1 /* 2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "memory/barrierSet.hpp" 31 #include "memory/cardTableModRefBS.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/intrinsicnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 47 //----------------------------GraphKit----------------------------------------- 48 // Main utility constructor. 49 GraphKit::GraphKit(JVMState* jvms) 50 : Phase(Phase::Parser), 51 _env(C->env()), 52 _gvn(*C->initial_gvn()) 53 { 54 _exceptions = jvms->map()->next_exception(); 55 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 56 set_jvms(jvms); 57 } 58 59 // Private constructor for parser. 60 GraphKit::GraphKit() 61 : Phase(Phase::Parser), 62 _env(C->env()), 63 _gvn(*C->initial_gvn()) 64 { 65 _exceptions = NULL; 66 set_map(NULL); 67 debug_only(_sp = -99); 68 debug_only(set_bci(-99)); 69 } 70 71 72 73 //---------------------------clean_stack--------------------------------------- 74 // Clear away rubbish from the stack area of the JVM state. 75 // This destroys any arguments that may be waiting on the stack. 76 void GraphKit::clean_stack(int from_sp) { 77 SafePointNode* map = this->map(); 78 JVMState* jvms = this->jvms(); 79 int stk_size = jvms->stk_size(); 80 int stkoff = jvms->stkoff(); 81 Node* top = this->top(); 82 for (int i = from_sp; i < stk_size; i++) { 83 if (map->in(stkoff + i) != top) { 84 map->set_req(stkoff + i, top); 85 } 86 } 87 } 88 89 90 //--------------------------------sync_jvms----------------------------------- 91 // Make sure our current jvms agrees with our parse state. 92 JVMState* GraphKit::sync_jvms() const { 93 JVMState* jvms = this->jvms(); 94 jvms->set_bci(bci()); // Record the new bci in the JVMState 95 jvms->set_sp(sp()); // Record the new sp in the JVMState 96 assert(jvms_in_sync(), "jvms is now in sync"); 97 return jvms; 98 } 99 100 //--------------------------------sync_jvms_for_reexecute--------------------- 101 // Make sure our current jvms agrees with our parse state. This version 102 // uses the reexecute_sp for reexecuting bytecodes. 103 JVMState* GraphKit::sync_jvms_for_reexecute() { 104 JVMState* jvms = this->jvms(); 105 jvms->set_bci(bci()); // Record the new bci in the JVMState 106 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 107 return jvms; 108 } 109 110 #ifdef ASSERT 111 bool GraphKit::jvms_in_sync() const { 112 Parse* parse = is_Parse(); 113 if (parse == NULL) { 114 if (bci() != jvms()->bci()) return false; 115 if (sp() != (int)jvms()->sp()) return false; 116 return true; 117 } 118 if (jvms()->method() != parse->method()) return false; 119 if (jvms()->bci() != parse->bci()) return false; 120 int jvms_sp = jvms()->sp(); 121 if (jvms_sp != parse->sp()) return false; 122 int jvms_depth = jvms()->depth(); 123 if (jvms_depth != parse->depth()) return false; 124 return true; 125 } 126 127 // Local helper checks for special internal merge points 128 // used to accumulate and merge exception states. 129 // They are marked by the region's in(0) edge being the map itself. 130 // Such merge points must never "escape" into the parser at large, 131 // until they have been handed to gvn.transform. 132 static bool is_hidden_merge(Node* reg) { 133 if (reg == NULL) return false; 134 if (reg->is_Phi()) { 135 reg = reg->in(0); 136 if (reg == NULL) return false; 137 } 138 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 139 } 140 141 void GraphKit::verify_map() const { 142 if (map() == NULL) return; // null map is OK 143 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 144 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 145 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 146 } 147 148 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 149 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 150 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 151 } 152 #endif 153 154 //---------------------------stop_and_kill_map--------------------------------- 155 // Set _map to NULL, signalling a stop to further bytecode execution. 156 // First smash the current map's control to a constant, to mark it dead. 157 void GraphKit::stop_and_kill_map() { 158 SafePointNode* dead_map = stop(); 159 if (dead_map != NULL) { 160 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 161 assert(dead_map->is_killed(), "must be so marked"); 162 } 163 } 164 165 166 //--------------------------------stopped-------------------------------------- 167 // Tell if _map is NULL, or control is top. 168 bool GraphKit::stopped() { 169 if (map() == NULL) return true; 170 else if (control() == top()) return true; 171 else return false; 172 } 173 174 175 //-----------------------------has_ex_handler---------------------------------- 176 // Tell if this method or any caller method has exception handlers. 177 bool GraphKit::has_ex_handler() { 178 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 179 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 180 return true; 181 } 182 } 183 return false; 184 } 185 186 //------------------------------save_ex_oop------------------------------------ 187 // Save an exception without blowing stack contents or other JVM state. 188 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 189 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 190 ex_map->add_req(ex_oop); 191 debug_only(verify_exception_state(ex_map)); 192 } 193 194 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 195 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 196 Node* ex_oop = ex_map->in(ex_map->req()-1); 197 if (clear_it) ex_map->del_req(ex_map->req()-1); 198 return ex_oop; 199 } 200 201 //-----------------------------saved_ex_oop------------------------------------ 202 // Recover a saved exception from its map. 203 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 204 return common_saved_ex_oop(ex_map, false); 205 } 206 207 //--------------------------clear_saved_ex_oop--------------------------------- 208 // Erase a previously saved exception from its map. 209 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, true); 211 } 212 213 #ifdef ASSERT 214 //---------------------------has_saved_ex_oop---------------------------------- 215 // Erase a previously saved exception from its map. 216 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 217 return ex_map->req() == ex_map->jvms()->endoff()+1; 218 } 219 #endif 220 221 //-------------------------make_exception_state-------------------------------- 222 // Turn the current JVM state into an exception state, appending the ex_oop. 223 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 224 sync_jvms(); 225 SafePointNode* ex_map = stop(); // do not manipulate this map any more 226 set_saved_ex_oop(ex_map, ex_oop); 227 return ex_map; 228 } 229 230 231 //--------------------------add_exception_state-------------------------------- 232 // Add an exception to my list of exceptions. 233 void GraphKit::add_exception_state(SafePointNode* ex_map) { 234 if (ex_map == NULL || ex_map->control() == top()) { 235 return; 236 } 237 #ifdef ASSERT 238 verify_exception_state(ex_map); 239 if (has_exceptions()) { 240 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 241 } 242 #endif 243 244 // If there is already an exception of exactly this type, merge with it. 245 // In particular, null-checks and other low-level exceptions common up here. 246 Node* ex_oop = saved_ex_oop(ex_map); 247 const Type* ex_type = _gvn.type(ex_oop); 248 if (ex_oop == top()) { 249 // No action needed. 250 return; 251 } 252 assert(ex_type->isa_instptr(), "exception must be an instance"); 253 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 254 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 255 // We check sp also because call bytecodes can generate exceptions 256 // both before and after arguments are popped! 257 if (ex_type2 == ex_type 258 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 259 combine_exception_states(ex_map, e2); 260 return; 261 } 262 } 263 264 // No pre-existing exception of the same type. Chain it on the list. 265 push_exception_state(ex_map); 266 } 267 268 //-----------------------add_exception_states_from----------------------------- 269 void GraphKit::add_exception_states_from(JVMState* jvms) { 270 SafePointNode* ex_map = jvms->map()->next_exception(); 271 if (ex_map != NULL) { 272 jvms->map()->set_next_exception(NULL); 273 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 274 next_map = ex_map->next_exception(); 275 ex_map->set_next_exception(NULL); 276 add_exception_state(ex_map); 277 } 278 } 279 } 280 281 //-----------------------transfer_exceptions_into_jvms------------------------- 282 JVMState* GraphKit::transfer_exceptions_into_jvms() { 283 if (map() == NULL) { 284 // We need a JVMS to carry the exceptions, but the map has gone away. 285 // Create a scratch JVMS, cloned from any of the exception states... 286 if (has_exceptions()) { 287 _map = _exceptions; 288 _map = clone_map(); 289 _map->set_next_exception(NULL); 290 clear_saved_ex_oop(_map); 291 debug_only(verify_map()); 292 } else { 293 // ...or created from scratch 294 JVMState* jvms = new (C) JVMState(_method, NULL); 295 jvms->set_bci(_bci); 296 jvms->set_sp(_sp); 297 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 298 set_jvms(jvms); 299 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 300 set_all_memory(top()); 301 while (map()->req() < jvms->endoff()) map()->add_req(top()); 302 } 303 // (This is a kludge, in case you didn't notice.) 304 set_control(top()); 305 } 306 JVMState* jvms = sync_jvms(); 307 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 308 jvms->map()->set_next_exception(_exceptions); 309 _exceptions = NULL; // done with this set of exceptions 310 return jvms; 311 } 312 313 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 314 assert(is_hidden_merge(dstphi), "must be a special merge node"); 315 assert(is_hidden_merge(srcphi), "must be a special merge node"); 316 uint limit = srcphi->req(); 317 for (uint i = PhiNode::Input; i < limit; i++) { 318 dstphi->add_req(srcphi->in(i)); 319 } 320 } 321 static inline void add_one_req(Node* dstphi, Node* src) { 322 assert(is_hidden_merge(dstphi), "must be a special merge node"); 323 assert(!is_hidden_merge(src), "must not be a special merge node"); 324 dstphi->add_req(src); 325 } 326 327 //-----------------------combine_exception_states------------------------------ 328 // This helper function combines exception states by building phis on a 329 // specially marked state-merging region. These regions and phis are 330 // untransformed, and can build up gradually. The region is marked by 331 // having a control input of its exception map, rather than NULL. Such 332 // regions do not appear except in this function, and in use_exception_state. 333 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 334 if (failing()) return; // dying anyway... 335 JVMState* ex_jvms = ex_map->_jvms; 336 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 337 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 338 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 339 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 340 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 341 assert(ex_map->req() == phi_map->req(), "matching maps"); 342 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 343 Node* hidden_merge_mark = root(); 344 Node* region = phi_map->control(); 345 MergeMemNode* phi_mem = phi_map->merged_memory(); 346 MergeMemNode* ex_mem = ex_map->merged_memory(); 347 if (region->in(0) != hidden_merge_mark) { 348 // The control input is not (yet) a specially-marked region in phi_map. 349 // Make it so, and build some phis. 350 region = new RegionNode(2); 351 _gvn.set_type(region, Type::CONTROL); 352 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 353 region->init_req(1, phi_map->control()); 354 phi_map->set_control(region); 355 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 356 record_for_igvn(io_phi); 357 _gvn.set_type(io_phi, Type::ABIO); 358 phi_map->set_i_o(io_phi); 359 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 360 Node* m = mms.memory(); 361 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 362 record_for_igvn(m_phi); 363 _gvn.set_type(m_phi, Type::MEMORY); 364 mms.set_memory(m_phi); 365 } 366 } 367 368 // Either or both of phi_map and ex_map might already be converted into phis. 369 Node* ex_control = ex_map->control(); 370 // if there is special marking on ex_map also, we add multiple edges from src 371 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 372 // how wide was the destination phi_map, originally? 373 uint orig_width = region->req(); 374 375 if (add_multiple) { 376 add_n_reqs(region, ex_control); 377 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 378 } else { 379 // ex_map has no merges, so we just add single edges everywhere 380 add_one_req(region, ex_control); 381 add_one_req(phi_map->i_o(), ex_map->i_o()); 382 } 383 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 384 if (mms.is_empty()) { 385 // get a copy of the base memory, and patch some inputs into it 386 const TypePtr* adr_type = mms.adr_type(C); 387 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 388 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 389 mms.set_memory(phi); 390 // Prepare to append interesting stuff onto the newly sliced phi: 391 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 392 } 393 // Append stuff from ex_map: 394 if (add_multiple) { 395 add_n_reqs(mms.memory(), mms.memory2()); 396 } else { 397 add_one_req(mms.memory(), mms.memory2()); 398 } 399 } 400 uint limit = ex_map->req(); 401 for (uint i = TypeFunc::Parms; i < limit; i++) { 402 // Skip everything in the JVMS after tos. (The ex_oop follows.) 403 if (i == tos) i = ex_jvms->monoff(); 404 Node* src = ex_map->in(i); 405 Node* dst = phi_map->in(i); 406 if (src != dst) { 407 PhiNode* phi; 408 if (dst->in(0) != region) { 409 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 410 record_for_igvn(phi); 411 _gvn.set_type(phi, phi->type()); 412 phi_map->set_req(i, dst); 413 // Prepare to append interesting stuff onto the new phi: 414 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 415 } else { 416 assert(dst->is_Phi(), "nobody else uses a hidden region"); 417 phi = dst->as_Phi(); 418 } 419 if (add_multiple && src->in(0) == ex_control) { 420 // Both are phis. 421 add_n_reqs(dst, src); 422 } else { 423 while (dst->req() < region->req()) add_one_req(dst, src); 424 } 425 const Type* srctype = _gvn.type(src); 426 if (phi->type() != srctype) { 427 const Type* dsttype = phi->type()->meet_speculative(srctype); 428 if (phi->type() != dsttype) { 429 phi->set_type(dsttype); 430 _gvn.set_type(phi, dsttype); 431 } 432 } 433 } 434 } 435 phi_map->merge_replaced_nodes_with(ex_map); 436 } 437 438 //--------------------------use_exception_state-------------------------------- 439 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 440 if (failing()) { stop(); return top(); } 441 Node* region = phi_map->control(); 442 Node* hidden_merge_mark = root(); 443 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 444 Node* ex_oop = clear_saved_ex_oop(phi_map); 445 if (region->in(0) == hidden_merge_mark) { 446 // Special marking for internal ex-states. Process the phis now. 447 region->set_req(0, region); // now it's an ordinary region 448 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 449 // Note: Setting the jvms also sets the bci and sp. 450 set_control(_gvn.transform(region)); 451 uint tos = jvms()->stkoff() + sp(); 452 for (uint i = 1; i < tos; i++) { 453 Node* x = phi_map->in(i); 454 if (x->in(0) == region) { 455 assert(x->is_Phi(), "expected a special phi"); 456 phi_map->set_req(i, _gvn.transform(x)); 457 } 458 } 459 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 460 Node* x = mms.memory(); 461 if (x->in(0) == region) { 462 assert(x->is_Phi(), "nobody else uses a hidden region"); 463 mms.set_memory(_gvn.transform(x)); 464 } 465 } 466 if (ex_oop->in(0) == region) { 467 assert(ex_oop->is_Phi(), "expected a special phi"); 468 ex_oop = _gvn.transform(ex_oop); 469 } 470 } else { 471 set_jvms(phi_map->jvms()); 472 } 473 474 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 475 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 476 return ex_oop; 477 } 478 479 //---------------------------------java_bc------------------------------------- 480 Bytecodes::Code GraphKit::java_bc() const { 481 ciMethod* method = this->method(); 482 int bci = this->bci(); 483 if (method != NULL && bci != InvocationEntryBci) 484 return method->java_code_at_bci(bci); 485 else 486 return Bytecodes::_illegal; 487 } 488 489 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 490 bool must_throw) { 491 // if the exception capability is set, then we will generate code 492 // to check the JavaThread.should_post_on_exceptions flag to see 493 // if we actually need to report exception events (for this 494 // thread). If we don't need to report exception events, we will 495 // take the normal fast path provided by add_exception_events. If 496 // exception event reporting is enabled for this thread, we will 497 // take the uncommon_trap in the BuildCutout below. 498 499 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 500 Node* jthread = _gvn.transform(new ThreadLocalNode()); 501 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 502 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 503 504 // Test the should_post_on_exceptions_flag vs. 0 505 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 506 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 507 508 // Branch to slow_path if should_post_on_exceptions_flag was true 509 { BuildCutout unless(this, tst, PROB_MAX); 510 // Do not try anything fancy if we're notifying the VM on every throw. 511 // Cf. case Bytecodes::_athrow in parse2.cpp. 512 uncommon_trap(reason, Deoptimization::Action_none, 513 (ciKlass*)NULL, (char*)NULL, must_throw); 514 } 515 516 } 517 518 //------------------------------builtin_throw---------------------------------- 519 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 520 bool must_throw = true; 521 522 if (env()->jvmti_can_post_on_exceptions()) { 523 // check if we must post exception events, take uncommon trap if so 524 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 525 // here if should_post_on_exceptions is false 526 // continue on with the normal codegen 527 } 528 529 // If this particular condition has not yet happened at this 530 // bytecode, then use the uncommon trap mechanism, and allow for 531 // a future recompilation if several traps occur here. 532 // If the throw is hot, try to use a more complicated inline mechanism 533 // which keeps execution inside the compiled code. 534 bool treat_throw_as_hot = false; 535 ciMethodData* md = method()->method_data(); 536 537 if (ProfileTraps) { 538 if (too_many_traps(reason)) { 539 treat_throw_as_hot = true; 540 } 541 // (If there is no MDO at all, assume it is early in 542 // execution, and that any deopts are part of the 543 // startup transient, and don't need to be remembered.) 544 545 // Also, if there is a local exception handler, treat all throws 546 // as hot if there has been at least one in this method. 547 if (C->trap_count(reason) != 0 548 && method()->method_data()->trap_count(reason) != 0 549 && has_ex_handler()) { 550 treat_throw_as_hot = true; 551 } 552 } 553 554 // If this throw happens frequently, an uncommon trap might cause 555 // a performance pothole. If there is a local exception handler, 556 // and if this particular bytecode appears to be deoptimizing often, 557 // let us handle the throw inline, with a preconstructed instance. 558 // Note: If the deopt count has blown up, the uncommon trap 559 // runtime is going to flush this nmethod, not matter what. 560 if (treat_throw_as_hot 561 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 562 // If the throw is local, we use a pre-existing instance and 563 // punt on the backtrace. This would lead to a missing backtrace 564 // (a repeat of 4292742) if the backtrace object is ever asked 565 // for its backtrace. 566 // Fixing this remaining case of 4292742 requires some flavor of 567 // escape analysis. Leave that for the future. 568 ciInstance* ex_obj = NULL; 569 switch (reason) { 570 case Deoptimization::Reason_null_check: 571 ex_obj = env()->NullPointerException_instance(); 572 break; 573 case Deoptimization::Reason_div0_check: 574 ex_obj = env()->ArithmeticException_instance(); 575 break; 576 case Deoptimization::Reason_range_check: 577 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 578 break; 579 case Deoptimization::Reason_class_check: 580 if (java_bc() == Bytecodes::_aastore) { 581 ex_obj = env()->ArrayStoreException_instance(); 582 } else { 583 ex_obj = env()->ClassCastException_instance(); 584 } 585 break; 586 } 587 if (failing()) { stop(); return; } // exception allocation might fail 588 if (ex_obj != NULL) { 589 // Cheat with a preallocated exception object. 590 if (C->log() != NULL) 591 C->log()->elem("hot_throw preallocated='1' reason='%s'", 592 Deoptimization::trap_reason_name(reason)); 593 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 594 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 595 596 // Clear the detail message of the preallocated exception object. 597 // Weblogic sometimes mutates the detail message of exceptions 598 // using reflection. 599 int offset = java_lang_Throwable::get_detailMessage_offset(); 600 const TypePtr* adr_typ = ex_con->add_offset(offset); 601 602 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 603 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 604 // Conservatively release stores of object references. 605 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 606 607 add_exception_state(make_exception_state(ex_node)); 608 return; 609 } 610 } 611 612 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 613 // It won't be much cheaper than bailing to the interp., since we'll 614 // have to pass up all the debug-info, and the runtime will have to 615 // create the stack trace. 616 617 // Usual case: Bail to interpreter. 618 // Reserve the right to recompile if we haven't seen anything yet. 619 620 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 621 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 622 if (treat_throw_as_hot 623 && (method()->method_data()->trap_recompiled_at(bci(), m) 624 || C->too_many_traps(reason))) { 625 // We cannot afford to take more traps here. Suffer in the interpreter. 626 if (C->log() != NULL) 627 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 628 Deoptimization::trap_reason_name(reason), 629 C->trap_count(reason)); 630 action = Deoptimization::Action_none; 631 } 632 633 // "must_throw" prunes the JVM state to include only the stack, if there 634 // are no local exception handlers. This should cut down on register 635 // allocation time and code size, by drastically reducing the number 636 // of in-edges on the call to the uncommon trap. 637 638 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 639 } 640 641 642 //----------------------------PreserveJVMState--------------------------------- 643 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 644 debug_only(kit->verify_map()); 645 _kit = kit; 646 _map = kit->map(); // preserve the map 647 _sp = kit->sp(); 648 kit->set_map(clone_map ? kit->clone_map() : NULL); 649 #ifdef ASSERT 650 _bci = kit->bci(); 651 Parse* parser = kit->is_Parse(); 652 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 653 _block = block; 654 #endif 655 } 656 PreserveJVMState::~PreserveJVMState() { 657 GraphKit* kit = _kit; 658 #ifdef ASSERT 659 assert(kit->bci() == _bci, "bci must not shift"); 660 Parse* parser = kit->is_Parse(); 661 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 662 assert(block == _block, "block must not shift"); 663 #endif 664 kit->set_map(_map); 665 kit->set_sp(_sp); 666 } 667 668 669 //-----------------------------BuildCutout------------------------------------- 670 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 671 : PreserveJVMState(kit) 672 { 673 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 674 SafePointNode* outer_map = _map; // preserved map is caller's 675 SafePointNode* inner_map = kit->map(); 676 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 677 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 678 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 679 } 680 BuildCutout::~BuildCutout() { 681 GraphKit* kit = _kit; 682 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 683 } 684 685 //---------------------------PreserveReexecuteState---------------------------- 686 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 687 assert(!kit->stopped(), "must call stopped() before"); 688 _kit = kit; 689 _sp = kit->sp(); 690 _reexecute = kit->jvms()->_reexecute; 691 } 692 PreserveReexecuteState::~PreserveReexecuteState() { 693 if (_kit->stopped()) return; 694 _kit->jvms()->_reexecute = _reexecute; 695 _kit->set_sp(_sp); 696 } 697 698 //------------------------------clone_map-------------------------------------- 699 // Implementation of PreserveJVMState 700 // 701 // Only clone_map(...) here. If this function is only used in the 702 // PreserveJVMState class we may want to get rid of this extra 703 // function eventually and do it all there. 704 705 SafePointNode* GraphKit::clone_map() { 706 if (map() == NULL) return NULL; 707 708 // Clone the memory edge first 709 Node* mem = MergeMemNode::make(map()->memory()); 710 gvn().set_type_bottom(mem); 711 712 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 713 JVMState* jvms = this->jvms(); 714 JVMState* clonejvms = jvms->clone_shallow(C); 715 clonemap->set_memory(mem); 716 clonemap->set_jvms(clonejvms); 717 clonejvms->set_map(clonemap); 718 record_for_igvn(clonemap); 719 gvn().set_type_bottom(clonemap); 720 return clonemap; 721 } 722 723 724 //-----------------------------set_map_clone----------------------------------- 725 void GraphKit::set_map_clone(SafePointNode* m) { 726 _map = m; 727 _map = clone_map(); 728 _map->set_next_exception(NULL); 729 debug_only(verify_map()); 730 } 731 732 733 //----------------------------kill_dead_locals--------------------------------- 734 // Detect any locals which are known to be dead, and force them to top. 735 void GraphKit::kill_dead_locals() { 736 // Consult the liveness information for the locals. If any 737 // of them are unused, then they can be replaced by top(). This 738 // should help register allocation time and cut down on the size 739 // of the deoptimization information. 740 741 // This call is made from many of the bytecode handling 742 // subroutines called from the Big Switch in do_one_bytecode. 743 // Every bytecode which might include a slow path is responsible 744 // for killing its dead locals. The more consistent we 745 // are about killing deads, the fewer useless phis will be 746 // constructed for them at various merge points. 747 748 // bci can be -1 (InvocationEntryBci). We return the entry 749 // liveness for the method. 750 751 if (method() == NULL || method()->code_size() == 0) { 752 // We are building a graph for a call to a native method. 753 // All locals are live. 754 return; 755 } 756 757 ResourceMark rm; 758 759 // Consult the liveness information for the locals. If any 760 // of them are unused, then they can be replaced by top(). This 761 // should help register allocation time and cut down on the size 762 // of the deoptimization information. 763 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 764 765 int len = (int)live_locals.size(); 766 assert(len <= jvms()->loc_size(), "too many live locals"); 767 for (int local = 0; local < len; local++) { 768 if (!live_locals.at(local)) { 769 set_local(local, top()); 770 } 771 } 772 } 773 774 #ifdef ASSERT 775 //-------------------------dead_locals_are_killed------------------------------ 776 // Return true if all dead locals are set to top in the map. 777 // Used to assert "clean" debug info at various points. 778 bool GraphKit::dead_locals_are_killed() { 779 if (method() == NULL || method()->code_size() == 0) { 780 // No locals need to be dead, so all is as it should be. 781 return true; 782 } 783 784 // Make sure somebody called kill_dead_locals upstream. 785 ResourceMark rm; 786 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 787 if (jvms->loc_size() == 0) continue; // no locals to consult 788 SafePointNode* map = jvms->map(); 789 ciMethod* method = jvms->method(); 790 int bci = jvms->bci(); 791 if (jvms == this->jvms()) { 792 bci = this->bci(); // it might not yet be synched 793 } 794 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 795 int len = (int)live_locals.size(); 796 if (!live_locals.is_valid() || len == 0) 797 // This method is trivial, or is poisoned by a breakpoint. 798 return true; 799 assert(len == jvms->loc_size(), "live map consistent with locals map"); 800 for (int local = 0; local < len; local++) { 801 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 802 if (PrintMiscellaneous && (Verbose || WizardMode)) { 803 tty->print_cr("Zombie local %d: ", local); 804 jvms->dump(); 805 } 806 return false; 807 } 808 } 809 } 810 return true; 811 } 812 813 #endif //ASSERT 814 815 // Helper function for enforcing certain bytecodes to reexecute if 816 // deoptimization happens 817 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 818 ciMethod* cur_method = jvms->method(); 819 int cur_bci = jvms->bci(); 820 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 821 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 822 return Interpreter::bytecode_should_reexecute(code) || 823 is_anewarray && code == Bytecodes::_multianewarray; 824 // Reexecute _multianewarray bytecode which was replaced with 825 // sequence of [a]newarray. See Parse::do_multianewarray(). 826 // 827 // Note: interpreter should not have it set since this optimization 828 // is limited by dimensions and guarded by flag so in some cases 829 // multianewarray() runtime calls will be generated and 830 // the bytecode should not be reexecutes (stack will not be reset). 831 } else 832 return false; 833 } 834 835 // Helper function for adding JVMState and debug information to node 836 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 837 // Add the safepoint edges to the call (or other safepoint). 838 839 // Make sure dead locals are set to top. This 840 // should help register allocation time and cut down on the size 841 // of the deoptimization information. 842 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 843 844 // Walk the inline list to fill in the correct set of JVMState's 845 // Also fill in the associated edges for each JVMState. 846 847 // If the bytecode needs to be reexecuted we need to put 848 // the arguments back on the stack. 849 const bool should_reexecute = jvms()->should_reexecute(); 850 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 851 852 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 853 // undefined if the bci is different. This is normal for Parse but it 854 // should not happen for LibraryCallKit because only one bci is processed. 855 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 856 "in LibraryCallKit the reexecute bit should not change"); 857 858 // If we are guaranteed to throw, we can prune everything but the 859 // input to the current bytecode. 860 bool can_prune_locals = false; 861 uint stack_slots_not_pruned = 0; 862 int inputs = 0, depth = 0; 863 if (must_throw) { 864 assert(method() == youngest_jvms->method(), "sanity"); 865 if (compute_stack_effects(inputs, depth)) { 866 can_prune_locals = true; 867 stack_slots_not_pruned = inputs; 868 } 869 } 870 871 if (env()->should_retain_local_variables()) { 872 // At any safepoint, this method can get breakpointed, which would 873 // then require an immediate deoptimization. 874 can_prune_locals = false; // do not prune locals 875 stack_slots_not_pruned = 0; 876 } 877 878 // do not scribble on the input jvms 879 JVMState* out_jvms = youngest_jvms->clone_deep(C); 880 call->set_jvms(out_jvms); // Start jvms list for call node 881 882 // For a known set of bytecodes, the interpreter should reexecute them if 883 // deoptimization happens. We set the reexecute state for them here 884 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 885 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 886 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 887 } 888 889 // Presize the call: 890 DEBUG_ONLY(uint non_debug_edges = call->req()); 891 call->add_req_batch(top(), youngest_jvms->debug_depth()); 892 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 893 894 // Set up edges so that the call looks like this: 895 // Call [state:] ctl io mem fptr retadr 896 // [parms:] parm0 ... parmN 897 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 898 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 899 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 900 // Note that caller debug info precedes callee debug info. 901 902 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 903 uint debug_ptr = call->req(); 904 905 // Loop over the map input edges associated with jvms, add them 906 // to the call node, & reset all offsets to match call node array. 907 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 908 uint debug_end = debug_ptr; 909 uint debug_start = debug_ptr - in_jvms->debug_size(); 910 debug_ptr = debug_start; // back up the ptr 911 912 uint p = debug_start; // walks forward in [debug_start, debug_end) 913 uint j, k, l; 914 SafePointNode* in_map = in_jvms->map(); 915 out_jvms->set_map(call); 916 917 if (can_prune_locals) { 918 assert(in_jvms->method() == out_jvms->method(), "sanity"); 919 // If the current throw can reach an exception handler in this JVMS, 920 // then we must keep everything live that can reach that handler. 921 // As a quick and dirty approximation, we look for any handlers at all. 922 if (in_jvms->method()->has_exception_handlers()) { 923 can_prune_locals = false; 924 } 925 } 926 927 // Add the Locals 928 k = in_jvms->locoff(); 929 l = in_jvms->loc_size(); 930 out_jvms->set_locoff(p); 931 if (!can_prune_locals) { 932 for (j = 0; j < l; j++) 933 call->set_req(p++, in_map->in(k+j)); 934 } else { 935 p += l; // already set to top above by add_req_batch 936 } 937 938 // Add the Expression Stack 939 k = in_jvms->stkoff(); 940 l = in_jvms->sp(); 941 out_jvms->set_stkoff(p); 942 if (!can_prune_locals) { 943 for (j = 0; j < l; j++) 944 call->set_req(p++, in_map->in(k+j)); 945 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 946 // Divide stack into {S0,...,S1}, where S0 is set to top. 947 uint s1 = stack_slots_not_pruned; 948 stack_slots_not_pruned = 0; // for next iteration 949 if (s1 > l) s1 = l; 950 uint s0 = l - s1; 951 p += s0; // skip the tops preinstalled by add_req_batch 952 for (j = s0; j < l; j++) 953 call->set_req(p++, in_map->in(k+j)); 954 } else { 955 p += l; // already set to top above by add_req_batch 956 } 957 958 // Add the Monitors 959 k = in_jvms->monoff(); 960 l = in_jvms->mon_size(); 961 out_jvms->set_monoff(p); 962 for (j = 0; j < l; j++) 963 call->set_req(p++, in_map->in(k+j)); 964 965 // Copy any scalar object fields. 966 k = in_jvms->scloff(); 967 l = in_jvms->scl_size(); 968 out_jvms->set_scloff(p); 969 for (j = 0; j < l; j++) 970 call->set_req(p++, in_map->in(k+j)); 971 972 // Finish the new jvms. 973 out_jvms->set_endoff(p); 974 975 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 976 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 977 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 978 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 979 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 980 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 981 982 // Update the two tail pointers in parallel. 983 out_jvms = out_jvms->caller(); 984 in_jvms = in_jvms->caller(); 985 } 986 987 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 988 989 // Test the correctness of JVMState::debug_xxx accessors: 990 assert(call->jvms()->debug_start() == non_debug_edges, ""); 991 assert(call->jvms()->debug_end() == call->req(), ""); 992 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 993 } 994 995 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 996 Bytecodes::Code code = java_bc(); 997 if (code == Bytecodes::_wide) { 998 code = method()->java_code_at_bci(bci() + 1); 999 } 1000 1001 BasicType rtype = T_ILLEGAL; 1002 int rsize = 0; 1003 1004 if (code != Bytecodes::_illegal) { 1005 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1006 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1007 if (rtype < T_CONFLICT) 1008 rsize = type2size[rtype]; 1009 } 1010 1011 switch (code) { 1012 case Bytecodes::_illegal: 1013 return false; 1014 1015 case Bytecodes::_ldc: 1016 case Bytecodes::_ldc_w: 1017 case Bytecodes::_ldc2_w: 1018 inputs = 0; 1019 break; 1020 1021 case Bytecodes::_dup: inputs = 1; break; 1022 case Bytecodes::_dup_x1: inputs = 2; break; 1023 case Bytecodes::_dup_x2: inputs = 3; break; 1024 case Bytecodes::_dup2: inputs = 2; break; 1025 case Bytecodes::_dup2_x1: inputs = 3; break; 1026 case Bytecodes::_dup2_x2: inputs = 4; break; 1027 case Bytecodes::_swap: inputs = 2; break; 1028 case Bytecodes::_arraylength: inputs = 1; break; 1029 1030 case Bytecodes::_getstatic: 1031 case Bytecodes::_putstatic: 1032 case Bytecodes::_getfield: 1033 case Bytecodes::_putfield: 1034 { 1035 bool ignored_will_link; 1036 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1037 int size = field->type()->size(); 1038 bool is_get = (depth >= 0), is_static = (depth & 1); 1039 inputs = (is_static ? 0 : 1); 1040 if (is_get) { 1041 depth = size - inputs; 1042 } else { 1043 inputs += size; // putxxx pops the value from the stack 1044 depth = - inputs; 1045 } 1046 } 1047 break; 1048 1049 case Bytecodes::_invokevirtual: 1050 case Bytecodes::_invokespecial: 1051 case Bytecodes::_invokestatic: 1052 case Bytecodes::_invokedynamic: 1053 case Bytecodes::_invokeinterface: 1054 { 1055 bool ignored_will_link; 1056 ciSignature* declared_signature = NULL; 1057 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1058 assert(declared_signature != NULL, "cannot be null"); 1059 inputs = declared_signature->arg_size_for_bc(code); 1060 int size = declared_signature->return_type()->size(); 1061 depth = size - inputs; 1062 } 1063 break; 1064 1065 case Bytecodes::_multianewarray: 1066 { 1067 ciBytecodeStream iter(method()); 1068 iter.reset_to_bci(bci()); 1069 iter.next(); 1070 inputs = iter.get_dimensions(); 1071 assert(rsize == 1, ""); 1072 depth = rsize - inputs; 1073 } 1074 break; 1075 1076 case Bytecodes::_ireturn: 1077 case Bytecodes::_lreturn: 1078 case Bytecodes::_freturn: 1079 case Bytecodes::_dreturn: 1080 case Bytecodes::_areturn: 1081 assert(rsize = -depth, ""); 1082 inputs = rsize; 1083 break; 1084 1085 case Bytecodes::_jsr: 1086 case Bytecodes::_jsr_w: 1087 inputs = 0; 1088 depth = 1; // S.B. depth=1, not zero 1089 break; 1090 1091 default: 1092 // bytecode produces a typed result 1093 inputs = rsize - depth; 1094 assert(inputs >= 0, ""); 1095 break; 1096 } 1097 1098 #ifdef ASSERT 1099 // spot check 1100 int outputs = depth + inputs; 1101 assert(outputs >= 0, "sanity"); 1102 switch (code) { 1103 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1104 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1105 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1106 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1107 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1108 } 1109 #endif //ASSERT 1110 1111 return true; 1112 } 1113 1114 1115 1116 //------------------------------basic_plus_adr--------------------------------- 1117 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1118 // short-circuit a common case 1119 if (offset == intcon(0)) return ptr; 1120 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1121 } 1122 1123 Node* GraphKit::ConvI2L(Node* offset) { 1124 // short-circuit a common case 1125 jint offset_con = find_int_con(offset, Type::OffsetBot); 1126 if (offset_con != Type::OffsetBot) { 1127 return longcon((jlong) offset_con); 1128 } 1129 return _gvn.transform( new ConvI2LNode(offset)); 1130 } 1131 1132 Node* GraphKit::ConvI2UL(Node* offset) { 1133 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1134 if (offset_con != (juint) Type::OffsetBot) { 1135 return longcon((julong) offset_con); 1136 } 1137 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1138 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1139 return _gvn.transform( new AndLNode(conv, mask) ); 1140 } 1141 1142 Node* GraphKit::ConvL2I(Node* offset) { 1143 // short-circuit a common case 1144 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1145 if (offset_con != (jlong)Type::OffsetBot) { 1146 return intcon((int) offset_con); 1147 } 1148 return _gvn.transform( new ConvL2INode(offset)); 1149 } 1150 1151 //-------------------------load_object_klass----------------------------------- 1152 Node* GraphKit::load_object_klass(Node* obj) { 1153 // Special-case a fresh allocation to avoid building nodes: 1154 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1155 if (akls != NULL) return akls; 1156 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1157 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1158 } 1159 1160 //-------------------------load_array_length----------------------------------- 1161 Node* GraphKit::load_array_length(Node* array) { 1162 // Special-case a fresh allocation to avoid building nodes: 1163 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1164 Node *alen; 1165 if (alloc == NULL) { 1166 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1167 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1168 } else { 1169 alen = alloc->Ideal_length(); 1170 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1171 if (ccast != alen) { 1172 alen = _gvn.transform(ccast); 1173 } 1174 } 1175 return alen; 1176 } 1177 1178 //------------------------------do_null_check---------------------------------- 1179 // Helper function to do a NULL pointer check. Returned value is 1180 // the incoming address with NULL casted away. You are allowed to use the 1181 // not-null value only if you are control dependent on the test. 1182 extern int explicit_null_checks_inserted, 1183 explicit_null_checks_elided; 1184 Node* GraphKit::null_check_common(Node* value, BasicType type, 1185 // optional arguments for variations: 1186 bool assert_null, 1187 Node* *null_control, 1188 bool speculative) { 1189 assert(!assert_null || null_control == NULL, "not both at once"); 1190 if (stopped()) return top(); 1191 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1192 // For some performance testing, we may wish to suppress null checking. 1193 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1194 return value; 1195 } 1196 explicit_null_checks_inserted++; 1197 1198 // Construct NULL check 1199 Node *chk = NULL; 1200 switch(type) { 1201 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1202 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1203 case T_ARRAY : // fall through 1204 type = T_OBJECT; // simplify further tests 1205 case T_OBJECT : { 1206 const Type *t = _gvn.type( value ); 1207 1208 const TypeOopPtr* tp = t->isa_oopptr(); 1209 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1210 // Only for do_null_check, not any of its siblings: 1211 && !assert_null && null_control == NULL) { 1212 // Usually, any field access or invocation on an unloaded oop type 1213 // will simply fail to link, since the statically linked class is 1214 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1215 // the static class is loaded but the sharper oop type is not. 1216 // Rather than checking for this obscure case in lots of places, 1217 // we simply observe that a null check on an unloaded class 1218 // will always be followed by a nonsense operation, so we 1219 // can just issue the uncommon trap here. 1220 // Our access to the unloaded class will only be correct 1221 // after it has been loaded and initialized, which requires 1222 // a trip through the interpreter. 1223 #ifndef PRODUCT 1224 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1225 #endif 1226 uncommon_trap(Deoptimization::Reason_unloaded, 1227 Deoptimization::Action_reinterpret, 1228 tp->klass(), "!loaded"); 1229 return top(); 1230 } 1231 1232 if (assert_null) { 1233 // See if the type is contained in NULL_PTR. 1234 // If so, then the value is already null. 1235 if (t->higher_equal(TypePtr::NULL_PTR)) { 1236 explicit_null_checks_elided++; 1237 return value; // Elided null assert quickly! 1238 } 1239 } else { 1240 // See if mixing in the NULL pointer changes type. 1241 // If so, then the NULL pointer was not allowed in the original 1242 // type. In other words, "value" was not-null. 1243 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1244 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1245 explicit_null_checks_elided++; 1246 return value; // Elided null check quickly! 1247 } 1248 } 1249 chk = new CmpPNode( value, null() ); 1250 break; 1251 } 1252 1253 default: 1254 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1255 } 1256 assert(chk != NULL, "sanity check"); 1257 chk = _gvn.transform(chk); 1258 1259 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1260 BoolNode *btst = new BoolNode( chk, btest); 1261 Node *tst = _gvn.transform( btst ); 1262 1263 //----------- 1264 // if peephole optimizations occurred, a prior test existed. 1265 // If a prior test existed, maybe it dominates as we can avoid this test. 1266 if (tst != btst && type == T_OBJECT) { 1267 // At this point we want to scan up the CFG to see if we can 1268 // find an identical test (and so avoid this test altogether). 1269 Node *cfg = control(); 1270 int depth = 0; 1271 while( depth < 16 ) { // Limit search depth for speed 1272 if( cfg->Opcode() == Op_IfTrue && 1273 cfg->in(0)->in(1) == tst ) { 1274 // Found prior test. Use "cast_not_null" to construct an identical 1275 // CastPP (and hence hash to) as already exists for the prior test. 1276 // Return that casted value. 1277 if (assert_null) { 1278 replace_in_map(value, null()); 1279 return null(); // do not issue the redundant test 1280 } 1281 Node *oldcontrol = control(); 1282 set_control(cfg); 1283 Node *res = cast_not_null(value); 1284 set_control(oldcontrol); 1285 explicit_null_checks_elided++; 1286 return res; 1287 } 1288 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1289 if (cfg == NULL) break; // Quit at region nodes 1290 depth++; 1291 } 1292 } 1293 1294 //----------- 1295 // Branch to failure if null 1296 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1297 Deoptimization::DeoptReason reason; 1298 if (assert_null) { 1299 reason = Deoptimization::Reason_null_assert; 1300 } else if (type == T_OBJECT) { 1301 reason = Deoptimization::reason_null_check(speculative); 1302 } else { 1303 reason = Deoptimization::Reason_div0_check; 1304 } 1305 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1306 // ciMethodData::has_trap_at will return a conservative -1 if any 1307 // must-be-null assertion has failed. This could cause performance 1308 // problems for a method after its first do_null_assert failure. 1309 // Consider using 'Reason_class_check' instead? 1310 1311 // To cause an implicit null check, we set the not-null probability 1312 // to the maximum (PROB_MAX). For an explicit check the probability 1313 // is set to a smaller value. 1314 if (null_control != NULL || too_many_traps(reason)) { 1315 // probability is less likely 1316 ok_prob = PROB_LIKELY_MAG(3); 1317 } else if (!assert_null && 1318 (ImplicitNullCheckThreshold > 0) && 1319 method() != NULL && 1320 (method()->method_data()->trap_count(reason) 1321 >= (uint)ImplicitNullCheckThreshold)) { 1322 ok_prob = PROB_LIKELY_MAG(3); 1323 } 1324 1325 if (null_control != NULL) { 1326 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1327 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1328 set_control( _gvn.transform( new IfTrueNode(iff))); 1329 if (null_true == top()) 1330 explicit_null_checks_elided++; 1331 (*null_control) = null_true; 1332 } else { 1333 BuildCutout unless(this, tst, ok_prob); 1334 // Check for optimizer eliding test at parse time 1335 if (stopped()) { 1336 // Failure not possible; do not bother making uncommon trap. 1337 explicit_null_checks_elided++; 1338 } else if (assert_null) { 1339 uncommon_trap(reason, 1340 Deoptimization::Action_make_not_entrant, 1341 NULL, "assert_null"); 1342 } else { 1343 replace_in_map(value, zerocon(type)); 1344 builtin_throw(reason); 1345 } 1346 } 1347 1348 // Must throw exception, fall-thru not possible? 1349 if (stopped()) { 1350 return top(); // No result 1351 } 1352 1353 if (assert_null) { 1354 // Cast obj to null on this path. 1355 replace_in_map(value, zerocon(type)); 1356 return zerocon(type); 1357 } 1358 1359 // Cast obj to not-null on this path, if there is no null_control. 1360 // (If there is a null_control, a non-null value may come back to haunt us.) 1361 if (type == T_OBJECT) { 1362 Node* cast = cast_not_null(value, false); 1363 if (null_control == NULL || (*null_control) == top()) 1364 replace_in_map(value, cast); 1365 value = cast; 1366 } 1367 1368 return value; 1369 } 1370 1371 1372 //------------------------------cast_not_null---------------------------------- 1373 // Cast obj to not-null on this path 1374 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1375 const Type *t = _gvn.type(obj); 1376 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1377 // Object is already not-null? 1378 if( t == t_not_null ) return obj; 1379 1380 Node *cast = new CastPPNode(obj,t_not_null); 1381 cast->init_req(0, control()); 1382 cast = _gvn.transform( cast ); 1383 1384 // Scan for instances of 'obj' in the current JVM mapping. 1385 // These instances are known to be not-null after the test. 1386 if (do_replace_in_map) 1387 replace_in_map(obj, cast); 1388 1389 return cast; // Return casted value 1390 } 1391 1392 1393 //--------------------------replace_in_map------------------------------------- 1394 void GraphKit::replace_in_map(Node* old, Node* neww) { 1395 if (old == neww) { 1396 return; 1397 } 1398 1399 map()->replace_edge(old, neww); 1400 1401 // Note: This operation potentially replaces any edge 1402 // on the map. This includes locals, stack, and monitors 1403 // of the current (innermost) JVM state. 1404 1405 // don't let inconsistent types from profiling escape this 1406 // method 1407 1408 const Type* told = _gvn.type(old); 1409 const Type* tnew = _gvn.type(neww); 1410 1411 if (!tnew->higher_equal(told)) { 1412 return; 1413 } 1414 1415 map()->record_replaced_node(old, neww); 1416 } 1417 1418 1419 //============================================================================= 1420 //--------------------------------memory--------------------------------------- 1421 Node* GraphKit::memory(uint alias_idx) { 1422 MergeMemNode* mem = merged_memory(); 1423 Node* p = mem->memory_at(alias_idx); 1424 _gvn.set_type(p, Type::MEMORY); // must be mapped 1425 return p; 1426 } 1427 1428 //-----------------------------reset_memory------------------------------------ 1429 Node* GraphKit::reset_memory() { 1430 Node* mem = map()->memory(); 1431 // do not use this node for any more parsing! 1432 debug_only( map()->set_memory((Node*)NULL) ); 1433 return _gvn.transform( mem ); 1434 } 1435 1436 //------------------------------set_all_memory--------------------------------- 1437 void GraphKit::set_all_memory(Node* newmem) { 1438 Node* mergemem = MergeMemNode::make(newmem); 1439 gvn().set_type_bottom(mergemem); 1440 map()->set_memory(mergemem); 1441 } 1442 1443 //------------------------------set_all_memory_call---------------------------- 1444 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1445 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1446 set_all_memory(newmem); 1447 } 1448 1449 //============================================================================= 1450 // 1451 // parser factory methods for MemNodes 1452 // 1453 // These are layered on top of the factory methods in LoadNode and StoreNode, 1454 // and integrate with the parser's memory state and _gvn engine. 1455 // 1456 1457 // factory methods in "int adr_idx" 1458 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1459 int adr_idx, 1460 MemNode::MemOrd mo, bool require_atomic_access) { 1461 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1462 const TypePtr* adr_type = NULL; // debug-mode-only argument 1463 debug_only(adr_type = C->get_adr_type(adr_idx)); 1464 Node* mem = memory(adr_idx); 1465 Node* ld; 1466 if (require_atomic_access && bt == T_LONG) { 1467 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo); 1468 } else if (require_atomic_access && bt == T_DOUBLE) { 1469 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo); 1470 } else { 1471 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); 1472 } 1473 ld = _gvn.transform(ld); 1474 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1475 // Improve graph before escape analysis and boxing elimination. 1476 record_for_igvn(ld); 1477 } 1478 return ld; 1479 } 1480 1481 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1482 int adr_idx, 1483 MemNode::MemOrd mo, 1484 bool require_atomic_access) { 1485 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1486 const TypePtr* adr_type = NULL; 1487 debug_only(adr_type = C->get_adr_type(adr_idx)); 1488 Node *mem = memory(adr_idx); 1489 Node* st; 1490 if (require_atomic_access && bt == T_LONG) { 1491 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1492 } else if (require_atomic_access && bt == T_DOUBLE) { 1493 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1494 } else { 1495 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1496 } 1497 st = _gvn.transform(st); 1498 set_memory(st, adr_idx); 1499 // Back-to-back stores can only remove intermediate store with DU info 1500 // so push on worklist for optimizer. 1501 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1502 record_for_igvn(st); 1503 1504 return st; 1505 } 1506 1507 1508 void GraphKit::pre_barrier(bool do_load, 1509 Node* ctl, 1510 Node* obj, 1511 Node* adr, 1512 uint adr_idx, 1513 Node* val, 1514 const TypeOopPtr* val_type, 1515 Node* pre_val, 1516 BasicType bt) { 1517 1518 BarrierSet* bs = Universe::heap()->barrier_set(); 1519 set_control(ctl); 1520 switch (bs->kind()) { 1521 case BarrierSet::G1SATBCT: 1522 case BarrierSet::G1SATBCTLogging: 1523 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1524 break; 1525 1526 case BarrierSet::CardTableModRef: 1527 case BarrierSet::CardTableExtension: 1528 case BarrierSet::ModRef: 1529 break; 1530 1531 case BarrierSet::Other: 1532 default : 1533 ShouldNotReachHere(); 1534 1535 } 1536 } 1537 1538 bool GraphKit::can_move_pre_barrier() const { 1539 BarrierSet* bs = Universe::heap()->barrier_set(); 1540 switch (bs->kind()) { 1541 case BarrierSet::G1SATBCT: 1542 case BarrierSet::G1SATBCTLogging: 1543 return true; // Can move it if no safepoint 1544 1545 case BarrierSet::CardTableModRef: 1546 case BarrierSet::CardTableExtension: 1547 case BarrierSet::ModRef: 1548 return true; // There is no pre-barrier 1549 1550 case BarrierSet::Other: 1551 default : 1552 ShouldNotReachHere(); 1553 } 1554 return false; 1555 } 1556 1557 void GraphKit::post_barrier(Node* ctl, 1558 Node* store, 1559 Node* obj, 1560 Node* adr, 1561 uint adr_idx, 1562 Node* val, 1563 BasicType bt, 1564 bool use_precise) { 1565 BarrierSet* bs = Universe::heap()->barrier_set(); 1566 set_control(ctl); 1567 switch (bs->kind()) { 1568 case BarrierSet::G1SATBCT: 1569 case BarrierSet::G1SATBCTLogging: 1570 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1571 break; 1572 1573 case BarrierSet::CardTableModRef: 1574 case BarrierSet::CardTableExtension: 1575 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1576 break; 1577 1578 case BarrierSet::ModRef: 1579 break; 1580 1581 case BarrierSet::Other: 1582 default : 1583 ShouldNotReachHere(); 1584 1585 } 1586 } 1587 1588 Node* GraphKit::store_oop(Node* ctl, 1589 Node* obj, 1590 Node* adr, 1591 const TypePtr* adr_type, 1592 Node* val, 1593 const TypeOopPtr* val_type, 1594 BasicType bt, 1595 bool use_precise, 1596 MemNode::MemOrd mo) { 1597 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1598 // could be delayed during Parse (for example, in adjust_map_after_if()). 1599 // Execute transformation here to avoid barrier generation in such case. 1600 if (_gvn.type(val) == TypePtr::NULL_PTR) 1601 val = _gvn.makecon(TypePtr::NULL_PTR); 1602 1603 set_control(ctl); 1604 if (stopped()) return top(); // Dead path ? 1605 1606 assert(bt == T_OBJECT, "sanity"); 1607 assert(val != NULL, "not dead path"); 1608 uint adr_idx = C->get_alias_index(adr_type); 1609 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1610 1611 pre_barrier(true /* do_load */, 1612 control(), obj, adr, adr_idx, val, val_type, 1613 NULL /* pre_val */, 1614 bt); 1615 1616 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1617 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1618 return store; 1619 } 1620 1621 // Could be an array or object we don't know at compile time (unsafe ref.) 1622 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1623 Node* obj, // containing obj 1624 Node* adr, // actual adress to store val at 1625 const TypePtr* adr_type, 1626 Node* val, 1627 BasicType bt, 1628 MemNode::MemOrd mo) { 1629 Compile::AliasType* at = C->alias_type(adr_type); 1630 const TypeOopPtr* val_type = NULL; 1631 if (adr_type->isa_instptr()) { 1632 if (at->field() != NULL) { 1633 // known field. This code is a copy of the do_put_xxx logic. 1634 ciField* field = at->field(); 1635 if (!field->type()->is_loaded()) { 1636 val_type = TypeInstPtr::BOTTOM; 1637 } else { 1638 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1639 } 1640 } 1641 } else if (adr_type->isa_aryptr()) { 1642 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1643 } 1644 if (val_type == NULL) { 1645 val_type = TypeInstPtr::BOTTOM; 1646 } 1647 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1648 } 1649 1650 1651 //-------------------------array_element_address------------------------- 1652 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1653 const TypeInt* sizetype) { 1654 uint shift = exact_log2(type2aelembytes(elembt)); 1655 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1656 1657 // short-circuit a common case (saves lots of confusing waste motion) 1658 jint idx_con = find_int_con(idx, -1); 1659 if (idx_con >= 0) { 1660 intptr_t offset = header + ((intptr_t)idx_con << shift); 1661 return basic_plus_adr(ary, offset); 1662 } 1663 1664 // must be correct type for alignment purposes 1665 Node* base = basic_plus_adr(ary, header); 1666 #ifdef _LP64 1667 // The scaled index operand to AddP must be a clean 64-bit value. 1668 // Java allows a 32-bit int to be incremented to a negative 1669 // value, which appears in a 64-bit register as a large 1670 // positive number. Using that large positive number as an 1671 // operand in pointer arithmetic has bad consequences. 1672 // On the other hand, 32-bit overflow is rare, and the possibility 1673 // can often be excluded, if we annotate the ConvI2L node with 1674 // a type assertion that its value is known to be a small positive 1675 // number. (The prior range check has ensured this.) 1676 // This assertion is used by ConvI2LNode::Ideal. 1677 int index_max = max_jint - 1; // array size is max_jint, index is one less 1678 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1679 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1680 idx = _gvn.transform( new ConvI2LNode(idx, lidxtype) ); 1681 #endif 1682 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1683 return basic_plus_adr(ary, base, scale); 1684 } 1685 1686 //-------------------------load_array_element------------------------- 1687 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1688 const Type* elemtype = arytype->elem(); 1689 BasicType elembt = elemtype->array_element_basic_type(); 1690 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1691 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1692 return ld; 1693 } 1694 1695 //-------------------------set_arguments_for_java_call------------------------- 1696 // Arguments (pre-popped from the stack) are taken from the JVMS. 1697 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1698 // Add the call arguments: 1699 uint nargs = call->method()->arg_size(); 1700 for (uint i = 0; i < nargs; i++) { 1701 Node* arg = argument(i); 1702 call->init_req(i + TypeFunc::Parms, arg); 1703 } 1704 } 1705 1706 //---------------------------set_edges_for_java_call--------------------------- 1707 // Connect a newly created call into the current JVMS. 1708 // A return value node (if any) is returned from set_edges_for_java_call. 1709 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1710 1711 // Add the predefined inputs: 1712 call->init_req( TypeFunc::Control, control() ); 1713 call->init_req( TypeFunc::I_O , i_o() ); 1714 call->init_req( TypeFunc::Memory , reset_memory() ); 1715 call->init_req( TypeFunc::FramePtr, frameptr() ); 1716 call->init_req( TypeFunc::ReturnAdr, top() ); 1717 1718 add_safepoint_edges(call, must_throw); 1719 1720 Node* xcall = _gvn.transform(call); 1721 1722 if (xcall == top()) { 1723 set_control(top()); 1724 return; 1725 } 1726 assert(xcall == call, "call identity is stable"); 1727 1728 // Re-use the current map to produce the result. 1729 1730 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1731 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1732 set_all_memory_call(xcall, separate_io_proj); 1733 1734 //return xcall; // no need, caller already has it 1735 } 1736 1737 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1738 if (stopped()) return top(); // maybe the call folded up? 1739 1740 // Capture the return value, if any. 1741 Node* ret; 1742 if (call->method() == NULL || 1743 call->method()->return_type()->basic_type() == T_VOID) 1744 ret = top(); 1745 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1746 1747 // Note: Since any out-of-line call can produce an exception, 1748 // we always insert an I_O projection from the call into the result. 1749 1750 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1751 1752 if (separate_io_proj) { 1753 // The caller requested separate projections be used by the fall 1754 // through and exceptional paths, so replace the projections for 1755 // the fall through path. 1756 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1757 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1758 } 1759 return ret; 1760 } 1761 1762 //--------------------set_predefined_input_for_runtime_call-------------------- 1763 // Reading and setting the memory state is way conservative here. 1764 // The real problem is that I am not doing real Type analysis on memory, 1765 // so I cannot distinguish card mark stores from other stores. Across a GC 1766 // point the Store Barrier and the card mark memory has to agree. I cannot 1767 // have a card mark store and its barrier split across the GC point from 1768 // either above or below. Here I get that to happen by reading ALL of memory. 1769 // A better answer would be to separate out card marks from other memory. 1770 // For now, return the input memory state, so that it can be reused 1771 // after the call, if this call has restricted memory effects. 1772 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1773 // Set fixed predefined input arguments 1774 Node* memory = reset_memory(); 1775 call->init_req( TypeFunc::Control, control() ); 1776 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1777 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1778 call->init_req( TypeFunc::FramePtr, frameptr() ); 1779 call->init_req( TypeFunc::ReturnAdr, top() ); 1780 return memory; 1781 } 1782 1783 //-------------------set_predefined_output_for_runtime_call-------------------- 1784 // Set control and memory (not i_o) from the call. 1785 // If keep_mem is not NULL, use it for the output state, 1786 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1787 // If hook_mem is NULL, this call produces no memory effects at all. 1788 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1789 // then only that memory slice is taken from the call. 1790 // In the last case, we must put an appropriate memory barrier before 1791 // the call, so as to create the correct anti-dependencies on loads 1792 // preceding the call. 1793 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1794 Node* keep_mem, 1795 const TypePtr* hook_mem) { 1796 // no i/o 1797 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1798 if (keep_mem) { 1799 // First clone the existing memory state 1800 set_all_memory(keep_mem); 1801 if (hook_mem != NULL) { 1802 // Make memory for the call 1803 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1804 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1805 // We also use hook_mem to extract specific effects from arraycopy stubs. 1806 set_memory(mem, hook_mem); 1807 } 1808 // ...else the call has NO memory effects. 1809 1810 // Make sure the call advertises its memory effects precisely. 1811 // This lets us build accurate anti-dependences in gcm.cpp. 1812 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1813 "call node must be constructed correctly"); 1814 } else { 1815 assert(hook_mem == NULL, ""); 1816 // This is not a "slow path" call; all memory comes from the call. 1817 set_all_memory_call(call); 1818 } 1819 } 1820 1821 1822 // Replace the call with the current state of the kit. 1823 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1824 JVMState* ejvms = NULL; 1825 if (has_exceptions()) { 1826 ejvms = transfer_exceptions_into_jvms(); 1827 } 1828 1829 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1830 ReplacedNodes replaced_nodes_exception; 1831 Node* ex_ctl = top(); 1832 1833 SafePointNode* final_state = stop(); 1834 1835 // Find all the needed outputs of this call 1836 CallProjections callprojs; 1837 call->extract_projections(&callprojs, true); 1838 1839 Node* init_mem = call->in(TypeFunc::Memory); 1840 Node* final_mem = final_state->in(TypeFunc::Memory); 1841 Node* final_ctl = final_state->in(TypeFunc::Control); 1842 Node* final_io = final_state->in(TypeFunc::I_O); 1843 1844 // Replace all the old call edges with the edges from the inlining result 1845 if (callprojs.fallthrough_catchproj != NULL) { 1846 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1847 } 1848 if (callprojs.fallthrough_memproj != NULL) { 1849 if (final_mem->is_MergeMem()) { 1850 // Parser's exits MergeMem was not transformed but may be optimized 1851 final_mem = _gvn.transform(final_mem); 1852 } 1853 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1854 } 1855 if (callprojs.fallthrough_ioproj != NULL) { 1856 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1857 } 1858 1859 // Replace the result with the new result if it exists and is used 1860 if (callprojs.resproj != NULL && result != NULL) { 1861 C->gvn_replace_by(callprojs.resproj, result); 1862 } 1863 1864 if (ejvms == NULL) { 1865 // No exception edges to simply kill off those paths 1866 if (callprojs.catchall_catchproj != NULL) { 1867 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1868 } 1869 if (callprojs.catchall_memproj != NULL) { 1870 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1871 } 1872 if (callprojs.catchall_ioproj != NULL) { 1873 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1874 } 1875 // Replace the old exception object with top 1876 if (callprojs.exobj != NULL) { 1877 C->gvn_replace_by(callprojs.exobj, C->top()); 1878 } 1879 } else { 1880 GraphKit ekit(ejvms); 1881 1882 // Load my combined exception state into the kit, with all phis transformed: 1883 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1884 replaced_nodes_exception = ex_map->replaced_nodes(); 1885 1886 Node* ex_oop = ekit.use_exception_state(ex_map); 1887 1888 if (callprojs.catchall_catchproj != NULL) { 1889 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1890 ex_ctl = ekit.control(); 1891 } 1892 if (callprojs.catchall_memproj != NULL) { 1893 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1894 } 1895 if (callprojs.catchall_ioproj != NULL) { 1896 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1897 } 1898 1899 // Replace the old exception object with the newly created one 1900 if (callprojs.exobj != NULL) { 1901 C->gvn_replace_by(callprojs.exobj, ex_oop); 1902 } 1903 } 1904 1905 // Disconnect the call from the graph 1906 call->disconnect_inputs(NULL, C); 1907 C->gvn_replace_by(call, C->top()); 1908 1909 // Clean up any MergeMems that feed other MergeMems since the 1910 // optimizer doesn't like that. 1911 if (final_mem->is_MergeMem()) { 1912 Node_List wl; 1913 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1914 Node* m = i.get(); 1915 if (m->is_MergeMem() && !wl.contains(m)) { 1916 wl.push(m); 1917 } 1918 } 1919 while (wl.size() > 0) { 1920 _gvn.transform(wl.pop()); 1921 } 1922 } 1923 1924 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1925 replaced_nodes.apply(C, final_ctl); 1926 } 1927 if (!ex_ctl->is_top() && do_replaced_nodes) { 1928 replaced_nodes_exception.apply(C, ex_ctl); 1929 } 1930 } 1931 1932 1933 //------------------------------increment_counter------------------------------ 1934 // for statistics: increment a VM counter by 1 1935 1936 void GraphKit::increment_counter(address counter_addr) { 1937 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1938 increment_counter(adr1); 1939 } 1940 1941 void GraphKit::increment_counter(Node* counter_addr) { 1942 int adr_type = Compile::AliasIdxRaw; 1943 Node* ctrl = control(); 1944 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1945 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1946 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1947 } 1948 1949 1950 //------------------------------uncommon_trap---------------------------------- 1951 // Bail out to the interpreter in mid-method. Implemented by calling the 1952 // uncommon_trap blob. This helper function inserts a runtime call with the 1953 // right debug info. 1954 void GraphKit::uncommon_trap(int trap_request, 1955 ciKlass* klass, const char* comment, 1956 bool must_throw, 1957 bool keep_exact_action) { 1958 if (failing()) stop(); 1959 if (stopped()) return; // trap reachable? 1960 1961 // Note: If ProfileTraps is true, and if a deopt. actually 1962 // occurs here, the runtime will make sure an MDO exists. There is 1963 // no need to call method()->ensure_method_data() at this point. 1964 1965 // Set the stack pointer to the right value for reexecution: 1966 set_sp(reexecute_sp()); 1967 1968 #ifdef ASSERT 1969 if (!must_throw) { 1970 // Make sure the stack has at least enough depth to execute 1971 // the current bytecode. 1972 int inputs, ignored_depth; 1973 if (compute_stack_effects(inputs, ignored_depth)) { 1974 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1975 Bytecodes::name(java_bc()), sp(), inputs)); 1976 } 1977 } 1978 #endif 1979 1980 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1981 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1982 1983 switch (action) { 1984 case Deoptimization::Action_maybe_recompile: 1985 case Deoptimization::Action_reinterpret: 1986 // Temporary fix for 6529811 to allow virtual calls to be sure they 1987 // get the chance to go from mono->bi->mega 1988 if (!keep_exact_action && 1989 Deoptimization::trap_request_index(trap_request) < 0 && 1990 too_many_recompiles(reason)) { 1991 // This BCI is causing too many recompilations. 1992 if (C->log() != NULL) { 1993 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1994 Deoptimization::trap_reason_name(reason), 1995 Deoptimization::trap_action_name(action)); 1996 } 1997 action = Deoptimization::Action_none; 1998 trap_request = Deoptimization::make_trap_request(reason, action); 1999 } else { 2000 C->set_trap_can_recompile(true); 2001 } 2002 break; 2003 case Deoptimization::Action_make_not_entrant: 2004 C->set_trap_can_recompile(true); 2005 break; 2006 #ifdef ASSERT 2007 case Deoptimization::Action_none: 2008 case Deoptimization::Action_make_not_compilable: 2009 break; 2010 default: 2011 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2012 break; 2013 #endif 2014 } 2015 2016 if (TraceOptoParse) { 2017 char buf[100]; 2018 tty->print_cr("Uncommon trap %s at bci:%d", 2019 Deoptimization::format_trap_request(buf, sizeof(buf), 2020 trap_request), bci()); 2021 } 2022 2023 CompileLog* log = C->log(); 2024 if (log != NULL) { 2025 int kid = (klass == NULL)? -1: log->identify(klass); 2026 log->begin_elem("uncommon_trap bci='%d'", bci()); 2027 char buf[100]; 2028 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2029 trap_request)); 2030 if (kid >= 0) log->print(" klass='%d'", kid); 2031 if (comment != NULL) log->print(" comment='%s'", comment); 2032 log->end_elem(); 2033 } 2034 2035 // Make sure any guarding test views this path as very unlikely 2036 Node *i0 = control()->in(0); 2037 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2038 IfNode *iff = i0->as_If(); 2039 float f = iff->_prob; // Get prob 2040 if (control()->Opcode() == Op_IfTrue) { 2041 if (f > PROB_UNLIKELY_MAG(4)) 2042 iff->_prob = PROB_MIN; 2043 } else { 2044 if (f < PROB_LIKELY_MAG(4)) 2045 iff->_prob = PROB_MAX; 2046 } 2047 } 2048 2049 // Clear out dead values from the debug info. 2050 kill_dead_locals(); 2051 2052 // Now insert the uncommon trap subroutine call 2053 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2054 const TypePtr* no_memory_effects = NULL; 2055 // Pass the index of the class to be loaded 2056 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2057 (must_throw ? RC_MUST_THROW : 0), 2058 OptoRuntime::uncommon_trap_Type(), 2059 call_addr, "uncommon_trap", no_memory_effects, 2060 intcon(trap_request)); 2061 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2062 "must extract request correctly from the graph"); 2063 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2064 2065 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2066 // The debug info is the only real input to this call. 2067 2068 // Halt-and-catch fire here. The above call should never return! 2069 HaltNode* halt = new HaltNode(control(), frameptr()); 2070 _gvn.set_type_bottom(halt); 2071 root()->add_req(halt); 2072 2073 stop_and_kill_map(); 2074 } 2075 2076 2077 //--------------------------just_allocated_object------------------------------ 2078 // Report the object that was just allocated. 2079 // It must be the case that there are no intervening safepoints. 2080 // We use this to determine if an object is so "fresh" that 2081 // it does not require card marks. 2082 Node* GraphKit::just_allocated_object(Node* current_control) { 2083 if (C->recent_alloc_ctl() == current_control) 2084 return C->recent_alloc_obj(); 2085 return NULL; 2086 } 2087 2088 2089 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2090 // (Note: TypeFunc::make has a cache that makes this fast.) 2091 const TypeFunc* tf = TypeFunc::make(dest_method); 2092 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2093 for (int j = 0; j < nargs; j++) { 2094 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2095 if( targ->basic_type() == T_DOUBLE ) { 2096 // If any parameters are doubles, they must be rounded before 2097 // the call, dstore_rounding does gvn.transform 2098 Node *arg = argument(j); 2099 arg = dstore_rounding(arg); 2100 set_argument(j, arg); 2101 } 2102 } 2103 } 2104 2105 /** 2106 * Record profiling data exact_kls for Node n with the type system so 2107 * that it can propagate it (speculation) 2108 * 2109 * @param n node that the type applies to 2110 * @param exact_kls type from profiling 2111 * @param maybe_null did profiling see null? 2112 * 2113 * @return node with improved type 2114 */ 2115 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2116 const Type* current_type = _gvn.type(n); 2117 assert(UseTypeSpeculation, "type speculation must be on"); 2118 2119 const TypePtr* speculative = current_type->speculative(); 2120 2121 // Should the klass from the profile be recorded in the speculative type? 2122 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2123 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2124 const TypeOopPtr* xtype = tklass->as_instance_type(); 2125 assert(xtype->klass_is_exact(), "Should be exact"); 2126 // Any reason to believe n is not null (from this profiling or a previous one)? 2127 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2128 // record the new speculative type's depth 2129 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2130 speculative = speculative->with_inline_depth(jvms()->depth()); 2131 } else if (current_type->would_improve_ptr(maybe_null)) { 2132 // Profiling report that null was never seen so we can change the 2133 // speculative type to non null ptr. 2134 assert(!maybe_null, "nothing to improve"); 2135 if (speculative == NULL) { 2136 speculative = TypePtr::NOTNULL; 2137 } else { 2138 const TypePtr* ptr = TypePtr::NOTNULL; 2139 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2140 } 2141 } 2142 2143 if (speculative != current_type->speculative()) { 2144 // Build a type with a speculative type (what we think we know 2145 // about the type but will need a guard when we use it) 2146 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2147 // We're changing the type, we need a new CheckCast node to carry 2148 // the new type. The new type depends on the control: what 2149 // profiling tells us is only valid from here as far as we can 2150 // tell. 2151 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2152 cast = _gvn.transform(cast); 2153 replace_in_map(n, cast); 2154 n = cast; 2155 } 2156 2157 return n; 2158 } 2159 2160 /** 2161 * Record profiling data from receiver profiling at an invoke with the 2162 * type system so that it can propagate it (speculation) 2163 * 2164 * @param n receiver node 2165 * 2166 * @return node with improved type 2167 */ 2168 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2169 if (!UseTypeSpeculation) { 2170 return n; 2171 } 2172 ciKlass* exact_kls = profile_has_unique_klass(); 2173 bool maybe_null = true; 2174 if (java_bc() == Bytecodes::_checkcast || 2175 java_bc() == Bytecodes::_instanceof || 2176 java_bc() == Bytecodes::_aastore) { 2177 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2178 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2179 } 2180 return record_profile_for_speculation(n, exact_kls, maybe_null); 2181 return n; 2182 } 2183 2184 /** 2185 * Record profiling data from argument profiling at an invoke with the 2186 * type system so that it can propagate it (speculation) 2187 * 2188 * @param dest_method target method for the call 2189 * @param bc what invoke bytecode is this? 2190 */ 2191 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2192 if (!UseTypeSpeculation) { 2193 return; 2194 } 2195 const TypeFunc* tf = TypeFunc::make(dest_method); 2196 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2197 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2198 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2199 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2200 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2201 bool maybe_null = true; 2202 ciKlass* better_type = NULL; 2203 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2204 record_profile_for_speculation(argument(j), better_type, maybe_null); 2205 } 2206 i++; 2207 } 2208 } 2209 } 2210 2211 /** 2212 * Record profiling data from parameter profiling at an invoke with 2213 * the type system so that it can propagate it (speculation) 2214 */ 2215 void GraphKit::record_profiled_parameters_for_speculation() { 2216 if (!UseTypeSpeculation) { 2217 return; 2218 } 2219 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2220 if (_gvn.type(local(i))->isa_oopptr()) { 2221 bool maybe_null = true; 2222 ciKlass* better_type = NULL; 2223 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2224 record_profile_for_speculation(local(i), better_type, maybe_null); 2225 } 2226 j++; 2227 } 2228 } 2229 } 2230 2231 /** 2232 * Record profiling data from return value profiling at an invoke with 2233 * the type system so that it can propagate it (speculation) 2234 */ 2235 void GraphKit::record_profiled_return_for_speculation() { 2236 if (!UseTypeSpeculation) { 2237 return; 2238 } 2239 bool maybe_null = true; 2240 ciKlass* better_type = NULL; 2241 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2242 // If profiling reports a single type for the return value, 2243 // feed it to the type system so it can propagate it as a 2244 // speculative type 2245 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2246 } 2247 } 2248 2249 void GraphKit::round_double_result(ciMethod* dest_method) { 2250 // A non-strict method may return a double value which has an extended 2251 // exponent, but this must not be visible in a caller which is 'strict' 2252 // If a strict caller invokes a non-strict callee, round a double result 2253 2254 BasicType result_type = dest_method->return_type()->basic_type(); 2255 assert( method() != NULL, "must have caller context"); 2256 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2257 // Destination method's return value is on top of stack 2258 // dstore_rounding() does gvn.transform 2259 Node *result = pop_pair(); 2260 result = dstore_rounding(result); 2261 push_pair(result); 2262 } 2263 } 2264 2265 // rounding for strict float precision conformance 2266 Node* GraphKit::precision_rounding(Node* n) { 2267 return UseStrictFP && _method->flags().is_strict() 2268 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2269 ? _gvn.transform( new RoundFloatNode(0, n) ) 2270 : n; 2271 } 2272 2273 // rounding for strict double precision conformance 2274 Node* GraphKit::dprecision_rounding(Node *n) { 2275 return UseStrictFP && _method->flags().is_strict() 2276 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2277 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2278 : n; 2279 } 2280 2281 // rounding for non-strict double stores 2282 Node* GraphKit::dstore_rounding(Node* n) { 2283 return Matcher::strict_fp_requires_explicit_rounding 2284 && UseSSE <= 1 2285 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2286 : n; 2287 } 2288 2289 //============================================================================= 2290 // Generate a fast path/slow path idiom. Graph looks like: 2291 // [foo] indicates that 'foo' is a parameter 2292 // 2293 // [in] NULL 2294 // \ / 2295 // CmpP 2296 // Bool ne 2297 // If 2298 // / \ 2299 // True False-<2> 2300 // / | 2301 // / cast_not_null 2302 // Load | | ^ 2303 // [fast_test] | | 2304 // gvn to opt_test | | 2305 // / \ | <1> 2306 // True False | 2307 // | \\ | 2308 // [slow_call] \[fast_result] 2309 // Ctl Val \ \ 2310 // | \ \ 2311 // Catch <1> \ \ 2312 // / \ ^ \ \ 2313 // Ex No_Ex | \ \ 2314 // | \ \ | \ <2> \ 2315 // ... \ [slow_res] | | \ [null_result] 2316 // \ \--+--+--- | | 2317 // \ | / \ | / 2318 // --------Region Phi 2319 // 2320 //============================================================================= 2321 // Code is structured as a series of driver functions all called 'do_XXX' that 2322 // call a set of helper functions. Helper functions first, then drivers. 2323 2324 //------------------------------null_check_oop--------------------------------- 2325 // Null check oop. Set null-path control into Region in slot 3. 2326 // Make a cast-not-nullness use the other not-null control. Return cast. 2327 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2328 bool never_see_null, 2329 bool safe_for_replace, 2330 bool speculative) { 2331 // Initial NULL check taken path 2332 (*null_control) = top(); 2333 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2334 2335 // Generate uncommon_trap: 2336 if (never_see_null && (*null_control) != top()) { 2337 // If we see an unexpected null at a check-cast we record it and force a 2338 // recompile; the offending check-cast will be compiled to handle NULLs. 2339 // If we see more than one offending BCI, then all checkcasts in the 2340 // method will be compiled to handle NULLs. 2341 PreserveJVMState pjvms(this); 2342 set_control(*null_control); 2343 replace_in_map(value, null()); 2344 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2345 uncommon_trap(reason, 2346 Deoptimization::Action_make_not_entrant); 2347 (*null_control) = top(); // NULL path is dead 2348 } 2349 if ((*null_control) == top() && safe_for_replace) { 2350 replace_in_map(value, cast); 2351 } 2352 2353 // Cast away null-ness on the result 2354 return cast; 2355 } 2356 2357 //------------------------------opt_iff---------------------------------------- 2358 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2359 // Return slow-path control. 2360 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2361 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2362 2363 // Fast path taken; set region slot 2 2364 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2365 region->init_req(2,fast_taken); // Capture fast-control 2366 2367 // Fast path not-taken, i.e. slow path 2368 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2369 return slow_taken; 2370 } 2371 2372 //-----------------------------make_runtime_call------------------------------- 2373 Node* GraphKit::make_runtime_call(int flags, 2374 const TypeFunc* call_type, address call_addr, 2375 const char* call_name, 2376 const TypePtr* adr_type, 2377 // The following parms are all optional. 2378 // The first NULL ends the list. 2379 Node* parm0, Node* parm1, 2380 Node* parm2, Node* parm3, 2381 Node* parm4, Node* parm5, 2382 Node* parm6, Node* parm7) { 2383 // Slow-path call 2384 bool is_leaf = !(flags & RC_NO_LEAF); 2385 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2386 if (call_name == NULL) { 2387 assert(!is_leaf, "must supply name for leaf"); 2388 call_name = OptoRuntime::stub_name(call_addr); 2389 } 2390 CallNode* call; 2391 if (!is_leaf) { 2392 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2393 bci(), adr_type); 2394 } else if (flags & RC_NO_FP) { 2395 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2396 } else { 2397 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2398 } 2399 2400 // The following is similar to set_edges_for_java_call, 2401 // except that the memory effects of the call are restricted to AliasIdxRaw. 2402 2403 // Slow path call has no side-effects, uses few values 2404 bool wide_in = !(flags & RC_NARROW_MEM); 2405 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2406 2407 Node* prev_mem = NULL; 2408 if (wide_in) { 2409 prev_mem = set_predefined_input_for_runtime_call(call); 2410 } else { 2411 assert(!wide_out, "narrow in => narrow out"); 2412 Node* narrow_mem = memory(adr_type); 2413 prev_mem = reset_memory(); 2414 map()->set_memory(narrow_mem); 2415 set_predefined_input_for_runtime_call(call); 2416 } 2417 2418 // Hook each parm in order. Stop looking at the first NULL. 2419 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2420 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2421 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2422 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2423 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2424 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2425 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2426 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2427 /* close each nested if ===> */ } } } } } } } } 2428 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2429 2430 if (!is_leaf) { 2431 // Non-leaves can block and take safepoints: 2432 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2433 } 2434 // Non-leaves can throw exceptions: 2435 if (has_io) { 2436 call->set_req(TypeFunc::I_O, i_o()); 2437 } 2438 2439 if (flags & RC_UNCOMMON) { 2440 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2441 // (An "if" probability corresponds roughly to an unconditional count. 2442 // Sort of.) 2443 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2444 } 2445 2446 Node* c = _gvn.transform(call); 2447 assert(c == call, "cannot disappear"); 2448 2449 if (wide_out) { 2450 // Slow path call has full side-effects. 2451 set_predefined_output_for_runtime_call(call); 2452 } else { 2453 // Slow path call has few side-effects, and/or sets few values. 2454 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2455 } 2456 2457 if (has_io) { 2458 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2459 } 2460 return call; 2461 2462 } 2463 2464 //------------------------------merge_memory----------------------------------- 2465 // Merge memory from one path into the current memory state. 2466 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2467 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2468 Node* old_slice = mms.force_memory(); 2469 Node* new_slice = mms.memory2(); 2470 if (old_slice != new_slice) { 2471 PhiNode* phi; 2472 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2473 if (mms.is_empty()) { 2474 // clone base memory Phi's inputs for this memory slice 2475 assert(old_slice == mms.base_memory(), "sanity"); 2476 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2477 _gvn.set_type(phi, Type::MEMORY); 2478 for (uint i = 1; i < phi->req(); i++) { 2479 phi->init_req(i, old_slice->in(i)); 2480 } 2481 } else { 2482 phi = old_slice->as_Phi(); // Phi was generated already 2483 } 2484 } else { 2485 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2486 _gvn.set_type(phi, Type::MEMORY); 2487 } 2488 phi->set_req(new_path, new_slice); 2489 mms.set_memory(phi); 2490 } 2491 } 2492 } 2493 2494 //------------------------------make_slow_call_ex------------------------------ 2495 // Make the exception handler hookups for the slow call 2496 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2497 if (stopped()) return; 2498 2499 // Make a catch node with just two handlers: fall-through and catch-all 2500 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2501 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2502 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2503 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2504 2505 { PreserveJVMState pjvms(this); 2506 set_control(excp); 2507 set_i_o(i_o); 2508 2509 if (excp != top()) { 2510 if (deoptimize) { 2511 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2512 uncommon_trap(Deoptimization::Reason_unhandled, 2513 Deoptimization::Action_none); 2514 } else { 2515 // Create an exception state also. 2516 // Use an exact type if the caller has specified a specific exception. 2517 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2518 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2519 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2520 } 2521 } 2522 } 2523 2524 // Get the no-exception control from the CatchNode. 2525 set_control(norm); 2526 } 2527 2528 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2529 Node* cmp = NULL; 2530 switch(bt) { 2531 case T_INT: cmp = new CmpINode(in1, in2); break; 2532 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2533 default: fatal(err_msg("unexpected comparison type %s", type2name(bt))); 2534 } 2535 gvn->transform(cmp); 2536 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2537 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2538 gvn->transform(iff); 2539 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2540 return iff; 2541 } 2542 2543 2544 //-------------------------------gen_subtype_check----------------------------- 2545 // Generate a subtyping check. Takes as input the subtype and supertype. 2546 // Returns 2 values: sets the default control() to the true path and returns 2547 // the false path. Only reads invariant memory; sets no (visible) memory. 2548 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2549 // but that's not exposed to the optimizer. This call also doesn't take in an 2550 // Object; if you wish to check an Object you need to load the Object's class 2551 // prior to coming here. 2552 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2553 Compile* C = gvn->C; 2554 // Fast check for identical types, perhaps identical constants. 2555 // The types can even be identical non-constants, in cases 2556 // involving Array.newInstance, Object.clone, etc. 2557 if (subklass == superklass) 2558 return C->top(); // false path is dead; no test needed. 2559 2560 if (gvn->type(superklass)->singleton()) { 2561 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2562 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2563 2564 // In the common case of an exact superklass, try to fold up the 2565 // test before generating code. You may ask, why not just generate 2566 // the code and then let it fold up? The answer is that the generated 2567 // code will necessarily include null checks, which do not always 2568 // completely fold away. If they are also needless, then they turn 2569 // into a performance loss. Example: 2570 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2571 // Here, the type of 'fa' is often exact, so the store check 2572 // of fa[1]=x will fold up, without testing the nullness of x. 2573 switch (C->static_subtype_check(superk, subk)) { 2574 case Compile::SSC_always_false: 2575 { 2576 Node* always_fail = *ctrl; 2577 *ctrl = gvn->C->top(); 2578 return always_fail; 2579 } 2580 case Compile::SSC_always_true: 2581 return C->top(); 2582 case Compile::SSC_easy_test: 2583 { 2584 // Just do a direct pointer compare and be done. 2585 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2586 *ctrl = gvn->transform(new IfTrueNode(iff)); 2587 return gvn->transform(new IfFalseNode(iff)); 2588 } 2589 case Compile::SSC_full_test: 2590 break; 2591 default: 2592 ShouldNotReachHere(); 2593 } 2594 } 2595 2596 // %%% Possible further optimization: Even if the superklass is not exact, 2597 // if the subklass is the unique subtype of the superklass, the check 2598 // will always succeed. We could leave a dependency behind to ensure this. 2599 2600 // First load the super-klass's check-offset 2601 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2602 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2603 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2604 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2605 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2606 2607 // Load from the sub-klass's super-class display list, or a 1-word cache of 2608 // the secondary superclass list, or a failing value with a sentinel offset 2609 // if the super-klass is an interface or exceptionally deep in the Java 2610 // hierarchy and we have to scan the secondary superclass list the hard way. 2611 // Worst-case type is a little odd: NULL is allowed as a result (usually 2612 // klass loads can never produce a NULL). 2613 Node *chk_off_X = chk_off; 2614 #ifdef _LP64 2615 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2616 #endif 2617 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2618 // For some types like interfaces the following loadKlass is from a 1-word 2619 // cache which is mutable so can't use immutable memory. Other 2620 // types load from the super-class display table which is immutable. 2621 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2622 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2623 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2624 2625 // Compile speed common case: ARE a subtype and we canNOT fail 2626 if( superklass == nkls ) 2627 return C->top(); // false path is dead; no test needed. 2628 2629 // See if we get an immediate positive hit. Happens roughly 83% of the 2630 // time. Test to see if the value loaded just previously from the subklass 2631 // is exactly the superklass. 2632 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2633 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2634 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2635 2636 // Compile speed common case: Check for being deterministic right now. If 2637 // chk_off is a constant and not equal to cacheoff then we are NOT a 2638 // subklass. In this case we need exactly the 1 test above and we can 2639 // return those results immediately. 2640 if (!might_be_cache) { 2641 Node* not_subtype_ctrl = *ctrl; 2642 *ctrl = iftrue1; // We need exactly the 1 test above 2643 return not_subtype_ctrl; 2644 } 2645 2646 // Gather the various success & failures here 2647 RegionNode *r_ok_subtype = new RegionNode(4); 2648 gvn->record_for_igvn(r_ok_subtype); 2649 RegionNode *r_not_subtype = new RegionNode(3); 2650 gvn->record_for_igvn(r_not_subtype); 2651 2652 r_ok_subtype->init_req(1, iftrue1); 2653 2654 // Check for immediate negative hit. Happens roughly 11% of the time (which 2655 // is roughly 63% of the remaining cases). Test to see if the loaded 2656 // check-offset points into the subklass display list or the 1-element 2657 // cache. If it points to the display (and NOT the cache) and the display 2658 // missed then it's not a subtype. 2659 Node *cacheoff = gvn->intcon(cacheoff_con); 2660 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2661 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2662 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2663 2664 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2665 // No performance impact (too rare) but allows sharing of secondary arrays 2666 // which has some footprint reduction. 2667 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2668 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2669 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2670 2671 // -- Roads not taken here: -- 2672 // We could also have chosen to perform the self-check at the beginning 2673 // of this code sequence, as the assembler does. This would not pay off 2674 // the same way, since the optimizer, unlike the assembler, can perform 2675 // static type analysis to fold away many successful self-checks. 2676 // Non-foldable self checks work better here in second position, because 2677 // the initial primary superclass check subsumes a self-check for most 2678 // types. An exception would be a secondary type like array-of-interface, 2679 // which does not appear in its own primary supertype display. 2680 // Finally, we could have chosen to move the self-check into the 2681 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2682 // dependent manner. But it is worthwhile to have the check here, 2683 // where it can be perhaps be optimized. The cost in code space is 2684 // small (register compare, branch). 2685 2686 // Now do a linear scan of the secondary super-klass array. Again, no real 2687 // performance impact (too rare) but it's gotta be done. 2688 // Since the code is rarely used, there is no penalty for moving it 2689 // out of line, and it can only improve I-cache density. 2690 // The decision to inline or out-of-line this final check is platform 2691 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2692 Node* psc = gvn->transform( 2693 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2694 2695 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2696 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2697 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2698 2699 // Return false path; set default control to true path. 2700 *ctrl = gvn->transform(r_ok_subtype); 2701 return gvn->transform(r_not_subtype); 2702 } 2703 2704 // Profile-driven exact type check: 2705 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2706 float prob, 2707 Node* *casted_receiver) { 2708 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2709 Node* recv_klass = load_object_klass(receiver); 2710 Node* want_klass = makecon(tklass); 2711 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2712 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2713 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2714 set_control( _gvn.transform( new IfTrueNode (iff) )); 2715 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2716 2717 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2718 assert(recv_xtype->klass_is_exact(), ""); 2719 2720 // Subsume downstream occurrences of receiver with a cast to 2721 // recv_xtype, since now we know what the type will be. 2722 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2723 (*casted_receiver) = _gvn.transform(cast); 2724 // (User must make the replace_in_map call.) 2725 2726 return fail; 2727 } 2728 2729 2730 //------------------------------seems_never_null------------------------------- 2731 // Use null_seen information if it is available from the profile. 2732 // If we see an unexpected null at a type check we record it and force a 2733 // recompile; the offending check will be recompiled to handle NULLs. 2734 // If we see several offending BCIs, then all checks in the 2735 // method will be recompiled. 2736 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2737 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2738 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2739 if (UncommonNullCast // Cutout for this technique 2740 && obj != null() // And not the -Xcomp stupid case? 2741 && !too_many_traps(reason) 2742 ) { 2743 if (speculating) { 2744 return true; 2745 } 2746 if (data == NULL) 2747 // Edge case: no mature data. Be optimistic here. 2748 return true; 2749 // If the profile has not seen a null, assume it won't happen. 2750 assert(java_bc() == Bytecodes::_checkcast || 2751 java_bc() == Bytecodes::_instanceof || 2752 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2753 return !data->as_BitData()->null_seen(); 2754 } 2755 speculating = false; 2756 return false; 2757 } 2758 2759 //------------------------maybe_cast_profiled_receiver------------------------- 2760 // If the profile has seen exactly one type, narrow to exactly that type. 2761 // Subsequent type checks will always fold up. 2762 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2763 ciKlass* require_klass, 2764 ciKlass* spec_klass, 2765 bool safe_for_replace) { 2766 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2767 2768 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2769 2770 // Make sure we haven't already deoptimized from this tactic. 2771 if (too_many_traps(reason) || too_many_recompiles(reason)) 2772 return NULL; 2773 2774 // (No, this isn't a call, but it's enough like a virtual call 2775 // to use the same ciMethod accessor to get the profile info...) 2776 // If we have a speculative type use it instead of profiling (which 2777 // may not help us) 2778 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2779 if (exact_kls != NULL) {// no cast failures here 2780 if (require_klass == NULL || 2781 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2782 // If we narrow the type to match what the type profile sees or 2783 // the speculative type, we can then remove the rest of the 2784 // cast. 2785 // This is a win, even if the exact_kls is very specific, 2786 // because downstream operations, such as method calls, 2787 // will often benefit from the sharper type. 2788 Node* exact_obj = not_null_obj; // will get updated in place... 2789 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2790 &exact_obj); 2791 { PreserveJVMState pjvms(this); 2792 set_control(slow_ctl); 2793 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2794 } 2795 if (safe_for_replace) { 2796 replace_in_map(not_null_obj, exact_obj); 2797 } 2798 return exact_obj; 2799 } 2800 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2801 } 2802 2803 return NULL; 2804 } 2805 2806 /** 2807 * Cast obj to type and emit guard unless we had too many traps here 2808 * already 2809 * 2810 * @param obj node being casted 2811 * @param type type to cast the node to 2812 * @param not_null true if we know node cannot be null 2813 */ 2814 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2815 ciKlass* type, 2816 bool not_null, 2817 SafePointNode* sfpt) { 2818 // type == NULL if profiling tells us this object is always null 2819 if (type != NULL) { 2820 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2821 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2822 ciMethod* trap_method = (sfpt == NULL) ? method() : sfpt->jvms()->method(); 2823 int trap_bci = (sfpt == NULL) ? bci() : sfpt->jvms()->bci(); 2824 2825 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2826 !C->too_many_traps(trap_method, trap_bci, class_reason) && 2827 !C->too_many_recompiles(trap_method, trap_bci, class_reason)) { 2828 Node* not_null_obj = NULL; 2829 // not_null is true if we know the object is not null and 2830 // there's no need for a null check 2831 if (!not_null) { 2832 Node* null_ctl = top(); 2833 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2834 assert(null_ctl->is_top(), "no null control here"); 2835 } else { 2836 not_null_obj = obj; 2837 } 2838 2839 Node* exact_obj = not_null_obj; 2840 ciKlass* exact_kls = type; 2841 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2842 &exact_obj); 2843 if (sfpt != NULL) { 2844 GraphKit kit(sfpt->jvms()); 2845 PreserveJVMState pjvms(&kit); 2846 kit.set_control(slow_ctl); 2847 kit.uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2848 } else { 2849 PreserveJVMState pjvms(this); 2850 set_control(slow_ctl); 2851 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2852 } 2853 replace_in_map(not_null_obj, exact_obj); 2854 obj = exact_obj; 2855 } 2856 } else { 2857 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2858 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2859 Node* exact_obj = null_assert(obj); 2860 replace_in_map(obj, exact_obj); 2861 obj = exact_obj; 2862 } 2863 } 2864 return obj; 2865 } 2866 2867 //-------------------------------gen_instanceof-------------------------------- 2868 // Generate an instance-of idiom. Used by both the instance-of bytecode 2869 // and the reflective instance-of call. 2870 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2871 kill_dead_locals(); // Benefit all the uncommon traps 2872 assert( !stopped(), "dead parse path should be checked in callers" ); 2873 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2874 "must check for not-null not-dead klass in callers"); 2875 2876 // Make the merge point 2877 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2878 RegionNode* region = new RegionNode(PATH_LIMIT); 2879 Node* phi = new PhiNode(region, TypeInt::BOOL); 2880 C->set_has_split_ifs(true); // Has chance for split-if optimization 2881 2882 ciProfileData* data = NULL; 2883 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2884 data = method()->method_data()->bci_to_data(bci()); 2885 } 2886 bool speculative_not_null = false; 2887 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2888 && seems_never_null(obj, data, speculative_not_null)); 2889 2890 // Null check; get casted pointer; set region slot 3 2891 Node* null_ctl = top(); 2892 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2893 2894 // If not_null_obj is dead, only null-path is taken 2895 if (stopped()) { // Doing instance-of on a NULL? 2896 set_control(null_ctl); 2897 return intcon(0); 2898 } 2899 region->init_req(_null_path, null_ctl); 2900 phi ->init_req(_null_path, intcon(0)); // Set null path value 2901 if (null_ctl == top()) { 2902 // Do this eagerly, so that pattern matches like is_diamond_phi 2903 // will work even during parsing. 2904 assert(_null_path == PATH_LIMIT-1, "delete last"); 2905 region->del_req(_null_path); 2906 phi ->del_req(_null_path); 2907 } 2908 2909 // Do we know the type check always succeed? 2910 bool known_statically = false; 2911 if (_gvn.type(superklass)->singleton()) { 2912 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2913 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2914 if (subk != NULL && subk->is_loaded()) { 2915 int static_res = C->static_subtype_check(superk, subk); 2916 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2917 } 2918 } 2919 2920 if (known_statically && UseTypeSpeculation) { 2921 // If we know the type check always succeeds then we don't use the 2922 // profiling data at this bytecode. Don't lose it, feed it to the 2923 // type system as a speculative type. 2924 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2925 } else { 2926 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2927 // We may not have profiling here or it may not help us. If we 2928 // have a speculative type use it to perform an exact cast. 2929 ciKlass* spec_obj_type = obj_type->speculative_type(); 2930 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2931 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2932 if (stopped()) { // Profile disagrees with this path. 2933 set_control(null_ctl); // Null is the only remaining possibility. 2934 return intcon(0); 2935 } 2936 if (cast_obj != NULL) { 2937 not_null_obj = cast_obj; 2938 } 2939 } 2940 } 2941 2942 // Load the object's klass 2943 Node* obj_klass = load_object_klass(not_null_obj); 2944 2945 // Generate the subtype check 2946 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2947 2948 // Plug in the success path to the general merge in slot 1. 2949 region->init_req(_obj_path, control()); 2950 phi ->init_req(_obj_path, intcon(1)); 2951 2952 // Plug in the failing path to the general merge in slot 2. 2953 region->init_req(_fail_path, not_subtype_ctrl); 2954 phi ->init_req(_fail_path, intcon(0)); 2955 2956 // Return final merged results 2957 set_control( _gvn.transform(region) ); 2958 record_for_igvn(region); 2959 return _gvn.transform(phi); 2960 } 2961 2962 //-------------------------------gen_checkcast--------------------------------- 2963 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2964 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2965 // uncommon-trap paths work. Adjust stack after this call. 2966 // If failure_control is supplied and not null, it is filled in with 2967 // the control edge for the cast failure. Otherwise, an appropriate 2968 // uncommon trap or exception is thrown. 2969 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2970 Node* *failure_control) { 2971 kill_dead_locals(); // Benefit all the uncommon traps 2972 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2973 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2974 2975 // Fast cutout: Check the case that the cast is vacuously true. 2976 // This detects the common cases where the test will short-circuit 2977 // away completely. We do this before we perform the null check, 2978 // because if the test is going to turn into zero code, we don't 2979 // want a residual null check left around. (Causes a slowdown, 2980 // for example, in some objArray manipulations, such as a[i]=a[j].) 2981 if (tk->singleton()) { 2982 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2983 if (objtp != NULL && objtp->klass() != NULL) { 2984 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2985 case Compile::SSC_always_true: 2986 // If we know the type check always succeed then we don't use 2987 // the profiling data at this bytecode. Don't lose it, feed it 2988 // to the type system as a speculative type. 2989 return record_profiled_receiver_for_speculation(obj); 2990 case Compile::SSC_always_false: 2991 // It needs a null check because a null will *pass* the cast check. 2992 // A non-null value will always produce an exception. 2993 return null_assert(obj); 2994 } 2995 } 2996 } 2997 2998 ciProfileData* data = NULL; 2999 bool safe_for_replace = false; 3000 if (failure_control == NULL) { // use MDO in regular case only 3001 assert(java_bc() == Bytecodes::_aastore || 3002 java_bc() == Bytecodes::_checkcast, 3003 "interpreter profiles type checks only for these BCs"); 3004 data = method()->method_data()->bci_to_data(bci()); 3005 safe_for_replace = true; 3006 } 3007 3008 // Make the merge point 3009 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3010 RegionNode* region = new RegionNode(PATH_LIMIT); 3011 Node* phi = new PhiNode(region, toop); 3012 C->set_has_split_ifs(true); // Has chance for split-if optimization 3013 3014 // Use null-cast information if it is available 3015 bool speculative_not_null = false; 3016 bool never_see_null = ((failure_control == NULL) // regular case only 3017 && seems_never_null(obj, data, speculative_not_null)); 3018 3019 // Null check; get casted pointer; set region slot 3 3020 Node* null_ctl = top(); 3021 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3022 3023 // If not_null_obj is dead, only null-path is taken 3024 if (stopped()) { // Doing instance-of on a NULL? 3025 set_control(null_ctl); 3026 return null(); 3027 } 3028 region->init_req(_null_path, null_ctl); 3029 phi ->init_req(_null_path, null()); // Set null path value 3030 if (null_ctl == top()) { 3031 // Do this eagerly, so that pattern matches like is_diamond_phi 3032 // will work even during parsing. 3033 assert(_null_path == PATH_LIMIT-1, "delete last"); 3034 region->del_req(_null_path); 3035 phi ->del_req(_null_path); 3036 } 3037 3038 Node* cast_obj = NULL; 3039 if (tk->klass_is_exact()) { 3040 // The following optimization tries to statically cast the speculative type of the object 3041 // (for example obtained during profiling) to the type of the superklass and then do a 3042 // dynamic check that the type of the object is what we expect. To work correctly 3043 // for checkcast and aastore the type of superklass should be exact. 3044 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3045 // We may not have profiling here or it may not help us. If we have 3046 // a speculative type use it to perform an exact cast. 3047 ciKlass* spec_obj_type = obj_type->speculative_type(); 3048 if (spec_obj_type != NULL || 3049 (data != NULL && 3050 // Counter has never been decremented (due to cast failure). 3051 // ...This is a reasonable thing to expect. It is true of 3052 // all casts inserted by javac to implement generic types. 3053 data->as_CounterData()->count() >= 0)) { 3054 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3055 if (cast_obj != NULL) { 3056 if (failure_control != NULL) // failure is now impossible 3057 (*failure_control) = top(); 3058 // adjust the type of the phi to the exact klass: 3059 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3060 } 3061 } 3062 } 3063 3064 if (cast_obj == NULL) { 3065 // Load the object's klass 3066 Node* obj_klass = load_object_klass(not_null_obj); 3067 3068 // Generate the subtype check 3069 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3070 3071 // Plug in success path into the merge 3072 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3073 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3074 if (failure_control == NULL) { 3075 if (not_subtype_ctrl != top()) { // If failure is possible 3076 PreserveJVMState pjvms(this); 3077 set_control(not_subtype_ctrl); 3078 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3079 } 3080 } else { 3081 (*failure_control) = not_subtype_ctrl; 3082 } 3083 } 3084 3085 region->init_req(_obj_path, control()); 3086 phi ->init_req(_obj_path, cast_obj); 3087 3088 // A merge of NULL or Casted-NotNull obj 3089 Node* res = _gvn.transform(phi); 3090 3091 // Note I do NOT always 'replace_in_map(obj,result)' here. 3092 // if( tk->klass()->can_be_primary_super() ) 3093 // This means that if I successfully store an Object into an array-of-String 3094 // I 'forget' that the Object is really now known to be a String. I have to 3095 // do this because we don't have true union types for interfaces - if I store 3096 // a Baz into an array-of-Interface and then tell the optimizer it's an 3097 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3098 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3099 // replace_in_map( obj, res ); 3100 3101 // Return final merged results 3102 set_control( _gvn.transform(region) ); 3103 record_for_igvn(region); 3104 return res; 3105 } 3106 3107 //------------------------------next_monitor----------------------------------- 3108 // What number should be given to the next monitor? 3109 int GraphKit::next_monitor() { 3110 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3111 int next = current + C->sync_stack_slots(); 3112 // Keep the toplevel high water mark current: 3113 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3114 return current; 3115 } 3116 3117 //------------------------------insert_mem_bar--------------------------------- 3118 // Memory barrier to avoid floating things around 3119 // The membar serves as a pinch point between both control and all memory slices. 3120 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3121 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3122 mb->init_req(TypeFunc::Control, control()); 3123 mb->init_req(TypeFunc::Memory, reset_memory()); 3124 Node* membar = _gvn.transform(mb); 3125 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3126 set_all_memory_call(membar); 3127 return membar; 3128 } 3129 3130 //-------------------------insert_mem_bar_volatile---------------------------- 3131 // Memory barrier to avoid floating things around 3132 // The membar serves as a pinch point between both control and memory(alias_idx). 3133 // If you want to make a pinch point on all memory slices, do not use this 3134 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3135 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3136 // When Parse::do_put_xxx updates a volatile field, it appends a series 3137 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3138 // The first membar is on the same memory slice as the field store opcode. 3139 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3140 // All the other membars (for other volatile slices, including AliasIdxBot, 3141 // which stands for all unknown volatile slices) are control-dependent 3142 // on the first membar. This prevents later volatile loads or stores 3143 // from sliding up past the just-emitted store. 3144 3145 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3146 mb->set_req(TypeFunc::Control,control()); 3147 if (alias_idx == Compile::AliasIdxBot) { 3148 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3149 } else { 3150 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3151 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3152 } 3153 Node* membar = _gvn.transform(mb); 3154 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3155 if (alias_idx == Compile::AliasIdxBot) { 3156 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3157 } else { 3158 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3159 } 3160 return membar; 3161 } 3162 3163 //------------------------------shared_lock------------------------------------ 3164 // Emit locking code. 3165 FastLockNode* GraphKit::shared_lock(Node* obj) { 3166 // bci is either a monitorenter bc or InvocationEntryBci 3167 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3168 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3169 3170 if( !GenerateSynchronizationCode ) 3171 return NULL; // Not locking things? 3172 if (stopped()) // Dead monitor? 3173 return NULL; 3174 3175 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3176 3177 // Box the stack location 3178 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3179 Node* mem = reset_memory(); 3180 3181 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3182 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3183 // Create the counters for this fast lock. 3184 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3185 } 3186 3187 // Create the rtm counters for this fast lock if needed. 3188 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3189 3190 // Add monitor to debug info for the slow path. If we block inside the 3191 // slow path and de-opt, we need the monitor hanging around 3192 map()->push_monitor( flock ); 3193 3194 const TypeFunc *tf = LockNode::lock_type(); 3195 LockNode *lock = new LockNode(C, tf); 3196 3197 lock->init_req( TypeFunc::Control, control() ); 3198 lock->init_req( TypeFunc::Memory , mem ); 3199 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3200 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3201 lock->init_req( TypeFunc::ReturnAdr, top() ); 3202 3203 lock->init_req(TypeFunc::Parms + 0, obj); 3204 lock->init_req(TypeFunc::Parms + 1, box); 3205 lock->init_req(TypeFunc::Parms + 2, flock); 3206 add_safepoint_edges(lock); 3207 3208 lock = _gvn.transform( lock )->as_Lock(); 3209 3210 // lock has no side-effects, sets few values 3211 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3212 3213 insert_mem_bar(Op_MemBarAcquireLock); 3214 3215 // Add this to the worklist so that the lock can be eliminated 3216 record_for_igvn(lock); 3217 3218 #ifndef PRODUCT 3219 if (PrintLockStatistics) { 3220 // Update the counter for this lock. Don't bother using an atomic 3221 // operation since we don't require absolute accuracy. 3222 lock->create_lock_counter(map()->jvms()); 3223 increment_counter(lock->counter()->addr()); 3224 } 3225 #endif 3226 3227 return flock; 3228 } 3229 3230 3231 //------------------------------shared_unlock---------------------------------- 3232 // Emit unlocking code. 3233 void GraphKit::shared_unlock(Node* box, Node* obj) { 3234 // bci is either a monitorenter bc or InvocationEntryBci 3235 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3236 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3237 3238 if( !GenerateSynchronizationCode ) 3239 return; 3240 if (stopped()) { // Dead monitor? 3241 map()->pop_monitor(); // Kill monitor from debug info 3242 return; 3243 } 3244 3245 // Memory barrier to avoid floating things down past the locked region 3246 insert_mem_bar(Op_MemBarReleaseLock); 3247 3248 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3249 UnlockNode *unlock = new UnlockNode(C, tf); 3250 uint raw_idx = Compile::AliasIdxRaw; 3251 unlock->init_req( TypeFunc::Control, control() ); 3252 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3253 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3254 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3255 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3256 3257 unlock->init_req(TypeFunc::Parms + 0, obj); 3258 unlock->init_req(TypeFunc::Parms + 1, box); 3259 unlock = _gvn.transform(unlock)->as_Unlock(); 3260 3261 Node* mem = reset_memory(); 3262 3263 // unlock has no side-effects, sets few values 3264 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3265 3266 // Kill monitor from debug info 3267 map()->pop_monitor( ); 3268 } 3269 3270 //-------------------------------get_layout_helper----------------------------- 3271 // If the given klass is a constant or known to be an array, 3272 // fetch the constant layout helper value into constant_value 3273 // and return (Node*)NULL. Otherwise, load the non-constant 3274 // layout helper value, and return the node which represents it. 3275 // This two-faced routine is useful because allocation sites 3276 // almost always feature constant types. 3277 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3278 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3279 if (!StressReflectiveCode && inst_klass != NULL) { 3280 ciKlass* klass = inst_klass->klass(); 3281 bool xklass = inst_klass->klass_is_exact(); 3282 if (xklass || klass->is_array_klass()) { 3283 jint lhelper = klass->layout_helper(); 3284 if (lhelper != Klass::_lh_neutral_value) { 3285 constant_value = lhelper; 3286 return (Node*) NULL; 3287 } 3288 } 3289 } 3290 constant_value = Klass::_lh_neutral_value; // put in a known value 3291 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3292 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3293 } 3294 3295 // We just put in an allocate/initialize with a big raw-memory effect. 3296 // Hook selected additional alias categories on the initialization. 3297 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3298 MergeMemNode* init_in_merge, 3299 Node* init_out_raw) { 3300 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3301 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3302 3303 Node* prevmem = kit.memory(alias_idx); 3304 init_in_merge->set_memory_at(alias_idx, prevmem); 3305 kit.set_memory(init_out_raw, alias_idx); 3306 } 3307 3308 //---------------------------set_output_for_allocation------------------------- 3309 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3310 const TypeOopPtr* oop_type, 3311 bool deoptimize_on_exception) { 3312 int rawidx = Compile::AliasIdxRaw; 3313 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3314 add_safepoint_edges(alloc); 3315 Node* allocx = _gvn.transform(alloc); 3316 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3317 // create memory projection for i_o 3318 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3319 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3320 3321 // create a memory projection as for the normal control path 3322 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3323 set_memory(malloc, rawidx); 3324 3325 // a normal slow-call doesn't change i_o, but an allocation does 3326 // we create a separate i_o projection for the normal control path 3327 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3328 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3329 3330 // put in an initialization barrier 3331 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3332 rawoop)->as_Initialize(); 3333 assert(alloc->initialization() == init, "2-way macro link must work"); 3334 assert(init ->allocation() == alloc, "2-way macro link must work"); 3335 { 3336 // Extract memory strands which may participate in the new object's 3337 // initialization, and source them from the new InitializeNode. 3338 // This will allow us to observe initializations when they occur, 3339 // and link them properly (as a group) to the InitializeNode. 3340 assert(init->in(InitializeNode::Memory) == malloc, ""); 3341 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3342 init->set_req(InitializeNode::Memory, minit_in); 3343 record_for_igvn(minit_in); // fold it up later, if possible 3344 Node* minit_out = memory(rawidx); 3345 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3346 if (oop_type->isa_aryptr()) { 3347 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3348 int elemidx = C->get_alias_index(telemref); 3349 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3350 } else if (oop_type->isa_instptr()) { 3351 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3352 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3353 ciField* field = ik->nonstatic_field_at(i); 3354 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3355 continue; // do not bother to track really large numbers of fields 3356 // Find (or create) the alias category for this field: 3357 int fieldidx = C->alias_type(field)->index(); 3358 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3359 } 3360 } 3361 } 3362 3363 // Cast raw oop to the real thing... 3364 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3365 javaoop = _gvn.transform(javaoop); 3366 C->set_recent_alloc(control(), javaoop); 3367 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3368 3369 #ifdef ASSERT 3370 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3371 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3372 "Ideal_allocation works"); 3373 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3374 "Ideal_allocation works"); 3375 if (alloc->is_AllocateArray()) { 3376 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3377 "Ideal_allocation works"); 3378 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3379 "Ideal_allocation works"); 3380 } else { 3381 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3382 } 3383 } 3384 #endif //ASSERT 3385 3386 return javaoop; 3387 } 3388 3389 //---------------------------new_instance-------------------------------------- 3390 // This routine takes a klass_node which may be constant (for a static type) 3391 // or may be non-constant (for reflective code). It will work equally well 3392 // for either, and the graph will fold nicely if the optimizer later reduces 3393 // the type to a constant. 3394 // The optional arguments are for specialized use by intrinsics: 3395 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3396 // - If 'return_size_val', report the the total object size to the caller. 3397 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3398 Node* GraphKit::new_instance(Node* klass_node, 3399 Node* extra_slow_test, 3400 Node* *return_size_val, 3401 bool deoptimize_on_exception) { 3402 // Compute size in doublewords 3403 // The size is always an integral number of doublewords, represented 3404 // as a positive bytewise size stored in the klass's layout_helper. 3405 // The layout_helper also encodes (in a low bit) the need for a slow path. 3406 jint layout_con = Klass::_lh_neutral_value; 3407 Node* layout_val = get_layout_helper(klass_node, layout_con); 3408 int layout_is_con = (layout_val == NULL); 3409 3410 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3411 // Generate the initial go-slow test. It's either ALWAYS (return a 3412 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3413 // case) a computed value derived from the layout_helper. 3414 Node* initial_slow_test = NULL; 3415 if (layout_is_con) { 3416 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3417 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3418 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3419 3420 } else { // reflective case 3421 // This reflective path is used by Unsafe.allocateInstance. 3422 // (It may be stress-tested by specifying StressReflectiveCode.) 3423 // Basically, we want to get into the VM is there's an illegal argument. 3424 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3425 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3426 if (extra_slow_test != intcon(0)) { 3427 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3428 } 3429 // (Macro-expander will further convert this to a Bool, if necessary.) 3430 } 3431 3432 // Find the size in bytes. This is easy; it's the layout_helper. 3433 // The size value must be valid even if the slow path is taken. 3434 Node* size = NULL; 3435 if (layout_is_con) { 3436 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3437 } else { // reflective case 3438 // This reflective path is used by clone and Unsafe.allocateInstance. 3439 size = ConvI2X(layout_val); 3440 3441 // Clear the low bits to extract layout_helper_size_in_bytes: 3442 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3443 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3444 size = _gvn.transform( new AndXNode(size, mask) ); 3445 } 3446 if (return_size_val != NULL) { 3447 (*return_size_val) = size; 3448 } 3449 3450 // This is a precise notnull oop of the klass. 3451 // (Actually, it need not be precise if this is a reflective allocation.) 3452 // It's what we cast the result to. 3453 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3454 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3455 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3456 3457 // Now generate allocation code 3458 3459 // The entire memory state is needed for slow path of the allocation 3460 // since GC and deoptimization can happened. 3461 Node *mem = reset_memory(); 3462 set_all_memory(mem); // Create new memory state 3463 3464 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3465 control(), mem, i_o(), 3466 size, klass_node, 3467 initial_slow_test); 3468 3469 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3470 } 3471 3472 //-------------------------------new_array------------------------------------- 3473 // helper for both newarray and anewarray 3474 // The 'length' parameter is (obviously) the length of the array. 3475 // See comments on new_instance for the meaning of the other arguments. 3476 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3477 Node* length, // number of array elements 3478 int nargs, // number of arguments to push back for uncommon trap 3479 Node* *return_size_val, 3480 bool deoptimize_on_exception) { 3481 jint layout_con = Klass::_lh_neutral_value; 3482 Node* layout_val = get_layout_helper(klass_node, layout_con); 3483 int layout_is_con = (layout_val == NULL); 3484 3485 if (!layout_is_con && !StressReflectiveCode && 3486 !too_many_traps(Deoptimization::Reason_class_check)) { 3487 // This is a reflective array creation site. 3488 // Optimistically assume that it is a subtype of Object[], 3489 // so that we can fold up all the address arithmetic. 3490 layout_con = Klass::array_layout_helper(T_OBJECT); 3491 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3492 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3493 { BuildCutout unless(this, bol_lh, PROB_MAX); 3494 inc_sp(nargs); 3495 uncommon_trap(Deoptimization::Reason_class_check, 3496 Deoptimization::Action_maybe_recompile); 3497 } 3498 layout_val = NULL; 3499 layout_is_con = true; 3500 } 3501 3502 // Generate the initial go-slow test. Make sure we do not overflow 3503 // if length is huge (near 2Gig) or negative! We do not need 3504 // exact double-words here, just a close approximation of needed 3505 // double-words. We can't add any offset or rounding bits, lest we 3506 // take a size -1 of bytes and make it positive. Use an unsigned 3507 // compare, so negative sizes look hugely positive. 3508 int fast_size_limit = FastAllocateSizeLimit; 3509 if (layout_is_con) { 3510 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3511 // Increase the size limit if we have exact knowledge of array type. 3512 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3513 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3514 } 3515 3516 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3517 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3518 if (initial_slow_test->is_Bool()) { 3519 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3520 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3521 } 3522 3523 // --- Size Computation --- 3524 // array_size = round_to_heap(array_header + (length << elem_shift)); 3525 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3526 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3527 // The rounding mask is strength-reduced, if possible. 3528 int round_mask = MinObjAlignmentInBytes - 1; 3529 Node* header_size = NULL; 3530 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3531 // (T_BYTE has the weakest alignment and size restrictions...) 3532 if (layout_is_con) { 3533 int hsize = Klass::layout_helper_header_size(layout_con); 3534 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3535 BasicType etype = Klass::layout_helper_element_type(layout_con); 3536 if ((round_mask & ~right_n_bits(eshift)) == 0) 3537 round_mask = 0; // strength-reduce it if it goes away completely 3538 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3539 assert(header_size_min <= hsize, "generic minimum is smallest"); 3540 header_size_min = hsize; 3541 header_size = intcon(hsize + round_mask); 3542 } else { 3543 Node* hss = intcon(Klass::_lh_header_size_shift); 3544 Node* hsm = intcon(Klass::_lh_header_size_mask); 3545 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3546 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3547 Node* mask = intcon(round_mask); 3548 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3549 } 3550 3551 Node* elem_shift = NULL; 3552 if (layout_is_con) { 3553 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3554 if (eshift != 0) 3555 elem_shift = intcon(eshift); 3556 } else { 3557 // There is no need to mask or shift this value. 3558 // The semantics of LShiftINode include an implicit mask to 0x1F. 3559 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3560 elem_shift = layout_val; 3561 } 3562 3563 // Transition to native address size for all offset calculations: 3564 Node* lengthx = ConvI2X(length); 3565 Node* headerx = ConvI2X(header_size); 3566 #ifdef _LP64 3567 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3568 if (tllen != NULL && tllen->_lo < 0) { 3569 // Add a manual constraint to a positive range. Cf. array_element_address. 3570 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3571 if (size_max > tllen->_hi) size_max = tllen->_hi; 3572 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3573 lengthx = _gvn.transform( new ConvI2LNode(length, tlcon)); 3574 } 3575 } 3576 #endif 3577 3578 // Combine header size (plus rounding) and body size. Then round down. 3579 // This computation cannot overflow, because it is used only in two 3580 // places, one where the length is sharply limited, and the other 3581 // after a successful allocation. 3582 Node* abody = lengthx; 3583 if (elem_shift != NULL) 3584 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3585 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3586 if (round_mask != 0) { 3587 Node* mask = MakeConX(~round_mask); 3588 size = _gvn.transform( new AndXNode(size, mask) ); 3589 } 3590 // else if round_mask == 0, the size computation is self-rounding 3591 3592 if (return_size_val != NULL) { 3593 // This is the size 3594 (*return_size_val) = size; 3595 } 3596 3597 // Now generate allocation code 3598 3599 // The entire memory state is needed for slow path of the allocation 3600 // since GC and deoptimization can happened. 3601 Node *mem = reset_memory(); 3602 set_all_memory(mem); // Create new memory state 3603 3604 // Create the AllocateArrayNode and its result projections 3605 AllocateArrayNode* alloc 3606 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3607 control(), mem, i_o(), 3608 size, klass_node, 3609 initial_slow_test, 3610 length); 3611 3612 // Cast to correct type. Note that the klass_node may be constant or not, 3613 // and in the latter case the actual array type will be inexact also. 3614 // (This happens via a non-constant argument to inline_native_newArray.) 3615 // In any case, the value of klass_node provides the desired array type. 3616 const TypeInt* length_type = _gvn.find_int_type(length); 3617 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3618 if (ary_type->isa_aryptr() && length_type != NULL) { 3619 // Try to get a better type than POS for the size 3620 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3621 } 3622 3623 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3624 3625 // Cast length on remaining path to be as narrow as possible 3626 if (map()->find_edge(length) >= 0) { 3627 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3628 if (ccast != length) { 3629 _gvn.set_type_bottom(ccast); 3630 record_for_igvn(ccast); 3631 replace_in_map(length, ccast); 3632 } 3633 } 3634 3635 return javaoop; 3636 } 3637 3638 // The following "Ideal_foo" functions are placed here because they recognize 3639 // the graph shapes created by the functions immediately above. 3640 3641 //---------------------------Ideal_allocation---------------------------------- 3642 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3643 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3644 if (ptr == NULL) { // reduce dumb test in callers 3645 return NULL; 3646 } 3647 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3648 ptr = ptr->in(1); 3649 if (ptr == NULL) return NULL; 3650 } 3651 // Return NULL for allocations with several casts: 3652 // j.l.reflect.Array.newInstance(jobject, jint) 3653 // Object.clone() 3654 // to keep more precise type from last cast. 3655 if (ptr->is_Proj()) { 3656 Node* allo = ptr->in(0); 3657 if (allo != NULL && allo->is_Allocate()) { 3658 return allo->as_Allocate(); 3659 } 3660 } 3661 // Report failure to match. 3662 return NULL; 3663 } 3664 3665 // Fancy version which also strips off an offset (and reports it to caller). 3666 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3667 intptr_t& offset) { 3668 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3669 if (base == NULL) return NULL; 3670 return Ideal_allocation(base, phase); 3671 } 3672 3673 // Trace Initialize <- Proj[Parm] <- Allocate 3674 AllocateNode* InitializeNode::allocation() { 3675 Node* rawoop = in(InitializeNode::RawAddress); 3676 if (rawoop->is_Proj()) { 3677 Node* alloc = rawoop->in(0); 3678 if (alloc->is_Allocate()) { 3679 return alloc->as_Allocate(); 3680 } 3681 } 3682 return NULL; 3683 } 3684 3685 // Trace Allocate -> Proj[Parm] -> Initialize 3686 InitializeNode* AllocateNode::initialization() { 3687 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3688 if (rawoop == NULL) return NULL; 3689 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3690 Node* init = rawoop->fast_out(i); 3691 if (init->is_Initialize()) { 3692 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3693 return init->as_Initialize(); 3694 } 3695 } 3696 return NULL; 3697 } 3698 3699 //----------------------------- loop predicates --------------------------- 3700 3701 //------------------------------add_predicate_impl---------------------------- 3702 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3703 // Too many traps seen? 3704 if (too_many_traps(reason)) { 3705 #ifdef ASSERT 3706 if (TraceLoopPredicate) { 3707 int tc = C->trap_count(reason); 3708 tty->print("too many traps=%s tcount=%d in ", 3709 Deoptimization::trap_reason_name(reason), tc); 3710 method()->print(); // which method has too many predicate traps 3711 tty->cr(); 3712 } 3713 #endif 3714 // We cannot afford to take more traps here, 3715 // do not generate predicate. 3716 return; 3717 } 3718 3719 Node *cont = _gvn.intcon(1); 3720 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3721 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3722 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3723 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3724 C->add_predicate_opaq(opq); 3725 { 3726 PreserveJVMState pjvms(this); 3727 set_control(iffalse); 3728 inc_sp(nargs); 3729 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3730 } 3731 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3732 set_control(iftrue); 3733 } 3734 3735 //------------------------------add_predicate--------------------------------- 3736 void GraphKit::add_predicate(int nargs) { 3737 if (UseLoopPredicate) { 3738 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3739 } 3740 // loop's limit check predicate should be near the loop. 3741 if (LoopLimitCheck) { 3742 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3743 } 3744 } 3745 3746 //----------------------------- store barriers ---------------------------- 3747 #define __ ideal. 3748 3749 void GraphKit::sync_kit(IdealKit& ideal) { 3750 set_all_memory(__ merged_memory()); 3751 set_i_o(__ i_o()); 3752 set_control(__ ctrl()); 3753 } 3754 3755 void GraphKit::final_sync(IdealKit& ideal) { 3756 // Final sync IdealKit and graphKit. 3757 sync_kit(ideal); 3758 } 3759 3760 // vanilla/CMS post barrier 3761 // Insert a write-barrier store. This is to let generational GC work; we have 3762 // to flag all oop-stores before the next GC point. 3763 void GraphKit::write_barrier_post(Node* oop_store, 3764 Node* obj, 3765 Node* adr, 3766 uint adr_idx, 3767 Node* val, 3768 bool use_precise) { 3769 // No store check needed if we're storing a NULL or an old object 3770 // (latter case is probably a string constant). The concurrent 3771 // mark sweep garbage collector, however, needs to have all nonNull 3772 // oop updates flagged via card-marks. 3773 if (val != NULL && val->is_Con()) { 3774 // must be either an oop or NULL 3775 const Type* t = val->bottom_type(); 3776 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3777 // stores of null never (?) need barriers 3778 return; 3779 } 3780 3781 if (use_ReduceInitialCardMarks() 3782 && obj == just_allocated_object(control())) { 3783 // We can skip marks on a freshly-allocated object in Eden. 3784 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3785 // That routine informs GC to take appropriate compensating steps, 3786 // upon a slow-path allocation, so as to make this card-mark 3787 // elision safe. 3788 return; 3789 } 3790 3791 if (!use_precise) { 3792 // All card marks for a (non-array) instance are in one place: 3793 adr = obj; 3794 } 3795 // (Else it's an array (or unknown), and we want more precise card marks.) 3796 assert(adr != NULL, ""); 3797 3798 IdealKit ideal(this, true); 3799 3800 // Convert the pointer to an int prior to doing math on it 3801 Node* cast = __ CastPX(__ ctrl(), adr); 3802 3803 // Divide by card size 3804 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3805 "Only one we handle so far."); 3806 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3807 3808 // Combine card table base and card offset 3809 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3810 3811 // Get the alias_index for raw card-mark memory 3812 int adr_type = Compile::AliasIdxRaw; 3813 Node* zero = __ ConI(0); // Dirty card value 3814 BasicType bt = T_BYTE; 3815 3816 if (UseCondCardMark) { 3817 // The classic GC reference write barrier is typically implemented 3818 // as a store into the global card mark table. Unfortunately 3819 // unconditional stores can result in false sharing and excessive 3820 // coherence traffic as well as false transactional aborts. 3821 // UseCondCardMark enables MP "polite" conditional card mark 3822 // stores. In theory we could relax the load from ctrl() to 3823 // no_ctrl, but that doesn't buy much latitude. 3824 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3825 __ if_then(card_val, BoolTest::ne, zero); 3826 } 3827 3828 // Smash zero into card 3829 if( !UseConcMarkSweepGC ) { 3830 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); 3831 } else { 3832 // Specialized path for CM store barrier 3833 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3834 } 3835 3836 if (UseCondCardMark) { 3837 __ end_if(); 3838 } 3839 3840 // Final sync IdealKit and GraphKit. 3841 final_sync(ideal); 3842 } 3843 /* 3844 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3845 * required by SATB to make sure all objects live at the start of the 3846 * marking are kept alive, all reference updates need to any previous 3847 * reference stored before writing. 3848 * 3849 * If the previous value is NULL there is no need to save the old value. 3850 * References that are NULL are filtered during runtime by the barrier 3851 * code to avoid unnecessary queuing. 3852 * 3853 * However in the case of newly allocated objects it might be possible to 3854 * prove that the reference about to be overwritten is NULL during compile 3855 * time and avoid adding the barrier code completely. 3856 * 3857 * The compiler needs to determine that the object in which a field is about 3858 * to be written is newly allocated, and that no prior store to the same field 3859 * has happened since the allocation. 3860 * 3861 * Returns true if the pre-barrier can be removed 3862 */ 3863 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3864 BasicType bt, uint adr_idx) { 3865 intptr_t offset = 0; 3866 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3867 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3868 3869 if (offset == Type::OffsetBot) { 3870 return false; // cannot unalias unless there are precise offsets 3871 } 3872 3873 if (alloc == NULL) { 3874 return false; // No allocation found 3875 } 3876 3877 intptr_t size_in_bytes = type2aelembytes(bt); 3878 3879 Node* mem = memory(adr_idx); // start searching here... 3880 3881 for (int cnt = 0; cnt < 50; cnt++) { 3882 3883 if (mem->is_Store()) { 3884 3885 Node* st_adr = mem->in(MemNode::Address); 3886 intptr_t st_offset = 0; 3887 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3888 3889 if (st_base == NULL) { 3890 break; // inscrutable pointer 3891 } 3892 3893 // Break we have found a store with same base and offset as ours so break 3894 if (st_base == base && st_offset == offset) { 3895 break; 3896 } 3897 3898 if (st_offset != offset && st_offset != Type::OffsetBot) { 3899 const int MAX_STORE = BytesPerLong; 3900 if (st_offset >= offset + size_in_bytes || 3901 st_offset <= offset - MAX_STORE || 3902 st_offset <= offset - mem->as_Store()->memory_size()) { 3903 // Success: The offsets are provably independent. 3904 // (You may ask, why not just test st_offset != offset and be done? 3905 // The answer is that stores of different sizes can co-exist 3906 // in the same sequence of RawMem effects. We sometimes initialize 3907 // a whole 'tile' of array elements with a single jint or jlong.) 3908 mem = mem->in(MemNode::Memory); 3909 continue; // advance through independent store memory 3910 } 3911 } 3912 3913 if (st_base != base 3914 && MemNode::detect_ptr_independence(base, alloc, st_base, 3915 AllocateNode::Ideal_allocation(st_base, phase), 3916 phase)) { 3917 // Success: The bases are provably independent. 3918 mem = mem->in(MemNode::Memory); 3919 continue; // advance through independent store memory 3920 } 3921 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3922 3923 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3924 AllocateNode* st_alloc = st_init->allocation(); 3925 3926 // Make sure that we are looking at the same allocation site. 3927 // The alloc variable is guaranteed to not be null here from earlier check. 3928 if (alloc == st_alloc) { 3929 // Check that the initialization is storing NULL so that no previous store 3930 // has been moved up and directly write a reference 3931 Node* captured_store = st_init->find_captured_store(offset, 3932 type2aelembytes(T_OBJECT), 3933 phase); 3934 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3935 return true; 3936 } 3937 } 3938 } 3939 3940 // Unless there is an explicit 'continue', we must bail out here, 3941 // because 'mem' is an inscrutable memory state (e.g., a call). 3942 break; 3943 } 3944 3945 return false; 3946 } 3947 3948 // G1 pre/post barriers 3949 void GraphKit::g1_write_barrier_pre(bool do_load, 3950 Node* obj, 3951 Node* adr, 3952 uint alias_idx, 3953 Node* val, 3954 const TypeOopPtr* val_type, 3955 Node* pre_val, 3956 BasicType bt) { 3957 3958 // Some sanity checks 3959 // Note: val is unused in this routine. 3960 3961 if (do_load) { 3962 // We need to generate the load of the previous value 3963 assert(obj != NULL, "must have a base"); 3964 assert(adr != NULL, "where are loading from?"); 3965 assert(pre_val == NULL, "loaded already?"); 3966 assert(val_type != NULL, "need a type"); 3967 3968 if (use_ReduceInitialCardMarks() 3969 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 3970 return; 3971 } 3972 3973 } else { 3974 // In this case both val_type and alias_idx are unused. 3975 assert(pre_val != NULL, "must be loaded already"); 3976 // Nothing to be done if pre_val is null. 3977 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3978 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3979 } 3980 assert(bt == T_OBJECT, "or we shouldn't be here"); 3981 3982 IdealKit ideal(this, true); 3983 3984 Node* tls = __ thread(); // ThreadLocalStorage 3985 3986 Node* no_ctrl = NULL; 3987 Node* no_base = __ top(); 3988 Node* zero = __ ConI(0); 3989 Node* zeroX = __ ConX(0); 3990 3991 float likely = PROB_LIKELY(0.999); 3992 float unlikely = PROB_UNLIKELY(0.999); 3993 3994 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3995 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3996 3997 // Offsets into the thread 3998 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3999 PtrQueue::byte_offset_of_active()); 4000 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4001 PtrQueue::byte_offset_of_index()); 4002 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4003 PtrQueue::byte_offset_of_buf()); 4004 4005 // Now the actual pointers into the thread 4006 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4007 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4008 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4009 4010 // Now some of the values 4011 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4012 4013 // if (!marking) 4014 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4015 BasicType index_bt = TypeX_X->basic_type(); 4016 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 4017 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4018 4019 if (do_load) { 4020 // load original value 4021 // alias_idx correct?? 4022 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4023 } 4024 4025 // if (pre_val != NULL) 4026 __ if_then(pre_val, BoolTest::ne, null()); { 4027 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4028 4029 // is the queue for this thread full? 4030 __ if_then(index, BoolTest::ne, zeroX, likely); { 4031 4032 // decrement the index 4033 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4034 4035 // Now get the buffer location we will log the previous value into and store it 4036 Node *log_addr = __ AddP(no_base, buffer, next_index); 4037 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4038 // update the index 4039 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4040 4041 } __ else_(); { 4042 4043 // logging buffer is full, call the runtime 4044 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4045 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4046 } __ end_if(); // (!index) 4047 } __ end_if(); // (pre_val != NULL) 4048 } __ end_if(); // (!marking) 4049 4050 // Final sync IdealKit and GraphKit. 4051 final_sync(ideal); 4052 } 4053 4054 /* 4055 * G1 similar to any GC with a Young Generation requires a way to keep track of 4056 * references from Old Generation to Young Generation to make sure all live 4057 * objects are found. G1 also requires to keep track of object references 4058 * between different regions to enable evacuation of old regions, which is done 4059 * as part of mixed collections. References are tracked in remembered sets and 4060 * is continuously updated as reference are written to with the help of the 4061 * post-barrier. 4062 * 4063 * To reduce the number of updates to the remembered set the post-barrier 4064 * filters updates to fields in objects located in the Young Generation, 4065 * the same region as the reference, when the NULL is being written or 4066 * if the card is already marked as dirty by an earlier write. 4067 * 4068 * Under certain circumstances it is possible to avoid generating the 4069 * post-barrier completely if it is possible during compile time to prove 4070 * the object is newly allocated and that no safepoint exists between the 4071 * allocation and the store. 4072 * 4073 * In the case of slow allocation the allocation code must handle the barrier 4074 * as part of the allocation in the case the allocated object is not located 4075 * in the nursery, this would happen for humongous objects. This is similar to 4076 * how CMS is required to handle this case, see the comments for the method 4077 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4078 * A deferred card mark is required for these objects and handled in the above 4079 * mentioned methods. 4080 * 4081 * Returns true if the post barrier can be removed 4082 */ 4083 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4084 Node* adr) { 4085 intptr_t offset = 0; 4086 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4087 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4088 4089 if (offset == Type::OffsetBot) { 4090 return false; // cannot unalias unless there are precise offsets 4091 } 4092 4093 if (alloc == NULL) { 4094 return false; // No allocation found 4095 } 4096 4097 // Start search from Store node 4098 Node* mem = store->in(MemNode::Control); 4099 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4100 4101 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4102 AllocateNode* st_alloc = st_init->allocation(); 4103 4104 // Make sure we are looking at the same allocation 4105 if (alloc == st_alloc) { 4106 return true; 4107 } 4108 } 4109 4110 return false; 4111 } 4112 4113 // 4114 // Update the card table and add card address to the queue 4115 // 4116 void GraphKit::g1_mark_card(IdealKit& ideal, 4117 Node* card_adr, 4118 Node* oop_store, 4119 uint oop_alias_idx, 4120 Node* index, 4121 Node* index_adr, 4122 Node* buffer, 4123 const TypeFunc* tf) { 4124 4125 Node* zero = __ ConI(0); 4126 Node* zeroX = __ ConX(0); 4127 Node* no_base = __ top(); 4128 BasicType card_bt = T_BYTE; 4129 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4130 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4131 4132 // Now do the queue work 4133 __ if_then(index, BoolTest::ne, zeroX); { 4134 4135 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4136 Node* log_addr = __ AddP(no_base, buffer, next_index); 4137 4138 // Order, see storeCM. 4139 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4140 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4141 4142 } __ else_(); { 4143 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4144 } __ end_if(); 4145 4146 } 4147 4148 void GraphKit::g1_write_barrier_post(Node* oop_store, 4149 Node* obj, 4150 Node* adr, 4151 uint alias_idx, 4152 Node* val, 4153 BasicType bt, 4154 bool use_precise) { 4155 // If we are writing a NULL then we need no post barrier 4156 4157 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4158 // Must be NULL 4159 const Type* t = val->bottom_type(); 4160 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4161 // No post barrier if writing NULLx 4162 return; 4163 } 4164 4165 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4166 // We can skip marks on a freshly-allocated object in Eden. 4167 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4168 // That routine informs GC to take appropriate compensating steps, 4169 // upon a slow-path allocation, so as to make this card-mark 4170 // elision safe. 4171 return; 4172 } 4173 4174 if (use_ReduceInitialCardMarks() 4175 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4176 return; 4177 } 4178 4179 if (!use_precise) { 4180 // All card marks for a (non-array) instance are in one place: 4181 adr = obj; 4182 } 4183 // (Else it's an array (or unknown), and we want more precise card marks.) 4184 assert(adr != NULL, ""); 4185 4186 IdealKit ideal(this, true); 4187 4188 Node* tls = __ thread(); // ThreadLocalStorage 4189 4190 Node* no_base = __ top(); 4191 float likely = PROB_LIKELY(0.999); 4192 float unlikely = PROB_UNLIKELY(0.999); 4193 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4194 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4195 Node* zeroX = __ ConX(0); 4196 4197 // Get the alias_index for raw card-mark memory 4198 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4199 4200 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4201 4202 // Offsets into the thread 4203 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4204 PtrQueue::byte_offset_of_index()); 4205 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4206 PtrQueue::byte_offset_of_buf()); 4207 4208 // Pointers into the thread 4209 4210 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4211 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4212 4213 // Now some values 4214 // Use ctrl to avoid hoisting these values past a safepoint, which could 4215 // potentially reset these fields in the JavaThread. 4216 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4217 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4218 4219 // Convert the store obj pointer to an int prior to doing math on it 4220 // Must use ctrl to prevent "integerized oop" existing across safepoint 4221 Node* cast = __ CastPX(__ ctrl(), adr); 4222 4223 // Divide pointer by card size 4224 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4225 4226 // Combine card table base and card offset 4227 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4228 4229 // If we know the value being stored does it cross regions? 4230 4231 if (val != NULL) { 4232 // Does the store cause us to cross regions? 4233 4234 // Should be able to do an unsigned compare of region_size instead of 4235 // and extra shift. Do we have an unsigned compare?? 4236 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4237 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4238 4239 // if (xor_res == 0) same region so skip 4240 __ if_then(xor_res, BoolTest::ne, zeroX); { 4241 4242 // No barrier if we are storing a NULL 4243 __ if_then(val, BoolTest::ne, null(), unlikely); { 4244 4245 // Ok must mark the card if not already dirty 4246 4247 // load the original value of the card 4248 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4249 4250 __ if_then(card_val, BoolTest::ne, young_card); { 4251 sync_kit(ideal); 4252 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4253 insert_mem_bar(Op_MemBarVolatile, oop_store); 4254 __ sync_kit(this); 4255 4256 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4257 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4258 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4259 } __ end_if(); 4260 } __ end_if(); 4261 } __ end_if(); 4262 } __ end_if(); 4263 } else { 4264 // Object.clone() instrinsic uses this path. 4265 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4266 } 4267 4268 // Final sync IdealKit and GraphKit. 4269 final_sync(ideal); 4270 } 4271 #undef __ 4272 4273 4274 4275 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4276 if (java_lang_String::has_offset_field()) { 4277 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4278 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4279 false, NULL, 0); 4280 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4281 int offset_field_idx = C->get_alias_index(offset_field_type); 4282 return make_load(ctrl, 4283 basic_plus_adr(str, str, offset_offset), 4284 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4285 } else { 4286 return intcon(0); 4287 } 4288 } 4289 4290 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4291 if (java_lang_String::has_count_field()) { 4292 int count_offset = java_lang_String::count_offset_in_bytes(); 4293 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4294 false, NULL, 0); 4295 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4296 int count_field_idx = C->get_alias_index(count_field_type); 4297 return make_load(ctrl, 4298 basic_plus_adr(str, str, count_offset), 4299 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4300 } else { 4301 return load_array_length(load_String_value(ctrl, str)); 4302 } 4303 } 4304 4305 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4306 int value_offset = java_lang_String::value_offset_in_bytes(); 4307 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4308 false, NULL, 0); 4309 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4310 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4311 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4312 ciTypeArrayKlass::make(T_CHAR), true, 0); 4313 int value_field_idx = C->get_alias_index(value_field_type); 4314 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4315 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4316 // String.value field is known to be @Stable. 4317 if (UseImplicitStableValues) { 4318 load = cast_array_to_stable(load, value_type); 4319 } 4320 return load; 4321 } 4322 4323 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4324 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4325 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4326 false, NULL, 0); 4327 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4328 int offset_field_idx = C->get_alias_index(offset_field_type); 4329 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4330 value, T_INT, offset_field_idx, MemNode::unordered); 4331 } 4332 4333 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4334 int value_offset = java_lang_String::value_offset_in_bytes(); 4335 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4336 false, NULL, 0); 4337 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4338 4339 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4340 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4341 } 4342 4343 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4344 int count_offset = java_lang_String::count_offset_in_bytes(); 4345 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4346 false, NULL, 0); 4347 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4348 int count_field_idx = C->get_alias_index(count_field_type); 4349 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4350 value, T_INT, count_field_idx, MemNode::unordered); 4351 } 4352 4353 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4354 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4355 // assumption of CCP analysis. 4356 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4357 }