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 823 // deoptimization happens 824 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 825 ciMethod* cur_method = jvms->method(); 826 int cur_bci = jvms->bci(); 827 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 828 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 829 return Interpreter::bytecode_should_reexecute(code) || 830 (is_anewarray && code == Bytecodes::_multianewarray); 831 // Reexecute _multianewarray bytecode which was replaced with 832 // sequence of [a]newarray. See Parse::do_multianewarray(). 833 // 834 // Note: interpreter should not have it set since this optimization 835 // is limited by dimensions and guarded by flag so in some cases 836 // multianewarray() runtime calls will be generated and 837 // the bytecode should not be reexecutes (stack will not be reset). 838 } else 839 return false; 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 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 894 } 895 896 // Presize the call: 897 DEBUG_ONLY(uint non_debug_edges = call->req()); 898 call->add_req_batch(top(), youngest_jvms->debug_depth()); 899 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 900 901 // Set up edges so that the call looks like this: 902 // Call [state:] ctl io mem fptr retadr 903 // [parms:] parm0 ... parmN 904 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 905 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 906 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 907 // Note that caller debug info precedes callee debug info. 908 909 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 910 uint debug_ptr = call->req(); 911 912 // Loop over the map input edges associated with jvms, add them 913 // to the call node, & reset all offsets to match call node array. 914 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 915 uint debug_end = debug_ptr; 916 uint debug_start = debug_ptr - in_jvms->debug_size(); 917 debug_ptr = debug_start; // back up the ptr 918 919 uint p = debug_start; // walks forward in [debug_start, debug_end) 920 uint j, k, l; 921 SafePointNode* in_map = in_jvms->map(); 922 out_jvms->set_map(call); 923 924 if (can_prune_locals) { 925 assert(in_jvms->method() == out_jvms->method(), "sanity"); 926 // If the current throw can reach an exception handler in this JVMS, 927 // then we must keep everything live that can reach that handler. 928 // As a quick and dirty approximation, we look for any handlers at all. 929 if (in_jvms->method()->has_exception_handlers()) { 930 can_prune_locals = false; 931 } 932 } 933 934 // Add the Locals 935 k = in_jvms->locoff(); 936 l = in_jvms->loc_size(); 937 out_jvms->set_locoff(p); 938 if (!can_prune_locals) { 939 for (j = 0; j < l; j++) 940 call->set_req(p++, in_map->in(k+j)); 941 } else { 942 p += l; // already set to top above by add_req_batch 943 } 944 945 // Add the Expression Stack 946 k = in_jvms->stkoff(); 947 l = in_jvms->sp(); 948 out_jvms->set_stkoff(p); 949 if (!can_prune_locals) { 950 for (j = 0; j < l; j++) 951 call->set_req(p++, in_map->in(k+j)); 952 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 953 // Divide stack into {S0,...,S1}, where S0 is set to top. 954 uint s1 = stack_slots_not_pruned; 955 stack_slots_not_pruned = 0; // for next iteration 956 if (s1 > l) s1 = l; 957 uint s0 = l - s1; 958 p += s0; // skip the tops preinstalled by add_req_batch 959 for (j = s0; j < l; j++) 960 call->set_req(p++, in_map->in(k+j)); 961 } else { 962 p += l; // already set to top above by add_req_batch 963 } 964 965 // Add the Monitors 966 k = in_jvms->monoff(); 967 l = in_jvms->mon_size(); 968 out_jvms->set_monoff(p); 969 for (j = 0; j < l; j++) 970 call->set_req(p++, in_map->in(k+j)); 971 972 // Copy any scalar object fields. 973 k = in_jvms->scloff(); 974 l = in_jvms->scl_size(); 975 out_jvms->set_scloff(p); 976 for (j = 0; j < l; j++) 977 call->set_req(p++, in_map->in(k+j)); 978 979 // Finish the new jvms. 980 out_jvms->set_endoff(p); 981 982 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 983 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 984 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 985 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 986 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 987 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 988 989 // Update the two tail pointers in parallel. 990 out_jvms = out_jvms->caller(); 991 in_jvms = in_jvms->caller(); 992 } 993 994 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 995 996 // Test the correctness of JVMState::debug_xxx accessors: 997 assert(call->jvms()->debug_start() == non_debug_edges, ""); 998 assert(call->jvms()->debug_end() == call->req(), ""); 999 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1000 } 1001 1002 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1003 Bytecodes::Code code = java_bc(); 1004 if (code == Bytecodes::_wide) { 1005 code = method()->java_code_at_bci(bci() + 1); 1006 } 1007 1008 BasicType rtype = T_ILLEGAL; 1009 int rsize = 0; 1010 1011 if (code != Bytecodes::_illegal) { 1012 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1013 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1014 if (rtype < T_CONFLICT) 1015 rsize = type2size[rtype]; 1016 } 1017 1018 switch (code) { 1019 case Bytecodes::_illegal: 1020 return false; 1021 1022 case Bytecodes::_ldc: 1023 case Bytecodes::_ldc_w: 1024 case Bytecodes::_ldc2_w: 1025 inputs = 0; 1026 break; 1027 1028 case Bytecodes::_dup: inputs = 1; break; 1029 case Bytecodes::_dup_x1: inputs = 2; break; 1030 case Bytecodes::_dup_x2: inputs = 3; break; 1031 case Bytecodes::_dup2: inputs = 2; break; 1032 case Bytecodes::_dup2_x1: inputs = 3; break; 1033 case Bytecodes::_dup2_x2: inputs = 4; break; 1034 case Bytecodes::_swap: inputs = 2; break; 1035 case Bytecodes::_arraylength: inputs = 1; break; 1036 1037 case Bytecodes::_getstatic: 1038 case Bytecodes::_putstatic: 1039 case Bytecodes::_getfield: 1040 case Bytecodes::_putfield: 1041 { 1042 bool ignored_will_link; 1043 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1044 int size = field->type()->size(); 1045 bool is_get = (depth >= 0), is_static = (depth & 1); 1046 inputs = (is_static ? 0 : 1); 1047 if (is_get) { 1048 depth = size - inputs; 1049 } else { 1050 inputs += size; // putxxx pops the value from the stack 1051 depth = - inputs; 1052 } 1053 } 1054 break; 1055 1056 case Bytecodes::_invokevirtual: 1057 case Bytecodes::_invokespecial: 1058 case Bytecodes::_invokestatic: 1059 case Bytecodes::_invokedynamic: 1060 case Bytecodes::_invokeinterface: 1061 { 1062 bool ignored_will_link; 1063 ciSignature* declared_signature = NULL; 1064 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1065 assert(declared_signature != NULL, "cannot be null"); 1066 inputs = declared_signature->arg_size_for_bc(code); 1067 int size = declared_signature->return_type()->size(); 1068 depth = size - inputs; 1069 } 1070 break; 1071 1072 case Bytecodes::_multianewarray: 1073 { 1074 ciBytecodeStream iter(method()); 1075 iter.reset_to_bci(bci()); 1076 iter.next(); 1077 inputs = iter.get_dimensions(); 1078 assert(rsize == 1, ""); 1079 depth = rsize - inputs; 1080 } 1081 break; 1082 1083 case Bytecodes::_ireturn: 1084 case Bytecodes::_lreturn: 1085 case Bytecodes::_freturn: 1086 case Bytecodes::_dreturn: 1087 case Bytecodes::_areturn: 1088 assert(rsize == -depth, ""); 1089 inputs = rsize; 1090 break; 1091 1092 case Bytecodes::_jsr: 1093 case Bytecodes::_jsr_w: 1094 inputs = 0; 1095 depth = 1; // S.B. depth=1, not zero 1096 break; 1097 1098 default: 1099 // bytecode produces a typed result 1100 inputs = rsize - depth; 1101 assert(inputs >= 0, ""); 1102 break; 1103 } 1104 1105 #ifdef ASSERT 1106 // spot check 1107 int outputs = depth + inputs; 1108 assert(outputs >= 0, "sanity"); 1109 switch (code) { 1110 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1111 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1112 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1113 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1114 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1115 default: break; 1116 } 1117 #endif //ASSERT 1118 1119 return true; 1120 } 1121 1122 1123 1124 //------------------------------basic_plus_adr--------------------------------- 1125 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1126 // short-circuit a common case 1127 if (offset == intcon(0)) return ptr; 1128 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1129 } 1130 1131 Node* GraphKit::ConvI2L(Node* offset) { 1132 // short-circuit a common case 1133 jint offset_con = find_int_con(offset, Type::OffsetBot); 1134 if (offset_con != Type::OffsetBot) { 1135 return longcon((jlong) offset_con); 1136 } 1137 return _gvn.transform( new ConvI2LNode(offset)); 1138 } 1139 1140 Node* GraphKit::ConvI2UL(Node* offset) { 1141 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1142 if (offset_con != (juint) Type::OffsetBot) { 1143 return longcon((julong) offset_con); 1144 } 1145 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1146 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1147 return _gvn.transform( new AndLNode(conv, mask) ); 1148 } 1149 1150 Node* GraphKit::ConvL2I(Node* offset) { 1151 // short-circuit a common case 1152 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1153 if (offset_con != (jlong)Type::OffsetBot) { 1154 return intcon((int) offset_con); 1155 } 1156 return _gvn.transform( new ConvL2INode(offset)); 1157 } 1158 1159 //-------------------------load_object_klass----------------------------------- 1160 Node* GraphKit::load_object_klass(Node* obj) { 1161 // Special-case a fresh allocation to avoid building nodes: 1162 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1163 if (akls != NULL) return akls; 1164 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1165 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1166 } 1167 1168 //-------------------------load_array_length----------------------------------- 1169 Node* GraphKit::load_array_length(Node* array) { 1170 // Special-case a fresh allocation to avoid building nodes: 1171 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1172 Node *alen; 1173 if (alloc == NULL) { 1174 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1175 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1176 } else { 1177 alen = alloc->Ideal_length(); 1178 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1179 if (ccast != alen) { 1180 alen = _gvn.transform(ccast); 1181 } 1182 } 1183 return alen; 1184 } 1185 1186 //------------------------------do_null_check---------------------------------- 1187 // Helper function to do a NULL pointer check. Returned value is 1188 // the incoming address with NULL casted away. You are allowed to use the 1189 // not-null value only if you are control dependent on the test. 1190 #ifndef PRODUCT 1191 extern int explicit_null_checks_inserted, 1192 explicit_null_checks_elided; 1193 #endif 1194 Node* GraphKit::null_check_common(Node* value, BasicType type, 1195 // optional arguments for variations: 1196 bool assert_null, 1197 Node* *null_control, 1198 bool speculative) { 1199 assert(!assert_null || null_control == NULL, "not both at once"); 1200 if (stopped()) return top(); 1201 NOT_PRODUCT(explicit_null_checks_inserted++); 1202 1203 // Construct NULL check 1204 Node *chk = NULL; 1205 switch(type) { 1206 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1207 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1208 case T_ARRAY : // fall through 1209 type = T_OBJECT; // simplify further tests 1210 case T_OBJECT : { 1211 const Type *t = _gvn.type( value ); 1212 1213 const TypeOopPtr* tp = t->isa_oopptr(); 1214 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1215 // Only for do_null_check, not any of its siblings: 1216 && !assert_null && null_control == NULL) { 1217 // Usually, any field access or invocation on an unloaded oop type 1218 // will simply fail to link, since the statically linked class is 1219 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1220 // the static class is loaded but the sharper oop type is not. 1221 // Rather than checking for this obscure case in lots of places, 1222 // we simply observe that a null check on an unloaded class 1223 // will always be followed by a nonsense operation, so we 1224 // can just issue the uncommon trap here. 1225 // Our access to the unloaded class will only be correct 1226 // after it has been loaded and initialized, which requires 1227 // a trip through the interpreter. 1228 #ifndef PRODUCT 1229 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1230 #endif 1231 uncommon_trap(Deoptimization::Reason_unloaded, 1232 Deoptimization::Action_reinterpret, 1233 tp->klass(), "!loaded"); 1234 return top(); 1235 } 1236 1237 if (assert_null) { 1238 // See if the type is contained in NULL_PTR. 1239 // If so, then the value is already null. 1240 if (t->higher_equal(TypePtr::NULL_PTR)) { 1241 NOT_PRODUCT(explicit_null_checks_elided++); 1242 return value; // Elided null assert quickly! 1243 } 1244 } else { 1245 // See if mixing in the NULL pointer changes type. 1246 // If so, then the NULL pointer was not allowed in the original 1247 // type. In other words, "value" was not-null. 1248 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1249 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1250 NOT_PRODUCT(explicit_null_checks_elided++); 1251 return value; // Elided null check quickly! 1252 } 1253 } 1254 chk = new CmpPNode( value, null() ); 1255 break; 1256 } 1257 1258 default: 1259 fatal("unexpected type: %s", type2name(type)); 1260 } 1261 assert(chk != NULL, "sanity check"); 1262 chk = _gvn.transform(chk); 1263 1264 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1265 BoolNode *btst = new BoolNode( chk, btest); 1266 Node *tst = _gvn.transform( btst ); 1267 1268 //----------- 1269 // if peephole optimizations occurred, a prior test existed. 1270 // If a prior test existed, maybe it dominates as we can avoid this test. 1271 if (tst != btst && type == T_OBJECT) { 1272 // At this point we want to scan up the CFG to see if we can 1273 // find an identical test (and so avoid this test altogether). 1274 Node *cfg = control(); 1275 int depth = 0; 1276 while( depth < 16 ) { // Limit search depth for speed 1277 if( cfg->Opcode() == Op_IfTrue && 1278 cfg->in(0)->in(1) == tst ) { 1279 // Found prior test. Use "cast_not_null" to construct an identical 1280 // CastPP (and hence hash to) as already exists for the prior test. 1281 // Return that casted value. 1282 if (assert_null) { 1283 replace_in_map(value, null()); 1284 return null(); // do not issue the redundant test 1285 } 1286 Node *oldcontrol = control(); 1287 set_control(cfg); 1288 Node *res = cast_not_null(value); 1289 set_control(oldcontrol); 1290 NOT_PRODUCT(explicit_null_checks_elided++); 1291 return res; 1292 } 1293 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1294 if (cfg == NULL) break; // Quit at region nodes 1295 depth++; 1296 } 1297 } 1298 1299 //----------- 1300 // Branch to failure if null 1301 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1302 Deoptimization::DeoptReason reason; 1303 if (assert_null) { 1304 reason = Deoptimization::reason_null_assert(speculative); 1305 } else if (type == T_OBJECT) { 1306 reason = Deoptimization::reason_null_check(speculative); 1307 } else { 1308 reason = Deoptimization::Reason_div0_check; 1309 } 1310 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1311 // ciMethodData::has_trap_at will return a conservative -1 if any 1312 // must-be-null assertion has failed. This could cause performance 1313 // problems for a method after its first do_null_assert failure. 1314 // Consider using 'Reason_class_check' instead? 1315 1316 // To cause an implicit null check, we set the not-null probability 1317 // to the maximum (PROB_MAX). For an explicit check the probability 1318 // is set to a smaller value. 1319 if (null_control != NULL || too_many_traps(reason)) { 1320 // probability is less likely 1321 ok_prob = PROB_LIKELY_MAG(3); 1322 } else if (!assert_null && 1323 (ImplicitNullCheckThreshold > 0) && 1324 method() != NULL && 1325 (method()->method_data()->trap_count(reason) 1326 >= (uint)ImplicitNullCheckThreshold)) { 1327 ok_prob = PROB_LIKELY_MAG(3); 1328 } 1329 1330 if (null_control != NULL) { 1331 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1332 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1333 set_control( _gvn.transform( new IfTrueNode(iff))); 1334 #ifndef PRODUCT 1335 if (null_true == top()) { 1336 explicit_null_checks_elided++; 1337 } 1338 #endif 1339 (*null_control) = null_true; 1340 } else { 1341 BuildCutout unless(this, tst, ok_prob); 1342 // Check for optimizer eliding test at parse time 1343 if (stopped()) { 1344 // Failure not possible; do not bother making uncommon trap. 1345 NOT_PRODUCT(explicit_null_checks_elided++); 1346 } else if (assert_null) { 1347 uncommon_trap(reason, 1348 Deoptimization::Action_make_not_entrant, 1349 NULL, "assert_null"); 1350 } else { 1351 replace_in_map(value, zerocon(type)); 1352 builtin_throw(reason); 1353 } 1354 } 1355 1356 // Must throw exception, fall-thru not possible? 1357 if (stopped()) { 1358 return top(); // No result 1359 } 1360 1361 if (assert_null) { 1362 // Cast obj to null on this path. 1363 replace_in_map(value, zerocon(type)); 1364 return zerocon(type); 1365 } 1366 1367 // Cast obj to not-null on this path, if there is no null_control. 1368 // (If there is a null_control, a non-null value may come back to haunt us.) 1369 return cast_not_null(value, (null_control == NULL || (*null_control) == top())); 1370 } 1371 1372 1373 //------------------------------cast_not_null---------------------------------- 1374 // Cast obj to not-null on this path 1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1376 Node* cast = NULL; 1377 const Type* t = _gvn.type(obj); 1378 if (t->make_ptr() != NULL) { 1379 const Type* t_not_null = t->join_speculative(TypePtr::NOTNULL); 1380 // Object is already not-null? 1381 if (t == t_not_null) { 1382 return obj; 1383 } 1384 cast = ConstraintCastNode::make_cast(Op_CastPP, control(), obj, t_not_null, false); 1385 } else if (t->isa_int() != NULL) { 1386 cast = ConstraintCastNode::make_cast(Op_CastII, control(), obj, TypeInt::INT, true); 1387 } else if (t->isa_long() != NULL) { 1388 cast = ConstraintCastNode::make_cast(Op_CastLL, control(), obj, TypeLong::LONG, true); 1389 } else { 1390 fatal("unexpected type: %s", type2name(t->basic_type())); 1391 } 1392 cast = _gvn.transform(cast); 1393 1394 // Scan for instances of 'obj' in the current JVM mapping. 1395 // These instances are known to be not-null after the test. 1396 if (do_replace_in_map) { 1397 replace_in_map(obj, cast); 1398 } 1399 return cast; 1400 } 1401 1402 // Sometimes in intrinsics, we implicitly know an object is not null 1403 // (there's no actual null check) so we can cast it to not null. In 1404 // the course of optimizations, the input to the cast can become null. 1405 // In that case that data path will die and we need the control path 1406 // to become dead as well to keep the graph consistent. So we have to 1407 // add a check for null for which one branch can't be taken. It uses 1408 // an Opaque4 node that will cause the check to be removed after loop 1409 // opts so the test goes away and the compiled code doesn't execute a 1410 // useless check. 1411 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1412 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1413 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1414 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1415 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1416 _gvn.set_type(iff, iff->Value(&_gvn)); 1417 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1418 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1419 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1420 C->root()->add_req(halt); 1421 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1422 set_control(if_t); 1423 return cast_not_null(value, do_replace_in_map); 1424 } 1425 1426 1427 //--------------------------replace_in_map------------------------------------- 1428 void GraphKit::replace_in_map(Node* old, Node* neww) { 1429 if (old == neww) { 1430 return; 1431 } 1432 1433 map()->replace_edge(old, neww); 1434 1435 // Note: This operation potentially replaces any edge 1436 // on the map. This includes locals, stack, and monitors 1437 // of the current (innermost) JVM state. 1438 1439 // don't let inconsistent types from profiling escape this 1440 // method 1441 1442 const Type* told = _gvn.type(old); 1443 const Type* tnew = _gvn.type(neww); 1444 1445 if (!tnew->higher_equal(told)) { 1446 return; 1447 } 1448 1449 map()->record_replaced_node(old, neww); 1450 } 1451 1452 1453 //============================================================================= 1454 //--------------------------------memory--------------------------------------- 1455 Node* GraphKit::memory(uint alias_idx) { 1456 MergeMemNode* mem = merged_memory(); 1457 Node* p = mem->memory_at(alias_idx); 1458 _gvn.set_type(p, Type::MEMORY); // must be mapped 1459 return p; 1460 } 1461 1462 //-----------------------------reset_memory------------------------------------ 1463 Node* GraphKit::reset_memory() { 1464 Node* mem = map()->memory(); 1465 // do not use this node for any more parsing! 1466 debug_only( map()->set_memory((Node*)NULL) ); 1467 return _gvn.transform( mem ); 1468 } 1469 1470 //------------------------------set_all_memory--------------------------------- 1471 void GraphKit::set_all_memory(Node* newmem) { 1472 Node* mergemem = MergeMemNode::make(newmem); 1473 gvn().set_type_bottom(mergemem); 1474 map()->set_memory(mergemem); 1475 } 1476 1477 //------------------------------set_all_memory_call---------------------------- 1478 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1479 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1480 set_all_memory(newmem); 1481 } 1482 1483 //============================================================================= 1484 // 1485 // parser factory methods for MemNodes 1486 // 1487 // These are layered on top of the factory methods in LoadNode and StoreNode, 1488 // and integrate with the parser's memory state and _gvn engine. 1489 // 1490 1491 // factory methods in "int adr_idx" 1492 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1493 int adr_idx, 1494 MemNode::MemOrd mo, 1495 LoadNode::ControlDependency control_dependency, 1496 bool require_atomic_access, 1497 bool unaligned, 1498 bool mismatched, 1499 bool unsafe) { 1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1501 const TypePtr* adr_type = NULL; // debug-mode-only argument 1502 debug_only(adr_type = C->get_adr_type(adr_idx)); 1503 Node* mem = memory(adr_idx); 1504 Node* ld; 1505 if (require_atomic_access && bt == T_LONG) { 1506 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe); 1507 } else if (require_atomic_access && bt == T_DOUBLE) { 1508 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe); 1509 } else { 1510 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe); 1511 } 1512 ld = _gvn.transform(ld); 1513 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1514 // Improve graph before escape analysis and boxing elimination. 1515 record_for_igvn(ld); 1516 } 1517 return ld; 1518 } 1519 1520 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1521 int adr_idx, 1522 MemNode::MemOrd mo, 1523 bool require_atomic_access, 1524 bool unaligned, 1525 bool mismatched, 1526 bool unsafe) { 1527 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1528 const TypePtr* adr_type = NULL; 1529 debug_only(adr_type = C->get_adr_type(adr_idx)); 1530 Node *mem = memory(adr_idx); 1531 Node* st; 1532 if (require_atomic_access && bt == T_LONG) { 1533 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1534 } else if (require_atomic_access && bt == T_DOUBLE) { 1535 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1536 } else { 1537 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1538 } 1539 if (unaligned) { 1540 st->as_Store()->set_unaligned_access(); 1541 } 1542 if (mismatched) { 1543 st->as_Store()->set_mismatched_access(); 1544 } 1545 if (unsafe) { 1546 st->as_Store()->set_unsafe_access(); 1547 } 1548 st = _gvn.transform(st); 1549 set_memory(st, adr_idx); 1550 // Back-to-back stores can only remove intermediate store with DU info 1551 // so push on worklist for optimizer. 1552 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1553 record_for_igvn(st); 1554 1555 return st; 1556 } 1557 1558 Node* GraphKit::access_store_at(Node* ctl, 1559 Node* obj, 1560 Node* adr, 1561 const TypePtr* adr_type, 1562 Node* val, 1563 const Type* val_type, 1564 BasicType bt, 1565 DecoratorSet decorators) { 1566 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1567 // could be delayed during Parse (for example, in adjust_map_after_if()). 1568 // Execute transformation here to avoid barrier generation in such case. 1569 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1570 val = _gvn.makecon(TypePtr::NULL_PTR); 1571 } 1572 1573 set_control(ctl); 1574 if (stopped()) { 1575 return top(); // Dead path ? 1576 } 1577 1578 assert(val != NULL, "not dead path"); 1579 1580 C2AccessValuePtr addr(adr, adr_type); 1581 C2AccessValue value(val, val_type); 1582 C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1583 if (access.is_raw()) { 1584 return _barrier_set->BarrierSetC2::store_at(access, value); 1585 } else { 1586 return _barrier_set->store_at(access, value); 1587 } 1588 } 1589 1590 Node* GraphKit::access_load_at(Node* obj, // containing obj 1591 Node* adr, // actual adress to store val at 1592 const TypePtr* adr_type, 1593 const Type* val_type, 1594 BasicType bt, 1595 DecoratorSet decorators) { 1596 if (stopped()) { 1597 return top(); // Dead path ? 1598 } 1599 1600 C2AccessValuePtr addr(adr, adr_type); 1601 C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr); 1602 if (access.is_raw()) { 1603 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1604 } else { 1605 return _barrier_set->load_at(access, val_type); 1606 } 1607 } 1608 1609 Node* GraphKit::access_load(Node* adr, // actual adress to load val at 1610 const Type* val_type, 1611 BasicType bt, 1612 DecoratorSet decorators) { 1613 if (stopped()) { 1614 return top(); // Dead path ? 1615 } 1616 1617 C2AccessValuePtr addr(adr, NULL); 1618 C2Access access(this, decorators | C2_READ_ACCESS, bt, NULL, addr); 1619 if (access.is_raw()) { 1620 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1621 } else { 1622 return _barrier_set->load_at(access, val_type); 1623 } 1624 } 1625 1626 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl, 1627 Node* obj, 1628 Node* adr, 1629 const TypePtr* adr_type, 1630 int alias_idx, 1631 Node* expected_val, 1632 Node* new_val, 1633 const Type* value_type, 1634 BasicType bt, 1635 DecoratorSet decorators) { 1636 set_control(ctl); 1637 C2AccessValuePtr addr(adr, adr_type); 1638 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1639 bt, obj, addr, alias_idx); 1640 if (access.is_raw()) { 1641 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1642 } else { 1643 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1644 } 1645 } 1646 1647 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl, 1648 Node* obj, 1649 Node* adr, 1650 const TypePtr* adr_type, 1651 int alias_idx, 1652 Node* expected_val, 1653 Node* new_val, 1654 const Type* value_type, 1655 BasicType bt, 1656 DecoratorSet decorators) { 1657 set_control(ctl); 1658 C2AccessValuePtr addr(adr, adr_type); 1659 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1660 bt, obj, addr, alias_idx); 1661 if (access.is_raw()) { 1662 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1663 } else { 1664 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1665 } 1666 } 1667 1668 Node* GraphKit::access_atomic_xchg_at(Node* ctl, 1669 Node* obj, 1670 Node* adr, 1671 const TypePtr* adr_type, 1672 int alias_idx, 1673 Node* new_val, 1674 const Type* value_type, 1675 BasicType bt, 1676 DecoratorSet decorators) { 1677 set_control(ctl); 1678 C2AccessValuePtr addr(adr, adr_type); 1679 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1680 bt, obj, addr, alias_idx); 1681 if (access.is_raw()) { 1682 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1683 } else { 1684 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1685 } 1686 } 1687 1688 Node* GraphKit::access_atomic_add_at(Node* ctl, 1689 Node* obj, 1690 Node* adr, 1691 const TypePtr* adr_type, 1692 int alias_idx, 1693 Node* new_val, 1694 const Type* value_type, 1695 BasicType bt, 1696 DecoratorSet decorators) { 1697 set_control(ctl); 1698 C2AccessValuePtr addr(adr, adr_type); 1699 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1700 if (access.is_raw()) { 1701 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1702 } else { 1703 return _barrier_set->atomic_add_at(access, new_val, value_type); 1704 } 1705 } 1706 1707 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) { 1708 set_control(ctl); 1709 return _barrier_set->clone(this, src, dst, size, is_array); 1710 } 1711 1712 //-------------------------array_element_address------------------------- 1713 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1714 const TypeInt* sizetype, Node* ctrl) { 1715 uint shift = exact_log2(type2aelembytes(elembt)); 1716 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1717 1718 // short-circuit a common case (saves lots of confusing waste motion) 1719 jint idx_con = find_int_con(idx, -1); 1720 if (idx_con >= 0) { 1721 intptr_t offset = header + ((intptr_t)idx_con << shift); 1722 return basic_plus_adr(ary, offset); 1723 } 1724 1725 // must be correct type for alignment purposes 1726 Node* base = basic_plus_adr(ary, header); 1727 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1728 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1729 return basic_plus_adr(ary, base, scale); 1730 } 1731 1732 //-------------------------load_array_element------------------------- 1733 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1734 const Type* elemtype = arytype->elem(); 1735 BasicType elembt = elemtype->array_element_basic_type(); 1736 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1737 if (elembt == T_NARROWOOP) { 1738 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1739 } 1740 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1741 return ld; 1742 } 1743 1744 //-------------------------set_arguments_for_java_call------------------------- 1745 // Arguments (pre-popped from the stack) are taken from the JVMS. 1746 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1747 // Add the call arguments: 1748 uint nargs = call->method()->arg_size(); 1749 for (uint i = 0; i < nargs; i++) { 1750 Node* arg = argument(i); 1751 call->init_req(i + TypeFunc::Parms, arg); 1752 } 1753 } 1754 1755 //---------------------------set_edges_for_java_call--------------------------- 1756 // Connect a newly created call into the current JVMS. 1757 // A return value node (if any) is returned from set_edges_for_java_call. 1758 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1759 1760 // Add the predefined inputs: 1761 call->init_req( TypeFunc::Control, control() ); 1762 call->init_req( TypeFunc::I_O , i_o() ); 1763 call->init_req( TypeFunc::Memory , reset_memory() ); 1764 call->init_req( TypeFunc::FramePtr, frameptr() ); 1765 call->init_req( TypeFunc::ReturnAdr, top() ); 1766 1767 add_safepoint_edges(call, must_throw); 1768 1769 Node* xcall = _gvn.transform(call); 1770 1771 if (xcall == top()) { 1772 set_control(top()); 1773 return; 1774 } 1775 assert(xcall == call, "call identity is stable"); 1776 1777 // Re-use the current map to produce the result. 1778 1779 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1780 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1781 set_all_memory_call(xcall, separate_io_proj); 1782 1783 //return xcall; // no need, caller already has it 1784 } 1785 1786 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1787 if (stopped()) return top(); // maybe the call folded up? 1788 1789 // Capture the return value, if any. 1790 Node* ret; 1791 if (call->method() == NULL || 1792 call->method()->return_type()->basic_type() == T_VOID) 1793 ret = top(); 1794 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1795 1796 // Note: Since any out-of-line call can produce an exception, 1797 // we always insert an I_O projection from the call into the result. 1798 1799 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1800 1801 if (separate_io_proj) { 1802 // The caller requested separate projections be used by the fall 1803 // through and exceptional paths, so replace the projections for 1804 // the fall through path. 1805 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1806 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1807 } 1808 return ret; 1809 } 1810 1811 //--------------------set_predefined_input_for_runtime_call-------------------- 1812 // Reading and setting the memory state is way conservative here. 1813 // The real problem is that I am not doing real Type analysis on memory, 1814 // so I cannot distinguish card mark stores from other stores. Across a GC 1815 // point the Store Barrier and the card mark memory has to agree. I cannot 1816 // have a card mark store and its barrier split across the GC point from 1817 // either above or below. Here I get that to happen by reading ALL of memory. 1818 // A better answer would be to separate out card marks from other memory. 1819 // For now, return the input memory state, so that it can be reused 1820 // after the call, if this call has restricted memory effects. 1821 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1822 // Set fixed predefined input arguments 1823 Node* memory = reset_memory(); 1824 Node* m = narrow_mem == NULL ? memory : narrow_mem; 1825 call->init_req( TypeFunc::Control, control() ); 1826 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1827 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1828 call->init_req( TypeFunc::FramePtr, frameptr() ); 1829 call->init_req( TypeFunc::ReturnAdr, top() ); 1830 return memory; 1831 } 1832 1833 //-------------------set_predefined_output_for_runtime_call-------------------- 1834 // Set control and memory (not i_o) from the call. 1835 // If keep_mem is not NULL, use it for the output state, 1836 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1837 // If hook_mem is NULL, this call produces no memory effects at all. 1838 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1839 // then only that memory slice is taken from the call. 1840 // In the last case, we must put an appropriate memory barrier before 1841 // the call, so as to create the correct anti-dependencies on loads 1842 // preceding the call. 1843 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1844 Node* keep_mem, 1845 const TypePtr* hook_mem) { 1846 // no i/o 1847 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1848 if (keep_mem) { 1849 // First clone the existing memory state 1850 set_all_memory(keep_mem); 1851 if (hook_mem != NULL) { 1852 // Make memory for the call 1853 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1854 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1855 // We also use hook_mem to extract specific effects from arraycopy stubs. 1856 set_memory(mem, hook_mem); 1857 } 1858 // ...else the call has NO memory effects. 1859 1860 // Make sure the call advertises its memory effects precisely. 1861 // This lets us build accurate anti-dependences in gcm.cpp. 1862 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1863 "call node must be constructed correctly"); 1864 } else { 1865 assert(hook_mem == NULL, ""); 1866 // This is not a "slow path" call; all memory comes from the call. 1867 set_all_memory_call(call); 1868 } 1869 } 1870 1871 1872 // Replace the call with the current state of the kit. 1873 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1874 JVMState* ejvms = NULL; 1875 if (has_exceptions()) { 1876 ejvms = transfer_exceptions_into_jvms(); 1877 } 1878 1879 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1880 ReplacedNodes replaced_nodes_exception; 1881 Node* ex_ctl = top(); 1882 1883 SafePointNode* final_state = stop(); 1884 1885 // Find all the needed outputs of this call 1886 CallProjections callprojs; 1887 call->extract_projections(&callprojs, true); 1888 1889 Node* init_mem = call->in(TypeFunc::Memory); 1890 Node* final_mem = final_state->in(TypeFunc::Memory); 1891 Node* final_ctl = final_state->in(TypeFunc::Control); 1892 Node* final_io = final_state->in(TypeFunc::I_O); 1893 1894 // Replace all the old call edges with the edges from the inlining result 1895 if (callprojs.fallthrough_catchproj != NULL) { 1896 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1897 } 1898 if (callprojs.fallthrough_memproj != NULL) { 1899 if (final_mem->is_MergeMem()) { 1900 // Parser's exits MergeMem was not transformed but may be optimized 1901 final_mem = _gvn.transform(final_mem); 1902 } 1903 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1904 } 1905 if (callprojs.fallthrough_ioproj != NULL) { 1906 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1907 } 1908 1909 // Replace the result with the new result if it exists and is used 1910 if (callprojs.resproj != NULL && result != NULL) { 1911 C->gvn_replace_by(callprojs.resproj, result); 1912 } 1913 1914 if (ejvms == NULL) { 1915 // No exception edges to simply kill off those paths 1916 if (callprojs.catchall_catchproj != NULL) { 1917 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1918 } 1919 if (callprojs.catchall_memproj != NULL) { 1920 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1921 } 1922 if (callprojs.catchall_ioproj != NULL) { 1923 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1924 } 1925 // Replace the old exception object with top 1926 if (callprojs.exobj != NULL) { 1927 C->gvn_replace_by(callprojs.exobj, C->top()); 1928 } 1929 } else { 1930 GraphKit ekit(ejvms); 1931 1932 // Load my combined exception state into the kit, with all phis transformed: 1933 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1934 replaced_nodes_exception = ex_map->replaced_nodes(); 1935 1936 Node* ex_oop = ekit.use_exception_state(ex_map); 1937 1938 if (callprojs.catchall_catchproj != NULL) { 1939 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1940 ex_ctl = ekit.control(); 1941 } 1942 if (callprojs.catchall_memproj != NULL) { 1943 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1944 } 1945 if (callprojs.catchall_ioproj != NULL) { 1946 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1947 } 1948 1949 // Replace the old exception object with the newly created one 1950 if (callprojs.exobj != NULL) { 1951 C->gvn_replace_by(callprojs.exobj, ex_oop); 1952 } 1953 } 1954 1955 // Disconnect the call from the graph 1956 call->disconnect_inputs(NULL, C); 1957 C->gvn_replace_by(call, C->top()); 1958 1959 // Clean up any MergeMems that feed other MergeMems since the 1960 // optimizer doesn't like that. 1961 if (final_mem->is_MergeMem()) { 1962 Node_List wl; 1963 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1964 Node* m = i.get(); 1965 if (m->is_MergeMem() && !wl.contains(m)) { 1966 wl.push(m); 1967 } 1968 } 1969 while (wl.size() > 0) { 1970 _gvn.transform(wl.pop()); 1971 } 1972 } 1973 1974 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1975 replaced_nodes.apply(C, final_ctl); 1976 } 1977 if (!ex_ctl->is_top() && do_replaced_nodes) { 1978 replaced_nodes_exception.apply(C, ex_ctl); 1979 } 1980 } 1981 1982 1983 //------------------------------increment_counter------------------------------ 1984 // for statistics: increment a VM counter by 1 1985 1986 void GraphKit::increment_counter(address counter_addr) { 1987 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1988 increment_counter(adr1); 1989 } 1990 1991 void GraphKit::increment_counter(Node* counter_addr) { 1992 int adr_type = Compile::AliasIdxRaw; 1993 Node* ctrl = control(); 1994 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1995 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1996 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1997 } 1998 1999 2000 //------------------------------uncommon_trap---------------------------------- 2001 // Bail out to the interpreter in mid-method. Implemented by calling the 2002 // uncommon_trap blob. This helper function inserts a runtime call with the 2003 // right debug info. 2004 void GraphKit::uncommon_trap(int trap_request, 2005 ciKlass* klass, const char* comment, 2006 bool must_throw, 2007 bool keep_exact_action) { 2008 if (failing()) stop(); 2009 if (stopped()) return; // trap reachable? 2010 2011 // Note: If ProfileTraps is true, and if a deopt. actually 2012 // occurs here, the runtime will make sure an MDO exists. There is 2013 // no need to call method()->ensure_method_data() at this point. 2014 2015 // Set the stack pointer to the right value for reexecution: 2016 set_sp(reexecute_sp()); 2017 2018 #ifdef ASSERT 2019 if (!must_throw) { 2020 // Make sure the stack has at least enough depth to execute 2021 // the current bytecode. 2022 int inputs, ignored_depth; 2023 if (compute_stack_effects(inputs, ignored_depth)) { 2024 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2025 Bytecodes::name(java_bc()), sp(), inputs); 2026 } 2027 } 2028 #endif 2029 2030 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2031 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2032 2033 switch (action) { 2034 case Deoptimization::Action_maybe_recompile: 2035 case Deoptimization::Action_reinterpret: 2036 // Temporary fix for 6529811 to allow virtual calls to be sure they 2037 // get the chance to go from mono->bi->mega 2038 if (!keep_exact_action && 2039 Deoptimization::trap_request_index(trap_request) < 0 && 2040 too_many_recompiles(reason)) { 2041 // This BCI is causing too many recompilations. 2042 if (C->log() != NULL) { 2043 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2044 Deoptimization::trap_reason_name(reason), 2045 Deoptimization::trap_action_name(action)); 2046 } 2047 action = Deoptimization::Action_none; 2048 trap_request = Deoptimization::make_trap_request(reason, action); 2049 } else { 2050 C->set_trap_can_recompile(true); 2051 } 2052 break; 2053 case Deoptimization::Action_make_not_entrant: 2054 C->set_trap_can_recompile(true); 2055 break; 2056 case Deoptimization::Action_none: 2057 case Deoptimization::Action_make_not_compilable: 2058 break; 2059 default: 2060 #ifdef ASSERT 2061 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2062 #endif 2063 break; 2064 } 2065 2066 if (TraceOptoParse) { 2067 char buf[100]; 2068 tty->print_cr("Uncommon trap %s at bci:%d", 2069 Deoptimization::format_trap_request(buf, sizeof(buf), 2070 trap_request), bci()); 2071 } 2072 2073 CompileLog* log = C->log(); 2074 if (log != NULL) { 2075 int kid = (klass == NULL)? -1: log->identify(klass); 2076 log->begin_elem("uncommon_trap bci='%d'", bci()); 2077 char buf[100]; 2078 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2079 trap_request)); 2080 if (kid >= 0) log->print(" klass='%d'", kid); 2081 if (comment != NULL) log->print(" comment='%s'", comment); 2082 log->end_elem(); 2083 } 2084 2085 // Make sure any guarding test views this path as very unlikely 2086 Node *i0 = control()->in(0); 2087 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2088 IfNode *iff = i0->as_If(); 2089 float f = iff->_prob; // Get prob 2090 if (control()->Opcode() == Op_IfTrue) { 2091 if (f > PROB_UNLIKELY_MAG(4)) 2092 iff->_prob = PROB_MIN; 2093 } else { 2094 if (f < PROB_LIKELY_MAG(4)) 2095 iff->_prob = PROB_MAX; 2096 } 2097 } 2098 2099 // Clear out dead values from the debug info. 2100 kill_dead_locals(); 2101 2102 // Now insert the uncommon trap subroutine call 2103 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2104 const TypePtr* no_memory_effects = NULL; 2105 // Pass the index of the class to be loaded 2106 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2107 (must_throw ? RC_MUST_THROW : 0), 2108 OptoRuntime::uncommon_trap_Type(), 2109 call_addr, "uncommon_trap", no_memory_effects, 2110 intcon(trap_request)); 2111 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2112 "must extract request correctly from the graph"); 2113 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2114 2115 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2116 // The debug info is the only real input to this call. 2117 2118 // Halt-and-catch fire here. The above call should never return! 2119 HaltNode* halt = new HaltNode(control(), frameptr()); 2120 _gvn.set_type_bottom(halt); 2121 root()->add_req(halt); 2122 2123 stop_and_kill_map(); 2124 } 2125 2126 2127 //--------------------------just_allocated_object------------------------------ 2128 // Report the object that was just allocated. 2129 // It must be the case that there are no intervening safepoints. 2130 // We use this to determine if an object is so "fresh" that 2131 // it does not require card marks. 2132 Node* GraphKit::just_allocated_object(Node* current_control) { 2133 if (C->recent_alloc_ctl() == current_control) 2134 return C->recent_alloc_obj(); 2135 return NULL; 2136 } 2137 2138 2139 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2140 // (Note: TypeFunc::make has a cache that makes this fast.) 2141 const TypeFunc* tf = TypeFunc::make(dest_method); 2142 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2143 for (int j = 0; j < nargs; j++) { 2144 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2145 if( targ->basic_type() == T_DOUBLE ) { 2146 // If any parameters are doubles, they must be rounded before 2147 // the call, dstore_rounding does gvn.transform 2148 Node *arg = argument(j); 2149 arg = dstore_rounding(arg); 2150 set_argument(j, arg); 2151 } 2152 } 2153 } 2154 2155 /** 2156 * Record profiling data exact_kls for Node n with the type system so 2157 * that it can propagate it (speculation) 2158 * 2159 * @param n node that the type applies to 2160 * @param exact_kls type from profiling 2161 * @param maybe_null did profiling see null? 2162 * 2163 * @return node with improved type 2164 */ 2165 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2166 const Type* current_type = _gvn.type(n); 2167 assert(UseTypeSpeculation, "type speculation must be on"); 2168 2169 const TypePtr* speculative = current_type->speculative(); 2170 2171 // Should the klass from the profile be recorded in the speculative type? 2172 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2173 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2174 const TypeOopPtr* xtype = tklass->as_instance_type(); 2175 assert(xtype->klass_is_exact(), "Should be exact"); 2176 // Any reason to believe n is not null (from this profiling or a previous one)? 2177 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2178 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2179 // record the new speculative type's depth 2180 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2181 speculative = speculative->with_inline_depth(jvms()->depth()); 2182 } else if (current_type->would_improve_ptr(ptr_kind)) { 2183 // Profiling report that null was never seen so we can change the 2184 // speculative type to non null ptr. 2185 if (ptr_kind == ProfileAlwaysNull) { 2186 speculative = TypePtr::NULL_PTR; 2187 } else { 2188 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2189 const TypePtr* ptr = TypePtr::NOTNULL; 2190 if (speculative != NULL) { 2191 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2192 } else { 2193 speculative = ptr; 2194 } 2195 } 2196 } 2197 2198 if (speculative != current_type->speculative()) { 2199 // Build a type with a speculative type (what we think we know 2200 // about the type but will need a guard when we use it) 2201 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2202 // We're changing the type, we need a new CheckCast node to carry 2203 // the new type. The new type depends on the control: what 2204 // profiling tells us is only valid from here as far as we can 2205 // tell. 2206 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2207 cast = _gvn.transform(cast); 2208 replace_in_map(n, cast); 2209 n = cast; 2210 } 2211 2212 return n; 2213 } 2214 2215 /** 2216 * Record profiling data from receiver profiling at an invoke with the 2217 * type system so that it can propagate it (speculation) 2218 * 2219 * @param n receiver node 2220 * 2221 * @return node with improved type 2222 */ 2223 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2224 if (!UseTypeSpeculation) { 2225 return n; 2226 } 2227 ciKlass* exact_kls = profile_has_unique_klass(); 2228 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2229 if ((java_bc() == Bytecodes::_checkcast || 2230 java_bc() == Bytecodes::_instanceof || 2231 java_bc() == Bytecodes::_aastore) && 2232 method()->method_data()->is_mature()) { 2233 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2234 if (data != NULL) { 2235 if (!data->as_BitData()->null_seen()) { 2236 ptr_kind = ProfileNeverNull; 2237 } else { 2238 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2239 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2240 uint i = 0; 2241 for (; i < call->row_limit(); i++) { 2242 ciKlass* receiver = call->receiver(i); 2243 if (receiver != NULL) { 2244 break; 2245 } 2246 } 2247 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2248 } 2249 } 2250 } 2251 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2252 } 2253 2254 /** 2255 * Record profiling data from argument profiling at an invoke with the 2256 * type system so that it can propagate it (speculation) 2257 * 2258 * @param dest_method target method for the call 2259 * @param bc what invoke bytecode is this? 2260 */ 2261 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2262 if (!UseTypeSpeculation) { 2263 return; 2264 } 2265 const TypeFunc* tf = TypeFunc::make(dest_method); 2266 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2267 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2268 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2269 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2270 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2271 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2272 ciKlass* better_type = NULL; 2273 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2274 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2275 } 2276 i++; 2277 } 2278 } 2279 } 2280 2281 /** 2282 * Record profiling data from parameter profiling at an invoke with 2283 * the type system so that it can propagate it (speculation) 2284 */ 2285 void GraphKit::record_profiled_parameters_for_speculation() { 2286 if (!UseTypeSpeculation) { 2287 return; 2288 } 2289 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2290 if (_gvn.type(local(i))->isa_oopptr()) { 2291 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2292 ciKlass* better_type = NULL; 2293 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2294 record_profile_for_speculation(local(i), better_type, ptr_kind); 2295 } 2296 j++; 2297 } 2298 } 2299 } 2300 2301 /** 2302 * Record profiling data from return value profiling at an invoke with 2303 * the type system so that it can propagate it (speculation) 2304 */ 2305 void GraphKit::record_profiled_return_for_speculation() { 2306 if (!UseTypeSpeculation) { 2307 return; 2308 } 2309 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2310 ciKlass* better_type = NULL; 2311 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2312 // If profiling reports a single type for the return value, 2313 // feed it to the type system so it can propagate it as a 2314 // speculative type 2315 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2316 } 2317 } 2318 2319 void GraphKit::round_double_result(ciMethod* dest_method) { 2320 // A non-strict method may return a double value which has an extended 2321 // exponent, but this must not be visible in a caller which is 'strict' 2322 // If a strict caller invokes a non-strict callee, round a double result 2323 2324 BasicType result_type = dest_method->return_type()->basic_type(); 2325 assert( method() != NULL, "must have caller context"); 2326 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2327 // Destination method's return value is on top of stack 2328 // dstore_rounding() does gvn.transform 2329 Node *result = pop_pair(); 2330 result = dstore_rounding(result); 2331 push_pair(result); 2332 } 2333 } 2334 2335 // rounding for strict float precision conformance 2336 Node* GraphKit::precision_rounding(Node* n) { 2337 return UseStrictFP && _method->flags().is_strict() 2338 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2339 ? _gvn.transform( new RoundFloatNode(0, n) ) 2340 : n; 2341 } 2342 2343 // rounding for strict double precision conformance 2344 Node* GraphKit::dprecision_rounding(Node *n) { 2345 return UseStrictFP && _method->flags().is_strict() 2346 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2347 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2348 : n; 2349 } 2350 2351 // rounding for non-strict double stores 2352 Node* GraphKit::dstore_rounding(Node* n) { 2353 return Matcher::strict_fp_requires_explicit_rounding 2354 && UseSSE <= 1 2355 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2356 : n; 2357 } 2358 2359 //============================================================================= 2360 // Generate a fast path/slow path idiom. Graph looks like: 2361 // [foo] indicates that 'foo' is a parameter 2362 // 2363 // [in] NULL 2364 // \ / 2365 // CmpP 2366 // Bool ne 2367 // If 2368 // / \ 2369 // True False-<2> 2370 // / | 2371 // / cast_not_null 2372 // Load | | ^ 2373 // [fast_test] | | 2374 // gvn to opt_test | | 2375 // / \ | <1> 2376 // True False | 2377 // | \\ | 2378 // [slow_call] \[fast_result] 2379 // Ctl Val \ \ 2380 // | \ \ 2381 // Catch <1> \ \ 2382 // / \ ^ \ \ 2383 // Ex No_Ex | \ \ 2384 // | \ \ | \ <2> \ 2385 // ... \ [slow_res] | | \ [null_result] 2386 // \ \--+--+--- | | 2387 // \ | / \ | / 2388 // --------Region Phi 2389 // 2390 //============================================================================= 2391 // Code is structured as a series of driver functions all called 'do_XXX' that 2392 // call a set of helper functions. Helper functions first, then drivers. 2393 2394 //------------------------------null_check_oop--------------------------------- 2395 // Null check oop. Set null-path control into Region in slot 3. 2396 // Make a cast-not-nullness use the other not-null control. Return cast. 2397 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2398 bool never_see_null, 2399 bool safe_for_replace, 2400 bool speculative) { 2401 // Initial NULL check taken path 2402 (*null_control) = top(); 2403 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2404 2405 // Generate uncommon_trap: 2406 if (never_see_null && (*null_control) != top()) { 2407 // If we see an unexpected null at a check-cast we record it and force a 2408 // recompile; the offending check-cast will be compiled to handle NULLs. 2409 // If we see more than one offending BCI, then all checkcasts in the 2410 // method will be compiled to handle NULLs. 2411 PreserveJVMState pjvms(this); 2412 set_control(*null_control); 2413 replace_in_map(value, null()); 2414 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2415 uncommon_trap(reason, 2416 Deoptimization::Action_make_not_entrant); 2417 (*null_control) = top(); // NULL path is dead 2418 } 2419 if ((*null_control) == top() && safe_for_replace) { 2420 replace_in_map(value, cast); 2421 } 2422 2423 // Cast away null-ness on the result 2424 return cast; 2425 } 2426 2427 //------------------------------opt_iff---------------------------------------- 2428 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2429 // Return slow-path control. 2430 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2431 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2432 2433 // Fast path taken; set region slot 2 2434 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2435 region->init_req(2,fast_taken); // Capture fast-control 2436 2437 // Fast path not-taken, i.e. slow path 2438 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2439 return slow_taken; 2440 } 2441 2442 //-----------------------------make_runtime_call------------------------------- 2443 Node* GraphKit::make_runtime_call(int flags, 2444 const TypeFunc* call_type, address call_addr, 2445 const char* call_name, 2446 const TypePtr* adr_type, 2447 // The following parms are all optional. 2448 // The first NULL ends the list. 2449 Node* parm0, Node* parm1, 2450 Node* parm2, Node* parm3, 2451 Node* parm4, Node* parm5, 2452 Node* parm6, Node* parm7) { 2453 assert(call_addr != NULL, "must not call NULL targets"); 2454 2455 // Slow-path call 2456 bool is_leaf = !(flags & RC_NO_LEAF); 2457 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2458 if (call_name == NULL) { 2459 assert(!is_leaf, "must supply name for leaf"); 2460 call_name = OptoRuntime::stub_name(call_addr); 2461 } 2462 CallNode* call; 2463 if (!is_leaf) { 2464 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2465 bci(), adr_type); 2466 } else if (flags & RC_NO_FP) { 2467 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2468 } else { 2469 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2470 } 2471 2472 // The following is similar to set_edges_for_java_call, 2473 // except that the memory effects of the call are restricted to AliasIdxRaw. 2474 2475 // Slow path call has no side-effects, uses few values 2476 bool wide_in = !(flags & RC_NARROW_MEM); 2477 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2478 2479 Node* prev_mem = NULL; 2480 if (wide_in) { 2481 prev_mem = set_predefined_input_for_runtime_call(call); 2482 } else { 2483 assert(!wide_out, "narrow in => narrow out"); 2484 Node* narrow_mem = memory(adr_type); 2485 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2486 } 2487 2488 // Hook each parm in order. Stop looking at the first NULL. 2489 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2490 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2491 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2492 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2493 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2494 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2495 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2496 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2497 /* close each nested if ===> */ } } } } } } } } 2498 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2499 2500 if (!is_leaf) { 2501 // Non-leaves can block and take safepoints: 2502 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2503 } 2504 // Non-leaves can throw exceptions: 2505 if (has_io) { 2506 call->set_req(TypeFunc::I_O, i_o()); 2507 } 2508 2509 if (flags & RC_UNCOMMON) { 2510 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2511 // (An "if" probability corresponds roughly to an unconditional count. 2512 // Sort of.) 2513 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2514 } 2515 2516 Node* c = _gvn.transform(call); 2517 assert(c == call, "cannot disappear"); 2518 2519 if (wide_out) { 2520 // Slow path call has full side-effects. 2521 set_predefined_output_for_runtime_call(call); 2522 } else { 2523 // Slow path call has few side-effects, and/or sets few values. 2524 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2525 } 2526 2527 if (has_io) { 2528 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2529 } 2530 return call; 2531 2532 } 2533 2534 //------------------------------merge_memory----------------------------------- 2535 // Merge memory from one path into the current memory state. 2536 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2537 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2538 Node* old_slice = mms.force_memory(); 2539 Node* new_slice = mms.memory2(); 2540 if (old_slice != new_slice) { 2541 PhiNode* phi; 2542 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2543 if (mms.is_empty()) { 2544 // clone base memory Phi's inputs for this memory slice 2545 assert(old_slice == mms.base_memory(), "sanity"); 2546 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2547 _gvn.set_type(phi, Type::MEMORY); 2548 for (uint i = 1; i < phi->req(); i++) { 2549 phi->init_req(i, old_slice->in(i)); 2550 } 2551 } else { 2552 phi = old_slice->as_Phi(); // Phi was generated already 2553 } 2554 } else { 2555 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2556 _gvn.set_type(phi, Type::MEMORY); 2557 } 2558 phi->set_req(new_path, new_slice); 2559 mms.set_memory(phi); 2560 } 2561 } 2562 } 2563 2564 //------------------------------make_slow_call_ex------------------------------ 2565 // Make the exception handler hookups for the slow call 2566 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2567 if (stopped()) return; 2568 2569 // Make a catch node with just two handlers: fall-through and catch-all 2570 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2571 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2572 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2573 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2574 2575 { PreserveJVMState pjvms(this); 2576 set_control(excp); 2577 set_i_o(i_o); 2578 2579 if (excp != top()) { 2580 if (deoptimize) { 2581 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2582 uncommon_trap(Deoptimization::Reason_unhandled, 2583 Deoptimization::Action_none); 2584 } else { 2585 // Create an exception state also. 2586 // Use an exact type if the caller has a specific exception. 2587 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2588 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2589 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2590 } 2591 } 2592 } 2593 2594 // Get the no-exception control from the CatchNode. 2595 set_control(norm); 2596 } 2597 2598 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2599 Node* cmp = NULL; 2600 switch(bt) { 2601 case T_INT: cmp = new CmpINode(in1, in2); break; 2602 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2603 default: fatal("unexpected comparison type %s", type2name(bt)); 2604 } 2605 gvn->transform(cmp); 2606 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2607 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2608 gvn->transform(iff); 2609 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2610 return iff; 2611 } 2612 2613 2614 //-------------------------------gen_subtype_check----------------------------- 2615 // Generate a subtyping check. Takes as input the subtype and supertype. 2616 // Returns 2 values: sets the default control() to the true path and returns 2617 // the false path. Only reads invariant memory; sets no (visible) memory. 2618 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2619 // but that's not exposed to the optimizer. This call also doesn't take in an 2620 // Object; if you wish to check an Object you need to load the Object's class 2621 // prior to coming here. 2622 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2623 Compile* C = gvn->C; 2624 2625 if ((*ctrl)->is_top()) { 2626 return C->top(); 2627 } 2628 2629 // Fast check for identical types, perhaps identical constants. 2630 // The types can even be identical non-constants, in cases 2631 // involving Array.newInstance, Object.clone, etc. 2632 if (subklass == superklass) 2633 return C->top(); // false path is dead; no test needed. 2634 2635 if (gvn->type(superklass)->singleton()) { 2636 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2637 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2638 2639 // In the common case of an exact superklass, try to fold up the 2640 // test before generating code. You may ask, why not just generate 2641 // the code and then let it fold up? The answer is that the generated 2642 // code will necessarily include null checks, which do not always 2643 // completely fold away. If they are also needless, then they turn 2644 // into a performance loss. Example: 2645 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2646 // Here, the type of 'fa' is often exact, so the store check 2647 // of fa[1]=x will fold up, without testing the nullness of x. 2648 switch (C->static_subtype_check(superk, subk)) { 2649 case Compile::SSC_always_false: 2650 { 2651 Node* always_fail = *ctrl; 2652 *ctrl = gvn->C->top(); 2653 return always_fail; 2654 } 2655 case Compile::SSC_always_true: 2656 return C->top(); 2657 case Compile::SSC_easy_test: 2658 { 2659 // Just do a direct pointer compare and be done. 2660 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2661 *ctrl = gvn->transform(new IfTrueNode(iff)); 2662 return gvn->transform(new IfFalseNode(iff)); 2663 } 2664 case Compile::SSC_full_test: 2665 break; 2666 default: 2667 ShouldNotReachHere(); 2668 } 2669 } 2670 2671 // %%% Possible further optimization: Even if the superklass is not exact, 2672 // if the subklass is the unique subtype of the superklass, the check 2673 // will always succeed. We could leave a dependency behind to ensure this. 2674 2675 // First load the super-klass's check-offset 2676 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2677 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2678 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2679 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2680 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2681 2682 // Load from the sub-klass's super-class display list, or a 1-word cache of 2683 // the secondary superclass list, or a failing value with a sentinel offset 2684 // if the super-klass is an interface or exceptionally deep in the Java 2685 // hierarchy and we have to scan the secondary superclass list the hard way. 2686 // Worst-case type is a little odd: NULL is allowed as a result (usually 2687 // klass loads can never produce a NULL). 2688 Node *chk_off_X = chk_off; 2689 #ifdef _LP64 2690 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2691 #endif 2692 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2693 // For some types like interfaces the following loadKlass is from a 1-word 2694 // cache which is mutable so can't use immutable memory. Other 2695 // types load from the super-class display table which is immutable. 2696 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2697 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2698 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2699 2700 // Compile speed common case: ARE a subtype and we canNOT fail 2701 if( superklass == nkls ) 2702 return C->top(); // false path is dead; no test needed. 2703 2704 // See if we get an immediate positive hit. Happens roughly 83% of the 2705 // time. Test to see if the value loaded just previously from the subklass 2706 // is exactly the superklass. 2707 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2708 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2709 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2710 2711 // Compile speed common case: Check for being deterministic right now. If 2712 // chk_off is a constant and not equal to cacheoff then we are NOT a 2713 // subklass. In this case we need exactly the 1 test above and we can 2714 // return those results immediately. 2715 if (!might_be_cache) { 2716 Node* not_subtype_ctrl = *ctrl; 2717 *ctrl = iftrue1; // We need exactly the 1 test above 2718 return not_subtype_ctrl; 2719 } 2720 2721 // Gather the various success & failures here 2722 RegionNode *r_ok_subtype = new RegionNode(4); 2723 gvn->record_for_igvn(r_ok_subtype); 2724 RegionNode *r_not_subtype = new RegionNode(3); 2725 gvn->record_for_igvn(r_not_subtype); 2726 2727 r_ok_subtype->init_req(1, iftrue1); 2728 2729 // Check for immediate negative hit. Happens roughly 11% of the time (which 2730 // is roughly 63% of the remaining cases). Test to see if the loaded 2731 // check-offset points into the subklass display list or the 1-element 2732 // cache. If it points to the display (and NOT the cache) and the display 2733 // missed then it's not a subtype. 2734 Node *cacheoff = gvn->intcon(cacheoff_con); 2735 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2736 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2737 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2738 2739 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2740 // No performance impact (too rare) but allows sharing of secondary arrays 2741 // which has some footprint reduction. 2742 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2743 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2744 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2745 2746 // -- Roads not taken here: -- 2747 // We could also have chosen to perform the self-check at the beginning 2748 // of this code sequence, as the assembler does. This would not pay off 2749 // the same way, since the optimizer, unlike the assembler, can perform 2750 // static type analysis to fold away many successful self-checks. 2751 // Non-foldable self checks work better here in second position, because 2752 // the initial primary superclass check subsumes a self-check for most 2753 // types. An exception would be a secondary type like array-of-interface, 2754 // which does not appear in its own primary supertype display. 2755 // Finally, we could have chosen to move the self-check into the 2756 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2757 // dependent manner. But it is worthwhile to have the check here, 2758 // where it can be perhaps be optimized. The cost in code space is 2759 // small (register compare, branch). 2760 2761 // Now do a linear scan of the secondary super-klass array. Again, no real 2762 // performance impact (too rare) but it's gotta be done. 2763 // Since the code is rarely used, there is no penalty for moving it 2764 // out of line, and it can only improve I-cache density. 2765 // The decision to inline or out-of-line this final check is platform 2766 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2767 Node* psc = gvn->transform( 2768 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2769 2770 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2771 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2772 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2773 2774 // Return false path; set default control to true path. 2775 *ctrl = gvn->transform(r_ok_subtype); 2776 return gvn->transform(r_not_subtype); 2777 } 2778 2779 // Profile-driven exact type check: 2780 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2781 float prob, 2782 Node* *casted_receiver) { 2783 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2784 Node* recv_klass = load_object_klass(receiver); 2785 Node* want_klass = makecon(tklass); 2786 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2787 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2788 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2789 set_control( _gvn.transform( new IfTrueNode (iff) )); 2790 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2791 2792 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2793 assert(recv_xtype->klass_is_exact(), ""); 2794 2795 // Subsume downstream occurrences of receiver with a cast to 2796 // recv_xtype, since now we know what the type will be. 2797 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2798 (*casted_receiver) = _gvn.transform(cast); 2799 // (User must make the replace_in_map call.) 2800 2801 return fail; 2802 } 2803 2804 2805 //------------------------------seems_never_null------------------------------- 2806 // Use null_seen information if it is available from the profile. 2807 // If we see an unexpected null at a type check we record it and force a 2808 // recompile; the offending check will be recompiled to handle NULLs. 2809 // If we see several offending BCIs, then all checks in the 2810 // method will be recompiled. 2811 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2812 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2813 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2814 if (UncommonNullCast // Cutout for this technique 2815 && obj != null() // And not the -Xcomp stupid case? 2816 && !too_many_traps(reason) 2817 ) { 2818 if (speculating) { 2819 return true; 2820 } 2821 if (data == NULL) 2822 // Edge case: no mature data. Be optimistic here. 2823 return true; 2824 // If the profile has not seen a null, assume it won't happen. 2825 assert(java_bc() == Bytecodes::_checkcast || 2826 java_bc() == Bytecodes::_instanceof || 2827 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2828 return !data->as_BitData()->null_seen(); 2829 } 2830 speculating = false; 2831 return false; 2832 } 2833 2834 //------------------------maybe_cast_profiled_receiver------------------------- 2835 // If the profile has seen exactly one type, narrow to exactly that type. 2836 // Subsequent type checks will always fold up. 2837 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2838 ciKlass* require_klass, 2839 ciKlass* spec_klass, 2840 bool safe_for_replace) { 2841 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2842 2843 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2844 2845 // Make sure we haven't already deoptimized from this tactic. 2846 if (too_many_traps_or_recompiles(reason)) 2847 return NULL; 2848 2849 // (No, this isn't a call, but it's enough like a virtual call 2850 // to use the same ciMethod accessor to get the profile info...) 2851 // If we have a speculative type use it instead of profiling (which 2852 // may not help us) 2853 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2854 if (exact_kls != NULL) {// no cast failures here 2855 if (require_klass == NULL || 2856 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2857 // If we narrow the type to match what the type profile sees or 2858 // the speculative type, we can then remove the rest of the 2859 // cast. 2860 // This is a win, even if the exact_kls is very specific, 2861 // because downstream operations, such as method calls, 2862 // will often benefit from the sharper type. 2863 Node* exact_obj = not_null_obj; // will get updated in place... 2864 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2865 &exact_obj); 2866 { PreserveJVMState pjvms(this); 2867 set_control(slow_ctl); 2868 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2869 } 2870 if (safe_for_replace) { 2871 replace_in_map(not_null_obj, exact_obj); 2872 } 2873 return exact_obj; 2874 } 2875 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2876 } 2877 2878 return NULL; 2879 } 2880 2881 /** 2882 * Cast obj to type and emit guard unless we had too many traps here 2883 * already 2884 * 2885 * @param obj node being casted 2886 * @param type type to cast the node to 2887 * @param not_null true if we know node cannot be null 2888 */ 2889 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2890 ciKlass* type, 2891 bool not_null) { 2892 if (stopped()) { 2893 return obj; 2894 } 2895 2896 // type == NULL if profiling tells us this object is always null 2897 if (type != NULL) { 2898 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2899 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2900 2901 if (!too_many_traps_or_recompiles(null_reason) && 2902 !too_many_traps_or_recompiles(class_reason)) { 2903 Node* not_null_obj = NULL; 2904 // not_null is true if we know the object is not null and 2905 // there's no need for a null check 2906 if (!not_null) { 2907 Node* null_ctl = top(); 2908 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2909 assert(null_ctl->is_top(), "no null control here"); 2910 } else { 2911 not_null_obj = obj; 2912 } 2913 2914 Node* exact_obj = not_null_obj; 2915 ciKlass* exact_kls = type; 2916 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2917 &exact_obj); 2918 { 2919 PreserveJVMState pjvms(this); 2920 set_control(slow_ctl); 2921 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2922 } 2923 replace_in_map(not_null_obj, exact_obj); 2924 obj = exact_obj; 2925 } 2926 } else { 2927 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 2928 Node* exact_obj = null_assert(obj); 2929 replace_in_map(obj, exact_obj); 2930 obj = exact_obj; 2931 } 2932 } 2933 return obj; 2934 } 2935 2936 //-------------------------------gen_instanceof-------------------------------- 2937 // Generate an instance-of idiom. Used by both the instance-of bytecode 2938 // and the reflective instance-of call. 2939 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2940 kill_dead_locals(); // Benefit all the uncommon traps 2941 assert( !stopped(), "dead parse path should be checked in callers" ); 2942 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2943 "must check for not-null not-dead klass in callers"); 2944 2945 // Make the merge point 2946 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2947 RegionNode* region = new RegionNode(PATH_LIMIT); 2948 Node* phi = new PhiNode(region, TypeInt::BOOL); 2949 C->set_has_split_ifs(true); // Has chance for split-if optimization 2950 2951 ciProfileData* data = NULL; 2952 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2953 data = method()->method_data()->bci_to_data(bci()); 2954 } 2955 bool speculative_not_null = false; 2956 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2957 && seems_never_null(obj, data, speculative_not_null)); 2958 2959 // Null check; get casted pointer; set region slot 3 2960 Node* null_ctl = top(); 2961 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2962 2963 // If not_null_obj is dead, only null-path is taken 2964 if (stopped()) { // Doing instance-of on a NULL? 2965 set_control(null_ctl); 2966 return intcon(0); 2967 } 2968 region->init_req(_null_path, null_ctl); 2969 phi ->init_req(_null_path, intcon(0)); // Set null path value 2970 if (null_ctl == top()) { 2971 // Do this eagerly, so that pattern matches like is_diamond_phi 2972 // will work even during parsing. 2973 assert(_null_path == PATH_LIMIT-1, "delete last"); 2974 region->del_req(_null_path); 2975 phi ->del_req(_null_path); 2976 } 2977 2978 // Do we know the type check always succeed? 2979 bool known_statically = false; 2980 if (_gvn.type(superklass)->singleton()) { 2981 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2982 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2983 if (subk != NULL && subk->is_loaded()) { 2984 int static_res = C->static_subtype_check(superk, subk); 2985 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2986 } 2987 } 2988 2989 if (!known_statically) { 2990 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2991 // We may not have profiling here or it may not help us. If we 2992 // have a speculative type use it to perform an exact cast. 2993 ciKlass* spec_obj_type = obj_type->speculative_type(); 2994 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2995 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2996 if (stopped()) { // Profile disagrees with this path. 2997 set_control(null_ctl); // Null is the only remaining possibility. 2998 return intcon(0); 2999 } 3000 if (cast_obj != NULL) { 3001 not_null_obj = cast_obj; 3002 } 3003 } 3004 } 3005 3006 // Load the object's klass 3007 Node* obj_klass = load_object_klass(not_null_obj); 3008 3009 // Generate the subtype check 3010 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 3011 3012 // Plug in the success path to the general merge in slot 1. 3013 region->init_req(_obj_path, control()); 3014 phi ->init_req(_obj_path, intcon(1)); 3015 3016 // Plug in the failing path to the general merge in slot 2. 3017 region->init_req(_fail_path, not_subtype_ctrl); 3018 phi ->init_req(_fail_path, intcon(0)); 3019 3020 // Return final merged results 3021 set_control( _gvn.transform(region) ); 3022 record_for_igvn(region); 3023 3024 // If we know the type check always succeeds then we don't use the 3025 // profiling data at this bytecode. Don't lose it, feed it to the 3026 // type system as a speculative type. 3027 if (safe_for_replace) { 3028 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3029 replace_in_map(obj, casted_obj); 3030 } 3031 3032 return _gvn.transform(phi); 3033 } 3034 3035 //-------------------------------gen_checkcast--------------------------------- 3036 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3037 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3038 // uncommon-trap paths work. Adjust stack after this call. 3039 // If failure_control is supplied and not null, it is filled in with 3040 // the control edge for the cast failure. Otherwise, an appropriate 3041 // uncommon trap or exception is thrown. 3042 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3043 Node* *failure_control) { 3044 kill_dead_locals(); // Benefit all the uncommon traps 3045 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3046 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3047 3048 // Fast cutout: Check the case that the cast is vacuously true. 3049 // This detects the common cases where the test will short-circuit 3050 // away completely. We do this before we perform the null check, 3051 // because if the test is going to turn into zero code, we don't 3052 // want a residual null check left around. (Causes a slowdown, 3053 // for example, in some objArray manipulations, such as a[i]=a[j].) 3054 if (tk->singleton()) { 3055 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3056 if (objtp != NULL && objtp->klass() != NULL) { 3057 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3058 case Compile::SSC_always_true: 3059 // If we know the type check always succeed then we don't use 3060 // the profiling data at this bytecode. Don't lose it, feed it 3061 // to the type system as a speculative type. 3062 return record_profiled_receiver_for_speculation(obj); 3063 case Compile::SSC_always_false: 3064 // It needs a null check because a null will *pass* the cast check. 3065 // A non-null value will always produce an exception. 3066 return null_assert(obj); 3067 } 3068 } 3069 } 3070 3071 ciProfileData* data = NULL; 3072 bool safe_for_replace = false; 3073 if (failure_control == NULL) { // use MDO in regular case only 3074 assert(java_bc() == Bytecodes::_aastore || 3075 java_bc() == Bytecodes::_checkcast, 3076 "interpreter profiles type checks only for these BCs"); 3077 data = method()->method_data()->bci_to_data(bci()); 3078 safe_for_replace = true; 3079 } 3080 3081 // Make the merge point 3082 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3083 RegionNode* region = new RegionNode(PATH_LIMIT); 3084 Node* phi = new PhiNode(region, toop); 3085 C->set_has_split_ifs(true); // Has chance for split-if optimization 3086 3087 // Use null-cast information if it is available 3088 bool speculative_not_null = false; 3089 bool never_see_null = ((failure_control == NULL) // regular case only 3090 && seems_never_null(obj, data, speculative_not_null)); 3091 3092 // Null check; get casted pointer; set region slot 3 3093 Node* null_ctl = top(); 3094 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3095 3096 // If not_null_obj is dead, only null-path is taken 3097 if (stopped()) { // Doing instance-of on a NULL? 3098 set_control(null_ctl); 3099 return null(); 3100 } 3101 region->init_req(_null_path, null_ctl); 3102 phi ->init_req(_null_path, null()); // Set null path value 3103 if (null_ctl == top()) { 3104 // Do this eagerly, so that pattern matches like is_diamond_phi 3105 // will work even during parsing. 3106 assert(_null_path == PATH_LIMIT-1, "delete last"); 3107 region->del_req(_null_path); 3108 phi ->del_req(_null_path); 3109 } 3110 3111 Node* cast_obj = NULL; 3112 if (tk->klass_is_exact()) { 3113 // The following optimization tries to statically cast the speculative type of the object 3114 // (for example obtained during profiling) to the type of the superklass and then do a 3115 // dynamic check that the type of the object is what we expect. To work correctly 3116 // for checkcast and aastore the type of superklass should be exact. 3117 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3118 // We may not have profiling here or it may not help us. If we have 3119 // a speculative type use it to perform an exact cast. 3120 ciKlass* spec_obj_type = obj_type->speculative_type(); 3121 if (spec_obj_type != NULL || data != NULL) { 3122 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3123 if (cast_obj != NULL) { 3124 if (failure_control != NULL) // failure is now impossible 3125 (*failure_control) = top(); 3126 // adjust the type of the phi to the exact klass: 3127 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3128 } 3129 } 3130 } 3131 3132 if (cast_obj == NULL) { 3133 // Load the object's klass 3134 Node* obj_klass = load_object_klass(not_null_obj); 3135 3136 // Generate the subtype check 3137 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3138 3139 // Plug in success path into the merge 3140 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3141 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3142 if (failure_control == NULL) { 3143 if (not_subtype_ctrl != top()) { // If failure is possible 3144 PreserveJVMState pjvms(this); 3145 set_control(not_subtype_ctrl); 3146 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3147 } 3148 } else { 3149 (*failure_control) = not_subtype_ctrl; 3150 } 3151 } 3152 3153 region->init_req(_obj_path, control()); 3154 phi ->init_req(_obj_path, cast_obj); 3155 3156 // A merge of NULL or Casted-NotNull obj 3157 Node* res = _gvn.transform(phi); 3158 3159 // Note I do NOT always 'replace_in_map(obj,result)' here. 3160 // if( tk->klass()->can_be_primary_super() ) 3161 // This means that if I successfully store an Object into an array-of-String 3162 // I 'forget' that the Object is really now known to be a String. I have to 3163 // do this because we don't have true union types for interfaces - if I store 3164 // a Baz into an array-of-Interface and then tell the optimizer it's an 3165 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3166 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3167 // replace_in_map( obj, res ); 3168 3169 // Return final merged results 3170 set_control( _gvn.transform(region) ); 3171 record_for_igvn(region); 3172 3173 return record_profiled_receiver_for_speculation(res); 3174 } 3175 3176 //------------------------------next_monitor----------------------------------- 3177 // What number should be given to the next monitor? 3178 int GraphKit::next_monitor() { 3179 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3180 int next = current + C->sync_stack_slots(); 3181 // Keep the toplevel high water mark current: 3182 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3183 return current; 3184 } 3185 3186 //------------------------------insert_mem_bar--------------------------------- 3187 // Memory barrier to avoid floating things around 3188 // The membar serves as a pinch point between both control and all memory slices. 3189 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3190 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3191 mb->init_req(TypeFunc::Control, control()); 3192 mb->init_req(TypeFunc::Memory, reset_memory()); 3193 Node* membar = _gvn.transform(mb); 3194 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3195 set_all_memory_call(membar); 3196 return membar; 3197 } 3198 3199 //-------------------------insert_mem_bar_volatile---------------------------- 3200 // Memory barrier to avoid floating things around 3201 // The membar serves as a pinch point between both control and memory(alias_idx). 3202 // If you want to make a pinch point on all memory slices, do not use this 3203 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3204 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3205 // When Parse::do_put_xxx updates a volatile field, it appends a series 3206 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3207 // The first membar is on the same memory slice as the field store opcode. 3208 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3209 // All the other membars (for other volatile slices, including AliasIdxBot, 3210 // which stands for all unknown volatile slices) are control-dependent 3211 // on the first membar. This prevents later volatile loads or stores 3212 // from sliding up past the just-emitted store. 3213 3214 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3215 mb->set_req(TypeFunc::Control,control()); 3216 if (alias_idx == Compile::AliasIdxBot) { 3217 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3218 } else { 3219 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3220 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3221 } 3222 Node* membar = _gvn.transform(mb); 3223 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3224 if (alias_idx == Compile::AliasIdxBot) { 3225 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3226 } else { 3227 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3228 } 3229 return membar; 3230 } 3231 3232 //------------------------------shared_lock------------------------------------ 3233 // Emit locking code. 3234 FastLockNode* GraphKit::shared_lock(Node* obj) { 3235 // bci is either a monitorenter bc or InvocationEntryBci 3236 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3237 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3238 3239 if( !GenerateSynchronizationCode ) 3240 return NULL; // Not locking things? 3241 if (stopped()) // Dead monitor? 3242 return NULL; 3243 3244 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3245 3246 // Box the stack location 3247 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3248 Node* mem = reset_memory(); 3249 3250 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3251 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3252 // Create the counters for this fast lock. 3253 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3254 } 3255 3256 // Create the rtm counters for this fast lock if needed. 3257 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3258 3259 // Add monitor to debug info for the slow path. If we block inside the 3260 // slow path and de-opt, we need the monitor hanging around 3261 map()->push_monitor( flock ); 3262 3263 const TypeFunc *tf = LockNode::lock_type(); 3264 LockNode *lock = new LockNode(C, tf); 3265 3266 lock->init_req( TypeFunc::Control, control() ); 3267 lock->init_req( TypeFunc::Memory , mem ); 3268 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3269 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3270 lock->init_req( TypeFunc::ReturnAdr, top() ); 3271 3272 lock->init_req(TypeFunc::Parms + 0, obj); 3273 lock->init_req(TypeFunc::Parms + 1, box); 3274 lock->init_req(TypeFunc::Parms + 2, flock); 3275 add_safepoint_edges(lock); 3276 3277 lock = _gvn.transform( lock )->as_Lock(); 3278 3279 // lock has no side-effects, sets few values 3280 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3281 3282 insert_mem_bar(Op_MemBarAcquireLock); 3283 3284 // Add this to the worklist so that the lock can be eliminated 3285 record_for_igvn(lock); 3286 3287 #ifndef PRODUCT 3288 if (PrintLockStatistics) { 3289 // Update the counter for this lock. Don't bother using an atomic 3290 // operation since we don't require absolute accuracy. 3291 lock->create_lock_counter(map()->jvms()); 3292 increment_counter(lock->counter()->addr()); 3293 } 3294 #endif 3295 3296 return flock; 3297 } 3298 3299 3300 //------------------------------shared_unlock---------------------------------- 3301 // Emit unlocking code. 3302 void GraphKit::shared_unlock(Node* box, Node* obj) { 3303 // bci is either a monitorenter bc or InvocationEntryBci 3304 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3305 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3306 3307 if( !GenerateSynchronizationCode ) 3308 return; 3309 if (stopped()) { // Dead monitor? 3310 map()->pop_monitor(); // Kill monitor from debug info 3311 return; 3312 } 3313 3314 // Memory barrier to avoid floating things down past the locked region 3315 insert_mem_bar(Op_MemBarReleaseLock); 3316 3317 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3318 UnlockNode *unlock = new UnlockNode(C, tf); 3319 #ifdef ASSERT 3320 unlock->set_dbg_jvms(sync_jvms()); 3321 #endif 3322 uint raw_idx = Compile::AliasIdxRaw; 3323 unlock->init_req( TypeFunc::Control, control() ); 3324 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3325 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3326 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3327 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3328 3329 unlock->init_req(TypeFunc::Parms + 0, obj); 3330 unlock->init_req(TypeFunc::Parms + 1, box); 3331 unlock = _gvn.transform(unlock)->as_Unlock(); 3332 3333 Node* mem = reset_memory(); 3334 3335 // unlock has no side-effects, sets few values 3336 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3337 3338 // Kill monitor from debug info 3339 map()->pop_monitor( ); 3340 } 3341 3342 //-------------------------------get_layout_helper----------------------------- 3343 // If the given klass is a constant or known to be an array, 3344 // fetch the constant layout helper value into constant_value 3345 // and return (Node*)NULL. Otherwise, load the non-constant 3346 // layout helper value, and return the node which represents it. 3347 // This two-faced routine is useful because allocation sites 3348 // almost always feature constant types. 3349 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3350 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3351 if (!StressReflectiveCode && inst_klass != NULL) { 3352 ciKlass* klass = inst_klass->klass(); 3353 bool xklass = inst_klass->klass_is_exact(); 3354 if (xklass || klass->is_array_klass()) { 3355 jint lhelper = klass->layout_helper(); 3356 if (lhelper != Klass::_lh_neutral_value) { 3357 constant_value = lhelper; 3358 return (Node*) NULL; 3359 } 3360 } 3361 } 3362 constant_value = Klass::_lh_neutral_value; // put in a known value 3363 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3364 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3365 } 3366 3367 // We just put in an allocate/initialize with a big raw-memory effect. 3368 // Hook selected additional alias categories on the initialization. 3369 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3370 MergeMemNode* init_in_merge, 3371 Node* init_out_raw) { 3372 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3373 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3374 3375 Node* prevmem = kit.memory(alias_idx); 3376 init_in_merge->set_memory_at(alias_idx, prevmem); 3377 kit.set_memory(init_out_raw, alias_idx); 3378 } 3379 3380 //---------------------------set_output_for_allocation------------------------- 3381 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3382 const TypeOopPtr* oop_type, 3383 bool deoptimize_on_exception) { 3384 int rawidx = Compile::AliasIdxRaw; 3385 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3386 add_safepoint_edges(alloc); 3387 Node* allocx = _gvn.transform(alloc); 3388 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3389 // create memory projection for i_o 3390 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3391 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3392 3393 // create a memory projection as for the normal control path 3394 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3395 set_memory(malloc, rawidx); 3396 3397 // a normal slow-call doesn't change i_o, but an allocation does 3398 // we create a separate i_o projection for the normal control path 3399 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3400 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3401 3402 // put in an initialization barrier 3403 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3404 rawoop)->as_Initialize(); 3405 assert(alloc->initialization() == init, "2-way macro link must work"); 3406 assert(init ->allocation() == alloc, "2-way macro link must work"); 3407 { 3408 // Extract memory strands which may participate in the new object's 3409 // initialization, and source them from the new InitializeNode. 3410 // This will allow us to observe initializations when they occur, 3411 // and link them properly (as a group) to the InitializeNode. 3412 assert(init->in(InitializeNode::Memory) == malloc, ""); 3413 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3414 init->set_req(InitializeNode::Memory, minit_in); 3415 record_for_igvn(minit_in); // fold it up later, if possible 3416 Node* minit_out = memory(rawidx); 3417 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3418 // Add an edge in the MergeMem for the header fields so an access 3419 // to one of those has correct memory state 3420 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 3421 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 3422 if (oop_type->isa_aryptr()) { 3423 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3424 int elemidx = C->get_alias_index(telemref); 3425 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3426 } else if (oop_type->isa_instptr()) { 3427 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3428 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3429 ciField* field = ik->nonstatic_field_at(i); 3430 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3431 continue; // do not bother to track really large numbers of fields 3432 // Find (or create) the alias category for this field: 3433 int fieldidx = C->alias_type(field)->index(); 3434 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3435 } 3436 } 3437 } 3438 3439 // Cast raw oop to the real thing... 3440 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3441 javaoop = _gvn.transform(javaoop); 3442 C->set_recent_alloc(control(), javaoop); 3443 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3444 3445 #ifdef ASSERT 3446 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3447 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3448 "Ideal_allocation works"); 3449 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3450 "Ideal_allocation works"); 3451 if (alloc->is_AllocateArray()) { 3452 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3453 "Ideal_allocation works"); 3454 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3455 "Ideal_allocation works"); 3456 } else { 3457 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3458 } 3459 } 3460 #endif //ASSERT 3461 3462 return javaoop; 3463 } 3464 3465 //---------------------------new_instance-------------------------------------- 3466 // This routine takes a klass_node which may be constant (for a static type) 3467 // or may be non-constant (for reflective code). It will work equally well 3468 // for either, and the graph will fold nicely if the optimizer later reduces 3469 // the type to a constant. 3470 // The optional arguments are for specialized use by intrinsics: 3471 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3472 // - If 'return_size_val', report the the total object size to the caller. 3473 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3474 Node* GraphKit::new_instance(Node* klass_node, 3475 Node* extra_slow_test, 3476 Node* *return_size_val, 3477 bool deoptimize_on_exception) { 3478 // Compute size in doublewords 3479 // The size is always an integral number of doublewords, represented 3480 // as a positive bytewise size stored in the klass's layout_helper. 3481 // The layout_helper also encodes (in a low bit) the need for a slow path. 3482 jint layout_con = Klass::_lh_neutral_value; 3483 Node* layout_val = get_layout_helper(klass_node, layout_con); 3484 int layout_is_con = (layout_val == NULL); 3485 3486 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3487 // Generate the initial go-slow test. It's either ALWAYS (return a 3488 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3489 // case) a computed value derived from the layout_helper. 3490 Node* initial_slow_test = NULL; 3491 if (layout_is_con) { 3492 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3493 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3494 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3495 } else { // reflective case 3496 // This reflective path is used by Unsafe.allocateInstance. 3497 // (It may be stress-tested by specifying StressReflectiveCode.) 3498 // Basically, we want to get into the VM is there's an illegal argument. 3499 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3500 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3501 if (extra_slow_test != intcon(0)) { 3502 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3503 } 3504 // (Macro-expander will further convert this to a Bool, if necessary.) 3505 } 3506 3507 // Find the size in bytes. This is easy; it's the layout_helper. 3508 // The size value must be valid even if the slow path is taken. 3509 Node* size = NULL; 3510 if (layout_is_con) { 3511 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3512 } else { // reflective case 3513 // This reflective path is used by clone and Unsafe.allocateInstance. 3514 size = ConvI2X(layout_val); 3515 3516 // Clear the low bits to extract layout_helper_size_in_bytes: 3517 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3518 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3519 size = _gvn.transform( new AndXNode(size, mask) ); 3520 } 3521 if (return_size_val != NULL) { 3522 (*return_size_val) = size; 3523 } 3524 3525 // This is a precise notnull oop of the klass. 3526 // (Actually, it need not be precise if this is a reflective allocation.) 3527 // It's what we cast the result to. 3528 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3529 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3530 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3531 3532 // Now generate allocation code 3533 3534 // The entire memory state is needed for slow path of the allocation 3535 // since GC and deoptimization can happened. 3536 Node *mem = reset_memory(); 3537 set_all_memory(mem); // Create new memory state 3538 3539 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3540 control(), mem, i_o(), 3541 size, klass_node, 3542 initial_slow_test); 3543 3544 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3545 } 3546 3547 //-------------------------------new_array------------------------------------- 3548 // helper for both newarray and anewarray 3549 // The 'length' parameter is (obviously) the length of the array. 3550 // See comments on new_instance for the meaning of the other arguments. 3551 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3552 Node* length, // number of array elements 3553 int nargs, // number of arguments to push back for uncommon trap 3554 Node* *return_size_val, 3555 bool deoptimize_on_exception) { 3556 jint layout_con = Klass::_lh_neutral_value; 3557 Node* layout_val = get_layout_helper(klass_node, layout_con); 3558 int layout_is_con = (layout_val == NULL); 3559 3560 if (!layout_is_con && !StressReflectiveCode && 3561 !too_many_traps(Deoptimization::Reason_class_check)) { 3562 // This is a reflective array creation site. 3563 // Optimistically assume that it is a subtype of Object[], 3564 // so that we can fold up all the address arithmetic. 3565 layout_con = Klass::array_layout_helper(T_OBJECT); 3566 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3567 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3568 { BuildCutout unless(this, bol_lh, PROB_MAX); 3569 inc_sp(nargs); 3570 uncommon_trap(Deoptimization::Reason_class_check, 3571 Deoptimization::Action_maybe_recompile); 3572 } 3573 layout_val = NULL; 3574 layout_is_con = true; 3575 } 3576 3577 // Generate the initial go-slow test. Make sure we do not overflow 3578 // if length is huge (near 2Gig) or negative! We do not need 3579 // exact double-words here, just a close approximation of needed 3580 // double-words. We can't add any offset or rounding bits, lest we 3581 // take a size -1 of bytes and make it positive. Use an unsigned 3582 // compare, so negative sizes look hugely positive. 3583 int fast_size_limit = FastAllocateSizeLimit; 3584 if (layout_is_con) { 3585 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3586 // Increase the size limit if we have exact knowledge of array type. 3587 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3588 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3589 } 3590 3591 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3592 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3593 3594 // --- Size Computation --- 3595 // array_size = round_to_heap(array_header + (length << elem_shift)); 3596 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3597 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3598 // The rounding mask is strength-reduced, if possible. 3599 int round_mask = MinObjAlignmentInBytes - 1; 3600 Node* header_size = NULL; 3601 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3602 // (T_BYTE has the weakest alignment and size restrictions...) 3603 if (layout_is_con) { 3604 int hsize = Klass::layout_helper_header_size(layout_con); 3605 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3606 BasicType etype = Klass::layout_helper_element_type(layout_con); 3607 if ((round_mask & ~right_n_bits(eshift)) == 0) 3608 round_mask = 0; // strength-reduce it if it goes away completely 3609 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3610 assert(header_size_min <= hsize, "generic minimum is smallest"); 3611 header_size_min = hsize; 3612 header_size = intcon(hsize + round_mask); 3613 } else { 3614 Node* hss = intcon(Klass::_lh_header_size_shift); 3615 Node* hsm = intcon(Klass::_lh_header_size_mask); 3616 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3617 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3618 Node* mask = intcon(round_mask); 3619 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3620 } 3621 3622 Node* elem_shift = NULL; 3623 if (layout_is_con) { 3624 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3625 if (eshift != 0) 3626 elem_shift = intcon(eshift); 3627 } else { 3628 // There is no need to mask or shift this value. 3629 // The semantics of LShiftINode include an implicit mask to 0x1F. 3630 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3631 elem_shift = layout_val; 3632 } 3633 3634 // Transition to native address size for all offset calculations: 3635 Node* lengthx = ConvI2X(length); 3636 Node* headerx = ConvI2X(header_size); 3637 #ifdef _LP64 3638 { const TypeInt* tilen = _gvn.find_int_type(length); 3639 if (tilen != NULL && tilen->_lo < 0) { 3640 // Add a manual constraint to a positive range. Cf. array_element_address. 3641 jint size_max = fast_size_limit; 3642 if (size_max > tilen->_hi) size_max = tilen->_hi; 3643 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3644 3645 // Only do a narrow I2L conversion if the range check passed. 3646 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3647 _gvn.transform(iff); 3648 RegionNode* region = new RegionNode(3); 3649 _gvn.set_type(region, Type::CONTROL); 3650 lengthx = new PhiNode(region, TypeLong::LONG); 3651 _gvn.set_type(lengthx, TypeLong::LONG); 3652 3653 // Range check passed. Use ConvI2L node with narrow type. 3654 Node* passed = IfFalse(iff); 3655 region->init_req(1, passed); 3656 // Make I2L conversion control dependent to prevent it from 3657 // floating above the range check during loop optimizations. 3658 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3659 3660 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3661 region->init_req(2, IfTrue(iff)); 3662 lengthx->init_req(2, ConvI2X(length)); 3663 3664 set_control(region); 3665 record_for_igvn(region); 3666 record_for_igvn(lengthx); 3667 } 3668 } 3669 #endif 3670 3671 // Combine header size (plus rounding) and body size. Then round down. 3672 // This computation cannot overflow, because it is used only in two 3673 // places, one where the length is sharply limited, and the other 3674 // after a successful allocation. 3675 Node* abody = lengthx; 3676 if (elem_shift != NULL) 3677 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3678 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3679 if (round_mask != 0) { 3680 Node* mask = MakeConX(~round_mask); 3681 size = _gvn.transform( new AndXNode(size, mask) ); 3682 } 3683 // else if round_mask == 0, the size computation is self-rounding 3684 3685 if (return_size_val != NULL) { 3686 // This is the size 3687 (*return_size_val) = size; 3688 } 3689 3690 // Now generate allocation code 3691 3692 // The entire memory state is needed for slow path of the allocation 3693 // since GC and deoptimization can happened. 3694 Node *mem = reset_memory(); 3695 set_all_memory(mem); // Create new memory state 3696 3697 if (initial_slow_test->is_Bool()) { 3698 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3699 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3700 } 3701 3702 // Create the AllocateArrayNode and its result projections 3703 AllocateArrayNode* alloc 3704 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3705 control(), mem, i_o(), 3706 size, klass_node, 3707 initial_slow_test, 3708 length); 3709 3710 // Cast to correct type. Note that the klass_node may be constant or not, 3711 // and in the latter case the actual array type will be inexact also. 3712 // (This happens via a non-constant argument to inline_native_newArray.) 3713 // In any case, the value of klass_node provides the desired array type. 3714 const TypeInt* length_type = _gvn.find_int_type(length); 3715 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3716 if (ary_type->isa_aryptr() && length_type != NULL) { 3717 // Try to get a better type than POS for the size 3718 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3719 } 3720 3721 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3722 3723 // Cast length on remaining path to be as narrow as possible 3724 if (map()->find_edge(length) >= 0) { 3725 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3726 if (ccast != length) { 3727 _gvn.set_type_bottom(ccast); 3728 record_for_igvn(ccast); 3729 replace_in_map(length, ccast); 3730 } 3731 } 3732 3733 return javaoop; 3734 } 3735 3736 // The following "Ideal_foo" functions are placed here because they recognize 3737 // the graph shapes created by the functions immediately above. 3738 3739 //---------------------------Ideal_allocation---------------------------------- 3740 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3741 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3742 if (ptr == NULL) { // reduce dumb test in callers 3743 return NULL; 3744 } 3745 3746 #if INCLUDE_SHENANDOAHGC 3747 if (UseShenandoahGC) { 3748 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 3749 ptr = bs->step_over_gc_barrier(ptr); 3750 } 3751 #endif 3752 3753 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3754 ptr = ptr->in(1); 3755 if (ptr == NULL) return NULL; 3756 } 3757 // Return NULL for allocations with several casts: 3758 // j.l.reflect.Array.newInstance(jobject, jint) 3759 // Object.clone() 3760 // to keep more precise type from last cast. 3761 if (ptr->is_Proj()) { 3762 Node* allo = ptr->in(0); 3763 if (allo != NULL && allo->is_Allocate()) { 3764 return allo->as_Allocate(); 3765 } 3766 } 3767 // Report failure to match. 3768 return NULL; 3769 } 3770 3771 // Fancy version which also strips off an offset (and reports it to caller). 3772 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3773 intptr_t& offset) { 3774 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3775 if (base == NULL) return NULL; 3776 return Ideal_allocation(base, phase); 3777 } 3778 3779 // Trace Initialize <- Proj[Parm] <- Allocate 3780 AllocateNode* InitializeNode::allocation() { 3781 Node* rawoop = in(InitializeNode::RawAddress); 3782 if (rawoop->is_Proj()) { 3783 Node* alloc = rawoop->in(0); 3784 if (alloc->is_Allocate()) { 3785 return alloc->as_Allocate(); 3786 } 3787 } 3788 return NULL; 3789 } 3790 3791 // Trace Allocate -> Proj[Parm] -> Initialize 3792 InitializeNode* AllocateNode::initialization() { 3793 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3794 if (rawoop == NULL) return NULL; 3795 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3796 Node* init = rawoop->fast_out(i); 3797 if (init->is_Initialize()) { 3798 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3799 return init->as_Initialize(); 3800 } 3801 } 3802 return NULL; 3803 } 3804 3805 //----------------------------- loop predicates --------------------------- 3806 3807 //------------------------------add_predicate_impl---------------------------- 3808 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3809 // Too many traps seen? 3810 if (too_many_traps(reason)) { 3811 #ifdef ASSERT 3812 if (TraceLoopPredicate) { 3813 int tc = C->trap_count(reason); 3814 tty->print("too many traps=%s tcount=%d in ", 3815 Deoptimization::trap_reason_name(reason), tc); 3816 method()->print(); // which method has too many predicate traps 3817 tty->cr(); 3818 } 3819 #endif 3820 // We cannot afford to take more traps here, 3821 // do not generate predicate. 3822 return; 3823 } 3824 3825 Node *cont = _gvn.intcon(1); 3826 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3827 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3828 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3829 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3830 C->add_predicate_opaq(opq); 3831 { 3832 PreserveJVMState pjvms(this); 3833 set_control(iffalse); 3834 inc_sp(nargs); 3835 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3836 } 3837 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3838 set_control(iftrue); 3839 } 3840 3841 //------------------------------add_predicate--------------------------------- 3842 void GraphKit::add_predicate(int nargs) { 3843 if (UseLoopPredicate) { 3844 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3845 } 3846 if (UseProfiledLoopPredicate) { 3847 add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs); 3848 } 3849 // loop's limit check predicate should be near the loop. 3850 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3851 } 3852 3853 void GraphKit::sync_kit(IdealKit& ideal) { 3854 set_all_memory(ideal.merged_memory()); 3855 set_i_o(ideal.i_o()); 3856 set_control(ideal.ctrl()); 3857 } 3858 3859 void GraphKit::final_sync(IdealKit& ideal) { 3860 // Final sync IdealKit and graphKit. 3861 sync_kit(ideal); 3862 } 3863 3864 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 3865 Node* len = load_array_length(load_String_value(ctrl, str)); 3866 Node* coder = load_String_coder(ctrl, str); 3867 // Divide length by 2 if coder is UTF16 3868 return _gvn.transform(new RShiftINode(len, coder)); 3869 } 3870 3871 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 3872 int value_offset = java_lang_String::value_offset_in_bytes(); 3873 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3874 false, NULL, 0); 3875 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3876 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 3877 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 3878 ciTypeArrayKlass::make(T_BYTE), true, 0); 3879 Node* p = basic_plus_adr(str, str, value_offset); 3880 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 3881 IN_HEAP | C2_CONTROL_DEPENDENT_LOAD); 3882 // String.value field is known to be @Stable. 3883 if (UseImplicitStableValues) { 3884 load = cast_array_to_stable(load, value_type); 3885 } 3886 return load; 3887 } 3888 3889 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 3890 if (!CompactStrings) { 3891 return intcon(java_lang_String::CODER_UTF16); 3892 } 3893 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3894 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3895 false, NULL, 0); 3896 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3897 int coder_field_idx = C->get_alias_index(coder_field_type); 3898 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 3899 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 3900 } 3901 3902 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 3903 int value_offset = java_lang_String::value_offset_in_bytes(); 3904 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3905 false, NULL, 0); 3906 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3907 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 3908 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP); 3909 } 3910 3911 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 3912 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3913 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3914 false, NULL, 0); 3915 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3916 int coder_field_idx = C->get_alias_index(coder_field_type); 3917 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 3918 value, T_BYTE, coder_field_idx, MemNode::unordered); 3919 } 3920 3921 // Capture src and dst memory state with a MergeMemNode 3922 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 3923 if (src_type == dst_type) { 3924 // Types are equal, we don't need a MergeMemNode 3925 return memory(src_type); 3926 } 3927 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 3928 record_for_igvn(merge); // fold it up later, if possible 3929 int src_idx = C->get_alias_index(src_type); 3930 int dst_idx = C->get_alias_index(dst_type); 3931 merge->set_memory_at(src_idx, memory(src_idx)); 3932 merge->set_memory_at(dst_idx, memory(dst_idx)); 3933 return merge; 3934 } 3935 3936 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 3937 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 3938 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 3939 // If input and output memory types differ, capture both states to preserve 3940 // the dependency between preceding and subsequent loads/stores. 3941 // For example, the following program: 3942 // StoreB 3943 // compress_string 3944 // LoadB 3945 // has this memory graph (use->def): 3946 // LoadB -> compress_string -> CharMem 3947 // ... -> StoreB -> ByteMem 3948 // The intrinsic hides the dependency between LoadB and StoreB, causing 3949 // the load to read from memory not containing the result of the StoreB. 3950 // The correct memory graph should look like this: 3951 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 3952 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 3953 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 3954 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 3955 set_memory(res_mem, TypeAryPtr::BYTES); 3956 return str; 3957 } 3958 3959 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 3960 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 3961 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 3962 // Capture src and dst memory (see comment in 'compress_string'). 3963 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 3964 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 3965 set_memory(_gvn.transform(str), dst_type); 3966 } 3967 3968 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 3969 /** 3970 * int i_char = start; 3971 * for (int i_byte = 0; i_byte < count; i_byte++) { 3972 * dst[i_char++] = (char)(src[i_byte] & 0xff); 3973 * } 3974 */ 3975 add_predicate(); 3976 RegionNode* head = new RegionNode(3); 3977 head->init_req(1, control()); 3978 gvn().set_type(head, Type::CONTROL); 3979 record_for_igvn(head); 3980 3981 Node* i_byte = new PhiNode(head, TypeInt::INT); 3982 i_byte->init_req(1, intcon(0)); 3983 gvn().set_type(i_byte, TypeInt::INT); 3984 record_for_igvn(i_byte); 3985 3986 Node* i_char = new PhiNode(head, TypeInt::INT); 3987 i_char->init_req(1, start); 3988 gvn().set_type(i_char, TypeInt::INT); 3989 record_for_igvn(i_char); 3990 3991 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 3992 gvn().set_type(mem, Type::MEMORY); 3993 record_for_igvn(mem); 3994 set_control(head); 3995 set_memory(mem, TypeAryPtr::BYTES); 3996 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 3997 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 3998 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 3999 false, false, true /* mismatched */); 4000 4001 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4002 head->init_req(2, IfTrue(iff)); 4003 mem->init_req(2, st); 4004 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4005 i_char->init_req(2, AddI(i_char, intcon(2))); 4006 4007 set_control(IfFalse(iff)); 4008 set_memory(st, TypeAryPtr::BYTES); 4009 } 4010 4011 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4012 if (!field->is_constant()) { 4013 return NULL; // Field not marked as constant. 4014 } 4015 ciInstance* holder = NULL; 4016 if (!field->is_static()) { 4017 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4018 if (const_oop != NULL && const_oop->is_instance()) { 4019 holder = const_oop->as_instance(); 4020 } 4021 } 4022 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4023 /*is_unsigned_load=*/false); 4024 if (con_type != NULL) { 4025 return makecon(con_type); 4026 } 4027 return NULL; 4028 } 4029 4030 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4031 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4032 // assumption of CCP analysis. 4033 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4034 }