 1260 chk = new CmpPNode( value, null() ); 1261 break; 1262 } 1263 1264 default: 1265 fatal("unexpected type: %s", type2name(type)); 1266 } 1267 assert(chk != NULL, "sanity check"); 1268 chk = _gvn.transform(chk); 1269 1270 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1271 BoolNode *btst = new BoolNode( chk, btest); 1272 Node *tst = _gvn.transform( btst ); 1273 1274 //----------- 1275 // if peephole optimizations occurred, a prior test existed. 1276 // If a prior test existed, maybe it dominates as we can avoid this test. 1277 if (tst != btst && type == T_OBJECT) { 1278 // At this point we want to scan up the CFG to see if we can 1279 // find an identical test (and so avoid this test altogether). 1280 Node *cfg = control(); 1281 int depth = 0; 1282 while( depth < 16 ) { // Limit search depth for speed 1283 if( cfg->Opcode() == Op_IfTrue && 1284 cfg->in(0)->in(1) == tst ) { 1285 // Found prior test. Use "cast_not_null" to construct an identical 1286 // CastPP (and hence hash to) as already exists for the prior test. 1287 // Return that casted value. 1288 if (assert_null) { 1289 replace_in_map(value, null()); 1290 return null(); // do not issue the redundant test 1291 } 1292 Node *oldcontrol = control(); 1293 set_control(cfg); 1294 Node *res = cast_not_null(value); 1295 set_control(oldcontrol); 1296 NOT_PRODUCT(explicit_null_checks_elided++); 1297 return res; 1298 } 1299 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1300 if (cfg == NULL) break; // Quit at region nodes 1301 depth++; 1302 } 1303 } 1304 1305 //----------- 1306 // Branch to failure if null 1307 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1308 Deoptimization::DeoptReason reason; 1309 if (assert_null) { 1310 reason = Deoptimization::Reason_null_assert; 1311 } else if (type == T_OBJECT) { 1312 reason = Deoptimization::reason_null_check(speculative); 1313 } else { 1314 reason = Deoptimization::Reason_div0_check; 1315 } 1316 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1317 // ciMethodData::has_trap_at will return a conservative -1 if any 1318 // must-be-null assertion has failed. This could cause performance 1319 // problems for a method after its first do_null_assert failure. 1320 // Consider using 'Reason_class_check' instead? 1321 1322 // To cause an implicit null check, we set the not-null probability 1323 // to the maximum (PROB_MAX). For an explicit check the probability 1324 // is set to a smaller value. 1325 if (null_control != NULL || too_many_traps(reason)) { 1326 // probability is less likely 1327 ok_prob = PROB_LIKELY_MAG(3); 1328 } else if (!assert_null && 1329 (ImplicitNullCheckThreshold > 0) && 1330 method() != NULL && 1331 (method()->method_data()->trap_count(reason) 1332 >= (uint)ImplicitNullCheckThreshold)) { 1333 ok_prob = PROB_LIKELY_MAG(3); 1334 } 1335 1336 if (null_control != NULL) { 1337 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1338 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1339 set_control( _gvn.transform( new IfTrueNode(iff))); 1340 #ifndef PRODUCT 1341 if (null_true == top()) { 1342 explicit_null_checks_elided++; 1343 } 1344 #endif 1345 (*null_control) = null_true; 1346 } else { 1347 BuildCutout unless(this, tst, ok_prob); 1348 // Check for optimizer eliding test at parse time 1349 if (stopped()) { 1350 // Failure not possible; do not bother making uncommon trap. 1351 NOT_PRODUCT(explicit_null_checks_elided++); 1352 } else if (assert_null) { 1353 uncommon_trap(reason, 1354 Deoptimization::Action_make_not_entrant, 1355 NULL, "assert_null"); 1356 } else { 1357 replace_in_map(value, zerocon(type)); 1358 builtin_throw(reason); 1359 } 1360 } 1361 1362 // Must throw exception, fall-thru not possible? 1363 if (stopped()) { 1364 return top(); // No result 1365 } 1366 1367 if (assert_null) { 1368 // Cast obj to null on this path. 1369 replace_in_map(value, zerocon(type)); 1370 return zerocon(type); 1371 } 1372 1373 // Cast obj to not-null on this path, if there is no null_control. 1374 // (If there is a null_control, a non-null value may come back to haunt us.) 1375 if (type == T_OBJECT) { 1376 Node* cast = cast_not_null(value, false); 1377 if (null_control == NULL || (*null_control) == top()) 1378 replace_in_map(value, cast); 1379 value = cast; 1380 } 1381 1382 return value; 1383 } 1384 1385 1386 //------------------------------cast_not_null---------------------------------- 1387 // Cast obj to not-null on this path 1388 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1389 const Type *t = _gvn.type(obj); 1390 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1391 // Object is already not-null? 1392 if( t == t_not_null ) return obj; 1393 1394 Node *cast = new CastPPNode(obj,t_not_null); 1395 cast->init_req(0, control()); 1396 cast = _gvn.transform( cast ); 1397 1398 // Scan for instances of 'obj' in the current JVM mapping. 1399 // These instances are known to be not-null after the test. 1400 if (do_replace_in_map) 1401 replace_in_map(obj, cast); 1402 1403 return cast; // Return casted value 1404 } 1405 1406 1407 //--------------------------replace_in_map------------------------------------- 1408 void GraphKit::replace_in_map(Node* old, Node* neww) { 1409 if (old == neww) { 1410 return; 1411 } 1412 1413 map()->replace_edge(old, neww); 1414 1415 // Note: This operation potentially replaces any edge 1416 // on the map. This includes locals, stack, and monitors 1417 // of the current (innermost) JVM state. 1418 1419 // don't let inconsistent types from profiling escape this 1420 // method 1421 1422 const Type* told = _gvn.type(old); 1423 const Type* tnew = _gvn.type(neww); 1424 1425 if (!tnew->higher_equal(told)) { 1426 return; 1427 } 1428 1429 map()->record_replaced_node(old, neww); 1430 } 1431 1432 1433 //============================================================================= 1434 //--------------------------------memory--------------------------------------- 1435 Node* GraphKit::memory(uint alias_idx) { 1436 MergeMemNode* mem = merged_memory(); 1437 Node* p = mem->memory_at(alias_idx); 1438 _gvn.set_type(p, Type::MEMORY); // must be mapped 1439 return p; 1440 } 1441 1442 //-----------------------------reset_memory------------------------------------ 1443 Node* GraphKit::reset_memory() { 1444 Node* mem = map()->memory(); 1445 // do not use this node for any more parsing! 1446 debug_only( map()->set_memory((Node*)NULL) ); 1447 return _gvn.transform( mem ); 1448 } 1449 1450 //------------------------------set_all_memory--------------------------------- 1451 void GraphKit::set_all_memory(Node* newmem) { 1452 Node* mergemem = MergeMemNode::make(newmem); 1453 gvn().set_type_bottom(mergemem); 1454 map()->set_memory(mergemem); 1455 } 1456 1457 //------------------------------set_all_memory_call---------------------------- 1458 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1459 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1460 set_all_memory(newmem); 1461 } 1462 1463 //============================================================================= 1464 // 1465 // parser factory methods for MemNodes 1466 // 1467 // These are layered on top of the factory methods in LoadNode and StoreNode, 1468 // and integrate with the parser's memory state and _gvn engine. 1469 // 1470 1471 // factory methods in "int adr_idx" 1472 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1473 int adr_idx, 1474 MemNode::MemOrd mo, 1475 LoadNode::ControlDependency control_dependency, 1476 bool require_atomic_access, 1477 bool unaligned, 1478 bool mismatched) { 1479 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1480 const TypePtr* adr_type = NULL; // debug-mode-only argument 1481 debug_only(adr_type = C->get_adr_type(adr_idx)); 1482 Node* mem = memory(adr_idx); 1483 Node* ld; 1484 if (require_atomic_access && bt == T_LONG) { 1485 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1486 } else if (require_atomic_access && bt == T_DOUBLE) { 1487 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1488 } else { 1489 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1490 } 1491 ld = _gvn.transform(ld); 1492 1493 if (bt == T_VALUETYPE) { 1494 // Load value type from oop 1495 ld = ValueTypeNode::make(gvn(), map()->memory(), ld); 1496 } else if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1497 // Improve graph before escape analysis and boxing elimination. 1498 record_for_igvn(ld); 1499 } 1500 return ld; 1501 } 1502 1503 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1504 int adr_idx, 1505 MemNode::MemOrd mo, 1506 bool require_atomic_access, 1507 bool unaligned, 1508 bool mismatched) { 1509 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1510 const TypePtr* adr_type = NULL; 1511 debug_only(adr_type = C->get_adr_type(adr_idx)); 1512 Node *mem = memory(adr_idx); 1513 Node* st; 1514 if (require_atomic_access && bt == T_LONG) { 1515 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1516 } else if (require_atomic_access && bt == T_DOUBLE) { 1517 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1518 } else { 1519 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1520 } 1521 if (unaligned) { 1522 st->as_Store()->set_unaligned_access(); 1523 } 1524 if (mismatched) { 1525 st->as_Store()->set_mismatched_access(); 1526 } 1527 st = _gvn.transform(st); 1528 set_memory(st, adr_idx); 1529 // Back-to-back stores can only remove intermediate store with DU info 1530 // so push on worklist for optimizer. 1531 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1532 record_for_igvn(st); 1533 1534 return st; 1535 } 1536 1537 1538 void GraphKit::pre_barrier(bool do_load, 1539 Node* ctl, 1540 Node* obj, 1541 Node* adr, 1542 uint adr_idx, 1543 Node* val, 1544 const TypeOopPtr* val_type, 1545 Node* pre_val, 1546 BasicType bt) { 1547 1548 BarrierSet* bs = Universe::heap()->barrier_set(); 1549 set_control(ctl); 1550 switch (bs->kind()) { 1551 case BarrierSet::G1SATBCTLogging: 1552 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1553 break; 1554 1555 case BarrierSet::CardTableForRS: 1556 case BarrierSet::CardTableExtension: 1557 case BarrierSet::ModRef: 1558 break; 1559 1560 default : 1561 ShouldNotReachHere(); 1562 1563 } 1564 } 1565 1566 bool GraphKit::can_move_pre_barrier() const { 1567 BarrierSet* bs = Universe::heap()->barrier_set(); 1568 switch (bs->kind()) { 1569 case BarrierSet::G1SATBCTLogging: 1570 return true; // Can move it if no safepoint 1571 1572 case BarrierSet::CardTableForRS: 1573 case BarrierSet::CardTableExtension: 1574 case BarrierSet::ModRef: 1575 return true; // There is no pre-barrier 1576 1577 default : 1578 ShouldNotReachHere(); 1579 } 1580 return false; 1581 } 1582 1583 void GraphKit::post_barrier(Node* ctl, 1584 Node* store, 1585 Node* obj, 1586 Node* adr, 1587 uint adr_idx, 1588 Node* val, 1589 BasicType bt, 1590 bool use_precise) { 1591 BarrierSet* bs = Universe::heap()->barrier_set(); 1592 set_control(ctl); 1593 switch (bs->kind()) { 1594 case BarrierSet::G1SATBCTLogging: 1595 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1596 break; 1597 1598 case BarrierSet::CardTableForRS: 1599 case BarrierSet::CardTableExtension: 1600 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1601 break; 1602 1603 case BarrierSet::ModRef: 1604 break; 1605 1606 default : 1607 ShouldNotReachHere(); 1608 1609 } 1610 } 1611 1612 Node* GraphKit::store_oop(Node* ctl, 1613 Node* obj, 1614 Node* adr, 1615 const TypePtr* adr_type, 1616 Node* val, 1617 const TypeOopPtr* val_type, 1618 BasicType bt, 1619 bool use_precise, 1620 MemNode::MemOrd mo, 1621 bool mismatched) { 1622 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1623 // could be delayed during Parse (for example, in adjust_map_after_if()). 1624 // Execute transformation here to avoid barrier generation in such case. 1625 if (_gvn.type(val) == TypePtr::NULL_PTR) 1626 val = _gvn.makecon(TypePtr::NULL_PTR); 1627 1628 set_control(ctl); 1629 if (stopped()) return top(); // Dead path ? 1630 1631 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity"); 1632 assert(val != NULL, "not dead path"); 1633 uint adr_idx = C->get_alias_index(adr_type); 1634 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1635 1636 if (bt == T_VALUETYPE) { 1637 // Allocate value type and store oop 1638 val = val->as_ValueType()->store_to_memory(this); 1639 } 1640 1641 pre_barrier(true /* do_load */, 1642 control(), obj, adr, adr_idx, val, val_type, 1643 NULL /* pre_val */, 1644 bt); 1645 1646 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1647 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1648 return store; 1649 } 1650 1651 // Could be an array or object we don't know at compile time (unsafe ref.) 1652 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1653 Node* obj, // containing obj 1654 Node* adr, // actual adress to store val at 1655 const TypePtr* adr_type, 1656 Node* val, 1657 BasicType bt, 1658 MemNode::MemOrd mo, 1659 bool mismatched) { 1660 Compile::AliasType* at = C->alias_type(adr_type); 1661 const TypeOopPtr* val_type = NULL; 1662 if (adr_type->isa_instptr()) { 1663 if (at->field() != NULL) { 1664 // known field. This code is a copy of the do_put_xxx logic. 1665 ciField* field = at->field(); 1666 if (!field->type()->is_loaded()) { 1667 val_type = TypeInstPtr::BOTTOM; 1668 } else { 1669 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1670 } 1671 } 1672 } else if (adr_type->isa_aryptr()) { 1673 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1674 } 1675 if (val_type == NULL) { 1676 val_type = TypeInstPtr::BOTTOM; 1677 } 1678 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1679 } 1680 1681 1682 //-------------------------array_element_address------------------------- 1683 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1684 const TypeInt* sizetype, Node* ctrl) { 1685 uint shift = exact_log2(type2aelembytes(elembt)); 1686 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1687 if (arytype_klass->is_value_array_klass()) { 1688 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass(); 1689 shift = vak->log2_element_size(); 1690 } 1691 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1692 1693 // short-circuit a common case (saves lots of confusing waste motion) 1694 jint idx_con = find_int_con(idx, -1); 1695 if (idx_con >= 0) { 1696 intptr_t offset = header + ((intptr_t)idx_con << shift); 1697 return basic_plus_adr(ary, offset); 1698 } 1699 1700 // must be correct type for alignment purposes 1701 Node* base = basic_plus_adr(ary, header); 1702 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1703 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1704 return basic_plus_adr(ary, base, scale); 1705 } 1706 1707 //-------------------------load_array_element------------------------- 1708 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1709 const Type* elemtype = arytype->elem(); 1710 BasicType elembt = elemtype->array_element_basic_type(); 1711 assert(elembt != T_VALUETYPE, "value types are not supported by this method"); 1712 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1713 if (elembt == T_NARROWOOP) { 1714 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1715 } 1716 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1717 return ld; 1718 } 1719 1720 //-------------------------set_arguments_for_java_call------------------------- 1721 // Arguments (pre-popped from the stack) are taken from the JVMS. 1722 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1723 // Add the call arguments: 1724 const TypeTuple* domain = call->tf()->domain_sig(); 1725 uint nargs = domain->cnt(); 1726 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1727 Node* arg = argument(i-TypeFunc::Parms); 1728 if (ValueTypePassFieldsAsArgs) { 1729 if (arg->is_ValueType()) { 1730 ValueTypeNode* vt = arg->as_ValueType(); 1731 if (domain->field_at(i)->is_valuetypeptr()->klass() != C->env()->___Value_klass()) { 1732 // We don't pass value type arguments by reference but instead 1733 // pass each field of the value type 1734 idx += vt->set_arguments_for_java_call(call, idx, *this); 1735 } else { 1736 arg = arg->as_ValueType()->store_to_memory(this); 1737 call->init_req(idx, arg); 1738 idx++; 1739 } 1740 // If a value type argument is passed as fields, attach the Method* to the call site 1741 // to be able to access the extended signature later via attached_method_before_pc(). 1742 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1743 call->set_override_symbolic_info(true); 1744 } else { 1745 call->init_req(idx, arg); 1746 idx++; 1747 } 1748 } else { 1749 if (arg->is_ValueType()) { 1750 // Pass value type argument via oop to callee 1751 arg = arg->as_ValueType()->store_to_memory(this); 1752 } 1753 call->init_req(i, arg); 1754 } 1755 } 1756 } 1757 1758 //---------------------------set_edges_for_java_call--------------------------- 1759 // Connect a newly created call into the current JVMS. 1760 // A return value node (if any) is returned from set_edges_for_java_call. 1761 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1762 1763 // Add the predefined inputs: 1764 call->init_req( TypeFunc::Control, control() ); 1765 call->init_req( TypeFunc::I_O , i_o() ); 1766 call->init_req( TypeFunc::Memory , reset_memory() ); 1767 call->init_req( TypeFunc::FramePtr, frameptr() ); 1768 call->init_req( TypeFunc::ReturnAdr, top() ); 1769 1770 add_safepoint_edges(call, must_throw); 1771 1772 Node* xcall = _gvn.transform(call); 1773 1774 if (xcall == top()) { 1775 set_control(top()); 1776 return; 1777 } 1778 assert(xcall == call, "call identity is stable"); 1779 1780 // Re-use the current map to produce the result. 1781 1782 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1783 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1784 set_all_memory_call(xcall, separate_io_proj); 1785 1786 //return xcall; // no need, caller already has it 1787 } 1788 1789 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1790 if (stopped()) return top(); // maybe the call folded up? 1791 1792 // Capture the return value, if any. 1793 Node* ret; 1794 if (call->method() == NULL || 1795 call->method()->return_type()->basic_type() == T_VOID) 1796 ret = top(); 1797 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1798 1799 // Note: Since any out-of-line call can produce an exception, 1800 // we always insert an I_O projection from the call into the result. 1801 1802 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1803 1804 if (separate_io_proj) { 1805 // The caller requested separate projections be used by the fall 1806 // through and exceptional paths, so replace the projections for 1807 // the fall through path. 1808 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1809 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1810 } 1811 return ret; 1812 } 1813 1814 //--------------------set_predefined_input_for_runtime_call-------------------- 1815 // Reading and setting the memory state is way conservative here. 1816 // The real problem is that I am not doing real Type analysis on memory, 1817 // so I cannot distinguish card mark stores from other stores. Across a GC 1818 // point the Store Barrier and the card mark memory has to agree. I cannot 1819 // have a card mark store and its barrier split across the GC point from 1820 // either above or below. Here I get that to happen by reading ALL of memory. 1821 // A better answer would be to separate out card marks from other memory. 1822 // For now, return the input memory state, so that it can be reused 1823 // after the call, if this call has restricted memory effects. 1824 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1825 // Set fixed predefined input arguments 1826 Node* memory = reset_memory(); 1827 call->init_req( TypeFunc::Control, control() ); 1828 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1829 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1830 call->init_req( TypeFunc::FramePtr, frameptr() ); 1831 call->init_req( TypeFunc::ReturnAdr, top() ); 1832 return memory; 1833 } 1834 1835 //-------------------set_predefined_output_for_runtime_call-------------------- 1836 // Set control and memory (not i_o) from the call. 1837 // If keep_mem is not NULL, use it for the output state, 1838 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1839 // If hook_mem is NULL, this call produces no memory effects at all. 1840 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1841 // then only that memory slice is taken from the call. 1842 // In the last case, we must put an appropriate memory barrier before 1843 // the call, so as to create the correct anti-dependencies on loads 1844 // preceding the call. 1845 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1846 Node* keep_mem, 1847 const TypePtr* hook_mem) { 1848 // no i/o 1849 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1850 if (keep_mem) { 1851 // First clone the existing memory state 1852 set_all_memory(keep_mem); 1853 if (hook_mem != NULL) { 1854 // Make memory for the call 1855 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1856 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1857 // We also use hook_mem to extract specific effects from arraycopy stubs. 1858 set_memory(mem, hook_mem); 1859 } 1860 // ...else the call has NO memory effects. 1861 1862 // Make sure the call advertises its memory effects precisely. 1863 // This lets us build accurate anti-dependences in gcm.cpp. 1864 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1865 "call node must be constructed correctly"); 1866 } else { 1867 assert(hook_mem == NULL, ""); 1868 // This is not a "slow path" call; all memory comes from the call. 1869 set_all_memory_call(call); 1870 } 1871 } 1872 1873 1874 // Replace the call with the current state of the kit. 1875 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1876 JVMState* ejvms = NULL; 1877 if (has_exceptions()) { 1878 ejvms = transfer_exceptions_into_jvms(); 1879 } 1880 1881 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1882 ReplacedNodes replaced_nodes_exception; 1883 Node* ex_ctl = top(); 1884 1885 SafePointNode* final_state = stop(); 1886 1887 // Find all the needed outputs of this call 1888 CallProjections callprojs; 1889 call->extract_projections(&callprojs, true); 1890 1891 Node* init_mem = call->in(TypeFunc::Memory); 1892 Node* final_mem = final_state->in(TypeFunc::Memory); 1893 Node* final_ctl = final_state->in(TypeFunc::Control); 1894 Node* final_io = final_state->in(TypeFunc::I_O); 1895 1896 // Replace all the old call edges with the edges from the inlining result 1897 if (callprojs.fallthrough_catchproj != NULL) { 1898 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1899 } 1900 if (callprojs.fallthrough_memproj != NULL) { 1901 if (final_mem->is_MergeMem()) { 1902 // Parser's exits MergeMem was not transformed but may be optimized 1903 final_mem = _gvn.transform(final_mem); 1904 } 1905 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1906 } 1907 if (callprojs.fallthrough_ioproj != NULL) { 1908 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1909 } 1910 1911 // Replace the result with the new result if it exists and is used 1912 if (callprojs.resproj != NULL && result != NULL) { 1913 C->gvn_replace_by(callprojs.resproj, result); 1914 } 1915 1916 if (ejvms == NULL) { 1917 // No exception edges to simply kill off those paths 1918 if (callprojs.catchall_catchproj != NULL) { 1919 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1920 } 1921 if (callprojs.catchall_memproj != NULL) { 1922 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1923 } 1924 if (callprojs.catchall_ioproj != NULL) { 1925 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1926 } 1927 // Replace the old exception object with top 1928 if (callprojs.exobj != NULL) { 1929 C->gvn_replace_by(callprojs.exobj, C->top()); 1930 } 1931 } else { 1932 GraphKit ekit(ejvms); 1933 1934 // Load my combined exception state into the kit, with all phis transformed: 1935 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1936 replaced_nodes_exception = ex_map->replaced_nodes(); 1937 1938 Node* ex_oop = ekit.use_exception_state(ex_map); 1939 1940 if (callprojs.catchall_catchproj != NULL) { 1941 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1942 ex_ctl = ekit.control(); 1943 } 1944 if (callprojs.catchall_memproj != NULL) { 1945 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1946 } 1947 if (callprojs.catchall_ioproj != NULL) { 1948 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1949 } 1950 1951 // Replace the old exception object with the newly created one 1952 if (callprojs.exobj != NULL) { 1953 C->gvn_replace_by(callprojs.exobj, ex_oop); 1954 } 1955 } 1956 1957 // Disconnect the call from the graph 1958 call->disconnect_inputs(NULL, C); 1959 C->gvn_replace_by(call, C->top()); 1960 1961 // Clean up any MergeMems that feed other MergeMems since the 1962 // optimizer doesn't like that. 1963 if (final_mem->is_MergeMem()) { 1964 Node_List wl; 1965 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1966 Node* m = i.get(); 1967 if (m->is_MergeMem() && !wl.contains(m)) { 1968 wl.push(m); 1969 } 1970 } 1971 while (wl.size() > 0) { 1972 _gvn.transform(wl.pop()); 1973 } 1974 } 1975 1976 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1977 replaced_nodes.apply(C, final_ctl); 1978 } 1979 if (!ex_ctl->is_top() && do_replaced_nodes) { 1980 replaced_nodes_exception.apply(C, ex_ctl); 1981 } 1982 } 1983 1984 1985 //------------------------------increment_counter------------------------------ 1986 // for statistics: increment a VM counter by 1 1987 1988 void GraphKit::increment_counter(address counter_addr) { 1989 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1990 increment_counter(adr1); 1991 } 1992 1993 void GraphKit::increment_counter(Node* counter_addr) { 1994 int adr_type = Compile::AliasIdxRaw; 1995 Node* ctrl = control(); 1996 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1997 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1998 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1999 } 2000 2001 2002 //------------------------------uncommon_trap---------------------------------- 2003 // Bail out to the interpreter in mid-method. Implemented by calling the 2004 // uncommon_trap blob. This helper function inserts a runtime call with the 2005 // right debug info. 2006 void GraphKit::uncommon_trap(int trap_request, 2007 ciKlass* klass, const char* comment, 2008 bool must_throw, 2009 bool keep_exact_action) { 2010 if (failing()) stop(); 2011 if (stopped()) return; // trap reachable? 2012 2013 // Note: If ProfileTraps is true, and if a deopt. actually 2014 // occurs here, the runtime will make sure an MDO exists. There is 2015 // no need to call method()->ensure_method_data() at this point. 2016 2017 // Set the stack pointer to the right value for reexecution: 2018 set_sp(reexecute_sp()); 2019 2020 #ifdef ASSERT 2021 if (!must_throw) { 2022 // Make sure the stack has at least enough depth to execute 2023 // the current bytecode. 2024 int inputs, ignored_depth; 2025 if (compute_stack_effects(inputs, ignored_depth)) { 2026 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2027 Bytecodes::name(java_bc()), sp(), inputs); 2028 } 2029 } 2030 #endif 2031 2032 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2033 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2034 2035 switch (action) { 2036 case Deoptimization::Action_maybe_recompile: 2037 case Deoptimization::Action_reinterpret: 2038 // Temporary fix for 6529811 to allow virtual calls to be sure they 2039 // get the chance to go from mono->bi->mega 2040 if (!keep_exact_action && 2041 Deoptimization::trap_request_index(trap_request) < 0 && 2042 too_many_recompiles(reason)) { 2043 // This BCI is causing too many recompilations. 2044 if (C->log() != NULL) { 2045 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2046 Deoptimization::trap_reason_name(reason), 2047 Deoptimization::trap_action_name(action)); 2048 } 2049 action = Deoptimization::Action_none; 2050 trap_request = Deoptimization::make_trap_request(reason, action); 2051 } else { 2052 C->set_trap_can_recompile(true); 2053 } 2054 break; 2055 case Deoptimization::Action_make_not_entrant: 2056 C->set_trap_can_recompile(true); 2057 break; 2058 #ifdef ASSERT 2059 case Deoptimization::Action_none: 2060 case Deoptimization::Action_make_not_compilable: 2061 break; 2062 default: 2063 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2064 break; 2065 #endif 2066 } 2067 2068 if (TraceOptoParse) { 2069 char buf[100]; 2070 tty->print_cr("Uncommon trap %s at bci:%d", 2071 Deoptimization::format_trap_request(buf, sizeof(buf), 2072 trap_request), bci()); 2073 } 2074 2075 CompileLog* log = C->log(); 2076 if (log != NULL) { 2077 int kid = (klass == NULL)? -1: log->identify(klass); 2078 log->begin_elem("uncommon_trap bci='%d'", bci()); 2079 char buf[100]; 2080 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2081 trap_request)); 2082 if (kid >= 0) log->print(" klass='%d'", kid); 2083 if (comment != NULL) log->print(" comment='%s'", comment); 2084 log->end_elem(); 2085 } 2086 2087 // Make sure any guarding test views this path as very unlikely 2088 Node *i0 = control()->in(0); 2089 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2090 IfNode *iff = i0->as_If(); 2091 float f = iff->_prob; // Get prob 2092 if (control()->Opcode() == Op_IfTrue) { 2093 if (f > PROB_UNLIKELY_MAG(4)) 2094 iff->_prob = PROB_MIN; 2095 } else { 2096 if (f < PROB_LIKELY_MAG(4)) 2097 iff->_prob = PROB_MAX; 2098 } 2099 } 2100 2101 // Clear out dead values from the debug info. 2102 kill_dead_locals(); 2103 2104 // Now insert the uncommon trap subroutine call 2105 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2106 const TypePtr* no_memory_effects = NULL; 2107 // Pass the index of the class to be loaded 2108 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2109 (must_throw ? RC_MUST_THROW : 0), 2110 OptoRuntime::uncommon_trap_Type(), 2111 call_addr, "uncommon_trap", no_memory_effects, 2112 intcon(trap_request)); 2113 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2114 "must extract request correctly from the graph"); 2115 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2116 2117 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2118 // The debug info is the only real input to this call. 2119 2120 // Halt-and-catch fire here. The above call should never return! 2121 HaltNode* halt = new HaltNode(control(), frameptr()); 2122 _gvn.set_type_bottom(halt); 2123 root()->add_req(halt); 2124 2125 stop_and_kill_map(); 2126 } 2127 2128 2129 //--------------------------just_allocated_object------------------------------ 2130 // Report the object that was just allocated. 2131 // It must be the case that there are no intervening safepoints. 2132 // We use this to determine if an object is so "fresh" that 2133 // it does not require card marks. 2134 Node* GraphKit::just_allocated_object(Node* current_control) { 2135 if (C->recent_alloc_ctl() == current_control) 2136 return C->recent_alloc_obj(); 2137 return NULL; 2138 } 2139 2140 2141 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2142 // (Note: TypeFunc::make has a cache that makes this fast.) 2143 const TypeFunc* tf = TypeFunc::make(dest_method); 2144 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2145 for (int j = 0; j < nargs; j++) { 2146 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2147 if( targ->basic_type() == T_DOUBLE ) { 2148 // If any parameters are doubles, they must be rounded before 2149 // the call, dstore_rounding does gvn.transform 2150 Node *arg = argument(j); 2151 arg = dstore_rounding(arg); 2152 set_argument(j, arg); 2153 } 2154 } 2155 } 2156 2157 /** 2158 * Record profiling data exact_kls for Node n with the type system so 2159 * that it can propagate it (speculation) 2160 * 2161 * @param n node that the type applies to 2162 * @param exact_kls type from profiling 2163 * @param maybe_null did profiling see null? 2164 * 2165 * @return node with improved type 2166 */ 2167 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2168 const Type* current_type = _gvn.type(n); 2169 assert(UseTypeSpeculation, "type speculation must be on"); 2170 2171 const TypePtr* speculative = current_type->speculative(); 2172 2173 // Should the klass from the profile be recorded in the speculative type? 2174 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2175 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2176 const TypeOopPtr* xtype = tklass->as_instance_type(); 2177 assert(xtype->klass_is_exact(), "Should be exact"); 2178 // Any reason to believe n is not null (from this profiling or a previous one)? 2179 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2180 // record the new speculative type's depth 2181 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2182 speculative = speculative->with_inline_depth(jvms()->depth()); 2183 } else if (current_type->would_improve_ptr(maybe_null)) { 2184 // Profiling report that null was never seen so we can change the 2185 // speculative type to non null ptr. 2186 assert(!maybe_null, "nothing to improve"); 2187 if (speculative == NULL) { 2188 speculative = TypePtr::NOTNULL; 2189 } else { 2190 const TypePtr* ptr = TypePtr::NOTNULL; 2191 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2192 } 2193 } 2194 2195 if (speculative != current_type->speculative()) { 2196 // Build a type with a speculative type (what we think we know 2197 // about the type but will need a guard when we use it) 2198 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2199 // We're changing the type, we need a new CheckCast node to carry 2200 // the new type. The new type depends on the control: what 2201 // profiling tells us is only valid from here as far as we can 2202 // tell. 2203 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2204 cast = _gvn.transform(cast); 2205 replace_in_map(n, cast); 2206 n = cast; 2207 } 2208 2209 return n; 2210 } 2211 2212 /** 2213 * Record profiling data from receiver profiling at an invoke with the 2214 * type system so that it can propagate it (speculation) 2215 * 2216 * @param n receiver node 2217 * 2218 * @return node with improved type 2219 */ 2220 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2221 if (!UseTypeSpeculation) { 2222 return n; 2223 } 2224 ciKlass* exact_kls = profile_has_unique_klass(); 2225 bool maybe_null = true; 2226 if (java_bc() == Bytecodes::_checkcast || 2227 java_bc() == Bytecodes::_instanceof || 2228 java_bc() == Bytecodes::_aastore) { 2229 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2230 maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2231 } 2232 return record_profile_for_speculation(n, exact_kls, maybe_null); 2233 } 2234 2235 /** 2236 * Record profiling data from argument profiling at an invoke with the 2237 * type system so that it can propagate it (speculation) 2238 * 2239 * @param dest_method target method for the call 2240 * @param bc what invoke bytecode is this? 2241 */ 2242 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2243 if (!UseTypeSpeculation) { 2244 return; 2245 } 2246 const TypeFunc* tf = TypeFunc::make(dest_method); 2247 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2248 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2249 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2250 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2251 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2252 bool maybe_null = true; 2253 ciKlass* better_type = NULL; 2254 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2255 record_profile_for_speculation(argument(j), better_type, maybe_null); 2256 } 2257 i++; 2258 } 2259 } 2260 } 2261 2262 /** 2263 * Record profiling data from parameter profiling at an invoke with 2264 * the type system so that it can propagate it (speculation) 2265 */ 2266 void GraphKit::record_profiled_parameters_for_speculation() { 2267 if (!UseTypeSpeculation) { 2268 return; 2269 } 2270 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2271 if (_gvn.type(local(i))->isa_oopptr()) { 2272 bool maybe_null = true; 2273 ciKlass* better_type = NULL; 2274 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2275 record_profile_for_speculation(local(i), better_type, maybe_null); 2276 } 2277 j++; 2278 } 2279 } 2280 } 2281 2282 /** 2283 * Record profiling data from return value profiling at an invoke with 2284 * the type system so that it can propagate it (speculation) 2285 */ 2286 void GraphKit::record_profiled_return_for_speculation() { 2287 if (!UseTypeSpeculation) { 2288 return; 2289 } 2290 bool maybe_null = true; 2291 ciKlass* better_type = NULL; 2292 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2293 // If profiling reports a single type for the return value, 2294 // feed it to the type system so it can propagate it as a 2295 // speculative type 2296 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2297 } 2298 } 2299 2300 void GraphKit::round_double_result(ciMethod* dest_method) { 2301 // A non-strict method may return a double value which has an extended 2302 // exponent, but this must not be visible in a caller which is 'strict' 2303 // If a strict caller invokes a non-strict callee, round a double result 2304 2305 BasicType result_type = dest_method->return_type()->basic_type(); 2306 assert( method() != NULL, "must have caller context"); 2307 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2308 // Destination method's return value is on top of stack 2309 // dstore_rounding() does gvn.transform 2310 Node *result = pop_pair(); 2311 result = dstore_rounding(result); 2312 push_pair(result); 2313 } 2314 } 2315 2316 // rounding for strict float precision conformance 2317 Node* GraphKit::precision_rounding(Node* n) { 2318 return UseStrictFP && _method->flags().is_strict() 2319 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2320 ? _gvn.transform( new RoundFloatNode(0, n) ) 2321 : n; 2322 } 2323 2324 // rounding for strict double precision conformance 2325 Node* GraphKit::dprecision_rounding(Node *n) { 2326 return UseStrictFP && _method->flags().is_strict() 2327 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2328 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2329 : n; 2330 } 2331 2332 // rounding for non-strict double stores 2333 Node* GraphKit::dstore_rounding(Node* n) { 2334 return Matcher::strict_fp_requires_explicit_rounding 2335 && UseSSE <= 1 2336 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2337 : n; 2338 } 2339 2340 //============================================================================= 2341 // Generate a fast path/slow path idiom. Graph looks like: 2342 // [foo] indicates that 'foo' is a parameter 2343 // 2344 // [in] NULL 2345 // \ / 2346 // CmpP 2347 // Bool ne 2348 // If 2349 // / \ 2350 // True False-<2> 2351 // / | 2352 // / cast_not_null 2353 // Load | | ^ 2354 // [fast_test] | | 2355 // gvn to opt_test | | 2356 // / \ | <1> 2357 // True False | 2358 // | \\ | 2359 // [slow_call] \[fast_result] 2360 // Ctl Val \ \ 2361 // | \ \ 2362 // Catch <1> \ \ 2363 // / \ ^ \ \ 2364 // Ex No_Ex | \ \ 2365 // | \ \ | \ <2> \ 2366 // ... \ [slow_res] | | \ [null_result] 2367 // \ \--+--+--- | | 2368 // \ | / \ | / 2369 // --------Region Phi 2370 // 2371 //============================================================================= 2372 // Code is structured as a series of driver functions all called 'do_XXX' that 2373 // call a set of helper functions. Helper functions first, then drivers. 2374 2375 //------------------------------null_check_oop--------------------------------- 2376 // Null check oop. Set null-path control into Region in slot 3. 2377 // Make a cast-not-nullness use the other not-null control. Return cast. 2378 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2379 bool never_see_null, 2380 bool safe_for_replace, 2381 bool speculative) { 2382 // Initial NULL check taken path 2383 (*null_control) = top(); 2384 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2385 2386 // Generate uncommon_trap: 2387 if (never_see_null && (*null_control) != top()) { 2388 // If we see an unexpected null at a check-cast we record it and force a 2389 // recompile; the offending check-cast will be compiled to handle NULLs. 2390 // If we see more than one offending BCI, then all checkcasts in the 2391 // method will be compiled to handle NULLs. 2392 PreserveJVMState pjvms(this); 2393 set_control(*null_control); 2394 replace_in_map(value, null()); 2395 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2396 uncommon_trap(reason, 2397 Deoptimization::Action_make_not_entrant); 2398 (*null_control) = top(); // NULL path is dead 2399 } 2400 if ((*null_control) == top() && safe_for_replace) { 2401 replace_in_map(value, cast); 2402 } 2403 2404 // Cast away null-ness on the result 2405 return cast; 2406 } 2407 2408 //------------------------------opt_iff---------------------------------------- 2409 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2410 // Return slow-path control. 2411 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2412 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2413 2414 // Fast path taken; set region slot 2 2415 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2416 region->init_req(2,fast_taken); // Capture fast-control 2417 2418 // Fast path not-taken, i.e. slow path 2419 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2420 return slow_taken; 2421 } 2422 2423 //-----------------------------make_runtime_call------------------------------- 2424 Node* GraphKit::make_runtime_call(int flags, 2425 const TypeFunc* call_type, address call_addr, 2426 const char* call_name, 2427 const TypePtr* adr_type, 2428 // The following parms are all optional. 2429 // The first NULL ends the list. 2430 Node* parm0, Node* parm1, 2431 Node* parm2, Node* parm3, 2432 Node* parm4, Node* parm5, 2433 Node* parm6, Node* parm7) { 2434 // Slow-path call 2435 bool is_leaf = !(flags & RC_NO_LEAF); 2436 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2437 if (call_name == NULL) { 2438 assert(!is_leaf, "must supply name for leaf"); 2439 call_name = OptoRuntime::stub_name(call_addr); 2440 } 2441 CallNode* call; 2442 if (!is_leaf) { 2443 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2444 bci(), adr_type); 2445 } else if (flags & RC_NO_FP) { 2446 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2447 } else { 2448 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2449 } 2450 2451 // The following is similar to set_edges_for_java_call, 2452 // except that the memory effects of the call are restricted to AliasIdxRaw. 2453 2454 // Slow path call has no side-effects, uses few values 2455 bool wide_in = !(flags & RC_NARROW_MEM); 2456 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2457 2458 Node* prev_mem = NULL; 2459 if (wide_in) { 2460 prev_mem = set_predefined_input_for_runtime_call(call); 2461 } else { 2462 assert(!wide_out, "narrow in => narrow out"); 2463 Node* narrow_mem = memory(adr_type); 2464 prev_mem = reset_memory(); 2465 map()->set_memory(narrow_mem); 2466 set_predefined_input_for_runtime_call(call); 2467 } 2468 2469 // Hook each parm in order. Stop looking at the first NULL. 2470 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2471 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2472 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2473 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2474 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2475 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2476 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2477 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2478 /* close each nested if ===> */ } } } } } } } } 2479 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2480 2481 if (!is_leaf) { 2482 // Non-leaves can block and take safepoints: 2483 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2484 } 2485 // Non-leaves can throw exceptions: 2486 if (has_io) { 2487 call->set_req(TypeFunc::I_O, i_o()); 2488 } 2489 2490 if (flags & RC_UNCOMMON) { 2491 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2492 // (An "if" probability corresponds roughly to an unconditional count. 2493 // Sort of.) 2494 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2495 } 2496 2497 Node* c = _gvn.transform(call); 2498 assert(c == call, "cannot disappear"); 2499 2500 if (wide_out) { 2501 // Slow path call has full side-effects. 2502 set_predefined_output_for_runtime_call(call); 2503 } else { 2504 // Slow path call has few side-effects, and/or sets few values. 2505 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2506 } 2507 2508 if (has_io) { 2509 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2510 } 2511 return call; 2512 2513 } 2514 2515 //------------------------------merge_memory----------------------------------- 2516 // Merge memory from one path into the current memory state. 2517 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2518 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2519 Node* old_slice = mms.force_memory(); 2520 Node* new_slice = mms.memory2(); 2521 if (old_slice != new_slice) { 2522 PhiNode* phi; 2523 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2524 if (mms.is_empty()) { 2525 // clone base memory Phi's inputs for this memory slice 2526 assert(old_slice == mms.base_memory(), "sanity"); 2527 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2528 _gvn.set_type(phi, Type::MEMORY); 2529 for (uint i = 1; i < phi->req(); i++) { 2530 phi->init_req(i, old_slice->in(i)); 2531 } 2532 } else { 2533 phi = old_slice->as_Phi(); // Phi was generated already 2534 } 2535 } else { 2536 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2537 _gvn.set_type(phi, Type::MEMORY); 2538 } 2539 phi->set_req(new_path, new_slice); 2540 mms.set_memory(phi); 2541 } 2542 } 2543 } 2544 2545 //------------------------------make_slow_call_ex------------------------------ 2546 // Make the exception handler hookups for the slow call 2547 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2548 if (stopped()) return; 2549 2550 // Make a catch node with just two handlers: fall-through and catch-all 2551 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2552 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2553 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2554 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2555 2556 { PreserveJVMState pjvms(this); 2557 set_control(excp); 2558 set_i_o(i_o); 2559 2560 if (excp != top()) { 2561 if (deoptimize) { 2562 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2563 uncommon_trap(Deoptimization::Reason_unhandled, 2564 Deoptimization::Action_none); 2565 } else { 2566 // Create an exception state also. 2567 // Use an exact type if the caller has specified a specific exception. 2568 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2569 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2570 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2571 } 2572 } 2573 } 2574 2575 // Get the no-exception control from the CatchNode. 2576 set_control(norm); 2577 } 2578 2579 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2580 Node* cmp = NULL; 2581 switch(bt) { 2582 case T_INT: cmp = new CmpINode(in1, in2); break; 2583 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2584 default: fatal("unexpected comparison type %s", type2name(bt)); 2585 } 2586 gvn->transform(cmp); 2587 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2588 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2589 gvn->transform(iff); 2590 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2591 return iff; 2592 } 2593 2594 2595 //-------------------------------gen_subtype_check----------------------------- 2596 // Generate a subtyping check. Takes as input the subtype and supertype. 2597 // Returns 2 values: sets the default control() to the true path and returns 2598 // the false path. Only reads invariant memory; sets no (visible) memory. 2599 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2600 // but that's not exposed to the optimizer. This call also doesn't take in an 2601 // Object; if you wish to check an Object you need to load the Object's class 2602 // prior to coming here. 2603 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2604 Compile* C = gvn->C; 2605 2606 if ((*ctrl)->is_top()) { 2607 return C->top(); 2608 } 2609 2610 // Fast check for identical types, perhaps identical constants. 2611 // The types can even be identical non-constants, in cases 2612 // involving Array.newInstance, Object.clone, etc. 2613 if (subklass == superklass) 2614 return C->top(); // false path is dead; no test needed. 2615 2616 if (gvn->type(superklass)->singleton()) { 2617 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2618 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2619 2620 // In the common case of an exact superklass, try to fold up the 2621 // test before generating code. You may ask, why not just generate 2622 // the code and then let it fold up? The answer is that the generated 2623 // code will necessarily include null checks, which do not always 2624 // completely fold away. If they are also needless, then they turn 2625 // into a performance loss. Example: 2626 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2627 // Here, the type of 'fa' is often exact, so the store check 2628 // of fa[1]=x will fold up, without testing the nullness of x. 2629 switch (C->static_subtype_check(superk, subk)) { 2630 case Compile::SSC_always_false: 2631 { 2632 Node* always_fail = *ctrl; 2633 *ctrl = gvn->C->top(); 2634 return always_fail; 2635 } 2636 case Compile::SSC_always_true: 2637 return C->top(); 2638 case Compile::SSC_easy_test: 2639 { 2640 // Just do a direct pointer compare and be done. 2641 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2642 *ctrl = gvn->transform(new IfTrueNode(iff)); 2643 return gvn->transform(new IfFalseNode(iff)); 2644 } 2645 case Compile::SSC_full_test: 2646 break; 2647 default: 2648 ShouldNotReachHere(); 2649 } 2650 } 2651 2652 // %%% Possible further optimization: Even if the superklass is not exact, 2653 // if the subklass is the unique subtype of the superklass, the check 2654 // will always succeed. We could leave a dependency behind to ensure this. 2655 2656 // First load the super-klass's check-offset 2657 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2658 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2659 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2660 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2661 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2662 2663 // Load from the sub-klass's super-class display list, or a 1-word cache of 2664 // the secondary superclass list, or a failing value with a sentinel offset 2665 // if the super-klass is an interface or exceptionally deep in the Java 2666 // hierarchy and we have to scan the secondary superclass list the hard way. 2667 // Worst-case type is a little odd: NULL is allowed as a result (usually 2668 // klass loads can never produce a NULL). 2669 Node *chk_off_X = chk_off; 2670 #ifdef _LP64 2671 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2672 #endif 2673 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2674 // For some types like interfaces the following loadKlass is from a 1-word 2675 // cache which is mutable so can't use immutable memory. Other 2676 // types load from the super-class display table which is immutable. 2677 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2678 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2679 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2680 2681 // Compile speed common case: ARE a subtype and we canNOT fail 2682 if( superklass == nkls ) 2683 return C->top(); // false path is dead; no test needed. 2684 2685 // See if we get an immediate positive hit. Happens roughly 83% of the 2686 // time. Test to see if the value loaded just previously from the subklass 2687 // is exactly the superklass. 2688 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2689 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2690 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2691 2692 // Compile speed common case: Check for being deterministic right now. If 2693 // chk_off is a constant and not equal to cacheoff then we are NOT a 2694 // subklass. In this case we need exactly the 1 test above and we can 2695 // return those results immediately. 2696 if (!might_be_cache) { 2697 Node* not_subtype_ctrl = *ctrl; 2698 *ctrl = iftrue1; // We need exactly the 1 test above 2699 return not_subtype_ctrl; 2700 } 2701 2702 // Gather the various success & failures here 2703 RegionNode *r_ok_subtype = new RegionNode(4); 2704 gvn->record_for_igvn(r_ok_subtype); 2705 RegionNode *r_not_subtype = new RegionNode(3); 2706 gvn->record_for_igvn(r_not_subtype); 2707 2708 r_ok_subtype->init_req(1, iftrue1); 2709 2710 // Check for immediate negative hit. Happens roughly 11% of the time (which 2711 // is roughly 63% of the remaining cases). Test to see if the loaded 2712 // check-offset points into the subklass display list or the 1-element 2713 // cache. If it points to the display (and NOT the cache) and the display 2714 // missed then it's not a subtype. 2715 Node *cacheoff = gvn->intcon(cacheoff_con); 2716 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2717 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2718 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2719 2720 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2721 // No performance impact (too rare) but allows sharing of secondary arrays 2722 // which has some footprint reduction. 2723 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2724 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2725 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2726 2727 // -- Roads not taken here: -- 2728 // We could also have chosen to perform the self-check at the beginning 2729 // of this code sequence, as the assembler does. This would not pay off 2730 // the same way, since the optimizer, unlike the assembler, can perform 2731 // static type analysis to fold away many successful self-checks. 2732 // Non-foldable self checks work better here in second position, because 2733 // the initial primary superclass check subsumes a self-check for most 2734 // types. An exception would be a secondary type like array-of-interface, 2735 // which does not appear in its own primary supertype display. 2736 // Finally, we could have chosen to move the self-check into the 2737 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2738 // dependent manner. But it is worthwhile to have the check here, 2739 // where it can be perhaps be optimized. The cost in code space is 2740 // small (register compare, branch). 2741 2742 // Now do a linear scan of the secondary super-klass array. Again, no real 2743 // performance impact (too rare) but it's gotta be done. 2744 // Since the code is rarely used, there is no penalty for moving it 2745 // out of line, and it can only improve I-cache density. 2746 // The decision to inline or out-of-line this final check is platform 2747 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2748 Node* psc = gvn->transform( 2749 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2750 2751 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2752 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2753 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2754 2755 // Return false path; set default control to true path. 2756 *ctrl = gvn->transform(r_ok_subtype); 2757 return gvn->transform(r_not_subtype); 2758 } 2759 2760 // Profile-driven exact type check: 2761 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2762 float prob, 2763 Node* *casted_receiver) { 2764 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2765 Node* recv_klass = load_object_klass(receiver); 2766 Node* want_klass = makecon(tklass); 2767 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2768 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2769 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2770 set_control( _gvn.transform( new IfTrueNode (iff) )); 2771 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2772 2773 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2774 assert(recv_xtype->klass_is_exact(), ""); 2775 2776 // Subsume downstream occurrences of receiver with a cast to 2777 // recv_xtype, since now we know what the type will be. 2778 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2779 (*casted_receiver) = _gvn.transform(cast); 2780 // (User must make the replace_in_map call.) 2781 2782 return fail; 2783 } 2784 2785 2786 //------------------------------seems_never_null------------------------------- 2787 // Use null_seen information if it is available from the profile. 2788 // If we see an unexpected null at a type check we record it and force a 2789 // recompile; the offending check will be recompiled to handle NULLs. 2790 // If we see several offending BCIs, then all checks in the 2791 // method will be recompiled. 2792 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2793 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2794 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2795 if (UncommonNullCast // Cutout for this technique 2796 && obj != null() // And not the -Xcomp stupid case? 2797 && !too_many_traps(reason) 2798 ) { 2799 if (speculating) { 2800 return true; 2801 } 2802 if (data == NULL) 2803 // Edge case: no mature data. Be optimistic here. 2804 return true; 2805 // If the profile has not seen a null, assume it won't happen. 2806 assert(java_bc() == Bytecodes::_checkcast || 2807 java_bc() == Bytecodes::_instanceof || 2808 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2809 return !data->as_BitData()->null_seen(); 2810 } 2811 speculating = false; 2812 return false; 2813 } 2814 2815 //------------------------maybe_cast_profiled_receiver------------------------- 2816 // If the profile has seen exactly one type, narrow to exactly that type. 2817 // Subsequent type checks will always fold up. 2818 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2819 ciKlass* require_klass, 2820 ciKlass* spec_klass, 2821 bool safe_for_replace) { 2822 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2823 2824 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2825 2826 // Make sure we haven't already deoptimized from this tactic. 2827 if (too_many_traps(reason) || too_many_recompiles(reason)) 2828 return NULL; 2829 2830 // (No, this isn't a call, but it's enough like a virtual call 2831 // to use the same ciMethod accessor to get the profile info...) 2832 // If we have a speculative type use it instead of profiling (which 2833 // may not help us) 2834 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2835 if (exact_kls != NULL) {// no cast failures here 2836 if (require_klass == NULL || 2837 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2838 // If we narrow the type to match what the type profile sees or 2839 // the speculative type, we can then remove the rest of the 2840 // cast. 2841 // This is a win, even if the exact_kls is very specific, 2842 // because downstream operations, such as method calls, 2843 // will often benefit from the sharper type. 2844 Node* exact_obj = not_null_obj; // will get updated in place... 2845 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2846 &exact_obj); 2847 { PreserveJVMState pjvms(this); 2848 set_control(slow_ctl); 2849 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2850 } 2851 if (safe_for_replace) { 2852 replace_in_map(not_null_obj, exact_obj); 2853 } 2854 return exact_obj; 2855 } 2856 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2857 } 2858 2859 return NULL; 2860 } 2861 2862 /** 2863 * Cast obj to type and emit guard unless we had too many traps here 2864 * already 2865 * 2866 * @param obj node being casted 2867 * @param type type to cast the node to 2868 * @param not_null true if we know node cannot be null 2869 */ 2870 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2871 ciKlass* type, 2872 bool not_null) { 2873 if (stopped()) { 2874 return obj; 2875 } 2876 2877 // type == NULL if profiling tells us this object is always null 2878 if (type != NULL) { 2879 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2880 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2881 2882 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2883 !too_many_traps(class_reason) && 2884 !too_many_recompiles(class_reason)) { 2885 Node* not_null_obj = NULL; 2886 // not_null is true if we know the object is not null and 2887 // there's no need for a null check 2888 if (!not_null) { 2889 Node* null_ctl = top(); 2890 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2891 assert(null_ctl->is_top(), "no null control here"); 2892 } else { 2893 not_null_obj = obj; 2894 } 2895 2896 Node* exact_obj = not_null_obj; 2897 ciKlass* exact_kls = type; 2898 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2899 &exact_obj); 2900 { 2901 PreserveJVMState pjvms(this); 2902 set_control(slow_ctl); 2903 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2904 } 2905 replace_in_map(not_null_obj, exact_obj); 2906 obj = exact_obj; 2907 } 2908 } else { 2909 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2910 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2911 Node* exact_obj = null_assert(obj); 2912 replace_in_map(obj, exact_obj); 2913 obj = exact_obj; 2914 } 2915 } 2916 return obj; 2917 } 2918 2919 //-------------------------------gen_instanceof-------------------------------- 2920 // Generate an instance-of idiom. Used by both the instance-of bytecode 2921 // and the reflective instance-of call. 2922 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2923 kill_dead_locals(); // Benefit all the uncommon traps 2924 assert( !stopped(), "dead parse path should be checked in callers" ); 2925 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2926 "must check for not-null not-dead klass in callers"); 2927 2928 // Make the merge point 2929 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2930 RegionNode* region = new RegionNode(PATH_LIMIT); 2931 Node* phi = new PhiNode(region, TypeInt::BOOL); 2932 C->set_has_split_ifs(true); // Has chance for split-if optimization 2933 2934 ciProfileData* data = NULL; 2935 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2936 data = method()->method_data()->bci_to_data(bci()); 2937 } 2938 bool speculative_not_null = false; 2939 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2940 && seems_never_null(obj, data, speculative_not_null)); 2941 2942 // Null check; get casted pointer; set region slot 3 2943 Node* null_ctl = top(); 2944 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2945 2946 // If not_null_obj is dead, only null-path is taken 2947 if (stopped()) { // Doing instance-of on a NULL? 2948 set_control(null_ctl); 2949 return intcon(0); 2950 } 2951 region->init_req(_null_path, null_ctl); 2952 phi ->init_req(_null_path, intcon(0)); // Set null path value 2953 if (null_ctl == top()) { 2954 // Do this eagerly, so that pattern matches like is_diamond_phi 2955 // will work even during parsing. 2956 assert(_null_path == PATH_LIMIT-1, "delete last"); 2957 region->del_req(_null_path); 2958 phi ->del_req(_null_path); 2959 } 2960 2961 // Do we know the type check always succeed? 2962 bool known_statically = false; 2963 if (_gvn.type(superklass)->singleton()) { 2964 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2965 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2966 if (subk != NULL && subk->is_loaded()) { 2967 int static_res = C->static_subtype_check(superk, subk); 2968 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2969 } 2970 } 2971 2972 if (known_statically && UseTypeSpeculation) { 2973 // If we know the type check always succeeds then we don't use the 2974 // profiling data at this bytecode. Don't lose it, feed it to the 2975 // type system as a speculative type. 2976 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2977 } else { 2978 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2979 // We may not have profiling here or it may not help us. If we 2980 // have a speculative type use it to perform an exact cast. 2981 ciKlass* spec_obj_type = obj_type->speculative_type(); 2982 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2983 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2984 if (stopped()) { // Profile disagrees with this path. 2985 set_control(null_ctl); // Null is the only remaining possibility. 2986 return intcon(0); 2987 } 2988 if (cast_obj != NULL) { 2989 not_null_obj = cast_obj; 2990 } 2991 } 2992 } 2993 2994 // Load the object's klass 2995 Node* obj_klass = load_object_klass(not_null_obj); 2996 2997 // Generate the subtype check 2998 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2999 3000 // Plug in the success path to the general merge in slot 1. 3001 region->init_req(_obj_path, control()); 3002 phi ->init_req(_obj_path, intcon(1)); 3003 3004 // Plug in the failing path to the general merge in slot 2. 3005 region->init_req(_fail_path, not_subtype_ctrl); 3006 phi ->init_req(_fail_path, intcon(0)); 3007 3008 // Return final merged results 3009 set_control( _gvn.transform(region) ); 3010 record_for_igvn(region); 3011 return _gvn.transform(phi); 3012 } 3013 3014 //-------------------------------gen_checkcast--------------------------------- 3015 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3016 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3017 // uncommon-trap paths work. Adjust stack after this call. 3018 // If failure_control is supplied and not null, it is filled in with 3019 // the control edge for the cast failure. Otherwise, an appropriate 3020 // uncommon trap or exception is thrown. 3021 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3022 Node* *failure_control) { 3023 kill_dead_locals(); // Benefit all the uncommon traps 3024 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3025 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3026 3027 // Fast cutout: Check the case that the cast is vacuously true. 3028 // This detects the common cases where the test will short-circuit 3029 // away completely. We do this before we perform the null check, 3030 // because if the test is going to turn into zero code, we don't 3031 // want a residual null check left around. (Causes a slowdown, 3032 // for example, in some objArray manipulations, such as a[i]=a[j].) 3033 if (tk->singleton()) { 3034 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3035 if (objtp != NULL && objtp->klass() != NULL) { 3036 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3037 case Compile::SSC_always_true: 3038 // If we know the type check always succeed then we don't use 3039 // the profiling data at this bytecode. Don't lose it, feed it 3040 // to the type system as a speculative type. 3041 return record_profiled_receiver_for_speculation(obj); 3042 case Compile::SSC_always_false: 3043 // It needs a null check because a null will *pass* the cast check. 3044 // A non-null value will always produce an exception. 3045 return null_assert(obj); 3046 } 3047 } 3048 } 3049 3050 ciProfileData* data = NULL; 3051 bool safe_for_replace = false; 3052 if (failure_control == NULL) { // use MDO in regular case only 3053 assert(java_bc() == Bytecodes::_aastore || 3054 java_bc() == Bytecodes::_checkcast, 3055 "interpreter profiles type checks only for these BCs"); 3056 data = method()->method_data()->bci_to_data(bci()); 3057 safe_for_replace = true; 3058 } 3059 3060 // Make the merge point 3061 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3062 RegionNode* region = new RegionNode(PATH_LIMIT); 3063 Node* phi = new PhiNode(region, toop); 3064 C->set_has_split_ifs(true); // Has chance for split-if optimization 3065 3066 // Use null-cast information if it is available 3067 bool speculative_not_null = false; 3068 bool never_see_null = ((failure_control == NULL) // regular case only 3069 && seems_never_null(obj, data, speculative_not_null)); 3070 3071 // Null check; get casted pointer; set region slot 3 3072 Node* null_ctl = top(); 3073 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3074 3075 // If not_null_obj is dead, only null-path is taken 3076 if (stopped()) { // Doing instance-of on a NULL? 3077 set_control(null_ctl); 3078 return null(); 3079 } 3080 region->init_req(_null_path, null_ctl); 3081 phi ->init_req(_null_path, null()); // Set null path value 3082 if (null_ctl == top()) { 3083 // Do this eagerly, so that pattern matches like is_diamond_phi 3084 // will work even during parsing. 3085 assert(_null_path == PATH_LIMIT-1, "delete last"); 3086 region->del_req(_null_path); 3087 phi ->del_req(_null_path); 3088 } 3089 3090 Node* cast_obj = NULL; 3091 if (tk->klass_is_exact()) { 3092 // The following optimization tries to statically cast the speculative type of the object 3093 // (for example obtained during profiling) to the type of the superklass and then do a 3094 // dynamic check that the type of the object is what we expect. To work correctly 3095 // for checkcast and aastore the type of superklass should be exact. 3096 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3097 // We may not have profiling here or it may not help us. If we have 3098 // a speculative type use it to perform an exact cast. 3099 ciKlass* spec_obj_type = obj_type->speculative_type(); 3100 if (spec_obj_type != NULL || data != NULL) { 3101 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3102 if (cast_obj != NULL) { 3103 if (failure_control != NULL) // failure is now impossible 3104 (*failure_control) = top(); 3105 // adjust the type of the phi to the exact klass: 3106 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3107 } 3108 } 3109 } 3110 3111 if (cast_obj == NULL) { 3112 // Load the object's klass 3113 Node* obj_klass = load_object_klass(not_null_obj); 3114 3115 // Generate the subtype check 3116 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3117 3118 // Plug in success path into the merge 3119 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3120 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3121 if (failure_control == NULL) { 3122 if (not_subtype_ctrl != top()) { // If failure is possible 3123 PreserveJVMState pjvms(this); 3124 set_control(not_subtype_ctrl); 3125 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3126 } 3127 } else { 3128 (*failure_control) = not_subtype_ctrl; 3129 } 3130 } 3131 3132 region->init_req(_obj_path, control()); 3133 phi ->init_req(_obj_path, cast_obj); 3134 3135 // A merge of NULL or Casted-NotNull obj 3136 Node* res = _gvn.transform(phi); 3137 3138 // Note I do NOT always 'replace_in_map(obj,result)' here. 3139 // if( tk->klass()->can_be_primary_super() ) 3140 // This means that if I successfully store an Object into an array-of-String 3141 // I 'forget' that the Object is really now known to be a String. I have to 3142 // do this because we don't have true union types for interfaces - if I store 3143 // a Baz into an array-of-Interface and then tell the optimizer it's an 3144 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3145 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3146 // replace_in_map( obj, res ); 3147 3148 // Return final merged results 3149 set_control( _gvn.transform(region) ); 3150 record_for_igvn(region); 3151 return res; 3152 } 3153 3154 //------------------------------next_monitor----------------------------------- 3155 // What number should be given to the next monitor? 3156 int GraphKit::next_monitor() { 3157 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3158 int next = current + C->sync_stack_slots(); 3159 // Keep the toplevel high water mark current: 3160 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3161 return current; 3162 } 3163 3164 //------------------------------insert_mem_bar--------------------------------- 3165 // Memory barrier to avoid floating things around 3166 // The membar serves as a pinch point between both control and all memory slices. 3167 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3168 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3169 mb->init_req(TypeFunc::Control, control()); 3170 mb->init_req(TypeFunc::Memory, reset_memory()); 3171 Node* membar = _gvn.transform(mb); 3172 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3173 set_all_memory_call(membar); 3174 return membar; 3175 } 3176 3177 //-------------------------insert_mem_bar_volatile---------------------------- 3178 // Memory barrier to avoid floating things around 3179 // The membar serves as a pinch point between both control and memory(alias_idx). 3180 // If you want to make a pinch point on all memory slices, do not use this 3181 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3182 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3183 // When Parse::do_put_xxx updates a volatile field, it appends a series 3184 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3185 // The first membar is on the same memory slice as the field store opcode. 3186 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3187 // All the other membars (for other volatile slices, including AliasIdxBot, 3188 // which stands for all unknown volatile slices) are control-dependent 3189 // on the first membar. This prevents later volatile loads or stores 3190 // from sliding up past the just-emitted store. 3191 3192 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3193 mb->set_req(TypeFunc::Control,control()); 3194 if (alias_idx == Compile::AliasIdxBot) { 3195 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3196 } else { 3197 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3198 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3199 } 3200 Node* membar = _gvn.transform(mb); 3201 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3202 if (alias_idx == Compile::AliasIdxBot) { 3203 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3204 } else { 3205 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3206 } 3207 return membar; 3208 } 3209 3210 //------------------------------shared_lock------------------------------------ 3211 // Emit locking code. 3212 FastLockNode* GraphKit::shared_lock(Node* obj) { 3213 // bci is either a monitorenter bc or InvocationEntryBci 3214 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3215 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3216 3217 if( !GenerateSynchronizationCode ) 3218 return NULL; // Not locking things? 3219 if (stopped()) // Dead monitor? 3220 return NULL; 3221 3222 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3223 3224 // Box the stack location 3225 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3226 Node* mem = reset_memory(); 3227 3228 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3229 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3230 // Create the counters for this fast lock. 3231 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3232 } 3233 3234 // Create the rtm counters for this fast lock if needed. 3235 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3236 3237 // Add monitor to debug info for the slow path. If we block inside the 3238 // slow path and de-opt, we need the monitor hanging around 3239 map()->push_monitor( flock ); 3240 3241 const TypeFunc *tf = LockNode::lock_type(); 3242 LockNode *lock = new LockNode(C, tf); 3243 3244 lock->init_req( TypeFunc::Control, control() ); 3245 lock->init_req( TypeFunc::Memory , mem ); 3246 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3247 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3248 lock->init_req( TypeFunc::ReturnAdr, top() ); 3249 3250 lock->init_req(TypeFunc::Parms + 0, obj); 3251 lock->init_req(TypeFunc::Parms + 1, box); 3252 lock->init_req(TypeFunc::Parms + 2, flock); 3253 add_safepoint_edges(lock); 3254 3255 lock = _gvn.transform( lock )->as_Lock(); 3256 3257 // lock has no side-effects, sets few values 3258 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3259 3260 insert_mem_bar(Op_MemBarAcquireLock); 3261 3262 // Add this to the worklist so that the lock can be eliminated 3263 record_for_igvn(lock); 3264 3265 #ifndef PRODUCT 3266 if (PrintLockStatistics) { 3267 // Update the counter for this lock. Don't bother using an atomic 3268 // operation since we don't require absolute accuracy. 3269 lock->create_lock_counter(map()->jvms()); 3270 increment_counter(lock->counter()->addr()); 3271 } 3272 #endif 3273 3274 return flock; 3275 } 3276 3277 3278 //------------------------------shared_unlock---------------------------------- 3279 // Emit unlocking code. 3280 void GraphKit::shared_unlock(Node* box, Node* obj) { 3281 // bci is either a monitorenter bc or InvocationEntryBci 3282 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3283 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3284 3285 if( !GenerateSynchronizationCode ) 3286 return; 3287 if (stopped()) { // Dead monitor? 3288 map()->pop_monitor(); // Kill monitor from debug info 3289 return; 3290 } 3291 3292 // Memory barrier to avoid floating things down past the locked region 3293 insert_mem_bar(Op_MemBarReleaseLock); 3294 3295 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3296 UnlockNode *unlock = new UnlockNode(C, tf); 3297 #ifdef ASSERT 3298 unlock->set_dbg_jvms(sync_jvms()); 3299 #endif 3300 uint raw_idx = Compile::AliasIdxRaw; 3301 unlock->init_req( TypeFunc::Control, control() ); 3302 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3303 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3304 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3305 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3306 3307 unlock->init_req(TypeFunc::Parms + 0, obj); 3308 unlock->init_req(TypeFunc::Parms + 1, box); 3309 unlock = _gvn.transform(unlock)->as_Unlock(); 3310 3311 Node* mem = reset_memory(); 3312 3313 // unlock has no side-effects, sets few values 3314 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3315 3316 // Kill monitor from debug info 3317 map()->pop_monitor( ); 3318 } 3319 3320 //-------------------------------get_layout_helper----------------------------- 3321 // If the given klass is a constant or known to be an array, 3322 // fetch the constant layout helper value into constant_value 3323 // and return (Node*)NULL. Otherwise, load the non-constant 3324 // layout helper value, and return the node which represents it. 3325 // This two-faced routine is useful because allocation sites 3326 // almost always feature constant types. 3327 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3328 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3329 if (!StressReflectiveCode && inst_klass != NULL) { 3330 ciKlass* klass = inst_klass->klass(); 3331 bool xklass = inst_klass->klass_is_exact(); 3332 if (xklass || klass->is_array_klass()) { 3333 jint lhelper = klass->layout_helper(); 3334 if (lhelper != Klass::_lh_neutral_value) { 3335 constant_value = lhelper; 3336 return (Node*) NULL; 3337 } 3338 } 3339 } 3340 constant_value = Klass::_lh_neutral_value; // put in a known value 3341 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3342 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3343 } 3344 3345 // We just put in an allocate/initialize with a big raw-memory effect. 3346 // Hook selected additional alias categories on the initialization. 3347 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3348 MergeMemNode* init_in_merge, 3349 Node* init_out_raw) { 3350 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3351 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3352 3353 Node* prevmem = kit.memory(alias_idx); 3354 init_in_merge->set_memory_at(alias_idx, prevmem); 3355 kit.set_memory(init_out_raw, alias_idx); 3356 } 3357 3358 //---------------------------set_output_for_allocation------------------------- 3359 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3360 const TypeOopPtr* oop_type, 3361 bool deoptimize_on_exception) { 3362 int rawidx = Compile::AliasIdxRaw; 3363 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3364 add_safepoint_edges(alloc); 3365 Node* allocx = _gvn.transform(alloc); 3366 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3367 // create memory projection for i_o 3368 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3369 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3370 3371 // create a memory projection as for the normal control path 3372 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3373 set_memory(malloc, rawidx); 3374 3375 // a normal slow-call doesn't change i_o, but an allocation does 3376 // we create a separate i_o projection for the normal control path 3377 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3378 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3379 3380 // put in an initialization barrier 3381 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3382 rawoop)->as_Initialize(); 3383 assert(alloc->initialization() == init, "2-way macro link must work"); 3384 assert(init ->allocation() == alloc, "2-way macro link must work"); 3385 { 3386 // Extract memory strands which may participate in the new object's 3387 // initialization, and source them from the new InitializeNode. 3388 // This will allow us to observe initializations when they occur, 3389 // and link them properly (as a group) to the InitializeNode. 3390 assert(init->in(InitializeNode::Memory) == malloc, ""); 3391 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3392 init->set_req(InitializeNode::Memory, minit_in); 3393 record_for_igvn(minit_in); // fold it up later, if possible 3394 Node* minit_out = memory(rawidx); 3395 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3396 if (oop_type->isa_aryptr()) { 3397 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3398 int elemidx = C->get_alias_index(telemref); 3399 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3400 } else if (oop_type->isa_instptr() || oop_type->isa_valuetypeptr()) { 3401 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3402 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3403 ciField* field = ik->nonstatic_field_at(i); 3404 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3405 continue; // do not bother to track really large numbers of fields 3406 // Find (or create) the alias category for this field: 3407 int fieldidx = C->alias_type(field)->index(); 3408 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3409 } 3410 } 3411 } 3412 3413 // Cast raw oop to the real thing... 3414 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3415 javaoop = _gvn.transform(javaoop); 3416 C->set_recent_alloc(control(), javaoop); 3417 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3418 3419 #ifdef ASSERT 3420 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3421 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3422 "Ideal_allocation works"); 3423 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3424 "Ideal_allocation works"); 3425 if (alloc->is_AllocateArray()) { 3426 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3427 "Ideal_allocation works"); 3428 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3429 "Ideal_allocation works"); 3430 } else { 3431 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3432 } 3433 } 3434 #endif //ASSERT 3435 3436 return javaoop; 3437 } 3438 3439 //---------------------------new_instance-------------------------------------- 3440 // This routine takes a klass_node which may be constant (for a static type) 3441 // or may be non-constant (for reflective code). It will work equally well 3442 // for either, and the graph will fold nicely if the optimizer later reduces 3443 // the type to a constant. 3444 // The optional arguments are for specialized use by intrinsics: 3445 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3446 // - If 'return_size_val', report the the total object size to the caller. 3447 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3448 Node* GraphKit::new_instance(Node* klass_node, 3449 Node* extra_slow_test, 3450 Node* *return_size_val, 3451 bool deoptimize_on_exception) { 3452 // Compute size in doublewords 3453 // The size is always an integral number of doublewords, represented 3454 // as a positive bytewise size stored in the klass's layout_helper. 3455 // The layout_helper also encodes (in a low bit) the need for a slow path. 3456 jint layout_con = Klass::_lh_neutral_value; 3457 Node* layout_val = get_layout_helper(klass_node, layout_con); 3458 int layout_is_con = (layout_val == NULL); 3459 3460 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3461 // Generate the initial go-slow test. It's either ALWAYS (return a 3462 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3463 // case) a computed value derived from the layout_helper. 3464 Node* initial_slow_test = NULL; 3465 if (layout_is_con) { 3466 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3467 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3468 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3469 } else { // reflective case 3470 // This reflective path is used by Unsafe.allocateInstance. 3471 // (It may be stress-tested by specifying StressReflectiveCode.) 3472 // Basically, we want to get into the VM is there's an illegal argument. 3473 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3474 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3475 if (extra_slow_test != intcon(0)) { 3476 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3477 } 3478 // (Macro-expander will further convert this to a Bool, if necessary.) 3479 } 3480 3481 // Find the size in bytes. This is easy; it's the layout_helper. 3482 // The size value must be valid even if the slow path is taken. 3483 Node* size = NULL; 3484 if (layout_is_con) { 3485 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3486 } else { // reflective case 3487 // This reflective path is used by clone and Unsafe.allocateInstance. 3488 size = ConvI2X(layout_val); 3489 3490 // Clear the low bits to extract layout_helper_size_in_bytes: 3491 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3492 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3493 size = _gvn.transform( new AndXNode(size, mask) ); 3494 } 3495 if (return_size_val != NULL) { 3496 (*return_size_val) = size; 3497 } 3498 3499 // This is a precise notnull oop of the klass. 3500 // (Actually, it need not be precise if this is a reflective allocation.) 3501 // It's what we cast the result to. 3502 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3503 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3504 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3505 3506 // Now generate allocation code 3507 3508 // The entire memory state is needed for slow path of the allocation 3509 // since GC and deoptimization can happen. 3510 Node *mem = reset_memory(); 3511 set_all_memory(mem); // Create new memory state 3512 3513 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3514 control(), mem, i_o(), 3515 size, klass_node, 3516 initial_slow_test); 3517 3518 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3519 } 3520 3521 //-------------------------------new_array------------------------------------- 3522 // helper for newarray, anewarray and vnewarray 3523 // The 'length' parameter is (obviously) the length of the array. 3524 // See comments on new_instance for the meaning of the other arguments. 3525 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3526 Node* length, // number of array elements 3527 int nargs, // number of arguments to push back for uncommon trap 3528 Node* *return_size_val, 3529 bool deoptimize_on_exception) { 3530 jint layout_con = Klass::_lh_neutral_value; 3531 Node* layout_val = get_layout_helper(klass_node, layout_con); 3532 int layout_is_con = (layout_val == NULL); 3533 3534 if (!layout_is_con && !StressReflectiveCode && 3535 !too_many_traps(Deoptimization::Reason_class_check)) { 3536 // This is a reflective array creation site. 3537 // Optimistically assume that it is a subtype of Object[], 3538 // so that we can fold up all the address arithmetic. 3539 layout_con = Klass::array_layout_helper(T_OBJECT); 3540 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3541 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3542 { BuildCutout unless(this, bol_lh, PROB_MAX); 3543 inc_sp(nargs); 3544 uncommon_trap(Deoptimization::Reason_class_check, 3545 Deoptimization::Action_maybe_recompile); 3546 } 3547 layout_val = NULL; 3548 layout_is_con = true; 3549 } 3550 3551 // Generate the initial go-slow test. Make sure we do not overflow 3552 // if length is huge (near 2Gig) or negative! We do not need 3553 // exact double-words here, just a close approximation of needed 3554 // double-words. We can't add any offset or rounding bits, lest we 3555 // take a size -1 of bytes and make it positive. Use an unsigned 3556 // compare, so negative sizes look hugely positive. 3557 int fast_size_limit = FastAllocateSizeLimit; 3558 if (layout_is_con) { 3559 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3560 // Increase the size limit if we have exact knowledge of array type. 3561 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3562 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3563 } 3564 3565 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3566 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3567 3568 // --- Size Computation --- 3569 // array_size = round_to_heap(array_header + (length << elem_shift)); 3570 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3571 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3572 // The rounding mask is strength-reduced, if possible. 3573 int round_mask = MinObjAlignmentInBytes - 1; 3574 Node* header_size = NULL; 3575 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3576 // (T_BYTE has the weakest alignment and size restrictions...) 3577 if (layout_is_con) { 3578 int hsize = Klass::layout_helper_header_size(layout_con); 3579 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3580 BasicType etype = Klass::layout_helper_element_type(layout_con); 3581 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con); 3582 if ((round_mask & ~right_n_bits(eshift)) == 0) 3583 round_mask = 0; // strength-reduce it if it goes away completely 3584 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3585 assert(header_size_min <= hsize, "generic minimum is smallest"); 3586 header_size_min = hsize; 3587 header_size = intcon(hsize + round_mask); 3588 } else { 3589 Node* hss = intcon(Klass::_lh_header_size_shift); 3590 Node* hsm = intcon(Klass::_lh_header_size_mask); 3591 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3592 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3593 Node* mask = intcon(round_mask); 3594 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3595 } 3596 3597 Node* elem_shift = NULL; 3598 if (layout_is_con) { 3599 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3600 if (eshift != 0) 3601 elem_shift = intcon(eshift); 3602 } else { 3603 // There is no need to mask or shift this value. 3604 // The semantics of LShiftINode include an implicit mask to 0x1F. 3605 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3606 elem_shift = layout_val; 3607 } 3608 3609 // Transition to native address size for all offset calculations: 3610 Node* lengthx = ConvI2X(length); 3611 Node* headerx = ConvI2X(header_size); 3612 #ifdef _LP64 3613 { const TypeInt* tilen = _gvn.find_int_type(length); 3614 if (tilen != NULL && tilen->_lo < 0) { 3615 // Add a manual constraint to a positive range. Cf. array_element_address. 3616 jint size_max = fast_size_limit; 3617 if (size_max > tilen->_hi) size_max = tilen->_hi; 3618 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3619 3620 // Only do a narrow I2L conversion if the range check passed. 3621 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3622 _gvn.transform(iff); 3623 RegionNode* region = new RegionNode(3); 3624 _gvn.set_type(region, Type::CONTROL); 3625 lengthx = new PhiNode(region, TypeLong::LONG); 3626 _gvn.set_type(lengthx, TypeLong::LONG); 3627 3628 // Range check passed. Use ConvI2L node with narrow type. 3629 Node* passed = IfFalse(iff); 3630 region->init_req(1, passed); 3631 // Make I2L conversion control dependent to prevent it from 3632 // floating above the range check during loop optimizations. 3633 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3634 3635 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3636 region->init_req(2, IfTrue(iff)); 3637 lengthx->init_req(2, ConvI2X(length)); 3638 3639 set_control(region); 3640 record_for_igvn(region); 3641 record_for_igvn(lengthx); 3642 } 3643 } 3644 #endif 3645 3646 // Combine header size (plus rounding) and body size. Then round down. 3647 // This computation cannot overflow, because it is used only in two 3648 // places, one where the length is sharply limited, and the other 3649 // after a successful allocation. 3650 Node* abody = lengthx; 3651 if (elem_shift != NULL) 3652 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3653 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3654 if (round_mask != 0) { 3655 Node* mask = MakeConX(~round_mask); 3656 size = _gvn.transform( new AndXNode(size, mask) ); 3657 } 3658 // else if round_mask == 0, the size computation is self-rounding 3659 3660 if (return_size_val != NULL) { 3661 // This is the size 3662 (*return_size_val) = size; 3663 } 3664 3665 // Now generate allocation code 3666 3667 // The entire memory state is needed for slow path of the allocation 3668 // since GC and deoptimization can happen. 3669 Node *mem = reset_memory(); 3670 set_all_memory(mem); // Create new memory state 3671 3672 if (initial_slow_test->is_Bool()) { 3673 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3674 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3675 } 3676 3677 // Create the AllocateArrayNode and its result projections 3678 AllocateArrayNode* alloc 3679 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3680 control(), mem, i_o(), 3681 size, klass_node, 3682 initial_slow_test, 3683 length); 3684 3685 // Cast to correct type. Note that the klass_node may be constant or not, 3686 // and in the latter case the actual array type will be inexact also. 3687 // (This happens via a non-constant argument to inline_native_newArray.) 3688 // In any case, the value of klass_node provides the desired array type. 3689 const TypeInt* length_type = _gvn.find_int_type(length); 3690 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3691 if (ary_type->isa_aryptr() && length_type != NULL) { 3692 // Try to get a better type than POS for the size 3693 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3694 } 3695 3696 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3697 3698 // Cast length on remaining path to be as narrow as possible 3699 if (map()->find_edge(length) >= 0) { 3700 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3701 if (ccast != length) { 3702 _gvn.set_type_bottom(ccast); 3703 record_for_igvn(ccast); 3704 replace_in_map(length, ccast); 3705 } 3706 } 3707 3708 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 3709 ciKlass* elem_klass = ary_ptr != NULL ? ary_ptr->klass()->as_array_klass()->element_klass() : NULL; 3710 if (elem_klass != NULL && elem_klass->is_valuetype()) { 3711 ciValueKlass* vk = elem_klass->as_value_klass(); 3712 if (!vk->flatten_array()) { 3713 // Non-flattened value type arrays need to be initialized with default value type oops 3714 initialize_value_type_array(javaoop, length, elem_klass->as_value_klass(), nargs); 3715 InitializeNode* init = alloc->initialization(); 3716 init->set_complete_with_arraycopy(); 3717 } 3718 } 3719 3720 return javaoop; 3721 } 3722 3723 void GraphKit::initialize_value_type_array(Node* array, Node* length, ciValueKlass* vk, int nargs) { 3724 // Check for zero length 3725 Node* null_ctl = top(); 3726 null_check_common(length, T_INT, false, &null_ctl, false); 3727 if (stopped()) { 3728 set_control(null_ctl); // Always zero 3729 return; 3730 } 3731 3732 // Prepare for merging control and IO 3733 RegionNode* res_ctl = new RegionNode(3); 3734 res_ctl->init_req(1, null_ctl); 3735 gvn().set_type(res_ctl, Type::CONTROL); 3736 record_for_igvn(res_ctl); 3737 Node* res_io = PhiNode::make(res_ctl, i_o(), Type::ABIO); 3738 gvn().set_type(res_io, Type::ABIO); 3739 record_for_igvn(res_io); 3740 3741 // TODO comment 3742 SafePointNode* loop_map = NULL; 3743 { 3744 PreserveJVMState pjvms(this); 3745 // Create default value type and store it to memory 3746 Node* oop = ValueTypeNode::make_default(gvn(), vk); 3747 oop = oop->as_ValueType()->store_to_memory(this); 3748 3749 length = SubI(length, intcon(1)); 3750 add_predicate(nargs); 3751 RegionNode* loop = new RegionNode(3); 3752 loop->init_req(1, control()); 3753 gvn().set_type(loop, Type::CONTROL); 3754 record_for_igvn(loop); 3755 3756 Node* index = new PhiNode(loop, TypeInt::INT); 3757 index->init_req(1, intcon(0)); 3758 gvn().set_type(index, TypeInt::INT); 3759 record_for_igvn(index); 3760 3761 // TODO explain why we need to capture all memory 3762 PhiNode* mem = new PhiNode(loop, Type::MEMORY, TypePtr::BOTTOM); 3763 mem->init_req(1, reset_memory()); 3764 gvn().set_type(mem, Type::MEMORY); 3765 record_for_igvn(mem); 3766 set_control(loop); 3767 set_all_memory(mem); 3768 // Initialize array element 3769 Node* adr = array_element_address(array, index, T_OBJECT); 3770 const TypeOopPtr* elemtype = TypeValueTypePtr::make(TypePtr::NotNull, vk); 3771 Node* store = store_oop_to_array(control(), array, adr, TypeAryPtr::OOPS, oop, elemtype, T_OBJECT, MemNode::release); 3772 3773 IfNode* iff = create_and_map_if(control(), Bool(CmpI(index, length), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 3774 loop->init_req(2, IfTrue(iff)); 3775 mem->init_req(2, merged_memory()); 3776 index->init_req(2, AddI(index, intcon(1))); 3777 3778 res_ctl->init_req(2, IfFalse(iff)); 3779 res_io->set_req(2, i_o()); 3780 loop_map = stop(); 3781 } 3782 // Set merged control, IO and memory 3783 set_control(res_ctl); 3784 set_i_o(res_io); 3785 merge_memory(loop_map->merged_memory(), res_ctl, 2); 3786 3787 // Transform new memory Phis. 3788 for (MergeMemStream mms(merged_memory()); mms.next_non_empty();) { 3789 Node* phi = mms.memory(); 3790 if (phi->is_Phi() && phi->in(0) == res_ctl) { 3791 mms.set_memory(gvn().transform(phi)); 3792 } 3793 } 3794 } 3795 3796 // The following "Ideal_foo" functions are placed here because they recognize 3797 // the graph shapes created by the functions immediately above. 3798 3799 //---------------------------Ideal_allocation---------------------------------- 3800 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3801 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3802 if (ptr == NULL) { // reduce dumb test in callers 3803 return NULL; 3804 } 3805 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3806 ptr = ptr->in(1); 3807 if (ptr == NULL) return NULL; 3808 } 3809 // Return NULL for allocations with several casts: 3810 // j.l.reflect.Array.newInstance(jobject, jint) 3811 // Object.clone() 3812 // to keep more precise type from last cast. 3813 if (ptr->is_Proj()) { 3814 Node* allo = ptr->in(0); 3815 if (allo != NULL && allo->is_Allocate()) { 3816 return allo->as_Allocate(); 3817 } 3818 } 3819 // Report failure to match. 3820 return NULL; 3821 } 3822 3823 // Fancy version which also strips off an offset (and reports it to caller). 3824 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3825 intptr_t& offset) { 3826 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3827 if (base == NULL) return NULL; 3828 return Ideal_allocation(base, phase); 3829 } 3830 3831 // Trace Initialize <- Proj[Parm] <- Allocate 3832 AllocateNode* InitializeNode::allocation() { 3833 Node* rawoop = in(InitializeNode::RawAddress); 3834 if (rawoop->is_Proj()) { 3835 Node* alloc = rawoop->in(0); 3836 if (alloc->is_Allocate()) { 3837 return alloc->as_Allocate(); 3838 } 3839 } 3840 return NULL; 3841 } 3842 3843 // Trace Allocate -> Proj[Parm] -> Initialize 3844 InitializeNode* AllocateNode::initialization() { 3845 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3846 if (rawoop == NULL) return NULL; 3847 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3848 Node* init = rawoop->fast_out(i); 3849 if (init->is_Initialize()) { 3850 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3851 return init->as_Initialize(); 3852 } 3853 } 3854 return NULL; 3855 } 3856 3857 //----------------------------- loop predicates --------------------------- 3858 3859 //------------------------------add_predicate_impl---------------------------- 3860 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3861 // Too many traps seen? 3862 if (too_many_traps(reason)) { 3863 #ifdef ASSERT 3864 if (TraceLoopPredicate) { 3865 int tc = C->trap_count(reason); 3866 tty->print("too many traps=%s tcount=%d in ", 3867 Deoptimization::trap_reason_name(reason), tc); 3868 method()->print(); // which method has too many predicate traps 3869 tty->cr(); 3870 } 3871 #endif 3872 // We cannot afford to take more traps here, 3873 // do not generate predicate. 3874 return; 3875 } 3876 3877 Node *cont = _gvn.intcon(1); 3878 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3879 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3880 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3881 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3882 C->add_predicate_opaq(opq); 3883 { 3884 PreserveJVMState pjvms(this); 3885 set_control(iffalse); 3886 inc_sp(nargs); 3887 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3888 } 3889 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3890 set_control(iftrue); 3891 } 3892 3893 //------------------------------add_predicate--------------------------------- 3894 void GraphKit::add_predicate(int nargs) { 3895 if (UseLoopPredicate) { 3896 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3897 } 3898 // loop's limit check predicate should be near the loop. 3899 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3900 } 3901 3902 //----------------------------- store barriers ---------------------------- 3903 #define __ ideal. 3904 3905 void GraphKit::sync_kit(IdealKit& ideal) { 3906 set_all_memory(__ merged_memory()); 3907 set_i_o(__ i_o()); 3908 set_control(__ ctrl()); 3909 } 3910 3911 void GraphKit::final_sync(IdealKit& ideal) { 3912 // Final sync IdealKit and graphKit. 3913 sync_kit(ideal); 3914 } 3915 3916 Node* GraphKit::byte_map_base_node() { 3917 // Get base of card map 3918 CardTableModRefBS* ct = 3919 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3920 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3921 if (ct->byte_map_base != NULL) { 3922 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3923 } else { 3924 return null(); 3925 } 3926 } 3927 3928 // vanilla/CMS post barrier 3929 // Insert a write-barrier store. This is to let generational GC work; we have 3930 // to flag all oop-stores before the next GC point. 3931 void GraphKit::write_barrier_post(Node* oop_store, 3932 Node* obj, 3933 Node* adr, 3934 uint adr_idx, 3935 Node* val, 3936 bool use_precise) { 3937 // No store check needed if we're storing a NULL or an old object 3938 // (latter case is probably a string constant). The concurrent 3939 // mark sweep garbage collector, however, needs to have all nonNull 3940 // oop updates flagged via card-marks. 3941 if (val != NULL && val->is_Con()) { 3942 // must be either an oop or NULL 3943 const Type* t = val->bottom_type(); 3944 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3945 // stores of null never (?) need barriers 3946 return; 3947 } 3948 3949 if (use_ReduceInitialCardMarks() 3950 && obj == just_allocated_object(control())) { 3951 // We can skip marks on a freshly-allocated object in Eden. 3952 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3953 // That routine informs GC to take appropriate compensating steps, 3954 // upon a slow-path allocation, so as to make this card-mark 3955 // elision safe. 3956 return; 3957 } 3958 3959 if (!use_precise) { 3960 // All card marks for a (non-array) instance are in one place: 3961 adr = obj; 3962 } 3963 // (Else it's an array (or unknown), and we want more precise card marks.) 3964 assert(adr != NULL, ""); 3965 3966 IdealKit ideal(this, true); 3967 3968 // Convert the pointer to an int prior to doing math on it 3969 Node* cast = __ CastPX(__ ctrl(), adr); 3970 3971 // Divide by card size 3972 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3973 "Only one we handle so far."); 3974 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3975 3976 // Combine card table base and card offset 3977 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3978 3979 // Get the alias_index for raw card-mark memory 3980 int adr_type = Compile::AliasIdxRaw; 3981 Node* zero = __ ConI(0); // Dirty card value 3982 BasicType bt = T_BYTE; 3983 3984 if (UseConcMarkSweepGC && UseCondCardMark) { 3985 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3986 __ sync_kit(this); 3987 } 3988 3989 if (UseCondCardMark) { 3990 // The classic GC reference write barrier is typically implemented 3991 // as a store into the global card mark table. Unfortunately 3992 // unconditional stores can result in false sharing and excessive 3993 // coherence traffic as well as false transactional aborts. 3994 // UseCondCardMark enables MP "polite" conditional card mark 3995 // stores. In theory we could relax the load from ctrl() to 3996 // no_ctrl, but that doesn't buy much latitude. 3997 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3998 __ if_then(card_val, BoolTest::ne, zero); 3999 } 4000 4001 // Smash zero into card 4002 if( !UseConcMarkSweepGC ) { 4003 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 4004 } else { 4005 // Specialized path for CM store barrier 4006 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 4007 } 4008 4009 if (UseCondCardMark) { 4010 __ end_if(); 4011 } 4012 4013 // Final sync IdealKit and GraphKit. 4014 final_sync(ideal); 4015 } 4016 /* 4017 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 4018 * required by SATB to make sure all objects live at the start of the 4019 * marking are kept alive, all reference updates need to any previous 4020 * reference stored before writing. 4021 * 4022 * If the previous value is NULL there is no need to save the old value. 4023 * References that are NULL are filtered during runtime by the barrier 4024 * code to avoid unnecessary queuing. 4025 * 4026 * However in the case of newly allocated objects it might be possible to 4027 * prove that the reference about to be overwritten is NULL during compile 4028 * time and avoid adding the barrier code completely. 4029 * 4030 * The compiler needs to determine that the object in which a field is about 4031 * to be written is newly allocated, and that no prior store to the same field 4032 * has happened since the allocation. 4033 * 4034 * Returns true if the pre-barrier can be removed 4035 */ 4036 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 4037 BasicType bt, uint adr_idx) { 4038 intptr_t offset = 0; 4039 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4040 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4041 4042 if (offset == Type::OffsetBot) { 4043 return false; // cannot unalias unless there are precise offsets 4044 } 4045 4046 if (alloc == NULL) { 4047 return false; // No allocation found 4048 } 4049 4050 intptr_t size_in_bytes = type2aelembytes(bt); 4051 4052 Node* mem = memory(adr_idx); // start searching here... 4053 4054 for (int cnt = 0; cnt < 50; cnt++) { 4055 4056 if (mem->is_Store()) { 4057 4058 Node* st_adr = mem->in(MemNode::Address); 4059 intptr_t st_offset = 0; 4060 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 4061 4062 if (st_base == NULL) { 4063 break; // inscrutable pointer 4064 } 4065 4066 // Break we have found a store with same base and offset as ours so break 4067 if (st_base == base && st_offset == offset) { 4068 break; 4069 } 4070 4071 if (st_offset != offset && st_offset != Type::OffsetBot) { 4072 const int MAX_STORE = BytesPerLong; 4073 if (st_offset >= offset + size_in_bytes || 4074 st_offset <= offset - MAX_STORE || 4075 st_offset <= offset - mem->as_Store()->memory_size()) { 4076 // Success: The offsets are provably independent. 4077 // (You may ask, why not just test st_offset != offset and be done? 4078 // The answer is that stores of different sizes can co-exist 4079 // in the same sequence of RawMem effects. We sometimes initialize 4080 // a whole 'tile' of array elements with a single jint or jlong.) 4081 mem = mem->in(MemNode::Memory); 4082 continue; // advance through independent store memory 4083 } 4084 } 4085 4086 if (st_base != base 4087 && MemNode::detect_ptr_independence(base, alloc, st_base, 4088 AllocateNode::Ideal_allocation(st_base, phase), 4089 phase)) { 4090 // Success: The bases are provably independent. 4091 mem = mem->in(MemNode::Memory); 4092 continue; // advance through independent store memory 4093 } 4094 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4095 4096 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4097 AllocateNode* st_alloc = st_init->allocation(); 4098 4099 // Make sure that we are looking at the same allocation site. 4100 // The alloc variable is guaranteed to not be null here from earlier check. 4101 if (alloc == st_alloc) { 4102 // Check that the initialization is storing NULL so that no previous store 4103 // has been moved up and directly write a reference 4104 Node* captured_store = st_init->find_captured_store(offset, 4105 type2aelembytes(T_OBJECT), 4106 phase); 4107 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4108 return true; 4109 } 4110 } 4111 } 4112 4113 // Unless there is an explicit 'continue', we must bail out here, 4114 // because 'mem' is an inscrutable memory state (e.g., a call). 4115 break; 4116 } 4117 4118 return false; 4119 } 4120 4121 // G1 pre/post barriers 4122 void GraphKit::g1_write_barrier_pre(bool do_load, 4123 Node* obj, 4124 Node* adr, 4125 uint alias_idx, 4126 Node* val, 4127 const TypeOopPtr* val_type, 4128 Node* pre_val, 4129 BasicType bt) { 4130 4131 // Some sanity checks 4132 // Note: val is unused in this routine. 4133 4134 if (do_load) { 4135 // We need to generate the load of the previous value 4136 assert(obj != NULL, "must have a base"); 4137 assert(adr != NULL, "where are loading from?"); 4138 assert(pre_val == NULL, "loaded already?"); 4139 assert(val_type != NULL, "need a type"); 4140 4141 if (use_ReduceInitialCardMarks() 4142 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4143 return; 4144 } 4145 4146 } else { 4147 // In this case both val_type and alias_idx are unused. 4148 assert(pre_val != NULL, "must be loaded already"); 4149 // Nothing to be done if pre_val is null. 4150 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4151 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4152 } 4153 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here"); 4154 4155 IdealKit ideal(this, true); 4156 4157 Node* tls = __ thread(); // ThreadLocalStorage 4158 4159 Node* no_ctrl = NULL; 4160 Node* no_base = __ top(); 4161 Node* zero = __ ConI(0); 4162 Node* zeroX = __ ConX(0); 4163 4164 float likely = PROB_LIKELY(0.999); 4165 float unlikely = PROB_UNLIKELY(0.999); 4166 4167 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4168 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4169 4170 // Offsets into the thread 4171 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4172 SATBMarkQueue::byte_offset_of_active()); 4173 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4174 SATBMarkQueue::byte_offset_of_index()); 4175 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4176 SATBMarkQueue::byte_offset_of_buf()); 4177 4178 // Now the actual pointers into the thread 4179 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4180 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4181 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4182 4183 // Now some of the values 4184 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4185 4186 // if (!marking) 4187 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4188 BasicType index_bt = TypeX_X->basic_type(); 4189 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4190 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4191 4192 if (do_load) { 4193 // load original value 4194 // alias_idx correct?? 4195 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4196 } 4197 4198 // if (pre_val != NULL) 4199 __ if_then(pre_val, BoolTest::ne, null()); { 4200 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4201 4202 // is the queue for this thread full? 4203 __ if_then(index, BoolTest::ne, zeroX, likely); { 4204 4205 // decrement the index 4206 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4207 4208 // Now get the buffer location we will log the previous value into and store it 4209 Node *log_addr = __ AddP(no_base, buffer, next_index); 4210 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4211 // update the index 4212 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4213 4214 } __ else_(); { 4215 4216 // logging buffer is full, call the runtime 4217 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4218 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4219 } __ end_if(); // (!index) 4220 } __ end_if(); // (pre_val != NULL) 4221 } __ end_if(); // (!marking) 4222 4223 // Final sync IdealKit and GraphKit. 4224 final_sync(ideal); 4225 } 4226 4227 /* 4228 * G1 similar to any GC with a Young Generation requires a way to keep track of 4229 * references from Old Generation to Young Generation to make sure all live 4230 * objects are found. G1 also requires to keep track of object references 4231 * between different regions to enable evacuation of old regions, which is done 4232 * as part of mixed collections. References are tracked in remembered sets and 4233 * is continuously updated as reference are written to with the help of the 4234 * post-barrier. 4235 * 4236 * To reduce the number of updates to the remembered set the post-barrier 4237 * filters updates to fields in objects located in the Young Generation, 4238 * the same region as the reference, when the NULL is being written or 4239 * if the card is already marked as dirty by an earlier write. 4240 * 4241 * Under certain circumstances it is possible to avoid generating the 4242 * post-barrier completely if it is possible during compile time to prove 4243 * the object is newly allocated and that no safepoint exists between the 4244 * allocation and the store. 4245 * 4246 * In the case of slow allocation the allocation code must handle the barrier 4247 * as part of the allocation in the case the allocated object is not located 4248 * in the nursery, this would happen for humongous objects. This is similar to 4249 * how CMS is required to handle this case, see the comments for the method 4250 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4251 * A deferred card mark is required for these objects and handled in the above 4252 * mentioned methods. 4253 * 4254 * Returns true if the post barrier can be removed 4255 */ 4256 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4257 Node* adr) { 4258 intptr_t offset = 0; 4259 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4260 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4261 4262 if (offset == Type::OffsetBot) { 4263 return false; // cannot unalias unless there are precise offsets 4264 } 4265 4266 if (alloc == NULL) { 4267 return false; // No allocation found 4268 } 4269 4270 // Start search from Store node 4271 Node* mem = store->in(MemNode::Control); 4272 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4273 4274 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4275 AllocateNode* st_alloc = st_init->allocation(); 4276 4277 // Make sure we are looking at the same allocation 4278 if (alloc == st_alloc) { 4279 return true; 4280 } 4281 } 4282 4283 return false; 4284 } 4285 4286 // 4287 // Update the card table and add card address to the queue 4288 // 4289 void GraphKit::g1_mark_card(IdealKit& ideal, 4290 Node* card_adr, 4291 Node* oop_store, 4292 uint oop_alias_idx, 4293 Node* index, 4294 Node* index_adr, 4295 Node* buffer, 4296 const TypeFunc* tf) { 4297 4298 Node* zero = __ ConI(0); 4299 Node* zeroX = __ ConX(0); 4300 Node* no_base = __ top(); 4301 BasicType card_bt = T_BYTE; 4302 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4303 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4304 4305 // Now do the queue work 4306 __ if_then(index, BoolTest::ne, zeroX); { 4307 4308 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4309 Node* log_addr = __ AddP(no_base, buffer, next_index); 4310 4311 // Order, see storeCM. 4312 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4313 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4314 4315 } __ else_(); { 4316 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4317 } __ end_if(); 4318 4319 } 4320 4321 void GraphKit::g1_write_barrier_post(Node* oop_store, 4322 Node* obj, 4323 Node* adr, 4324 uint alias_idx, 4325 Node* val, 4326 BasicType bt, 4327 bool use_precise) { 4328 // If we are writing a NULL then we need no post barrier 4329 4330 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4331 // Must be NULL 4332 const Type* t = val->bottom_type(); 4333 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4334 // No post barrier if writing NULLx 4335 return; 4336 } 4337 4338 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4339 // We can skip marks on a freshly-allocated object in Eden. 4340 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4341 // That routine informs GC to take appropriate compensating steps, 4342 // upon a slow-path allocation, so as to make this card-mark 4343 // elision safe. 4344 return; 4345 } 4346 4347 if (use_ReduceInitialCardMarks() 4348 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4349 return; 4350 } 4351 4352 if (!use_precise) { 4353 // All card marks for a (non-array) instance are in one place: 4354 adr = obj; 4355 } 4356 // (Else it's an array (or unknown), and we want more precise card marks.) 4357 assert(adr != NULL, ""); 4358 4359 IdealKit ideal(this, true); 4360 4361 Node* tls = __ thread(); // ThreadLocalStorage 4362 4363 Node* no_base = __ top(); 4364 float likely = PROB_LIKELY(0.999); 4365 float unlikely = PROB_UNLIKELY(0.999); 4366 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4367 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4368 Node* zeroX = __ ConX(0); 4369 4370 // Get the alias_index for raw card-mark memory 4371 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4372 4373 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4374 4375 // Offsets into the thread 4376 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4377 DirtyCardQueue::byte_offset_of_index()); 4378 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4379 DirtyCardQueue::byte_offset_of_buf()); 4380 4381 // Pointers into the thread 4382 4383 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4384 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4385 4386 // Now some values 4387 // Use ctrl to avoid hoisting these values past a safepoint, which could 4388 // potentially reset these fields in the JavaThread. 4389 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4390 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4391 4392 // Convert the store obj pointer to an int prior to doing math on it 4393 // Must use ctrl to prevent "integerized oop" existing across safepoint 4394 Node* cast = __ CastPX(__ ctrl(), adr); 4395 4396 // Divide pointer by card size 4397 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4398 4399 // Combine card table base and card offset 4400 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4401 4402 // If we know the value being stored does it cross regions? 4403 4404 if (val != NULL) { 4405 // Does the store cause us to cross regions? 4406 4407 // Should be able to do an unsigned compare of region_size instead of 4408 // and extra shift. Do we have an unsigned compare?? 4409 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4410 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4411 4412 // if (xor_res == 0) same region so skip 4413 __ if_then(xor_res, BoolTest::ne, zeroX); { 4414 4415 // No barrier if we are storing a NULL 4416 __ if_then(val, BoolTest::ne, null(), unlikely); { 4417 4418 // Ok must mark the card if not already dirty 4419 4420 // load the original value of the card 4421 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4422 4423 __ if_then(card_val, BoolTest::ne, young_card); { 4424 sync_kit(ideal); 4425 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4426 insert_mem_bar(Op_MemBarVolatile, oop_store); 4427 __ sync_kit(this); 4428 4429 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4430 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4431 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4432 } __ end_if(); 4433 } __ end_if(); 4434 } __ end_if(); 4435 } __ end_if(); 4436 } else { 4437 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4438 // We don't need a barrier here if the destination is a newly allocated object 4439 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4440 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4441 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4442 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4443 __ if_then(card_val, BoolTest::ne, young_card); { 4444 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4445 } __ end_if(); 4446 } 4447 4448 // Final sync IdealKit and GraphKit. 4449 final_sync(ideal); 4450 } 4451 #undef __ 4452 4453 4454 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4455 Node* len = load_array_length(load_String_value(ctrl, str)); 4456 Node* coder = load_String_coder(ctrl, str); 4457 // Divide length by 2 if coder is UTF16 4458 return _gvn.transform(new RShiftINode(len, coder)); 4459 } 4460 4461 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4462 int value_offset = java_lang_String::value_offset_in_bytes(); 4463 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4464 false, NULL, Type::Offset(0)); 4465 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4466 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4467 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4468 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4469 int value_field_idx = C->get_alias_index(value_field_type); 4470 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4471 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4472 // String.value field is known to be @Stable. 4473 if (UseImplicitStableValues) { 4474 load = cast_array_to_stable(load, value_type); 4475 } 4476 return load; 4477 } 4478 4479 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4480 if (!CompactStrings) { 4481 return intcon(java_lang_String::CODER_UTF16); 4482 } 4483 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4484 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4485 false, NULL, Type::Offset(0)); 4486 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4487 int coder_field_idx = C->get_alias_index(coder_field_type); 4488 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4489 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4490 } 4491 4492 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4493 int value_offset = java_lang_String::value_offset_in_bytes(); 4494 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4495 false, NULL, Type::Offset(0)); 4496 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4497 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4498 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4499 } 4500 4501 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4502 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4503 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4504 false, NULL, Type::Offset(0)); 4505 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4506 int coder_field_idx = C->get_alias_index(coder_field_type); 4507 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4508 value, T_BYTE, coder_field_idx, MemNode::unordered); 4509 } 4510 4511 // Capture src and dst memory state with a MergeMemNode 4512 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4513 if (src_type == dst_type) { 4514 // Types are equal, we don't need a MergeMemNode 4515 return memory(src_type); 4516 } 4517 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4518 record_for_igvn(merge); // fold it up later, if possible 4519 int src_idx = C->get_alias_index(src_type); 4520 int dst_idx = C->get_alias_index(dst_type); 4521 merge->set_memory_at(src_idx, memory(src_idx)); 4522 merge->set_memory_at(dst_idx, memory(dst_idx)); 4523 return merge; 4524 } 4525 4526 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4527 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4528 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4529 // If input and output memory types differ, capture both states to preserve 4530 // the dependency between preceding and subsequent loads/stores. 4531 // For example, the following program: 4532 // StoreB 4533 // compress_string 4534 // LoadB 4535 // has this memory graph (use->def): 4536 // LoadB -> compress_string -> CharMem 4537 // ... -> StoreB -> ByteMem 4538 // The intrinsic hides the dependency between LoadB and StoreB, causing 4539 // the load to read from memory not containing the result of the StoreB. 4540 // The correct memory graph should look like this: 4541 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4542 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4543 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4544 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4545 set_memory(res_mem, TypeAryPtr::BYTES); 4546 return str; 4547 } 4548 4549 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4550 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4551 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4552 // Capture src and dst memory (see comment in 'compress_string'). 4553 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4554 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4555 set_memory(_gvn.transform(str), dst_type); 4556 } 4557 4558 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4559 /** 4560 * int i_char = start; 4561 * for (int i_byte = 0; i_byte < count; i_byte++) { 4562 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4563 * } 4564 */ 4565 add_predicate(); 4566 RegionNode* head = new RegionNode(3); 4567 head->init_req(1, control()); 4568 gvn().set_type(head, Type::CONTROL); 4569 record_for_igvn(head); 4570 4571 Node* i_byte = new PhiNode(head, TypeInt::INT); 4572 i_byte->init_req(1, intcon(0)); 4573 gvn().set_type(i_byte, TypeInt::INT); 4574 record_for_igvn(i_byte); 4575 4576 Node* i_char = new PhiNode(head, TypeInt::INT); 4577 i_char->init_req(1, start); 4578 gvn().set_type(i_char, TypeInt::INT); 4579 record_for_igvn(i_char); 4580 4581 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4582 gvn().set_type(mem, Type::MEMORY); 4583 record_for_igvn(mem); 4584 set_control(head); 4585 set_memory(mem, TypeAryPtr::BYTES); 4586 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4587 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4588 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4589 false, false, true /* mismatched */); 4590 4591 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4592 head->init_req(2, IfTrue(iff)); 4593 mem->init_req(2, st); 4594 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4595 i_char->init_req(2, AddI(i_char, intcon(2))); 4596 4597 set_control(IfFalse(iff)); 4598 set_memory(st, TypeAryPtr::BYTES); 4599 } 4600 4601 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4602 if (!field->is_constant()) { 4603 return NULL; // Field not marked as constant. 4604 } 4605 ciInstance* holder = NULL; 4606 if (!field->is_static()) { 4607 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4608 if (const_oop != NULL && const_oop->is_instance()) { 4609 holder = const_oop->as_instance(); 4610 } 4611 } 4612 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4613 /*is_unsigned_load=*/false); 4614 if (con_type != NULL) { 4615 Node* con = makecon(con_type); 4616 if (field->layout_type() == T_VALUETYPE) { 4617 // Load value type from constant oop 4618 con = ValueTypeNode::make(gvn(), map()->memory(), con); 4619 } 4620 return con; 4621 } 4622 return NULL; 4623 } 4624 4625 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4626 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4627 // assumption of CCP analysis. 4628 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4629 }