1 /* 2 * Copyright (c) 2001, 2022, 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/ciFlatArrayKlass.hpp" 27 #include "ci/ciInlineKlass.hpp" 28 #include "ci/ciUtilities.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "ci/ciNativeEntryPoint.hpp" 31 #include "ci/ciObjArray.hpp" 32 #include "asm/register.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "gc/shared/barrierSet.hpp" 35 #include "gc/shared/c2/barrierSetC2.hpp" 36 #include "interpreter/interpreter.hpp" 37 #include "memory/resourceArea.hpp" 38 #include "opto/addnode.hpp" 39 #include "opto/castnode.hpp" 40 #include "opto/convertnode.hpp" 41 #include "opto/graphKit.hpp" 42 #include "opto/idealKit.hpp" 43 #include "opto/inlinetypenode.hpp" 44 #include "opto/intrinsicnode.hpp" 45 #include "opto/locknode.hpp" 46 #include "opto/machnode.hpp" 47 #include "opto/narrowptrnode.hpp" 48 #include "opto/opaquenode.hpp" 49 #include "opto/parse.hpp" 50 #include "opto/rootnode.hpp" 51 #include "opto/runtime.hpp" 52 #include "opto/subtypenode.hpp" 53 #include "runtime/deoptimization.hpp" 54 #include "runtime/sharedRuntime.hpp" 55 #include "utilities/bitMap.inline.hpp" 56 #include "utilities/powerOfTwo.hpp" 57 #include "utilities/growableArray.hpp" 58 59 //----------------------------GraphKit----------------------------------------- 60 // Main utility constructor. 61 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn) 62 : Phase(Phase::Parser), 63 _env(C->env()), 64 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()), 65 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 66 { 67 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled"); 68 _exceptions = jvms->map()->next_exception(); 69 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 70 set_jvms(jvms); 71 #ifdef ASSERT 72 if (_gvn.is_IterGVN() != NULL) { 73 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used"); 74 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit) 75 _worklist_size = _gvn.C->for_igvn()->size(); 76 } 77 #endif 78 } 79 80 // Private constructor for parser. 81 GraphKit::GraphKit() 82 : Phase(Phase::Parser), 83 _env(C->env()), 84 _gvn(*C->initial_gvn()), 85 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 86 { 87 _exceptions = NULL; 88 set_map(NULL); 89 debug_only(_sp = -99); 90 debug_only(set_bci(-99)); 91 } 92 93 94 95 //---------------------------clean_stack--------------------------------------- 96 // Clear away rubbish from the stack area of the JVM state. 97 // This destroys any arguments that may be waiting on the stack. 98 void GraphKit::clean_stack(int from_sp) { 99 SafePointNode* map = this->map(); 100 JVMState* jvms = this->jvms(); 101 int stk_size = jvms->stk_size(); 102 int stkoff = jvms->stkoff(); 103 Node* top = this->top(); 104 for (int i = from_sp; i < stk_size; i++) { 105 if (map->in(stkoff + i) != top) { 106 map->set_req(stkoff + i, top); 107 } 108 } 109 } 110 111 112 //--------------------------------sync_jvms----------------------------------- 113 // Make sure our current jvms agrees with our parse state. 114 JVMState* GraphKit::sync_jvms() const { 115 JVMState* jvms = this->jvms(); 116 jvms->set_bci(bci()); // Record the new bci in the JVMState 117 jvms->set_sp(sp()); // Record the new sp in the JVMState 118 assert(jvms_in_sync(), "jvms is now in sync"); 119 return jvms; 120 } 121 122 //--------------------------------sync_jvms_for_reexecute--------------------- 123 // Make sure our current jvms agrees with our parse state. This version 124 // uses the reexecute_sp for reexecuting bytecodes. 125 JVMState* GraphKit::sync_jvms_for_reexecute() { 126 JVMState* jvms = this->jvms(); 127 jvms->set_bci(bci()); // Record the new bci in the JVMState 128 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 129 return jvms; 130 } 131 132 #ifdef ASSERT 133 bool GraphKit::jvms_in_sync() const { 134 Parse* parse = is_Parse(); 135 if (parse == NULL) { 136 if (bci() != jvms()->bci()) return false; 137 if (sp() != (int)jvms()->sp()) return false; 138 return true; 139 } 140 if (jvms()->method() != parse->method()) return false; 141 if (jvms()->bci() != parse->bci()) return false; 142 int jvms_sp = jvms()->sp(); 143 if (jvms_sp != parse->sp()) return false; 144 int jvms_depth = jvms()->depth(); 145 if (jvms_depth != parse->depth()) return false; 146 return true; 147 } 148 149 // Local helper checks for special internal merge points 150 // used to accumulate and merge exception states. 151 // They are marked by the region's in(0) edge being the map itself. 152 // Such merge points must never "escape" into the parser at large, 153 // until they have been handed to gvn.transform. 154 static bool is_hidden_merge(Node* reg) { 155 if (reg == NULL) return false; 156 if (reg->is_Phi()) { 157 reg = reg->in(0); 158 if (reg == NULL) return false; 159 } 160 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 161 } 162 163 void GraphKit::verify_map() const { 164 if (map() == NULL) return; // null map is OK 165 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 166 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 167 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 168 } 169 170 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 171 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 172 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 173 } 174 #endif 175 176 //---------------------------stop_and_kill_map--------------------------------- 177 // Set _map to NULL, signalling a stop to further bytecode execution. 178 // First smash the current map's control to a constant, to mark it dead. 179 void GraphKit::stop_and_kill_map() { 180 SafePointNode* dead_map = stop(); 181 if (dead_map != NULL) { 182 dead_map->disconnect_inputs(C); // Mark the map as killed. 183 assert(dead_map->is_killed(), "must be so marked"); 184 } 185 } 186 187 188 //--------------------------------stopped-------------------------------------- 189 // Tell if _map is NULL, or control is top. 190 bool GraphKit::stopped() { 191 if (map() == NULL) return true; 192 else if (control() == top()) return true; 193 else return false; 194 } 195 196 197 //-----------------------------has_ex_handler---------------------------------- 198 // Tell if this method or any caller method has exception handlers. 199 bool GraphKit::has_ex_handler() { 200 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 201 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 202 return true; 203 } 204 } 205 return false; 206 } 207 208 //------------------------------save_ex_oop------------------------------------ 209 // Save an exception without blowing stack contents or other JVM state. 210 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 211 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 212 ex_map->add_req(ex_oop); 213 debug_only(verify_exception_state(ex_map)); 214 } 215 216 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 217 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 218 Node* ex_oop = ex_map->in(ex_map->req()-1); 219 if (clear_it) ex_map->del_req(ex_map->req()-1); 220 return ex_oop; 221 } 222 223 //-----------------------------saved_ex_oop------------------------------------ 224 // Recover a saved exception from its map. 225 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 226 return common_saved_ex_oop(ex_map, false); 227 } 228 229 //--------------------------clear_saved_ex_oop--------------------------------- 230 // Erase a previously saved exception from its map. 231 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 232 return common_saved_ex_oop(ex_map, true); 233 } 234 235 #ifdef ASSERT 236 //---------------------------has_saved_ex_oop---------------------------------- 237 // Erase a previously saved exception from its map. 238 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 239 return ex_map->req() == ex_map->jvms()->endoff()+1; 240 } 241 #endif 242 243 //-------------------------make_exception_state-------------------------------- 244 // Turn the current JVM state into an exception state, appending the ex_oop. 245 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 246 sync_jvms(); 247 SafePointNode* ex_map = stop(); // do not manipulate this map any more 248 set_saved_ex_oop(ex_map, ex_oop); 249 return ex_map; 250 } 251 252 253 //--------------------------add_exception_state-------------------------------- 254 // Add an exception to my list of exceptions. 255 void GraphKit::add_exception_state(SafePointNode* ex_map) { 256 if (ex_map == NULL || ex_map->control() == top()) { 257 return; 258 } 259 #ifdef ASSERT 260 verify_exception_state(ex_map); 261 if (has_exceptions()) { 262 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 263 } 264 #endif 265 266 // If there is already an exception of exactly this type, merge with it. 267 // In particular, null-checks and other low-level exceptions common up here. 268 Node* ex_oop = saved_ex_oop(ex_map); 269 const Type* ex_type = _gvn.type(ex_oop); 270 if (ex_oop == top()) { 271 // No action needed. 272 return; 273 } 274 assert(ex_type->isa_instptr(), "exception must be an instance"); 275 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 276 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 277 // We check sp also because call bytecodes can generate exceptions 278 // both before and after arguments are popped! 279 if (ex_type2 == ex_type 280 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 281 combine_exception_states(ex_map, e2); 282 return; 283 } 284 } 285 286 // No pre-existing exception of the same type. Chain it on the list. 287 push_exception_state(ex_map); 288 } 289 290 //-----------------------add_exception_states_from----------------------------- 291 void GraphKit::add_exception_states_from(JVMState* jvms) { 292 SafePointNode* ex_map = jvms->map()->next_exception(); 293 if (ex_map != NULL) { 294 jvms->map()->set_next_exception(NULL); 295 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 296 next_map = ex_map->next_exception(); 297 ex_map->set_next_exception(NULL); 298 add_exception_state(ex_map); 299 } 300 } 301 } 302 303 //-----------------------transfer_exceptions_into_jvms------------------------- 304 JVMState* GraphKit::transfer_exceptions_into_jvms() { 305 if (map() == NULL) { 306 // We need a JVMS to carry the exceptions, but the map has gone away. 307 // Create a scratch JVMS, cloned from any of the exception states... 308 if (has_exceptions()) { 309 _map = _exceptions; 310 _map = clone_map(); 311 _map->set_next_exception(NULL); 312 clear_saved_ex_oop(_map); 313 debug_only(verify_map()); 314 } else { 315 // ...or created from scratch 316 JVMState* jvms = new (C) JVMState(_method, NULL); 317 jvms->set_bci(_bci); 318 jvms->set_sp(_sp); 319 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 320 set_jvms(jvms); 321 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 322 set_all_memory(top()); 323 while (map()->req() < jvms->endoff()) map()->add_req(top()); 324 } 325 // (This is a kludge, in case you didn't notice.) 326 set_control(top()); 327 } 328 JVMState* jvms = sync_jvms(); 329 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 330 jvms->map()->set_next_exception(_exceptions); 331 _exceptions = NULL; // done with this set of exceptions 332 return jvms; 333 } 334 335 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 336 assert(is_hidden_merge(dstphi), "must be a special merge node"); 337 assert(is_hidden_merge(srcphi), "must be a special merge node"); 338 uint limit = srcphi->req(); 339 for (uint i = PhiNode::Input; i < limit; i++) { 340 dstphi->add_req(srcphi->in(i)); 341 } 342 } 343 static inline void add_one_req(Node* dstphi, Node* src) { 344 assert(is_hidden_merge(dstphi), "must be a special merge node"); 345 assert(!is_hidden_merge(src), "must not be a special merge node"); 346 dstphi->add_req(src); 347 } 348 349 //-----------------------combine_exception_states------------------------------ 350 // This helper function combines exception states by building phis on a 351 // specially marked state-merging region. These regions and phis are 352 // untransformed, and can build up gradually. The region is marked by 353 // having a control input of its exception map, rather than NULL. Such 354 // regions do not appear except in this function, and in use_exception_state. 355 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 356 if (failing()) return; // dying anyway... 357 JVMState* ex_jvms = ex_map->_jvms; 358 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 359 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 360 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 361 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 362 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 363 assert(ex_map->req() == phi_map->req(), "matching maps"); 364 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 365 Node* hidden_merge_mark = root(); 366 Node* region = phi_map->control(); 367 MergeMemNode* phi_mem = phi_map->merged_memory(); 368 MergeMemNode* ex_mem = ex_map->merged_memory(); 369 if (region->in(0) != hidden_merge_mark) { 370 // The control input is not (yet) a specially-marked region in phi_map. 371 // Make it so, and build some phis. 372 region = new RegionNode(2); 373 _gvn.set_type(region, Type::CONTROL); 374 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 375 region->init_req(1, phi_map->control()); 376 phi_map->set_control(region); 377 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 378 record_for_igvn(io_phi); 379 _gvn.set_type(io_phi, Type::ABIO); 380 phi_map->set_i_o(io_phi); 381 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 382 Node* m = mms.memory(); 383 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 384 record_for_igvn(m_phi); 385 _gvn.set_type(m_phi, Type::MEMORY); 386 mms.set_memory(m_phi); 387 } 388 } 389 390 // Either or both of phi_map and ex_map might already be converted into phis. 391 Node* ex_control = ex_map->control(); 392 // if there is special marking on ex_map also, we add multiple edges from src 393 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 394 // how wide was the destination phi_map, originally? 395 uint orig_width = region->req(); 396 397 if (add_multiple) { 398 add_n_reqs(region, ex_control); 399 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 400 } else { 401 // ex_map has no merges, so we just add single edges everywhere 402 add_one_req(region, ex_control); 403 add_one_req(phi_map->i_o(), ex_map->i_o()); 404 } 405 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 406 if (mms.is_empty()) { 407 // get a copy of the base memory, and patch some inputs into it 408 const TypePtr* adr_type = mms.adr_type(C); 409 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 410 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 411 mms.set_memory(phi); 412 // Prepare to append interesting stuff onto the newly sliced phi: 413 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 414 } 415 // Append stuff from ex_map: 416 if (add_multiple) { 417 add_n_reqs(mms.memory(), mms.memory2()); 418 } else { 419 add_one_req(mms.memory(), mms.memory2()); 420 } 421 } 422 uint limit = ex_map->req(); 423 for (uint i = TypeFunc::Parms; i < limit; i++) { 424 // Skip everything in the JVMS after tos. (The ex_oop follows.) 425 if (i == tos) i = ex_jvms->monoff(); 426 Node* src = ex_map->in(i); 427 Node* dst = phi_map->in(i); 428 if (src != dst) { 429 PhiNode* phi; 430 if (dst->in(0) != region) { 431 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 432 record_for_igvn(phi); 433 _gvn.set_type(phi, phi->type()); 434 phi_map->set_req(i, dst); 435 // Prepare to append interesting stuff onto the new phi: 436 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 437 } else { 438 assert(dst->is_Phi(), "nobody else uses a hidden region"); 439 phi = dst->as_Phi(); 440 } 441 if (add_multiple && src->in(0) == ex_control) { 442 // Both are phis. 443 add_n_reqs(dst, src); 444 } else { 445 while (dst->req() < region->req()) add_one_req(dst, src); 446 } 447 const Type* srctype = _gvn.type(src); 448 if (phi->type() != srctype) { 449 const Type* dsttype = phi->type()->meet_speculative(srctype); 450 if (phi->type() != dsttype) { 451 phi->set_type(dsttype); 452 _gvn.set_type(phi, dsttype); 453 } 454 } 455 } 456 } 457 phi_map->merge_replaced_nodes_with(ex_map); 458 } 459 460 //--------------------------use_exception_state-------------------------------- 461 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 462 if (failing()) { stop(); return top(); } 463 Node* region = phi_map->control(); 464 Node* hidden_merge_mark = root(); 465 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 466 Node* ex_oop = clear_saved_ex_oop(phi_map); 467 if (region->in(0) == hidden_merge_mark) { 468 // Special marking for internal ex-states. Process the phis now. 469 region->set_req(0, region); // now it's an ordinary region 470 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 471 // Note: Setting the jvms also sets the bci and sp. 472 set_control(_gvn.transform(region)); 473 uint tos = jvms()->stkoff() + sp(); 474 for (uint i = 1; i < tos; i++) { 475 Node* x = phi_map->in(i); 476 if (x->in(0) == region) { 477 assert(x->is_Phi(), "expected a special phi"); 478 phi_map->set_req(i, _gvn.transform(x)); 479 } 480 } 481 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 482 Node* x = mms.memory(); 483 if (x->in(0) == region) { 484 assert(x->is_Phi(), "nobody else uses a hidden region"); 485 mms.set_memory(_gvn.transform(x)); 486 } 487 } 488 if (ex_oop->in(0) == region) { 489 assert(ex_oop->is_Phi(), "expected a special phi"); 490 ex_oop = _gvn.transform(ex_oop); 491 } 492 } else { 493 set_jvms(phi_map->jvms()); 494 } 495 496 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 497 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 498 return ex_oop; 499 } 500 501 //---------------------------------java_bc------------------------------------- 502 Bytecodes::Code GraphKit::java_bc() const { 503 ciMethod* method = this->method(); 504 int bci = this->bci(); 505 if (method != NULL && bci != InvocationEntryBci) 506 return method->java_code_at_bci(bci); 507 else 508 return Bytecodes::_illegal; 509 } 510 511 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 512 bool must_throw) { 513 // if the exception capability is set, then we will generate code 514 // to check the JavaThread.should_post_on_exceptions flag to see 515 // if we actually need to report exception events (for this 516 // thread). If we don't need to report exception events, we will 517 // take the normal fast path provided by add_exception_events. If 518 // exception event reporting is enabled for this thread, we will 519 // take the uncommon_trap in the BuildCutout below. 520 521 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 522 Node* jthread = _gvn.transform(new ThreadLocalNode()); 523 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 524 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 525 526 // Test the should_post_on_exceptions_flag vs. 0 527 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 528 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 529 530 // Branch to slow_path if should_post_on_exceptions_flag was true 531 { BuildCutout unless(this, tst, PROB_MAX); 532 // Do not try anything fancy if we're notifying the VM on every throw. 533 // Cf. case Bytecodes::_athrow in parse2.cpp. 534 uncommon_trap(reason, Deoptimization::Action_none, 535 (ciKlass*)NULL, (char*)NULL, must_throw); 536 } 537 538 } 539 540 //------------------------------builtin_throw---------------------------------- 541 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 542 bool must_throw = true; 543 544 // If this particular condition has not yet happened at this 545 // bytecode, then use the uncommon trap mechanism, and allow for 546 // a future recompilation if several traps occur here. 547 // If the throw is hot, try to use a more complicated inline mechanism 548 // which keeps execution inside the compiled code. 549 bool treat_throw_as_hot = false; 550 ciMethodData* md = method()->method_data(); 551 552 if (ProfileTraps) { 553 if (too_many_traps(reason)) { 554 treat_throw_as_hot = true; 555 } 556 // (If there is no MDO at all, assume it is early in 557 // execution, and that any deopts are part of the 558 // startup transient, and don't need to be remembered.) 559 560 // Also, if there is a local exception handler, treat all throws 561 // as hot if there has been at least one in this method. 562 if (C->trap_count(reason) != 0 563 && method()->method_data()->trap_count(reason) != 0 564 && has_ex_handler()) { 565 treat_throw_as_hot = true; 566 } 567 } 568 569 // If this throw happens frequently, an uncommon trap might cause 570 // a performance pothole. If there is a local exception handler, 571 // and if this particular bytecode appears to be deoptimizing often, 572 // let us handle the throw inline, with a preconstructed instance. 573 // Note: If the deopt count has blown up, the uncommon trap 574 // runtime is going to flush this nmethod, not matter what. 575 if (treat_throw_as_hot && method()->can_omit_stack_trace()) { 576 // If the throw is local, we use a pre-existing instance and 577 // punt on the backtrace. This would lead to a missing backtrace 578 // (a repeat of 4292742) if the backtrace object is ever asked 579 // for its backtrace. 580 // Fixing this remaining case of 4292742 requires some flavor of 581 // escape analysis. Leave that for the future. 582 ciInstance* ex_obj = NULL; 583 switch (reason) { 584 case Deoptimization::Reason_null_check: 585 ex_obj = env()->NullPointerException_instance(); 586 break; 587 case Deoptimization::Reason_div0_check: 588 ex_obj = env()->ArithmeticException_instance(); 589 break; 590 case Deoptimization::Reason_range_check: 591 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 592 break; 593 case Deoptimization::Reason_class_check: 594 ex_obj = env()->ClassCastException_instance(); 595 break; 596 case Deoptimization::Reason_array_check: 597 ex_obj = env()->ArrayStoreException_instance(); 598 break; 599 default: 600 break; 601 } 602 if (failing()) { stop(); return; } // exception allocation might fail 603 if (ex_obj != NULL) { 604 if (env()->jvmti_can_post_on_exceptions()) { 605 // check if we must post exception events, take uncommon trap if so 606 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 607 // here if should_post_on_exceptions is false 608 // continue on with the normal codegen 609 } 610 611 // Cheat with a preallocated exception object. 612 if (C->log() != NULL) 613 C->log()->elem("hot_throw preallocated='1' reason='%s'", 614 Deoptimization::trap_reason_name(reason)); 615 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 616 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 617 618 // Clear the detail message of the preallocated exception object. 619 // Weblogic sometimes mutates the detail message of exceptions 620 // using reflection. 621 int offset = java_lang_Throwable::get_detailMessage_offset(); 622 const TypePtr* adr_typ = ex_con->add_offset(offset); 623 624 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 625 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 626 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 627 628 if (!method()->has_exception_handlers()) { 629 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway. 630 // This prevents issues with exception handling and late inlining. 631 set_sp(0); 632 clean_stack(0); 633 } 634 635 add_exception_state(make_exception_state(ex_node)); 636 return; 637 } 638 } 639 640 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 641 // It won't be much cheaper than bailing to the interp., since we'll 642 // have to pass up all the debug-info, and the runtime will have to 643 // create the stack trace. 644 645 // Usual case: Bail to interpreter. 646 // Reserve the right to recompile if we haven't seen anything yet. 647 648 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 649 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 650 if (treat_throw_as_hot 651 && (method()->method_data()->trap_recompiled_at(bci(), m) 652 || C->too_many_traps(reason))) { 653 // We cannot afford to take more traps here. Suffer in the interpreter. 654 if (C->log() != NULL) 655 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 656 Deoptimization::trap_reason_name(reason), 657 C->trap_count(reason)); 658 action = Deoptimization::Action_none; 659 } 660 661 // "must_throw" prunes the JVM state to include only the stack, if there 662 // are no local exception handlers. This should cut down on register 663 // allocation time and code size, by drastically reducing the number 664 // of in-edges on the call to the uncommon trap. 665 666 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 667 } 668 669 670 //----------------------------PreserveJVMState--------------------------------- 671 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 672 debug_only(kit->verify_map()); 673 _kit = kit; 674 _map = kit->map(); // preserve the map 675 _sp = kit->sp(); 676 kit->set_map(clone_map ? kit->clone_map() : NULL); 677 #ifdef ASSERT 678 _bci = kit->bci(); 679 Parse* parser = kit->is_Parse(); 680 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 681 _block = block; 682 #endif 683 } 684 PreserveJVMState::~PreserveJVMState() { 685 GraphKit* kit = _kit; 686 #ifdef ASSERT 687 assert(kit->bci() == _bci, "bci must not shift"); 688 Parse* parser = kit->is_Parse(); 689 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 690 assert(block == _block, "block must not shift"); 691 #endif 692 kit->set_map(_map); 693 kit->set_sp(_sp); 694 } 695 696 697 //-----------------------------BuildCutout------------------------------------- 698 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 699 : PreserveJVMState(kit) 700 { 701 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 702 SafePointNode* outer_map = _map; // preserved map is caller's 703 SafePointNode* inner_map = kit->map(); 704 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 705 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 706 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 707 } 708 BuildCutout::~BuildCutout() { 709 GraphKit* kit = _kit; 710 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 711 } 712 713 //---------------------------PreserveReexecuteState---------------------------- 714 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 715 assert(!kit->stopped(), "must call stopped() before"); 716 _kit = kit; 717 _sp = kit->sp(); 718 _reexecute = kit->jvms()->_reexecute; 719 } 720 PreserveReexecuteState::~PreserveReexecuteState() { 721 if (_kit->stopped()) return; 722 _kit->jvms()->_reexecute = _reexecute; 723 _kit->set_sp(_sp); 724 } 725 726 //------------------------------clone_map-------------------------------------- 727 // Implementation of PreserveJVMState 728 // 729 // Only clone_map(...) here. If this function is only used in the 730 // PreserveJVMState class we may want to get rid of this extra 731 // function eventually and do it all there. 732 733 SafePointNode* GraphKit::clone_map() { 734 if (map() == NULL) return NULL; 735 736 // Clone the memory edge first 737 Node* mem = MergeMemNode::make(map()->memory()); 738 gvn().set_type_bottom(mem); 739 740 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 741 JVMState* jvms = this->jvms(); 742 JVMState* clonejvms = jvms->clone_shallow(C); 743 clonemap->set_memory(mem); 744 clonemap->set_jvms(clonejvms); 745 clonejvms->set_map(clonemap); 746 record_for_igvn(clonemap); 747 gvn().set_type_bottom(clonemap); 748 return clonemap; 749 } 750 751 752 //-----------------------------set_map_clone----------------------------------- 753 void GraphKit::set_map_clone(SafePointNode* m) { 754 _map = m; 755 _map = clone_map(); 756 _map->set_next_exception(NULL); 757 debug_only(verify_map()); 758 } 759 760 761 //----------------------------kill_dead_locals--------------------------------- 762 // Detect any locals which are known to be dead, and force them to top. 763 void GraphKit::kill_dead_locals() { 764 // Consult the liveness information for the locals. If any 765 // of them are unused, then they can be replaced by top(). This 766 // should help register allocation time and cut down on the size 767 // of the deoptimization information. 768 769 // This call is made from many of the bytecode handling 770 // subroutines called from the Big Switch in do_one_bytecode. 771 // Every bytecode which might include a slow path is responsible 772 // for killing its dead locals. The more consistent we 773 // are about killing deads, the fewer useless phis will be 774 // constructed for them at various merge points. 775 776 // bci can be -1 (InvocationEntryBci). We return the entry 777 // liveness for the method. 778 779 if (method() == NULL || method()->code_size() == 0) { 780 // We are building a graph for a call to a native method. 781 // All locals are live. 782 return; 783 } 784 785 ResourceMark rm; 786 787 // Consult the liveness information for the locals. If any 788 // of them are unused, then they can be replaced by top(). This 789 // should help register allocation time and cut down on the size 790 // of the deoptimization information. 791 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 792 793 int len = (int)live_locals.size(); 794 assert(len <= jvms()->loc_size(), "too many live locals"); 795 for (int local = 0; local < len; local++) { 796 if (!live_locals.at(local)) { 797 set_local(local, top()); 798 } 799 } 800 } 801 802 #ifdef ASSERT 803 //-------------------------dead_locals_are_killed------------------------------ 804 // Return true if all dead locals are set to top in the map. 805 // Used to assert "clean" debug info at various points. 806 bool GraphKit::dead_locals_are_killed() { 807 if (method() == NULL || method()->code_size() == 0) { 808 // No locals need to be dead, so all is as it should be. 809 return true; 810 } 811 812 // Make sure somebody called kill_dead_locals upstream. 813 ResourceMark rm; 814 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 815 if (jvms->loc_size() == 0) continue; // no locals to consult 816 SafePointNode* map = jvms->map(); 817 ciMethod* method = jvms->method(); 818 int bci = jvms->bci(); 819 if (jvms == this->jvms()) { 820 bci = this->bci(); // it might not yet be synched 821 } 822 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 823 int len = (int)live_locals.size(); 824 if (!live_locals.is_valid() || len == 0) 825 // This method is trivial, or is poisoned by a breakpoint. 826 return true; 827 assert(len == jvms->loc_size(), "live map consistent with locals map"); 828 for (int local = 0; local < len; local++) { 829 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 830 if (PrintMiscellaneous && (Verbose || WizardMode)) { 831 tty->print_cr("Zombie local %d: ", local); 832 jvms->dump(); 833 } 834 return false; 835 } 836 } 837 } 838 return true; 839 } 840 841 #endif //ASSERT 842 843 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens. 844 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 845 ciMethod* cur_method = jvms->method(); 846 int cur_bci = jvms->bci(); 847 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 848 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 849 return Interpreter::bytecode_should_reexecute(code) || 850 (is_anewarray && (code == Bytecodes::_multianewarray)); 851 // Reexecute _multianewarray bytecode which was replaced with 852 // sequence of [a]newarray. See Parse::do_multianewarray(). 853 // 854 // Note: interpreter should not have it set since this optimization 855 // is limited by dimensions and guarded by flag so in some cases 856 // multianewarray() runtime calls will be generated and 857 // the bytecode should not be reexecutes (stack will not be reset). 858 } else { 859 return false; 860 } 861 } 862 863 // Helper function for adding JVMState and debug information to node 864 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 865 // Add the safepoint edges to the call (or other safepoint). 866 867 // Make sure dead locals are set to top. This 868 // should help register allocation time and cut down on the size 869 // of the deoptimization information. 870 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 871 872 // Walk the inline list to fill in the correct set of JVMState's 873 // Also fill in the associated edges for each JVMState. 874 875 // If the bytecode needs to be reexecuted we need to put 876 // the arguments back on the stack. 877 const bool should_reexecute = jvms()->should_reexecute(); 878 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 879 880 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 881 // undefined if the bci is different. This is normal for Parse but it 882 // should not happen for LibraryCallKit because only one bci is processed. 883 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 884 "in LibraryCallKit the reexecute bit should not change"); 885 886 // If we are guaranteed to throw, we can prune everything but the 887 // input to the current bytecode. 888 bool can_prune_locals = false; 889 uint stack_slots_not_pruned = 0; 890 int inputs = 0, depth = 0; 891 if (must_throw) { 892 assert(method() == youngest_jvms->method(), "sanity"); 893 if (compute_stack_effects(inputs, depth)) { 894 can_prune_locals = true; 895 stack_slots_not_pruned = inputs; 896 } 897 } 898 899 if (env()->should_retain_local_variables()) { 900 // At any safepoint, this method can get breakpointed, which would 901 // then require an immediate deoptimization. 902 can_prune_locals = false; // do not prune locals 903 stack_slots_not_pruned = 0; 904 } 905 906 // do not scribble on the input jvms 907 JVMState* out_jvms = youngest_jvms->clone_deep(C); 908 call->set_jvms(out_jvms); // Start jvms list for call node 909 910 // For a known set of bytecodes, the interpreter should reexecute them if 911 // deoptimization happens. We set the reexecute state for them here 912 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 913 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 914 #ifdef ASSERT 915 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects() 916 assert(method() == youngest_jvms->method(), "sanity"); 917 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc())); 918 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution"); 919 #endif // ASSERT 920 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 921 } 922 923 // Presize the call: 924 DEBUG_ONLY(uint non_debug_edges = call->req()); 925 call->add_req_batch(top(), youngest_jvms->debug_depth()); 926 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 927 928 // Set up edges so that the call looks like this: 929 // Call [state:] ctl io mem fptr retadr 930 // [parms:] parm0 ... parmN 931 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 932 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 933 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 934 // Note that caller debug info precedes callee debug info. 935 936 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 937 uint debug_ptr = call->req(); 938 939 // Loop over the map input edges associated with jvms, add them 940 // to the call node, & reset all offsets to match call node array. 941 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 942 uint debug_end = debug_ptr; 943 uint debug_start = debug_ptr - in_jvms->debug_size(); 944 debug_ptr = debug_start; // back up the ptr 945 946 uint p = debug_start; // walks forward in [debug_start, debug_end) 947 uint j, k, l; 948 SafePointNode* in_map = in_jvms->map(); 949 out_jvms->set_map(call); 950 951 if (can_prune_locals) { 952 assert(in_jvms->method() == out_jvms->method(), "sanity"); 953 // If the current throw can reach an exception handler in this JVMS, 954 // then we must keep everything live that can reach that handler. 955 // As a quick and dirty approximation, we look for any handlers at all. 956 if (in_jvms->method()->has_exception_handlers()) { 957 can_prune_locals = false; 958 } 959 } 960 961 // Add the Locals 962 k = in_jvms->locoff(); 963 l = in_jvms->loc_size(); 964 out_jvms->set_locoff(p); 965 if (!can_prune_locals) { 966 for (j = 0; j < l; j++) 967 call->set_req(p++, in_map->in(k+j)); 968 } else { 969 p += l; // already set to top above by add_req_batch 970 } 971 972 // Add the Expression Stack 973 k = in_jvms->stkoff(); 974 l = in_jvms->sp(); 975 out_jvms->set_stkoff(p); 976 if (!can_prune_locals) { 977 for (j = 0; j < l; j++) 978 call->set_req(p++, in_map->in(k+j)); 979 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 980 // Divide stack into {S0,...,S1}, where S0 is set to top. 981 uint s1 = stack_slots_not_pruned; 982 stack_slots_not_pruned = 0; // for next iteration 983 if (s1 > l) s1 = l; 984 uint s0 = l - s1; 985 p += s0; // skip the tops preinstalled by add_req_batch 986 for (j = s0; j < l; j++) 987 call->set_req(p++, in_map->in(k+j)); 988 } else { 989 p += l; // already set to top above by add_req_batch 990 } 991 992 // Add the Monitors 993 k = in_jvms->monoff(); 994 l = in_jvms->mon_size(); 995 out_jvms->set_monoff(p); 996 for (j = 0; j < l; j++) 997 call->set_req(p++, in_map->in(k+j)); 998 999 // Copy any scalar object fields. 1000 k = in_jvms->scloff(); 1001 l = in_jvms->scl_size(); 1002 out_jvms->set_scloff(p); 1003 for (j = 0; j < l; j++) 1004 call->set_req(p++, in_map->in(k+j)); 1005 1006 // Finish the new jvms. 1007 out_jvms->set_endoff(p); 1008 1009 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 1010 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 1011 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 1012 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 1013 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 1014 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 1015 1016 // Update the two tail pointers in parallel. 1017 out_jvms = out_jvms->caller(); 1018 in_jvms = in_jvms->caller(); 1019 } 1020 1021 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 1022 1023 // Test the correctness of JVMState::debug_xxx accessors: 1024 assert(call->jvms()->debug_start() == non_debug_edges, ""); 1025 assert(call->jvms()->debug_end() == call->req(), ""); 1026 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 1027 } 1028 1029 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1030 Bytecodes::Code code = java_bc(); 1031 if (code == Bytecodes::_wide) { 1032 code = method()->java_code_at_bci(bci() + 1); 1033 } 1034 1035 BasicType rtype = T_ILLEGAL; 1036 int rsize = 0; 1037 1038 if (code != Bytecodes::_illegal) { 1039 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1040 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1041 if (rtype < T_CONFLICT) 1042 rsize = type2size[rtype]; 1043 } 1044 1045 switch (code) { 1046 case Bytecodes::_illegal: 1047 return false; 1048 1049 case Bytecodes::_ldc: 1050 case Bytecodes::_ldc_w: 1051 case Bytecodes::_ldc2_w: 1052 inputs = 0; 1053 break; 1054 1055 case Bytecodes::_dup: inputs = 1; break; 1056 case Bytecodes::_dup_x1: inputs = 2; break; 1057 case Bytecodes::_dup_x2: inputs = 3; break; 1058 case Bytecodes::_dup2: inputs = 2; break; 1059 case Bytecodes::_dup2_x1: inputs = 3; break; 1060 case Bytecodes::_dup2_x2: inputs = 4; break; 1061 case Bytecodes::_swap: inputs = 2; break; 1062 case Bytecodes::_arraylength: inputs = 1; break; 1063 1064 case Bytecodes::_getstatic: 1065 case Bytecodes::_putstatic: 1066 case Bytecodes::_getfield: 1067 case Bytecodes::_putfield: 1068 { 1069 bool ignored_will_link; 1070 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1071 int size = field->type()->size(); 1072 bool is_get = (depth >= 0), is_static = (depth & 1); 1073 inputs = (is_static ? 0 : 1); 1074 if (is_get) { 1075 depth = size - inputs; 1076 } else { 1077 inputs += size; // putxxx pops the value from the stack 1078 depth = - inputs; 1079 } 1080 } 1081 break; 1082 1083 case Bytecodes::_invokevirtual: 1084 case Bytecodes::_invokespecial: 1085 case Bytecodes::_invokestatic: 1086 case Bytecodes::_invokedynamic: 1087 case Bytecodes::_invokeinterface: 1088 { 1089 bool ignored_will_link; 1090 ciSignature* declared_signature = NULL; 1091 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1092 assert(declared_signature != NULL, "cannot be null"); 1093 inputs = declared_signature->arg_size_for_bc(code); 1094 int size = declared_signature->return_type()->size(); 1095 depth = size - inputs; 1096 } 1097 break; 1098 1099 case Bytecodes::_multianewarray: 1100 { 1101 ciBytecodeStream iter(method()); 1102 iter.reset_to_bci(bci()); 1103 iter.next(); 1104 inputs = iter.get_dimensions(); 1105 assert(rsize == 1, ""); 1106 depth = rsize - inputs; 1107 } 1108 break; 1109 1110 case Bytecodes::_withfield: { 1111 bool ignored_will_link; 1112 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1113 int size = field->type()->size(); 1114 inputs = size+1; 1115 depth = rsize - inputs; 1116 break; 1117 } 1118 1119 case Bytecodes::_ireturn: 1120 case Bytecodes::_lreturn: 1121 case Bytecodes::_freturn: 1122 case Bytecodes::_dreturn: 1123 case Bytecodes::_areturn: 1124 assert(rsize == -depth, ""); 1125 inputs = rsize; 1126 break; 1127 1128 case Bytecodes::_jsr: 1129 case Bytecodes::_jsr_w: 1130 inputs = 0; 1131 depth = 1; // S.B. depth=1, not zero 1132 break; 1133 1134 default: 1135 // bytecode produces a typed result 1136 inputs = rsize - depth; 1137 assert(inputs >= 0, ""); 1138 break; 1139 } 1140 1141 #ifdef ASSERT 1142 // spot check 1143 int outputs = depth + inputs; 1144 assert(outputs >= 0, "sanity"); 1145 switch (code) { 1146 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1147 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1148 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1149 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1150 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1151 default: break; 1152 } 1153 #endif //ASSERT 1154 1155 return true; 1156 } 1157 1158 1159 1160 //------------------------------basic_plus_adr--------------------------------- 1161 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1162 // short-circuit a common case 1163 if (offset == intcon(0)) return ptr; 1164 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1165 } 1166 1167 Node* GraphKit::ConvI2L(Node* offset) { 1168 // short-circuit a common case 1169 jint offset_con = find_int_con(offset, Type::OffsetBot); 1170 if (offset_con != Type::OffsetBot) { 1171 return longcon((jlong) offset_con); 1172 } 1173 return _gvn.transform( new ConvI2LNode(offset)); 1174 } 1175 1176 Node* GraphKit::ConvI2UL(Node* offset) { 1177 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1178 if (offset_con != (juint) Type::OffsetBot) { 1179 return longcon((julong) offset_con); 1180 } 1181 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1182 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1183 return _gvn.transform( new AndLNode(conv, mask) ); 1184 } 1185 1186 Node* GraphKit::ConvL2I(Node* offset) { 1187 // short-circuit a common case 1188 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1189 if (offset_con != (jlong)Type::OffsetBot) { 1190 return intcon((int) offset_con); 1191 } 1192 return _gvn.transform( new ConvL2INode(offset)); 1193 } 1194 1195 //-------------------------load_object_klass----------------------------------- 1196 Node* GraphKit::load_object_klass(Node* obj) { 1197 // Special-case a fresh allocation to avoid building nodes: 1198 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1199 if (akls != NULL) return akls; 1200 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1201 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT)); 1202 } 1203 1204 //-------------------------load_array_length----------------------------------- 1205 Node* GraphKit::load_array_length(Node* array) { 1206 // Special-case a fresh allocation to avoid building nodes: 1207 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1208 Node *alen; 1209 if (alloc == NULL) { 1210 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1211 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1212 } else { 1213 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false); 1214 } 1215 return alen; 1216 } 1217 1218 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc, 1219 const TypeOopPtr* oop_type, 1220 bool replace_length_in_map) { 1221 Node* length = alloc->Ideal_length(); 1222 if (replace_length_in_map == false || map()->find_edge(length) >= 0) { 1223 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn); 1224 if (ccast != length) { 1225 // do not transfrom ccast here, it might convert to top node for 1226 // negative array length and break assumptions in parsing stage. 1227 _gvn.set_type_bottom(ccast); 1228 record_for_igvn(ccast); 1229 if (replace_length_in_map) { 1230 replace_in_map(length, ccast); 1231 } 1232 return ccast; 1233 } 1234 } 1235 return length; 1236 } 1237 1238 //------------------------------do_null_check---------------------------------- 1239 // Helper function to do a NULL pointer check. Returned value is 1240 // the incoming address with NULL casted away. You are allowed to use the 1241 // not-null value only if you are control dependent on the test. 1242 #ifndef PRODUCT 1243 extern int explicit_null_checks_inserted, 1244 explicit_null_checks_elided; 1245 #endif 1246 Node* GraphKit::null_check_common(Node* value, BasicType type, 1247 // optional arguments for variations: 1248 bool assert_null, 1249 Node* *null_control, 1250 bool speculative, 1251 bool is_init_check) { 1252 assert(!assert_null || null_control == NULL, "not both at once"); 1253 if (stopped()) return top(); 1254 NOT_PRODUCT(explicit_null_checks_inserted++); 1255 1256 if (value->is_InlineType()) { 1257 InlineTypeNode* vt = value->as_InlineType(); 1258 null_check_common(vt->get_is_init(), T_INT, assert_null, null_control, speculative, true); 1259 if (stopped()) { 1260 return top(); 1261 } 1262 if (assert_null) { 1263 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts) 1264 // vt = InlineTypeNode::make_null(_gvn, vt->type()->inline_klass()); 1265 // replace_in_map(value, vt); 1266 // return vt; 1267 return null(); 1268 } 1269 bool do_replace_in_map = (null_control == NULL || (*null_control) == top()); 1270 return cast_not_null(value, do_replace_in_map); 1271 } else if (value->is_InlineTypePtr()) { 1272 // Null checking a scalarized but nullable inline type. Check the IsInit 1273 // input instead of the oop input to avoid keeping buffer allocations alive. 1274 InlineTypePtrNode* vtptr = value->as_InlineTypePtr(); 1275 while (vtptr->get_oop()->is_InlineTypePtr()) { 1276 vtptr = vtptr->get_oop()->as_InlineTypePtr(); 1277 } 1278 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true); 1279 if (stopped()) { 1280 return top(); 1281 } 1282 if (assert_null) { 1283 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts) 1284 // vtptr = InlineTypePtrNode::make_null(_gvn, vtptr->type()->inline_klass()); 1285 // replace_in_map(value, vtptr); 1286 // return vtptr; 1287 return null(); 1288 } 1289 bool do_replace_in_map = (null_control == NULL || (*null_control) == top()); 1290 return cast_not_null(value, do_replace_in_map); 1291 } 1292 1293 // Construct NULL check 1294 Node *chk = NULL; 1295 switch(type) { 1296 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1297 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1298 case T_PRIMITIVE_OBJECT : // fall through 1299 case T_ARRAY : // fall through 1300 type = T_OBJECT; // simplify further tests 1301 case T_OBJECT : { 1302 const Type *t = _gvn.type( value ); 1303 1304 const TypeOopPtr* tp = t->isa_oopptr(); 1305 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1306 // Only for do_null_check, not any of its siblings: 1307 && !assert_null && null_control == NULL) { 1308 // Usually, any field access or invocation on an unloaded oop type 1309 // will simply fail to link, since the statically linked class is 1310 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1311 // the static class is loaded but the sharper oop type is not. 1312 // Rather than checking for this obscure case in lots of places, 1313 // we simply observe that a null check on an unloaded class 1314 // will always be followed by a nonsense operation, so we 1315 // can just issue the uncommon trap here. 1316 // Our access to the unloaded class will only be correct 1317 // after it has been loaded and initialized, which requires 1318 // a trip through the interpreter. 1319 #ifndef PRODUCT 1320 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1321 #endif 1322 uncommon_trap(Deoptimization::Reason_unloaded, 1323 Deoptimization::Action_reinterpret, 1324 tp->klass(), "!loaded"); 1325 return top(); 1326 } 1327 1328 if (assert_null) { 1329 // See if the type is contained in NULL_PTR. 1330 // If so, then the value is already null. 1331 if (t->higher_equal(TypePtr::NULL_PTR)) { 1332 NOT_PRODUCT(explicit_null_checks_elided++); 1333 return value; // Elided null assert quickly! 1334 } 1335 } else { 1336 // See if mixing in the NULL pointer changes type. 1337 // If so, then the NULL pointer was not allowed in the original 1338 // type. In other words, "value" was not-null. 1339 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1340 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1341 NOT_PRODUCT(explicit_null_checks_elided++); 1342 return value; // Elided null check quickly! 1343 } 1344 } 1345 chk = new CmpPNode( value, null() ); 1346 break; 1347 } 1348 1349 default: 1350 fatal("unexpected type: %s", type2name(type)); 1351 } 1352 assert(chk != NULL, "sanity check"); 1353 chk = _gvn.transform(chk); 1354 1355 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1356 BoolNode *btst = new BoolNode( chk, btest); 1357 Node *tst = _gvn.transform( btst ); 1358 1359 //----------- 1360 // if peephole optimizations occurred, a prior test existed. 1361 // If a prior test existed, maybe it dominates as we can avoid this test. 1362 if (tst != btst && type == T_OBJECT) { 1363 // At this point we want to scan up the CFG to see if we can 1364 // find an identical test (and so avoid this test altogether). 1365 Node *cfg = control(); 1366 int depth = 0; 1367 while( depth < 16 ) { // Limit search depth for speed 1368 if( cfg->Opcode() == Op_IfTrue && 1369 cfg->in(0)->in(1) == tst ) { 1370 // Found prior test. Use "cast_not_null" to construct an identical 1371 // CastPP (and hence hash to) as already exists for the prior test. 1372 // Return that casted value. 1373 if (assert_null) { 1374 replace_in_map(value, null()); 1375 return null(); // do not issue the redundant test 1376 } 1377 Node *oldcontrol = control(); 1378 set_control(cfg); 1379 Node *res = cast_not_null(value); 1380 set_control(oldcontrol); 1381 NOT_PRODUCT(explicit_null_checks_elided++); 1382 return res; 1383 } 1384 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1385 if (cfg == NULL) break; // Quit at region nodes 1386 depth++; 1387 } 1388 } 1389 1390 //----------- 1391 // Branch to failure if null 1392 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1393 Deoptimization::DeoptReason reason; 1394 if (assert_null) { 1395 reason = Deoptimization::reason_null_assert(speculative); 1396 } else if (type == T_OBJECT || is_init_check) { 1397 reason = Deoptimization::reason_null_check(speculative); 1398 } else { 1399 reason = Deoptimization::Reason_div0_check; 1400 } 1401 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1402 // ciMethodData::has_trap_at will return a conservative -1 if any 1403 // must-be-null assertion has failed. This could cause performance 1404 // problems for a method after its first do_null_assert failure. 1405 // Consider using 'Reason_class_check' instead? 1406 1407 // To cause an implicit null check, we set the not-null probability 1408 // to the maximum (PROB_MAX). For an explicit check the probability 1409 // is set to a smaller value. 1410 if (null_control != NULL || too_many_traps(reason)) { 1411 // probability is less likely 1412 ok_prob = PROB_LIKELY_MAG(3); 1413 } else if (!assert_null && 1414 (ImplicitNullCheckThreshold > 0) && 1415 method() != NULL && 1416 (method()->method_data()->trap_count(reason) 1417 >= (uint)ImplicitNullCheckThreshold)) { 1418 ok_prob = PROB_LIKELY_MAG(3); 1419 } 1420 1421 if (null_control != NULL) { 1422 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1423 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1424 set_control( _gvn.transform( new IfTrueNode(iff))); 1425 #ifndef PRODUCT 1426 if (null_true == top()) { 1427 explicit_null_checks_elided++; 1428 } 1429 #endif 1430 (*null_control) = null_true; 1431 } else { 1432 BuildCutout unless(this, tst, ok_prob); 1433 // Check for optimizer eliding test at parse time 1434 if (stopped()) { 1435 // Failure not possible; do not bother making uncommon trap. 1436 NOT_PRODUCT(explicit_null_checks_elided++); 1437 } else if (assert_null) { 1438 uncommon_trap(reason, 1439 Deoptimization::Action_make_not_entrant, 1440 NULL, "assert_null"); 1441 } else { 1442 replace_in_map(value, zerocon(type)); 1443 builtin_throw(reason); 1444 } 1445 } 1446 1447 // Must throw exception, fall-thru not possible? 1448 if (stopped()) { 1449 return top(); // No result 1450 } 1451 1452 if (assert_null) { 1453 // Cast obj to null on this path. 1454 replace_in_map(value, zerocon(type)); 1455 return zerocon(type); 1456 } 1457 1458 // Cast obj to not-null on this path, if there is no null_control. 1459 // (If there is a null_control, a non-null value may come back to haunt us.) 1460 if (type == T_OBJECT) { 1461 Node* cast = cast_not_null(value, false); 1462 if (null_control == NULL || (*null_control) == top()) 1463 replace_in_map(value, cast); 1464 value = cast; 1465 } 1466 1467 return value; 1468 } 1469 1470 //------------------------------cast_not_null---------------------------------- 1471 // Cast obj to not-null on this path 1472 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1473 if (obj->is_InlineType()) { 1474 InlineTypeNode* vt = obj->clone()->as_InlineType(); 1475 vt->set_is_init(_gvn); 1476 vt = _gvn.transform(vt)->as_InlineType(); 1477 if (do_replace_in_map) { 1478 replace_in_map(obj, vt); 1479 } 1480 return vt; 1481 } else if (obj->is_InlineTypePtr()) { 1482 // Cast oop input instead 1483 Node* cast = cast_not_null(obj->as_InlineTypePtr()->get_oop(), do_replace_in_map); 1484 if (cast->is_top()) { 1485 // Always null 1486 return top(); 1487 } 1488 // Create a new node with the casted oop input and is_init set 1489 InlineTypeBaseNode* vt = obj->clone()->as_InlineTypePtr(); 1490 vt->set_oop(cast); 1491 vt->set_is_init(_gvn); 1492 vt = _gvn.transform(vt)->as_InlineTypePtr(); 1493 if (do_replace_in_map) { 1494 replace_in_map(obj, vt); 1495 } 1496 return vt; 1497 } 1498 const Type *t = _gvn.type(obj); 1499 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1500 // Object is already not-null? 1501 if( t == t_not_null ) return obj; 1502 1503 Node *cast = new CastPPNode(obj,t_not_null); 1504 cast->init_req(0, control()); 1505 cast = _gvn.transform( cast ); 1506 1507 // Scan for instances of 'obj' in the current JVM mapping. 1508 // These instances are known to be not-null after the test. 1509 if (do_replace_in_map) 1510 replace_in_map(obj, cast); 1511 1512 return cast; // Return casted value 1513 } 1514 1515 // Sometimes in intrinsics, we implicitly know an object is not null 1516 // (there's no actual null check) so we can cast it to not null. In 1517 // the course of optimizations, the input to the cast can become null. 1518 // In that case that data path will die and we need the control path 1519 // to become dead as well to keep the graph consistent. So we have to 1520 // add a check for null for which one branch can't be taken. It uses 1521 // an Opaque4 node that will cause the check to be removed after loop 1522 // opts so the test goes away and the compiled code doesn't execute a 1523 // useless check. 1524 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1525 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) { 1526 return value; 1527 } 1528 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1529 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1530 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1531 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1532 _gvn.set_type(iff, iff->Value(&_gvn)); 1533 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1534 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1535 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic")); 1536 C->root()->add_req(halt); 1537 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1538 set_control(if_t); 1539 return cast_not_null(value, do_replace_in_map); 1540 } 1541 1542 1543 //--------------------------replace_in_map------------------------------------- 1544 void GraphKit::replace_in_map(Node* old, Node* neww) { 1545 if (old == neww) { 1546 return; 1547 } 1548 1549 map()->replace_edge(old, neww); 1550 1551 // Note: This operation potentially replaces any edge 1552 // on the map. This includes locals, stack, and monitors 1553 // of the current (innermost) JVM state. 1554 1555 // don't let inconsistent types from profiling escape this 1556 // method 1557 1558 const Type* told = _gvn.type(old); 1559 const Type* tnew = _gvn.type(neww); 1560 1561 if (!tnew->higher_equal(told)) { 1562 return; 1563 } 1564 1565 map()->record_replaced_node(old, neww); 1566 } 1567 1568 1569 //============================================================================= 1570 //--------------------------------memory--------------------------------------- 1571 Node* GraphKit::memory(uint alias_idx) { 1572 MergeMemNode* mem = merged_memory(); 1573 Node* p = mem->memory_at(alias_idx); 1574 assert(p != mem->empty_memory(), "empty"); 1575 _gvn.set_type(p, Type::MEMORY); // must be mapped 1576 return p; 1577 } 1578 1579 //-----------------------------reset_memory------------------------------------ 1580 Node* GraphKit::reset_memory() { 1581 Node* mem = map()->memory(); 1582 // do not use this node for any more parsing! 1583 debug_only( map()->set_memory((Node*)NULL) ); 1584 return _gvn.transform( mem ); 1585 } 1586 1587 //------------------------------set_all_memory--------------------------------- 1588 void GraphKit::set_all_memory(Node* newmem) { 1589 Node* mergemem = MergeMemNode::make(newmem); 1590 gvn().set_type_bottom(mergemem); 1591 map()->set_memory(mergemem); 1592 } 1593 1594 //------------------------------set_all_memory_call---------------------------- 1595 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1596 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1597 set_all_memory(newmem); 1598 } 1599 1600 //============================================================================= 1601 // 1602 // parser factory methods for MemNodes 1603 // 1604 // These are layered on top of the factory methods in LoadNode and StoreNode, 1605 // and integrate with the parser's memory state and _gvn engine. 1606 // 1607 1608 // factory methods in "int adr_idx" 1609 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1610 int adr_idx, 1611 MemNode::MemOrd mo, 1612 LoadNode::ControlDependency control_dependency, 1613 bool require_atomic_access, 1614 bool unaligned, 1615 bool mismatched, 1616 bool unsafe, 1617 uint8_t barrier_data) { 1618 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1619 const TypePtr* adr_type = NULL; // debug-mode-only argument 1620 debug_only(adr_type = C->get_adr_type(adr_idx)); 1621 Node* mem = memory(adr_idx); 1622 Node* ld; 1623 if (require_atomic_access && bt == T_LONG) { 1624 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data); 1625 } else if (require_atomic_access && bt == T_DOUBLE) { 1626 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data); 1627 } else { 1628 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data); 1629 } 1630 ld = _gvn.transform(ld); 1631 1632 if (((bt == T_OBJECT || bt == T_PRIMITIVE_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1633 // Improve graph before escape analysis and boxing elimination. 1634 record_for_igvn(ld); 1635 } 1636 return ld; 1637 } 1638 1639 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1640 int adr_idx, 1641 MemNode::MemOrd mo, 1642 bool require_atomic_access, 1643 bool unaligned, 1644 bool mismatched, 1645 bool unsafe) { 1646 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1647 const TypePtr* adr_type = NULL; 1648 debug_only(adr_type = C->get_adr_type(adr_idx)); 1649 Node *mem = memory(adr_idx); 1650 Node* st; 1651 if (require_atomic_access && bt == T_LONG) { 1652 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1653 } else if (require_atomic_access && bt == T_DOUBLE) { 1654 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1655 } else { 1656 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1657 } 1658 if (unaligned) { 1659 st->as_Store()->set_unaligned_access(); 1660 } 1661 if (mismatched) { 1662 st->as_Store()->set_mismatched_access(); 1663 } 1664 if (unsafe) { 1665 st->as_Store()->set_unsafe_access(); 1666 } 1667 st = _gvn.transform(st); 1668 set_memory(st, adr_idx); 1669 // Back-to-back stores can only remove intermediate store with DU info 1670 // so push on worklist for optimizer. 1671 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1672 record_for_igvn(st); 1673 1674 return st; 1675 } 1676 1677 Node* GraphKit::access_store_at(Node* obj, 1678 Node* adr, 1679 const TypePtr* adr_type, 1680 Node* val, 1681 const Type* val_type, 1682 BasicType bt, 1683 DecoratorSet decorators, 1684 bool safe_for_replace) { 1685 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1686 // could be delayed during Parse (for example, in adjust_map_after_if()). 1687 // Execute transformation here to avoid barrier generation in such case. 1688 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1689 val = _gvn.makecon(TypePtr::NULL_PTR); 1690 } 1691 1692 if (stopped()) { 1693 return top(); // Dead path ? 1694 } 1695 1696 assert(val != NULL, "not dead path"); 1697 if (val->is_InlineType()) { 1698 // Store to non-flattened field. Buffer the inline type and make sure 1699 // the store is re-executed if the allocation triggers deoptimization. 1700 PreserveReexecuteState preexecs(this); 1701 jvms()->set_should_reexecute(true); 1702 val = val->as_InlineType()->buffer(this, safe_for_replace); 1703 } 1704 1705 C2AccessValuePtr addr(adr, adr_type); 1706 C2AccessValue value(val, val_type); 1707 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1708 if (access.is_raw()) { 1709 return _barrier_set->BarrierSetC2::store_at(access, value); 1710 } else { 1711 return _barrier_set->store_at(access, value); 1712 } 1713 } 1714 1715 Node* GraphKit::access_load_at(Node* obj, // containing obj 1716 Node* adr, // actual adress to store val at 1717 const TypePtr* adr_type, 1718 const Type* val_type, 1719 BasicType bt, 1720 DecoratorSet decorators, 1721 Node* ctl) { 1722 if (stopped()) { 1723 return top(); // Dead path ? 1724 } 1725 1726 C2AccessValuePtr addr(adr, adr_type); 1727 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl); 1728 if (access.is_raw()) { 1729 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1730 } else { 1731 return _barrier_set->load_at(access, val_type); 1732 } 1733 } 1734 1735 Node* GraphKit::access_load(Node* adr, // actual adress to load val at 1736 const Type* val_type, 1737 BasicType bt, 1738 DecoratorSet decorators) { 1739 if (stopped()) { 1740 return top(); // Dead path ? 1741 } 1742 1743 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr()); 1744 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr); 1745 if (access.is_raw()) { 1746 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1747 } else { 1748 return _barrier_set->load_at(access, val_type); 1749 } 1750 } 1751 1752 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj, 1753 Node* adr, 1754 const TypePtr* adr_type, 1755 int alias_idx, 1756 Node* expected_val, 1757 Node* new_val, 1758 const Type* value_type, 1759 BasicType bt, 1760 DecoratorSet decorators) { 1761 C2AccessValuePtr addr(adr, adr_type); 1762 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1763 bt, obj, addr, alias_idx); 1764 if (access.is_raw()) { 1765 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1766 } else { 1767 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1768 } 1769 } 1770 1771 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj, 1772 Node* adr, 1773 const TypePtr* adr_type, 1774 int alias_idx, 1775 Node* expected_val, 1776 Node* new_val, 1777 const Type* value_type, 1778 BasicType bt, 1779 DecoratorSet decorators) { 1780 C2AccessValuePtr addr(adr, adr_type); 1781 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1782 bt, obj, addr, alias_idx); 1783 if (access.is_raw()) { 1784 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1785 } else { 1786 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1787 } 1788 } 1789 1790 Node* GraphKit::access_atomic_xchg_at(Node* obj, 1791 Node* adr, 1792 const TypePtr* adr_type, 1793 int alias_idx, 1794 Node* new_val, 1795 const Type* value_type, 1796 BasicType bt, 1797 DecoratorSet decorators) { 1798 C2AccessValuePtr addr(adr, adr_type); 1799 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1800 bt, obj, addr, alias_idx); 1801 if (access.is_raw()) { 1802 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1803 } else { 1804 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1805 } 1806 } 1807 1808 Node* GraphKit::access_atomic_add_at(Node* obj, 1809 Node* adr, 1810 const TypePtr* adr_type, 1811 int alias_idx, 1812 Node* new_val, 1813 const Type* value_type, 1814 BasicType bt, 1815 DecoratorSet decorators) { 1816 C2AccessValuePtr addr(adr, adr_type); 1817 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1818 if (access.is_raw()) { 1819 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1820 } else { 1821 return _barrier_set->atomic_add_at(access, new_val, value_type); 1822 } 1823 } 1824 1825 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) { 1826 return _barrier_set->clone(this, src, dst, size, is_array); 1827 } 1828 1829 //-------------------------array_element_address------------------------- 1830 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1831 const TypeInt* sizetype, Node* ctrl) { 1832 uint shift = exact_log2(type2aelembytes(elembt)); 1833 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass(); 1834 if (arytype_klass != NULL && arytype_klass->is_flat_array_klass()) { 1835 ciFlatArrayKlass* vak = arytype_klass->as_flat_array_klass(); 1836 shift = vak->log2_element_size(); 1837 } 1838 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1839 1840 // short-circuit a common case (saves lots of confusing waste motion) 1841 jint idx_con = find_int_con(idx, -1); 1842 if (idx_con >= 0) { 1843 intptr_t offset = header + ((intptr_t)idx_con << shift); 1844 return basic_plus_adr(ary, offset); 1845 } 1846 1847 // must be correct type for alignment purposes 1848 Node* base = basic_plus_adr(ary, header); 1849 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1850 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1851 return basic_plus_adr(ary, base, scale); 1852 } 1853 1854 //-------------------------load_array_element------------------------- 1855 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) { 1856 const Type* elemtype = arytype->elem(); 1857 BasicType elembt = elemtype->array_element_basic_type(); 1858 assert(elembt != T_PRIMITIVE_OBJECT, "inline types are not supported by this method"); 1859 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1860 if (elembt == T_NARROWOOP) { 1861 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1862 } 1863 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt, 1864 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0)); 1865 return ld; 1866 } 1867 1868 //-------------------------set_arguments_for_java_call------------------------- 1869 // Arguments (pre-popped from the stack) are taken from the JVMS. 1870 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) { 1871 PreserveReexecuteState preexecs(this); 1872 if (EnableValhalla) { 1873 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization. 1874 // At this point, the call hasn't been executed yet, so we will only ever execute the call once. 1875 jvms()->set_should_reexecute(true); 1876 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc()); 1877 inc_sp(arg_size); 1878 } 1879 // Add the call arguments 1880 const TypeTuple* domain = call->tf()->domain_sig(); 1881 uint nargs = domain->cnt(); 1882 int arg_num = 0; 1883 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) { 1884 Node* arg = argument(i-TypeFunc::Parms); 1885 const Type* t = domain->field_at(i); 1886 if (t->is_inlinetypeptr() && call->method()->is_scalarized_arg(arg_num)) { 1887 // We don't pass inline type arguments by reference but instead pass each field of the inline type 1888 if (!arg->is_InlineTypeBase()) { 1889 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type"); 1890 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free()); 1891 } 1892 InlineTypeBaseNode* vt = arg->as_InlineTypeBase(); 1893 vt->pass_fields(this, call, idx, true, !t->maybe_null()); 1894 // If an inline type argument is passed as fields, attach the Method* to the call site 1895 // to be able to access the extended signature later via attached_method_before_pc(). 1896 // For example, see CompiledMethod::preserve_callee_argument_oops(). 1897 call->set_override_symbolic_info(true); 1898 arg_num++; 1899 continue; 1900 } else if (arg->is_InlineType()) { 1901 // Pass inline type argument via oop to callee 1902 arg = arg->as_InlineType()->buffer(this); 1903 if (!is_late_inline) { 1904 arg = arg->as_InlineTypePtr()->get_oop(); 1905 } 1906 } 1907 if (t != Type::HALF) { 1908 arg_num++; 1909 } 1910 call->init_req(idx++, arg); 1911 } 1912 } 1913 1914 //---------------------------set_edges_for_java_call--------------------------- 1915 // Connect a newly created call into the current JVMS. 1916 // A return value node (if any) is returned from set_edges_for_java_call. 1917 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1918 1919 // Add the predefined inputs: 1920 call->init_req( TypeFunc::Control, control() ); 1921 call->init_req( TypeFunc::I_O , i_o() ); 1922 call->init_req( TypeFunc::Memory , reset_memory() ); 1923 call->init_req( TypeFunc::FramePtr, frameptr() ); 1924 call->init_req( TypeFunc::ReturnAdr, top() ); 1925 1926 add_safepoint_edges(call, must_throw); 1927 1928 Node* xcall = _gvn.transform(call); 1929 1930 if (xcall == top()) { 1931 set_control(top()); 1932 return; 1933 } 1934 assert(xcall == call, "call identity is stable"); 1935 1936 // Re-use the current map to produce the result. 1937 1938 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1939 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1940 set_all_memory_call(xcall, separate_io_proj); 1941 1942 //return xcall; // no need, caller already has it 1943 } 1944 1945 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) { 1946 if (stopped()) return top(); // maybe the call folded up? 1947 1948 // Note: Since any out-of-line call can produce an exception, 1949 // we always insert an I_O projection from the call into the result. 1950 1951 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize); 1952 1953 if (separate_io_proj) { 1954 // The caller requested separate projections be used by the fall 1955 // through and exceptional paths, so replace the projections for 1956 // the fall through path. 1957 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1958 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1959 } 1960 1961 // Capture the return value, if any. 1962 Node* ret; 1963 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) { 1964 ret = top(); 1965 } else if (call->tf()->returns_inline_type_as_fields()) { 1966 // Return of multiple values (inline type fields): we create a 1967 // InlineType node, each field is a projection from the call. 1968 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass(); 1969 uint base_input = TypeFunc::Parms; 1970 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, call->method()->signature()->returns_null_free_inline_type()); 1971 } else { 1972 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1973 } 1974 1975 return ret; 1976 } 1977 1978 //--------------------set_predefined_input_for_runtime_call-------------------- 1979 // Reading and setting the memory state is way conservative here. 1980 // The real problem is that I am not doing real Type analysis on memory, 1981 // so I cannot distinguish card mark stores from other stores. Across a GC 1982 // point the Store Barrier and the card mark memory has to agree. I cannot 1983 // have a card mark store and its barrier split across the GC point from 1984 // either above or below. Here I get that to happen by reading ALL of memory. 1985 // A better answer would be to separate out card marks from other memory. 1986 // For now, return the input memory state, so that it can be reused 1987 // after the call, if this call has restricted memory effects. 1988 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) { 1989 // Set fixed predefined input arguments 1990 Node* memory = reset_memory(); 1991 Node* m = narrow_mem == NULL ? memory : narrow_mem; 1992 call->init_req( TypeFunc::Control, control() ); 1993 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1994 call->init_req( TypeFunc::Memory, m ); // may gc ptrs 1995 call->init_req( TypeFunc::FramePtr, frameptr() ); 1996 call->init_req( TypeFunc::ReturnAdr, top() ); 1997 return memory; 1998 } 1999 2000 //-------------------set_predefined_output_for_runtime_call-------------------- 2001 // Set control and memory (not i_o) from the call. 2002 // If keep_mem is not NULL, use it for the output state, 2003 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 2004 // If hook_mem is NULL, this call produces no memory effects at all. 2005 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 2006 // then only that memory slice is taken from the call. 2007 // In the last case, we must put an appropriate memory barrier before 2008 // the call, so as to create the correct anti-dependencies on loads 2009 // preceding the call. 2010 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 2011 Node* keep_mem, 2012 const TypePtr* hook_mem) { 2013 // no i/o 2014 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 2015 if (keep_mem) { 2016 // First clone the existing memory state 2017 set_all_memory(keep_mem); 2018 if (hook_mem != NULL) { 2019 // Make memory for the call 2020 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 2021 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 2022 // We also use hook_mem to extract specific effects from arraycopy stubs. 2023 set_memory(mem, hook_mem); 2024 } 2025 // ...else the call has NO memory effects. 2026 2027 // Make sure the call advertises its memory effects precisely. 2028 // This lets us build accurate anti-dependences in gcm.cpp. 2029 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 2030 "call node must be constructed correctly"); 2031 } else { 2032 assert(hook_mem == NULL, ""); 2033 // This is not a "slow path" call; all memory comes from the call. 2034 set_all_memory_call(call); 2035 } 2036 } 2037 2038 // Keep track of MergeMems feeding into other MergeMems 2039 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) { 2040 if (!mem->is_MergeMem()) { 2041 return; 2042 } 2043 for (SimpleDUIterator i(mem); i.has_next(); i.next()) { 2044 Node* use = i.get(); 2045 if (use->is_MergeMem()) { 2046 wl.push(use); 2047 } 2048 } 2049 } 2050 2051 // Replace the call with the current state of the kit. 2052 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 2053 JVMState* ejvms = NULL; 2054 if (has_exceptions()) { 2055 ejvms = transfer_exceptions_into_jvms(); 2056 } 2057 2058 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 2059 ReplacedNodes replaced_nodes_exception; 2060 Node* ex_ctl = top(); 2061 2062 SafePointNode* final_state = stop(); 2063 2064 // Find all the needed outputs of this call 2065 CallProjections* callprojs = call->extract_projections(true); 2066 2067 Unique_Node_List wl; 2068 Node* init_mem = call->in(TypeFunc::Memory); 2069 Node* final_mem = final_state->in(TypeFunc::Memory); 2070 Node* final_ctl = final_state->in(TypeFunc::Control); 2071 Node* final_io = final_state->in(TypeFunc::I_O); 2072 2073 // Replace all the old call edges with the edges from the inlining result 2074 if (callprojs->fallthrough_catchproj != NULL) { 2075 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl); 2076 } 2077 if (callprojs->fallthrough_memproj != NULL) { 2078 if (final_mem->is_MergeMem()) { 2079 // Parser's exits MergeMem was not transformed but may be optimized 2080 final_mem = _gvn.transform(final_mem); 2081 } 2082 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem); 2083 add_mergemem_users_to_worklist(wl, final_mem); 2084 } 2085 if (callprojs->fallthrough_ioproj != NULL) { 2086 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io); 2087 } 2088 2089 // Replace the result with the new result if it exists and is used 2090 if (callprojs->resproj[0] != NULL && result != NULL) { 2091 // If the inlined code is dead, the result projections for an inline type returned as 2092 // fields have not been replaced. They will go away once the call is replaced by TOP below. 2093 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()), 2094 "unexpected number of results"); 2095 C->gvn_replace_by(callprojs->resproj[0], result); 2096 } 2097 2098 if (ejvms == NULL) { 2099 // No exception edges to simply kill off those paths 2100 if (callprojs->catchall_catchproj != NULL) { 2101 C->gvn_replace_by(callprojs->catchall_catchproj, C->top()); 2102 } 2103 if (callprojs->catchall_memproj != NULL) { 2104 C->gvn_replace_by(callprojs->catchall_memproj, C->top()); 2105 } 2106 if (callprojs->catchall_ioproj != NULL) { 2107 C->gvn_replace_by(callprojs->catchall_ioproj, C->top()); 2108 } 2109 // Replace the old exception object with top 2110 if (callprojs->exobj != NULL) { 2111 C->gvn_replace_by(callprojs->exobj, C->top()); 2112 } 2113 } else { 2114 GraphKit ekit(ejvms); 2115 2116 // Load my combined exception state into the kit, with all phis transformed: 2117 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 2118 replaced_nodes_exception = ex_map->replaced_nodes(); 2119 2120 Node* ex_oop = ekit.use_exception_state(ex_map); 2121 2122 if (callprojs->catchall_catchproj != NULL) { 2123 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control()); 2124 ex_ctl = ekit.control(); 2125 } 2126 if (callprojs->catchall_memproj != NULL) { 2127 Node* ex_mem = ekit.reset_memory(); 2128 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem); 2129 add_mergemem_users_to_worklist(wl, ex_mem); 2130 } 2131 if (callprojs->catchall_ioproj != NULL) { 2132 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o()); 2133 } 2134 2135 // Replace the old exception object with the newly created one 2136 if (callprojs->exobj != NULL) { 2137 C->gvn_replace_by(callprojs->exobj, ex_oop); 2138 } 2139 } 2140 2141 // Disconnect the call from the graph 2142 call->disconnect_inputs(C); 2143 C->gvn_replace_by(call, C->top()); 2144 2145 // Clean up any MergeMems that feed other MergeMems since the 2146 // optimizer doesn't like that. 2147 while (wl.size() > 0) { 2148 _gvn.transform(wl.pop()); 2149 } 2150 2151 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 2152 replaced_nodes.apply(C, final_ctl); 2153 } 2154 if (!ex_ctl->is_top() && do_replaced_nodes) { 2155 replaced_nodes_exception.apply(C, ex_ctl); 2156 } 2157 } 2158 2159 2160 //------------------------------increment_counter------------------------------ 2161 // for statistics: increment a VM counter by 1 2162 2163 void GraphKit::increment_counter(address counter_addr) { 2164 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 2165 increment_counter(adr1); 2166 } 2167 2168 void GraphKit::increment_counter(Node* counter_addr) { 2169 int adr_type = Compile::AliasIdxRaw; 2170 Node* ctrl = control(); 2171 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered); 2172 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1))); 2173 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered); 2174 } 2175 2176 2177 //------------------------------uncommon_trap---------------------------------- 2178 // Bail out to the interpreter in mid-method. Implemented by calling the 2179 // uncommon_trap blob. This helper function inserts a runtime call with the 2180 // right debug info. 2181 void GraphKit::uncommon_trap(int trap_request, 2182 ciKlass* klass, const char* comment, 2183 bool must_throw, 2184 bool keep_exact_action) { 2185 if (failing()) stop(); 2186 if (stopped()) return; // trap reachable? 2187 2188 // Note: If ProfileTraps is true, and if a deopt. actually 2189 // occurs here, the runtime will make sure an MDO exists. There is 2190 // no need to call method()->ensure_method_data() at this point. 2191 2192 // Set the stack pointer to the right value for reexecution: 2193 set_sp(reexecute_sp()); 2194 2195 #ifdef ASSERT 2196 if (!must_throw) { 2197 // Make sure the stack has at least enough depth to execute 2198 // the current bytecode. 2199 int inputs, ignored_depth; 2200 if (compute_stack_effects(inputs, ignored_depth)) { 2201 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2202 Bytecodes::name(java_bc()), sp(), inputs); 2203 } 2204 } 2205 #endif 2206 2207 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2208 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2209 2210 switch (action) { 2211 case Deoptimization::Action_maybe_recompile: 2212 case Deoptimization::Action_reinterpret: 2213 // Temporary fix for 6529811 to allow virtual calls to be sure they 2214 // get the chance to go from mono->bi->mega 2215 if (!keep_exact_action && 2216 Deoptimization::trap_request_index(trap_request) < 0 && 2217 too_many_recompiles(reason)) { 2218 // This BCI is causing too many recompilations. 2219 if (C->log() != NULL) { 2220 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2221 Deoptimization::trap_reason_name(reason), 2222 Deoptimization::trap_action_name(action)); 2223 } 2224 action = Deoptimization::Action_none; 2225 trap_request = Deoptimization::make_trap_request(reason, action); 2226 } else { 2227 C->set_trap_can_recompile(true); 2228 } 2229 break; 2230 case Deoptimization::Action_make_not_entrant: 2231 C->set_trap_can_recompile(true); 2232 break; 2233 case Deoptimization::Action_none: 2234 case Deoptimization::Action_make_not_compilable: 2235 break; 2236 default: 2237 #ifdef ASSERT 2238 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2239 #endif 2240 break; 2241 } 2242 2243 if (TraceOptoParse) { 2244 char buf[100]; 2245 tty->print_cr("Uncommon trap %s at bci:%d", 2246 Deoptimization::format_trap_request(buf, sizeof(buf), 2247 trap_request), bci()); 2248 } 2249 2250 CompileLog* log = C->log(); 2251 if (log != NULL) { 2252 int kid = (klass == NULL)? -1: log->identify(klass); 2253 log->begin_elem("uncommon_trap bci='%d'", bci()); 2254 char buf[100]; 2255 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2256 trap_request)); 2257 if (kid >= 0) log->print(" klass='%d'", kid); 2258 if (comment != NULL) log->print(" comment='%s'", comment); 2259 log->end_elem(); 2260 } 2261 2262 // Make sure any guarding test views this path as very unlikely 2263 Node *i0 = control()->in(0); 2264 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2265 IfNode *iff = i0->as_If(); 2266 float f = iff->_prob; // Get prob 2267 if (control()->Opcode() == Op_IfTrue) { 2268 if (f > PROB_UNLIKELY_MAG(4)) 2269 iff->_prob = PROB_MIN; 2270 } else { 2271 if (f < PROB_LIKELY_MAG(4)) 2272 iff->_prob = PROB_MAX; 2273 } 2274 } 2275 2276 // Clear out dead values from the debug info. 2277 kill_dead_locals(); 2278 2279 // Now insert the uncommon trap subroutine call 2280 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2281 const TypePtr* no_memory_effects = NULL; 2282 // Pass the index of the class to be loaded 2283 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2284 (must_throw ? RC_MUST_THROW : 0), 2285 OptoRuntime::uncommon_trap_Type(), 2286 call_addr, "uncommon_trap", no_memory_effects, 2287 intcon(trap_request)); 2288 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2289 "must extract request correctly from the graph"); 2290 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2291 2292 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2293 // The debug info is the only real input to this call. 2294 2295 // Halt-and-catch fire here. The above call should never return! 2296 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen" 2297 PRODUCT_ONLY(COMMA /*reachable*/false)); 2298 _gvn.set_type_bottom(halt); 2299 root()->add_req(halt); 2300 2301 stop_and_kill_map(); 2302 } 2303 2304 2305 //--------------------------just_allocated_object------------------------------ 2306 // Report the object that was just allocated. 2307 // It must be the case that there are no intervening safepoints. 2308 // We use this to determine if an object is so "fresh" that 2309 // it does not require card marks. 2310 Node* GraphKit::just_allocated_object(Node* current_control) { 2311 Node* ctrl = current_control; 2312 // Object::<init> is invoked after allocation, most of invoke nodes 2313 // will be reduced, but a region node is kept in parse time, we check 2314 // the pattern and skip the region node if it degraded to a copy. 2315 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 && 2316 ctrl->as_Region()->is_copy()) { 2317 ctrl = ctrl->as_Region()->is_copy(); 2318 } 2319 if (C->recent_alloc_ctl() == ctrl) { 2320 return C->recent_alloc_obj(); 2321 } 2322 return NULL; 2323 } 2324 2325 2326 /** 2327 * Record profiling data exact_kls for Node n with the type system so 2328 * that it can propagate it (speculation) 2329 * 2330 * @param n node that the type applies to 2331 * @param exact_kls type from profiling 2332 * @param maybe_null did profiling see null? 2333 * 2334 * @return node with improved type 2335 */ 2336 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2337 const Type* current_type = _gvn.type(n); 2338 assert(UseTypeSpeculation, "type speculation must be on"); 2339 2340 const TypePtr* speculative = current_type->speculative(); 2341 2342 // Should the klass from the profile be recorded in the speculative type? 2343 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2344 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2345 const TypeOopPtr* xtype = tklass->as_instance_type(); 2346 assert(xtype->klass_is_exact(), "Should be exact"); 2347 // Any reason to believe n is not null (from this profiling or a previous one)? 2348 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2349 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2350 // record the new speculative type's depth 2351 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2352 speculative = speculative->with_inline_depth(jvms()->depth()); 2353 } else if (current_type->would_improve_ptr(ptr_kind)) { 2354 // Profiling report that null was never seen so we can change the 2355 // speculative type to non null ptr. 2356 if (ptr_kind == ProfileAlwaysNull) { 2357 speculative = TypePtr::NULL_PTR; 2358 } else { 2359 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2360 const TypePtr* ptr = TypePtr::NOTNULL; 2361 if (speculative != NULL) { 2362 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2363 } else { 2364 speculative = ptr; 2365 } 2366 } 2367 } 2368 2369 if (speculative != current_type->speculative()) { 2370 // Build a type with a speculative type (what we think we know 2371 // about the type but will need a guard when we use it) 2372 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative); 2373 // We're changing the type, we need a new CheckCast node to carry 2374 // the new type. The new type depends on the control: what 2375 // profiling tells us is only valid from here as far as we can 2376 // tell. 2377 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2378 cast = _gvn.transform(cast); 2379 replace_in_map(n, cast); 2380 n = cast; 2381 } 2382 2383 return n; 2384 } 2385 2386 /** 2387 * Record profiling data from receiver profiling at an invoke with the 2388 * type system so that it can propagate it (speculation) 2389 * 2390 * @param n receiver node 2391 * 2392 * @return node with improved type 2393 */ 2394 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2395 if (!UseTypeSpeculation) { 2396 return n; 2397 } 2398 ciKlass* exact_kls = profile_has_unique_klass(); 2399 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2400 if ((java_bc() == Bytecodes::_checkcast || 2401 java_bc() == Bytecodes::_instanceof || 2402 java_bc() == Bytecodes::_aastore) && 2403 method()->method_data()->is_mature()) { 2404 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2405 if (data != NULL) { 2406 if (java_bc() == Bytecodes::_aastore) { 2407 ciKlass* array_type = NULL; 2408 ciKlass* element_type = NULL; 2409 ProfilePtrKind element_ptr = ProfileMaybeNull; 2410 bool flat_array = true; 2411 bool null_free_array = true; 2412 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 2413 exact_kls = element_type; 2414 ptr_kind = element_ptr; 2415 } else { 2416 if (!data->as_BitData()->null_seen()) { 2417 ptr_kind = ProfileNeverNull; 2418 } else { 2419 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2420 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2421 uint i = 0; 2422 for (; i < call->row_limit(); i++) { 2423 ciKlass* receiver = call->receiver(i); 2424 if (receiver != NULL) { 2425 break; 2426 } 2427 } 2428 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2429 } 2430 } 2431 } 2432 } 2433 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2434 } 2435 2436 /** 2437 * Record profiling data from argument profiling at an invoke with the 2438 * type system so that it can propagate it (speculation) 2439 * 2440 * @param dest_method target method for the call 2441 * @param bc what invoke bytecode is this? 2442 */ 2443 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2444 if (!UseTypeSpeculation) { 2445 return; 2446 } 2447 const TypeFunc* tf = TypeFunc::make(dest_method); 2448 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2449 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2450 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2451 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2452 if (is_reference_type(targ->basic_type())) { 2453 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2454 ciKlass* better_type = NULL; 2455 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2456 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2457 } 2458 i++; 2459 } 2460 } 2461 } 2462 2463 /** 2464 * Record profiling data from parameter profiling at an invoke with 2465 * the type system so that it can propagate it (speculation) 2466 */ 2467 void GraphKit::record_profiled_parameters_for_speculation() { 2468 if (!UseTypeSpeculation) { 2469 return; 2470 } 2471 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2472 if (_gvn.type(local(i))->isa_oopptr()) { 2473 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2474 ciKlass* better_type = NULL; 2475 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2476 record_profile_for_speculation(local(i), better_type, ptr_kind); 2477 } 2478 j++; 2479 } 2480 } 2481 } 2482 2483 /** 2484 * Record profiling data from return value profiling at an invoke with 2485 * the type system so that it can propagate it (speculation) 2486 */ 2487 void GraphKit::record_profiled_return_for_speculation() { 2488 if (!UseTypeSpeculation) { 2489 return; 2490 } 2491 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2492 ciKlass* better_type = NULL; 2493 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2494 // If profiling reports a single type for the return value, 2495 // feed it to the type system so it can propagate it as a 2496 // speculative type 2497 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2498 } 2499 } 2500 2501 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2502 if (Matcher::strict_fp_requires_explicit_rounding) { 2503 // (Note: TypeFunc::make has a cache that makes this fast.) 2504 const TypeFunc* tf = TypeFunc::make(dest_method); 2505 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms; 2506 for (int j = 0; j < nargs; j++) { 2507 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms); 2508 if (targ->basic_type() == T_DOUBLE) { 2509 // If any parameters are doubles, they must be rounded before 2510 // the call, dprecision_rounding does gvn.transform 2511 Node *arg = argument(j); 2512 arg = dprecision_rounding(arg); 2513 set_argument(j, arg); 2514 } 2515 } 2516 } 2517 } 2518 2519 // rounding for strict float precision conformance 2520 Node* GraphKit::precision_rounding(Node* n) { 2521 if (Matcher::strict_fp_requires_explicit_rounding) { 2522 #ifdef IA32 2523 if (UseSSE == 0) { 2524 return _gvn.transform(new RoundFloatNode(0, n)); 2525 } 2526 #else 2527 Unimplemented(); 2528 #endif // IA32 2529 } 2530 return n; 2531 } 2532 2533 // rounding for strict double precision conformance 2534 Node* GraphKit::dprecision_rounding(Node *n) { 2535 if (Matcher::strict_fp_requires_explicit_rounding) { 2536 #ifdef IA32 2537 if (UseSSE < 2) { 2538 return _gvn.transform(new RoundDoubleNode(0, n)); 2539 } 2540 #else 2541 Unimplemented(); 2542 #endif // IA32 2543 } 2544 return n; 2545 } 2546 2547 //============================================================================= 2548 // Generate a fast path/slow path idiom. Graph looks like: 2549 // [foo] indicates that 'foo' is a parameter 2550 // 2551 // [in] NULL 2552 // \ / 2553 // CmpP 2554 // Bool ne 2555 // If 2556 // / \ 2557 // True False-<2> 2558 // / | 2559 // / cast_not_null 2560 // Load | | ^ 2561 // [fast_test] | | 2562 // gvn to opt_test | | 2563 // / \ | <1> 2564 // True False | 2565 // | \\ | 2566 // [slow_call] \[fast_result] 2567 // Ctl Val \ \ 2568 // | \ \ 2569 // Catch <1> \ \ 2570 // / \ ^ \ \ 2571 // Ex No_Ex | \ \ 2572 // | \ \ | \ <2> \ 2573 // ... \ [slow_res] | | \ [null_result] 2574 // \ \--+--+--- | | 2575 // \ | / \ | / 2576 // --------Region Phi 2577 // 2578 //============================================================================= 2579 // Code is structured as a series of driver functions all called 'do_XXX' that 2580 // call a set of helper functions. Helper functions first, then drivers. 2581 2582 //------------------------------null_check_oop--------------------------------- 2583 // Null check oop. Set null-path control into Region in slot 3. 2584 // Make a cast-not-nullness use the other not-null control. Return cast. 2585 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2586 bool never_see_null, 2587 bool safe_for_replace, 2588 bool speculative) { 2589 // Initial NULL check taken path 2590 (*null_control) = top(); 2591 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2592 2593 // Generate uncommon_trap: 2594 if (never_see_null && (*null_control) != top()) { 2595 // If we see an unexpected null at a check-cast we record it and force a 2596 // recompile; the offending check-cast will be compiled to handle NULLs. 2597 // If we see more than one offending BCI, then all checkcasts in the 2598 // method will be compiled to handle NULLs. 2599 PreserveJVMState pjvms(this); 2600 set_control(*null_control); 2601 replace_in_map(value, null()); 2602 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2603 uncommon_trap(reason, 2604 Deoptimization::Action_make_not_entrant); 2605 (*null_control) = top(); // NULL path is dead 2606 } 2607 if ((*null_control) == top() && safe_for_replace) { 2608 replace_in_map(value, cast); 2609 } 2610 2611 // Cast away null-ness on the result 2612 return cast; 2613 } 2614 2615 //------------------------------opt_iff---------------------------------------- 2616 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2617 // Return slow-path control. 2618 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2619 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2620 2621 // Fast path taken; set region slot 2 2622 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2623 region->init_req(2,fast_taken); // Capture fast-control 2624 2625 // Fast path not-taken, i.e. slow path 2626 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2627 return slow_taken; 2628 } 2629 2630 //-----------------------------make_runtime_call------------------------------- 2631 Node* GraphKit::make_runtime_call(int flags, 2632 const TypeFunc* call_type, address call_addr, 2633 const char* call_name, 2634 const TypePtr* adr_type, 2635 // The following parms are all optional. 2636 // The first NULL ends the list. 2637 Node* parm0, Node* parm1, 2638 Node* parm2, Node* parm3, 2639 Node* parm4, Node* parm5, 2640 Node* parm6, Node* parm7) { 2641 assert(call_addr != NULL, "must not call NULL targets"); 2642 2643 // Slow-path call 2644 bool is_leaf = !(flags & RC_NO_LEAF); 2645 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2646 if (call_name == NULL) { 2647 assert(!is_leaf, "must supply name for leaf"); 2648 call_name = OptoRuntime::stub_name(call_addr); 2649 } 2650 CallNode* call; 2651 if (!is_leaf) { 2652 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type); 2653 } else if (flags & RC_NO_FP) { 2654 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2655 } else if (flags & RC_VECTOR){ 2656 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte; 2657 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits); 2658 } else { 2659 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2660 } 2661 2662 // The following is similar to set_edges_for_java_call, 2663 // except that the memory effects of the call are restricted to AliasIdxRaw. 2664 2665 // Slow path call has no side-effects, uses few values 2666 bool wide_in = !(flags & RC_NARROW_MEM); 2667 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2668 2669 Node* prev_mem = NULL; 2670 if (wide_in) { 2671 prev_mem = set_predefined_input_for_runtime_call(call); 2672 } else { 2673 assert(!wide_out, "narrow in => narrow out"); 2674 Node* narrow_mem = memory(adr_type); 2675 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem); 2676 } 2677 2678 // Hook each parm in order. Stop looking at the first NULL. 2679 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2680 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2681 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2682 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2683 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2684 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2685 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2686 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2687 /* close each nested if ===> */ } } } } } } } } 2688 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2689 2690 if (!is_leaf) { 2691 // Non-leaves can block and take safepoints: 2692 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2693 } 2694 // Non-leaves can throw exceptions: 2695 if (has_io) { 2696 call->set_req(TypeFunc::I_O, i_o()); 2697 } 2698 2699 if (flags & RC_UNCOMMON) { 2700 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2701 // (An "if" probability corresponds roughly to an unconditional count. 2702 // Sort of.) 2703 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2704 } 2705 2706 Node* c = _gvn.transform(call); 2707 assert(c == call, "cannot disappear"); 2708 2709 if (wide_out) { 2710 // Slow path call has full side-effects. 2711 set_predefined_output_for_runtime_call(call); 2712 } else { 2713 // Slow path call has few side-effects, and/or sets few values. 2714 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2715 } 2716 2717 if (has_io) { 2718 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2719 } 2720 return call; 2721 2722 } 2723 2724 // i2b 2725 Node* GraphKit::sign_extend_byte(Node* in) { 2726 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24))); 2727 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24))); 2728 } 2729 2730 // i2s 2731 Node* GraphKit::sign_extend_short(Node* in) { 2732 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16))); 2733 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16))); 2734 } 2735 2736 //-----------------------------make_native_call------------------------------- 2737 Node* GraphKit::make_native_call(address call_addr, const TypeFunc* call_type, uint nargs, ciNativeEntryPoint* nep) { 2738 // Select just the actual call args to pass on 2739 // [MethodHandle fallback, long addr, HALF addr, ... args , NativeEntryPoint nep] 2740 // | | 2741 // V V 2742 // [ ... args ] 2743 uint n_filtered_args = nargs - 4; // -fallback, -addr (2), -nep; 2744 ResourceMark rm; 2745 Node** argument_nodes = NEW_RESOURCE_ARRAY(Node*, n_filtered_args); 2746 const Type** arg_types = TypeTuple::fields(n_filtered_args); 2747 GrowableArray<VMReg> arg_regs(C->comp_arena(), n_filtered_args, n_filtered_args, VMRegImpl::Bad()); 2748 2749 VMReg* argRegs = nep->argMoves(); 2750 { 2751 for (uint vm_arg_pos = 0, java_arg_read_pos = 0; 2752 vm_arg_pos < n_filtered_args; vm_arg_pos++) { 2753 uint vm_unfiltered_arg_pos = vm_arg_pos + 3; // +3 to skip fallback handle argument and addr (2 since long) 2754 Node* node = argument(vm_unfiltered_arg_pos); 2755 const Type* type = call_type->domain_sig()->field_at(TypeFunc::Parms + vm_unfiltered_arg_pos); 2756 VMReg reg = type == Type::HALF 2757 ? VMRegImpl::Bad() 2758 : argRegs[java_arg_read_pos++]; 2759 2760 argument_nodes[vm_arg_pos] = node; 2761 arg_types[TypeFunc::Parms + vm_arg_pos] = type; 2762 arg_regs.at_put(vm_arg_pos, reg); 2763 } 2764 } 2765 2766 uint n_returns = call_type->range_sig()->cnt() - TypeFunc::Parms; 2767 GrowableArray<VMReg> ret_regs(C->comp_arena(), n_returns, n_returns, VMRegImpl::Bad()); 2768 const Type** ret_types = TypeTuple::fields(n_returns); 2769 2770 VMReg* retRegs = nep->returnMoves(); 2771 { 2772 for (uint vm_ret_pos = 0, java_ret_read_pos = 0; 2773 vm_ret_pos < n_returns; vm_ret_pos++) { // 0 or 1 2774 const Type* type = call_type->range_sig()->field_at(TypeFunc::Parms + vm_ret_pos); 2775 VMReg reg = type == Type::HALF 2776 ? VMRegImpl::Bad() 2777 : retRegs[java_ret_read_pos++]; 2778 2779 ret_regs.at_put(vm_ret_pos, reg); 2780 ret_types[TypeFunc::Parms + vm_ret_pos] = type; 2781 } 2782 } 2783 2784 const TypeFunc* new_call_type = TypeFunc::make( 2785 TypeTuple::make(TypeFunc::Parms + n_filtered_args, arg_types), 2786 TypeTuple::make(TypeFunc::Parms + n_returns, ret_types) 2787 ); 2788 2789 if (nep->need_transition()) { 2790 RuntimeStub* invoker = SharedRuntime::make_native_invoker(call_addr, 2791 nep->shadow_space(), 2792 arg_regs, ret_regs); 2793 if (invoker == NULL) { 2794 C->record_failure("native invoker not implemented on this platform"); 2795 return NULL; 2796 } 2797 C->add_native_invoker(invoker); 2798 call_addr = invoker->code_begin(); 2799 } 2800 assert(call_addr != NULL, "sanity"); 2801 2802 CallNativeNode* call = new CallNativeNode(new_call_type, call_addr, nep->name(), TypePtr::BOTTOM, 2803 arg_regs, 2804 ret_regs, 2805 nep->shadow_space(), 2806 nep->need_transition()); 2807 2808 if (call->_need_transition) { 2809 add_safepoint_edges(call); 2810 } 2811 2812 set_predefined_input_for_runtime_call(call); 2813 2814 for (uint i = 0; i < n_filtered_args; i++) { 2815 call->init_req(i + TypeFunc::Parms, argument_nodes[i]); 2816 } 2817 2818 Node* c = gvn().transform(call); 2819 assert(c == call, "cannot disappear"); 2820 2821 set_predefined_output_for_runtime_call(call); 2822 2823 Node* ret; 2824 if (method() == NULL || method()->return_type()->basic_type() == T_VOID) { 2825 ret = top(); 2826 } else { 2827 ret = gvn().transform(new ProjNode(call, TypeFunc::Parms)); 2828 // Unpack native results if needed 2829 // Need this method type since it's unerased 2830 switch (nep->method_type()->rtype()->basic_type()) { 2831 case T_CHAR: 2832 ret = _gvn.transform(new AndINode(ret, _gvn.intcon(0xFFFF))); 2833 break; 2834 case T_BYTE: 2835 ret = sign_extend_byte(ret); 2836 break; 2837 case T_SHORT: 2838 ret = sign_extend_short(ret); 2839 break; 2840 default: // do nothing 2841 break; 2842 } 2843 } 2844 2845 push_node(method()->return_type()->basic_type(), ret); 2846 2847 return call; 2848 } 2849 2850 //------------------------------merge_memory----------------------------------- 2851 // Merge memory from one path into the current memory state. 2852 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2853 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2854 Node* old_slice = mms.force_memory(); 2855 Node* new_slice = mms.memory2(); 2856 if (old_slice != new_slice) { 2857 PhiNode* phi; 2858 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2859 if (mms.is_empty()) { 2860 // clone base memory Phi's inputs for this memory slice 2861 assert(old_slice == mms.base_memory(), "sanity"); 2862 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2863 _gvn.set_type(phi, Type::MEMORY); 2864 for (uint i = 1; i < phi->req(); i++) { 2865 phi->init_req(i, old_slice->in(i)); 2866 } 2867 } else { 2868 phi = old_slice->as_Phi(); // Phi was generated already 2869 } 2870 } else { 2871 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2872 _gvn.set_type(phi, Type::MEMORY); 2873 } 2874 phi->set_req(new_path, new_slice); 2875 mms.set_memory(phi); 2876 } 2877 } 2878 } 2879 2880 //------------------------------make_slow_call_ex------------------------------ 2881 // Make the exception handler hookups for the slow call 2882 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2883 if (stopped()) return; 2884 2885 // Make a catch node with just two handlers: fall-through and catch-all 2886 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2887 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2888 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2889 _gvn.set_type_bottom(norm); 2890 C->record_for_igvn(norm); 2891 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2892 2893 { PreserveJVMState pjvms(this); 2894 set_control(excp); 2895 set_i_o(i_o); 2896 2897 if (excp != top()) { 2898 if (deoptimize) { 2899 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2900 uncommon_trap(Deoptimization::Reason_unhandled, 2901 Deoptimization::Action_none); 2902 } else { 2903 // Create an exception state also. 2904 // Use an exact type if the caller has a specific exception. 2905 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2906 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2907 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2908 } 2909 } 2910 } 2911 2912 // Get the no-exception control from the CatchNode. 2913 set_control(norm); 2914 } 2915 2916 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) { 2917 Node* cmp = NULL; 2918 switch(bt) { 2919 case T_INT: cmp = new CmpINode(in1, in2); break; 2920 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2921 default: fatal("unexpected comparison type %s", type2name(bt)); 2922 } 2923 gvn.transform(cmp); 2924 Node* bol = gvn.transform(new BoolNode(cmp, test)); 2925 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2926 gvn.transform(iff); 2927 if (!bol->is_Con()) gvn.record_for_igvn(iff); 2928 return iff; 2929 } 2930 2931 //-------------------------------gen_subtype_check----------------------------- 2932 // Generate a subtyping check. Takes as input the subtype and supertype. 2933 // Returns 2 values: sets the default control() to the true path and returns 2934 // the false path. Only reads invariant memory; sets no (visible) memory. 2935 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2936 // but that's not exposed to the optimizer. This call also doesn't take in an 2937 // Object; if you wish to check an Object you need to load the Object's class 2938 // prior to coming here. 2939 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) { 2940 Compile* C = gvn.C; 2941 if ((*ctrl)->is_top()) { 2942 return C->top(); 2943 } 2944 2945 // Fast check for identical types, perhaps identical constants. 2946 // The types can even be identical non-constants, in cases 2947 // involving Array.newInstance, Object.clone, etc. 2948 if (subklass == superklass) 2949 return C->top(); // false path is dead; no test needed. 2950 2951 if (gvn.type(superklass)->singleton()) { 2952 ciKlass* superk = gvn.type(superklass)->is_klassptr()->klass(); 2953 ciKlass* subk = gvn.type(subklass)->is_klassptr()->klass(); 2954 2955 // In the common case of an exact superklass, try to fold up the 2956 // test before generating code. You may ask, why not just generate 2957 // the code and then let it fold up? The answer is that the generated 2958 // code will necessarily include null checks, which do not always 2959 // completely fold away. If they are also needless, then they turn 2960 // into a performance loss. Example: 2961 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2962 // Here, the type of 'fa' is often exact, so the store check 2963 // of fa[1]=x will fold up, without testing the nullness of x. 2964 switch (C->static_subtype_check(superk, subk)) { 2965 case Compile::SSC_always_false: 2966 { 2967 Node* always_fail = *ctrl; 2968 *ctrl = gvn.C->top(); 2969 return always_fail; 2970 } 2971 case Compile::SSC_always_true: 2972 return C->top(); 2973 case Compile::SSC_easy_test: 2974 { 2975 // Just do a direct pointer compare and be done. 2976 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2977 *ctrl = gvn.transform(new IfTrueNode(iff)); 2978 return gvn.transform(new IfFalseNode(iff)); 2979 } 2980 case Compile::SSC_full_test: 2981 break; 2982 default: 2983 ShouldNotReachHere(); 2984 } 2985 } 2986 2987 // %%% Possible further optimization: Even if the superklass is not exact, 2988 // if the subklass is the unique subtype of the superklass, the check 2989 // will always succeed. We could leave a dependency behind to ensure this. 2990 2991 // First load the super-klass's check-offset 2992 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2993 Node* m = C->immutable_memory(); 2994 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2995 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2996 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2997 2998 // Load from the sub-klass's super-class display list, or a 1-word cache of 2999 // the secondary superclass list, or a failing value with a sentinel offset 3000 // if the super-klass is an interface or exceptionally deep in the Java 3001 // hierarchy and we have to scan the secondary superclass list the hard way. 3002 // Worst-case type is a little odd: NULL is allowed as a result (usually 3003 // klass loads can never produce a NULL). 3004 Node *chk_off_X = chk_off; 3005 #ifdef _LP64 3006 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X)); 3007 #endif 3008 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X)); 3009 // For some types like interfaces the following loadKlass is from a 1-word 3010 // cache which is mutable so can't use immutable memory. Other 3011 // types load from the super-class display table which is immutable. 3012 Node *kmem = C->immutable_memory(); 3013 // secondary_super_cache is not immutable but can be treated as such because: 3014 // - no ideal node writes to it in a way that could cause an 3015 // incorrect/missed optimization of the following Load. 3016 // - it's a cache so, worse case, not reading the latest value 3017 // wouldn't cause incorrect execution 3018 if (might_be_cache && mem != NULL) { 3019 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem; 3020 } 3021 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL)); 3022 3023 // Compile speed common case: ARE a subtype and we canNOT fail 3024 if( superklass == nkls ) 3025 return C->top(); // false path is dead; no test needed. 3026 3027 // See if we get an immediate positive hit. Happens roughly 83% of the 3028 // time. Test to see if the value loaded just previously from the subklass 3029 // is exactly the superklass. 3030 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 3031 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1)); 3032 *ctrl = gvn.transform(new IfFalseNode(iff1)); 3033 3034 // Compile speed common case: Check for being deterministic right now. If 3035 // chk_off is a constant and not equal to cacheoff then we are NOT a 3036 // subklass. In this case we need exactly the 1 test above and we can 3037 // return those results immediately. 3038 if (!might_be_cache) { 3039 Node* not_subtype_ctrl = *ctrl; 3040 *ctrl = iftrue1; // We need exactly the 1 test above 3041 return not_subtype_ctrl; 3042 } 3043 3044 // Gather the various success & failures here 3045 RegionNode *r_ok_subtype = new RegionNode(4); 3046 gvn.record_for_igvn(r_ok_subtype); 3047 RegionNode *r_not_subtype = new RegionNode(3); 3048 gvn.record_for_igvn(r_not_subtype); 3049 3050 r_ok_subtype->init_req(1, iftrue1); 3051 3052 // Check for immediate negative hit. Happens roughly 11% of the time (which 3053 // is roughly 63% of the remaining cases). Test to see if the loaded 3054 // check-offset points into the subklass display list or the 1-element 3055 // cache. If it points to the display (and NOT the cache) and the display 3056 // missed then it's not a subtype. 3057 Node *cacheoff = gvn.intcon(cacheoff_con); 3058 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 3059 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2))); 3060 *ctrl = gvn.transform(new IfFalseNode(iff2)); 3061 3062 // Check for self. Very rare to get here, but it is taken 1/3 the time. 3063 // No performance impact (too rare) but allows sharing of secondary arrays 3064 // which has some footprint reduction. 3065 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 3066 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3))); 3067 *ctrl = gvn.transform(new IfFalseNode(iff3)); 3068 3069 // -- Roads not taken here: -- 3070 // We could also have chosen to perform the self-check at the beginning 3071 // of this code sequence, as the assembler does. This would not pay off 3072 // the same way, since the optimizer, unlike the assembler, can perform 3073 // static type analysis to fold away many successful self-checks. 3074 // Non-foldable self checks work better here in second position, because 3075 // the initial primary superclass check subsumes a self-check for most 3076 // types. An exception would be a secondary type like array-of-interface, 3077 // which does not appear in its own primary supertype display. 3078 // Finally, we could have chosen to move the self-check into the 3079 // PartialSubtypeCheckNode, and from there out-of-line in a platform 3080 // dependent manner. But it is worthwhile to have the check here, 3081 // where it can be perhaps be optimized. The cost in code space is 3082 // small (register compare, branch). 3083 3084 // Now do a linear scan of the secondary super-klass array. Again, no real 3085 // performance impact (too rare) but it's gotta be done. 3086 // Since the code is rarely used, there is no penalty for moving it 3087 // out of line, and it can only improve I-cache density. 3088 // The decision to inline or out-of-line this final check is platform 3089 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 3090 Node* psc = gvn.transform( 3091 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 3092 3093 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 3094 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4))); 3095 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4))); 3096 3097 // Return false path; set default control to true path. 3098 *ctrl = gvn.transform(r_ok_subtype); 3099 return gvn.transform(r_not_subtype); 3100 } 3101 3102 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) { 3103 const Type* sub_t = _gvn.type(obj_or_subklass); 3104 if (sub_t->isa_inlinetype()) { 3105 obj_or_subklass = makecon(TypeKlassPtr::make(sub_t->inline_klass())); 3106 } 3107 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over 3108 ExpandSubTypeCheckAtParseTime; // forced expansion 3109 if (expand_subtype_check) { 3110 MergeMemNode* mem = merged_memory(); 3111 Node* ctrl = control(); 3112 Node* subklass = obj_or_subklass; 3113 if (!sub_t->isa_klassptr() && !sub_t->isa_inlinetype()) { 3114 subklass = load_object_klass(obj_or_subklass); 3115 } 3116 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn); 3117 set_control(ctrl); 3118 return n; 3119 } 3120 3121 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass)); 3122 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq)); 3123 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 3124 set_control(_gvn.transform(new IfTrueNode(iff))); 3125 return _gvn.transform(new IfFalseNode(iff)); 3126 } 3127 3128 // Profile-driven exact type check: 3129 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 3130 float prob, Node* *casted_receiver) { 3131 assert(!klass->is_interface(), "no exact type check on interfaces"); 3132 Node* fail = top(); 3133 const Type* rec_t = _gvn.type(receiver); 3134 if (rec_t->isa_inlinetype()) { 3135 if (klass->equals(rec_t->inline_klass())) { 3136 (*casted_receiver) = receiver; // Always passes 3137 } else { 3138 (*casted_receiver) = top(); // Always fails 3139 fail = control(); 3140 set_control(top()); 3141 } 3142 return fail; 3143 } 3144 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 3145 Node* recv_klass = load_object_klass(receiver); 3146 fail = type_check(recv_klass, tklass, prob); 3147 3148 if (!stopped()) { 3149 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3150 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 3151 assert(recv_xtype->klass_is_exact(), ""); 3152 3153 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts 3154 // Subsume downstream occurrences of receiver with a cast to 3155 // recv_xtype, since now we know what the type will be. 3156 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 3157 Node* res = _gvn.transform(cast); 3158 if (recv_xtype->is_inlinetypeptr()) { 3159 assert(!gvn().type(res)->maybe_null(), "receiver should never be null"); 3160 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass())->as_InlineTypeBase()->as_ptr(&gvn()); 3161 } 3162 (*casted_receiver) = res; 3163 // (User must make the replace_in_map call.) 3164 } 3165 } 3166 3167 return fail; 3168 } 3169 3170 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass, 3171 float prob) { 3172 Node* want_klass = makecon(tklass); 3173 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass)); 3174 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3175 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 3176 set_control(_gvn.transform(new IfTrueNode (iff))); 3177 Node* fail = _gvn.transform(new IfFalseNode(iff)); 3178 return fail; 3179 } 3180 3181 //------------------------------subtype_check_receiver------------------------- 3182 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass, 3183 Node** casted_receiver) { 3184 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 3185 Node* want_klass = makecon(tklass); 3186 3187 Node* slow_ctl = gen_subtype_check(receiver, want_klass); 3188 3189 // Ignore interface type information until interface types are properly tracked. 3190 if (!stopped() && !klass->is_interface()) { 3191 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr(); 3192 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type(); 3193 if (receiver_type != NULL && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts 3194 Node* cast = new CheckCastPPNode(control(), receiver, recv_type); 3195 (*casted_receiver) = _gvn.transform(cast); 3196 } 3197 } 3198 3199 return slow_ctl; 3200 } 3201 3202 //------------------------------seems_never_null------------------------------- 3203 // Use null_seen information if it is available from the profile. 3204 // If we see an unexpected null at a type check we record it and force a 3205 // recompile; the offending check will be recompiled to handle NULLs. 3206 // If we see several offending BCIs, then all checks in the 3207 // method will be recompiled. 3208 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 3209 speculating = !_gvn.type(obj)->speculative_maybe_null(); 3210 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 3211 if (UncommonNullCast // Cutout for this technique 3212 && obj != null() // And not the -Xcomp stupid case? 3213 && !too_many_traps(reason) 3214 ) { 3215 if (speculating) { 3216 return true; 3217 } 3218 if (data == NULL) 3219 // Edge case: no mature data. Be optimistic here. 3220 return true; 3221 // If the profile has not seen a null, assume it won't happen. 3222 assert(java_bc() == Bytecodes::_checkcast || 3223 java_bc() == Bytecodes::_instanceof || 3224 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 3225 if (java_bc() == Bytecodes::_aastore) { 3226 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull; 3227 } 3228 return !data->as_BitData()->null_seen(); 3229 } 3230 speculating = false; 3231 return false; 3232 } 3233 3234 void GraphKit::guard_klass_being_initialized(Node* klass) { 3235 int init_state_off = in_bytes(InstanceKlass::init_state_offset()); 3236 Node* adr = basic_plus_adr(top(), klass, init_state_off); 3237 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr, 3238 adr->bottom_type()->is_ptr(), TypeInt::BYTE, 3239 T_BYTE, MemNode::unordered); 3240 init_state = _gvn.transform(init_state); 3241 3242 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized)); 3243 3244 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state)); 3245 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3246 3247 { BuildCutout unless(this, tst, PROB_MAX); 3248 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret); 3249 } 3250 } 3251 3252 void GraphKit::guard_init_thread(Node* klass) { 3253 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset()); 3254 Node* adr = basic_plus_adr(top(), klass, init_thread_off); 3255 3256 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr, 3257 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL, 3258 T_ADDRESS, MemNode::unordered); 3259 init_thread = _gvn.transform(init_thread); 3260 3261 Node* cur_thread = _gvn.transform(new ThreadLocalNode()); 3262 3263 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread)); 3264 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq)); 3265 3266 { BuildCutout unless(this, tst, PROB_MAX); 3267 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none); 3268 } 3269 } 3270 3271 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) { 3272 if (ik->is_being_initialized()) { 3273 if (C->needs_clinit_barrier(ik, context)) { 3274 Node* klass = makecon(TypeKlassPtr::make(ik)); 3275 guard_klass_being_initialized(klass); 3276 guard_init_thread(klass); 3277 insert_mem_bar(Op_MemBarCPUOrder); 3278 } 3279 } else if (ik->is_initialized()) { 3280 return; // no barrier needed 3281 } else { 3282 uncommon_trap(Deoptimization::Reason_uninitialized, 3283 Deoptimization::Action_reinterpret, 3284 NULL); 3285 } 3286 } 3287 3288 //------------------------maybe_cast_profiled_receiver------------------------- 3289 // If the profile has seen exactly one type, narrow to exactly that type. 3290 // Subsequent type checks will always fold up. 3291 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 3292 ciKlass* require_klass, 3293 ciKlass* spec_klass, 3294 bool safe_for_replace) { 3295 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 3296 3297 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 3298 3299 // Make sure we haven't already deoptimized from this tactic. 3300 if (too_many_traps_or_recompiles(reason)) 3301 return NULL; 3302 3303 // (No, this isn't a call, but it's enough like a virtual call 3304 // to use the same ciMethod accessor to get the profile info...) 3305 // If we have a speculative type use it instead of profiling (which 3306 // may not help us) 3307 ciKlass* exact_kls = spec_klass; 3308 if (exact_kls == NULL) { 3309 if (java_bc() == Bytecodes::_aastore) { 3310 ciKlass* array_type = NULL; 3311 ciKlass* element_type = NULL; 3312 ProfilePtrKind element_ptr = ProfileMaybeNull; 3313 bool flat_array = true; 3314 bool null_free_array = true; 3315 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 3316 exact_kls = element_type; 3317 } else { 3318 exact_kls = profile_has_unique_klass(); 3319 } 3320 } 3321 if (exact_kls != NULL) {// no cast failures here 3322 if (require_klass == NULL || 3323 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 3324 // If we narrow the type to match what the type profile sees or 3325 // the speculative type, we can then remove the rest of the 3326 // cast. 3327 // This is a win, even if the exact_kls is very specific, 3328 // because downstream operations, such as method calls, 3329 // will often benefit from the sharper type. 3330 Node* exact_obj = not_null_obj; // will get updated in place... 3331 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3332 &exact_obj); 3333 { PreserveJVMState pjvms(this); 3334 set_control(slow_ctl); 3335 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 3336 } 3337 if (safe_for_replace) { 3338 replace_in_map(not_null_obj, exact_obj); 3339 } 3340 return exact_obj; 3341 } 3342 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 3343 } 3344 3345 return NULL; 3346 } 3347 3348 /** 3349 * Cast obj to type and emit guard unless we had too many traps here 3350 * already 3351 * 3352 * @param obj node being casted 3353 * @param type type to cast the node to 3354 * @param not_null true if we know node cannot be null 3355 */ 3356 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 3357 ciKlass* type, 3358 bool not_null) { 3359 if (stopped()) { 3360 return obj; 3361 } 3362 3363 // type == NULL if profiling tells us this object is always null 3364 if (type != NULL) { 3365 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 3366 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 3367 3368 if (!too_many_traps_or_recompiles(null_reason) && 3369 !too_many_traps_or_recompiles(class_reason)) { 3370 Node* not_null_obj = NULL; 3371 // not_null is true if we know the object is not null and 3372 // there's no need for a null check 3373 if (!not_null) { 3374 Node* null_ctl = top(); 3375 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 3376 assert(null_ctl->is_top(), "no null control here"); 3377 } else { 3378 not_null_obj = obj; 3379 } 3380 3381 Node* exact_obj = not_null_obj; 3382 ciKlass* exact_kls = type; 3383 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 3384 &exact_obj); 3385 { 3386 PreserveJVMState pjvms(this); 3387 set_control(slow_ctl); 3388 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 3389 } 3390 replace_in_map(not_null_obj, exact_obj); 3391 obj = exact_obj; 3392 } 3393 } else { 3394 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3395 Node* exact_obj = null_assert(obj); 3396 replace_in_map(obj, exact_obj); 3397 obj = exact_obj; 3398 } 3399 } 3400 return obj; 3401 } 3402 3403 //-------------------------------gen_instanceof-------------------------------- 3404 // Generate an instance-of idiom. Used by both the instance-of bytecode 3405 // and the reflective instance-of call. 3406 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 3407 kill_dead_locals(); // Benefit all the uncommon traps 3408 assert( !stopped(), "dead parse path should be checked in callers" ); 3409 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 3410 "must check for not-null not-dead klass in callers"); 3411 3412 // Make the merge point 3413 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 3414 RegionNode* region = new RegionNode(PATH_LIMIT); 3415 Node* phi = new PhiNode(region, TypeInt::BOOL); 3416 C->set_has_split_ifs(true); // Has chance for split-if optimization 3417 3418 ciProfileData* data = NULL; 3419 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 3420 data = method()->method_data()->bci_to_data(bci()); 3421 } 3422 bool speculative_not_null = false; 3423 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 3424 && seems_never_null(obj, data, speculative_not_null)); 3425 bool is_value = obj->is_InlineType(); 3426 3427 // Null check; get casted pointer; set region slot 3 3428 Node* null_ctl = top(); 3429 if (is_value) { 3430 // TODO 8284443 Enable this 3431 safe_for_replace = false; 3432 never_see_null = false; 3433 } 3434 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3435 3436 // If not_null_obj is dead, only null-path is taken 3437 if (stopped()) { // Doing instance-of on a NULL? 3438 set_control(null_ctl); 3439 return intcon(0); 3440 } 3441 region->init_req(_null_path, null_ctl); 3442 phi ->init_req(_null_path, intcon(0)); // Set null path value 3443 if (null_ctl == top()) { 3444 // Do this eagerly, so that pattern matches like is_diamond_phi 3445 // will work even during parsing. 3446 assert(_null_path == PATH_LIMIT-1, "delete last"); 3447 region->del_req(_null_path); 3448 phi ->del_req(_null_path); 3449 } 3450 3451 // Do we know the type check always succeed? 3452 if (!is_value) { 3453 bool known_statically = false; 3454 if (_gvn.type(superklass)->singleton()) { 3455 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 3456 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 3457 if (subk != NULL && subk->is_loaded()) { 3458 int static_res = C->static_subtype_check(superk, subk); 3459 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 3460 } 3461 } 3462 3463 if (!known_statically) { 3464 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3465 // We may not have profiling here or it may not help us. If we 3466 // have a speculative type use it to perform an exact cast. 3467 ciKlass* spec_obj_type = obj_type->speculative_type(); 3468 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 3469 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 3470 if (stopped()) { // Profile disagrees with this path. 3471 set_control(null_ctl); // Null is the only remaining possibility. 3472 return intcon(0); 3473 } 3474 if (cast_obj != NULL) { 3475 not_null_obj = cast_obj; 3476 is_value = not_null_obj->is_InlineType(); 3477 } 3478 } 3479 } 3480 } 3481 3482 // Generate the subtype check 3483 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3484 3485 // Plug in the success path to the general merge in slot 1. 3486 region->init_req(_obj_path, control()); 3487 phi ->init_req(_obj_path, intcon(1)); 3488 3489 // Plug in the failing path to the general merge in slot 2. 3490 region->init_req(_fail_path, not_subtype_ctrl); 3491 phi ->init_req(_fail_path, intcon(0)); 3492 3493 // Return final merged results 3494 set_control( _gvn.transform(region) ); 3495 record_for_igvn(region); 3496 3497 // If we know the type check always succeeds then we don't use the 3498 // profiling data at this bytecode. Don't lose it, feed it to the 3499 // type system as a speculative type. 3500 if (safe_for_replace && !is_value) { 3501 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3502 replace_in_map(obj, casted_obj); 3503 } 3504 3505 return _gvn.transform(phi); 3506 } 3507 3508 //-------------------------------gen_checkcast--------------------------------- 3509 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3510 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3511 // uncommon-trap paths work. Adjust stack after this call. 3512 // If failure_control is supplied and not null, it is filled in with 3513 // the control edge for the cast failure. Otherwise, an appropriate 3514 // uncommon trap or exception is thrown. 3515 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) { 3516 kill_dead_locals(); // Benefit all the uncommon traps 3517 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr(); 3518 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass()); 3519 bool safe_for_replace = (failure_control == NULL); 3520 bool from_inline = obj->is_InlineType(); 3521 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer"); 3522 3523 // Fast cutout: Check the case that the cast is vacuously true. 3524 // This detects the common cases where the test will short-circuit 3525 // away completely. We do this before we perform the null check, 3526 // because if the test is going to turn into zero code, we don't 3527 // want a residual null check left around. (Causes a slowdown, 3528 // for example, in some objArray manipulations, such as a[i]=a[j].) 3529 if (tk->singleton()) { 3530 ciKlass* klass = NULL; 3531 if (obj->is_InlineTypeBase()) { 3532 klass = _gvn.type(obj)->inline_klass(); 3533 } else { 3534 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3535 if (objtp != NULL) { 3536 klass = objtp->klass(); 3537 } 3538 } 3539 if (klass != NULL) { 3540 switch (C->static_subtype_check(tk->klass(), klass)) { 3541 case Compile::SSC_always_true: 3542 // If we know the type check always succeed then we don't use 3543 // the profiling data at this bytecode. Don't lose it, feed it 3544 // to the type system as a speculative type. 3545 if (!from_inline) { 3546 obj = record_profiled_receiver_for_speculation(obj); 3547 } 3548 if (null_free) { 3549 assert(safe_for_replace, "must be"); 3550 obj = null_check(obj); 3551 } 3552 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineTypeBase(), "should have been scalarized"); 3553 return obj; 3554 case Compile::SSC_always_false: 3555 if (null_free) { 3556 assert(safe_for_replace, "must be"); 3557 obj = null_check(obj); 3558 } 3559 // It needs a null check because a null will *pass* the cast check. 3560 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3561 if (objtp != NULL && !objtp->maybe_null()) { 3562 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3563 Deoptimization::DeoptReason reason = is_aastore ? 3564 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3565 builtin_throw(reason, makecon(TypeKlassPtr::make(klass))); 3566 return top(); 3567 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) { 3568 return null_assert(obj); 3569 } 3570 break; // Fall through to full check 3571 } 3572 } 3573 } 3574 3575 ciProfileData* data = NULL; 3576 if (failure_control == NULL) { // use MDO in regular case only 3577 assert(java_bc() == Bytecodes::_aastore || 3578 java_bc() == Bytecodes::_checkcast, 3579 "interpreter profiles type checks only for these BCs"); 3580 if (method()->method_data()->is_mature()) { 3581 data = method()->method_data()->bci_to_data(bci()); 3582 } 3583 } 3584 3585 // Make the merge point 3586 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3587 RegionNode* region = new RegionNode(PATH_LIMIT); 3588 Node* phi = new PhiNode(region, toop); 3589 _gvn.set_type(region, Type::CONTROL); 3590 _gvn.set_type(phi, toop); 3591 3592 C->set_has_split_ifs(true); // Has chance for split-if optimization 3593 3594 // Use null-cast information if it is available 3595 bool speculative_not_null = false; 3596 bool never_see_null = ((failure_control == NULL) // regular case only 3597 && seems_never_null(obj, data, speculative_not_null)); 3598 3599 // Null check; get casted pointer; set region slot 3 3600 Node* null_ctl = top(); 3601 Node* not_null_obj = NULL; 3602 if (null_free) { 3603 assert(safe_for_replace, "must be"); 3604 not_null_obj = null_check(obj); 3605 } else if (from_inline) { 3606 // TODO 8284443 obj can be null and null should pass 3607 not_null_obj = obj; 3608 } else { 3609 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3610 } 3611 3612 // If not_null_obj is dead, only null-path is taken 3613 if (stopped()) { // Doing instance-of on a NULL? 3614 set_control(null_ctl); 3615 if (toop->is_inlinetypeptr()) { 3616 return InlineTypePtrNode::make_null(_gvn, toop->inline_klass()); 3617 } 3618 return null(); 3619 } 3620 region->init_req(_null_path, null_ctl); 3621 phi ->init_req(_null_path, null()); // Set null path value 3622 if (null_ctl == top()) { 3623 // Do this eagerly, so that pattern matches like is_diamond_phi 3624 // will work even during parsing. 3625 assert(_null_path == PATH_LIMIT-1, "delete last"); 3626 region->del_req(_null_path); 3627 phi ->del_req(_null_path); 3628 } 3629 3630 Node* cast_obj = NULL; 3631 if (!from_inline && tk->klass_is_exact()) { 3632 // The following optimization tries to statically cast the speculative type of the object 3633 // (for example obtained during profiling) to the type of the superklass and then do a 3634 // dynamic check that the type of the object is what we expect. To work correctly 3635 // for checkcast and aastore the type of superklass should be exact. 3636 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3637 // We may not have profiling here or it may not help us. If we have 3638 // a speculative type use it to perform an exact cast. 3639 ciKlass* spec_obj_type = obj_type->speculative_type(); 3640 if (spec_obj_type != NULL || data != NULL) { 3641 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3642 if (cast_obj != NULL) { 3643 if (failure_control != NULL) // failure is now impossible 3644 (*failure_control) = top(); 3645 // adjust the type of the phi to the exact klass: 3646 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3647 } 3648 } 3649 } 3650 3651 if (cast_obj == NULL) { 3652 // Generate the subtype check 3653 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass); 3654 3655 // Plug in success path into the merge 3656 cast_obj = from_inline ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3657 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3658 if (failure_control == NULL) { 3659 if (not_subtype_ctrl != top()) { // If failure is possible 3660 PreserveJVMState pjvms(this); 3661 set_control(not_subtype_ctrl); 3662 Node* obj_klass = NULL; 3663 if (not_null_obj->is_InlineTypeBase()) { 3664 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass())); 3665 } else { 3666 obj_klass = load_object_klass(not_null_obj); 3667 } 3668 bool is_aastore = (java_bc() == Bytecodes::_aastore); 3669 Deoptimization::DeoptReason reason = is_aastore ? 3670 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check; 3671 builtin_throw(reason, obj_klass); 3672 } 3673 } else { 3674 (*failure_control) = not_subtype_ctrl; 3675 } 3676 } 3677 3678 region->init_req(_obj_path, control()); 3679 phi ->init_req(_obj_path, cast_obj); 3680 3681 // A merge of NULL or Casted-NotNull obj 3682 Node* res = _gvn.transform(phi); 3683 3684 // Note I do NOT always 'replace_in_map(obj,result)' here. 3685 // if( tk->klass()->can_be_primary_super() ) 3686 // This means that if I successfully store an Object into an array-of-String 3687 // I 'forget' that the Object is really now known to be a String. I have to 3688 // do this because we don't have true union types for interfaces - if I store 3689 // a Baz into an array-of-Interface and then tell the optimizer it's an 3690 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3691 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3692 // replace_in_map( obj, res ); 3693 3694 // Return final merged results 3695 set_control( _gvn.transform(region) ); 3696 record_for_igvn(region); 3697 3698 bool not_inline = !toop->can_be_inline_type(); 3699 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array()); 3700 if (EnableValhalla && not_flattened) { 3701 // Check if obj has been loaded from an array 3702 obj = obj->isa_DecodeN() ? obj->in(1) : obj; 3703 Node* array = NULL; 3704 if (obj->isa_Load()) { 3705 Node* address = obj->in(MemNode::Address); 3706 if (address->isa_AddP()) { 3707 array = address->as_AddP()->in(AddPNode::Base); 3708 } 3709 } else if (obj->is_Phi()) { 3710 Node* region = obj->in(0); 3711 // TODO make this more robust (see JDK-8231346) 3712 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) { 3713 IfNode* iff = region->in(2)->in(0)->isa_If(); 3714 if (iff != NULL) { 3715 iff->is_flat_array_check(&_gvn, &array); 3716 } 3717 } 3718 } 3719 if (array != NULL) { 3720 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr(); 3721 if (ary_t != NULL) { 3722 if (!ary_t->is_not_null_free() && not_inline) { 3723 // Casting array element to a non-inline-type, mark array as not null-free. 3724 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free())); 3725 replace_in_map(array, cast); 3726 } else if (!ary_t->is_not_flat()) { 3727 // Casting array element to a non-flattened type, mark array as not flat. 3728 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat())); 3729 replace_in_map(array, cast); 3730 } 3731 } 3732 } 3733 } 3734 3735 if (!stopped() && !res->is_InlineTypeBase()) { 3736 res = record_profiled_receiver_for_speculation(res); 3737 if (toop->is_inlinetypeptr()) { 3738 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null()); 3739 res = vt; 3740 if (safe_for_replace) { 3741 if (vt->is_InlineType() && C->inlining_incrementally()) { 3742 vt = vt->as_InlineType()->as_ptr(&_gvn); 3743 } 3744 replace_in_map(obj, vt); 3745 replace_in_map(not_null_obj, vt); 3746 replace_in_map(res, vt); 3747 } 3748 } 3749 } 3750 return res; 3751 } 3752 3753 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) { 3754 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 3755 Node* mark = make_load(NULL, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 3756 Node* mask = MakeConX(markWord::inline_type_pattern); 3757 Node* masked = _gvn.transform(new AndXNode(mark, mask)); 3758 Node* cmp = _gvn.transform(new CmpXNode(masked, mask)); 3759 return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne)); 3760 } 3761 3762 Node* GraphKit::is_val_mirror(Node* mirror) { 3763 Node* p = basic_plus_adr(mirror, java_lang_Class::secondary_mirror_offset()); 3764 Node* secondary_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP); 3765 Node* cmp = _gvn.transform(new CmpPNode(mirror, secondary_mirror)); 3766 return _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 3767 } 3768 3769 Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) { 3770 Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset())); 3771 // Make sure to use immutable memory here to enable hoisting the check out of loops 3772 Node* lh_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered)); 3773 Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask))); 3774 Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val))); 3775 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne)); 3776 } 3777 3778 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) { 3779 // We can't use immutable memory here because the mark word is mutable. 3780 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the 3781 // check is moved out of loops (mainly to enable loop unswitching). 3782 Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory(); 3783 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, array_or_klass)); 3784 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN 3785 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne)); 3786 } 3787 3788 Node* GraphKit::null_free_array_test(Node* klass, bool null_free) { 3789 return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free); 3790 } 3791 3792 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null 3793 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) { 3794 RegionNode* region = new RegionNode(3); 3795 Node* null_ctl = top(); 3796 null_check_oop(val, &null_ctl); 3797 if (null_ctl != top()) { 3798 PreserveJVMState pjvms(this); 3799 set_control(null_ctl); 3800 { 3801 // Deoptimize if null-free array 3802 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX); 3803 inc_sp(nargs); 3804 uncommon_trap(Deoptimization::Reason_null_check, 3805 Deoptimization::Action_none); 3806 } 3807 region->init_req(1, control()); 3808 } 3809 region->init_req(2, control()); 3810 set_control(_gvn.transform(region)); 3811 record_for_igvn(region); 3812 if (_gvn.type(val) == TypePtr::NULL_PTR) { 3813 // Since we were just successfully storing null, the array can't be null free. 3814 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 3815 ary_t = ary_t->cast_to_not_null_free(); 3816 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t)); 3817 if (safe_for_replace) { 3818 replace_in_map(ary, cast); 3819 } 3820 ary = cast; 3821 } 3822 return ary; 3823 } 3824 3825 //------------------------------next_monitor----------------------------------- 3826 // What number should be given to the next monitor? 3827 int GraphKit::next_monitor() { 3828 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3829 int next = current + C->sync_stack_slots(); 3830 // Keep the toplevel high water mark current: 3831 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3832 return current; 3833 } 3834 3835 //------------------------------insert_mem_bar--------------------------------- 3836 // Memory barrier to avoid floating things around 3837 // The membar serves as a pinch point between both control and all memory slices. 3838 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3839 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3840 mb->init_req(TypeFunc::Control, control()); 3841 mb->init_req(TypeFunc::Memory, reset_memory()); 3842 Node* membar = _gvn.transform(mb); 3843 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3844 set_all_memory_call(membar); 3845 return membar; 3846 } 3847 3848 //-------------------------insert_mem_bar_volatile---------------------------- 3849 // Memory barrier to avoid floating things around 3850 // The membar serves as a pinch point between both control and memory(alias_idx). 3851 // If you want to make a pinch point on all memory slices, do not use this 3852 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3853 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3854 // When Parse::do_put_xxx updates a volatile field, it appends a series 3855 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3856 // The first membar is on the same memory slice as the field store opcode. 3857 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3858 // All the other membars (for other volatile slices, including AliasIdxBot, 3859 // which stands for all unknown volatile slices) are control-dependent 3860 // on the first membar. This prevents later volatile loads or stores 3861 // from sliding up past the just-emitted store. 3862 3863 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3864 mb->set_req(TypeFunc::Control,control()); 3865 if (alias_idx == Compile::AliasIdxBot) { 3866 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3867 } else { 3868 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3869 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3870 } 3871 Node* membar = _gvn.transform(mb); 3872 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3873 if (alias_idx == Compile::AliasIdxBot) { 3874 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3875 } else { 3876 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3877 } 3878 return membar; 3879 } 3880 3881 //------------------------------shared_lock------------------------------------ 3882 // Emit locking code. 3883 FastLockNode* GraphKit::shared_lock(Node* obj) { 3884 // bci is either a monitorenter bc or InvocationEntryBci 3885 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3886 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3887 3888 if( !GenerateSynchronizationCode ) 3889 return NULL; // Not locking things? 3890 3891 if (stopped()) // Dead monitor? 3892 return NULL; 3893 3894 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3895 3896 // Box the stack location 3897 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3898 Node* mem = reset_memory(); 3899 3900 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3901 3902 // Create the rtm counters for this fast lock if needed. 3903 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3904 3905 // Add monitor to debug info for the slow path. If we block inside the 3906 // slow path and de-opt, we need the monitor hanging around 3907 map()->push_monitor( flock ); 3908 3909 const TypeFunc *tf = LockNode::lock_type(); 3910 LockNode *lock = new LockNode(C, tf); 3911 3912 lock->init_req( TypeFunc::Control, control() ); 3913 lock->init_req( TypeFunc::Memory , mem ); 3914 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3915 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3916 lock->init_req( TypeFunc::ReturnAdr, top() ); 3917 3918 lock->init_req(TypeFunc::Parms + 0, obj); 3919 lock->init_req(TypeFunc::Parms + 1, box); 3920 lock->init_req(TypeFunc::Parms + 2, flock); 3921 add_safepoint_edges(lock); 3922 3923 lock = _gvn.transform( lock )->as_Lock(); 3924 3925 // lock has no side-effects, sets few values 3926 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3927 3928 insert_mem_bar(Op_MemBarAcquireLock); 3929 3930 // Add this to the worklist so that the lock can be eliminated 3931 record_for_igvn(lock); 3932 3933 #ifndef PRODUCT 3934 if (PrintLockStatistics) { 3935 // Update the counter for this lock. Don't bother using an atomic 3936 // operation since we don't require absolute accuracy. 3937 lock->create_lock_counter(map()->jvms()); 3938 increment_counter(lock->counter()->addr()); 3939 } 3940 #endif 3941 3942 return flock; 3943 } 3944 3945 3946 //------------------------------shared_unlock---------------------------------- 3947 // Emit unlocking code. 3948 void GraphKit::shared_unlock(Node* box, Node* obj) { 3949 // bci is either a monitorenter bc or InvocationEntryBci 3950 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3951 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3952 3953 if( !GenerateSynchronizationCode ) 3954 return; 3955 if (stopped()) { // Dead monitor? 3956 map()->pop_monitor(); // Kill monitor from debug info 3957 return; 3958 } 3959 assert(!obj->is_InlineTypeBase(), "should not unlock on inline type"); 3960 3961 // Memory barrier to avoid floating things down past the locked region 3962 insert_mem_bar(Op_MemBarReleaseLock); 3963 3964 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3965 UnlockNode *unlock = new UnlockNode(C, tf); 3966 #ifdef ASSERT 3967 unlock->set_dbg_jvms(sync_jvms()); 3968 #endif 3969 uint raw_idx = Compile::AliasIdxRaw; 3970 unlock->init_req( TypeFunc::Control, control() ); 3971 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3972 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3973 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3974 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3975 3976 unlock->init_req(TypeFunc::Parms + 0, obj); 3977 unlock->init_req(TypeFunc::Parms + 1, box); 3978 unlock = _gvn.transform(unlock)->as_Unlock(); 3979 3980 Node* mem = reset_memory(); 3981 3982 // unlock has no side-effects, sets few values 3983 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3984 3985 // Kill monitor from debug info 3986 map()->pop_monitor( ); 3987 } 3988 3989 //-------------------------------get_layout_helper----------------------------- 3990 // If the given klass is a constant or known to be an array, 3991 // fetch the constant layout helper value into constant_value 3992 // and return (Node*)NULL. Otherwise, load the non-constant 3993 // layout helper value, and return the node which represents it. 3994 // This two-faced routine is useful because allocation sites 3995 // almost always feature constant types. 3996 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3997 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3998 if (!StressReflectiveCode && inst_klass != NULL) { 3999 ciKlass* klass = inst_klass->klass(); 4000 assert(klass != NULL, "klass should not be NULL"); 4001 bool xklass = inst_klass->klass_is_exact(); 4002 bool can_be_flattened = false; 4003 if (UseFlatArray && klass->is_obj_array_klass() && !klass->as_obj_array_klass()->is_elem_null_free()) { 4004 // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue. 4005 ciKlass* elem = klass->as_obj_array_klass()->element_klass(); 4006 can_be_flattened = elem->can_be_inline_klass() && (!elem->is_inlinetype() || elem->flatten_array()); 4007 } 4008 if (!can_be_flattened && (xklass || klass->is_array_klass())) { 4009 jint lhelper = klass->layout_helper(); 4010 if (lhelper != Klass::_lh_neutral_value) { 4011 constant_value = lhelper; 4012 return (Node*) NULL; 4013 } 4014 } 4015 } 4016 constant_value = Klass::_lh_neutral_value; // put in a known value 4017 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 4018 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 4019 } 4020 4021 // We just put in an allocate/initialize with a big raw-memory effect. 4022 // Hook selected additional alias categories on the initialization. 4023 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 4024 MergeMemNode* init_in_merge, 4025 Node* init_out_raw) { 4026 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 4027 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 4028 4029 Node* prevmem = kit.memory(alias_idx); 4030 init_in_merge->set_memory_at(alias_idx, prevmem); 4031 if (init_out_raw != NULL) { 4032 kit.set_memory(init_out_raw, alias_idx); 4033 } 4034 } 4035 4036 //---------------------------set_output_for_allocation------------------------- 4037 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 4038 const TypeOopPtr* oop_type, 4039 bool deoptimize_on_exception) { 4040 int rawidx = Compile::AliasIdxRaw; 4041 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 4042 add_safepoint_edges(alloc); 4043 Node* allocx = _gvn.transform(alloc); 4044 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 4045 // create memory projection for i_o 4046 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 4047 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 4048 4049 // create a memory projection as for the normal control path 4050 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 4051 set_memory(malloc, rawidx); 4052 4053 // a normal slow-call doesn't change i_o, but an allocation does 4054 // we create a separate i_o projection for the normal control path 4055 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 4056 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 4057 4058 // put in an initialization barrier 4059 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 4060 rawoop)->as_Initialize(); 4061 assert(alloc->initialization() == init, "2-way macro link must work"); 4062 assert(init ->allocation() == alloc, "2-way macro link must work"); 4063 { 4064 // Extract memory strands which may participate in the new object's 4065 // initialization, and source them from the new InitializeNode. 4066 // This will allow us to observe initializations when they occur, 4067 // and link them properly (as a group) to the InitializeNode. 4068 assert(init->in(InitializeNode::Memory) == malloc, ""); 4069 MergeMemNode* minit_in = MergeMemNode::make(malloc); 4070 init->set_req(InitializeNode::Memory, minit_in); 4071 record_for_igvn(minit_in); // fold it up later, if possible 4072 _gvn.set_type(minit_in, Type::MEMORY); 4073 Node* minit_out = memory(rawidx); 4074 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 4075 // Add an edge in the MergeMem for the header fields so an access 4076 // to one of those has correct memory state 4077 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()))); 4078 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()))); 4079 if (oop_type->isa_aryptr()) { 4080 const TypeAryPtr* arytype = oop_type->is_aryptr(); 4081 if (arytype->klass()->is_flat_array_klass()) { 4082 // Initially all flattened array accesses share a single slice 4083 // but that changes after parsing. Prepare the memory graph so 4084 // it can optimize flattened array accesses properly once they 4085 // don't share a single slice. 4086 assert(C->flattened_accesses_share_alias(), "should be set at parse time"); 4087 C->set_flattened_accesses_share_alias(false); 4088 ciFlatArrayKlass* vak = arytype->klass()->as_flat_array_klass(); 4089 ciInlineKlass* vk = vak->element_klass()->as_inline_klass(); 4090 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) { 4091 ciField* field = vk->nonstatic_field_at(i); 4092 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 4093 continue; // do not bother to track really large numbers of fields 4094 int off_in_vt = field->offset() - vk->first_field_offset(); 4095 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot); 4096 int fieldidx = C->get_alias_index(adr_type, true); 4097 // Pass NULL for init_out. Having per flat array element field memory edges as uses of the Initialize node 4098 // can result in per flat array field Phis to be created which confuses the logic of 4099 // Compile::adjust_flattened_array_access_aliases(). 4100 hook_memory_on_init(*this, fieldidx, minit_in, NULL); 4101 } 4102 C->set_flattened_accesses_share_alias(true); 4103 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out); 4104 } else { 4105 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 4106 int elemidx = C->get_alias_index(telemref); 4107 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 4108 } 4109 } else if (oop_type->isa_instptr()) { 4110 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word 4111 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 4112 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 4113 ciField* field = ik->nonstatic_field_at(i); 4114 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 4115 continue; // do not bother to track really large numbers of fields 4116 // Find (or create) the alias category for this field: 4117 int fieldidx = C->alias_type(field)->index(); 4118 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 4119 } 4120 } 4121 } 4122 4123 // Cast raw oop to the real thing... 4124 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 4125 javaoop = _gvn.transform(javaoop); 4126 C->set_recent_alloc(control(), javaoop); 4127 assert(just_allocated_object(control()) == javaoop, "just allocated"); 4128 4129 #ifdef ASSERT 4130 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 4131 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 4132 "Ideal_allocation works"); 4133 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 4134 "Ideal_allocation works"); 4135 if (alloc->is_AllocateArray()) { 4136 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 4137 "Ideal_allocation works"); 4138 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 4139 "Ideal_allocation works"); 4140 } else { 4141 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 4142 } 4143 } 4144 #endif //ASSERT 4145 4146 return javaoop; 4147 } 4148 4149 //---------------------------new_instance-------------------------------------- 4150 // This routine takes a klass_node which may be constant (for a static type) 4151 // or may be non-constant (for reflective code). It will work equally well 4152 // for either, and the graph will fold nicely if the optimizer later reduces 4153 // the type to a constant. 4154 // The optional arguments are for specialized use by intrinsics: 4155 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 4156 // - If 'return_size_val', report the the total object size to the caller. 4157 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 4158 Node* GraphKit::new_instance(Node* klass_node, 4159 Node* extra_slow_test, 4160 Node* *return_size_val, 4161 bool deoptimize_on_exception, 4162 InlineTypeBaseNode* inline_type_node) { 4163 // Compute size in doublewords 4164 // The size is always an integral number of doublewords, represented 4165 // as a positive bytewise size stored in the klass's layout_helper. 4166 // The layout_helper also encodes (in a low bit) the need for a slow path. 4167 jint layout_con = Klass::_lh_neutral_value; 4168 Node* layout_val = get_layout_helper(klass_node, layout_con); 4169 bool layout_is_con = (layout_val == NULL); 4170 4171 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 4172 // Generate the initial go-slow test. It's either ALWAYS (return a 4173 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 4174 // case) a computed value derived from the layout_helper. 4175 Node* initial_slow_test = NULL; 4176 if (layout_is_con) { 4177 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4178 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 4179 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 4180 } else { // reflective case 4181 // This reflective path is used by Unsafe.allocateInstance. 4182 // (It may be stress-tested by specifying StressReflectiveCode.) 4183 // Basically, we want to get into the VM is there's an illegal argument. 4184 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 4185 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 4186 if (extra_slow_test != intcon(0)) { 4187 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 4188 } 4189 // (Macro-expander will further convert this to a Bool, if necessary.) 4190 } 4191 4192 // Find the size in bytes. This is easy; it's the layout_helper. 4193 // The size value must be valid even if the slow path is taken. 4194 Node* size = NULL; 4195 if (layout_is_con) { 4196 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 4197 } else { // reflective case 4198 // This reflective path is used by clone and Unsafe.allocateInstance. 4199 size = ConvI2X(layout_val); 4200 4201 // Clear the low bits to extract layout_helper_size_in_bytes: 4202 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 4203 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 4204 size = _gvn.transform( new AndXNode(size, mask) ); 4205 } 4206 if (return_size_val != NULL) { 4207 (*return_size_val) = size; 4208 } 4209 4210 // This is a precise notnull oop of the klass. 4211 // (Actually, it need not be precise if this is a reflective allocation.) 4212 // It's what we cast the result to. 4213 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 4214 if (!tklass) tklass = TypeInstKlassPtr::OBJECT; 4215 const TypeOopPtr* oop_type = tklass->as_instance_type(); 4216 4217 // Now generate allocation code 4218 4219 // The entire memory state is needed for slow path of the allocation 4220 // since GC and deoptimization can happen. 4221 Node *mem = reset_memory(); 4222 set_all_memory(mem); // Create new memory state 4223 4224 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 4225 control(), mem, i_o(), 4226 size, klass_node, 4227 initial_slow_test, inline_type_node); 4228 4229 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 4230 } 4231 4232 //-------------------------------new_array------------------------------------- 4233 // helper for newarray and anewarray 4234 // The 'length' parameter is (obviously) the length of the array. 4235 // See comments on new_instance for the meaning of the other arguments. 4236 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 4237 Node* length, // number of array elements 4238 int nargs, // number of arguments to push back for uncommon trap 4239 Node* *return_size_val, 4240 bool deoptimize_on_exception) { 4241 jint layout_con = Klass::_lh_neutral_value; 4242 Node* layout_val = get_layout_helper(klass_node, layout_con); 4243 bool layout_is_con = (layout_val == NULL); 4244 4245 if (!layout_is_con && !StressReflectiveCode && 4246 !too_many_traps(Deoptimization::Reason_class_check)) { 4247 // This is a reflective array creation site. 4248 // Optimistically assume that it is a subtype of Object[], 4249 // so that we can fold up all the address arithmetic. 4250 layout_con = Klass::array_layout_helper(T_OBJECT); 4251 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 4252 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 4253 { BuildCutout unless(this, bol_lh, PROB_MAX); 4254 inc_sp(nargs); 4255 uncommon_trap(Deoptimization::Reason_class_check, 4256 Deoptimization::Action_maybe_recompile); 4257 } 4258 layout_val = NULL; 4259 layout_is_con = true; 4260 } 4261 4262 // Generate the initial go-slow test. Make sure we do not overflow 4263 // if length is huge (near 2Gig) or negative! We do not need 4264 // exact double-words here, just a close approximation of needed 4265 // double-words. We can't add any offset or rounding bits, lest we 4266 // take a size -1 of bytes and make it positive. Use an unsigned 4267 // compare, so negative sizes look hugely positive. 4268 int fast_size_limit = FastAllocateSizeLimit; 4269 if (layout_is_con) { 4270 assert(!StressReflectiveCode, "stress mode does not use these paths"); 4271 // Increase the size limit if we have exact knowledge of array type. 4272 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 4273 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0); 4274 } 4275 4276 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 4277 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 4278 4279 // --- Size Computation --- 4280 // array_size = round_to_heap(array_header + (length << elem_shift)); 4281 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 4282 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 4283 // The rounding mask is strength-reduced, if possible. 4284 int round_mask = MinObjAlignmentInBytes - 1; 4285 Node* header_size = NULL; 4286 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 4287 // (T_BYTE has the weakest alignment and size restrictions...) 4288 if (layout_is_con) { 4289 int hsize = Klass::layout_helper_header_size(layout_con); 4290 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4291 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con); 4292 if ((round_mask & ~right_n_bits(eshift)) == 0) 4293 round_mask = 0; // strength-reduce it if it goes away completely 4294 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 4295 assert(header_size_min <= hsize, "generic minimum is smallest"); 4296 header_size_min = hsize; 4297 header_size = intcon(hsize + round_mask); 4298 } else { 4299 Node* hss = intcon(Klass::_lh_header_size_shift); 4300 Node* hsm = intcon(Klass::_lh_header_size_mask); 4301 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 4302 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 4303 Node* mask = intcon(round_mask); 4304 header_size = _gvn.transform( new AddINode(hsize, mask) ); 4305 } 4306 4307 Node* elem_shift = NULL; 4308 if (layout_is_con) { 4309 int eshift = Klass::layout_helper_log2_element_size(layout_con); 4310 if (eshift != 0) 4311 elem_shift = intcon(eshift); 4312 } else { 4313 // There is no need to mask or shift this value. 4314 // The semantics of LShiftINode include an implicit mask to 0x1F. 4315 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 4316 elem_shift = layout_val; 4317 } 4318 4319 // Transition to native address size for all offset calculations: 4320 Node* lengthx = ConvI2X(length); 4321 Node* headerx = ConvI2X(header_size); 4322 #ifdef _LP64 4323 { const TypeInt* tilen = _gvn.find_int_type(length); 4324 if (tilen != NULL && tilen->_lo < 0) { 4325 // Add a manual constraint to a positive range. Cf. array_element_address. 4326 jint size_max = fast_size_limit; 4327 if (size_max > tilen->_hi) size_max = tilen->_hi; 4328 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 4329 4330 // Only do a narrow I2L conversion if the range check passed. 4331 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 4332 _gvn.transform(iff); 4333 RegionNode* region = new RegionNode(3); 4334 _gvn.set_type(region, Type::CONTROL); 4335 lengthx = new PhiNode(region, TypeLong::LONG); 4336 _gvn.set_type(lengthx, TypeLong::LONG); 4337 4338 // Range check passed. Use ConvI2L node with narrow type. 4339 Node* passed = IfFalse(iff); 4340 region->init_req(1, passed); 4341 // Make I2L conversion control dependent to prevent it from 4342 // floating above the range check during loop optimizations. 4343 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 4344 4345 // Range check failed. Use ConvI2L with wide type because length may be invalid. 4346 region->init_req(2, IfTrue(iff)); 4347 lengthx->init_req(2, ConvI2X(length)); 4348 4349 set_control(region); 4350 record_for_igvn(region); 4351 record_for_igvn(lengthx); 4352 } 4353 } 4354 #endif 4355 4356 // Combine header size (plus rounding) and body size. Then round down. 4357 // This computation cannot overflow, because it is used only in two 4358 // places, one where the length is sharply limited, and the other 4359 // after a successful allocation. 4360 Node* abody = lengthx; 4361 if (elem_shift != NULL) 4362 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 4363 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 4364 if (round_mask != 0) { 4365 Node* mask = MakeConX(~round_mask); 4366 size = _gvn.transform( new AndXNode(size, mask) ); 4367 } 4368 // else if round_mask == 0, the size computation is self-rounding 4369 4370 if (return_size_val != NULL) { 4371 // This is the size 4372 (*return_size_val) = size; 4373 } 4374 4375 // Now generate allocation code 4376 4377 // The entire memory state is needed for slow path of the allocation 4378 // since GC and deoptimization can happen. 4379 Node *mem = reset_memory(); 4380 set_all_memory(mem); // Create new memory state 4381 4382 if (initial_slow_test->is_Bool()) { 4383 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 4384 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 4385 } 4386 4387 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr(); 4388 const TypeOopPtr* ary_type = ary_klass->as_instance_type(); 4389 Node* valid_length_test = _gvn.intcon(1); 4390 if (ary_type->klass()->is_array_klass()) { 4391 BasicType bt = ary_type->klass()->as_array_klass()->element_type()->basic_type(); 4392 jint max = TypeAryPtr::max_array_length(bt); 4393 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max))); 4394 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le)); 4395 } 4396 4397 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr(); 4398 4399 // Inline type array variants: 4400 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue") 4401 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue") 4402 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue") 4403 // Check if array is a null-free, non-flattened inline type array 4404 // that needs to be initialized with the default inline type. 4405 Node* default_value = NULL; 4406 Node* raw_default_value = NULL; 4407 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) { 4408 // Array type is known 4409 if (ary_ptr->klass()->as_array_klass()->is_elem_null_free()) { 4410 ciInlineKlass* vk = ary_ptr->klass()->as_array_klass()->element_klass()->as_inline_klass(); 4411 if (!vk->flatten_array()) { 4412 default_value = InlineTypeNode::default_oop(gvn(), vk); 4413 } 4414 } 4415 } else if (ary_klass->klass()->can_be_inline_array_klass()) { 4416 // Array type is not known, add runtime checks 4417 assert(!ary_klass->klass_is_exact(), "unexpected exact type"); 4418 Node* r = new RegionNode(3); 4419 default_value = new PhiNode(r, TypeInstPtr::BOTTOM); 4420 4421 Node* bol = array_lh_test(klass_node, Klass::_lh_array_tag_flat_value_bit_inplace | Klass::_lh_null_free_array_bit_inplace, Klass::_lh_null_free_array_bit_inplace); 4422 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN); 4423 4424 // Null-free, non-flattened inline type array, initialize with the default value 4425 set_control(_gvn.transform(new IfTrueNode(iff))); 4426 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset())); 4427 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS)); 4428 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset())); 4429 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4430 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset())); 4431 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered); 4432 Node* elem_mirror = load_mirror_from_klass(eklass); 4433 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset)); 4434 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP); 4435 r->init_req(1, control()); 4436 default_value->init_req(1, val); 4437 4438 // Otherwise initialize with all zero 4439 r->init_req(2, _gvn.transform(new IfFalseNode(iff))); 4440 default_value->init_req(2, null()); 4441 4442 set_control(_gvn.transform(r)); 4443 default_value = _gvn.transform(default_value); 4444 } 4445 if (default_value != NULL) { 4446 if (UseCompressedOops) { 4447 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops 4448 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop())); 4449 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value)); 4450 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32))); 4451 raw_default_value = _gvn.transform(new OrLNode(lower, upper)); 4452 } else { 4453 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value)); 4454 } 4455 } 4456 4457 // Create the AllocateArrayNode and its result projections 4458 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 4459 control(), mem, i_o(), 4460 size, klass_node, 4461 initial_slow_test, 4462 length, valid_length_test, 4463 default_value, raw_default_value); 4464 4465 // Cast to correct type. Note that the klass_node may be constant or not, 4466 // and in the latter case the actual array type will be inexact also. 4467 // (This happens via a non-constant argument to inline_native_newArray.) 4468 // In any case, the value of klass_node provides the desired array type. 4469 const TypeInt* length_type = _gvn.find_int_type(length); 4470 if (ary_type->isa_aryptr() && length_type != NULL) { 4471 // Try to get a better type than POS for the size 4472 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4473 } 4474 4475 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 4476 4477 array_ideal_length(alloc, ary_type, true); 4478 return javaoop; 4479 } 4480 4481 // The following "Ideal_foo" functions are placed here because they recognize 4482 // the graph shapes created by the functions immediately above. 4483 4484 //---------------------------Ideal_allocation---------------------------------- 4485 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 4486 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 4487 if (ptr == NULL) { // reduce dumb test in callers 4488 return NULL; 4489 } 4490 4491 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 4492 ptr = bs->step_over_gc_barrier(ptr); 4493 4494 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 4495 ptr = ptr->in(1); 4496 if (ptr == NULL) return NULL; 4497 } 4498 // Return NULL for allocations with several casts: 4499 // j.l.reflect.Array.newInstance(jobject, jint) 4500 // Object.clone() 4501 // to keep more precise type from last cast. 4502 if (ptr->is_Proj()) { 4503 Node* allo = ptr->in(0); 4504 if (allo != NULL && allo->is_Allocate()) { 4505 return allo->as_Allocate(); 4506 } 4507 } 4508 // Report failure to match. 4509 return NULL; 4510 } 4511 4512 // Fancy version which also strips off an offset (and reports it to caller). 4513 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 4514 intptr_t& offset) { 4515 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 4516 if (base == NULL) return NULL; 4517 return Ideal_allocation(base, phase); 4518 } 4519 4520 // Trace Initialize <- Proj[Parm] <- Allocate 4521 AllocateNode* InitializeNode::allocation() { 4522 Node* rawoop = in(InitializeNode::RawAddress); 4523 if (rawoop->is_Proj()) { 4524 Node* alloc = rawoop->in(0); 4525 if (alloc->is_Allocate()) { 4526 return alloc->as_Allocate(); 4527 } 4528 } 4529 return NULL; 4530 } 4531 4532 // Trace Allocate -> Proj[Parm] -> Initialize 4533 InitializeNode* AllocateNode::initialization() { 4534 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 4535 if (rawoop == NULL) return NULL; 4536 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 4537 Node* init = rawoop->fast_out(i); 4538 if (init->is_Initialize()) { 4539 assert(init->as_Initialize()->allocation() == this, "2-way link"); 4540 return init->as_Initialize(); 4541 } 4542 } 4543 return NULL; 4544 } 4545 4546 //----------------------------- loop predicates --------------------------- 4547 4548 //------------------------------add_predicate_impl---------------------------- 4549 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 4550 // Too many traps seen? 4551 if (too_many_traps(reason)) { 4552 #ifdef ASSERT 4553 if (TraceLoopPredicate) { 4554 int tc = C->trap_count(reason); 4555 tty->print("too many traps=%s tcount=%d in ", 4556 Deoptimization::trap_reason_name(reason), tc); 4557 method()->print(); // which method has too many predicate traps 4558 tty->cr(); 4559 } 4560 #endif 4561 // We cannot afford to take more traps here, 4562 // do not generate predicate. 4563 return; 4564 } 4565 4566 Node *cont = _gvn.intcon(1); 4567 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 4568 Node *bol = _gvn.transform(new Conv2BNode(opq)); 4569 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 4570 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 4571 C->add_predicate_opaq(opq); 4572 { 4573 PreserveJVMState pjvms(this); 4574 set_control(iffalse); 4575 inc_sp(nargs); 4576 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 4577 } 4578 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 4579 set_control(iftrue); 4580 } 4581 4582 //------------------------------add_predicate--------------------------------- 4583 void GraphKit::add_empty_predicates(int nargs) { 4584 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding 4585 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of 4586 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate. 4587 if (UseLoopPredicate) { 4588 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs); 4589 } 4590 if (UseProfiledLoopPredicate) { 4591 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs); 4592 } 4593 // loop's limit check predicate should be near the loop. 4594 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 4595 } 4596 4597 void GraphKit::sync_kit(IdealKit& ideal) { 4598 set_all_memory(ideal.merged_memory()); 4599 set_i_o(ideal.i_o()); 4600 set_control(ideal.ctrl()); 4601 } 4602 4603 void GraphKit::final_sync(IdealKit& ideal) { 4604 // Final sync IdealKit and graphKit. 4605 sync_kit(ideal); 4606 } 4607 4608 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) { 4609 Node* len = load_array_length(load_String_value(str, set_ctrl)); 4610 Node* coder = load_String_coder(str, set_ctrl); 4611 // Divide length by 2 if coder is UTF16 4612 return _gvn.transform(new RShiftINode(len, coder)); 4613 } 4614 4615 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) { 4616 int value_offset = java_lang_String::value_offset(); 4617 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4618 false, NULL, Type::Offset(0)); 4619 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4620 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4621 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true), 4622 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0)); 4623 Node* p = basic_plus_adr(str, str, value_offset); 4624 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 4625 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4626 return load; 4627 } 4628 4629 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) { 4630 if (!CompactStrings) { 4631 return intcon(java_lang_String::CODER_UTF16); 4632 } 4633 int coder_offset = java_lang_String::coder_offset(); 4634 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4635 false, NULL, Type::Offset(0)); 4636 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4637 4638 Node* p = basic_plus_adr(str, str, coder_offset); 4639 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE, 4640 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED); 4641 return load; 4642 } 4643 4644 void GraphKit::store_String_value(Node* str, Node* value) { 4645 int value_offset = java_lang_String::value_offset(); 4646 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4647 false, NULL, Type::Offset(0)); 4648 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4649 4650 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type, 4651 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED); 4652 } 4653 4654 void GraphKit::store_String_coder(Node* str, Node* value) { 4655 int coder_offset = java_lang_String::coder_offset(); 4656 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4657 false, NULL, Type::Offset(0)); 4658 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4659 4660 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type, 4661 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED); 4662 } 4663 4664 // Capture src and dst memory state with a MergeMemNode 4665 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4666 if (src_type == dst_type) { 4667 // Types are equal, we don't need a MergeMemNode 4668 return memory(src_type); 4669 } 4670 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4671 record_for_igvn(merge); // fold it up later, if possible 4672 int src_idx = C->get_alias_index(src_type); 4673 int dst_idx = C->get_alias_index(dst_type); 4674 merge->set_memory_at(src_idx, memory(src_idx)); 4675 merge->set_memory_at(dst_idx, memory(dst_idx)); 4676 return merge; 4677 } 4678 4679 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4680 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4681 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4682 // If input and output memory types differ, capture both states to preserve 4683 // the dependency between preceding and subsequent loads/stores. 4684 // For example, the following program: 4685 // StoreB 4686 // compress_string 4687 // LoadB 4688 // has this memory graph (use->def): 4689 // LoadB -> compress_string -> CharMem 4690 // ... -> StoreB -> ByteMem 4691 // The intrinsic hides the dependency between LoadB and StoreB, causing 4692 // the load to read from memory not containing the result of the StoreB. 4693 // The correct memory graph should look like this: 4694 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4695 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4696 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4697 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str))); 4698 set_memory(res_mem, TypeAryPtr::BYTES); 4699 return str; 4700 } 4701 4702 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4703 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4704 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4705 // Capture src and dst memory (see comment in 'compress_string'). 4706 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4707 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4708 set_memory(_gvn.transform(str), dst_type); 4709 } 4710 4711 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4712 /** 4713 * int i_char = start; 4714 * for (int i_byte = 0; i_byte < count; i_byte++) { 4715 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4716 * } 4717 */ 4718 add_empty_predicates(); 4719 C->set_has_loops(true); 4720 4721 RegionNode* head = new RegionNode(3); 4722 head->init_req(1, control()); 4723 gvn().set_type(head, Type::CONTROL); 4724 record_for_igvn(head); 4725 4726 Node* i_byte = new PhiNode(head, TypeInt::INT); 4727 i_byte->init_req(1, intcon(0)); 4728 gvn().set_type(i_byte, TypeInt::INT); 4729 record_for_igvn(i_byte); 4730 4731 Node* i_char = new PhiNode(head, TypeInt::INT); 4732 i_char->init_req(1, start); 4733 gvn().set_type(i_char, TypeInt::INT); 4734 record_for_igvn(i_char); 4735 4736 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4737 gvn().set_type(mem, Type::MEMORY); 4738 record_for_igvn(mem); 4739 set_control(head); 4740 set_memory(mem, TypeAryPtr::BYTES); 4741 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true); 4742 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4743 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4744 false, false, true /* mismatched */); 4745 4746 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4747 head->init_req(2, IfTrue(iff)); 4748 mem->init_req(2, st); 4749 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4750 i_char->init_req(2, AddI(i_char, intcon(2))); 4751 4752 set_control(IfFalse(iff)); 4753 set_memory(st, TypeAryPtr::BYTES); 4754 } 4755 4756 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4757 if (!field->is_constant()) { 4758 return NULL; // Field not marked as constant. 4759 } 4760 ciInstance* holder = NULL; 4761 if (!field->is_static()) { 4762 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4763 if (const_oop != NULL && const_oop->is_instance()) { 4764 holder = const_oop->as_instance(); 4765 } 4766 } 4767 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4768 /*is_unsigned_load=*/false); 4769 if (con_type != NULL) { 4770 Node* con = makecon(con_type); 4771 if (field->type()->is_inlinetype()) { 4772 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free()); 4773 } else if (con_type->is_inlinetypeptr()) { 4774 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free()); 4775 } 4776 return con; 4777 } 4778 return NULL; 4779 } 4780 4781 //---------------------------load_mirror_from_klass---------------------------- 4782 // Given a klass oop, load its java mirror (a java.lang.Class oop). 4783 Node* GraphKit::load_mirror_from_klass(Node* klass) { 4784 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 4785 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4786 // mirror = ((OopHandle)mirror)->resolve(); 4787 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE); 4788 }