1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciUtilities.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/g1/g1BarrierSet.hpp" 29 #include "gc/g1/g1CardTable.hpp" 30 #include "gc/g1/heapRegion.hpp" 31 #include "gc/shared/barrierSet.hpp" 32 #include "gc/shared/cardTable.hpp" 33 #include "gc/shared/cardTableBarrierSet.hpp" 34 #include "gc/shared/collectedHeap.hpp" 35 #include "memory/resourceArea.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/castnode.hpp" 38 #include "opto/convertnode.hpp" 39 #include "opto/graphKit.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/intrinsicnode.hpp" 42 #include "opto/locknode.hpp" 43 #include "opto/machnode.hpp" 44 #include "opto/opaquenode.hpp" 45 #include "opto/parse.hpp" 46 #include "opto/rootnode.hpp" 47 #include "opto/runtime.hpp" 48 #include "runtime/deoptimization.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 51 //----------------------------GraphKit----------------------------------------- 52 // Main utility constructor. 53 GraphKit::GraphKit(JVMState* jvms) 54 : Phase(Phase::Parser), 55 _env(C->env()), 56 _gvn(*C->initial_gvn()) 57 { 58 _exceptions = jvms->map()->next_exception(); 59 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 60 set_jvms(jvms); 61 } 62 63 // Private constructor for parser. 64 GraphKit::GraphKit() 65 : Phase(Phase::Parser), 66 _env(C->env()), 67 _gvn(*C->initial_gvn()) 68 { 69 _exceptions = NULL; 70 set_map(NULL); 71 debug_only(_sp = -99); 72 debug_only(set_bci(-99)); 73 } 74 75 76 77 //---------------------------clean_stack--------------------------------------- 78 // Clear away rubbish from the stack area of the JVM state. 79 // This destroys any arguments that may be waiting on the stack. 80 void GraphKit::clean_stack(int from_sp) { 81 SafePointNode* map = this->map(); 82 JVMState* jvms = this->jvms(); 83 int stk_size = jvms->stk_size(); 84 int stkoff = jvms->stkoff(); 85 Node* top = this->top(); 86 for (int i = from_sp; i < stk_size; i++) { 87 if (map->in(stkoff + i) != top) { 88 map->set_req(stkoff + i, top); 89 } 90 } 91 } 92 93 94 //--------------------------------sync_jvms----------------------------------- 95 // Make sure our current jvms agrees with our parse state. 96 JVMState* GraphKit::sync_jvms() const { 97 JVMState* jvms = this->jvms(); 98 jvms->set_bci(bci()); // Record the new bci in the JVMState 99 jvms->set_sp(sp()); // Record the new sp in the JVMState 100 assert(jvms_in_sync(), "jvms is now in sync"); 101 return jvms; 102 } 103 104 //--------------------------------sync_jvms_for_reexecute--------------------- 105 // Make sure our current jvms agrees with our parse state. This version 106 // uses the reexecute_sp for reexecuting bytecodes. 107 JVMState* GraphKit::sync_jvms_for_reexecute() { 108 JVMState* jvms = this->jvms(); 109 jvms->set_bci(bci()); // Record the new bci in the JVMState 110 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 111 return jvms; 112 } 113 114 #ifdef ASSERT 115 bool GraphKit::jvms_in_sync() const { 116 Parse* parse = is_Parse(); 117 if (parse == NULL) { 118 if (bci() != jvms()->bci()) return false; 119 if (sp() != (int)jvms()->sp()) return false; 120 return true; 121 } 122 if (jvms()->method() != parse->method()) return false; 123 if (jvms()->bci() != parse->bci()) return false; 124 int jvms_sp = jvms()->sp(); 125 if (jvms_sp != parse->sp()) return false; 126 int jvms_depth = jvms()->depth(); 127 if (jvms_depth != parse->depth()) return false; 128 return true; 129 } 130 131 // Local helper checks for special internal merge points 132 // used to accumulate and merge exception states. 133 // They are marked by the region's in(0) edge being the map itself. 134 // Such merge points must never "escape" into the parser at large, 135 // until they have been handed to gvn.transform. 136 static bool is_hidden_merge(Node* reg) { 137 if (reg == NULL) return false; 138 if (reg->is_Phi()) { 139 reg = reg->in(0); 140 if (reg == NULL) return false; 141 } 142 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 143 } 144 145 void GraphKit::verify_map() const { 146 if (map() == NULL) return; // null map is OK 147 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 148 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 149 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 150 } 151 152 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 153 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 154 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 155 } 156 #endif 157 158 //---------------------------stop_and_kill_map--------------------------------- 159 // Set _map to NULL, signalling a stop to further bytecode execution. 160 // First smash the current map's control to a constant, to mark it dead. 161 void GraphKit::stop_and_kill_map() { 162 SafePointNode* dead_map = stop(); 163 if (dead_map != NULL) { 164 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 165 assert(dead_map->is_killed(), "must be so marked"); 166 } 167 } 168 169 170 //--------------------------------stopped-------------------------------------- 171 // Tell if _map is NULL, or control is top. 172 bool GraphKit::stopped() { 173 if (map() == NULL) return true; 174 else if (control() == top()) return true; 175 else return false; 176 } 177 178 179 //-----------------------------has_ex_handler---------------------------------- 180 // Tell if this method or any caller method has exception handlers. 181 bool GraphKit::has_ex_handler() { 182 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 183 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 184 return true; 185 } 186 } 187 return false; 188 } 189 190 //------------------------------save_ex_oop------------------------------------ 191 // Save an exception without blowing stack contents or other JVM state. 192 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 193 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 194 ex_map->add_req(ex_oop); 195 debug_only(verify_exception_state(ex_map)); 196 } 197 198 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 199 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 200 Node* ex_oop = ex_map->in(ex_map->req()-1); 201 if (clear_it) ex_map->del_req(ex_map->req()-1); 202 return ex_oop; 203 } 204 205 //-----------------------------saved_ex_oop------------------------------------ 206 // Recover a saved exception from its map. 207 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 208 return common_saved_ex_oop(ex_map, false); 209 } 210 211 //--------------------------clear_saved_ex_oop--------------------------------- 212 // Erase a previously saved exception from its map. 213 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 214 return common_saved_ex_oop(ex_map, true); 215 } 216 217 #ifdef ASSERT 218 //---------------------------has_saved_ex_oop---------------------------------- 219 // Erase a previously saved exception from its map. 220 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 221 return ex_map->req() == ex_map->jvms()->endoff()+1; 222 } 223 #endif 224 225 //-------------------------make_exception_state-------------------------------- 226 // Turn the current JVM state into an exception state, appending the ex_oop. 227 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 228 sync_jvms(); 229 SafePointNode* ex_map = stop(); // do not manipulate this map any more 230 set_saved_ex_oop(ex_map, ex_oop); 231 return ex_map; 232 } 233 234 235 //--------------------------add_exception_state-------------------------------- 236 // Add an exception to my list of exceptions. 237 void GraphKit::add_exception_state(SafePointNode* ex_map) { 238 if (ex_map == NULL || ex_map->control() == top()) { 239 return; 240 } 241 #ifdef ASSERT 242 verify_exception_state(ex_map); 243 if (has_exceptions()) { 244 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 245 } 246 #endif 247 248 // If there is already an exception of exactly this type, merge with it. 249 // In particular, null-checks and other low-level exceptions common up here. 250 Node* ex_oop = saved_ex_oop(ex_map); 251 const Type* ex_type = _gvn.type(ex_oop); 252 if (ex_oop == top()) { 253 // No action needed. 254 return; 255 } 256 assert(ex_type->isa_instptr(), "exception must be an instance"); 257 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 258 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 259 // We check sp also because call bytecodes can generate exceptions 260 // both before and after arguments are popped! 261 if (ex_type2 == ex_type 262 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 263 combine_exception_states(ex_map, e2); 264 return; 265 } 266 } 267 268 // No pre-existing exception of the same type. Chain it on the list. 269 push_exception_state(ex_map); 270 } 271 272 //-----------------------add_exception_states_from----------------------------- 273 void GraphKit::add_exception_states_from(JVMState* jvms) { 274 SafePointNode* ex_map = jvms->map()->next_exception(); 275 if (ex_map != NULL) { 276 jvms->map()->set_next_exception(NULL); 277 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 278 next_map = ex_map->next_exception(); 279 ex_map->set_next_exception(NULL); 280 add_exception_state(ex_map); 281 } 282 } 283 } 284 285 //-----------------------transfer_exceptions_into_jvms------------------------- 286 JVMState* GraphKit::transfer_exceptions_into_jvms() { 287 if (map() == NULL) { 288 // We need a JVMS to carry the exceptions, but the map has gone away. 289 // Create a scratch JVMS, cloned from any of the exception states... 290 if (has_exceptions()) { 291 _map = _exceptions; 292 _map = clone_map(); 293 _map->set_next_exception(NULL); 294 clear_saved_ex_oop(_map); 295 debug_only(verify_map()); 296 } else { 297 // ...or created from scratch 298 JVMState* jvms = new (C) JVMState(_method, NULL); 299 jvms->set_bci(_bci); 300 jvms->set_sp(_sp); 301 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 302 set_jvms(jvms); 303 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 304 set_all_memory(top()); 305 while (map()->req() < jvms->endoff()) map()->add_req(top()); 306 } 307 // (This is a kludge, in case you didn't notice.) 308 set_control(top()); 309 } 310 JVMState* jvms = sync_jvms(); 311 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 312 jvms->map()->set_next_exception(_exceptions); 313 _exceptions = NULL; // done with this set of exceptions 314 return jvms; 315 } 316 317 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 318 assert(is_hidden_merge(dstphi), "must be a special merge node"); 319 assert(is_hidden_merge(srcphi), "must be a special merge node"); 320 uint limit = srcphi->req(); 321 for (uint i = PhiNode::Input; i < limit; i++) { 322 dstphi->add_req(srcphi->in(i)); 323 } 324 } 325 static inline void add_one_req(Node* dstphi, Node* src) { 326 assert(is_hidden_merge(dstphi), "must be a special merge node"); 327 assert(!is_hidden_merge(src), "must not be a special merge node"); 328 dstphi->add_req(src); 329 } 330 331 //-----------------------combine_exception_states------------------------------ 332 // This helper function combines exception states by building phis on a 333 // specially marked state-merging region. These regions and phis are 334 // untransformed, and can build up gradually. The region is marked by 335 // having a control input of its exception map, rather than NULL. Such 336 // regions do not appear except in this function, and in use_exception_state. 337 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 338 if (failing()) return; // dying anyway... 339 JVMState* ex_jvms = ex_map->_jvms; 340 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 341 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 342 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 343 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 344 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 345 assert(ex_map->req() == phi_map->req(), "matching maps"); 346 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 347 Node* hidden_merge_mark = root(); 348 Node* region = phi_map->control(); 349 MergeMemNode* phi_mem = phi_map->merged_memory(); 350 MergeMemNode* ex_mem = ex_map->merged_memory(); 351 if (region->in(0) != hidden_merge_mark) { 352 // The control input is not (yet) a specially-marked region in phi_map. 353 // Make it so, and build some phis. 354 region = new RegionNode(2); 355 _gvn.set_type(region, Type::CONTROL); 356 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 357 region->init_req(1, phi_map->control()); 358 phi_map->set_control(region); 359 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 360 record_for_igvn(io_phi); 361 _gvn.set_type(io_phi, Type::ABIO); 362 phi_map->set_i_o(io_phi); 363 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 364 Node* m = mms.memory(); 365 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 366 record_for_igvn(m_phi); 367 _gvn.set_type(m_phi, Type::MEMORY); 368 mms.set_memory(m_phi); 369 } 370 } 371 372 // Either or both of phi_map and ex_map might already be converted into phis. 373 Node* ex_control = ex_map->control(); 374 // if there is special marking on ex_map also, we add multiple edges from src 375 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 376 // how wide was the destination phi_map, originally? 377 uint orig_width = region->req(); 378 379 if (add_multiple) { 380 add_n_reqs(region, ex_control); 381 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 382 } else { 383 // ex_map has no merges, so we just add single edges everywhere 384 add_one_req(region, ex_control); 385 add_one_req(phi_map->i_o(), ex_map->i_o()); 386 } 387 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 388 if (mms.is_empty()) { 389 // get a copy of the base memory, and patch some inputs into it 390 const TypePtr* adr_type = mms.adr_type(C); 391 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 392 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 393 mms.set_memory(phi); 394 // Prepare to append interesting stuff onto the newly sliced phi: 395 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 396 } 397 // Append stuff from ex_map: 398 if (add_multiple) { 399 add_n_reqs(mms.memory(), mms.memory2()); 400 } else { 401 add_one_req(mms.memory(), mms.memory2()); 402 } 403 } 404 uint limit = ex_map->req(); 405 for (uint i = TypeFunc::Parms; i < limit; i++) { 406 // Skip everything in the JVMS after tos. (The ex_oop follows.) 407 if (i == tos) i = ex_jvms->monoff(); 408 Node* src = ex_map->in(i); 409 Node* dst = phi_map->in(i); 410 if (src != dst) { 411 PhiNode* phi; 412 if (dst->in(0) != region) { 413 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 414 record_for_igvn(phi); 415 _gvn.set_type(phi, phi->type()); 416 phi_map->set_req(i, dst); 417 // Prepare to append interesting stuff onto the new phi: 418 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 419 } else { 420 assert(dst->is_Phi(), "nobody else uses a hidden region"); 421 phi = dst->as_Phi(); 422 } 423 if (add_multiple && src->in(0) == ex_control) { 424 // Both are phis. 425 add_n_reqs(dst, src); 426 } else { 427 while (dst->req() < region->req()) add_one_req(dst, src); 428 } 429 const Type* srctype = _gvn.type(src); 430 if (phi->type() != srctype) { 431 const Type* dsttype = phi->type()->meet_speculative(srctype); 432 if (phi->type() != dsttype) { 433 phi->set_type(dsttype); 434 _gvn.set_type(phi, dsttype); 435 } 436 } 437 } 438 } 439 phi_map->merge_replaced_nodes_with(ex_map); 440 } 441 442 //--------------------------use_exception_state-------------------------------- 443 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 444 if (failing()) { stop(); return top(); } 445 Node* region = phi_map->control(); 446 Node* hidden_merge_mark = root(); 447 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 448 Node* ex_oop = clear_saved_ex_oop(phi_map); 449 if (region->in(0) == hidden_merge_mark) { 450 // Special marking for internal ex-states. Process the phis now. 451 region->set_req(0, region); // now it's an ordinary region 452 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 453 // Note: Setting the jvms also sets the bci and sp. 454 set_control(_gvn.transform(region)); 455 uint tos = jvms()->stkoff() + sp(); 456 for (uint i = 1; i < tos; i++) { 457 Node* x = phi_map->in(i); 458 if (x->in(0) == region) { 459 assert(x->is_Phi(), "expected a special phi"); 460 phi_map->set_req(i, _gvn.transform(x)); 461 } 462 } 463 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 464 Node* x = mms.memory(); 465 if (x->in(0) == region) { 466 assert(x->is_Phi(), "nobody else uses a hidden region"); 467 mms.set_memory(_gvn.transform(x)); 468 } 469 } 470 if (ex_oop->in(0) == region) { 471 assert(ex_oop->is_Phi(), "expected a special phi"); 472 ex_oop = _gvn.transform(ex_oop); 473 } 474 } else { 475 set_jvms(phi_map->jvms()); 476 } 477 478 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 479 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 480 return ex_oop; 481 } 482 483 //---------------------------------java_bc------------------------------------- 484 Bytecodes::Code GraphKit::java_bc() const { 485 ciMethod* method = this->method(); 486 int bci = this->bci(); 487 if (method != NULL && bci != InvocationEntryBci) 488 return method->java_code_at_bci(bci); 489 else 490 return Bytecodes::_illegal; 491 } 492 493 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 494 bool must_throw) { 495 // if the exception capability is set, then we will generate code 496 // to check the JavaThread.should_post_on_exceptions flag to see 497 // if we actually need to report exception events (for this 498 // thread). If we don't need to report exception events, we will 499 // take the normal fast path provided by add_exception_events. If 500 // exception event reporting is enabled for this thread, we will 501 // take the uncommon_trap in the BuildCutout below. 502 503 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 504 Node* jthread = _gvn.transform(new ThreadLocalNode()); 505 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 506 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 507 508 // Test the should_post_on_exceptions_flag vs. 0 509 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 510 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 511 512 // Branch to slow_path if should_post_on_exceptions_flag was true 513 { BuildCutout unless(this, tst, PROB_MAX); 514 // Do not try anything fancy if we're notifying the VM on every throw. 515 // Cf. case Bytecodes::_athrow in parse2.cpp. 516 uncommon_trap(reason, Deoptimization::Action_none, 517 (ciKlass*)NULL, (char*)NULL, must_throw); 518 } 519 520 } 521 522 //------------------------------builtin_throw---------------------------------- 523 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 524 bool must_throw = true; 525 526 if (env()->jvmti_can_post_on_exceptions()) { 527 // check if we must post exception events, take uncommon trap if so 528 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 529 // here if should_post_on_exceptions is false 530 // continue on with the normal codegen 531 } 532 533 // If this particular condition has not yet happened at this 534 // bytecode, then use the uncommon trap mechanism, and allow for 535 // a future recompilation if several traps occur here. 536 // If the throw is hot, try to use a more complicated inline mechanism 537 // which keeps execution inside the compiled code. 538 bool treat_throw_as_hot = false; 539 ciMethodData* md = method()->method_data(); 540 541 if (ProfileTraps) { 542 if (too_many_traps(reason)) { 543 treat_throw_as_hot = true; 544 } 545 // (If there is no MDO at all, assume it is early in 546 // execution, and that any deopts are part of the 547 // startup transient, and don't need to be remembered.) 548 549 // Also, if there is a local exception handler, treat all throws 550 // as hot if there has been at least one in this method. 551 if (C->trap_count(reason) != 0 552 && method()->method_data()->trap_count(reason) != 0 553 && has_ex_handler()) { 554 treat_throw_as_hot = true; 555 } 556 } 557 558 // If this throw happens frequently, an uncommon trap might cause 559 // a performance pothole. If there is a local exception handler, 560 // and if this particular bytecode appears to be deoptimizing often, 561 // let us handle the throw inline, with a preconstructed instance. 562 // Note: If the deopt count has blown up, the uncommon trap 563 // runtime is going to flush this nmethod, not matter what. 564 if (treat_throw_as_hot 565 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 566 // If the throw is local, we use a pre-existing instance and 567 // punt on the backtrace. This would lead to a missing backtrace 568 // (a repeat of 4292742) if the backtrace object is ever asked 569 // for its backtrace. 570 // Fixing this remaining case of 4292742 requires some flavor of 571 // escape analysis. Leave that for the future. 572 ciInstance* ex_obj = NULL; 573 switch (reason) { 574 case Deoptimization::Reason_null_check: 575 ex_obj = env()->NullPointerException_instance(); 576 break; 577 case Deoptimization::Reason_div0_check: 578 ex_obj = env()->ArithmeticException_instance(); 579 break; 580 case Deoptimization::Reason_range_check: 581 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 582 break; 583 case Deoptimization::Reason_class_check: 584 if (java_bc() == Bytecodes::_aastore) { 585 ex_obj = env()->ArrayStoreException_instance(); 586 } else { 587 ex_obj = env()->ClassCastException_instance(); 588 } 589 break; 590 default: 591 break; 592 } 593 if (failing()) { stop(); return; } // exception allocation might fail 594 if (ex_obj != NULL) { 595 // Cheat with a preallocated exception object. 596 if (C->log() != NULL) 597 C->log()->elem("hot_throw preallocated='1' reason='%s'", 598 Deoptimization::trap_reason_name(reason)); 599 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 600 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 601 602 // Clear the detail message of the preallocated exception object. 603 // Weblogic sometimes mutates the detail message of exceptions 604 // using reflection. 605 int offset = java_lang_Throwable::get_detailMessage_offset(); 606 const TypePtr* adr_typ = ex_con->add_offset(offset); 607 608 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 609 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 610 // Conservatively release stores of object references. 611 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 612 613 add_exception_state(make_exception_state(ex_node)); 614 return; 615 } 616 } 617 618 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 619 // It won't be much cheaper than bailing to the interp., since we'll 620 // have to pass up all the debug-info, and the runtime will have to 621 // create the stack trace. 622 623 // Usual case: Bail to interpreter. 624 // Reserve the right to recompile if we haven't seen anything yet. 625 626 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 627 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 628 if (treat_throw_as_hot 629 && (method()->method_data()->trap_recompiled_at(bci(), m) 630 || C->too_many_traps(reason))) { 631 // We cannot afford to take more traps here. Suffer in the interpreter. 632 if (C->log() != NULL) 633 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 634 Deoptimization::trap_reason_name(reason), 635 C->trap_count(reason)); 636 action = Deoptimization::Action_none; 637 } 638 639 // "must_throw" prunes the JVM state to include only the stack, if there 640 // are no local exception handlers. This should cut down on register 641 // allocation time and code size, by drastically reducing the number 642 // of in-edges on the call to the uncommon trap. 643 644 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 645 } 646 647 648 //----------------------------PreserveJVMState--------------------------------- 649 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 650 debug_only(kit->verify_map()); 651 _kit = kit; 652 _map = kit->map(); // preserve the map 653 _sp = kit->sp(); 654 kit->set_map(clone_map ? kit->clone_map() : NULL); 655 #ifdef ASSERT 656 _bci = kit->bci(); 657 Parse* parser = kit->is_Parse(); 658 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 659 _block = block; 660 #endif 661 } 662 PreserveJVMState::~PreserveJVMState() { 663 GraphKit* kit = _kit; 664 #ifdef ASSERT 665 assert(kit->bci() == _bci, "bci must not shift"); 666 Parse* parser = kit->is_Parse(); 667 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 668 assert(block == _block, "block must not shift"); 669 #endif 670 kit->set_map(_map); 671 kit->set_sp(_sp); 672 } 673 674 675 //-----------------------------BuildCutout------------------------------------- 676 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 677 : PreserveJVMState(kit) 678 { 679 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 680 SafePointNode* outer_map = _map; // preserved map is caller's 681 SafePointNode* inner_map = kit->map(); 682 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 683 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 684 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 685 } 686 BuildCutout::~BuildCutout() { 687 GraphKit* kit = _kit; 688 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 689 } 690 691 //---------------------------PreserveReexecuteState---------------------------- 692 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 693 assert(!kit->stopped(), "must call stopped() before"); 694 _kit = kit; 695 _sp = kit->sp(); 696 _reexecute = kit->jvms()->_reexecute; 697 } 698 PreserveReexecuteState::~PreserveReexecuteState() { 699 if (_kit->stopped()) return; 700 _kit->jvms()->_reexecute = _reexecute; 701 _kit->set_sp(_sp); 702 } 703 704 //------------------------------clone_map-------------------------------------- 705 // Implementation of PreserveJVMState 706 // 707 // Only clone_map(...) here. If this function is only used in the 708 // PreserveJVMState class we may want to get rid of this extra 709 // function eventually and do it all there. 710 711 SafePointNode* GraphKit::clone_map() { 712 if (map() == NULL) return NULL; 713 714 // Clone the memory edge first 715 Node* mem = MergeMemNode::make(map()->memory()); 716 gvn().set_type_bottom(mem); 717 718 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 719 JVMState* jvms = this->jvms(); 720 JVMState* clonejvms = jvms->clone_shallow(C); 721 clonemap->set_memory(mem); 722 clonemap->set_jvms(clonejvms); 723 clonejvms->set_map(clonemap); 724 record_for_igvn(clonemap); 725 gvn().set_type_bottom(clonemap); 726 return clonemap; 727 } 728 729 730 //-----------------------------set_map_clone----------------------------------- 731 void GraphKit::set_map_clone(SafePointNode* m) { 732 _map = m; 733 _map = clone_map(); 734 _map->set_next_exception(NULL); 735 debug_only(verify_map()); 736 } 737 738 739 //----------------------------kill_dead_locals--------------------------------- 740 // Detect any locals which are known to be dead, and force them to top. 741 void GraphKit::kill_dead_locals() { 742 // Consult the liveness information for the locals. If any 743 // of them are unused, then they can be replaced by top(). This 744 // should help register allocation time and cut down on the size 745 // of the deoptimization information. 746 747 // This call is made from many of the bytecode handling 748 // subroutines called from the Big Switch in do_one_bytecode. 749 // Every bytecode which might include a slow path is responsible 750 // for killing its dead locals. The more consistent we 751 // are about killing deads, the fewer useless phis will be 752 // constructed for them at various merge points. 753 754 // bci can be -1 (InvocationEntryBci). We return the entry 755 // liveness for the method. 756 757 if (method() == NULL || method()->code_size() == 0) { 758 // We are building a graph for a call to a native method. 759 // All locals are live. 760 return; 761 } 762 763 ResourceMark rm; 764 765 // Consult the liveness information for the locals. If any 766 // of them are unused, then they can be replaced by top(). This 767 // should help register allocation time and cut down on the size 768 // of the deoptimization information. 769 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 770 771 int len = (int)live_locals.size(); 772 assert(len <= jvms()->loc_size(), "too many live locals"); 773 for (int local = 0; local < len; local++) { 774 if (!live_locals.at(local)) { 775 set_local(local, top()); 776 } 777 } 778 } 779 780 #ifdef ASSERT 781 //-------------------------dead_locals_are_killed------------------------------ 782 // Return true if all dead locals are set to top in the map. 783 // Used to assert "clean" debug info at various points. 784 bool GraphKit::dead_locals_are_killed() { 785 if (method() == NULL || method()->code_size() == 0) { 786 // No locals need to be dead, so all is as it should be. 787 return true; 788 } 789 790 // Make sure somebody called kill_dead_locals upstream. 791 ResourceMark rm; 792 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 793 if (jvms->loc_size() == 0) continue; // no locals to consult 794 SafePointNode* map = jvms->map(); 795 ciMethod* method = jvms->method(); 796 int bci = jvms->bci(); 797 if (jvms == this->jvms()) { 798 bci = this->bci(); // it might not yet be synched 799 } 800 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 801 int len = (int)live_locals.size(); 802 if (!live_locals.is_valid() || len == 0) 803 // This method is trivial, or is poisoned by a breakpoint. 804 return true; 805 assert(len == jvms->loc_size(), "live map consistent with locals map"); 806 for (int local = 0; local < len; local++) { 807 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 808 if (PrintMiscellaneous && (Verbose || WizardMode)) { 809 tty->print_cr("Zombie local %d: ", local); 810 jvms->dump(); 811 } 812 return false; 813 } 814 } 815 } 816 return true; 817 } 818 819 #endif //ASSERT 820 821 // Helper function for enforcing certain bytecodes to reexecute if 822 // deoptimization happens 823 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 824 ciMethod* cur_method = jvms->method(); 825 int cur_bci = jvms->bci(); 826 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 827 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 828 return Interpreter::bytecode_should_reexecute(code) || 829 (is_anewarray && code == Bytecodes::_multianewarray); 830 // Reexecute _multianewarray bytecode which was replaced with 831 // sequence of [a]newarray. See Parse::do_multianewarray(). 832 // 833 // Note: interpreter should not have it set since this optimization 834 // is limited by dimensions and guarded by flag so in some cases 835 // multianewarray() runtime calls will be generated and 836 // the bytecode should not be reexecutes (stack will not be reset). 837 } else 838 return false; 839 } 840 841 // Helper function for adding JVMState and debug information to node 842 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 843 // Add the safepoint edges to the call (or other safepoint). 844 845 // Make sure dead locals are set to top. This 846 // should help register allocation time and cut down on the size 847 // of the deoptimization information. 848 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 849 850 // Walk the inline list to fill in the correct set of JVMState's 851 // Also fill in the associated edges for each JVMState. 852 853 // If the bytecode needs to be reexecuted we need to put 854 // the arguments back on the stack. 855 const bool should_reexecute = jvms()->should_reexecute(); 856 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 857 858 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 859 // undefined if the bci is different. This is normal for Parse but it 860 // should not happen for LibraryCallKit because only one bci is processed. 861 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 862 "in LibraryCallKit the reexecute bit should not change"); 863 864 // If we are guaranteed to throw, we can prune everything but the 865 // input to the current bytecode. 866 bool can_prune_locals = false; 867 uint stack_slots_not_pruned = 0; 868 int inputs = 0, depth = 0; 869 if (must_throw) { 870 assert(method() == youngest_jvms->method(), "sanity"); 871 if (compute_stack_effects(inputs, depth)) { 872 can_prune_locals = true; 873 stack_slots_not_pruned = inputs; 874 } 875 } 876 877 if (env()->should_retain_local_variables()) { 878 // At any safepoint, this method can get breakpointed, which would 879 // then require an immediate deoptimization. 880 can_prune_locals = false; // do not prune locals 881 stack_slots_not_pruned = 0; 882 } 883 884 // do not scribble on the input jvms 885 JVMState* out_jvms = youngest_jvms->clone_deep(C); 886 call->set_jvms(out_jvms); // Start jvms list for call node 887 888 // For a known set of bytecodes, the interpreter should reexecute them if 889 // deoptimization happens. We set the reexecute state for them here 890 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 891 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 892 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 893 } 894 895 // Presize the call: 896 DEBUG_ONLY(uint non_debug_edges = call->req()); 897 call->add_req_batch(top(), youngest_jvms->debug_depth()); 898 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 899 900 // Set up edges so that the call looks like this: 901 // Call [state:] ctl io mem fptr retadr 902 // [parms:] parm0 ... parmN 903 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 904 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 905 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 906 // Note that caller debug info precedes callee debug info. 907 908 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 909 uint debug_ptr = call->req(); 910 911 // Loop over the map input edges associated with jvms, add them 912 // to the call node, & reset all offsets to match call node array. 913 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 914 uint debug_end = debug_ptr; 915 uint debug_start = debug_ptr - in_jvms->debug_size(); 916 debug_ptr = debug_start; // back up the ptr 917 918 uint p = debug_start; // walks forward in [debug_start, debug_end) 919 uint j, k, l; 920 SafePointNode* in_map = in_jvms->map(); 921 out_jvms->set_map(call); 922 923 if (can_prune_locals) { 924 assert(in_jvms->method() == out_jvms->method(), "sanity"); 925 // If the current throw can reach an exception handler in this JVMS, 926 // then we must keep everything live that can reach that handler. 927 // As a quick and dirty approximation, we look for any handlers at all. 928 if (in_jvms->method()->has_exception_handlers()) { 929 can_prune_locals = false; 930 } 931 } 932 933 // Add the Locals 934 k = in_jvms->locoff(); 935 l = in_jvms->loc_size(); 936 out_jvms->set_locoff(p); 937 if (!can_prune_locals) { 938 for (j = 0; j < l; j++) 939 call->set_req(p++, in_map->in(k+j)); 940 } else { 941 p += l; // already set to top above by add_req_batch 942 } 943 944 // Add the Expression Stack 945 k = in_jvms->stkoff(); 946 l = in_jvms->sp(); 947 out_jvms->set_stkoff(p); 948 if (!can_prune_locals) { 949 for (j = 0; j < l; j++) 950 call->set_req(p++, in_map->in(k+j)); 951 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 952 // Divide stack into {S0,...,S1}, where S0 is set to top. 953 uint s1 = stack_slots_not_pruned; 954 stack_slots_not_pruned = 0; // for next iteration 955 if (s1 > l) s1 = l; 956 uint s0 = l - s1; 957 p += s0; // skip the tops preinstalled by add_req_batch 958 for (j = s0; j < l; j++) 959 call->set_req(p++, in_map->in(k+j)); 960 } else { 961 p += l; // already set to top above by add_req_batch 962 } 963 964 // Add the Monitors 965 k = in_jvms->monoff(); 966 l = in_jvms->mon_size(); 967 out_jvms->set_monoff(p); 968 for (j = 0; j < l; j++) 969 call->set_req(p++, in_map->in(k+j)); 970 971 // Copy any scalar object fields. 972 k = in_jvms->scloff(); 973 l = in_jvms->scl_size(); 974 out_jvms->set_scloff(p); 975 for (j = 0; j < l; j++) 976 call->set_req(p++, in_map->in(k+j)); 977 978 // Finish the new jvms. 979 out_jvms->set_endoff(p); 980 981 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 982 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 983 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 984 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 985 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 986 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 987 988 // Update the two tail pointers in parallel. 989 out_jvms = out_jvms->caller(); 990 in_jvms = in_jvms->caller(); 991 } 992 993 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 994 995 // Test the correctness of JVMState::debug_xxx accessors: 996 assert(call->jvms()->debug_start() == non_debug_edges, ""); 997 assert(call->jvms()->debug_end() == call->req(), ""); 998 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 999 } 1000 1001 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1002 Bytecodes::Code code = java_bc(); 1003 if (code == Bytecodes::_wide) { 1004 code = method()->java_code_at_bci(bci() + 1); 1005 } 1006 1007 BasicType rtype = T_ILLEGAL; 1008 int rsize = 0; 1009 1010 if (code != Bytecodes::_illegal) { 1011 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1012 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1013 if (rtype < T_CONFLICT) 1014 rsize = type2size[rtype]; 1015 } 1016 1017 switch (code) { 1018 case Bytecodes::_illegal: 1019 return false; 1020 1021 case Bytecodes::_ldc: 1022 case Bytecodes::_ldc_w: 1023 case Bytecodes::_ldc2_w: 1024 inputs = 0; 1025 break; 1026 1027 case Bytecodes::_dup: inputs = 1; break; 1028 case Bytecodes::_dup_x1: inputs = 2; break; 1029 case Bytecodes::_dup_x2: inputs = 3; break; 1030 case Bytecodes::_dup2: inputs = 2; break; 1031 case Bytecodes::_dup2_x1: inputs = 3; break; 1032 case Bytecodes::_dup2_x2: inputs = 4; break; 1033 case Bytecodes::_swap: inputs = 2; break; 1034 case Bytecodes::_arraylength: inputs = 1; break; 1035 1036 case Bytecodes::_getstatic: 1037 case Bytecodes::_putstatic: 1038 case Bytecodes::_getfield: 1039 case Bytecodes::_putfield: 1040 { 1041 bool ignored_will_link; 1042 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1043 int size = field->type()->size(); 1044 bool is_get = (depth >= 0), is_static = (depth & 1); 1045 inputs = (is_static ? 0 : 1); 1046 if (is_get) { 1047 depth = size - inputs; 1048 } else { 1049 inputs += size; // putxxx pops the value from the stack 1050 depth = - inputs; 1051 } 1052 } 1053 break; 1054 1055 case Bytecodes::_invokevirtual: 1056 case Bytecodes::_invokespecial: 1057 case Bytecodes::_invokestatic: 1058 case Bytecodes::_invokedynamic: 1059 case Bytecodes::_invokeinterface: 1060 { 1061 bool ignored_will_link; 1062 ciSignature* declared_signature = NULL; 1063 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1064 assert(declared_signature != NULL, "cannot be null"); 1065 inputs = declared_signature->arg_size_for_bc(code); 1066 int size = declared_signature->return_type()->size(); 1067 depth = size - inputs; 1068 } 1069 break; 1070 1071 case Bytecodes::_multianewarray: 1072 { 1073 ciBytecodeStream iter(method()); 1074 iter.reset_to_bci(bci()); 1075 iter.next(); 1076 inputs = iter.get_dimensions(); 1077 assert(rsize == 1, ""); 1078 depth = rsize - inputs; 1079 } 1080 break; 1081 1082 case Bytecodes::_ireturn: 1083 case Bytecodes::_lreturn: 1084 case Bytecodes::_freturn: 1085 case Bytecodes::_dreturn: 1086 case Bytecodes::_areturn: 1087 assert(rsize == -depth, ""); 1088 inputs = rsize; 1089 break; 1090 1091 case Bytecodes::_jsr: 1092 case Bytecodes::_jsr_w: 1093 inputs = 0; 1094 depth = 1; // S.B. depth=1, not zero 1095 break; 1096 1097 default: 1098 // bytecode produces a typed result 1099 inputs = rsize - depth; 1100 assert(inputs >= 0, ""); 1101 break; 1102 } 1103 1104 #ifdef ASSERT 1105 // spot check 1106 int outputs = depth + inputs; 1107 assert(outputs >= 0, "sanity"); 1108 switch (code) { 1109 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1110 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1111 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1112 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1113 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1114 default: break; 1115 } 1116 #endif //ASSERT 1117 1118 return true; 1119 } 1120 1121 1122 1123 //------------------------------basic_plus_adr--------------------------------- 1124 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1125 // short-circuit a common case 1126 if (offset == intcon(0)) return ptr; 1127 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1128 } 1129 1130 Node* GraphKit::ConvI2L(Node* offset) { 1131 // short-circuit a common case 1132 jint offset_con = find_int_con(offset, Type::OffsetBot); 1133 if (offset_con != Type::OffsetBot) { 1134 return longcon((jlong) offset_con); 1135 } 1136 return _gvn.transform( new ConvI2LNode(offset)); 1137 } 1138 1139 Node* GraphKit::ConvI2UL(Node* offset) { 1140 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1141 if (offset_con != (juint) Type::OffsetBot) { 1142 return longcon((julong) offset_con); 1143 } 1144 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1145 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1146 return _gvn.transform( new AndLNode(conv, mask) ); 1147 } 1148 1149 Node* GraphKit::ConvL2I(Node* offset) { 1150 // short-circuit a common case 1151 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1152 if (offset_con != (jlong)Type::OffsetBot) { 1153 return intcon((int) offset_con); 1154 } 1155 return _gvn.transform( new ConvL2INode(offset)); 1156 } 1157 1158 //-------------------------load_object_klass----------------------------------- 1159 Node* GraphKit::load_object_klass(Node* obj) { 1160 // Special-case a fresh allocation to avoid building nodes: 1161 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1162 if (akls != NULL) return akls; 1163 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1164 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1165 } 1166 1167 //-------------------------load_array_length----------------------------------- 1168 Node* GraphKit::load_array_length(Node* array) { 1169 // Special-case a fresh allocation to avoid building nodes: 1170 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1171 Node *alen; 1172 if (alloc == NULL) { 1173 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1174 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1175 } else { 1176 alen = alloc->Ideal_length(); 1177 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1178 if (ccast != alen) { 1179 alen = _gvn.transform(ccast); 1180 } 1181 } 1182 return alen; 1183 } 1184 1185 //------------------------------do_null_check---------------------------------- 1186 // Helper function to do a NULL pointer check. Returned value is 1187 // the incoming address with NULL casted away. You are allowed to use the 1188 // not-null value only if you are control dependent on the test. 1189 #ifndef PRODUCT 1190 extern int explicit_null_checks_inserted, 1191 explicit_null_checks_elided; 1192 #endif 1193 Node* GraphKit::null_check_common(Node* value, BasicType type, 1194 // optional arguments for variations: 1195 bool assert_null, 1196 Node* *null_control, 1197 bool speculative) { 1198 assert(!assert_null || null_control == NULL, "not both at once"); 1199 if (stopped()) return top(); 1200 NOT_PRODUCT(explicit_null_checks_inserted++); 1201 1202 // Construct NULL check 1203 Node *chk = NULL; 1204 switch(type) { 1205 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1206 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1207 case T_ARRAY : // fall through 1208 type = T_OBJECT; // simplify further tests 1209 case T_OBJECT : { 1210 const Type *t = _gvn.type( value ); 1211 1212 const TypeOopPtr* tp = t->isa_oopptr(); 1213 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1214 // Only for do_null_check, not any of its siblings: 1215 && !assert_null && null_control == NULL) { 1216 // Usually, any field access or invocation on an unloaded oop type 1217 // will simply fail to link, since the statically linked class is 1218 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1219 // the static class is loaded but the sharper oop type is not. 1220 // Rather than checking for this obscure case in lots of places, 1221 // we simply observe that a null check on an unloaded class 1222 // will always be followed by a nonsense operation, so we 1223 // can just issue the uncommon trap here. 1224 // Our access to the unloaded class will only be correct 1225 // after it has been loaded and initialized, which requires 1226 // a trip through the interpreter. 1227 #ifndef PRODUCT 1228 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1229 #endif 1230 uncommon_trap(Deoptimization::Reason_unloaded, 1231 Deoptimization::Action_reinterpret, 1232 tp->klass(), "!loaded"); 1233 return top(); 1234 } 1235 1236 if (assert_null) { 1237 // See if the type is contained in NULL_PTR. 1238 // If so, then the value is already null. 1239 if (t->higher_equal(TypePtr::NULL_PTR)) { 1240 NOT_PRODUCT(explicit_null_checks_elided++); 1241 return value; // Elided null assert quickly! 1242 } 1243 } else { 1244 // See if mixing in the NULL pointer changes type. 1245 // If so, then the NULL pointer was not allowed in the original 1246 // type. In other words, "value" was not-null. 1247 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1248 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1249 NOT_PRODUCT(explicit_null_checks_elided++); 1250 return value; // Elided null check quickly! 1251 } 1252 } 1253 chk = new CmpPNode( value, null() ); 1254 break; 1255 } 1256 1257 default: 1258 fatal("unexpected type: %s", type2name(type)); 1259 } 1260 assert(chk != NULL, "sanity check"); 1261 chk = _gvn.transform(chk); 1262 1263 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1264 BoolNode *btst = new BoolNode( chk, btest); 1265 Node *tst = _gvn.transform( btst ); 1266 1267 //----------- 1268 // if peephole optimizations occurred, a prior test existed. 1269 // If a prior test existed, maybe it dominates as we can avoid this test. 1270 if (tst != btst && type == T_OBJECT) { 1271 // At this point we want to scan up the CFG to see if we can 1272 // find an identical test (and so avoid this test altogether). 1273 Node *cfg = control(); 1274 int depth = 0; 1275 while( depth < 16 ) { // Limit search depth for speed 1276 if( cfg->Opcode() == Op_IfTrue && 1277 cfg->in(0)->in(1) == tst ) { 1278 // Found prior test. Use "cast_not_null" to construct an identical 1279 // CastPP (and hence hash to) as already exists for the prior test. 1280 // Return that casted value. 1281 if (assert_null) { 1282 replace_in_map(value, null()); 1283 return null(); // do not issue the redundant test 1284 } 1285 Node *oldcontrol = control(); 1286 set_control(cfg); 1287 Node *res = cast_not_null(value); 1288 set_control(oldcontrol); 1289 NOT_PRODUCT(explicit_null_checks_elided++); 1290 return res; 1291 } 1292 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1293 if (cfg == NULL) break; // Quit at region nodes 1294 depth++; 1295 } 1296 } 1297 1298 //----------- 1299 // Branch to failure if null 1300 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1301 Deoptimization::DeoptReason reason; 1302 if (assert_null) { 1303 reason = Deoptimization::reason_null_assert(speculative); 1304 } else if (type == T_OBJECT) { 1305 reason = Deoptimization::reason_null_check(speculative); 1306 } else { 1307 reason = Deoptimization::Reason_div0_check; 1308 } 1309 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1310 // ciMethodData::has_trap_at will return a conservative -1 if any 1311 // must-be-null assertion has failed. This could cause performance 1312 // problems for a method after its first do_null_assert failure. 1313 // Consider using 'Reason_class_check' instead? 1314 1315 // To cause an implicit null check, we set the not-null probability 1316 // to the maximum (PROB_MAX). For an explicit check the probability 1317 // is set to a smaller value. 1318 if (null_control != NULL || too_many_traps(reason)) { 1319 // probability is less likely 1320 ok_prob = PROB_LIKELY_MAG(3); 1321 } else if (!assert_null && 1322 (ImplicitNullCheckThreshold > 0) && 1323 method() != NULL && 1324 (method()->method_data()->trap_count(reason) 1325 >= (uint)ImplicitNullCheckThreshold)) { 1326 ok_prob = PROB_LIKELY_MAG(3); 1327 } 1328 1329 if (null_control != NULL) { 1330 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1331 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1332 set_control( _gvn.transform( new IfTrueNode(iff))); 1333 #ifndef PRODUCT 1334 if (null_true == top()) { 1335 explicit_null_checks_elided++; 1336 } 1337 #endif 1338 (*null_control) = null_true; 1339 } else { 1340 BuildCutout unless(this, tst, ok_prob); 1341 // Check for optimizer eliding test at parse time 1342 if (stopped()) { 1343 // Failure not possible; do not bother making uncommon trap. 1344 NOT_PRODUCT(explicit_null_checks_elided++); 1345 } else if (assert_null) { 1346 uncommon_trap(reason, 1347 Deoptimization::Action_make_not_entrant, 1348 NULL, "assert_null"); 1349 } else { 1350 replace_in_map(value, zerocon(type)); 1351 builtin_throw(reason); 1352 } 1353 } 1354 1355 // Must throw exception, fall-thru not possible? 1356 if (stopped()) { 1357 return top(); // No result 1358 } 1359 1360 if (assert_null) { 1361 // Cast obj to null on this path. 1362 replace_in_map(value, zerocon(type)); 1363 return zerocon(type); 1364 } 1365 1366 // Cast obj to not-null on this path, if there is no null_control. 1367 // (If there is a null_control, a non-null value may come back to haunt us.) 1368 if (type == T_OBJECT) { 1369 Node* cast = cast_not_null(value, false); 1370 if (null_control == NULL || (*null_control) == top()) 1371 replace_in_map(value, cast); 1372 value = cast; 1373 } 1374 1375 return value; 1376 } 1377 1378 1379 //------------------------------cast_not_null---------------------------------- 1380 // Cast obj to not-null on this path 1381 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1382 const Type *t = _gvn.type(obj); 1383 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1384 // Object is already not-null? 1385 if( t == t_not_null ) return obj; 1386 1387 Node *cast = new CastPPNode(obj,t_not_null); 1388 cast->init_req(0, control()); 1389 cast = _gvn.transform( cast ); 1390 1391 // Scan for instances of 'obj' in the current JVM mapping. 1392 // These instances are known to be not-null after the test. 1393 if (do_replace_in_map) 1394 replace_in_map(obj, cast); 1395 1396 return cast; // Return casted value 1397 } 1398 1399 // Sometimes in intrinsics, we implicitly know an object is not null 1400 // (there's no actual null check) so we can cast it to not null. In 1401 // the course of optimizations, the input to the cast can become null. 1402 // In that case that data path will die and we need the control path 1403 // to become dead as well to keep the graph consistent. So we have to 1404 // add a check for null for which one branch can't be taken. It uses 1405 // an Opaque4 node that will cause the check to be removed after loop 1406 // opts so the test goes away and the compiled code doesn't execute a 1407 // useless check. 1408 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1409 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1410 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1411 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1412 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1413 _gvn.set_type(iff, iff->Value(&_gvn)); 1414 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1415 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1416 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1417 C->root()->add_req(halt); 1418 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1419 set_control(if_t); 1420 return cast_not_null(value, do_replace_in_map); 1421 } 1422 1423 1424 //--------------------------replace_in_map------------------------------------- 1425 void GraphKit::replace_in_map(Node* old, Node* neww) { 1426 if (old == neww) { 1427 return; 1428 } 1429 1430 map()->replace_edge(old, neww); 1431 1432 // Note: This operation potentially replaces any edge 1433 // on the map. This includes locals, stack, and monitors 1434 // of the current (innermost) JVM state. 1435 1436 // don't let inconsistent types from profiling escape this 1437 // method 1438 1439 const Type* told = _gvn.type(old); 1440 const Type* tnew = _gvn.type(neww); 1441 1442 if (!tnew->higher_equal(told)) { 1443 return; 1444 } 1445 1446 map()->record_replaced_node(old, neww); 1447 } 1448 1449 1450 //============================================================================= 1451 //--------------------------------memory--------------------------------------- 1452 Node* GraphKit::memory(uint alias_idx) { 1453 MergeMemNode* mem = merged_memory(); 1454 Node* p = mem->memory_at(alias_idx); 1455 _gvn.set_type(p, Type::MEMORY); // must be mapped 1456 return p; 1457 } 1458 1459 //-----------------------------reset_memory------------------------------------ 1460 Node* GraphKit::reset_memory() { 1461 Node* mem = map()->memory(); 1462 // do not use this node for any more parsing! 1463 debug_only( map()->set_memory((Node*)NULL) ); 1464 return _gvn.transform( mem ); 1465 } 1466 1467 //------------------------------set_all_memory--------------------------------- 1468 void GraphKit::set_all_memory(Node* newmem) { 1469 Node* mergemem = MergeMemNode::make(newmem); 1470 gvn().set_type_bottom(mergemem); 1471 map()->set_memory(mergemem); 1472 } 1473 1474 //------------------------------set_all_memory_call---------------------------- 1475 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1476 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1477 set_all_memory(newmem); 1478 } 1479 1480 //============================================================================= 1481 // 1482 // parser factory methods for MemNodes 1483 // 1484 // These are layered on top of the factory methods in LoadNode and StoreNode, 1485 // and integrate with the parser's memory state and _gvn engine. 1486 // 1487 1488 // factory methods in "int adr_idx" 1489 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1490 int adr_idx, 1491 MemNode::MemOrd mo, 1492 LoadNode::ControlDependency control_dependency, 1493 bool require_atomic_access, 1494 bool unaligned, 1495 bool mismatched) { 1496 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1497 const TypePtr* adr_type = NULL; // debug-mode-only argument 1498 debug_only(adr_type = C->get_adr_type(adr_idx)); 1499 Node* mem = memory(adr_idx); 1500 Node* ld; 1501 if (require_atomic_access && bt == T_LONG) { 1502 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1503 } else if (require_atomic_access && bt == T_DOUBLE) { 1504 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1505 } else { 1506 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1507 } 1508 ld = _gvn.transform(ld); 1509 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1510 // Improve graph before escape analysis and boxing elimination. 1511 record_for_igvn(ld); 1512 } 1513 return ld; 1514 } 1515 1516 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1517 int adr_idx, 1518 MemNode::MemOrd mo, 1519 bool require_atomic_access, 1520 bool unaligned, 1521 bool mismatched) { 1522 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1523 const TypePtr* adr_type = NULL; 1524 debug_only(adr_type = C->get_adr_type(adr_idx)); 1525 Node *mem = memory(adr_idx); 1526 Node* st; 1527 if (require_atomic_access && bt == T_LONG) { 1528 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1529 } else if (require_atomic_access && bt == T_DOUBLE) { 1530 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1531 } else { 1532 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1533 } 1534 if (unaligned) { 1535 st->as_Store()->set_unaligned_access(); 1536 } 1537 if (mismatched) { 1538 st->as_Store()->set_mismatched_access(); 1539 } 1540 st = _gvn.transform(st); 1541 set_memory(st, adr_idx); 1542 // Back-to-back stores can only remove intermediate store with DU info 1543 // so push on worklist for optimizer. 1544 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1545 record_for_igvn(st); 1546 1547 return st; 1548 } 1549 1550 1551 void GraphKit::pre_barrier(bool do_load, 1552 Node* ctl, 1553 Node* obj, 1554 Node* adr, 1555 uint adr_idx, 1556 Node* val, 1557 const TypeOopPtr* val_type, 1558 Node* pre_val, 1559 BasicType bt) { 1560 1561 BarrierSet* bs = Universe::heap()->barrier_set(); 1562 set_control(ctl); 1563 switch (bs->kind()) { 1564 case BarrierSet::G1BarrierSet: 1565 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1566 break; 1567 1568 case BarrierSet::CardTableBarrierSet: 1569 break; 1570 1571 default : 1572 ShouldNotReachHere(); 1573 1574 } 1575 } 1576 1577 bool GraphKit::can_move_pre_barrier() const { 1578 BarrierSet* bs = Universe::heap()->barrier_set(); 1579 switch (bs->kind()) { 1580 case BarrierSet::G1BarrierSet: 1581 return true; // Can move it if no safepoint 1582 1583 case BarrierSet::CardTableBarrierSet: 1584 return true; // There is no pre-barrier 1585 1586 default : 1587 ShouldNotReachHere(); 1588 } 1589 return false; 1590 } 1591 1592 void GraphKit::post_barrier(Node* ctl, 1593 Node* store, 1594 Node* obj, 1595 Node* adr, 1596 uint adr_idx, 1597 Node* val, 1598 BasicType bt, 1599 bool use_precise) { 1600 BarrierSet* bs = Universe::heap()->barrier_set(); 1601 set_control(ctl); 1602 switch (bs->kind()) { 1603 case BarrierSet::G1BarrierSet: 1604 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1605 break; 1606 1607 case BarrierSet::CardTableBarrierSet: 1608 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1609 break; 1610 1611 default : 1612 ShouldNotReachHere(); 1613 1614 } 1615 } 1616 1617 Node* GraphKit::store_oop(Node* ctl, 1618 Node* obj, 1619 Node* adr, 1620 const TypePtr* adr_type, 1621 Node* val, 1622 const TypeOopPtr* val_type, 1623 BasicType bt, 1624 bool use_precise, 1625 MemNode::MemOrd mo, 1626 bool mismatched) { 1627 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1628 // could be delayed during Parse (for example, in adjust_map_after_if()). 1629 // Execute transformation here to avoid barrier generation in such case. 1630 if (_gvn.type(val) == TypePtr::NULL_PTR) 1631 val = _gvn.makecon(TypePtr::NULL_PTR); 1632 1633 set_control(ctl); 1634 if (stopped()) return top(); // Dead path ? 1635 1636 assert(bt == T_OBJECT, "sanity"); 1637 assert(val != NULL, "not dead path"); 1638 uint adr_idx = C->get_alias_index(adr_type); 1639 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1640 1641 pre_barrier(true /* do_load */, 1642 control(), obj, adr, adr_idx, val, val_type, 1643 NULL /* pre_val */, 1644 bt); 1645 1646 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1647 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1648 return store; 1649 } 1650 1651 // Could be an array or object we don't know at compile time (unsafe ref.) 1652 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1653 Node* obj, // containing obj 1654 Node* adr, // actual adress to store val at 1655 const TypePtr* adr_type, 1656 Node* val, 1657 BasicType bt, 1658 MemNode::MemOrd mo, 1659 bool mismatched) { 1660 Compile::AliasType* at = C->alias_type(adr_type); 1661 const TypeOopPtr* val_type = NULL; 1662 if (adr_type->isa_instptr()) { 1663 if (at->field() != NULL) { 1664 // known field. This code is a copy of the do_put_xxx logic. 1665 ciField* field = at->field(); 1666 if (!field->type()->is_loaded()) { 1667 val_type = TypeInstPtr::BOTTOM; 1668 } else { 1669 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1670 } 1671 } 1672 } else if (adr_type->isa_aryptr()) { 1673 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1674 } 1675 if (val_type == NULL) { 1676 val_type = TypeInstPtr::BOTTOM; 1677 } 1678 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1679 } 1680 1681 1682 //-------------------------array_element_address------------------------- 1683 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1684 const TypeInt* sizetype, Node* ctrl) { 1685 uint shift = exact_log2(type2aelembytes(elembt)); 1686 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1687 1688 // short-circuit a common case (saves lots of confusing waste motion) 1689 jint idx_con = find_int_con(idx, -1); 1690 if (idx_con >= 0) { 1691 intptr_t offset = header + ((intptr_t)idx_con << shift); 1692 return basic_plus_adr(ary, offset); 1693 } 1694 1695 // must be correct type for alignment purposes 1696 Node* base = basic_plus_adr(ary, header); 1697 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1698 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1699 return basic_plus_adr(ary, base, scale); 1700 } 1701 1702 //-------------------------load_array_element------------------------- 1703 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1704 const Type* elemtype = arytype->elem(); 1705 BasicType elembt = elemtype->array_element_basic_type(); 1706 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1707 if (elembt == T_NARROWOOP) { 1708 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1709 } 1710 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1711 return ld; 1712 } 1713 1714 //-------------------------set_arguments_for_java_call------------------------- 1715 // Arguments (pre-popped from the stack) are taken from the JVMS. 1716 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1717 // Add the call arguments: 1718 uint nargs = call->method()->arg_size(); 1719 for (uint i = 0; i < nargs; i++) { 1720 Node* arg = argument(i); 1721 call->init_req(i + TypeFunc::Parms, arg); 1722 } 1723 } 1724 1725 //---------------------------set_edges_for_java_call--------------------------- 1726 // Connect a newly created call into the current JVMS. 1727 // A return value node (if any) is returned from set_edges_for_java_call. 1728 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1729 1730 // Add the predefined inputs: 1731 call->init_req( TypeFunc::Control, control() ); 1732 call->init_req( TypeFunc::I_O , i_o() ); 1733 call->init_req( TypeFunc::Memory , reset_memory() ); 1734 call->init_req( TypeFunc::FramePtr, frameptr() ); 1735 call->init_req( TypeFunc::ReturnAdr, top() ); 1736 1737 add_safepoint_edges(call, must_throw); 1738 1739 Node* xcall = _gvn.transform(call); 1740 1741 if (xcall == top()) { 1742 set_control(top()); 1743 return; 1744 } 1745 assert(xcall == call, "call identity is stable"); 1746 1747 // Re-use the current map to produce the result. 1748 1749 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1750 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1751 set_all_memory_call(xcall, separate_io_proj); 1752 1753 //return xcall; // no need, caller already has it 1754 } 1755 1756 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1757 if (stopped()) return top(); // maybe the call folded up? 1758 1759 // Capture the return value, if any. 1760 Node* ret; 1761 if (call->method() == NULL || 1762 call->method()->return_type()->basic_type() == T_VOID) 1763 ret = top(); 1764 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1765 1766 // Note: Since any out-of-line call can produce an exception, 1767 // we always insert an I_O projection from the call into the result. 1768 1769 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1770 1771 if (separate_io_proj) { 1772 // The caller requested separate projections be used by the fall 1773 // through and exceptional paths, so replace the projections for 1774 // the fall through path. 1775 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1776 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1777 } 1778 return ret; 1779 } 1780 1781 //--------------------set_predefined_input_for_runtime_call-------------------- 1782 // Reading and setting the memory state is way conservative here. 1783 // The real problem is that I am not doing real Type analysis on memory, 1784 // so I cannot distinguish card mark stores from other stores. Across a GC 1785 // point the Store Barrier and the card mark memory has to agree. I cannot 1786 // have a card mark store and its barrier split across the GC point from 1787 // either above or below. Here I get that to happen by reading ALL of memory. 1788 // A better answer would be to separate out card marks from other memory. 1789 // For now, return the input memory state, so that it can be reused 1790 // after the call, if this call has restricted memory effects. 1791 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1792 // Set fixed predefined input arguments 1793 Node* memory = reset_memory(); 1794 call->init_req( TypeFunc::Control, control() ); 1795 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1796 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1797 call->init_req( TypeFunc::FramePtr, frameptr() ); 1798 call->init_req( TypeFunc::ReturnAdr, top() ); 1799 return memory; 1800 } 1801 1802 //-------------------set_predefined_output_for_runtime_call-------------------- 1803 // Set control and memory (not i_o) from the call. 1804 // If keep_mem is not NULL, use it for the output state, 1805 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1806 // If hook_mem is NULL, this call produces no memory effects at all. 1807 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1808 // then only that memory slice is taken from the call. 1809 // In the last case, we must put an appropriate memory barrier before 1810 // the call, so as to create the correct anti-dependencies on loads 1811 // preceding the call. 1812 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1813 Node* keep_mem, 1814 const TypePtr* hook_mem) { 1815 // no i/o 1816 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1817 if (keep_mem) { 1818 // First clone the existing memory state 1819 set_all_memory(keep_mem); 1820 if (hook_mem != NULL) { 1821 // Make memory for the call 1822 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1823 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1824 // We also use hook_mem to extract specific effects from arraycopy stubs. 1825 set_memory(mem, hook_mem); 1826 } 1827 // ...else the call has NO memory effects. 1828 1829 // Make sure the call advertises its memory effects precisely. 1830 // This lets us build accurate anti-dependences in gcm.cpp. 1831 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1832 "call node must be constructed correctly"); 1833 } else { 1834 assert(hook_mem == NULL, ""); 1835 // This is not a "slow path" call; all memory comes from the call. 1836 set_all_memory_call(call); 1837 } 1838 } 1839 1840 1841 // Replace the call with the current state of the kit. 1842 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1843 JVMState* ejvms = NULL; 1844 if (has_exceptions()) { 1845 ejvms = transfer_exceptions_into_jvms(); 1846 } 1847 1848 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1849 ReplacedNodes replaced_nodes_exception; 1850 Node* ex_ctl = top(); 1851 1852 SafePointNode* final_state = stop(); 1853 1854 // Find all the needed outputs of this call 1855 CallProjections callprojs; 1856 call->extract_projections(&callprojs, true); 1857 1858 Node* init_mem = call->in(TypeFunc::Memory); 1859 Node* final_mem = final_state->in(TypeFunc::Memory); 1860 Node* final_ctl = final_state->in(TypeFunc::Control); 1861 Node* final_io = final_state->in(TypeFunc::I_O); 1862 1863 // Replace all the old call edges with the edges from the inlining result 1864 if (callprojs.fallthrough_catchproj != NULL) { 1865 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1866 } 1867 if (callprojs.fallthrough_memproj != NULL) { 1868 if (final_mem->is_MergeMem()) { 1869 // Parser's exits MergeMem was not transformed but may be optimized 1870 final_mem = _gvn.transform(final_mem); 1871 } 1872 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1873 } 1874 if (callprojs.fallthrough_ioproj != NULL) { 1875 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1876 } 1877 1878 // Replace the result with the new result if it exists and is used 1879 if (callprojs.resproj != NULL && result != NULL) { 1880 C->gvn_replace_by(callprojs.resproj, result); 1881 } 1882 1883 if (ejvms == NULL) { 1884 // No exception edges to simply kill off those paths 1885 if (callprojs.catchall_catchproj != NULL) { 1886 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1887 } 1888 if (callprojs.catchall_memproj != NULL) { 1889 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1890 } 1891 if (callprojs.catchall_ioproj != NULL) { 1892 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1893 } 1894 // Replace the old exception object with top 1895 if (callprojs.exobj != NULL) { 1896 C->gvn_replace_by(callprojs.exobj, C->top()); 1897 } 1898 } else { 1899 GraphKit ekit(ejvms); 1900 1901 // Load my combined exception state into the kit, with all phis transformed: 1902 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1903 replaced_nodes_exception = ex_map->replaced_nodes(); 1904 1905 Node* ex_oop = ekit.use_exception_state(ex_map); 1906 1907 if (callprojs.catchall_catchproj != NULL) { 1908 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1909 ex_ctl = ekit.control(); 1910 } 1911 if (callprojs.catchall_memproj != NULL) { 1912 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1913 } 1914 if (callprojs.catchall_ioproj != NULL) { 1915 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1916 } 1917 1918 // Replace the old exception object with the newly created one 1919 if (callprojs.exobj != NULL) { 1920 C->gvn_replace_by(callprojs.exobj, ex_oop); 1921 } 1922 } 1923 1924 // Disconnect the call from the graph 1925 call->disconnect_inputs(NULL, C); 1926 C->gvn_replace_by(call, C->top()); 1927 1928 // Clean up any MergeMems that feed other MergeMems since the 1929 // optimizer doesn't like that. 1930 if (final_mem->is_MergeMem()) { 1931 Node_List wl; 1932 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1933 Node* m = i.get(); 1934 if (m->is_MergeMem() && !wl.contains(m)) { 1935 wl.push(m); 1936 } 1937 } 1938 while (wl.size() > 0) { 1939 _gvn.transform(wl.pop()); 1940 } 1941 } 1942 1943 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1944 replaced_nodes.apply(C, final_ctl); 1945 } 1946 if (!ex_ctl->is_top() && do_replaced_nodes) { 1947 replaced_nodes_exception.apply(C, ex_ctl); 1948 } 1949 } 1950 1951 1952 //------------------------------increment_counter------------------------------ 1953 // for statistics: increment a VM counter by 1 1954 1955 void GraphKit::increment_counter(address counter_addr) { 1956 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1957 increment_counter(adr1); 1958 } 1959 1960 void GraphKit::increment_counter(Node* counter_addr) { 1961 int adr_type = Compile::AliasIdxRaw; 1962 Node* ctrl = control(); 1963 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1964 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1965 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1966 } 1967 1968 1969 //------------------------------uncommon_trap---------------------------------- 1970 // Bail out to the interpreter in mid-method. Implemented by calling the 1971 // uncommon_trap blob. This helper function inserts a runtime call with the 1972 // right debug info. 1973 void GraphKit::uncommon_trap(int trap_request, 1974 ciKlass* klass, const char* comment, 1975 bool must_throw, 1976 bool keep_exact_action) { 1977 if (failing()) stop(); 1978 if (stopped()) return; // trap reachable? 1979 1980 // Note: If ProfileTraps is true, and if a deopt. actually 1981 // occurs here, the runtime will make sure an MDO exists. There is 1982 // no need to call method()->ensure_method_data() at this point. 1983 1984 // Set the stack pointer to the right value for reexecution: 1985 set_sp(reexecute_sp()); 1986 1987 #ifdef ASSERT 1988 if (!must_throw) { 1989 // Make sure the stack has at least enough depth to execute 1990 // the current bytecode. 1991 int inputs, ignored_depth; 1992 if (compute_stack_effects(inputs, ignored_depth)) { 1993 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1994 Bytecodes::name(java_bc()), sp(), inputs); 1995 } 1996 } 1997 #endif 1998 1999 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2000 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2001 2002 switch (action) { 2003 case Deoptimization::Action_maybe_recompile: 2004 case Deoptimization::Action_reinterpret: 2005 // Temporary fix for 6529811 to allow virtual calls to be sure they 2006 // get the chance to go from mono->bi->mega 2007 if (!keep_exact_action && 2008 Deoptimization::trap_request_index(trap_request) < 0 && 2009 too_many_recompiles(reason)) { 2010 // This BCI is causing too many recompilations. 2011 if (C->log() != NULL) { 2012 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2013 Deoptimization::trap_reason_name(reason), 2014 Deoptimization::trap_action_name(action)); 2015 } 2016 action = Deoptimization::Action_none; 2017 trap_request = Deoptimization::make_trap_request(reason, action); 2018 } else { 2019 C->set_trap_can_recompile(true); 2020 } 2021 break; 2022 case Deoptimization::Action_make_not_entrant: 2023 C->set_trap_can_recompile(true); 2024 break; 2025 case Deoptimization::Action_none: 2026 case Deoptimization::Action_make_not_compilable: 2027 break; 2028 default: 2029 #ifdef ASSERT 2030 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2031 #endif 2032 break; 2033 } 2034 2035 if (TraceOptoParse) { 2036 char buf[100]; 2037 tty->print_cr("Uncommon trap %s at bci:%d", 2038 Deoptimization::format_trap_request(buf, sizeof(buf), 2039 trap_request), bci()); 2040 } 2041 2042 CompileLog* log = C->log(); 2043 if (log != NULL) { 2044 int kid = (klass == NULL)? -1: log->identify(klass); 2045 log->begin_elem("uncommon_trap bci='%d'", bci()); 2046 char buf[100]; 2047 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2048 trap_request)); 2049 if (kid >= 0) log->print(" klass='%d'", kid); 2050 if (comment != NULL) log->print(" comment='%s'", comment); 2051 log->end_elem(); 2052 } 2053 2054 // Make sure any guarding test views this path as very unlikely 2055 Node *i0 = control()->in(0); 2056 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2057 IfNode *iff = i0->as_If(); 2058 float f = iff->_prob; // Get prob 2059 if (control()->Opcode() == Op_IfTrue) { 2060 if (f > PROB_UNLIKELY_MAG(4)) 2061 iff->_prob = PROB_MIN; 2062 } else { 2063 if (f < PROB_LIKELY_MAG(4)) 2064 iff->_prob = PROB_MAX; 2065 } 2066 } 2067 2068 // Clear out dead values from the debug info. 2069 kill_dead_locals(); 2070 2071 // Now insert the uncommon trap subroutine call 2072 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2073 const TypePtr* no_memory_effects = NULL; 2074 // Pass the index of the class to be loaded 2075 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2076 (must_throw ? RC_MUST_THROW : 0), 2077 OptoRuntime::uncommon_trap_Type(), 2078 call_addr, "uncommon_trap", no_memory_effects, 2079 intcon(trap_request)); 2080 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2081 "must extract request correctly from the graph"); 2082 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2083 2084 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2085 // The debug info is the only real input to this call. 2086 2087 // Halt-and-catch fire here. The above call should never return! 2088 HaltNode* halt = new HaltNode(control(), frameptr()); 2089 _gvn.set_type_bottom(halt); 2090 root()->add_req(halt); 2091 2092 stop_and_kill_map(); 2093 } 2094 2095 2096 //--------------------------just_allocated_object------------------------------ 2097 // Report the object that was just allocated. 2098 // It must be the case that there are no intervening safepoints. 2099 // We use this to determine if an object is so "fresh" that 2100 // it does not require card marks. 2101 Node* GraphKit::just_allocated_object(Node* current_control) { 2102 if (C->recent_alloc_ctl() == current_control) 2103 return C->recent_alloc_obj(); 2104 return NULL; 2105 } 2106 2107 2108 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2109 // (Note: TypeFunc::make has a cache that makes this fast.) 2110 const TypeFunc* tf = TypeFunc::make(dest_method); 2111 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2112 for (int j = 0; j < nargs; j++) { 2113 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2114 if( targ->basic_type() == T_DOUBLE ) { 2115 // If any parameters are doubles, they must be rounded before 2116 // the call, dstore_rounding does gvn.transform 2117 Node *arg = argument(j); 2118 arg = dstore_rounding(arg); 2119 set_argument(j, arg); 2120 } 2121 } 2122 } 2123 2124 /** 2125 * Record profiling data exact_kls for Node n with the type system so 2126 * that it can propagate it (speculation) 2127 * 2128 * @param n node that the type applies to 2129 * @param exact_kls type from profiling 2130 * @param maybe_null did profiling see null? 2131 * 2132 * @return node with improved type 2133 */ 2134 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2135 const Type* current_type = _gvn.type(n); 2136 assert(UseTypeSpeculation, "type speculation must be on"); 2137 2138 const TypePtr* speculative = current_type->speculative(); 2139 2140 // Should the klass from the profile be recorded in the speculative type? 2141 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2142 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2143 const TypeOopPtr* xtype = tklass->as_instance_type(); 2144 assert(xtype->klass_is_exact(), "Should be exact"); 2145 // Any reason to believe n is not null (from this profiling or a previous one)? 2146 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2147 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2148 // record the new speculative type's depth 2149 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2150 speculative = speculative->with_inline_depth(jvms()->depth()); 2151 } else if (current_type->would_improve_ptr(ptr_kind)) { 2152 // Profiling report that null was never seen so we can change the 2153 // speculative type to non null ptr. 2154 if (ptr_kind == ProfileAlwaysNull) { 2155 speculative = TypePtr::NULL_PTR; 2156 } else { 2157 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2158 const TypePtr* ptr = TypePtr::NOTNULL; 2159 if (speculative != NULL) { 2160 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2161 } else { 2162 speculative = ptr; 2163 } 2164 } 2165 } 2166 2167 if (speculative != current_type->speculative()) { 2168 // Build a type with a speculative type (what we think we know 2169 // about the type but will need a guard when we use it) 2170 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2171 // We're changing the type, we need a new CheckCast node to carry 2172 // the new type. The new type depends on the control: what 2173 // profiling tells us is only valid from here as far as we can 2174 // tell. 2175 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2176 cast = _gvn.transform(cast); 2177 replace_in_map(n, cast); 2178 n = cast; 2179 } 2180 2181 return n; 2182 } 2183 2184 /** 2185 * Record profiling data from receiver profiling at an invoke with the 2186 * type system so that it can propagate it (speculation) 2187 * 2188 * @param n receiver node 2189 * 2190 * @return node with improved type 2191 */ 2192 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2193 if (!UseTypeSpeculation) { 2194 return n; 2195 } 2196 ciKlass* exact_kls = profile_has_unique_klass(); 2197 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2198 if ((java_bc() == Bytecodes::_checkcast || 2199 java_bc() == Bytecodes::_instanceof || 2200 java_bc() == Bytecodes::_aastore) && 2201 method()->method_data()->is_mature()) { 2202 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2203 if (data != NULL) { 2204 if (!data->as_BitData()->null_seen()) { 2205 ptr_kind = ProfileNeverNull; 2206 } else { 2207 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2208 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2209 uint i = 0; 2210 for (; i < call->row_limit(); i++) { 2211 ciKlass* receiver = call->receiver(i); 2212 if (receiver != NULL) { 2213 break; 2214 } 2215 } 2216 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2217 } 2218 } 2219 } 2220 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2221 } 2222 2223 /** 2224 * Record profiling data from argument profiling at an invoke with the 2225 * type system so that it can propagate it (speculation) 2226 * 2227 * @param dest_method target method for the call 2228 * @param bc what invoke bytecode is this? 2229 */ 2230 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2231 if (!UseTypeSpeculation) { 2232 return; 2233 } 2234 const TypeFunc* tf = TypeFunc::make(dest_method); 2235 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2236 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2237 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2238 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2239 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2240 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2241 ciKlass* better_type = NULL; 2242 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2243 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2244 } 2245 i++; 2246 } 2247 } 2248 } 2249 2250 /** 2251 * Record profiling data from parameter profiling at an invoke with 2252 * the type system so that it can propagate it (speculation) 2253 */ 2254 void GraphKit::record_profiled_parameters_for_speculation() { 2255 if (!UseTypeSpeculation) { 2256 return; 2257 } 2258 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2259 if (_gvn.type(local(i))->isa_oopptr()) { 2260 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2261 ciKlass* better_type = NULL; 2262 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2263 record_profile_for_speculation(local(i), better_type, ptr_kind); 2264 } 2265 j++; 2266 } 2267 } 2268 } 2269 2270 /** 2271 * Record profiling data from return value profiling at an invoke with 2272 * the type system so that it can propagate it (speculation) 2273 */ 2274 void GraphKit::record_profiled_return_for_speculation() { 2275 if (!UseTypeSpeculation) { 2276 return; 2277 } 2278 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2279 ciKlass* better_type = NULL; 2280 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2281 // If profiling reports a single type for the return value, 2282 // feed it to the type system so it can propagate it as a 2283 // speculative type 2284 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2285 } 2286 } 2287 2288 void GraphKit::round_double_result(ciMethod* dest_method) { 2289 // A non-strict method may return a double value which has an extended 2290 // exponent, but this must not be visible in a caller which is 'strict' 2291 // If a strict caller invokes a non-strict callee, round a double result 2292 2293 BasicType result_type = dest_method->return_type()->basic_type(); 2294 assert( method() != NULL, "must have caller context"); 2295 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2296 // Destination method's return value is on top of stack 2297 // dstore_rounding() does gvn.transform 2298 Node *result = pop_pair(); 2299 result = dstore_rounding(result); 2300 push_pair(result); 2301 } 2302 } 2303 2304 // rounding for strict float precision conformance 2305 Node* GraphKit::precision_rounding(Node* n) { 2306 return UseStrictFP && _method->flags().is_strict() 2307 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2308 ? _gvn.transform( new RoundFloatNode(0, n) ) 2309 : n; 2310 } 2311 2312 // rounding for strict double precision conformance 2313 Node* GraphKit::dprecision_rounding(Node *n) { 2314 return UseStrictFP && _method->flags().is_strict() 2315 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2316 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2317 : n; 2318 } 2319 2320 // rounding for non-strict double stores 2321 Node* GraphKit::dstore_rounding(Node* n) { 2322 return Matcher::strict_fp_requires_explicit_rounding 2323 && UseSSE <= 1 2324 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2325 : n; 2326 } 2327 2328 //============================================================================= 2329 // Generate a fast path/slow path idiom. Graph looks like: 2330 // [foo] indicates that 'foo' is a parameter 2331 // 2332 // [in] NULL 2333 // \ / 2334 // CmpP 2335 // Bool ne 2336 // If 2337 // / \ 2338 // True False-<2> 2339 // / | 2340 // / cast_not_null 2341 // Load | | ^ 2342 // [fast_test] | | 2343 // gvn to opt_test | | 2344 // / \ | <1> 2345 // True False | 2346 // | \\ | 2347 // [slow_call] \[fast_result] 2348 // Ctl Val \ \ 2349 // | \ \ 2350 // Catch <1> \ \ 2351 // / \ ^ \ \ 2352 // Ex No_Ex | \ \ 2353 // | \ \ | \ <2> \ 2354 // ... \ [slow_res] | | \ [null_result] 2355 // \ \--+--+--- | | 2356 // \ | / \ | / 2357 // --------Region Phi 2358 // 2359 //============================================================================= 2360 // Code is structured as a series of driver functions all called 'do_XXX' that 2361 // call a set of helper functions. Helper functions first, then drivers. 2362 2363 //------------------------------null_check_oop--------------------------------- 2364 // Null check oop. Set null-path control into Region in slot 3. 2365 // Make a cast-not-nullness use the other not-null control. Return cast. 2366 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2367 bool never_see_null, 2368 bool safe_for_replace, 2369 bool speculative) { 2370 // Initial NULL check taken path 2371 (*null_control) = top(); 2372 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2373 2374 // Generate uncommon_trap: 2375 if (never_see_null && (*null_control) != top()) { 2376 // If we see an unexpected null at a check-cast we record it and force a 2377 // recompile; the offending check-cast will be compiled to handle NULLs. 2378 // If we see more than one offending BCI, then all checkcasts in the 2379 // method will be compiled to handle NULLs. 2380 PreserveJVMState pjvms(this); 2381 set_control(*null_control); 2382 replace_in_map(value, null()); 2383 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2384 uncommon_trap(reason, 2385 Deoptimization::Action_make_not_entrant); 2386 (*null_control) = top(); // NULL path is dead 2387 } 2388 if ((*null_control) == top() && safe_for_replace) { 2389 replace_in_map(value, cast); 2390 } 2391 2392 // Cast away null-ness on the result 2393 return cast; 2394 } 2395 2396 //------------------------------opt_iff---------------------------------------- 2397 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2398 // Return slow-path control. 2399 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2400 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2401 2402 // Fast path taken; set region slot 2 2403 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2404 region->init_req(2,fast_taken); // Capture fast-control 2405 2406 // Fast path not-taken, i.e. slow path 2407 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2408 return slow_taken; 2409 } 2410 2411 //-----------------------------make_runtime_call------------------------------- 2412 Node* GraphKit::make_runtime_call(int flags, 2413 const TypeFunc* call_type, address call_addr, 2414 const char* call_name, 2415 const TypePtr* adr_type, 2416 // The following parms are all optional. 2417 // The first NULL ends the list. 2418 Node* parm0, Node* parm1, 2419 Node* parm2, Node* parm3, 2420 Node* parm4, Node* parm5, 2421 Node* parm6, Node* parm7) { 2422 // Slow-path call 2423 bool is_leaf = !(flags & RC_NO_LEAF); 2424 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2425 if (call_name == NULL) { 2426 assert(!is_leaf, "must supply name for leaf"); 2427 call_name = OptoRuntime::stub_name(call_addr); 2428 } 2429 CallNode* call; 2430 if (!is_leaf) { 2431 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2432 bci(), adr_type); 2433 } else if (flags & RC_NO_FP) { 2434 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2435 } else { 2436 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2437 } 2438 2439 // The following is similar to set_edges_for_java_call, 2440 // except that the memory effects of the call are restricted to AliasIdxRaw. 2441 2442 // Slow path call has no side-effects, uses few values 2443 bool wide_in = !(flags & RC_NARROW_MEM); 2444 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2445 2446 Node* prev_mem = NULL; 2447 if (wide_in) { 2448 prev_mem = set_predefined_input_for_runtime_call(call); 2449 } else { 2450 assert(!wide_out, "narrow in => narrow out"); 2451 Node* narrow_mem = memory(adr_type); 2452 prev_mem = reset_memory(); 2453 map()->set_memory(narrow_mem); 2454 set_predefined_input_for_runtime_call(call); 2455 } 2456 2457 // Hook each parm in order. Stop looking at the first NULL. 2458 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2459 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2460 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2461 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2462 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2463 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2464 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2465 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2466 /* close each nested if ===> */ } } } } } } } } 2467 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2468 2469 if (!is_leaf) { 2470 // Non-leaves can block and take safepoints: 2471 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2472 } 2473 // Non-leaves can throw exceptions: 2474 if (has_io) { 2475 call->set_req(TypeFunc::I_O, i_o()); 2476 } 2477 2478 if (flags & RC_UNCOMMON) { 2479 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2480 // (An "if" probability corresponds roughly to an unconditional count. 2481 // Sort of.) 2482 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2483 } 2484 2485 Node* c = _gvn.transform(call); 2486 assert(c == call, "cannot disappear"); 2487 2488 if (wide_out) { 2489 // Slow path call has full side-effects. 2490 set_predefined_output_for_runtime_call(call); 2491 } else { 2492 // Slow path call has few side-effects, and/or sets few values. 2493 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2494 } 2495 2496 if (has_io) { 2497 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2498 } 2499 return call; 2500 2501 } 2502 2503 //------------------------------merge_memory----------------------------------- 2504 // Merge memory from one path into the current memory state. 2505 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2506 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2507 Node* old_slice = mms.force_memory(); 2508 Node* new_slice = mms.memory2(); 2509 if (old_slice != new_slice) { 2510 PhiNode* phi; 2511 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2512 if (mms.is_empty()) { 2513 // clone base memory Phi's inputs for this memory slice 2514 assert(old_slice == mms.base_memory(), "sanity"); 2515 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2516 _gvn.set_type(phi, Type::MEMORY); 2517 for (uint i = 1; i < phi->req(); i++) { 2518 phi->init_req(i, old_slice->in(i)); 2519 } 2520 } else { 2521 phi = old_slice->as_Phi(); // Phi was generated already 2522 } 2523 } else { 2524 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2525 _gvn.set_type(phi, Type::MEMORY); 2526 } 2527 phi->set_req(new_path, new_slice); 2528 mms.set_memory(phi); 2529 } 2530 } 2531 } 2532 2533 //------------------------------make_slow_call_ex------------------------------ 2534 // Make the exception handler hookups for the slow call 2535 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2536 if (stopped()) return; 2537 2538 // Make a catch node with just two handlers: fall-through and catch-all 2539 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2540 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2541 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2542 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2543 2544 { PreserveJVMState pjvms(this); 2545 set_control(excp); 2546 set_i_o(i_o); 2547 2548 if (excp != top()) { 2549 if (deoptimize) { 2550 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2551 uncommon_trap(Deoptimization::Reason_unhandled, 2552 Deoptimization::Action_none); 2553 } else { 2554 // Create an exception state also. 2555 // Use an exact type if the caller has specified a specific exception. 2556 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2557 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2558 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2559 } 2560 } 2561 } 2562 2563 // Get the no-exception control from the CatchNode. 2564 set_control(norm); 2565 } 2566 2567 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2568 Node* cmp = NULL; 2569 switch(bt) { 2570 case T_INT: cmp = new CmpINode(in1, in2); break; 2571 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2572 default: fatal("unexpected comparison type %s", type2name(bt)); 2573 } 2574 gvn->transform(cmp); 2575 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2576 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2577 gvn->transform(iff); 2578 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2579 return iff; 2580 } 2581 2582 2583 //-------------------------------gen_subtype_check----------------------------- 2584 // Generate a subtyping check. Takes as input the subtype and supertype. 2585 // Returns 2 values: sets the default control() to the true path and returns 2586 // the false path. Only reads invariant memory; sets no (visible) memory. 2587 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2588 // but that's not exposed to the optimizer. This call also doesn't take in an 2589 // Object; if you wish to check an Object you need to load the Object's class 2590 // prior to coming here. 2591 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2592 Compile* C = gvn->C; 2593 2594 if ((*ctrl)->is_top()) { 2595 return C->top(); 2596 } 2597 2598 // Fast check for identical types, perhaps identical constants. 2599 // The types can even be identical non-constants, in cases 2600 // involving Array.newInstance, Object.clone, etc. 2601 if (subklass == superklass) 2602 return C->top(); // false path is dead; no test needed. 2603 2604 if (gvn->type(superklass)->singleton()) { 2605 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2606 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2607 2608 // In the common case of an exact superklass, try to fold up the 2609 // test before generating code. You may ask, why not just generate 2610 // the code and then let it fold up? The answer is that the generated 2611 // code will necessarily include null checks, which do not always 2612 // completely fold away. If they are also needless, then they turn 2613 // into a performance loss. Example: 2614 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2615 // Here, the type of 'fa' is often exact, so the store check 2616 // of fa[1]=x will fold up, without testing the nullness of x. 2617 switch (C->static_subtype_check(superk, subk)) { 2618 case Compile::SSC_always_false: 2619 { 2620 Node* always_fail = *ctrl; 2621 *ctrl = gvn->C->top(); 2622 return always_fail; 2623 } 2624 case Compile::SSC_always_true: 2625 return C->top(); 2626 case Compile::SSC_easy_test: 2627 { 2628 // Just do a direct pointer compare and be done. 2629 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2630 *ctrl = gvn->transform(new IfTrueNode(iff)); 2631 return gvn->transform(new IfFalseNode(iff)); 2632 } 2633 case Compile::SSC_full_test: 2634 break; 2635 default: 2636 ShouldNotReachHere(); 2637 } 2638 } 2639 2640 // %%% Possible further optimization: Even if the superklass is not exact, 2641 // if the subklass is the unique subtype of the superklass, the check 2642 // will always succeed. We could leave a dependency behind to ensure this. 2643 2644 // First load the super-klass's check-offset 2645 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2646 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2647 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2648 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2649 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2650 2651 // Load from the sub-klass's super-class display list, or a 1-word cache of 2652 // the secondary superclass list, or a failing value with a sentinel offset 2653 // if the super-klass is an interface or exceptionally deep in the Java 2654 // hierarchy and we have to scan the secondary superclass list the hard way. 2655 // Worst-case type is a little odd: NULL is allowed as a result (usually 2656 // klass loads can never produce a NULL). 2657 Node *chk_off_X = chk_off; 2658 #ifdef _LP64 2659 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2660 #endif 2661 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2662 // For some types like interfaces the following loadKlass is from a 1-word 2663 // cache which is mutable so can't use immutable memory. Other 2664 // types load from the super-class display table which is immutable. 2665 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2666 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2667 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2668 2669 // Compile speed common case: ARE a subtype and we canNOT fail 2670 if( superklass == nkls ) 2671 return C->top(); // false path is dead; no test needed. 2672 2673 // See if we get an immediate positive hit. Happens roughly 83% of the 2674 // time. Test to see if the value loaded just previously from the subklass 2675 // is exactly the superklass. 2676 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2677 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2678 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2679 2680 // Compile speed common case: Check for being deterministic right now. If 2681 // chk_off is a constant and not equal to cacheoff then we are NOT a 2682 // subklass. In this case we need exactly the 1 test above and we can 2683 // return those results immediately. 2684 if (!might_be_cache) { 2685 Node* not_subtype_ctrl = *ctrl; 2686 *ctrl = iftrue1; // We need exactly the 1 test above 2687 return not_subtype_ctrl; 2688 } 2689 2690 // Gather the various success & failures here 2691 RegionNode *r_ok_subtype = new RegionNode(4); 2692 gvn->record_for_igvn(r_ok_subtype); 2693 RegionNode *r_not_subtype = new RegionNode(3); 2694 gvn->record_for_igvn(r_not_subtype); 2695 2696 r_ok_subtype->init_req(1, iftrue1); 2697 2698 // Check for immediate negative hit. Happens roughly 11% of the time (which 2699 // is roughly 63% of the remaining cases). Test to see if the loaded 2700 // check-offset points into the subklass display list or the 1-element 2701 // cache. If it points to the display (and NOT the cache) and the display 2702 // missed then it's not a subtype. 2703 Node *cacheoff = gvn->intcon(cacheoff_con); 2704 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2705 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2706 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2707 2708 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2709 // No performance impact (too rare) but allows sharing of secondary arrays 2710 // which has some footprint reduction. 2711 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2712 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2713 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2714 2715 // -- Roads not taken here: -- 2716 // We could also have chosen to perform the self-check at the beginning 2717 // of this code sequence, as the assembler does. This would not pay off 2718 // the same way, since the optimizer, unlike the assembler, can perform 2719 // static type analysis to fold away many successful self-checks. 2720 // Non-foldable self checks work better here in second position, because 2721 // the initial primary superclass check subsumes a self-check for most 2722 // types. An exception would be a secondary type like array-of-interface, 2723 // which does not appear in its own primary supertype display. 2724 // Finally, we could have chosen to move the self-check into the 2725 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2726 // dependent manner. But it is worthwhile to have the check here, 2727 // where it can be perhaps be optimized. The cost in code space is 2728 // small (register compare, branch). 2729 2730 // Now do a linear scan of the secondary super-klass array. Again, no real 2731 // performance impact (too rare) but it's gotta be done. 2732 // Since the code is rarely used, there is no penalty for moving it 2733 // out of line, and it can only improve I-cache density. 2734 // The decision to inline or out-of-line this final check is platform 2735 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2736 Node* psc = gvn->transform( 2737 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2738 2739 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2740 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2741 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2742 2743 // Return false path; set default control to true path. 2744 *ctrl = gvn->transform(r_ok_subtype); 2745 return gvn->transform(r_not_subtype); 2746 } 2747 2748 // Profile-driven exact type check: 2749 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2750 float prob, 2751 Node* *casted_receiver) { 2752 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2753 Node* recv_klass = load_object_klass(receiver); 2754 Node* want_klass = makecon(tklass); 2755 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2756 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2757 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2758 set_control( _gvn.transform( new IfTrueNode (iff) )); 2759 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2760 2761 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2762 assert(recv_xtype->klass_is_exact(), ""); 2763 2764 // Subsume downstream occurrences of receiver with a cast to 2765 // recv_xtype, since now we know what the type will be. 2766 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2767 (*casted_receiver) = _gvn.transform(cast); 2768 // (User must make the replace_in_map call.) 2769 2770 return fail; 2771 } 2772 2773 2774 //------------------------------seems_never_null------------------------------- 2775 // Use null_seen information if it is available from the profile. 2776 // If we see an unexpected null at a type check we record it and force a 2777 // recompile; the offending check will be recompiled to handle NULLs. 2778 // If we see several offending BCIs, then all checks in the 2779 // method will be recompiled. 2780 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2781 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2782 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2783 if (UncommonNullCast // Cutout for this technique 2784 && obj != null() // And not the -Xcomp stupid case? 2785 && !too_many_traps(reason) 2786 ) { 2787 if (speculating) { 2788 return true; 2789 } 2790 if (data == NULL) 2791 // Edge case: no mature data. Be optimistic here. 2792 return true; 2793 // If the profile has not seen a null, assume it won't happen. 2794 assert(java_bc() == Bytecodes::_checkcast || 2795 java_bc() == Bytecodes::_instanceof || 2796 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2797 return !data->as_BitData()->null_seen(); 2798 } 2799 speculating = false; 2800 return false; 2801 } 2802 2803 //------------------------maybe_cast_profiled_receiver------------------------- 2804 // If the profile has seen exactly one type, narrow to exactly that type. 2805 // Subsequent type checks will always fold up. 2806 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2807 ciKlass* require_klass, 2808 ciKlass* spec_klass, 2809 bool safe_for_replace) { 2810 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2811 2812 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2813 2814 // Make sure we haven't already deoptimized from this tactic. 2815 if (too_many_traps(reason) || too_many_recompiles(reason)) 2816 return NULL; 2817 2818 // (No, this isn't a call, but it's enough like a virtual call 2819 // to use the same ciMethod accessor to get the profile info...) 2820 // If we have a speculative type use it instead of profiling (which 2821 // may not help us) 2822 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2823 if (exact_kls != NULL) {// no cast failures here 2824 if (require_klass == NULL || 2825 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2826 // If we narrow the type to match what the type profile sees or 2827 // the speculative type, we can then remove the rest of the 2828 // cast. 2829 // This is a win, even if the exact_kls is very specific, 2830 // because downstream operations, such as method calls, 2831 // will often benefit from the sharper type. 2832 Node* exact_obj = not_null_obj; // will get updated in place... 2833 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2834 &exact_obj); 2835 { PreserveJVMState pjvms(this); 2836 set_control(slow_ctl); 2837 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2838 } 2839 if (safe_for_replace) { 2840 replace_in_map(not_null_obj, exact_obj); 2841 } 2842 return exact_obj; 2843 } 2844 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2845 } 2846 2847 return NULL; 2848 } 2849 2850 /** 2851 * Cast obj to type and emit guard unless we had too many traps here 2852 * already 2853 * 2854 * @param obj node being casted 2855 * @param type type to cast the node to 2856 * @param not_null true if we know node cannot be null 2857 */ 2858 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2859 ciKlass* type, 2860 bool not_null) { 2861 if (stopped()) { 2862 return obj; 2863 } 2864 2865 // type == NULL if profiling tells us this object is always null 2866 if (type != NULL) { 2867 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2868 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2869 2870 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2871 !too_many_traps(class_reason) && 2872 !too_many_recompiles(class_reason)) { 2873 Node* not_null_obj = NULL; 2874 // not_null is true if we know the object is not null and 2875 // there's no need for a null check 2876 if (!not_null) { 2877 Node* null_ctl = top(); 2878 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2879 assert(null_ctl->is_top(), "no null control here"); 2880 } else { 2881 not_null_obj = obj; 2882 } 2883 2884 Node* exact_obj = not_null_obj; 2885 ciKlass* exact_kls = type; 2886 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2887 &exact_obj); 2888 { 2889 PreserveJVMState pjvms(this); 2890 set_control(slow_ctl); 2891 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2892 } 2893 replace_in_map(not_null_obj, exact_obj); 2894 obj = exact_obj; 2895 } 2896 } else { 2897 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2898 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2899 Node* exact_obj = null_assert(obj); 2900 replace_in_map(obj, exact_obj); 2901 obj = exact_obj; 2902 } 2903 } 2904 return obj; 2905 } 2906 2907 //-------------------------------gen_instanceof-------------------------------- 2908 // Generate an instance-of idiom. Used by both the instance-of bytecode 2909 // and the reflective instance-of call. 2910 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2911 kill_dead_locals(); // Benefit all the uncommon traps 2912 assert( !stopped(), "dead parse path should be checked in callers" ); 2913 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2914 "must check for not-null not-dead klass in callers"); 2915 2916 // Make the merge point 2917 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2918 RegionNode* region = new RegionNode(PATH_LIMIT); 2919 Node* phi = new PhiNode(region, TypeInt::BOOL); 2920 C->set_has_split_ifs(true); // Has chance for split-if optimization 2921 2922 ciProfileData* data = NULL; 2923 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2924 data = method()->method_data()->bci_to_data(bci()); 2925 } 2926 bool speculative_not_null = false; 2927 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2928 && seems_never_null(obj, data, speculative_not_null)); 2929 2930 // Null check; get casted pointer; set region slot 3 2931 Node* null_ctl = top(); 2932 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2933 2934 // If not_null_obj is dead, only null-path is taken 2935 if (stopped()) { // Doing instance-of on a NULL? 2936 set_control(null_ctl); 2937 return intcon(0); 2938 } 2939 region->init_req(_null_path, null_ctl); 2940 phi ->init_req(_null_path, intcon(0)); // Set null path value 2941 if (null_ctl == top()) { 2942 // Do this eagerly, so that pattern matches like is_diamond_phi 2943 // will work even during parsing. 2944 assert(_null_path == PATH_LIMIT-1, "delete last"); 2945 region->del_req(_null_path); 2946 phi ->del_req(_null_path); 2947 } 2948 2949 // Do we know the type check always succeed? 2950 bool known_statically = false; 2951 if (_gvn.type(superklass)->singleton()) { 2952 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2953 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2954 if (subk != NULL && subk->is_loaded()) { 2955 int static_res = C->static_subtype_check(superk, subk); 2956 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2957 } 2958 } 2959 2960 if (!known_statically) { 2961 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2962 // We may not have profiling here or it may not help us. If we 2963 // have a speculative type use it to perform an exact cast. 2964 ciKlass* spec_obj_type = obj_type->speculative_type(); 2965 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2966 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2967 if (stopped()) { // Profile disagrees with this path. 2968 set_control(null_ctl); // Null is the only remaining possibility. 2969 return intcon(0); 2970 } 2971 if (cast_obj != NULL) { 2972 not_null_obj = cast_obj; 2973 } 2974 } 2975 } 2976 2977 // Load the object's klass 2978 Node* obj_klass = load_object_klass(not_null_obj); 2979 2980 // Generate the subtype check 2981 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2982 2983 // Plug in the success path to the general merge in slot 1. 2984 region->init_req(_obj_path, control()); 2985 phi ->init_req(_obj_path, intcon(1)); 2986 2987 // Plug in the failing path to the general merge in slot 2. 2988 region->init_req(_fail_path, not_subtype_ctrl); 2989 phi ->init_req(_fail_path, intcon(0)); 2990 2991 // Return final merged results 2992 set_control( _gvn.transform(region) ); 2993 record_for_igvn(region); 2994 2995 // If we know the type check always succeeds then we don't use the 2996 // profiling data at this bytecode. Don't lose it, feed it to the 2997 // type system as a speculative type. 2998 if (safe_for_replace) { 2999 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3000 replace_in_map(obj, casted_obj); 3001 } 3002 3003 return _gvn.transform(phi); 3004 } 3005 3006 //-------------------------------gen_checkcast--------------------------------- 3007 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3008 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3009 // uncommon-trap paths work. Adjust stack after this call. 3010 // If failure_control is supplied and not null, it is filled in with 3011 // the control edge for the cast failure. Otherwise, an appropriate 3012 // uncommon trap or exception is thrown. 3013 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3014 Node* *failure_control) { 3015 kill_dead_locals(); // Benefit all the uncommon traps 3016 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3017 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3018 3019 // Fast cutout: Check the case that the cast is vacuously true. 3020 // This detects the common cases where the test will short-circuit 3021 // away completely. We do this before we perform the null check, 3022 // because if the test is going to turn into zero code, we don't 3023 // want a residual null check left around. (Causes a slowdown, 3024 // for example, in some objArray manipulations, such as a[i]=a[j].) 3025 if (tk->singleton()) { 3026 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3027 if (objtp != NULL && objtp->klass() != NULL) { 3028 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3029 case Compile::SSC_always_true: 3030 // If we know the type check always succeed then we don't use 3031 // the profiling data at this bytecode. Don't lose it, feed it 3032 // to the type system as a speculative type. 3033 return record_profiled_receiver_for_speculation(obj); 3034 case Compile::SSC_always_false: 3035 // It needs a null check because a null will *pass* the cast check. 3036 // A non-null value will always produce an exception. 3037 return null_assert(obj); 3038 } 3039 } 3040 } 3041 3042 ciProfileData* data = NULL; 3043 bool safe_for_replace = false; 3044 if (failure_control == NULL) { // use MDO in regular case only 3045 assert(java_bc() == Bytecodes::_aastore || 3046 java_bc() == Bytecodes::_checkcast, 3047 "interpreter profiles type checks only for these BCs"); 3048 data = method()->method_data()->bci_to_data(bci()); 3049 safe_for_replace = true; 3050 } 3051 3052 // Make the merge point 3053 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3054 RegionNode* region = new RegionNode(PATH_LIMIT); 3055 Node* phi = new PhiNode(region, toop); 3056 C->set_has_split_ifs(true); // Has chance for split-if optimization 3057 3058 // Use null-cast information if it is available 3059 bool speculative_not_null = false; 3060 bool never_see_null = ((failure_control == NULL) // regular case only 3061 && seems_never_null(obj, data, speculative_not_null)); 3062 3063 // Null check; get casted pointer; set region slot 3 3064 Node* null_ctl = top(); 3065 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3066 3067 // If not_null_obj is dead, only null-path is taken 3068 if (stopped()) { // Doing instance-of on a NULL? 3069 set_control(null_ctl); 3070 return null(); 3071 } 3072 region->init_req(_null_path, null_ctl); 3073 phi ->init_req(_null_path, null()); // Set null path value 3074 if (null_ctl == top()) { 3075 // Do this eagerly, so that pattern matches like is_diamond_phi 3076 // will work even during parsing. 3077 assert(_null_path == PATH_LIMIT-1, "delete last"); 3078 region->del_req(_null_path); 3079 phi ->del_req(_null_path); 3080 } 3081 3082 Node* cast_obj = NULL; 3083 if (tk->klass_is_exact()) { 3084 // The following optimization tries to statically cast the speculative type of the object 3085 // (for example obtained during profiling) to the type of the superklass and then do a 3086 // dynamic check that the type of the object is what we expect. To work correctly 3087 // for checkcast and aastore the type of superklass should be exact. 3088 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3089 // We may not have profiling here or it may not help us. If we have 3090 // a speculative type use it to perform an exact cast. 3091 ciKlass* spec_obj_type = obj_type->speculative_type(); 3092 if (spec_obj_type != NULL || data != NULL) { 3093 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3094 if (cast_obj != NULL) { 3095 if (failure_control != NULL) // failure is now impossible 3096 (*failure_control) = top(); 3097 // adjust the type of the phi to the exact klass: 3098 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3099 } 3100 } 3101 } 3102 3103 if (cast_obj == NULL) { 3104 // Load the object's klass 3105 Node* obj_klass = load_object_klass(not_null_obj); 3106 3107 // Generate the subtype check 3108 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3109 3110 // Plug in success path into the merge 3111 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3112 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3113 if (failure_control == NULL) { 3114 if (not_subtype_ctrl != top()) { // If failure is possible 3115 PreserveJVMState pjvms(this); 3116 set_control(not_subtype_ctrl); 3117 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3118 } 3119 } else { 3120 (*failure_control) = not_subtype_ctrl; 3121 } 3122 } 3123 3124 region->init_req(_obj_path, control()); 3125 phi ->init_req(_obj_path, cast_obj); 3126 3127 // A merge of NULL or Casted-NotNull obj 3128 Node* res = _gvn.transform(phi); 3129 3130 // Note I do NOT always 'replace_in_map(obj,result)' here. 3131 // if( tk->klass()->can_be_primary_super() ) 3132 // This means that if I successfully store an Object into an array-of-String 3133 // I 'forget' that the Object is really now known to be a String. I have to 3134 // do this because we don't have true union types for interfaces - if I store 3135 // a Baz into an array-of-Interface and then tell the optimizer it's an 3136 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3137 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3138 // replace_in_map( obj, res ); 3139 3140 // Return final merged results 3141 set_control( _gvn.transform(region) ); 3142 record_for_igvn(region); 3143 3144 return record_profiled_receiver_for_speculation(res); 3145 } 3146 3147 //------------------------------next_monitor----------------------------------- 3148 // What number should be given to the next monitor? 3149 int GraphKit::next_monitor() { 3150 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3151 int next = current + C->sync_stack_slots(); 3152 // Keep the toplevel high water mark current: 3153 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3154 return current; 3155 } 3156 3157 //------------------------------insert_mem_bar--------------------------------- 3158 // Memory barrier to avoid floating things around 3159 // The membar serves as a pinch point between both control and all memory slices. 3160 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3161 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3162 mb->init_req(TypeFunc::Control, control()); 3163 mb->init_req(TypeFunc::Memory, reset_memory()); 3164 Node* membar = _gvn.transform(mb); 3165 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3166 set_all_memory_call(membar); 3167 return membar; 3168 } 3169 3170 //-------------------------insert_mem_bar_volatile---------------------------- 3171 // Memory barrier to avoid floating things around 3172 // The membar serves as a pinch point between both control and memory(alias_idx). 3173 // If you want to make a pinch point on all memory slices, do not use this 3174 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3175 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3176 // When Parse::do_put_xxx updates a volatile field, it appends a series 3177 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3178 // The first membar is on the same memory slice as the field store opcode. 3179 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3180 // All the other membars (for other volatile slices, including AliasIdxBot, 3181 // which stands for all unknown volatile slices) are control-dependent 3182 // on the first membar. This prevents later volatile loads or stores 3183 // from sliding up past the just-emitted store. 3184 3185 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3186 mb->set_req(TypeFunc::Control,control()); 3187 if (alias_idx == Compile::AliasIdxBot) { 3188 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3189 } else { 3190 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3191 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3192 } 3193 Node* membar = _gvn.transform(mb); 3194 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3195 if (alias_idx == Compile::AliasIdxBot) { 3196 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3197 } else { 3198 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3199 } 3200 return membar; 3201 } 3202 3203 //------------------------------shared_lock------------------------------------ 3204 // Emit locking code. 3205 FastLockNode* GraphKit::shared_lock(Node* obj) { 3206 // bci is either a monitorenter bc or InvocationEntryBci 3207 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3208 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3209 3210 if( !GenerateSynchronizationCode ) 3211 return NULL; // Not locking things? 3212 if (stopped()) // Dead monitor? 3213 return NULL; 3214 3215 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3216 3217 // Box the stack location 3218 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3219 Node* mem = reset_memory(); 3220 3221 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3222 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3223 // Create the counters for this fast lock. 3224 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3225 } 3226 3227 // Create the rtm counters for this fast lock if needed. 3228 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3229 3230 // Add monitor to debug info for the slow path. If we block inside the 3231 // slow path and de-opt, we need the monitor hanging around 3232 map()->push_monitor( flock ); 3233 3234 const TypeFunc *tf = LockNode::lock_type(); 3235 LockNode *lock = new LockNode(C, tf); 3236 3237 lock->init_req( TypeFunc::Control, control() ); 3238 lock->init_req( TypeFunc::Memory , mem ); 3239 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3240 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3241 lock->init_req( TypeFunc::ReturnAdr, top() ); 3242 3243 lock->init_req(TypeFunc::Parms + 0, obj); 3244 lock->init_req(TypeFunc::Parms + 1, box); 3245 lock->init_req(TypeFunc::Parms + 2, flock); 3246 add_safepoint_edges(lock); 3247 3248 lock = _gvn.transform( lock )->as_Lock(); 3249 3250 // lock has no side-effects, sets few values 3251 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3252 3253 insert_mem_bar(Op_MemBarAcquireLock); 3254 3255 // Add this to the worklist so that the lock can be eliminated 3256 record_for_igvn(lock); 3257 3258 #ifndef PRODUCT 3259 if (PrintLockStatistics) { 3260 // Update the counter for this lock. Don't bother using an atomic 3261 // operation since we don't require absolute accuracy. 3262 lock->create_lock_counter(map()->jvms()); 3263 increment_counter(lock->counter()->addr()); 3264 } 3265 #endif 3266 3267 return flock; 3268 } 3269 3270 3271 //------------------------------shared_unlock---------------------------------- 3272 // Emit unlocking code. 3273 void GraphKit::shared_unlock(Node* box, Node* obj) { 3274 // bci is either a monitorenter bc or InvocationEntryBci 3275 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3276 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3277 3278 if( !GenerateSynchronizationCode ) 3279 return; 3280 if (stopped()) { // Dead monitor? 3281 map()->pop_monitor(); // Kill monitor from debug info 3282 return; 3283 } 3284 3285 // Memory barrier to avoid floating things down past the locked region 3286 insert_mem_bar(Op_MemBarReleaseLock); 3287 3288 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3289 UnlockNode *unlock = new UnlockNode(C, tf); 3290 #ifdef ASSERT 3291 unlock->set_dbg_jvms(sync_jvms()); 3292 #endif 3293 uint raw_idx = Compile::AliasIdxRaw; 3294 unlock->init_req( TypeFunc::Control, control() ); 3295 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3296 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3297 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3298 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3299 3300 unlock->init_req(TypeFunc::Parms + 0, obj); 3301 unlock->init_req(TypeFunc::Parms + 1, box); 3302 unlock = _gvn.transform(unlock)->as_Unlock(); 3303 3304 Node* mem = reset_memory(); 3305 3306 // unlock has no side-effects, sets few values 3307 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3308 3309 // Kill monitor from debug info 3310 map()->pop_monitor( ); 3311 } 3312 3313 //-------------------------------get_layout_helper----------------------------- 3314 // If the given klass is a constant or known to be an array, 3315 // fetch the constant layout helper value into constant_value 3316 // and return (Node*)NULL. Otherwise, load the non-constant 3317 // layout helper value, and return the node which represents it. 3318 // This two-faced routine is useful because allocation sites 3319 // almost always feature constant types. 3320 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3321 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3322 if (!StressReflectiveCode && inst_klass != NULL) { 3323 ciKlass* klass = inst_klass->klass(); 3324 bool xklass = inst_klass->klass_is_exact(); 3325 if (xklass || klass->is_array_klass()) { 3326 jint lhelper = klass->layout_helper(); 3327 if (lhelper != Klass::_lh_neutral_value) { 3328 constant_value = lhelper; 3329 return (Node*) NULL; 3330 } 3331 } 3332 } 3333 constant_value = Klass::_lh_neutral_value; // put in a known value 3334 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3335 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3336 } 3337 3338 // We just put in an allocate/initialize with a big raw-memory effect. 3339 // Hook selected additional alias categories on the initialization. 3340 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3341 MergeMemNode* init_in_merge, 3342 Node* init_out_raw) { 3343 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3344 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3345 3346 Node* prevmem = kit.memory(alias_idx); 3347 init_in_merge->set_memory_at(alias_idx, prevmem); 3348 kit.set_memory(init_out_raw, alias_idx); 3349 } 3350 3351 //---------------------------set_output_for_allocation------------------------- 3352 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3353 const TypeOopPtr* oop_type, 3354 bool deoptimize_on_exception) { 3355 int rawidx = Compile::AliasIdxRaw; 3356 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3357 add_safepoint_edges(alloc); 3358 Node* allocx = _gvn.transform(alloc); 3359 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3360 // create memory projection for i_o 3361 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3362 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3363 3364 // create a memory projection as for the normal control path 3365 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3366 set_memory(malloc, rawidx); 3367 3368 // a normal slow-call doesn't change i_o, but an allocation does 3369 // we create a separate i_o projection for the normal control path 3370 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3371 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3372 3373 // put in an initialization barrier 3374 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3375 rawoop)->as_Initialize(); 3376 assert(alloc->initialization() == init, "2-way macro link must work"); 3377 assert(init ->allocation() == alloc, "2-way macro link must work"); 3378 { 3379 // Extract memory strands which may participate in the new object's 3380 // initialization, and source them from the new InitializeNode. 3381 // This will allow us to observe initializations when they occur, 3382 // and link them properly (as a group) to the InitializeNode. 3383 assert(init->in(InitializeNode::Memory) == malloc, ""); 3384 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3385 init->set_req(InitializeNode::Memory, minit_in); 3386 record_for_igvn(minit_in); // fold it up later, if possible 3387 Node* minit_out = memory(rawidx); 3388 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3389 if (oop_type->isa_aryptr()) { 3390 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3391 int elemidx = C->get_alias_index(telemref); 3392 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3393 } else if (oop_type->isa_instptr()) { 3394 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3395 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3396 ciField* field = ik->nonstatic_field_at(i); 3397 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3398 continue; // do not bother to track really large numbers of fields 3399 // Find (or create) the alias category for this field: 3400 int fieldidx = C->alias_type(field)->index(); 3401 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3402 } 3403 } 3404 } 3405 3406 // Cast raw oop to the real thing... 3407 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3408 javaoop = _gvn.transform(javaoop); 3409 C->set_recent_alloc(control(), javaoop); 3410 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3411 3412 #ifdef ASSERT 3413 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3414 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3415 "Ideal_allocation works"); 3416 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3417 "Ideal_allocation works"); 3418 if (alloc->is_AllocateArray()) { 3419 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3420 "Ideal_allocation works"); 3421 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3422 "Ideal_allocation works"); 3423 } else { 3424 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3425 } 3426 } 3427 #endif //ASSERT 3428 3429 return javaoop; 3430 } 3431 3432 //---------------------------new_instance-------------------------------------- 3433 // This routine takes a klass_node which may be constant (for a static type) 3434 // or may be non-constant (for reflective code). It will work equally well 3435 // for either, and the graph will fold nicely if the optimizer later reduces 3436 // the type to a constant. 3437 // The optional arguments are for specialized use by intrinsics: 3438 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3439 // - If 'return_size_val', report the the total object size to the caller. 3440 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3441 Node* GraphKit::new_instance(Node* klass_node, 3442 Node* extra_slow_test, 3443 Node* *return_size_val, 3444 bool deoptimize_on_exception) { 3445 // Compute size in doublewords 3446 // The size is always an integral number of doublewords, represented 3447 // as a positive bytewise size stored in the klass's layout_helper. 3448 // The layout_helper also encodes (in a low bit) the need for a slow path. 3449 jint layout_con = Klass::_lh_neutral_value; 3450 Node* layout_val = get_layout_helper(klass_node, layout_con); 3451 int layout_is_con = (layout_val == NULL); 3452 3453 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3454 // Generate the initial go-slow test. It's either ALWAYS (return a 3455 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3456 // case) a computed value derived from the layout_helper. 3457 Node* initial_slow_test = NULL; 3458 if (layout_is_con) { 3459 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3460 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3461 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3462 } else { // reflective case 3463 // This reflective path is used by Unsafe.allocateInstance. 3464 // (It may be stress-tested by specifying StressReflectiveCode.) 3465 // Basically, we want to get into the VM is there's an illegal argument. 3466 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3467 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3468 if (extra_slow_test != intcon(0)) { 3469 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3470 } 3471 // (Macro-expander will further convert this to a Bool, if necessary.) 3472 } 3473 3474 // Find the size in bytes. This is easy; it's the layout_helper. 3475 // The size value must be valid even if the slow path is taken. 3476 Node* size = NULL; 3477 if (layout_is_con) { 3478 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3479 } else { // reflective case 3480 // This reflective path is used by clone and Unsafe.allocateInstance. 3481 size = ConvI2X(layout_val); 3482 3483 // Clear the low bits to extract layout_helper_size_in_bytes: 3484 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3485 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3486 size = _gvn.transform( new AndXNode(size, mask) ); 3487 } 3488 if (return_size_val != NULL) { 3489 (*return_size_val) = size; 3490 } 3491 3492 // This is a precise notnull oop of the klass. 3493 // (Actually, it need not be precise if this is a reflective allocation.) 3494 // It's what we cast the result to. 3495 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3496 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3497 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3498 3499 // Now generate allocation code 3500 3501 // The entire memory state is needed for slow path of the allocation 3502 // since GC and deoptimization can happened. 3503 Node *mem = reset_memory(); 3504 set_all_memory(mem); // Create new memory state 3505 3506 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3507 control(), mem, i_o(), 3508 size, klass_node, 3509 initial_slow_test); 3510 3511 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3512 } 3513 3514 //-------------------------------new_array------------------------------------- 3515 // helper for both newarray and anewarray 3516 // The 'length' parameter is (obviously) the length of the array. 3517 // See comments on new_instance for the meaning of the other arguments. 3518 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3519 Node* length, // number of array elements 3520 int nargs, // number of arguments to push back for uncommon trap 3521 Node* *return_size_val, 3522 bool deoptimize_on_exception) { 3523 jint layout_con = Klass::_lh_neutral_value; 3524 Node* layout_val = get_layout_helper(klass_node, layout_con); 3525 int layout_is_con = (layout_val == NULL); 3526 3527 if (!layout_is_con && !StressReflectiveCode && 3528 !too_many_traps(Deoptimization::Reason_class_check)) { 3529 // This is a reflective array creation site. 3530 // Optimistically assume that it is a subtype of Object[], 3531 // so that we can fold up all the address arithmetic. 3532 layout_con = Klass::array_layout_helper(T_OBJECT); 3533 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3534 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3535 { BuildCutout unless(this, bol_lh, PROB_MAX); 3536 inc_sp(nargs); 3537 uncommon_trap(Deoptimization::Reason_class_check, 3538 Deoptimization::Action_maybe_recompile); 3539 } 3540 layout_val = NULL; 3541 layout_is_con = true; 3542 } 3543 3544 // Generate the initial go-slow test. Make sure we do not overflow 3545 // if length is huge (near 2Gig) or negative! We do not need 3546 // exact double-words here, just a close approximation of needed 3547 // double-words. We can't add any offset or rounding bits, lest we 3548 // take a size -1 of bytes and make it positive. Use an unsigned 3549 // compare, so negative sizes look hugely positive. 3550 int fast_size_limit = FastAllocateSizeLimit; 3551 if (layout_is_con) { 3552 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3553 // Increase the size limit if we have exact knowledge of array type. 3554 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3555 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3556 } 3557 3558 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3559 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3560 3561 // --- Size Computation --- 3562 // array_size = round_to_heap(array_header + (length << elem_shift)); 3563 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3564 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3565 // The rounding mask is strength-reduced, if possible. 3566 int round_mask = MinObjAlignmentInBytes - 1; 3567 Node* header_size = NULL; 3568 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3569 // (T_BYTE has the weakest alignment and size restrictions...) 3570 if (layout_is_con) { 3571 int hsize = Klass::layout_helper_header_size(layout_con); 3572 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3573 BasicType etype = Klass::layout_helper_element_type(layout_con); 3574 if ((round_mask & ~right_n_bits(eshift)) == 0) 3575 round_mask = 0; // strength-reduce it if it goes away completely 3576 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3577 assert(header_size_min <= hsize, "generic minimum is smallest"); 3578 header_size_min = hsize; 3579 header_size = intcon(hsize + round_mask); 3580 } else { 3581 Node* hss = intcon(Klass::_lh_header_size_shift); 3582 Node* hsm = intcon(Klass::_lh_header_size_mask); 3583 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3584 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3585 Node* mask = intcon(round_mask); 3586 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3587 } 3588 3589 Node* elem_shift = NULL; 3590 if (layout_is_con) { 3591 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3592 if (eshift != 0) 3593 elem_shift = intcon(eshift); 3594 } else { 3595 // There is no need to mask or shift this value. 3596 // The semantics of LShiftINode include an implicit mask to 0x1F. 3597 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3598 elem_shift = layout_val; 3599 } 3600 3601 // Transition to native address size for all offset calculations: 3602 Node* lengthx = ConvI2X(length); 3603 Node* headerx = ConvI2X(header_size); 3604 #ifdef _LP64 3605 { const TypeInt* tilen = _gvn.find_int_type(length); 3606 if (tilen != NULL && tilen->_lo < 0) { 3607 // Add a manual constraint to a positive range. Cf. array_element_address. 3608 jint size_max = fast_size_limit; 3609 if (size_max > tilen->_hi) size_max = tilen->_hi; 3610 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3611 3612 // Only do a narrow I2L conversion if the range check passed. 3613 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3614 _gvn.transform(iff); 3615 RegionNode* region = new RegionNode(3); 3616 _gvn.set_type(region, Type::CONTROL); 3617 lengthx = new PhiNode(region, TypeLong::LONG); 3618 _gvn.set_type(lengthx, TypeLong::LONG); 3619 3620 // Range check passed. Use ConvI2L node with narrow type. 3621 Node* passed = IfFalse(iff); 3622 region->init_req(1, passed); 3623 // Make I2L conversion control dependent to prevent it from 3624 // floating above the range check during loop optimizations. 3625 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3626 3627 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3628 region->init_req(2, IfTrue(iff)); 3629 lengthx->init_req(2, ConvI2X(length)); 3630 3631 set_control(region); 3632 record_for_igvn(region); 3633 record_for_igvn(lengthx); 3634 } 3635 } 3636 #endif 3637 3638 // Combine header size (plus rounding) and body size. Then round down. 3639 // This computation cannot overflow, because it is used only in two 3640 // places, one where the length is sharply limited, and the other 3641 // after a successful allocation. 3642 Node* abody = lengthx; 3643 if (elem_shift != NULL) 3644 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3645 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3646 if (round_mask != 0) { 3647 Node* mask = MakeConX(~round_mask); 3648 size = _gvn.transform( new AndXNode(size, mask) ); 3649 } 3650 // else if round_mask == 0, the size computation is self-rounding 3651 3652 if (return_size_val != NULL) { 3653 // This is the size 3654 (*return_size_val) = size; 3655 } 3656 3657 // Now generate allocation code 3658 3659 // The entire memory state is needed for slow path of the allocation 3660 // since GC and deoptimization can happened. 3661 Node *mem = reset_memory(); 3662 set_all_memory(mem); // Create new memory state 3663 3664 if (initial_slow_test->is_Bool()) { 3665 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3666 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3667 } 3668 3669 // Create the AllocateArrayNode and its result projections 3670 AllocateArrayNode* alloc 3671 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3672 control(), mem, i_o(), 3673 size, klass_node, 3674 initial_slow_test, 3675 length); 3676 3677 // Cast to correct type. Note that the klass_node may be constant or not, 3678 // and in the latter case the actual array type will be inexact also. 3679 // (This happens via a non-constant argument to inline_native_newArray.) 3680 // In any case, the value of klass_node provides the desired array type. 3681 const TypeInt* length_type = _gvn.find_int_type(length); 3682 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3683 if (ary_type->isa_aryptr() && length_type != NULL) { 3684 // Try to get a better type than POS for the size 3685 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3686 } 3687 3688 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3689 3690 // Cast length on remaining path to be as narrow as possible 3691 if (map()->find_edge(length) >= 0) { 3692 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3693 if (ccast != length) { 3694 _gvn.set_type_bottom(ccast); 3695 record_for_igvn(ccast); 3696 replace_in_map(length, ccast); 3697 } 3698 } 3699 3700 return javaoop; 3701 } 3702 3703 // The following "Ideal_foo" functions are placed here because they recognize 3704 // the graph shapes created by the functions immediately above. 3705 3706 //---------------------------Ideal_allocation---------------------------------- 3707 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3708 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3709 if (ptr == NULL) { // reduce dumb test in callers 3710 return NULL; 3711 } 3712 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3713 ptr = ptr->in(1); 3714 if (ptr == NULL) return NULL; 3715 } 3716 // Return NULL for allocations with several casts: 3717 // j.l.reflect.Array.newInstance(jobject, jint) 3718 // Object.clone() 3719 // to keep more precise type from last cast. 3720 if (ptr->is_Proj()) { 3721 Node* allo = ptr->in(0); 3722 if (allo != NULL && allo->is_Allocate()) { 3723 return allo->as_Allocate(); 3724 } 3725 } 3726 // Report failure to match. 3727 return NULL; 3728 } 3729 3730 // Fancy version which also strips off an offset (and reports it to caller). 3731 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3732 intptr_t& offset) { 3733 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3734 if (base == NULL) return NULL; 3735 return Ideal_allocation(base, phase); 3736 } 3737 3738 // Trace Initialize <- Proj[Parm] <- Allocate 3739 AllocateNode* InitializeNode::allocation() { 3740 Node* rawoop = in(InitializeNode::RawAddress); 3741 if (rawoop->is_Proj()) { 3742 Node* alloc = rawoop->in(0); 3743 if (alloc->is_Allocate()) { 3744 return alloc->as_Allocate(); 3745 } 3746 } 3747 return NULL; 3748 } 3749 3750 // Trace Allocate -> Proj[Parm] -> Initialize 3751 InitializeNode* AllocateNode::initialization() { 3752 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3753 if (rawoop == NULL) return NULL; 3754 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3755 Node* init = rawoop->fast_out(i); 3756 if (init->is_Initialize()) { 3757 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3758 return init->as_Initialize(); 3759 } 3760 } 3761 return NULL; 3762 } 3763 3764 //----------------------------- loop predicates --------------------------- 3765 3766 //------------------------------add_predicate_impl---------------------------- 3767 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3768 // Too many traps seen? 3769 if (too_many_traps(reason)) { 3770 #ifdef ASSERT 3771 if (TraceLoopPredicate) { 3772 int tc = C->trap_count(reason); 3773 tty->print("too many traps=%s tcount=%d in ", 3774 Deoptimization::trap_reason_name(reason), tc); 3775 method()->print(); // which method has too many predicate traps 3776 tty->cr(); 3777 } 3778 #endif 3779 // We cannot afford to take more traps here, 3780 // do not generate predicate. 3781 return; 3782 } 3783 3784 Node *cont = _gvn.intcon(1); 3785 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3786 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3787 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3788 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3789 C->add_predicate_opaq(opq); 3790 { 3791 PreserveJVMState pjvms(this); 3792 set_control(iffalse); 3793 inc_sp(nargs); 3794 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3795 } 3796 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3797 set_control(iftrue); 3798 } 3799 3800 //------------------------------add_predicate--------------------------------- 3801 void GraphKit::add_predicate(int nargs) { 3802 if (UseLoopPredicate) { 3803 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3804 } 3805 // loop's limit check predicate should be near the loop. 3806 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3807 } 3808 3809 //----------------------------- store barriers ---------------------------- 3810 #define __ ideal. 3811 3812 bool GraphKit::use_ReduceInitialCardMarks() { 3813 BarrierSet *bs = Universe::heap()->barrier_set(); 3814 return bs->is_a(BarrierSet::CardTableBarrierSet) 3815 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers() 3816 && ReduceInitialCardMarks; 3817 } 3818 3819 void GraphKit::sync_kit(IdealKit& ideal) { 3820 set_all_memory(__ merged_memory()); 3821 set_i_o(__ i_o()); 3822 set_control(__ ctrl()); 3823 } 3824 3825 void GraphKit::final_sync(IdealKit& ideal) { 3826 // Final sync IdealKit and graphKit. 3827 sync_kit(ideal); 3828 } 3829 3830 Node* GraphKit::byte_map_base_node() { 3831 // Get base of card map 3832 jbyte* card_table_base = ci_card_table_address(); 3833 if (card_table_base != NULL) { 3834 return makecon(TypeRawPtr::make((address)card_table_base)); 3835 } else { 3836 return null(); 3837 } 3838 } 3839 3840 // vanilla/CMS post barrier 3841 // Insert a write-barrier store. This is to let generational GC work; we have 3842 // to flag all oop-stores before the next GC point. 3843 void GraphKit::write_barrier_post(Node* oop_store, 3844 Node* obj, 3845 Node* adr, 3846 uint adr_idx, 3847 Node* val, 3848 bool use_precise) { 3849 // No store check needed if we're storing a NULL or an old object 3850 // (latter case is probably a string constant). The concurrent 3851 // mark sweep garbage collector, however, needs to have all nonNull 3852 // oop updates flagged via card-marks. 3853 if (val != NULL && val->is_Con()) { 3854 // must be either an oop or NULL 3855 const Type* t = val->bottom_type(); 3856 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3857 // stores of null never (?) need barriers 3858 return; 3859 } 3860 3861 if (use_ReduceInitialCardMarks() 3862 && obj == just_allocated_object(control())) { 3863 // We can skip marks on a freshly-allocated object in Eden. 3864 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 3865 // That routine informs GC to take appropriate compensating steps, 3866 // upon a slow-path allocation, so as to make this card-mark 3867 // elision safe. 3868 return; 3869 } 3870 3871 if (!use_precise) { 3872 // All card marks for a (non-array) instance are in one place: 3873 adr = obj; 3874 } 3875 // (Else it's an array (or unknown), and we want more precise card marks.) 3876 assert(adr != NULL, ""); 3877 3878 IdealKit ideal(this, true); 3879 3880 // Convert the pointer to an int prior to doing math on it 3881 Node* cast = __ CastPX(__ ctrl(), adr); 3882 3883 // Divide by card size 3884 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableBarrierSet), 3885 "Only one we handle so far."); 3886 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 3887 3888 // Combine card table base and card offset 3889 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3890 3891 // Get the alias_index for raw card-mark memory 3892 int adr_type = Compile::AliasIdxRaw; 3893 Node* zero = __ ConI(0); // Dirty card value 3894 BasicType bt = T_BYTE; 3895 3896 if (UseConcMarkSweepGC && UseCondCardMark) { 3897 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3898 __ sync_kit(this); 3899 } 3900 3901 if (UseCondCardMark) { 3902 // The classic GC reference write barrier is typically implemented 3903 // as a store into the global card mark table. Unfortunately 3904 // unconditional stores can result in false sharing and excessive 3905 // coherence traffic as well as false transactional aborts. 3906 // UseCondCardMark enables MP "polite" conditional card mark 3907 // stores. In theory we could relax the load from ctrl() to 3908 // no_ctrl, but that doesn't buy much latitude. 3909 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3910 __ if_then(card_val, BoolTest::ne, zero); 3911 } 3912 3913 // Smash zero into card 3914 if( !UseConcMarkSweepGC ) { 3915 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3916 } else { 3917 // Specialized path for CM store barrier 3918 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3919 } 3920 3921 if (UseCondCardMark) { 3922 __ end_if(); 3923 } 3924 3925 // Final sync IdealKit and GraphKit. 3926 final_sync(ideal); 3927 } 3928 /* 3929 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3930 * required by SATB to make sure all objects live at the start of the 3931 * marking are kept alive, all reference updates need to any previous 3932 * reference stored before writing. 3933 * 3934 * If the previous value is NULL there is no need to save the old value. 3935 * References that are NULL are filtered during runtime by the barrier 3936 * code to avoid unnecessary queuing. 3937 * 3938 * However in the case of newly allocated objects it might be possible to 3939 * prove that the reference about to be overwritten is NULL during compile 3940 * time and avoid adding the barrier code completely. 3941 * 3942 * The compiler needs to determine that the object in which a field is about 3943 * to be written is newly allocated, and that no prior store to the same field 3944 * has happened since the allocation. 3945 * 3946 * Returns true if the pre-barrier can be removed 3947 */ 3948 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3949 BasicType bt, uint adr_idx) { 3950 intptr_t offset = 0; 3951 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3952 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3953 3954 if (offset == Type::OffsetBot) { 3955 return false; // cannot unalias unless there are precise offsets 3956 } 3957 3958 if (alloc == NULL) { 3959 return false; // No allocation found 3960 } 3961 3962 intptr_t size_in_bytes = type2aelembytes(bt); 3963 3964 Node* mem = memory(adr_idx); // start searching here... 3965 3966 for (int cnt = 0; cnt < 50; cnt++) { 3967 3968 if (mem->is_Store()) { 3969 3970 Node* st_adr = mem->in(MemNode::Address); 3971 intptr_t st_offset = 0; 3972 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3973 3974 if (st_base == NULL) { 3975 break; // inscrutable pointer 3976 } 3977 3978 // Break we have found a store with same base and offset as ours so break 3979 if (st_base == base && st_offset == offset) { 3980 break; 3981 } 3982 3983 if (st_offset != offset && st_offset != Type::OffsetBot) { 3984 const int MAX_STORE = BytesPerLong; 3985 if (st_offset >= offset + size_in_bytes || 3986 st_offset <= offset - MAX_STORE || 3987 st_offset <= offset - mem->as_Store()->memory_size()) { 3988 // Success: The offsets are provably independent. 3989 // (You may ask, why not just test st_offset != offset and be done? 3990 // The answer is that stores of different sizes can co-exist 3991 // in the same sequence of RawMem effects. We sometimes initialize 3992 // a whole 'tile' of array elements with a single jint or jlong.) 3993 mem = mem->in(MemNode::Memory); 3994 continue; // advance through independent store memory 3995 } 3996 } 3997 3998 if (st_base != base 3999 && MemNode::detect_ptr_independence(base, alloc, st_base, 4000 AllocateNode::Ideal_allocation(st_base, phase), 4001 phase)) { 4002 // Success: The bases are provably independent. 4003 mem = mem->in(MemNode::Memory); 4004 continue; // advance through independent store memory 4005 } 4006 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4007 4008 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4009 AllocateNode* st_alloc = st_init->allocation(); 4010 4011 // Make sure that we are looking at the same allocation site. 4012 // The alloc variable is guaranteed to not be null here from earlier check. 4013 if (alloc == st_alloc) { 4014 // Check that the initialization is storing NULL so that no previous store 4015 // has been moved up and directly write a reference 4016 Node* captured_store = st_init->find_captured_store(offset, 4017 type2aelembytes(T_OBJECT), 4018 phase); 4019 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4020 return true; 4021 } 4022 } 4023 } 4024 4025 // Unless there is an explicit 'continue', we must bail out here, 4026 // because 'mem' is an inscrutable memory state (e.g., a call). 4027 break; 4028 } 4029 4030 return false; 4031 } 4032 4033 // G1 pre/post barriers 4034 void GraphKit::g1_write_barrier_pre(bool do_load, 4035 Node* obj, 4036 Node* adr, 4037 uint alias_idx, 4038 Node* val, 4039 const TypeOopPtr* val_type, 4040 Node* pre_val, 4041 BasicType bt) { 4042 4043 // Some sanity checks 4044 // Note: val is unused in this routine. 4045 4046 if (do_load) { 4047 // We need to generate the load of the previous value 4048 assert(obj != NULL, "must have a base"); 4049 assert(adr != NULL, "where are loading from?"); 4050 assert(pre_val == NULL, "loaded already?"); 4051 assert(val_type != NULL, "need a type"); 4052 4053 if (use_ReduceInitialCardMarks() 4054 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4055 return; 4056 } 4057 4058 } else { 4059 // In this case both val_type and alias_idx are unused. 4060 assert(pre_val != NULL, "must be loaded already"); 4061 // Nothing to be done if pre_val is null. 4062 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4063 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4064 } 4065 assert(bt == T_OBJECT, "or we shouldn't be here"); 4066 4067 IdealKit ideal(this, true); 4068 4069 Node* tls = __ thread(); // ThreadLocalStorage 4070 4071 Node* no_ctrl = NULL; 4072 Node* no_base = __ top(); 4073 Node* zero = __ ConI(0); 4074 Node* zeroX = __ ConX(0); 4075 4076 float likely = PROB_LIKELY(0.999); 4077 float unlikely = PROB_UNLIKELY(0.999); 4078 4079 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4080 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4081 4082 // Offsets into the thread 4083 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4084 SATBMarkQueue::byte_offset_of_active()); 4085 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4086 SATBMarkQueue::byte_offset_of_index()); 4087 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4088 SATBMarkQueue::byte_offset_of_buf()); 4089 4090 // Now the actual pointers into the thread 4091 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4092 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4093 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4094 4095 // Now some of the values 4096 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4097 4098 // if (!marking) 4099 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4100 BasicType index_bt = TypeX_X->basic_type(); 4101 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4102 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4103 4104 if (do_load) { 4105 // load original value 4106 // alias_idx correct?? 4107 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4108 } 4109 4110 // if (pre_val != NULL) 4111 __ if_then(pre_val, BoolTest::ne, null()); { 4112 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4113 4114 // is the queue for this thread full? 4115 __ if_then(index, BoolTest::ne, zeroX, likely); { 4116 4117 // decrement the index 4118 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4119 4120 // Now get the buffer location we will log the previous value into and store it 4121 Node *log_addr = __ AddP(no_base, buffer, next_index); 4122 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4123 // update the index 4124 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4125 4126 } __ else_(); { 4127 4128 // logging buffer is full, call the runtime 4129 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4130 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4131 } __ end_if(); // (!index) 4132 } __ end_if(); // (pre_val != NULL) 4133 } __ end_if(); // (!marking) 4134 4135 // Final sync IdealKit and GraphKit. 4136 final_sync(ideal); 4137 } 4138 4139 /* 4140 * G1 similar to any GC with a Young Generation requires a way to keep track of 4141 * references from Old Generation to Young Generation to make sure all live 4142 * objects are found. G1 also requires to keep track of object references 4143 * between different regions to enable evacuation of old regions, which is done 4144 * as part of mixed collections. References are tracked in remembered sets and 4145 * is continuously updated as reference are written to with the help of the 4146 * post-barrier. 4147 * 4148 * To reduce the number of updates to the remembered set the post-barrier 4149 * filters updates to fields in objects located in the Young Generation, 4150 * the same region as the reference, when the NULL is being written or 4151 * if the card is already marked as dirty by an earlier write. 4152 * 4153 * Under certain circumstances it is possible to avoid generating the 4154 * post-barrier completely if it is possible during compile time to prove 4155 * the object is newly allocated and that no safepoint exists between the 4156 * allocation and the store. 4157 * 4158 * In the case of slow allocation the allocation code must handle the barrier 4159 * as part of the allocation in the case the allocated object is not located 4160 * in the nursery, this would happen for humongous objects. This is similar to 4161 * how CMS is required to handle this case, see the comments for the method 4162 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier. 4163 * A deferred card mark is required for these objects and handled in the above 4164 * mentioned methods. 4165 * 4166 * Returns true if the post barrier can be removed 4167 */ 4168 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4169 Node* adr) { 4170 intptr_t offset = 0; 4171 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4172 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4173 4174 if (offset == Type::OffsetBot) { 4175 return false; // cannot unalias unless there are precise offsets 4176 } 4177 4178 if (alloc == NULL) { 4179 return false; // No allocation found 4180 } 4181 4182 // Start search from Store node 4183 Node* mem = store->in(MemNode::Control); 4184 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4185 4186 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4187 AllocateNode* st_alloc = st_init->allocation(); 4188 4189 // Make sure we are looking at the same allocation 4190 if (alloc == st_alloc) { 4191 return true; 4192 } 4193 } 4194 4195 return false; 4196 } 4197 4198 // 4199 // Update the card table and add card address to the queue 4200 // 4201 void GraphKit::g1_mark_card(IdealKit& ideal, 4202 Node* card_adr, 4203 Node* oop_store, 4204 uint oop_alias_idx, 4205 Node* index, 4206 Node* index_adr, 4207 Node* buffer, 4208 const TypeFunc* tf) { 4209 4210 Node* zero = __ ConI(0); 4211 Node* zeroX = __ ConX(0); 4212 Node* no_base = __ top(); 4213 BasicType card_bt = T_BYTE; 4214 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4215 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4216 4217 // Now do the queue work 4218 __ if_then(index, BoolTest::ne, zeroX); { 4219 4220 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4221 Node* log_addr = __ AddP(no_base, buffer, next_index); 4222 4223 // Order, see storeCM. 4224 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4225 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4226 4227 } __ else_(); { 4228 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4229 } __ end_if(); 4230 4231 } 4232 4233 void GraphKit::g1_write_barrier_post(Node* oop_store, 4234 Node* obj, 4235 Node* adr, 4236 uint alias_idx, 4237 Node* val, 4238 BasicType bt, 4239 bool use_precise) { 4240 // If we are writing a NULL then we need no post barrier 4241 4242 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4243 // Must be NULL 4244 const Type* t = val->bottom_type(); 4245 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4246 // No post barrier if writing NULLx 4247 return; 4248 } 4249 4250 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4251 // We can skip marks on a freshly-allocated object in Eden. 4252 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp. 4253 // That routine informs GC to take appropriate compensating steps, 4254 // upon a slow-path allocation, so as to make this card-mark 4255 // elision safe. 4256 return; 4257 } 4258 4259 if (use_ReduceInitialCardMarks() 4260 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4261 return; 4262 } 4263 4264 if (!use_precise) { 4265 // All card marks for a (non-array) instance are in one place: 4266 adr = obj; 4267 } 4268 // (Else it's an array (or unknown), and we want more precise card marks.) 4269 assert(adr != NULL, ""); 4270 4271 IdealKit ideal(this, true); 4272 4273 Node* tls = __ thread(); // ThreadLocalStorage 4274 4275 Node* no_base = __ top(); 4276 float likely = PROB_LIKELY(0.999); 4277 float unlikely = PROB_UNLIKELY(0.999); 4278 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val()); 4279 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val()); 4280 Node* zeroX = __ ConX(0); 4281 4282 // Get the alias_index for raw card-mark memory 4283 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4284 4285 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4286 4287 // Offsets into the thread 4288 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4289 DirtyCardQueue::byte_offset_of_index()); 4290 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4291 DirtyCardQueue::byte_offset_of_buf()); 4292 4293 // Pointers into the thread 4294 4295 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4296 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4297 4298 // Now some values 4299 // Use ctrl to avoid hoisting these values past a safepoint, which could 4300 // potentially reset these fields in the JavaThread. 4301 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4302 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4303 4304 // Convert the store obj pointer to an int prior to doing math on it 4305 // Must use ctrl to prevent "integerized oop" existing across safepoint 4306 Node* cast = __ CastPX(__ ctrl(), adr); 4307 4308 // Divide pointer by card size 4309 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) ); 4310 4311 // Combine card table base and card offset 4312 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4313 4314 // If we know the value being stored does it cross regions? 4315 4316 if (val != NULL) { 4317 // Does the store cause us to cross regions? 4318 4319 // Should be able to do an unsigned compare of region_size instead of 4320 // and extra shift. Do we have an unsigned compare?? 4321 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4322 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4323 4324 // if (xor_res == 0) same region so skip 4325 __ if_then(xor_res, BoolTest::ne, zeroX); { 4326 4327 // No barrier if we are storing a NULL 4328 __ if_then(val, BoolTest::ne, null(), unlikely); { 4329 4330 // Ok must mark the card if not already dirty 4331 4332 // load the original value of the card 4333 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4334 4335 __ if_then(card_val, BoolTest::ne, young_card); { 4336 sync_kit(ideal); 4337 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4338 insert_mem_bar(Op_MemBarVolatile, oop_store); 4339 __ sync_kit(this); 4340 4341 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4342 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4343 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4344 } __ end_if(); 4345 } __ end_if(); 4346 } __ end_if(); 4347 } __ end_if(); 4348 } else { 4349 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4350 // We don't need a barrier here if the destination is a newly allocated object 4351 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4352 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()). 4353 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4354 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4355 __ if_then(card_val, BoolTest::ne, young_card); { 4356 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4357 } __ end_if(); 4358 } 4359 4360 // Final sync IdealKit and GraphKit. 4361 final_sync(ideal); 4362 } 4363 #undef __ 4364 4365 4366 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4367 Node* len = load_array_length(load_String_value(ctrl, str)); 4368 Node* coder = load_String_coder(ctrl, str); 4369 // Divide length by 2 if coder is UTF16 4370 return _gvn.transform(new RShiftINode(len, coder)); 4371 } 4372 4373 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4374 int value_offset = java_lang_String::value_offset_in_bytes(); 4375 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4376 false, NULL, 0); 4377 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4378 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4379 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4380 ciTypeArrayKlass::make(T_BYTE), true, 0); 4381 int value_field_idx = C->get_alias_index(value_field_type); 4382 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4383 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4384 // String.value field is known to be @Stable. 4385 if (UseImplicitStableValues) { 4386 load = cast_array_to_stable(load, value_type); 4387 } 4388 return load; 4389 } 4390 4391 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4392 if (!CompactStrings) { 4393 return intcon(java_lang_String::CODER_UTF16); 4394 } 4395 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4396 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4397 false, NULL, 0); 4398 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4399 int coder_field_idx = C->get_alias_index(coder_field_type); 4400 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4401 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4402 } 4403 4404 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4405 int value_offset = java_lang_String::value_offset_in_bytes(); 4406 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4407 false, NULL, 0); 4408 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4409 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4410 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4411 } 4412 4413 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4414 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4415 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4416 false, NULL, 0); 4417 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4418 int coder_field_idx = C->get_alias_index(coder_field_type); 4419 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4420 value, T_BYTE, coder_field_idx, MemNode::unordered); 4421 } 4422 4423 // Capture src and dst memory state with a MergeMemNode 4424 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4425 if (src_type == dst_type) { 4426 // Types are equal, we don't need a MergeMemNode 4427 return memory(src_type); 4428 } 4429 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4430 record_for_igvn(merge); // fold it up later, if possible 4431 int src_idx = C->get_alias_index(src_type); 4432 int dst_idx = C->get_alias_index(dst_type); 4433 merge->set_memory_at(src_idx, memory(src_idx)); 4434 merge->set_memory_at(dst_idx, memory(dst_idx)); 4435 return merge; 4436 } 4437 4438 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4439 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4440 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4441 // If input and output memory types differ, capture both states to preserve 4442 // the dependency between preceding and subsequent loads/stores. 4443 // For example, the following program: 4444 // StoreB 4445 // compress_string 4446 // LoadB 4447 // has this memory graph (use->def): 4448 // LoadB -> compress_string -> CharMem 4449 // ... -> StoreB -> ByteMem 4450 // The intrinsic hides the dependency between LoadB and StoreB, causing 4451 // the load to read from memory not containing the result of the StoreB. 4452 // The correct memory graph should look like this: 4453 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4454 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4455 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4456 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4457 set_memory(res_mem, TypeAryPtr::BYTES); 4458 return str; 4459 } 4460 4461 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4462 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4463 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4464 // Capture src and dst memory (see comment in 'compress_string'). 4465 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4466 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4467 set_memory(_gvn.transform(str), dst_type); 4468 } 4469 4470 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4471 /** 4472 * int i_char = start; 4473 * for (int i_byte = 0; i_byte < count; i_byte++) { 4474 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4475 * } 4476 */ 4477 add_predicate(); 4478 RegionNode* head = new RegionNode(3); 4479 head->init_req(1, control()); 4480 gvn().set_type(head, Type::CONTROL); 4481 record_for_igvn(head); 4482 4483 Node* i_byte = new PhiNode(head, TypeInt::INT); 4484 i_byte->init_req(1, intcon(0)); 4485 gvn().set_type(i_byte, TypeInt::INT); 4486 record_for_igvn(i_byte); 4487 4488 Node* i_char = new PhiNode(head, TypeInt::INT); 4489 i_char->init_req(1, start); 4490 gvn().set_type(i_char, TypeInt::INT); 4491 record_for_igvn(i_char); 4492 4493 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4494 gvn().set_type(mem, Type::MEMORY); 4495 record_for_igvn(mem); 4496 set_control(head); 4497 set_memory(mem, TypeAryPtr::BYTES); 4498 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4499 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4500 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4501 false, false, true /* mismatched */); 4502 4503 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4504 head->init_req(2, IfTrue(iff)); 4505 mem->init_req(2, st); 4506 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4507 i_char->init_req(2, AddI(i_char, intcon(2))); 4508 4509 set_control(IfFalse(iff)); 4510 set_memory(st, TypeAryPtr::BYTES); 4511 } 4512 4513 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4514 if (!field->is_constant()) { 4515 return NULL; // Field not marked as constant. 4516 } 4517 ciInstance* holder = NULL; 4518 if (!field->is_static()) { 4519 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4520 if (const_oop != NULL && const_oop->is_instance()) { 4521 holder = const_oop->as_instance(); 4522 } 4523 } 4524 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4525 /*is_unsigned_load=*/false); 4526 if (con_type != NULL) { 4527 return makecon(con_type); 4528 } 4529 return NULL; 4530 } 4531 4532 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4533 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4534 // assumption of CCP analysis. 4535 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4536 }