1 /* 2 * Copyright (c) 2001, 2018, 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 "ci/ciUtilities.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "interpreter/interpreter.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/intrinsicnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 #include "utilities/macros.hpp" 47 #if INCLUDE_SHENANDOAHGC 48 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 49 #endif 50 51 //----------------------------GraphKit----------------------------------------- 52 // Main utility constructor. 53 GraphKit::GraphKit(JVMState* jvms) 54 : Phase(Phase::Parser), 55 _env(C->env()), 56 _gvn(*C->initial_gvn()), 57 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 58 { 59 _exceptions = jvms->map()->next_exception(); 60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 61 set_jvms(jvms); 62 } 63 64 // Private constructor for parser. 65 GraphKit::GraphKit() 66 : Phase(Phase::Parser), 67 _env(C->env()), 68 _gvn(*C->initial_gvn()), 69 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 70 { 71 _exceptions = NULL; 72 set_map(NULL); 73 debug_only(_sp = -99); 74 debug_only(set_bci(-99)); 75 } 76 77 78 79 //---------------------------clean_stack--------------------------------------- 80 // Clear away rubbish from the stack area of the JVM state. 81 // This destroys any arguments that may be waiting on the stack. 82 void GraphKit::clean_stack(int from_sp) { 83 SafePointNode* map = this->map(); 84 JVMState* jvms = this->jvms(); 85 int stk_size = jvms->stk_size(); 86 int stkoff = jvms->stkoff(); 87 Node* top = this->top(); 88 for (int i = from_sp; i < stk_size; i++) { 89 if (map->in(stkoff + i) != top) { 90 map->set_req(stkoff + i, top); 91 } 92 } 93 } 94 95 96 //--------------------------------sync_jvms----------------------------------- 97 // Make sure our current jvms agrees with our parse state. 98 JVMState* GraphKit::sync_jvms() const { 99 JVMState* jvms = this->jvms(); 100 jvms->set_bci(bci()); // Record the new bci in the JVMState 101 jvms->set_sp(sp()); // Record the new sp in the JVMState 102 assert(jvms_in_sync(), "jvms is now in sync"); 103 return jvms; 104 } 105 106 //--------------------------------sync_jvms_for_reexecute--------------------- 107 // Make sure our current jvms agrees with our parse state. This version 108 // uses the reexecute_sp for reexecuting bytecodes. 109 JVMState* GraphKit::sync_jvms_for_reexecute() { 110 JVMState* jvms = this->jvms(); 111 jvms->set_bci(bci()); // Record the new bci in the JVMState 112 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 113 return jvms; 114 } 115 116 #ifdef ASSERT 117 bool GraphKit::jvms_in_sync() const { 118 Parse* parse = is_Parse(); 119 if (parse == NULL) { 120 if (bci() != jvms()->bci()) return false; 121 if (sp() != (int)jvms()->sp()) return false; 122 return true; 123 } 124 if (jvms()->method() != parse->method()) return false; 125 if (jvms()->bci() != parse->bci()) return false; 126 int jvms_sp = jvms()->sp(); 127 if (jvms_sp != parse->sp()) return false; 128 int jvms_depth = jvms()->depth(); 129 if (jvms_depth != parse->depth()) return false; 130 return true; 131 } 132 133 // Local helper checks for special internal merge points 134 // used to accumulate and merge exception states. 135 // They are marked by the region's in(0) edge being the map itself. 136 // Such merge points must never "escape" into the parser at large, 137 // until they have been handed to gvn.transform. 138 static bool is_hidden_merge(Node* reg) { 139 if (reg == NULL) return false; 140 if (reg->is_Phi()) { 141 reg = reg->in(0); 142 if (reg == NULL) return false; 143 } 144 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 145 } 146 147 void GraphKit::verify_map() const { 148 if (map() == NULL) return; // null map is OK 149 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 150 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 151 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 152 } 153 154 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 155 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 156 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 157 } 158 #endif 159 160 //---------------------------stop_and_kill_map--------------------------------- 161 // Set _map to NULL, signalling a stop to further bytecode execution. 162 // First smash the current map's control to a constant, to mark it dead. 163 void GraphKit::stop_and_kill_map() { 164 SafePointNode* dead_map = stop(); 165 if (dead_map != NULL) { 166 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 167 assert(dead_map->is_killed(), "must be so marked"); 168 } 169 } 170 171 172 //--------------------------------stopped-------------------------------------- 173 // Tell if _map is NULL, or control is top. 174 bool GraphKit::stopped() { 175 if (map() == NULL) return true; 176 else if (control() == top()) return true; 177 else return false; 178 } 179 180 181 //-----------------------------has_ex_handler---------------------------------- 182 // Tell if this method or any caller method has exception handlers. 183 bool GraphKit::has_ex_handler() { 184 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 185 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 186 return true; 187 } 188 } 189 return false; 190 } 191 192 //------------------------------save_ex_oop------------------------------------ 193 // Save an exception without blowing stack contents or other JVM state. 194 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 195 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 196 ex_map->add_req(ex_oop); 197 debug_only(verify_exception_state(ex_map)); 198 } 199 200 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 201 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 202 Node* ex_oop = ex_map->in(ex_map->req()-1); 203 if (clear_it) ex_map->del_req(ex_map->req()-1); 204 return ex_oop; 205 } 206 207 //-----------------------------saved_ex_oop------------------------------------ 208 // Recover a saved exception from its map. 209 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, false); 211 } 212 213 //--------------------------clear_saved_ex_oop--------------------------------- 214 // Erase a previously saved exception from its map. 215 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 216 return common_saved_ex_oop(ex_map, true); 217 } 218 219 #ifdef ASSERT 220 //---------------------------has_saved_ex_oop---------------------------------- 221 // Erase a previously saved exception from its map. 222 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 223 return ex_map->req() == ex_map->jvms()->endoff()+1; 224 } 225 #endif 226 227 //-------------------------make_exception_state-------------------------------- 228 // Turn the current JVM state into an exception state, appending the ex_oop. 229 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 230 sync_jvms(); 231 SafePointNode* ex_map = stop(); // do not manipulate this map any more 232 set_saved_ex_oop(ex_map, ex_oop); 233 return ex_map; 234 } 235 236 237 //--------------------------add_exception_state-------------------------------- 238 // Add an exception to my list of exceptions. 239 void GraphKit::add_exception_state(SafePointNode* ex_map) { 240 if (ex_map == NULL || ex_map->control() == top()) { 241 return; 242 } 243 #ifdef ASSERT 244 verify_exception_state(ex_map); 245 if (has_exceptions()) { 246 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 247 } 248 #endif 249 250 // If there is already an exception of exactly this type, merge with it. 251 // In particular, null-checks and other low-level exceptions common up here. 252 Node* ex_oop = saved_ex_oop(ex_map); 253 const Type* ex_type = _gvn.type(ex_oop); 254 if (ex_oop == top()) { 255 // No action needed. 256 return; 257 } 258 assert(ex_type->isa_instptr(), "exception must be an instance"); 259 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 260 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 261 // We check sp also because call bytecodes can generate exceptions 262 // both before and after arguments are popped! 263 if (ex_type2 == ex_type 264 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 265 combine_exception_states(ex_map, e2); 266 return; 267 } 268 } 269 270 // No pre-existing exception of the same type. Chain it on the list. 271 push_exception_state(ex_map); 272 } 273 274 //-----------------------add_exception_states_from----------------------------- 275 void GraphKit::add_exception_states_from(JVMState* jvms) { 276 SafePointNode* ex_map = jvms->map()->next_exception(); 277 if (ex_map != NULL) { 278 jvms->map()->set_next_exception(NULL); 279 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 280 next_map = ex_map->next_exception(); 281 ex_map->set_next_exception(NULL); 282 add_exception_state(ex_map); 283 } 284 } 285 } 286 287 //-----------------------transfer_exceptions_into_jvms------------------------- 288 JVMState* GraphKit::transfer_exceptions_into_jvms() { 289 if (map() == NULL) { 290 // We need a JVMS to carry the exceptions, but the map has gone away. 291 // Create a scratch JVMS, cloned from any of the exception states... 292 if (has_exceptions()) { 293 _map = _exceptions; 294 _map = clone_map(); 295 _map->set_next_exception(NULL); 296 clear_saved_ex_oop(_map); 297 debug_only(verify_map()); 298 } else { 299 // ...or created from scratch 300 JVMState* jvms = new (C) JVMState(_method, NULL); 301 jvms->set_bci(_bci); 302 jvms->set_sp(_sp); 303 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 304 set_jvms(jvms); 305 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 306 set_all_memory(top()); 307 while (map()->req() < jvms->endoff()) map()->add_req(top()); 308 } 309 // (This is a kludge, in case you didn't notice.) 310 set_control(top()); 311 } 312 JVMState* jvms = sync_jvms(); 313 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 314 jvms->map()->set_next_exception(_exceptions); 315 _exceptions = NULL; // done with this set of exceptions 316 return jvms; 317 } 318 319 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 320 assert(is_hidden_merge(dstphi), "must be a special merge node"); 321 assert(is_hidden_merge(srcphi), "must be a special merge node"); 322 uint limit = srcphi->req(); 323 for (uint i = PhiNode::Input; i < limit; i++) { 324 dstphi->add_req(srcphi->in(i)); 325 } 326 } 327 static inline void add_one_req(Node* dstphi, Node* src) { 328 assert(is_hidden_merge(dstphi), "must be a special merge node"); 329 assert(!is_hidden_merge(src), "must not be a special merge node"); 330 dstphi->add_req(src); 331 } 332 333 //-----------------------combine_exception_states------------------------------ 334 // This helper function combines exception states by building phis on a 335 // specially marked state-merging region. These regions and phis are 336 // untransformed, and can build up gradually. The region is marked by 337 // having a control input of its exception map, rather than NULL. Such 338 // regions do not appear except in this function, and in use_exception_state. 339 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 340 if (failing()) return; // dying anyway... 341 JVMState* ex_jvms = ex_map->_jvms; 342 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 343 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 344 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 345 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 346 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 347 assert(ex_map->req() == phi_map->req(), "matching maps"); 348 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 349 Node* hidden_merge_mark = root(); 350 Node* region = phi_map->control(); 351 MergeMemNode* phi_mem = phi_map->merged_memory(); 352 MergeMemNode* ex_mem = ex_map->merged_memory(); 353 if (region->in(0) != hidden_merge_mark) { 354 // The control input is not (yet) a specially-marked region in phi_map. 355 // Make it so, and build some phis. 356 region = new RegionNode(2); 357 _gvn.set_type(region, Type::CONTROL); 358 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 359 region->init_req(1, phi_map->control()); 360 phi_map->set_control(region); 361 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 362 record_for_igvn(io_phi); 363 _gvn.set_type(io_phi, Type::ABIO); 364 phi_map->set_i_o(io_phi); 365 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 366 Node* m = mms.memory(); 367 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 368 record_for_igvn(m_phi); 369 _gvn.set_type(m_phi, Type::MEMORY); 370 mms.set_memory(m_phi); 371 } 372 } 373 374 // Either or both of phi_map and ex_map might already be converted into phis. 375 Node* ex_control = ex_map->control(); 376 // if there is special marking on ex_map also, we add multiple edges from src 377 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 378 // how wide was the destination phi_map, originally? 379 uint orig_width = region->req(); 380 381 if (add_multiple) { 382 add_n_reqs(region, ex_control); 383 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 384 } else { 385 // ex_map has no merges, so we just add single edges everywhere 386 add_one_req(region, ex_control); 387 add_one_req(phi_map->i_o(), ex_map->i_o()); 388 } 389 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 390 if (mms.is_empty()) { 391 // get a copy of the base memory, and patch some inputs into it 392 const TypePtr* adr_type = mms.adr_type(C); 393 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 394 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 395 mms.set_memory(phi); 396 // Prepare to append interesting stuff onto the newly sliced phi: 397 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 398 } 399 // Append stuff from ex_map: 400 if (add_multiple) { 401 add_n_reqs(mms.memory(), mms.memory2()); 402 } else { 403 add_one_req(mms.memory(), mms.memory2()); 404 } 405 } 406 uint limit = ex_map->req(); 407 for (uint i = TypeFunc::Parms; i < limit; i++) { 408 // Skip everything in the JVMS after tos. (The ex_oop follows.) 409 if (i == tos) i = ex_jvms->monoff(); 410 Node* src = ex_map->in(i); 411 Node* dst = phi_map->in(i); 412 if (src != dst) { 413 PhiNode* phi; 414 if (dst->in(0) != region) { 415 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 416 record_for_igvn(phi); 417 _gvn.set_type(phi, phi->type()); 418 phi_map->set_req(i, dst); 419 // Prepare to append interesting stuff onto the new phi: 420 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 421 } else { 422 assert(dst->is_Phi(), "nobody else uses a hidden region"); 423 phi = dst->as_Phi(); 424 } 425 if (add_multiple && src->in(0) == ex_control) { 426 // Both are phis. 427 add_n_reqs(dst, src); 428 } else { 429 while (dst->req() < region->req()) add_one_req(dst, src); 430 } 431 const Type* srctype = _gvn.type(src); 432 if (phi->type() != srctype) { 433 const Type* dsttype = phi->type()->meet_speculative(srctype); 434 if (phi->type() != dsttype) { 435 phi->set_type(dsttype); 436 _gvn.set_type(phi, dsttype); 437 } 438 } 439 } 440 } 441 phi_map->merge_replaced_nodes_with(ex_map); 442 } 443 444 //--------------------------use_exception_state-------------------------------- 445 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 446 if (failing()) { stop(); return top(); } 447 Node* region = phi_map->control(); 448 Node* hidden_merge_mark = root(); 449 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 450 Node* ex_oop = clear_saved_ex_oop(phi_map); 451 if (region->in(0) == hidden_merge_mark) { 452 // Special marking for internal ex-states. Process the phis now. 453 region->set_req(0, region); // now it's an ordinary region 454 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 455 // Note: Setting the jvms also sets the bci and sp. 456 set_control(_gvn.transform(region)); 457 uint tos = jvms()->stkoff() + sp(); 458 for (uint i = 1; i < tos; i++) { 459 Node* x = phi_map->in(i); 460 if (x->in(0) == region) { 461 assert(x->is_Phi(), "expected a special phi"); 462 phi_map->set_req(i, _gvn.transform(x)); 463 } 464 } 465 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 466 Node* x = mms.memory(); 467 if (x->in(0) == region) { 468 assert(x->is_Phi(), "nobody else uses a hidden region"); 469 mms.set_memory(_gvn.transform(x)); 470 } 471 } 472 if (ex_oop->in(0) == region) { 473 assert(ex_oop->is_Phi(), "expected a special phi"); 474 ex_oop = _gvn.transform(ex_oop); 475 } 476 } else { 477 set_jvms(phi_map->jvms()); 478 } 479 480 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 481 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 482 return ex_oop; 483 } 484 485 //---------------------------------java_bc------------------------------------- 486 Bytecodes::Code GraphKit::java_bc() const { 487 ciMethod* method = this->method(); 488 int bci = this->bci(); 489 if (method != NULL && bci != InvocationEntryBci) 490 return method->java_code_at_bci(bci); 491 else 492 return Bytecodes::_illegal; 493 } 494 495 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 496 bool must_throw) { 497 // if the exception capability is set, then we will generate code 498 // to check the JavaThread.should_post_on_exceptions flag to see 499 // if we actually need to report exception events (for this 500 // thread). If we don't need to report exception events, we will 501 // take the normal fast path provided by add_exception_events. If 502 // exception event reporting is enabled for this thread, we will 503 // take the uncommon_trap in the BuildCutout below. 504 505 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 506 Node* jthread = _gvn.transform(new ThreadLocalNode()); 507 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 508 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 509 510 // Test the should_post_on_exceptions_flag vs. 0 511 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 512 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 513 514 // Branch to slow_path if should_post_on_exceptions_flag was true 515 { BuildCutout unless(this, tst, PROB_MAX); 516 // Do not try anything fancy if we're notifying the VM on every throw. 517 // Cf. case Bytecodes::_athrow in parse2.cpp. 518 uncommon_trap(reason, Deoptimization::Action_none, 519 (ciKlass*)NULL, (char*)NULL, must_throw); 520 } 521 522 } 523 524 //------------------------------builtin_throw---------------------------------- 525 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 526 bool must_throw = true; 527 528 if (env()->jvmti_can_post_on_exceptions()) { 529 // check if we must post exception events, take uncommon trap if so 530 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 531 // here if should_post_on_exceptions is false 532 // continue on with the normal codegen 533 } 534 535 // If this particular condition has not yet happened at this 536 // bytecode, then use the uncommon trap mechanism, and allow for 537 // a future recompilation if several traps occur here. 538 // If the throw is hot, try to use a more complicated inline mechanism 539 // which keeps execution inside the compiled code. 540 bool treat_throw_as_hot = false; 541 ciMethodData* md = method()->method_data(); 542 543 if (ProfileTraps) { 544 if (too_many_traps(reason)) { 545 treat_throw_as_hot = true; 546 } 547 // (If there is no MDO at all, assume it is early in 548 // execution, and that any deopts are part of the 549 // startup transient, and don't need to be remembered.) 550 551 // Also, if there is a local exception handler, treat all throws 552 // as hot if there has been at least one in this method. 553 if (C->trap_count(reason) != 0 554 && method()->method_data()->trap_count(reason) != 0 555 && has_ex_handler()) { 556 treat_throw_as_hot = true; 557 } 558 } 559 560 // If this throw happens frequently, an uncommon trap might cause 561 // a performance pothole. If there is a local exception handler, 562 // and if this particular bytecode appears to be deoptimizing often, 563 // let us handle the throw inline, with a preconstructed instance. 564 // Note: If the deopt count has blown up, the uncommon trap 565 // runtime is going to flush this nmethod, not matter what. 566 if (treat_throw_as_hot 567 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 568 // If the throw is local, we use a pre-existing instance and 569 // punt on the backtrace. This would lead to a missing backtrace 570 // (a repeat of 4292742) if the backtrace object is ever asked 571 // for its backtrace. 572 // Fixing this remaining case of 4292742 requires some flavor of 573 // escape analysis. Leave that for the future. 574 ciInstance* ex_obj = NULL; 575 switch (reason) { 576 case Deoptimization::Reason_null_check: 577 ex_obj = env()->NullPointerException_instance(); 578 break; 579 case Deoptimization::Reason_div0_check: 580 ex_obj = env()->ArithmeticException_instance(); 581 break; 582 case Deoptimization::Reason_range_check: 583 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 584 break; 585 case Deoptimization::Reason_class_check: 586 if (java_bc() == Bytecodes::_aastore) { 587 ex_obj = env()->ArrayStoreException_instance(); 588 } else { 589 ex_obj = env()->ClassCastException_instance(); 590 } 591 break; 592 default: 593 break; 594 } 595 if (failing()) { stop(); return; } // exception allocation might fail 596 if (ex_obj != NULL) { 597 // Cheat with a preallocated exception object. 598 if (C->log() != NULL) 599 C->log()->elem("hot_throw preallocated='1' reason='%s'", 600 Deoptimization::trap_reason_name(reason)); 601 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 602 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 603 604 // Clear the detail message of the preallocated exception object. 605 // Weblogic sometimes mutates the detail message of exceptions 606 // using reflection. 607 int offset = java_lang_Throwable::get_detailMessage_offset(); 608 const TypePtr* adr_typ = ex_con->add_offset(offset); 609 610 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 611 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 612 Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 613 614 add_exception_state(make_exception_state(ex_node)); 615 return; 616 } 617 } 618 619 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 620 // It won't be much cheaper than bailing to the interp., since we'll 621 // have to pass up all the debug-info, and the runtime will have to 622 // create the stack trace. 623 624 // Usual case: Bail to interpreter. 625 // Reserve the right to recompile if we haven't seen anything yet. 626 627 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 628 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 629 if (treat_throw_as_hot 630 && (method()->method_data()->trap_recompiled_at(bci(), m) 631 || C->too_many_traps(reason))) { 632 // We cannot afford to take more traps here. Suffer in the interpreter. 633 if (C->log() != NULL) 634 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 635 Deoptimization::trap_reason_name(reason), 636 C->trap_count(reason)); 637 action = Deoptimization::Action_none; 638 } 639 640 // "must_throw" prunes the JVM state to include only the stack, if there 641 // are no local exception handlers. This should cut down on register 642 // allocation time and code size, by drastically reducing the number 643 // of in-edges on the call to the uncommon trap. 644 645 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 646 } 647 648 649 //----------------------------PreserveJVMState--------------------------------- 650 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 651 debug_only(kit->verify_map()); 652 _kit = kit; 653 _map = kit->map(); // preserve the map 654 _sp = kit->sp(); 655 kit->set_map(clone_map ? kit->clone_map() : NULL); 656 #ifdef ASSERT 657 _bci = kit->bci(); 658 Parse* parser = kit->is_Parse(); 659 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 660 _block = block; 661 #endif 662 } 663 PreserveJVMState::~PreserveJVMState() { 664 GraphKit* kit = _kit; 665 #ifdef ASSERT 666 assert(kit->bci() == _bci, "bci must not shift"); 667 Parse* parser = kit->is_Parse(); 668 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 669 assert(block == _block, "block must not shift"); 670 #endif 671 kit->set_map(_map); 672 kit->set_sp(_sp); 673 } 674 675 676 //-----------------------------BuildCutout------------------------------------- 677 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 678 : PreserveJVMState(kit) 679 { 680 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 681 SafePointNode* outer_map = _map; // preserved map is caller's 682 SafePointNode* inner_map = kit->map(); 683 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 684 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 685 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 686 } 687 BuildCutout::~BuildCutout() { 688 GraphKit* kit = _kit; 689 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 690 } 691 692 //---------------------------PreserveReexecuteState---------------------------- 693 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 694 assert(!kit->stopped(), "must call stopped() before"); 695 _kit = kit; 696 _sp = kit->sp(); 697 _reexecute = kit->jvms()->_reexecute; 698 } 699 PreserveReexecuteState::~PreserveReexecuteState() { 700 if (_kit->stopped()) return; 701 _kit->jvms()->_reexecute = _reexecute; 702 _kit->set_sp(_sp); 703 } 704 705 //------------------------------clone_map-------------------------------------- 706 // Implementation of PreserveJVMState 707 // 708 // Only clone_map(...) here. If this function is only used in the 709 // PreserveJVMState class we may want to get rid of this extra 710 // function eventually and do it all there. 711 712 SafePointNode* GraphKit::clone_map() { 713 if (map() == NULL) return NULL; 714 715 // Clone the memory edge first 716 Node* mem = MergeMemNode::make(map()->memory()); 717 gvn().set_type_bottom(mem); 718 719 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 720 JVMState* jvms = this->jvms(); 721 JVMState* clonejvms = jvms->clone_shallow(C); 722 clonemap->set_memory(mem); 723 clonemap->set_jvms(clonejvms); 724 clonejvms->set_map(clonemap); 725 record_for_igvn(clonemap); 726 gvn().set_type_bottom(clonemap); 727 return clonemap; 728 } 729 730 731 //-----------------------------set_map_clone----------------------------------- 732 void GraphKit::set_map_clone(SafePointNode* m) { 733 _map = m; 734 _map = clone_map(); 735 _map->set_next_exception(NULL); 736 debug_only(verify_map()); 737 } 738 739 740 //----------------------------kill_dead_locals--------------------------------- 741 // Detect any locals which are known to be dead, and force them to top. 742 void GraphKit::kill_dead_locals() { 743 // Consult the liveness information for the locals. If any 744 // of them are unused, then they can be replaced by top(). This 745 // should help register allocation time and cut down on the size 746 // of the deoptimization information. 747 748 // This call is made from many of the bytecode handling 749 // subroutines called from the Big Switch in do_one_bytecode. 750 // Every bytecode which might include a slow path is responsible 751 // for killing its dead locals. The more consistent we 752 // are about killing deads, the fewer useless phis will be 753 // constructed for them at various merge points. 754 755 // bci can be -1 (InvocationEntryBci). We return the entry 756 // liveness for the method. 757 758 if (method() == NULL || method()->code_size() == 0) { 759 // We are building a graph for a call to a native method. 760 // All locals are live. 761 return; 762 } 763 764 ResourceMark rm; 765 766 // Consult the liveness information for the locals. If any 767 // of them are unused, then they can be replaced by top(). This 768 // should help register allocation time and cut down on the size 769 // of the deoptimization information. 770 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 771 772 int len = (int)live_locals.size(); 773 assert(len <= jvms()->loc_size(), "too many live locals"); 774 for (int local = 0; local < len; local++) { 775 if (!live_locals.at(local)) { 776 set_local(local, top()); 777 } 778 } 779 } 780 781 #ifdef ASSERT 782 //-------------------------dead_locals_are_killed------------------------------ 783 // Return true if all dead locals are set to top in the map. 784 // Used to assert "clean" debug info at various points. 785 bool GraphKit::dead_locals_are_killed() { 786 if (method() == NULL || method()->code_size() == 0) { 787 // No locals need to be dead, so all is as it should be. 788 return true; 789 } 790 791 // Make sure somebody called kill_dead_locals upstream. 792 ResourceMark rm; 793 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 794 if (jvms->loc_size() == 0) continue; // no locals to consult 795 SafePointNode* map = jvms->map(); 796 ciMethod* method = jvms->method(); 797 int bci = jvms->bci(); 798 if (jvms == this->jvms()) { 799 bci = this->bci(); // it might not yet be synched 800 } 801 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 802 int len = (int)live_locals.size(); 803 if (!live_locals.is_valid() || len == 0) 804 // This method is trivial, or is poisoned by a breakpoint. 805 return true; 806 assert(len == jvms->loc_size(), "live map consistent with locals map"); 807 for (int local = 0; local < len; local++) { 808 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 809 if (PrintMiscellaneous && (Verbose || WizardMode)) { 810 tty->print_cr("Zombie local %d: ", local); 811 jvms->dump(); 812 } 813 return false; 814 } 815 } 816 } 817 return true; 818 } 819 820 #endif //ASSERT 821 822 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens. 823 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 824 ciMethod* cur_method = jvms->method(); 825 int cur_bci = jvms->bci(); 826 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 827 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 828 return Interpreter::bytecode_should_reexecute(code) || 829 (is_anewarray && code == Bytecodes::_multianewarray); 830 // Reexecute _multianewarray bytecode which was replaced with 831 // sequence of [a]newarray. See Parse::do_multianewarray(). 832 // 833 // Note: interpreter should not have it set since this optimization 834 // is limited by dimensions and guarded by flag so in some cases 835 // multianewarray() runtime calls will be generated and 836 // the bytecode should not be reexecutes (stack will not be reset). 837 } else { 838 return false; 839 } 840 } 841 842 // Helper function for adding JVMState and debug information to node 843 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 844 // Add the safepoint edges to the call (or other safepoint). 845 846 // Make sure dead locals are set to top. This 847 // should help register allocation time and cut down on the size 848 // of the deoptimization information. 849 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 850 851 // Walk the inline list to fill in the correct set of JVMState's 852 // Also fill in the associated edges for each JVMState. 853 854 // If the bytecode needs to be reexecuted we need to put 855 // the arguments back on the stack. 856 const bool should_reexecute = jvms()->should_reexecute(); 857 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 858 859 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 860 // undefined if the bci is different. This is normal for Parse but it 861 // should not happen for LibraryCallKit because only one bci is processed. 862 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 863 "in LibraryCallKit the reexecute bit should not change"); 864 865 // If we are guaranteed to throw, we can prune everything but the 866 // input to the current bytecode. 867 bool can_prune_locals = false; 868 uint stack_slots_not_pruned = 0; 869 int inputs = 0, depth = 0; 870 if (must_throw) { 871 assert(method() == youngest_jvms->method(), "sanity"); 872 if (compute_stack_effects(inputs, depth)) { 873 can_prune_locals = true; 874 stack_slots_not_pruned = inputs; 875 } 876 } 877 878 if (env()->should_retain_local_variables()) { 879 // At any safepoint, this method can get breakpointed, which would 880 // then require an immediate deoptimization. 881 can_prune_locals = false; // do not prune locals 882 stack_slots_not_pruned = 0; 883 } 884 885 // do not scribble on the input jvms 886 JVMState* out_jvms = youngest_jvms->clone_deep(C); 887 call->set_jvms(out_jvms); // Start jvms list for call node 888 889 // For a known set of bytecodes, the interpreter should reexecute them if 890 // deoptimization happens. We set the reexecute state for them here 891 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 892 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 893 #ifdef ASSERT 894 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects() 895 assert(method() == youngest_jvms->method(), "sanity"); 896 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc())); 897 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution"); 898 #endif // ASSERT 899 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 900 } 901 902 // Presize the call: 903 DEBUG_ONLY(uint non_debug_edges = call->req()); 904 call->add_req_batch(top(), youngest_jvms->debug_depth()); 905 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 906 907 // Set up edges so that the call looks like this: 908 // Call [state:] ctl io mem fptr retadr 909 // [parms:] parm0 ... parmN 910 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 911 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 912 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 913 // Note that caller debug info precedes callee debug info. 914 915 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 916 uint debug_ptr = call->req(); 917 918 // Loop over the map input edges associated with jvms, add them 919 // to the call node, & reset all offsets to match call node array. 920 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 921 uint debug_end = debug_ptr; 922 uint debug_start = debug_ptr - in_jvms->debug_size(); 923 debug_ptr = debug_start; // back up the ptr 924 925 uint p = debug_start; // walks forward in [debug_start, debug_end) 926 uint j, k, l; 927 SafePointNode* in_map = in_jvms->map(); 928 out_jvms->set_map(call); 929 930 if (can_prune_locals) { 931 assert(in_jvms->method() == out_jvms->method(), "sanity"); 932 // If the current throw can reach an exception handler in this JVMS, 933 // then we must keep everything live that can reach that handler. 934 // As a quick and dirty approximation, we look for any handlers at all. 935 if (in_jvms->method()->has_exception_handlers()) { 936 can_prune_locals = false; 937 } 938 } 939 940 // Add the Locals 941 k = in_jvms->locoff(); 942 l = in_jvms->loc_size(); 943 out_jvms->set_locoff(p); 944 if (!can_prune_locals) { 945 for (j = 0; j < l; j++) 946 call->set_req(p++, in_map->in(k+j)); 947 } else { 948 p += l; // already set to top above by add_req_batch 949 } 950 951 // Add the Expression Stack 952 k = in_jvms->stkoff(); 953 l = in_jvms->sp(); 954 out_jvms->set_stkoff(p); 955 if (!can_prune_locals) { 956 for (j = 0; j < l; j++) 957 call->set_req(p++, in_map->in(k+j)); 958 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 959 // Divide stack into {S0,...,S1}, where S0 is set to top. 960 uint s1 = stack_slots_not_pruned; 961 stack_slots_not_pruned = 0; // for next iteration 962 if (s1 > l) s1 = l; 963 uint s0 = l - s1; 964 p += s0; // skip the tops preinstalled by add_req_batch 965 for (j = s0; j < l; j++) 966 call->set_req(p++, in_map->in(k+j)); 967 } else { 968 p += l; // already set to top above by add_req_batch 969 } 970 971 // Add the Monitors 972 k = in_jvms->monoff(); 973 l = in_jvms->mon_size(); 974 out_jvms->set_monoff(p); 975 for (j = 0; j < l; j++) 976 call->set_req(p++, in_map->in(k+j)); 977 978 // Copy any scalar object fields. 979 k = in_jvms->scloff(); 980 l = in_jvms->scl_size(); 981 out_jvms->set_scloff(p); 982 for (j = 0; j < l; j++) 983 call->set_req(p++, in_map->in(k+j)); 984 985 // Finish the new jvms. 986 out_jvms->set_endoff(p); 987 988 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 989 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 990 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 991 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 992 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 993 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 994 995 // Update the two tail pointers in parallel. 996 out_jvms = out_jvms->caller(); 997 in_jvms = in_jvms->caller(); 998 } 999 1000 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1001 1002 // Test the correctness of JVMState::debug_xxx accessors: 1003 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1004 assert(call->jvms()->debug_end() == call->req(), ""); 1005 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1006 } 1007 1008 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1009 Bytecodes::Code code = java_bc(); 1010 if (code == Bytecodes::_wide) { 1011 code = method()->java_code_at_bci(bci() + 1); 1012 } 1013 1014 BasicType rtype = T_ILLEGAL; 1015 int rsize = 0; 1016 1017 if (code != Bytecodes::_illegal) { 1018 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1019 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1020 if (rtype < T_CONFLICT) 1021 rsize = type2size[rtype]; 1022 } 1023 1024 switch (code) { 1025 case Bytecodes::_illegal: 1026 return false; 1027 1028 case Bytecodes::_ldc: 1029 case Bytecodes::_ldc_w: 1030 case Bytecodes::_ldc2_w: 1031 inputs = 0; 1032 break; 1033 1034 case Bytecodes::_dup: inputs = 1; break; 1035 case Bytecodes::_dup_x1: inputs = 2; break; 1036 case Bytecodes::_dup_x2: inputs = 3; break; 1037 case Bytecodes::_dup2: inputs = 2; break; 1038 case Bytecodes::_dup2_x1: inputs = 3; break; 1039 case Bytecodes::_dup2_x2: inputs = 4; break; 1040 case Bytecodes::_swap: inputs = 2; break; 1041 case Bytecodes::_arraylength: inputs = 1; break; 1042 1043 case Bytecodes::_getstatic: 1044 case Bytecodes::_putstatic: 1045 case Bytecodes::_getfield: 1046 case Bytecodes::_putfield: 1047 { 1048 bool ignored_will_link; 1049 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1050 int size = field->type()->size(); 1051 bool is_get = (depth >= 0), is_static = (depth & 1); 1052 inputs = (is_static ? 0 : 1); 1053 if (is_get) { 1054 depth = size - inputs; 1055 } else { 1056 inputs += size; // putxxx pops the value from the stack 1057 depth = - inputs; 1058 } 1059 } 1060 break; 1061 1062 case Bytecodes::_invokevirtual: 1063 case Bytecodes::_invokespecial: 1064 case Bytecodes::_invokestatic: 1065 case Bytecodes::_invokedynamic: 1066 case Bytecodes::_invokeinterface: 1067 { 1068 bool ignored_will_link; 1069 ciSignature* declared_signature = NULL; 1070 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1071 assert(declared_signature != NULL, "cannot be null"); 1072 inputs = declared_signature->arg_size_for_bc(code); 1073 int size = declared_signature->return_type()->size(); 1074 depth = size - inputs; 1075 } 1076 break; 1077 1078 case Bytecodes::_multianewarray: 1079 { 1080 ciBytecodeStream iter(method()); 1081 iter.reset_to_bci(bci()); 1082 iter.next(); 1083 inputs = iter.get_dimensions(); 1084 assert(rsize == 1, ""); 1085 depth = rsize - inputs; 1086 } 1087 break; 1088 1089 case Bytecodes::_ireturn: 1090 case Bytecodes::_lreturn: 1091 case Bytecodes::_freturn: 1092 case Bytecodes::_dreturn: 1093 case Bytecodes::_areturn: 1094 assert(rsize == -depth, ""); 1095 inputs = rsize; 1096 break; 1097 1098 case Bytecodes::_jsr: 1099 case Bytecodes::_jsr_w: 1100 inputs = 0; 1101 depth = 1; // S.B. depth=1, not zero 1102 break; 1103 1104 default: 1105 // bytecode produces a typed result 1106 inputs = rsize - depth; 1107 assert(inputs >= 0, ""); 1108 break; 1109 } 1110 1111 #ifdef ASSERT 1112 // spot check 1113 int outputs = depth + inputs; 1114 assert(outputs >= 0, "sanity"); 1115 switch (code) { 1116 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1117 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1118 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1119 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1120 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1121 default: 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(speculative); 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 // Sometimes in intrinsics, we implicitly know an object is not null 1407 // (there's no actual null check) so we can cast it to not null. In 1408 // the course of optimizations, the input to the cast can become null. 1409 // In that case that data path will die and we need the control path 1410 // to become dead as well to keep the graph consistent. So we have to 1411 // add a check for null for which one branch can't be taken. It uses 1412 // an Opaque4 node that will cause the check to be removed after loop 1413 // opts so the test goes away and the compiled code doesn't execute a 1414 // useless check. 1415 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1416 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1417 return value; 1418 } 1419 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1420 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1421 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1422 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1423 _gvn.set_type(iff, iff->Value(&_gvn)); 1424 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1425 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1426 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1427 C->root()->add_req(halt); 1428 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1429 set_control(if_t); 1430 return cast_not_null(value, do_replace_in_map); 1431 } 1432 1433 1434 //--------------------------replace_in_map------------------------------------- 1435 void GraphKit::replace_in_map(Node* old, Node* neww) { 1436 if (old == neww) { 1437 return; 1438 } 1439 1440 map()->replace_edge(old, neww); 1441 1442 // Note: This operation potentially replaces any edge 1443 // on the map. This includes locals, stack, and monitors 1444 // of the current (innermost) JVM state. 1445 1446 // don't let inconsistent types from profiling escape this 1447 // method 1448 1449 const Type* told = _gvn.type(old); 1450 const Type* tnew = _gvn.type(neww); 1451 1452 if (!tnew->higher_equal(told)) { 1453 return; 1454 } 1455 1456 map()->record_replaced_node(old, neww); 1457 } 1458 1459 1460 //============================================================================= 1461 //--------------------------------memory--------------------------------------- 1462 Node* GraphKit::memory(uint alias_idx) { 1463 MergeMemNode* mem = merged_memory(); 1464 Node* p = mem->memory_at(alias_idx); 1465 _gvn.set_type(p, Type::MEMORY); // must be mapped 1466 return p; 1467 } 1468 1469 //-----------------------------reset_memory------------------------------------ 1470 Node* GraphKit::reset_memory() { 1471 Node* mem = map()->memory(); 1472 // do not use this node for any more parsing! 1473 debug_only( map()->set_memory((Node*)NULL) ); 1474 return _gvn.transform( mem ); 1475 } 1476 1477 //------------------------------set_all_memory--------------------------------- 1478 void GraphKit::set_all_memory(Node* newmem) { 1479 Node* mergemem = MergeMemNode::make(newmem); 1480 gvn().set_type_bottom(mergemem); 1481 map()->set_memory(mergemem); 1482 } 1483 1484 //------------------------------set_all_memory_call---------------------------- 1485 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1486 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1487 set_all_memory(newmem); 1488 } 1489 1490 //============================================================================= 1491 // 1492 // parser factory methods for MemNodes 1493 // 1494 // These are layered on top of the factory methods in LoadNode and StoreNode, 1495 // and integrate with the parser's memory state and _gvn engine. 1496 // 1497 1498 // factory methods in "int adr_idx" 1499 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1500 int adr_idx, 1501 MemNode::MemOrd mo, 1502 LoadNode::ControlDependency control_dependency, 1503 bool require_atomic_access, 1504 bool unaligned, 1505 bool mismatched, 1506 bool unsafe) { 1507 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1508 const TypePtr* adr_type = NULL; // debug-mode-only argument 1509 debug_only(adr_type = C->get_adr_type(adr_idx)); 1510 Node* mem = memory(adr_idx); 1511 Node* ld; 1512 if (require_atomic_access && bt == T_LONG) { 1513 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe); 1514 } else if (require_atomic_access && bt == T_DOUBLE) { 1515 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe); 1516 } else { 1517 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe); 1518 } 1519 ld = _gvn.transform(ld); 1520 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1521 // Improve graph before escape analysis and boxing elimination. 1522 record_for_igvn(ld); 1523 } 1524 return ld; 1525 } 1526 1527 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1528 int adr_idx, 1529 MemNode::MemOrd mo, 1530 bool require_atomic_access, 1531 bool unaligned, 1532 bool mismatched, 1533 bool unsafe) { 1534 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1535 const TypePtr* adr_type = NULL; 1536 debug_only(adr_type = C->get_adr_type(adr_idx)); 1537 Node *mem = memory(adr_idx); 1538 Node* st; 1539 if (require_atomic_access && bt == T_LONG) { 1540 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1541 } else if (require_atomic_access && bt == T_DOUBLE) { 1542 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1543 } else { 1544 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1545 } 1546 if (unaligned) { 1547 st->as_Store()->set_unaligned_access(); 1548 } 1549 if (mismatched) { 1550 st->as_Store()->set_mismatched_access(); 1551 } 1552 if (unsafe) { 1553 st->as_Store()->set_unsafe_access(); 1554 } 1555 st = _gvn.transform(st); 1556 set_memory(st, adr_idx); 1557 // Back-to-back stores can only remove intermediate store with DU info 1558 // so push on worklist for optimizer. 1559 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1560 record_for_igvn(st); 1561 1562 return st; 1563 } 1564 1565 Node* GraphKit::access_store_at(Node* ctl, 1566 Node* obj, 1567 Node* adr, 1568 const TypePtr* adr_type, 1569 Node* val, 1570 const Type* val_type, 1571 BasicType bt, 1572 DecoratorSet decorators) { 1573 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1574 // could be delayed during Parse (for example, in adjust_map_after_if()). 1575 // Execute transformation here to avoid barrier generation in such case. 1576 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1577 val = _gvn.makecon(TypePtr::NULL_PTR); 1578 } 1579 1580 set_control(ctl); 1581 if (stopped()) { 1582 return top(); // Dead path ? 1583 } 1584 1585 assert(val != NULL, "not dead path"); 1586 1587 C2AccessValuePtr addr(adr, adr_type); 1588 C2AccessValue value(val, val_type); 1589 C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1590 if (access.is_raw()) { 1591 return _barrier_set->BarrierSetC2::store_at(access, value); 1592 } else { 1593 return _barrier_set->store_at(access, value); 1594 } 1595 } 1596 1597 Node* GraphKit::access_load_at(Node* obj, // containing obj 1598 Node* adr, // actual adress to store val at 1599 const TypePtr* adr_type, 1600 const Type* val_type, 1601 BasicType bt, 1602 DecoratorSet decorators) { 1603 if (stopped()) { 1604 return top(); // Dead path ? 1605 } 1606 1607 C2AccessValuePtr addr(adr, adr_type); 1608 C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr); 1609 if (access.is_raw()) { 1610 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1611 } else { 1612 return _barrier_set->load_at(access, val_type); 1613 } 1614 } 1615 1616 Node* GraphKit::access_load(Node* adr, // actual adress to load val at 1617 const Type* val_type, 1618 BasicType bt, 1619 DecoratorSet decorators) { 1620 if (stopped()) { 1621 return top(); // Dead path ? 1622 } 1623 1624 C2AccessValuePtr addr(adr, NULL); 1625 C2Access access(this, decorators | C2_READ_ACCESS, bt, NULL, addr); 1626 if (access.is_raw()) { 1627 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1628 } else { 1629 return _barrier_set->load_at(access, val_type); 1630 } 1631 } 1632 1633 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl, 1634 Node* obj, 1635 Node* adr, 1636 const TypePtr* adr_type, 1637 int alias_idx, 1638 Node* expected_val, 1639 Node* new_val, 1640 const Type* value_type, 1641 BasicType bt, 1642 DecoratorSet decorators) { 1643 set_control(ctl); 1644 C2AccessValuePtr addr(adr, adr_type); 1645 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1646 bt, obj, addr, alias_idx); 1647 if (access.is_raw()) { 1648 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1649 } else { 1650 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1651 } 1652 } 1653 1654 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl, 1655 Node* obj, 1656 Node* adr, 1657 const TypePtr* adr_type, 1658 int alias_idx, 1659 Node* expected_val, 1660 Node* new_val, 1661 const Type* value_type, 1662 BasicType bt, 1663 DecoratorSet decorators) { 1664 set_control(ctl); 1665 C2AccessValuePtr addr(adr, adr_type); 1666 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1667 bt, obj, addr, alias_idx); 1668 if (access.is_raw()) { 1669 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1670 } else { 1671 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1672 } 1673 } 1674 1675 Node* GraphKit::access_atomic_xchg_at(Node* ctl, 1676 Node* obj, 1677 Node* adr, 1678 const TypePtr* adr_type, 1679 int alias_idx, 1680 Node* new_val, 1681 const Type* value_type, 1682 BasicType bt, 1683 DecoratorSet decorators) { 1684 set_control(ctl); 1685 C2AccessValuePtr addr(adr, adr_type); 1686 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1687 bt, obj, addr, alias_idx); 1688 if (access.is_raw()) { 1689 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1690 } else { 1691 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1692 } 1693 } 1694 1695 Node* GraphKit::access_atomic_add_at(Node* ctl, 1696 Node* obj, 1697 Node* adr, 1698 const TypePtr* adr_type, 1699 int alias_idx, 1700 Node* new_val, 1701 const Type* value_type, 1702 BasicType bt, 1703 DecoratorSet decorators) { 1704 set_control(ctl); 1705 C2AccessValuePtr addr(adr, adr_type); 1706 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1707 if (access.is_raw()) { 1708 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1709 } else { 1710 return _barrier_set->atomic_add_at(access, new_val, value_type); 1711 } 1712 } 1713 1714 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) { 1715 set_control(ctl); 1716 return _barrier_set->clone(this, src, dst, size, is_array); 1717 } 1718 1719 //-------------------------array_element_address------------------------- 1720 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1721 const TypeInt* sizetype, Node* ctrl) { 1722 uint shift = exact_log2(type2aelembytes(elembt)); 1723 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1724 1725 // short-circuit a common case (saves lots of confusing waste motion) 1726 jint idx_con = find_int_con(idx, -1); 1727 if (idx_con >= 0) { 1728 intptr_t offset = header + ((intptr_t)idx_con << shift); 1729 return basic_plus_adr(ary, offset); 1730 } 1731 1732 // must be correct type for alignment purposes 1733 Node* base = basic_plus_adr(ary, header); 1734 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1735 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1736 return basic_plus_adr(ary, base, scale); 1737 } 1738 1739 //-------------------------load_array_element------------------------- 1740 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1741 const Type* elemtype = arytype->elem(); 1742 BasicType elembt = elemtype->array_element_basic_type(); 1743 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1744 if (elembt == T_NARROWOOP) { 1745 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1746 } 1747 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1748 return ld; 1749 } 1750 1751 //-------------------------set_arguments_for_java_call------------------------- 1752 // Arguments (pre-popped from the stack) are taken from the JVMS. 1753 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1754 // Add the call arguments: 1755 uint nargs = call->method()->arg_size(); 1756 for (uint i = 0; i < nargs; i++) { 1757 Node* arg = argument(i); 1758 call->init_req(i + TypeFunc::Parms, arg); 1759 } 1760 } 1761 1762 //---------------------------set_edges_for_java_call--------------------------- 1763 // Connect a newly created call into the current JVMS. 1764 // A return value node (if any) is returned from set_edges_for_java_call. 1765 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1766 1767 // Add the predefined inputs: 1768 call->init_req( TypeFunc::Control, control() ); 1769 call->init_req( TypeFunc::I_O , i_o() ); 1770 call->init_req( TypeFunc::Memory , reset_memory() ); 1771 call->init_req( TypeFunc::FramePtr, frameptr() ); 1772 call->init_req( TypeFunc::ReturnAdr, top() ); 1773 1774 add_safepoint_edges(call, must_throw); 1775 1776 Node* xcall = _gvn.transform(call); 1777 1778 if (xcall == top()) { 1779 set_control(top()); 1780 return; 1781 } 1782 assert(xcall == call, "call identity is stable"); 1783 1784 // Re-use the current map to produce the result. 1785 1786 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1787 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1788 set_all_memory_call(xcall, separate_io_proj); 1789 1790 //return xcall; // no need, caller already has it 1791 } 1792 1793 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1794 if (stopped()) return top(); // maybe the call folded up? 1795 1796 // Capture the return value, if any. 1797 Node* ret; 1798 if (call->method() == NULL || 1799 call->method()->return_type()->basic_type() == T_VOID) 1800 ret = top(); 1801 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1802 1803 // Note: Since any out-of-line call can produce an exception, 1804 // we always insert an I_O projection from the call into the result. 1805 1806 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1807 1808 if (separate_io_proj) { 1809 // The caller requested separate projections be used by the fall 1810 // through and exceptional paths, so replace the projections for 1811 // the fall through path. 1812 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1813 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1814 } 1815 return ret; 1816 } 1817 1818 //--------------------set_predefined_input_for_runtime_call-------------------- 1819 // Reading and setting the memory state is way conservative here. 1820 // The real problem is that I am not doing real Type analysis on memory, 1821 // so I cannot distinguish card mark stores from other stores. Across a GC 1822 // point the Store Barrier and the card mark memory has to agree. I cannot 1823 // have a card mark store and its barrier split across the GC point from 1824 // either above or below. Here I get that to happen by reading ALL of memory. 1825 // A better answer would be to separate out card marks from other memory. 1826 // For now, return the input memory state, so that it can be reused 1827 // after the call, if this call has restricted memory effects. 1828 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1829 // Set fixed predefined input arguments 1830 Node* memory = reset_memory(); 1831 Node* m = narrow_mem == NULL ? memory : narrow_mem; 1832 call->init_req( TypeFunc::Control, control() ); 1833 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1834 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1835 call->init_req( TypeFunc::FramePtr, frameptr() ); 1836 call->init_req( TypeFunc::ReturnAdr, top() ); 1837 return memory; 1838 } 1839 1840 //-------------------set_predefined_output_for_runtime_call-------------------- 1841 // Set control and memory (not i_o) from the call. 1842 // If keep_mem is not NULL, use it for the output state, 1843 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1844 // If hook_mem is NULL, this call produces no memory effects at all. 1845 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1846 // then only that memory slice is taken from the call. 1847 // In the last case, we must put an appropriate memory barrier before 1848 // the call, so as to create the correct anti-dependencies on loads 1849 // preceding the call. 1850 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1851 Node* keep_mem, 1852 const TypePtr* hook_mem) { 1853 // no i/o 1854 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1855 if (keep_mem) { 1856 // First clone the existing memory state 1857 set_all_memory(keep_mem); 1858 if (hook_mem != NULL) { 1859 // Make memory for the call 1860 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1861 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1862 // We also use hook_mem to extract specific effects from arraycopy stubs. 1863 set_memory(mem, hook_mem); 1864 } 1865 // ...else the call has NO memory effects. 1866 1867 // Make sure the call advertises its memory effects precisely. 1868 // This lets us build accurate anti-dependences in gcm.cpp. 1869 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1870 "call node must be constructed correctly"); 1871 } else { 1872 assert(hook_mem == NULL, ""); 1873 // This is not a "slow path" call; all memory comes from the call. 1874 set_all_memory_call(call); 1875 } 1876 } 1877 1878 1879 // Replace the call with the current state of the kit. 1880 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1881 JVMState* ejvms = NULL; 1882 if (has_exceptions()) { 1883 ejvms = transfer_exceptions_into_jvms(); 1884 } 1885 1886 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1887 ReplacedNodes replaced_nodes_exception; 1888 Node* ex_ctl = top(); 1889 1890 SafePointNode* final_state = stop(); 1891 1892 // Find all the needed outputs of this call 1893 CallProjections callprojs; 1894 call->extract_projections(&callprojs, true); 1895 1896 Node* init_mem = call->in(TypeFunc::Memory); 1897 Node* final_mem = final_state->in(TypeFunc::Memory); 1898 Node* final_ctl = final_state->in(TypeFunc::Control); 1899 Node* final_io = final_state->in(TypeFunc::I_O); 1900 1901 // Replace all the old call edges with the edges from the inlining result 1902 if (callprojs.fallthrough_catchproj != NULL) { 1903 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1904 } 1905 if (callprojs.fallthrough_memproj != NULL) { 1906 if (final_mem->is_MergeMem()) { 1907 // Parser's exits MergeMem was not transformed but may be optimized 1908 final_mem = _gvn.transform(final_mem); 1909 } 1910 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1911 } 1912 if (callprojs.fallthrough_ioproj != NULL) { 1913 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1914 } 1915 1916 // Replace the result with the new result if it exists and is used 1917 if (callprojs.resproj != NULL && result != NULL) { 1918 C->gvn_replace_by(callprojs.resproj, result); 1919 } 1920 1921 if (ejvms == NULL) { 1922 // No exception edges to simply kill off those paths 1923 if (callprojs.catchall_catchproj != NULL) { 1924 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1925 } 1926 if (callprojs.catchall_memproj != NULL) { 1927 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1928 } 1929 if (callprojs.catchall_ioproj != NULL) { 1930 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1931 } 1932 // Replace the old exception object with top 1933 if (callprojs.exobj != NULL) { 1934 C->gvn_replace_by(callprojs.exobj, C->top()); 1935 } 1936 } else { 1937 GraphKit ekit(ejvms); 1938 1939 // Load my combined exception state into the kit, with all phis transformed: 1940 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1941 replaced_nodes_exception = ex_map->replaced_nodes(); 1942 1943 Node* ex_oop = ekit.use_exception_state(ex_map); 1944 1945 if (callprojs.catchall_catchproj != NULL) { 1946 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1947 ex_ctl = ekit.control(); 1948 } 1949 if (callprojs.catchall_memproj != NULL) { 1950 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1951 } 1952 if (callprojs.catchall_ioproj != NULL) { 1953 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1954 } 1955 1956 // Replace the old exception object with the newly created one 1957 if (callprojs.exobj != NULL) { 1958 C->gvn_replace_by(callprojs.exobj, ex_oop); 1959 } 1960 } 1961 1962 // Disconnect the call from the graph 1963 call->disconnect_inputs(NULL, C); 1964 C->gvn_replace_by(call, C->top()); 1965 1966 // Clean up any MergeMems that feed other MergeMems since the 1967 // optimizer doesn't like that. 1968 if (final_mem->is_MergeMem()) { 1969 Node_List wl; 1970 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1971 Node* m = i.get(); 1972 if (m->is_MergeMem() && !wl.contains(m)) { 1973 wl.push(m); 1974 } 1975 } 1976 while (wl.size() > 0) { 1977 _gvn.transform(wl.pop()); 1978 } 1979 } 1980 1981 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1982 replaced_nodes.apply(C, final_ctl); 1983 } 1984 if (!ex_ctl->is_top() && do_replaced_nodes) { 1985 replaced_nodes_exception.apply(C, ex_ctl); 1986 } 1987 } 1988 1989 1990 //------------------------------increment_counter------------------------------ 1991 // for statistics: increment a VM counter by 1 1992 1993 void GraphKit::increment_counter(address counter_addr) { 1994 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1995 increment_counter(adr1); 1996 } 1997 1998 void GraphKit::increment_counter(Node* counter_addr) { 1999 int adr_type = Compile::AliasIdxRaw; 2000 Node* ctrl = control(); 2001 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 2002 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 2003 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 2004 } 2005 2006 2007 //------------------------------uncommon_trap---------------------------------- 2008 // Bail out to the interpreter in mid-method. Implemented by calling the 2009 // uncommon_trap blob. This helper function inserts a runtime call with the 2010 // right debug info. 2011 void GraphKit::uncommon_trap(int trap_request, 2012 ciKlass* klass, const char* comment, 2013 bool must_throw, 2014 bool keep_exact_action) { 2015 if (failing()) stop(); 2016 if (stopped()) return; // trap reachable? 2017 2018 // Note: If ProfileTraps is true, and if a deopt. actually 2019 // occurs here, the runtime will make sure an MDO exists. There is 2020 // no need to call method()->ensure_method_data() at this point. 2021 2022 // Set the stack pointer to the right value for reexecution: 2023 set_sp(reexecute_sp()); 2024 2025 #ifdef ASSERT 2026 if (!must_throw) { 2027 // Make sure the stack has at least enough depth to execute 2028 // the current bytecode. 2029 int inputs, ignored_depth; 2030 if (compute_stack_effects(inputs, ignored_depth)) { 2031 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2032 Bytecodes::name(java_bc()), sp(), inputs); 2033 } 2034 } 2035 #endif 2036 2037 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2038 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2039 2040 switch (action) { 2041 case Deoptimization::Action_maybe_recompile: 2042 case Deoptimization::Action_reinterpret: 2043 // Temporary fix for 6529811 to allow virtual calls to be sure they 2044 // get the chance to go from mono->bi->mega 2045 if (!keep_exact_action && 2046 Deoptimization::trap_request_index(trap_request) < 0 && 2047 too_many_recompiles(reason)) { 2048 // This BCI is causing too many recompilations. 2049 if (C->log() != NULL) { 2050 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2051 Deoptimization::trap_reason_name(reason), 2052 Deoptimization::trap_action_name(action)); 2053 } 2054 action = Deoptimization::Action_none; 2055 trap_request = Deoptimization::make_trap_request(reason, action); 2056 } else { 2057 C->set_trap_can_recompile(true); 2058 } 2059 break; 2060 case Deoptimization::Action_make_not_entrant: 2061 C->set_trap_can_recompile(true); 2062 break; 2063 case Deoptimization::Action_none: 2064 case Deoptimization::Action_make_not_compilable: 2065 break; 2066 default: 2067 #ifdef ASSERT 2068 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2069 #endif 2070 break; 2071 } 2072 2073 if (TraceOptoParse) { 2074 char buf[100]; 2075 tty->print_cr("Uncommon trap %s at bci:%d", 2076 Deoptimization::format_trap_request(buf, sizeof(buf), 2077 trap_request), bci()); 2078 } 2079 2080 CompileLog* log = C->log(); 2081 if (log != NULL) { 2082 int kid = (klass == NULL)? -1: log->identify(klass); 2083 log->begin_elem("uncommon_trap bci='%d'", bci()); 2084 char buf[100]; 2085 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2086 trap_request)); 2087 if (kid >= 0) log->print(" klass='%d'", kid); 2088 if (comment != NULL) log->print(" comment='%s'", comment); 2089 log->end_elem(); 2090 } 2091 2092 // Make sure any guarding test views this path as very unlikely 2093 Node *i0 = control()->in(0); 2094 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2095 IfNode *iff = i0->as_If(); 2096 float f = iff->_prob; // Get prob 2097 if (control()->Opcode() == Op_IfTrue) { 2098 if (f > PROB_UNLIKELY_MAG(4)) 2099 iff->_prob = PROB_MIN; 2100 } else { 2101 if (f < PROB_LIKELY_MAG(4)) 2102 iff->_prob = PROB_MAX; 2103 } 2104 } 2105 2106 // Clear out dead values from the debug info. 2107 kill_dead_locals(); 2108 2109 // Now insert the uncommon trap subroutine call 2110 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2111 const TypePtr* no_memory_effects = NULL; 2112 // Pass the index of the class to be loaded 2113 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2114 (must_throw ? RC_MUST_THROW : 0), 2115 OptoRuntime::uncommon_trap_Type(), 2116 call_addr, "uncommon_trap", no_memory_effects, 2117 intcon(trap_request)); 2118 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2119 "must extract request correctly from the graph"); 2120 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2121 2122 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2123 // The debug info is the only real input to this call. 2124 2125 // Halt-and-catch fire here. The above call should never return! 2126 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2127 PRODUCT_ONLY(COMMA /*reachable*/false)); 2128 _gvn.set_type_bottom(halt); 2129 root()->add_req(halt); 2130 2131 stop_and_kill_map(); 2132 } 2133 2134 2135 //--------------------------just_allocated_object------------------------------ 2136 // Report the object that was just allocated. 2137 // It must be the case that there are no intervening safepoints. 2138 // We use this to determine if an object is so "fresh" that 2139 // it does not require card marks. 2140 Node* GraphKit::just_allocated_object(Node* current_control) { 2141 if (C->recent_alloc_ctl() == current_control) 2142 return C->recent_alloc_obj(); 2143 return NULL; 2144 } 2145 2146 2147 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2148 // (Note: TypeFunc::make has a cache that makes this fast.) 2149 const TypeFunc* tf = TypeFunc::make(dest_method); 2150 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2151 for (int j = 0; j < nargs; j++) { 2152 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2153 if( targ->basic_type() == T_DOUBLE ) { 2154 // If any parameters are doubles, they must be rounded before 2155 // the call, dstore_rounding does gvn.transform 2156 Node *arg = argument(j); 2157 arg = dstore_rounding(arg); 2158 set_argument(j, arg); 2159 } 2160 } 2161 } 2162 2163 /** 2164 * Record profiling data exact_kls for Node n with the type system so 2165 * that it can propagate it (speculation) 2166 * 2167 * @param n node that the type applies to 2168 * @param exact_kls type from profiling 2169 * @param maybe_null did profiling see null? 2170 * 2171 * @return node with improved type 2172 */ 2173 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2174 const Type* current_type = _gvn.type(n); 2175 assert(UseTypeSpeculation, "type speculation must be on"); 2176 2177 const TypePtr* speculative = current_type->speculative(); 2178 2179 // Should the klass from the profile be recorded in the speculative type? 2180 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2181 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2182 const TypeOopPtr* xtype = tklass->as_instance_type(); 2183 assert(xtype->klass_is_exact(), "Should be exact"); 2184 // Any reason to believe n is not null (from this profiling or a previous one)? 2185 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2186 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2187 // record the new speculative type's depth 2188 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2189 speculative = speculative->with_inline_depth(jvms()->depth()); 2190 } else if (current_type->would_improve_ptr(ptr_kind)) { 2191 // Profiling report that null was never seen so we can change the 2192 // speculative type to non null ptr. 2193 if (ptr_kind == ProfileAlwaysNull) { 2194 speculative = TypePtr::NULL_PTR; 2195 } else { 2196 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2197 const TypePtr* ptr = TypePtr::NOTNULL; 2198 if (speculative != NULL) { 2199 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2200 } else { 2201 speculative = ptr; 2202 } 2203 } 2204 } 2205 2206 if (speculative != current_type->speculative()) { 2207 // Build a type with a speculative type (what we think we know 2208 // about the type but will need a guard when we use it) 2209 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2210 // We're changing the type, we need a new CheckCast node to carry 2211 // the new type. The new type depends on the control: what 2212 // profiling tells us is only valid from here as far as we can 2213 // tell. 2214 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2215 cast = _gvn.transform(cast); 2216 replace_in_map(n, cast); 2217 n = cast; 2218 } 2219 2220 return n; 2221 } 2222 2223 /** 2224 * Record profiling data from receiver profiling at an invoke with the 2225 * type system so that it can propagate it (speculation) 2226 * 2227 * @param n receiver node 2228 * 2229 * @return node with improved type 2230 */ 2231 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2232 if (!UseTypeSpeculation) { 2233 return n; 2234 } 2235 ciKlass* exact_kls = profile_has_unique_klass(); 2236 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2237 if ((java_bc() == Bytecodes::_checkcast || 2238 java_bc() == Bytecodes::_instanceof || 2239 java_bc() == Bytecodes::_aastore) && 2240 method()->method_data()->is_mature()) { 2241 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2242 if (data != NULL) { 2243 if (!data->as_BitData()->null_seen()) { 2244 ptr_kind = ProfileNeverNull; 2245 } else { 2246 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2247 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2248 uint i = 0; 2249 for (; i < call->row_limit(); i++) { 2250 ciKlass* receiver = call->receiver(i); 2251 if (receiver != NULL) { 2252 break; 2253 } 2254 } 2255 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2256 } 2257 } 2258 } 2259 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2260 } 2261 2262 /** 2263 * Record profiling data from argument profiling at an invoke with the 2264 * type system so that it can propagate it (speculation) 2265 * 2266 * @param dest_method target method for the call 2267 * @param bc what invoke bytecode is this? 2268 */ 2269 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2270 if (!UseTypeSpeculation) { 2271 return; 2272 } 2273 const TypeFunc* tf = TypeFunc::make(dest_method); 2274 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2275 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2276 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2277 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2278 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2279 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2280 ciKlass* better_type = NULL; 2281 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2282 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2283 } 2284 i++; 2285 } 2286 } 2287 } 2288 2289 /** 2290 * Record profiling data from parameter profiling at an invoke with 2291 * the type system so that it can propagate it (speculation) 2292 */ 2293 void GraphKit::record_profiled_parameters_for_speculation() { 2294 if (!UseTypeSpeculation) { 2295 return; 2296 } 2297 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2298 if (_gvn.type(local(i))->isa_oopptr()) { 2299 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2300 ciKlass* better_type = NULL; 2301 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2302 record_profile_for_speculation(local(i), better_type, ptr_kind); 2303 } 2304 j++; 2305 } 2306 } 2307 } 2308 2309 /** 2310 * Record profiling data from return value profiling at an invoke with 2311 * the type system so that it can propagate it (speculation) 2312 */ 2313 void GraphKit::record_profiled_return_for_speculation() { 2314 if (!UseTypeSpeculation) { 2315 return; 2316 } 2317 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2318 ciKlass* better_type = NULL; 2319 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2320 // If profiling reports a single type for the return value, 2321 // feed it to the type system so it can propagate it as a 2322 // speculative type 2323 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2324 } 2325 } 2326 2327 void GraphKit::round_double_result(ciMethod* dest_method) { 2328 // A non-strict method may return a double value which has an extended 2329 // exponent, but this must not be visible in a caller which is 'strict' 2330 // If a strict caller invokes a non-strict callee, round a double result 2331 2332 BasicType result_type = dest_method->return_type()->basic_type(); 2333 assert( method() != NULL, "must have caller context"); 2334 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2335 // Destination method's return value is on top of stack 2336 // dstore_rounding() does gvn.transform 2337 Node *result = pop_pair(); 2338 result = dstore_rounding(result); 2339 push_pair(result); 2340 } 2341 } 2342 2343 // rounding for strict float precision conformance 2344 Node* GraphKit::precision_rounding(Node* n) { 2345 return UseStrictFP && _method->flags().is_strict() 2346 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2347 ? _gvn.transform( new RoundFloatNode(0, n) ) 2348 : n; 2349 } 2350 2351 // rounding for strict double precision conformance 2352 Node* GraphKit::dprecision_rounding(Node *n) { 2353 return UseStrictFP && _method->flags().is_strict() 2354 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2355 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2356 : n; 2357 } 2358 2359 // rounding for non-strict double stores 2360 Node* GraphKit::dstore_rounding(Node* n) { 2361 return Matcher::strict_fp_requires_explicit_rounding 2362 && UseSSE <= 1 2363 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2364 : n; 2365 } 2366 2367 //============================================================================= 2368 // Generate a fast path/slow path idiom. Graph looks like: 2369 // [foo] indicates that 'foo' is a parameter 2370 // 2371 // [in] NULL 2372 // \ / 2373 // CmpP 2374 // Bool ne 2375 // If 2376 // / \ 2377 // True False-<2> 2378 // / | 2379 // / cast_not_null 2380 // Load | | ^ 2381 // [fast_test] | | 2382 // gvn to opt_test | | 2383 // / \ | <1> 2384 // True False | 2385 // | \\ | 2386 // [slow_call] \[fast_result] 2387 // Ctl Val \ \ 2388 // | \ \ 2389 // Catch <1> \ \ 2390 // / \ ^ \ \ 2391 // Ex No_Ex | \ \ 2392 // | \ \ | \ <2> \ 2393 // ... \ [slow_res] | | \ [null_result] 2394 // \ \--+--+--- | | 2395 // \ | / \ | / 2396 // --------Region Phi 2397 // 2398 //============================================================================= 2399 // Code is structured as a series of driver functions all called 'do_XXX' that 2400 // call a set of helper functions. Helper functions first, then drivers. 2401 2402 //------------------------------null_check_oop--------------------------------- 2403 // Null check oop. Set null-path control into Region in slot 3. 2404 // Make a cast-not-nullness use the other not-null control. Return cast. 2405 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2406 bool never_see_null, 2407 bool safe_for_replace, 2408 bool speculative) { 2409 // Initial NULL check taken path 2410 (*null_control) = top(); 2411 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2412 2413 // Generate uncommon_trap: 2414 if (never_see_null && (*null_control) != top()) { 2415 // If we see an unexpected null at a check-cast we record it and force a 2416 // recompile; the offending check-cast will be compiled to handle NULLs. 2417 // If we see more than one offending BCI, then all checkcasts in the 2418 // method will be compiled to handle NULLs. 2419 PreserveJVMState pjvms(this); 2420 set_control(*null_control); 2421 replace_in_map(value, null()); 2422 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2423 uncommon_trap(reason, 2424 Deoptimization::Action_make_not_entrant); 2425 (*null_control) = top(); // NULL path is dead 2426 } 2427 if ((*null_control) == top() && safe_for_replace) { 2428 replace_in_map(value, cast); 2429 } 2430 2431 // Cast away null-ness on the result 2432 return cast; 2433 } 2434 2435 //------------------------------opt_iff---------------------------------------- 2436 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2437 // Return slow-path control. 2438 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2439 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2440 2441 // Fast path taken; set region slot 2 2442 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2443 region->init_req(2,fast_taken); // Capture fast-control 2444 2445 // Fast path not-taken, i.e. slow path 2446 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2447 return slow_taken; 2448 } 2449 2450 //-----------------------------make_runtime_call------------------------------- 2451 Node* GraphKit::make_runtime_call(int flags, 2452 const TypeFunc* call_type, address call_addr, 2453 const char* call_name, 2454 const TypePtr* adr_type, 2455 // The following parms are all optional. 2456 // The first NULL ends the list. 2457 Node* parm0, Node* parm1, 2458 Node* parm2, Node* parm3, 2459 Node* parm4, Node* parm5, 2460 Node* parm6, Node* parm7) { 2461 assert(call_addr != NULL, "must not call NULL targets"); 2462 2463 // Slow-path call 2464 bool is_leaf = !(flags & RC_NO_LEAF); 2465 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2466 if (call_name == NULL) { 2467 assert(!is_leaf, "must supply name for leaf"); 2468 call_name = OptoRuntime::stub_name(call_addr); 2469 } 2470 CallNode* call; 2471 if (!is_leaf) { 2472 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2473 bci(), adr_type); 2474 } else if (flags & RC_NO_FP) { 2475 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2476 } else { 2477 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2478 } 2479 2480 // The following is similar to set_edges_for_java_call, 2481 // except that the memory effects of the call are restricted to AliasIdxRaw. 2482 2483 // Slow path call has no side-effects, uses few values 2484 bool wide_in = !(flags & RC_NARROW_MEM); 2485 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2486 2487 Node* prev_mem = NULL; 2488 if (wide_in) { 2489 prev_mem = set_predefined_input_for_runtime_call(call); 2490 } else { 2491 assert(!wide_out, "narrow in => narrow out"); 2492 Node* narrow_mem = memory(adr_type); 2493 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2494 } 2495 2496 // Hook each parm in order. Stop looking at the first NULL. 2497 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2498 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2499 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2500 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2501 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2502 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2503 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2504 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2505 /* close each nested if ===> */ } } } } } } } } 2506 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2507 2508 if (!is_leaf) { 2509 // Non-leaves can block and take safepoints: 2510 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2511 } 2512 // Non-leaves can throw exceptions: 2513 if (has_io) { 2514 call->set_req(TypeFunc::I_O, i_o()); 2515 } 2516 2517 if (flags & RC_UNCOMMON) { 2518 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2519 // (An "if" probability corresponds roughly to an unconditional count. 2520 // Sort of.) 2521 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2522 } 2523 2524 Node* c = _gvn.transform(call); 2525 assert(c == call, "cannot disappear"); 2526 2527 if (wide_out) { 2528 // Slow path call has full side-effects. 2529 set_predefined_output_for_runtime_call(call); 2530 } else { 2531 // Slow path call has few side-effects, and/or sets few values. 2532 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2533 } 2534 2535 if (has_io) { 2536 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2537 } 2538 return call; 2539 2540 } 2541 2542 //------------------------------merge_memory----------------------------------- 2543 // Merge memory from one path into the current memory state. 2544 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2545 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2546 Node* old_slice = mms.force_memory(); 2547 Node* new_slice = mms.memory2(); 2548 if (old_slice != new_slice) { 2549 PhiNode* phi; 2550 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2551 if (mms.is_empty()) { 2552 // clone base memory Phi's inputs for this memory slice 2553 assert(old_slice == mms.base_memory(), "sanity"); 2554 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2555 _gvn.set_type(phi, Type::MEMORY); 2556 for (uint i = 1; i < phi->req(); i++) { 2557 phi->init_req(i, old_slice->in(i)); 2558 } 2559 } else { 2560 phi = old_slice->as_Phi(); // Phi was generated already 2561 } 2562 } else { 2563 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2564 _gvn.set_type(phi, Type::MEMORY); 2565 } 2566 phi->set_req(new_path, new_slice); 2567 mms.set_memory(phi); 2568 } 2569 } 2570 } 2571 2572 //------------------------------make_slow_call_ex------------------------------ 2573 // Make the exception handler hookups for the slow call 2574 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2575 if (stopped()) return; 2576 2577 // Make a catch node with just two handlers: fall-through and catch-all 2578 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2579 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2580 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2581 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2582 2583 { PreserveJVMState pjvms(this); 2584 set_control(excp); 2585 set_i_o(i_o); 2586 2587 if (excp != top()) { 2588 if (deoptimize) { 2589 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2590 uncommon_trap(Deoptimization::Reason_unhandled, 2591 Deoptimization::Action_none); 2592 } else { 2593 // Create an exception state also. 2594 // Use an exact type if the caller has a specific exception. 2595 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2596 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2597 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2598 } 2599 } 2600 } 2601 2602 // Get the no-exception control from the CatchNode. 2603 set_control(norm); 2604 } 2605 2606 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2607 Node* cmp = NULL; 2608 switch(bt) { 2609 case T_INT: cmp = new CmpINode(in1, in2); break; 2610 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2611 default: fatal("unexpected comparison type %s", type2name(bt)); 2612 } 2613 gvn->transform(cmp); 2614 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2615 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2616 gvn->transform(iff); 2617 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2618 return iff; 2619 } 2620 2621 2622 //-------------------------------gen_subtype_check----------------------------- 2623 // Generate a subtyping check. Takes as input the subtype and supertype. 2624 // Returns 2 values: sets the default control() to the true path and returns 2625 // the false path. Only reads invariant memory; sets no (visible) memory. 2626 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2627 // but that's not exposed to the optimizer. This call also doesn't take in an 2628 // Object; if you wish to check an Object you need to load the Object's class 2629 // prior to coming here. 2630 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2631 Compile* C = gvn->C; 2632 2633 if ((*ctrl)->is_top()) { 2634 return C->top(); 2635 } 2636 2637 // Fast check for identical types, perhaps identical constants. 2638 // The types can even be identical non-constants, in cases 2639 // involving Array.newInstance, Object.clone, etc. 2640 if (subklass == superklass) 2641 return C->top(); // false path is dead; no test needed. 2642 2643 if (gvn->type(superklass)->singleton()) { 2644 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2645 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2646 2647 // In the common case of an exact superklass, try to fold up the 2648 // test before generating code. You may ask, why not just generate 2649 // the code and then let it fold up? The answer is that the generated 2650 // code will necessarily include null checks, which do not always 2651 // completely fold away. If they are also needless, then they turn 2652 // into a performance loss. Example: 2653 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2654 // Here, the type of 'fa' is often exact, so the store check 2655 // of fa[1]=x will fold up, without testing the nullness of x. 2656 switch (C->static_subtype_check(superk, subk)) { 2657 case Compile::SSC_always_false: 2658 { 2659 Node* always_fail = *ctrl; 2660 *ctrl = gvn->C->top(); 2661 return always_fail; 2662 } 2663 case Compile::SSC_always_true: 2664 return C->top(); 2665 case Compile::SSC_easy_test: 2666 { 2667 // Just do a direct pointer compare and be done. 2668 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2669 *ctrl = gvn->transform(new IfTrueNode(iff)); 2670 return gvn->transform(new IfFalseNode(iff)); 2671 } 2672 case Compile::SSC_full_test: 2673 break; 2674 default: 2675 ShouldNotReachHere(); 2676 } 2677 } 2678 2679 // %%% Possible further optimization: Even if the superklass is not exact, 2680 // if the subklass is the unique subtype of the superklass, the check 2681 // will always succeed. We could leave a dependency behind to ensure this. 2682 2683 // First load the super-klass's check-offset 2684 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2685 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2686 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2687 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2688 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2689 2690 // Load from the sub-klass's super-class display list, or a 1-word cache of 2691 // the secondary superclass list, or a failing value with a sentinel offset 2692 // if the super-klass is an interface or exceptionally deep in the Java 2693 // hierarchy and we have to scan the secondary superclass list the hard way. 2694 // Worst-case type is a little odd: NULL is allowed as a result (usually 2695 // klass loads can never produce a NULL). 2696 Node *chk_off_X = chk_off; 2697 #ifdef _LP64 2698 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2699 #endif 2700 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2701 // For some types like interfaces the following loadKlass is from a 1-word 2702 // cache which is mutable so can't use immutable memory. Other 2703 // types load from the super-class display table which is immutable. 2704 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2705 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2706 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2707 2708 // Compile speed common case: ARE a subtype and we canNOT fail 2709 if( superklass == nkls ) 2710 return C->top(); // false path is dead; no test needed. 2711 2712 // See if we get an immediate positive hit. Happens roughly 83% of the 2713 // time. Test to see if the value loaded just previously from the subklass 2714 // is exactly the superklass. 2715 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2716 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2717 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2718 2719 // Compile speed common case: Check for being deterministic right now. If 2720 // chk_off is a constant and not equal to cacheoff then we are NOT a 2721 // subklass. In this case we need exactly the 1 test above and we can 2722 // return those results immediately. 2723 if (!might_be_cache) { 2724 Node* not_subtype_ctrl = *ctrl; 2725 *ctrl = iftrue1; // We need exactly the 1 test above 2726 return not_subtype_ctrl; 2727 } 2728 2729 // Gather the various success & failures here 2730 RegionNode *r_ok_subtype = new RegionNode(4); 2731 gvn->record_for_igvn(r_ok_subtype); 2732 RegionNode *r_not_subtype = new RegionNode(3); 2733 gvn->record_for_igvn(r_not_subtype); 2734 2735 r_ok_subtype->init_req(1, iftrue1); 2736 2737 // Check for immediate negative hit. Happens roughly 11% of the time (which 2738 // is roughly 63% of the remaining cases). Test to see if the loaded 2739 // check-offset points into the subklass display list or the 1-element 2740 // cache. If it points to the display (and NOT the cache) and the display 2741 // missed then it's not a subtype. 2742 Node *cacheoff = gvn->intcon(cacheoff_con); 2743 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2744 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2745 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2746 2747 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2748 // No performance impact (too rare) but allows sharing of secondary arrays 2749 // which has some footprint reduction. 2750 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2751 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2752 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2753 2754 // -- Roads not taken here: -- 2755 // We could also have chosen to perform the self-check at the beginning 2756 // of this code sequence, as the assembler does. This would not pay off 2757 // the same way, since the optimizer, unlike the assembler, can perform 2758 // static type analysis to fold away many successful self-checks. 2759 // Non-foldable self checks work better here in second position, because 2760 // the initial primary superclass check subsumes a self-check for most 2761 // types. An exception would be a secondary type like array-of-interface, 2762 // which does not appear in its own primary supertype display. 2763 // Finally, we could have chosen to move the self-check into the 2764 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2765 // dependent manner. But it is worthwhile to have the check here, 2766 // where it can be perhaps be optimized. The cost in code space is 2767 // small (register compare, branch). 2768 2769 // Now do a linear scan of the secondary super-klass array. Again, no real 2770 // performance impact (too rare) but it's gotta be done. 2771 // Since the code is rarely used, there is no penalty for moving it 2772 // out of line, and it can only improve I-cache density. 2773 // The decision to inline or out-of-line this final check is platform 2774 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2775 Node* psc = gvn->transform( 2776 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2777 2778 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2779 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2780 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2781 2782 // Return false path; set default control to true path. 2783 *ctrl = gvn->transform(r_ok_subtype); 2784 return gvn->transform(r_not_subtype); 2785 } 2786 2787 // Profile-driven exact type check: 2788 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2789 float prob, 2790 Node* *casted_receiver) { 2791 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2792 Node* recv_klass = load_object_klass(receiver); 2793 Node* want_klass = makecon(tklass); 2794 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2795 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2796 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2797 set_control( _gvn.transform( new IfTrueNode (iff) )); 2798 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2799 2800 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2801 assert(recv_xtype->klass_is_exact(), ""); 2802 2803 // Subsume downstream occurrences of receiver with a cast to 2804 // recv_xtype, since now we know what the type will be. 2805 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2806 (*casted_receiver) = _gvn.transform(cast); 2807 // (User must make the replace_in_map call.) 2808 2809 return fail; 2810 } 2811 2812 2813 //------------------------------seems_never_null------------------------------- 2814 // Use null_seen information if it is available from the profile. 2815 // If we see an unexpected null at a type check we record it and force a 2816 // recompile; the offending check will be recompiled to handle NULLs. 2817 // If we see several offending BCIs, then all checks in the 2818 // method will be recompiled. 2819 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2820 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2821 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2822 if (UncommonNullCast // Cutout for this technique 2823 && obj != null() // And not the -Xcomp stupid case? 2824 && !too_many_traps(reason) 2825 ) { 2826 if (speculating) { 2827 return true; 2828 } 2829 if (data == NULL) 2830 // Edge case: no mature data. Be optimistic here. 2831 return true; 2832 // If the profile has not seen a null, assume it won't happen. 2833 assert(java_bc() == Bytecodes::_checkcast || 2834 java_bc() == Bytecodes::_instanceof || 2835 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2836 return !data->as_BitData()->null_seen(); 2837 } 2838 speculating = false; 2839 return false; 2840 } 2841 2842 //------------------------maybe_cast_profiled_receiver------------------------- 2843 // If the profile has seen exactly one type, narrow to exactly that type. 2844 // Subsequent type checks will always fold up. 2845 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2846 ciKlass* require_klass, 2847 ciKlass* spec_klass, 2848 bool safe_for_replace) { 2849 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2850 2851 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2852 2853 // Make sure we haven't already deoptimized from this tactic. 2854 if (too_many_traps_or_recompiles(reason)) 2855 return NULL; 2856 2857 // (No, this isn't a call, but it's enough like a virtual call 2858 // to use the same ciMethod accessor to get the profile info...) 2859 // If we have a speculative type use it instead of profiling (which 2860 // may not help us) 2861 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2862 if (exact_kls != NULL) {// no cast failures here 2863 if (require_klass == NULL || 2864 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2865 // If we narrow the type to match what the type profile sees or 2866 // the speculative type, we can then remove the rest of the 2867 // cast. 2868 // This is a win, even if the exact_kls is very specific, 2869 // because downstream operations, such as method calls, 2870 // will often benefit from the sharper type. 2871 Node* exact_obj = not_null_obj; // will get updated in place... 2872 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2873 &exact_obj); 2874 { PreserveJVMState pjvms(this); 2875 set_control(slow_ctl); 2876 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2877 } 2878 if (safe_for_replace) { 2879 replace_in_map(not_null_obj, exact_obj); 2880 } 2881 return exact_obj; 2882 } 2883 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2884 } 2885 2886 return NULL; 2887 } 2888 2889 /** 2890 * Cast obj to type and emit guard unless we had too many traps here 2891 * already 2892 * 2893 * @param obj node being casted 2894 * @param type type to cast the node to 2895 * @param not_null true if we know node cannot be null 2896 */ 2897 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2898 ciKlass* type, 2899 bool not_null) { 2900 if (stopped()) { 2901 return obj; 2902 } 2903 2904 // type == NULL if profiling tells us this object is always null 2905 if (type != NULL) { 2906 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2907 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2908 2909 if (!too_many_traps_or_recompiles(null_reason) && 2910 !too_many_traps_or_recompiles(class_reason)) { 2911 Node* not_null_obj = NULL; 2912 // not_null is true if we know the object is not null and 2913 // there's no need for a null check 2914 if (!not_null) { 2915 Node* null_ctl = top(); 2916 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2917 assert(null_ctl->is_top(), "no null control here"); 2918 } else { 2919 not_null_obj = obj; 2920 } 2921 2922 Node* exact_obj = not_null_obj; 2923 ciKlass* exact_kls = type; 2924 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2925 &exact_obj); 2926 { 2927 PreserveJVMState pjvms(this); 2928 set_control(slow_ctl); 2929 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2930 } 2931 replace_in_map(not_null_obj, exact_obj); 2932 obj = exact_obj; 2933 } 2934 } else { 2935 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 2936 Node* exact_obj = null_assert(obj); 2937 replace_in_map(obj, exact_obj); 2938 obj = exact_obj; 2939 } 2940 } 2941 return obj; 2942 } 2943 2944 //-------------------------------gen_instanceof-------------------------------- 2945 // Generate an instance-of idiom. Used by both the instance-of bytecode 2946 // and the reflective instance-of call. 2947 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2948 kill_dead_locals(); // Benefit all the uncommon traps 2949 assert( !stopped(), "dead parse path should be checked in callers" ); 2950 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2951 "must check for not-null not-dead klass in callers"); 2952 2953 // Make the merge point 2954 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2955 RegionNode* region = new RegionNode(PATH_LIMIT); 2956 Node* phi = new PhiNode(region, TypeInt::BOOL); 2957 C->set_has_split_ifs(true); // Has chance for split-if optimization 2958 2959 ciProfileData* data = NULL; 2960 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2961 data = method()->method_data()->bci_to_data(bci()); 2962 } 2963 bool speculative_not_null = false; 2964 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2965 && seems_never_null(obj, data, speculative_not_null)); 2966 2967 // Null check; get casted pointer; set region slot 3 2968 Node* null_ctl = top(); 2969 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2970 2971 // If not_null_obj is dead, only null-path is taken 2972 if (stopped()) { // Doing instance-of on a NULL? 2973 set_control(null_ctl); 2974 return intcon(0); 2975 } 2976 region->init_req(_null_path, null_ctl); 2977 phi ->init_req(_null_path, intcon(0)); // Set null path value 2978 if (null_ctl == top()) { 2979 // Do this eagerly, so that pattern matches like is_diamond_phi 2980 // will work even during parsing. 2981 assert(_null_path == PATH_LIMIT-1, "delete last"); 2982 region->del_req(_null_path); 2983 phi ->del_req(_null_path); 2984 } 2985 2986 // Do we know the type check always succeed? 2987 bool known_statically = false; 2988 if (_gvn.type(superklass)->singleton()) { 2989 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2990 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2991 if (subk != NULL && subk->is_loaded()) { 2992 int static_res = C->static_subtype_check(superk, subk); 2993 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2994 } 2995 } 2996 2997 if (!known_statically) { 2998 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2999 // We may not have profiling here or it may not help us. If we 3000 // have a speculative type use it to perform an exact cast. 3001 ciKlass* spec_obj_type = obj_type->speculative_type(); 3002 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 3003 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 3004 if (stopped()) { // Profile disagrees with this path. 3005 set_control(null_ctl); // Null is the only remaining possibility. 3006 return intcon(0); 3007 } 3008 if (cast_obj != NULL) { 3009 not_null_obj = cast_obj; 3010 } 3011 } 3012 } 3013 3014 // Load the object's klass 3015 Node* obj_klass = load_object_klass(not_null_obj); 3016 3017 // Generate the subtype check 3018 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 3019 3020 // Plug in the success path to the general merge in slot 1. 3021 region->init_req(_obj_path, control()); 3022 phi ->init_req(_obj_path, intcon(1)); 3023 3024 // Plug in the failing path to the general merge in slot 2. 3025 region->init_req(_fail_path, not_subtype_ctrl); 3026 phi ->init_req(_fail_path, intcon(0)); 3027 3028 // Return final merged results 3029 set_control( _gvn.transform(region) ); 3030 record_for_igvn(region); 3031 3032 // If we know the type check always succeeds then we don't use the 3033 // profiling data at this bytecode. Don't lose it, feed it to the 3034 // type system as a speculative type. 3035 if (safe_for_replace) { 3036 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3037 replace_in_map(obj, casted_obj); 3038 } 3039 3040 return _gvn.transform(phi); 3041 } 3042 3043 //-------------------------------gen_checkcast--------------------------------- 3044 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3045 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3046 // uncommon-trap paths work. Adjust stack after this call. 3047 // If failure_control is supplied and not null, it is filled in with 3048 // the control edge for the cast failure. Otherwise, an appropriate 3049 // uncommon trap or exception is thrown. 3050 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3051 Node* *failure_control) { 3052 kill_dead_locals(); // Benefit all the uncommon traps 3053 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3054 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3055 3056 // Fast cutout: Check the case that the cast is vacuously true. 3057 // This detects the common cases where the test will short-circuit 3058 // away completely. We do this before we perform the null check, 3059 // because if the test is going to turn into zero code, we don't 3060 // want a residual null check left around. (Causes a slowdown, 3061 // for example, in some objArray manipulations, such as a[i]=a[j].) 3062 if (tk->singleton()) { 3063 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3064 if (objtp != NULL && objtp->klass() != NULL) { 3065 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3066 case Compile::SSC_always_true: 3067 // If we know the type check always succeed then we don't use 3068 // the profiling data at this bytecode. Don't lose it, feed it 3069 // to the type system as a speculative type. 3070 return record_profiled_receiver_for_speculation(obj); 3071 case Compile::SSC_always_false: 3072 // It needs a null check because a null will *pass* the cast check. 3073 // A non-null value will always produce an exception. 3074 if (!objtp->maybe_null()) { 3075 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(objtp->klass()))); 3076 return top(); 3077 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3078 return null_assert(obj); 3079 } 3080 break; // Fall through to full check 3081 } 3082 } 3083 } 3084 3085 ciProfileData* data = NULL; 3086 bool safe_for_replace = false; 3087 if (failure_control == NULL) { // use MDO in regular case only 3088 assert(java_bc() == Bytecodes::_aastore || 3089 java_bc() == Bytecodes::_checkcast, 3090 "interpreter profiles type checks only for these BCs"); 3091 data = method()->method_data()->bci_to_data(bci()); 3092 safe_for_replace = true; 3093 } 3094 3095 // Make the merge point 3096 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3097 RegionNode* region = new RegionNode(PATH_LIMIT); 3098 Node* phi = new PhiNode(region, toop); 3099 C->set_has_split_ifs(true); // Has chance for split-if optimization 3100 3101 // Use null-cast information if it is available 3102 bool speculative_not_null = false; 3103 bool never_see_null = ((failure_control == NULL) // regular case only 3104 && seems_never_null(obj, data, speculative_not_null)); 3105 3106 // Null check; get casted pointer; set region slot 3 3107 Node* null_ctl = top(); 3108 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3109 3110 // If not_null_obj is dead, only null-path is taken 3111 if (stopped()) { // Doing instance-of on a NULL? 3112 set_control(null_ctl); 3113 return null(); 3114 } 3115 region->init_req(_null_path, null_ctl); 3116 phi ->init_req(_null_path, null()); // Set null path value 3117 if (null_ctl == top()) { 3118 // Do this eagerly, so that pattern matches like is_diamond_phi 3119 // will work even during parsing. 3120 assert(_null_path == PATH_LIMIT-1, "delete last"); 3121 region->del_req(_null_path); 3122 phi ->del_req(_null_path); 3123 } 3124 3125 Node* cast_obj = NULL; 3126 if (tk->klass_is_exact()) { 3127 // The following optimization tries to statically cast the speculative type of the object 3128 // (for example obtained during profiling) to the type of the superklass and then do a 3129 // dynamic check that the type of the object is what we expect. To work correctly 3130 // for checkcast and aastore the type of superklass should be exact. 3131 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3132 // We may not have profiling here or it may not help us. If we have 3133 // a speculative type use it to perform an exact cast. 3134 ciKlass* spec_obj_type = obj_type->speculative_type(); 3135 if (spec_obj_type != NULL || data != NULL) { 3136 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3137 if (cast_obj != NULL) { 3138 if (failure_control != NULL) // failure is now impossible 3139 (*failure_control) = top(); 3140 // adjust the type of the phi to the exact klass: 3141 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3142 } 3143 } 3144 } 3145 3146 if (cast_obj == NULL) { 3147 // Load the object's klass 3148 Node* obj_klass = load_object_klass(not_null_obj); 3149 3150 // Generate the subtype check 3151 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3152 3153 // Plug in success path into the merge 3154 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3155 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3156 if (failure_control == NULL) { 3157 if (not_subtype_ctrl != top()) { // If failure is possible 3158 PreserveJVMState pjvms(this); 3159 set_control(not_subtype_ctrl); 3160 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3161 } 3162 } else { 3163 (*failure_control) = not_subtype_ctrl; 3164 } 3165 } 3166 3167 region->init_req(_obj_path, control()); 3168 phi ->init_req(_obj_path, cast_obj); 3169 3170 // A merge of NULL or Casted-NotNull obj 3171 Node* res = _gvn.transform(phi); 3172 3173 // Note I do NOT always 'replace_in_map(obj,result)' here. 3174 // if( tk->klass()->can_be_primary_super() ) 3175 // This means that if I successfully store an Object into an array-of-String 3176 // I 'forget' that the Object is really now known to be a String. I have to 3177 // do this because we don't have true union types for interfaces - if I store 3178 // a Baz into an array-of-Interface and then tell the optimizer it's an 3179 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3180 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3181 // replace_in_map( obj, res ); 3182 3183 // Return final merged results 3184 set_control( _gvn.transform(region) ); 3185 record_for_igvn(region); 3186 3187 return record_profiled_receiver_for_speculation(res); 3188 } 3189 3190 //------------------------------next_monitor----------------------------------- 3191 // What number should be given to the next monitor? 3192 int GraphKit::next_monitor() { 3193 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3194 int next = current + C->sync_stack_slots(); 3195 // Keep the toplevel high water mark current: 3196 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3197 return current; 3198 } 3199 3200 //------------------------------insert_mem_bar--------------------------------- 3201 // Memory barrier to avoid floating things around 3202 // The membar serves as a pinch point between both control and all memory slices. 3203 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3204 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3205 mb->init_req(TypeFunc::Control, control()); 3206 mb->init_req(TypeFunc::Memory, reset_memory()); 3207 Node* membar = _gvn.transform(mb); 3208 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3209 set_all_memory_call(membar); 3210 return membar; 3211 } 3212 3213 //-------------------------insert_mem_bar_volatile---------------------------- 3214 // Memory barrier to avoid floating things around 3215 // The membar serves as a pinch point between both control and memory(alias_idx). 3216 // If you want to make a pinch point on all memory slices, do not use this 3217 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3218 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3219 // When Parse::do_put_xxx updates a volatile field, it appends a series 3220 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3221 // The first membar is on the same memory slice as the field store opcode. 3222 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3223 // All the other membars (for other volatile slices, including AliasIdxBot, 3224 // which stands for all unknown volatile slices) are control-dependent 3225 // on the first membar. This prevents later volatile loads or stores 3226 // from sliding up past the just-emitted store. 3227 3228 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3229 mb->set_req(TypeFunc::Control,control()); 3230 if (alias_idx == Compile::AliasIdxBot) { 3231 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3232 } else { 3233 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3234 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3235 } 3236 Node* membar = _gvn.transform(mb); 3237 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3238 if (alias_idx == Compile::AliasIdxBot) { 3239 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3240 } else { 3241 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3242 } 3243 return membar; 3244 } 3245 3246 //------------------------------shared_lock------------------------------------ 3247 // Emit locking code. 3248 FastLockNode* GraphKit::shared_lock(Node* obj) { 3249 // bci is either a monitorenter bc or InvocationEntryBci 3250 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3251 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3252 3253 if( !GenerateSynchronizationCode ) 3254 return NULL; // Not locking things? 3255 if (stopped()) // Dead monitor? 3256 return NULL; 3257 3258 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3259 3260 // Box the stack location 3261 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3262 Node* mem = reset_memory(); 3263 3264 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3265 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3266 // Create the counters for this fast lock. 3267 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3268 } 3269 3270 // Create the rtm counters for this fast lock if needed. 3271 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3272 3273 // Add monitor to debug info for the slow path. If we block inside the 3274 // slow path and de-opt, we need the monitor hanging around 3275 map()->push_monitor( flock ); 3276 3277 const TypeFunc *tf = LockNode::lock_type(); 3278 LockNode *lock = new LockNode(C, tf); 3279 3280 lock->init_req( TypeFunc::Control, control() ); 3281 lock->init_req( TypeFunc::Memory , mem ); 3282 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3283 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3284 lock->init_req( TypeFunc::ReturnAdr, top() ); 3285 3286 lock->init_req(TypeFunc::Parms + 0, obj); 3287 lock->init_req(TypeFunc::Parms + 1, box); 3288 lock->init_req(TypeFunc::Parms + 2, flock); 3289 add_safepoint_edges(lock); 3290 3291 lock = _gvn.transform( lock )->as_Lock(); 3292 3293 // lock has no side-effects, sets few values 3294 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3295 3296 insert_mem_bar(Op_MemBarAcquireLock); 3297 3298 // Add this to the worklist so that the lock can be eliminated 3299 record_for_igvn(lock); 3300 3301 #ifndef PRODUCT 3302 if (PrintLockStatistics) { 3303 // Update the counter for this lock. Don't bother using an atomic 3304 // operation since we don't require absolute accuracy. 3305 lock->create_lock_counter(map()->jvms()); 3306 increment_counter(lock->counter()->addr()); 3307 } 3308 #endif 3309 3310 return flock; 3311 } 3312 3313 3314 //------------------------------shared_unlock---------------------------------- 3315 // Emit unlocking code. 3316 void GraphKit::shared_unlock(Node* box, Node* obj) { 3317 // bci is either a monitorenter bc or InvocationEntryBci 3318 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3319 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3320 3321 if( !GenerateSynchronizationCode ) 3322 return; 3323 if (stopped()) { // Dead monitor? 3324 map()->pop_monitor(); // Kill monitor from debug info 3325 return; 3326 } 3327 3328 // Memory barrier to avoid floating things down past the locked region 3329 insert_mem_bar(Op_MemBarReleaseLock); 3330 3331 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3332 UnlockNode *unlock = new UnlockNode(C, tf); 3333 #ifdef ASSERT 3334 unlock->set_dbg_jvms(sync_jvms()); 3335 #endif 3336 uint raw_idx = Compile::AliasIdxRaw; 3337 unlock->init_req( TypeFunc::Control, control() ); 3338 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3339 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3340 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3341 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3342 3343 unlock->init_req(TypeFunc::Parms + 0, obj); 3344 unlock->init_req(TypeFunc::Parms + 1, box); 3345 unlock = _gvn.transform(unlock)->as_Unlock(); 3346 3347 Node* mem = reset_memory(); 3348 3349 // unlock has no side-effects, sets few values 3350 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3351 3352 // Kill monitor from debug info 3353 map()->pop_monitor( ); 3354 } 3355 3356 //-------------------------------get_layout_helper----------------------------- 3357 // If the given klass is a constant or known to be an array, 3358 // fetch the constant layout helper value into constant_value 3359 // and return (Node*)NULL. Otherwise, load the non-constant 3360 // layout helper value, and return the node which represents it. 3361 // This two-faced routine is useful because allocation sites 3362 // almost always feature constant types. 3363 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3364 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3365 if (!StressReflectiveCode && inst_klass != NULL) { 3366 ciKlass* klass = inst_klass->klass(); 3367 bool xklass = inst_klass->klass_is_exact(); 3368 if (xklass || klass->is_array_klass()) { 3369 jint lhelper = klass->layout_helper(); 3370 if (lhelper != Klass::_lh_neutral_value) { 3371 constant_value = lhelper; 3372 return (Node*) NULL; 3373 } 3374 } 3375 } 3376 constant_value = Klass::_lh_neutral_value; // put in a known value 3377 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3378 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3379 } 3380 3381 // We just put in an allocate/initialize with a big raw-memory effect. 3382 // Hook selected additional alias categories on the initialization. 3383 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3384 MergeMemNode* init_in_merge, 3385 Node* init_out_raw) { 3386 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3387 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3388 3389 Node* prevmem = kit.memory(alias_idx); 3390 init_in_merge->set_memory_at(alias_idx, prevmem); 3391 kit.set_memory(init_out_raw, alias_idx); 3392 } 3393 3394 //---------------------------set_output_for_allocation------------------------- 3395 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3396 const TypeOopPtr* oop_type, 3397 bool deoptimize_on_exception) { 3398 int rawidx = Compile::AliasIdxRaw; 3399 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3400 add_safepoint_edges(alloc); 3401 Node* allocx = _gvn.transform(alloc); 3402 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3403 // create memory projection for i_o 3404 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3405 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3406 3407 // create a memory projection as for the normal control path 3408 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3409 set_memory(malloc, rawidx); 3410 3411 // a normal slow-call doesn't change i_o, but an allocation does 3412 // we create a separate i_o projection for the normal control path 3413 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3414 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3415 3416 // put in an initialization barrier 3417 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3418 rawoop)->as_Initialize(); 3419 assert(alloc->initialization() == init, "2-way macro link must work"); 3420 assert(init ->allocation() == alloc, "2-way macro link must work"); 3421 { 3422 // Extract memory strands which may participate in the new object's 3423 // initialization, and source them from the new InitializeNode. 3424 // This will allow us to observe initializations when they occur, 3425 // and link them properly (as a group) to the InitializeNode. 3426 assert(init->in(InitializeNode::Memory) == malloc, ""); 3427 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3428 init->set_req(InitializeNode::Memory, minit_in); 3429 record_for_igvn(minit_in); // fold it up later, if possible 3430 Node* minit_out = memory(rawidx); 3431 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3432 // Add an edge in the MergeMem for the header fields so an access 3433 // to one of those has correct memory state 3434 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 3435 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 3436 if (oop_type->isa_aryptr()) { 3437 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3438 int elemidx = C->get_alias_index(telemref); 3439 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3440 } else if (oop_type->isa_instptr()) { 3441 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3442 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3443 ciField* field = ik->nonstatic_field_at(i); 3444 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3445 continue; // do not bother to track really large numbers of fields 3446 // Find (or create) the alias category for this field: 3447 int fieldidx = C->alias_type(field)->index(); 3448 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3449 } 3450 } 3451 } 3452 3453 // Cast raw oop to the real thing... 3454 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3455 javaoop = _gvn.transform(javaoop); 3456 C->set_recent_alloc(control(), javaoop); 3457 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3458 3459 #ifdef ASSERT 3460 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3461 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3462 "Ideal_allocation works"); 3463 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3464 "Ideal_allocation works"); 3465 if (alloc->is_AllocateArray()) { 3466 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3467 "Ideal_allocation works"); 3468 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3469 "Ideal_allocation works"); 3470 } else { 3471 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3472 } 3473 } 3474 #endif //ASSERT 3475 3476 return javaoop; 3477 } 3478 3479 //---------------------------new_instance-------------------------------------- 3480 // This routine takes a klass_node which may be constant (for a static type) 3481 // or may be non-constant (for reflective code). It will work equally well 3482 // for either, and the graph will fold nicely if the optimizer later reduces 3483 // the type to a constant. 3484 // The optional arguments are for specialized use by intrinsics: 3485 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3486 // - If 'return_size_val', report the the total object size to the caller. 3487 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3488 Node* GraphKit::new_instance(Node* klass_node, 3489 Node* extra_slow_test, 3490 Node* *return_size_val, 3491 bool deoptimize_on_exception) { 3492 // Compute size in doublewords 3493 // The size is always an integral number of doublewords, represented 3494 // as a positive bytewise size stored in the klass's layout_helper. 3495 // The layout_helper also encodes (in a low bit) the need for a slow path. 3496 jint layout_con = Klass::_lh_neutral_value; 3497 Node* layout_val = get_layout_helper(klass_node, layout_con); 3498 int layout_is_con = (layout_val == NULL); 3499 3500 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3501 // Generate the initial go-slow test. It's either ALWAYS (return a 3502 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3503 // case) a computed value derived from the layout_helper. 3504 Node* initial_slow_test = NULL; 3505 if (layout_is_con) { 3506 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3507 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3508 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3509 } else { // reflective case 3510 // This reflective path is used by Unsafe.allocateInstance. 3511 // (It may be stress-tested by specifying StressReflectiveCode.) 3512 // Basically, we want to get into the VM is there's an illegal argument. 3513 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3514 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3515 if (extra_slow_test != intcon(0)) { 3516 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3517 } 3518 // (Macro-expander will further convert this to a Bool, if necessary.) 3519 } 3520 3521 // Find the size in bytes. This is easy; it's the layout_helper. 3522 // The size value must be valid even if the slow path is taken. 3523 Node* size = NULL; 3524 if (layout_is_con) { 3525 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3526 } else { // reflective case 3527 // This reflective path is used by clone and Unsafe.allocateInstance. 3528 size = ConvI2X(layout_val); 3529 3530 // Clear the low bits to extract layout_helper_size_in_bytes: 3531 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3532 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3533 size = _gvn.transform( new AndXNode(size, mask) ); 3534 } 3535 if (return_size_val != NULL) { 3536 (*return_size_val) = size; 3537 } 3538 3539 // This is a precise notnull oop of the klass. 3540 // (Actually, it need not be precise if this is a reflective allocation.) 3541 // It's what we cast the result to. 3542 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3543 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3544 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3545 3546 // Now generate allocation code 3547 3548 // The entire memory state is needed for slow path of the allocation 3549 // since GC and deoptimization can happened. 3550 Node *mem = reset_memory(); 3551 set_all_memory(mem); // Create new memory state 3552 3553 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3554 control(), mem, i_o(), 3555 size, klass_node, 3556 initial_slow_test); 3557 3558 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3559 } 3560 3561 //-------------------------------new_array------------------------------------- 3562 // helper for both newarray and anewarray 3563 // The 'length' parameter is (obviously) the length of the array. 3564 // See comments on new_instance for the meaning of the other arguments. 3565 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3566 Node* length, // number of array elements 3567 int nargs, // number of arguments to push back for uncommon trap 3568 Node* *return_size_val, 3569 bool deoptimize_on_exception) { 3570 jint layout_con = Klass::_lh_neutral_value; 3571 Node* layout_val = get_layout_helper(klass_node, layout_con); 3572 int layout_is_con = (layout_val == NULL); 3573 3574 if (!layout_is_con && !StressReflectiveCode && 3575 !too_many_traps(Deoptimization::Reason_class_check)) { 3576 // This is a reflective array creation site. 3577 // Optimistically assume that it is a subtype of Object[], 3578 // so that we can fold up all the address arithmetic. 3579 layout_con = Klass::array_layout_helper(T_OBJECT); 3580 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3581 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3582 { BuildCutout unless(this, bol_lh, PROB_MAX); 3583 inc_sp(nargs); 3584 uncommon_trap(Deoptimization::Reason_class_check, 3585 Deoptimization::Action_maybe_recompile); 3586 } 3587 layout_val = NULL; 3588 layout_is_con = true; 3589 } 3590 3591 // Generate the initial go-slow test. Make sure we do not overflow 3592 // if length is huge (near 2Gig) or negative! We do not need 3593 // exact double-words here, just a close approximation of needed 3594 // double-words. We can't add any offset or rounding bits, lest we 3595 // take a size -1 of bytes and make it positive. Use an unsigned 3596 // compare, so negative sizes look hugely positive. 3597 int fast_size_limit = FastAllocateSizeLimit; 3598 if (layout_is_con) { 3599 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3600 // Increase the size limit if we have exact knowledge of array type. 3601 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3602 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3603 } 3604 3605 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3606 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3607 3608 // --- Size Computation --- 3609 // array_size = round_to_heap(array_header + (length << elem_shift)); 3610 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3611 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3612 // The rounding mask is strength-reduced, if possible. 3613 int round_mask = MinObjAlignmentInBytes - 1; 3614 Node* header_size = NULL; 3615 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3616 // (T_BYTE has the weakest alignment and size restrictions...) 3617 if (layout_is_con) { 3618 int hsize = Klass::layout_helper_header_size(layout_con); 3619 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3620 BasicType etype = Klass::layout_helper_element_type(layout_con); 3621 if ((round_mask & ~right_n_bits(eshift)) == 0) 3622 round_mask = 0; // strength-reduce it if it goes away completely 3623 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3624 assert(header_size_min <= hsize, "generic minimum is smallest"); 3625 header_size_min = hsize; 3626 header_size = intcon(hsize + round_mask); 3627 } else { 3628 Node* hss = intcon(Klass::_lh_header_size_shift); 3629 Node* hsm = intcon(Klass::_lh_header_size_mask); 3630 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3631 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3632 Node* mask = intcon(round_mask); 3633 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3634 } 3635 3636 Node* elem_shift = NULL; 3637 if (layout_is_con) { 3638 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3639 if (eshift != 0) 3640 elem_shift = intcon(eshift); 3641 } else { 3642 // There is no need to mask or shift this value. 3643 // The semantics of LShiftINode include an implicit mask to 0x1F. 3644 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3645 elem_shift = layout_val; 3646 } 3647 3648 // Transition to native address size for all offset calculations: 3649 Node* lengthx = ConvI2X(length); 3650 Node* headerx = ConvI2X(header_size); 3651 #ifdef _LP64 3652 { const TypeInt* tilen = _gvn.find_int_type(length); 3653 if (tilen != NULL && tilen->_lo < 0) { 3654 // Add a manual constraint to a positive range. Cf. array_element_address. 3655 jint size_max = fast_size_limit; 3656 if (size_max > tilen->_hi) size_max = tilen->_hi; 3657 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3658 3659 // Only do a narrow I2L conversion if the range check passed. 3660 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3661 _gvn.transform(iff); 3662 RegionNode* region = new RegionNode(3); 3663 _gvn.set_type(region, Type::CONTROL); 3664 lengthx = new PhiNode(region, TypeLong::LONG); 3665 _gvn.set_type(lengthx, TypeLong::LONG); 3666 3667 // Range check passed. Use ConvI2L node with narrow type. 3668 Node* passed = IfFalse(iff); 3669 region->init_req(1, passed); 3670 // Make I2L conversion control dependent to prevent it from 3671 // floating above the range check during loop optimizations. 3672 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3673 3674 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3675 region->init_req(2, IfTrue(iff)); 3676 lengthx->init_req(2, ConvI2X(length)); 3677 3678 set_control(region); 3679 record_for_igvn(region); 3680 record_for_igvn(lengthx); 3681 } 3682 } 3683 #endif 3684 3685 // Combine header size (plus rounding) and body size. Then round down. 3686 // This computation cannot overflow, because it is used only in two 3687 // places, one where the length is sharply limited, and the other 3688 // after a successful allocation. 3689 Node* abody = lengthx; 3690 if (elem_shift != NULL) 3691 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3692 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3693 if (round_mask != 0) { 3694 Node* mask = MakeConX(~round_mask); 3695 size = _gvn.transform( new AndXNode(size, mask) ); 3696 } 3697 // else if round_mask == 0, the size computation is self-rounding 3698 3699 if (return_size_val != NULL) { 3700 // This is the size 3701 (*return_size_val) = size; 3702 } 3703 3704 // Now generate allocation code 3705 3706 // The entire memory state is needed for slow path of the allocation 3707 // since GC and deoptimization can happened. 3708 Node *mem = reset_memory(); 3709 set_all_memory(mem); // Create new memory state 3710 3711 if (initial_slow_test->is_Bool()) { 3712 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3713 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3714 } 3715 3716 // Create the AllocateArrayNode and its result projections 3717 AllocateArrayNode* alloc 3718 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3719 control(), mem, i_o(), 3720 size, klass_node, 3721 initial_slow_test, 3722 length); 3723 3724 // Cast to correct type. Note that the klass_node may be constant or not, 3725 // and in the latter case the actual array type will be inexact also. 3726 // (This happens via a non-constant argument to inline_native_newArray.) 3727 // In any case, the value of klass_node provides the desired array type. 3728 const TypeInt* length_type = _gvn.find_int_type(length); 3729 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3730 if (ary_type->isa_aryptr() && length_type != NULL) { 3731 // Try to get a better type than POS for the size 3732 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3733 } 3734 3735 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3736 3737 // Cast length on remaining path to be as narrow as possible 3738 if (map()->find_edge(length) >= 0) { 3739 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3740 if (ccast != length) { 3741 _gvn.set_type_bottom(ccast); 3742 record_for_igvn(ccast); 3743 replace_in_map(length, ccast); 3744 } 3745 } 3746 3747 return javaoop; 3748 } 3749 3750 // The following "Ideal_foo" functions are placed here because they recognize 3751 // the graph shapes created by the functions immediately above. 3752 3753 //---------------------------Ideal_allocation---------------------------------- 3754 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3755 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3756 if (ptr == NULL) { // reduce dumb test in callers 3757 return NULL; 3758 } 3759 3760 #if INCLUDE_SHENANDOAHGC 3761 if (UseShenandoahGC) { 3762 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 3763 ptr = bs->step_over_gc_barrier(ptr); 3764 } 3765 #endif 3766 3767 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3768 ptr = ptr->in(1); 3769 if (ptr == NULL) return NULL; 3770 } 3771 // Return NULL for allocations with several casts: 3772 // j.l.reflect.Array.newInstance(jobject, jint) 3773 // Object.clone() 3774 // to keep more precise type from last cast. 3775 if (ptr->is_Proj()) { 3776 Node* allo = ptr->in(0); 3777 if (allo != NULL && allo->is_Allocate()) { 3778 return allo->as_Allocate(); 3779 } 3780 } 3781 // Report failure to match. 3782 return NULL; 3783 } 3784 3785 // Fancy version which also strips off an offset (and reports it to caller). 3786 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3787 intptr_t& offset) { 3788 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3789 if (base == NULL) return NULL; 3790 return Ideal_allocation(base, phase); 3791 } 3792 3793 // Trace Initialize <- Proj[Parm] <- Allocate 3794 AllocateNode* InitializeNode::allocation() { 3795 Node* rawoop = in(InitializeNode::RawAddress); 3796 if (rawoop->is_Proj()) { 3797 Node* alloc = rawoop->in(0); 3798 if (alloc->is_Allocate()) { 3799 return alloc->as_Allocate(); 3800 } 3801 } 3802 return NULL; 3803 } 3804 3805 // Trace Allocate -> Proj[Parm] -> Initialize 3806 InitializeNode* AllocateNode::initialization() { 3807 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3808 if (rawoop == NULL) return NULL; 3809 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3810 Node* init = rawoop->fast_out(i); 3811 if (init->is_Initialize()) { 3812 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3813 return init->as_Initialize(); 3814 } 3815 } 3816 return NULL; 3817 } 3818 3819 //----------------------------- loop predicates --------------------------- 3820 3821 //------------------------------add_predicate_impl---------------------------- 3822 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3823 // Too many traps seen? 3824 if (too_many_traps(reason)) { 3825 #ifdef ASSERT 3826 if (TraceLoopPredicate) { 3827 int tc = C->trap_count(reason); 3828 tty->print("too many traps=%s tcount=%d in ", 3829 Deoptimization::trap_reason_name(reason), tc); 3830 method()->print(); // which method has too many predicate traps 3831 tty->cr(); 3832 } 3833 #endif 3834 // We cannot afford to take more traps here, 3835 // do not generate predicate. 3836 return; 3837 } 3838 3839 Node *cont = _gvn.intcon(1); 3840 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3841 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3842 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3843 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3844 C->add_predicate_opaq(opq); 3845 { 3846 PreserveJVMState pjvms(this); 3847 set_control(iffalse); 3848 inc_sp(nargs); 3849 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3850 } 3851 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3852 set_control(iftrue); 3853 } 3854 3855 //------------------------------add_predicate--------------------------------- 3856 void GraphKit::add_empty_predicates(int nargs) { 3857 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding 3858 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of 3859 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate. 3860 if (UseLoopPredicate) { 3861 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs); 3862 } 3863 if (UseProfiledLoopPredicate) { 3864 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs); 3865 } 3866 // loop's limit check predicate should be near the loop. 3867 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3868 } 3869 3870 void GraphKit::sync_kit(IdealKit& ideal) { 3871 set_all_memory(ideal.merged_memory()); 3872 set_i_o(ideal.i_o()); 3873 set_control(ideal.ctrl()); 3874 } 3875 3876 void GraphKit::final_sync(IdealKit& ideal) { 3877 // Final sync IdealKit and graphKit. 3878 sync_kit(ideal); 3879 } 3880 3881 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 3882 Node* len = load_array_length(load_String_value(ctrl, str)); 3883 Node* coder = load_String_coder(ctrl, str); 3884 // Divide length by 2 if coder is UTF16 3885 return _gvn.transform(new RShiftINode(len, coder)); 3886 } 3887 3888 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 3889 int value_offset = java_lang_String::value_offset_in_bytes(); 3890 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3891 false, NULL, 0); 3892 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3893 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 3894 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 3895 ciTypeArrayKlass::make(T_BYTE), true, 0); 3896 Node* p = basic_plus_adr(str, str, value_offset); 3897 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 3898 IN_HEAP | C2_CONTROL_DEPENDENT_LOAD); 3899 // String.value field is known to be @Stable. 3900 if (UseImplicitStableValues) { 3901 load = cast_array_to_stable(load, value_type); 3902 } 3903 return load; 3904 } 3905 3906 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 3907 if (!CompactStrings) { 3908 return intcon(java_lang_String::CODER_UTF16); 3909 } 3910 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3911 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3912 false, NULL, 0); 3913 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3914 int coder_field_idx = C->get_alias_index(coder_field_type); 3915 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 3916 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 3917 } 3918 3919 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 3920 int value_offset = java_lang_String::value_offset_in_bytes(); 3921 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3922 false, NULL, 0); 3923 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3924 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 3925 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP); 3926 } 3927 3928 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 3929 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3930 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3931 false, NULL, 0); 3932 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3933 int coder_field_idx = C->get_alias_index(coder_field_type); 3934 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 3935 value, T_BYTE, coder_field_idx, MemNode::unordered); 3936 } 3937 3938 // Capture src and dst memory state with a MergeMemNode 3939 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 3940 if (src_type == dst_type) { 3941 // Types are equal, we don't need a MergeMemNode 3942 return memory(src_type); 3943 } 3944 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 3945 record_for_igvn(merge); // fold it up later, if possible 3946 int src_idx = C->get_alias_index(src_type); 3947 int dst_idx = C->get_alias_index(dst_type); 3948 merge->set_memory_at(src_idx, memory(src_idx)); 3949 merge->set_memory_at(dst_idx, memory(dst_idx)); 3950 return merge; 3951 } 3952 3953 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 3954 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 3955 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 3956 // If input and output memory types differ, capture both states to preserve 3957 // the dependency between preceding and subsequent loads/stores. 3958 // For example, the following program: 3959 // StoreB 3960 // compress_string 3961 // LoadB 3962 // has this memory graph (use->def): 3963 // LoadB -> compress_string -> CharMem 3964 // ... -> StoreB -> ByteMem 3965 // The intrinsic hides the dependency between LoadB and StoreB, causing 3966 // the load to read from memory not containing the result of the StoreB. 3967 // The correct memory graph should look like this: 3968 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 3969 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 3970 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 3971 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 3972 set_memory(res_mem, TypeAryPtr::BYTES); 3973 return str; 3974 } 3975 3976 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 3977 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 3978 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 3979 // Capture src and dst memory (see comment in 'compress_string'). 3980 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 3981 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 3982 set_memory(_gvn.transform(str), dst_type); 3983 } 3984 3985 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 3986 /** 3987 * int i_char = start; 3988 * for (int i_byte = 0; i_byte < count; i_byte++) { 3989 * dst[i_char++] = (char)(src[i_byte] & 0xff); 3990 * } 3991 */ 3992 add_empty_predicates(); 3993 RegionNode* head = new RegionNode(3); 3994 head->init_req(1, control()); 3995 gvn().set_type(head, Type::CONTROL); 3996 record_for_igvn(head); 3997 3998 Node* i_byte = new PhiNode(head, TypeInt::INT); 3999 i_byte->init_req(1, intcon(0)); 4000 gvn().set_type(i_byte, TypeInt::INT); 4001 record_for_igvn(i_byte); 4002 4003 Node* i_char = new PhiNode(head, TypeInt::INT); 4004 i_char->init_req(1, start); 4005 gvn().set_type(i_char, TypeInt::INT); 4006 record_for_igvn(i_char); 4007 4008 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4009 gvn().set_type(mem, Type::MEMORY); 4010 record_for_igvn(mem); 4011 set_control(head); 4012 set_memory(mem, TypeAryPtr::BYTES); 4013 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4014 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4015 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4016 false, false, true /* mismatched */); 4017 4018 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4019 head->init_req(2, IfTrue(iff)); 4020 mem->init_req(2, st); 4021 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4022 i_char->init_req(2, AddI(i_char, intcon(2))); 4023 4024 set_control(IfFalse(iff)); 4025 set_memory(st, TypeAryPtr::BYTES); 4026 } 4027 4028 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4029 if (!field->is_constant()) { 4030 return NULL; // Field not marked as constant. 4031 } 4032 ciInstance* holder = NULL; 4033 if (!field->is_static()) { 4034 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4035 if (const_oop != NULL && const_oop->is_instance()) { 4036 holder = const_oop->as_instance(); 4037 } 4038 } 4039 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4040 /*is_unsigned_load=*/false); 4041 if (con_type != NULL) { 4042 return makecon(con_type); 4043 } 4044 return NULL; 4045 } 4046 4047 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4048 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4049 // assumption of CCP analysis. 4050 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4051 }