1 /* 2 * Copyright (c) 2001, 2015, 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/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc/g1/heapRegion.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/cardTableModRefBS.hpp" 31 #include "gc/shared/collectedHeap.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 #ifndef PRODUCT 1183 extern int explicit_null_checks_inserted, 1184 explicit_null_checks_elided; 1185 #endif 1186 Node* GraphKit::null_check_common(Node* value, BasicType type, 1187 // optional arguments for variations: 1188 bool assert_null, 1189 Node* *null_control, 1190 bool speculative) { 1191 assert(!assert_null || null_control == NULL, "not both at once"); 1192 if (stopped()) return top(); 1193 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1194 // For some performance testing, we may wish to suppress null checking. 1195 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1196 return value; 1197 } 1198 NOT_PRODUCT(explicit_null_checks_inserted++); 1199 1200 // Construct NULL check 1201 Node *chk = NULL; 1202 switch(type) { 1203 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1204 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1205 case T_ARRAY : // fall through 1206 type = T_OBJECT; // simplify further tests 1207 case T_OBJECT : { 1208 const Type *t = _gvn.type( value ); 1209 1210 const TypeOopPtr* tp = t->isa_oopptr(); 1211 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1212 // Only for do_null_check, not any of its siblings: 1213 && !assert_null && null_control == NULL) { 1214 // Usually, any field access or invocation on an unloaded oop type 1215 // will simply fail to link, since the statically linked class is 1216 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1217 // the static class is loaded but the sharper oop type is not. 1218 // Rather than checking for this obscure case in lots of places, 1219 // we simply observe that a null check on an unloaded class 1220 // will always be followed by a nonsense operation, so we 1221 // can just issue the uncommon trap here. 1222 // Our access to the unloaded class will only be correct 1223 // after it has been loaded and initialized, which requires 1224 // a trip through the interpreter. 1225 #ifndef PRODUCT 1226 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1227 #endif 1228 uncommon_trap(Deoptimization::Reason_unloaded, 1229 Deoptimization::Action_reinterpret, 1230 tp->klass(), "!loaded"); 1231 return top(); 1232 } 1233 1234 if (assert_null) { 1235 // See if the type is contained in NULL_PTR. 1236 // If so, then the value is already null. 1237 if (t->higher_equal(TypePtr::NULL_PTR)) { 1238 NOT_PRODUCT(explicit_null_checks_elided++); 1239 return value; // Elided null assert quickly! 1240 } 1241 } else { 1242 // See if mixing in the NULL pointer changes type. 1243 // If so, then the NULL pointer was not allowed in the original 1244 // type. In other words, "value" was not-null. 1245 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1246 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1247 NOT_PRODUCT(explicit_null_checks_elided++); 1248 return value; // Elided null check quickly! 1249 } 1250 } 1251 chk = new CmpPNode( value, null() ); 1252 break; 1253 } 1254 1255 default: 1256 fatal("unexpected type: %s", type2name(type)); 1257 } 1258 assert(chk != NULL, "sanity check"); 1259 chk = _gvn.transform(chk); 1260 1261 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1262 BoolNode *btst = new BoolNode( chk, btest); 1263 Node *tst = _gvn.transform( btst ); 1264 1265 //----------- 1266 // if peephole optimizations occurred, a prior test existed. 1267 // If a prior test existed, maybe it dominates as we can avoid this test. 1268 if (tst != btst && type == T_OBJECT) { 1269 // At this point we want to scan up the CFG to see if we can 1270 // find an identical test (and so avoid this test altogether). 1271 Node *cfg = control(); 1272 int depth = 0; 1273 while( depth < 16 ) { // Limit search depth for speed 1274 if( cfg->Opcode() == Op_IfTrue && 1275 cfg->in(0)->in(1) == tst ) { 1276 // Found prior test. Use "cast_not_null" to construct an identical 1277 // CastPP (and hence hash to) as already exists for the prior test. 1278 // Return that casted value. 1279 if (assert_null) { 1280 replace_in_map(value, null()); 1281 return null(); // do not issue the redundant test 1282 } 1283 Node *oldcontrol = control(); 1284 set_control(cfg); 1285 Node *res = cast_not_null(value); 1286 set_control(oldcontrol); 1287 NOT_PRODUCT(explicit_null_checks_elided++); 1288 return res; 1289 } 1290 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1291 if (cfg == NULL) break; // Quit at region nodes 1292 depth++; 1293 } 1294 } 1295 1296 //----------- 1297 // Branch to failure if null 1298 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1299 Deoptimization::DeoptReason reason; 1300 if (assert_null) { 1301 reason = Deoptimization::Reason_null_assert; 1302 } else if (type == T_OBJECT) { 1303 reason = Deoptimization::reason_null_check(speculative); 1304 } else { 1305 reason = Deoptimization::Reason_div0_check; 1306 } 1307 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1308 // ciMethodData::has_trap_at will return a conservative -1 if any 1309 // must-be-null assertion has failed. This could cause performance 1310 // problems for a method after its first do_null_assert failure. 1311 // Consider using 'Reason_class_check' instead? 1312 1313 // To cause an implicit null check, we set the not-null probability 1314 // to the maximum (PROB_MAX). For an explicit check the probability 1315 // is set to a smaller value. 1316 if (null_control != NULL || too_many_traps(reason)) { 1317 // probability is less likely 1318 ok_prob = PROB_LIKELY_MAG(3); 1319 } else if (!assert_null && 1320 (ImplicitNullCheckThreshold > 0) && 1321 method() != NULL && 1322 (method()->method_data()->trap_count(reason) 1323 >= (uint)ImplicitNullCheckThreshold)) { 1324 ok_prob = PROB_LIKELY_MAG(3); 1325 } 1326 1327 if (null_control != NULL) { 1328 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1329 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1330 set_control( _gvn.transform( new IfTrueNode(iff))); 1331 #ifndef PRODUCT 1332 if (null_true == top()) { 1333 explicit_null_checks_elided++; 1334 } 1335 #endif 1336 (*null_control) = null_true; 1337 } else { 1338 BuildCutout unless(this, tst, ok_prob); 1339 // Check for optimizer eliding test at parse time 1340 if (stopped()) { 1341 // Failure not possible; do not bother making uncommon trap. 1342 NOT_PRODUCT(explicit_null_checks_elided++); 1343 } else if (assert_null) { 1344 uncommon_trap(reason, 1345 Deoptimization::Action_make_not_entrant, 1346 NULL, "assert_null"); 1347 } else { 1348 replace_in_map(value, zerocon(type)); 1349 builtin_throw(reason); 1350 } 1351 } 1352 1353 // Must throw exception, fall-thru not possible? 1354 if (stopped()) { 1355 return top(); // No result 1356 } 1357 1358 if (assert_null) { 1359 // Cast obj to null on this path. 1360 replace_in_map(value, zerocon(type)); 1361 return zerocon(type); 1362 } 1363 1364 // Cast obj to not-null on this path, if there is no null_control. 1365 // (If there is a null_control, a non-null value may come back to haunt us.) 1366 if (type == T_OBJECT) { 1367 Node* cast = cast_not_null(value, false); 1368 if (null_control == NULL || (*null_control) == top()) 1369 replace_in_map(value, cast); 1370 value = cast; 1371 } 1372 1373 return value; 1374 } 1375 1376 1377 //------------------------------cast_not_null---------------------------------- 1378 // Cast obj to not-null on this path 1379 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1380 const Type *t = _gvn.type(obj); 1381 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1382 // Object is already not-null? 1383 if( t == t_not_null ) return obj; 1384 1385 Node *cast = new CastPPNode(obj,t_not_null); 1386 cast->init_req(0, control()); 1387 cast = _gvn.transform( cast ); 1388 1389 // Scan for instances of 'obj' in the current JVM mapping. 1390 // These instances are known to be not-null after the test. 1391 if (do_replace_in_map) 1392 replace_in_map(obj, cast); 1393 1394 return cast; // Return casted value 1395 } 1396 1397 1398 //--------------------------replace_in_map------------------------------------- 1399 void GraphKit::replace_in_map(Node* old, Node* neww) { 1400 if (old == neww) { 1401 return; 1402 } 1403 1404 map()->replace_edge(old, neww); 1405 1406 // Note: This operation potentially replaces any edge 1407 // on the map. This includes locals, stack, and monitors 1408 // of the current (innermost) JVM state. 1409 1410 // don't let inconsistent types from profiling escape this 1411 // method 1412 1413 const Type* told = _gvn.type(old); 1414 const Type* tnew = _gvn.type(neww); 1415 1416 if (!tnew->higher_equal(told)) { 1417 return; 1418 } 1419 1420 map()->record_replaced_node(old, neww); 1421 } 1422 1423 1424 //============================================================================= 1425 //--------------------------------memory--------------------------------------- 1426 Node* GraphKit::memory(uint alias_idx) { 1427 MergeMemNode* mem = merged_memory(); 1428 Node* p = mem->memory_at(alias_idx); 1429 _gvn.set_type(p, Type::MEMORY); // must be mapped 1430 return p; 1431 } 1432 1433 //-----------------------------reset_memory------------------------------------ 1434 Node* GraphKit::reset_memory() { 1435 Node* mem = map()->memory(); 1436 // do not use this node for any more parsing! 1437 debug_only( map()->set_memory((Node*)NULL) ); 1438 return _gvn.transform( mem ); 1439 } 1440 1441 //------------------------------set_all_memory--------------------------------- 1442 void GraphKit::set_all_memory(Node* newmem) { 1443 Node* mergemem = MergeMemNode::make(newmem); 1444 gvn().set_type_bottom(mergemem); 1445 map()->set_memory(mergemem); 1446 } 1447 1448 //------------------------------set_all_memory_call---------------------------- 1449 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1450 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1451 set_all_memory(newmem); 1452 } 1453 1454 //============================================================================= 1455 // 1456 // parser factory methods for MemNodes 1457 // 1458 // These are layered on top of the factory methods in LoadNode and StoreNode, 1459 // and integrate with the parser's memory state and _gvn engine. 1460 // 1461 1462 // factory methods in "int adr_idx" 1463 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1464 int adr_idx, 1465 MemNode::MemOrd mo, 1466 LoadNode::ControlDependency control_dependency, 1467 bool require_atomic_access, 1468 bool unaligned, 1469 bool mismatched) { 1470 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1471 const TypePtr* adr_type = NULL; // debug-mode-only argument 1472 debug_only(adr_type = C->get_adr_type(adr_idx)); 1473 Node* mem = memory(adr_idx); 1474 Node* ld; 1475 if (require_atomic_access && bt == T_LONG) { 1476 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1477 } else if (require_atomic_access && bt == T_DOUBLE) { 1478 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1479 } else { 1480 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1481 } 1482 ld = _gvn.transform(ld); 1483 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1484 // Improve graph before escape analysis and boxing elimination. 1485 record_for_igvn(ld); 1486 } 1487 return ld; 1488 } 1489 1490 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1491 int adr_idx, 1492 MemNode::MemOrd mo, 1493 bool require_atomic_access, 1494 bool unaligned, 1495 bool mismatched) { 1496 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1497 const TypePtr* adr_type = NULL; 1498 debug_only(adr_type = C->get_adr_type(adr_idx)); 1499 Node *mem = memory(adr_idx); 1500 Node* st; 1501 if (require_atomic_access && bt == T_LONG) { 1502 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1503 } else if (require_atomic_access && bt == T_DOUBLE) { 1504 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1505 } else { 1506 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1507 } 1508 if (unaligned) { 1509 st->as_Store()->set_unaligned_access(); 1510 } 1511 if (mismatched) { 1512 st->as_Store()->set_mismatched_access(); 1513 } 1514 st = _gvn.transform(st); 1515 set_memory(st, adr_idx); 1516 // Back-to-back stores can only remove intermediate store with DU info 1517 // so push on worklist for optimizer. 1518 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1519 record_for_igvn(st); 1520 1521 return st; 1522 } 1523 1524 1525 void GraphKit::pre_barrier(bool do_load, 1526 Node* ctl, 1527 Node* obj, 1528 Node* adr, 1529 uint adr_idx, 1530 Node* val, 1531 const TypeOopPtr* val_type, 1532 Node* pre_val, 1533 BasicType bt) { 1534 1535 BarrierSet* bs = Universe::heap()->barrier_set(); 1536 set_control(ctl); 1537 switch (bs->kind()) { 1538 case BarrierSet::G1SATBCTLogging: 1539 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1540 break; 1541 1542 case BarrierSet::CardTableForRS: 1543 case BarrierSet::CardTableExtension: 1544 case BarrierSet::ModRef: 1545 break; 1546 1547 default : 1548 ShouldNotReachHere(); 1549 1550 } 1551 } 1552 1553 bool GraphKit::can_move_pre_barrier() const { 1554 BarrierSet* bs = Universe::heap()->barrier_set(); 1555 switch (bs->kind()) { 1556 case BarrierSet::G1SATBCTLogging: 1557 return true; // Can move it if no safepoint 1558 1559 case BarrierSet::CardTableForRS: 1560 case BarrierSet::CardTableExtension: 1561 case BarrierSet::ModRef: 1562 return true; // There is no pre-barrier 1563 1564 default : 1565 ShouldNotReachHere(); 1566 } 1567 return false; 1568 } 1569 1570 void GraphKit::post_barrier(Node* ctl, 1571 Node* store, 1572 Node* obj, 1573 Node* adr, 1574 uint adr_idx, 1575 Node* val, 1576 BasicType bt, 1577 bool use_precise) { 1578 BarrierSet* bs = Universe::heap()->barrier_set(); 1579 set_control(ctl); 1580 switch (bs->kind()) { 1581 case BarrierSet::G1SATBCTLogging: 1582 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1583 break; 1584 1585 case BarrierSet::CardTableForRS: 1586 case BarrierSet::CardTableExtension: 1587 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1588 break; 1589 1590 case BarrierSet::ModRef: 1591 break; 1592 1593 default : 1594 ShouldNotReachHere(); 1595 1596 } 1597 } 1598 1599 Node* GraphKit::store_oop(Node* ctl, 1600 Node* obj, 1601 Node* adr, 1602 const TypePtr* adr_type, 1603 Node* val, 1604 const TypeOopPtr* val_type, 1605 BasicType bt, 1606 bool use_precise, 1607 MemNode::MemOrd mo, 1608 bool mismatched) { 1609 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1610 // could be delayed during Parse (for example, in adjust_map_after_if()). 1611 // Execute transformation here to avoid barrier generation in such case. 1612 if (_gvn.type(val) == TypePtr::NULL_PTR) 1613 val = _gvn.makecon(TypePtr::NULL_PTR); 1614 1615 set_control(ctl); 1616 if (stopped()) return top(); // Dead path ? 1617 1618 assert(bt == T_OBJECT, "sanity"); 1619 assert(val != NULL, "not dead path"); 1620 uint adr_idx = C->get_alias_index(adr_type); 1621 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1622 1623 pre_barrier(true /* do_load */, 1624 control(), obj, adr, adr_idx, val, val_type, 1625 NULL /* pre_val */, 1626 bt); 1627 1628 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1629 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1630 return store; 1631 } 1632 1633 // Could be an array or object we don't know at compile time (unsafe ref.) 1634 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1635 Node* obj, // containing obj 1636 Node* adr, // actual adress to store val at 1637 const TypePtr* adr_type, 1638 Node* val, 1639 BasicType bt, 1640 MemNode::MemOrd mo, 1641 bool mismatched) { 1642 Compile::AliasType* at = C->alias_type(adr_type); 1643 const TypeOopPtr* val_type = NULL; 1644 if (adr_type->isa_instptr()) { 1645 if (at->field() != NULL) { 1646 // known field. This code is a copy of the do_put_xxx logic. 1647 ciField* field = at->field(); 1648 if (!field->type()->is_loaded()) { 1649 val_type = TypeInstPtr::BOTTOM; 1650 } else { 1651 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1652 } 1653 } 1654 } else if (adr_type->isa_aryptr()) { 1655 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1656 } 1657 if (val_type == NULL) { 1658 val_type = TypeInstPtr::BOTTOM; 1659 } 1660 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1661 } 1662 1663 1664 //-------------------------array_element_address------------------------- 1665 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1666 const TypeInt* sizetype, Node* ctrl) { 1667 uint shift = exact_log2(type2aelembytes(elembt)); 1668 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1669 1670 // short-circuit a common case (saves lots of confusing waste motion) 1671 jint idx_con = find_int_con(idx, -1); 1672 if (idx_con >= 0) { 1673 intptr_t offset = header + ((intptr_t)idx_con << shift); 1674 return basic_plus_adr(ary, offset); 1675 } 1676 1677 // must be correct type for alignment purposes 1678 Node* base = basic_plus_adr(ary, header); 1679 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1680 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1681 return basic_plus_adr(ary, base, scale); 1682 } 1683 1684 //-------------------------load_array_element------------------------- 1685 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1686 const Type* elemtype = arytype->elem(); 1687 BasicType elembt = elemtype->array_element_basic_type(); 1688 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1689 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1690 return ld; 1691 } 1692 1693 //-------------------------set_arguments_for_java_call------------------------- 1694 // Arguments (pre-popped from the stack) are taken from the JVMS. 1695 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1696 // Add the call arguments: 1697 uint nargs = call->method()->arg_size(); 1698 for (uint i = 0; i < nargs; i++) { 1699 Node* arg = argument(i); 1700 call->init_req(i + TypeFunc::Parms, arg); 1701 } 1702 } 1703 1704 //---------------------------set_edges_for_java_call--------------------------- 1705 // Connect a newly created call into the current JVMS. 1706 // A return value node (if any) is returned from set_edges_for_java_call. 1707 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1708 1709 // Add the predefined inputs: 1710 call->init_req( TypeFunc::Control, control() ); 1711 call->init_req( TypeFunc::I_O , i_o() ); 1712 call->init_req( TypeFunc::Memory , reset_memory() ); 1713 call->init_req( TypeFunc::FramePtr, frameptr() ); 1714 call->init_req( TypeFunc::ReturnAdr, top() ); 1715 1716 add_safepoint_edges(call, must_throw); 1717 1718 Node* xcall = _gvn.transform(call); 1719 1720 if (xcall == top()) { 1721 set_control(top()); 1722 return; 1723 } 1724 assert(xcall == call, "call identity is stable"); 1725 1726 // Re-use the current map to produce the result. 1727 1728 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1729 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1730 set_all_memory_call(xcall, separate_io_proj); 1731 1732 //return xcall; // no need, caller already has it 1733 } 1734 1735 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1736 if (stopped()) return top(); // maybe the call folded up? 1737 1738 // Capture the return value, if any. 1739 Node* ret; 1740 if (call->method() == NULL || 1741 call->method()->return_type()->basic_type() == T_VOID) 1742 ret = top(); 1743 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1744 1745 // Note: Since any out-of-line call can produce an exception, 1746 // we always insert an I_O projection from the call into the result. 1747 1748 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1749 1750 if (separate_io_proj) { 1751 // The caller requested separate projections be used by the fall 1752 // through and exceptional paths, so replace the projections for 1753 // the fall through path. 1754 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1755 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1756 } 1757 return ret; 1758 } 1759 1760 //--------------------set_predefined_input_for_runtime_call-------------------- 1761 // Reading and setting the memory state is way conservative here. 1762 // The real problem is that I am not doing real Type analysis on memory, 1763 // so I cannot distinguish card mark stores from other stores. Across a GC 1764 // point the Store Barrier and the card mark memory has to agree. I cannot 1765 // have a card mark store and its barrier split across the GC point from 1766 // either above or below. Here I get that to happen by reading ALL of memory. 1767 // A better answer would be to separate out card marks from other memory. 1768 // For now, return the input memory state, so that it can be reused 1769 // after the call, if this call has restricted memory effects. 1770 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1771 // Set fixed predefined input arguments 1772 Node* memory = reset_memory(); 1773 call->init_req( TypeFunc::Control, control() ); 1774 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1775 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1776 call->init_req( TypeFunc::FramePtr, frameptr() ); 1777 call->init_req( TypeFunc::ReturnAdr, top() ); 1778 return memory; 1779 } 1780 1781 //-------------------set_predefined_output_for_runtime_call-------------------- 1782 // Set control and memory (not i_o) from the call. 1783 // If keep_mem is not NULL, use it for the output state, 1784 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1785 // If hook_mem is NULL, this call produces no memory effects at all. 1786 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1787 // then only that memory slice is taken from the call. 1788 // In the last case, we must put an appropriate memory barrier before 1789 // the call, so as to create the correct anti-dependencies on loads 1790 // preceding the call. 1791 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1792 Node* keep_mem, 1793 const TypePtr* hook_mem) { 1794 // no i/o 1795 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1796 if (keep_mem) { 1797 // First clone the existing memory state 1798 set_all_memory(keep_mem); 1799 if (hook_mem != NULL) { 1800 // Make memory for the call 1801 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1802 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1803 // We also use hook_mem to extract specific effects from arraycopy stubs. 1804 set_memory(mem, hook_mem); 1805 } 1806 // ...else the call has NO memory effects. 1807 1808 // Make sure the call advertises its memory effects precisely. 1809 // This lets us build accurate anti-dependences in gcm.cpp. 1810 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1811 "call node must be constructed correctly"); 1812 } else { 1813 assert(hook_mem == NULL, ""); 1814 // This is not a "slow path" call; all memory comes from the call. 1815 set_all_memory_call(call); 1816 } 1817 } 1818 1819 1820 // Replace the call with the current state of the kit. 1821 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1822 JVMState* ejvms = NULL; 1823 if (has_exceptions()) { 1824 ejvms = transfer_exceptions_into_jvms(); 1825 } 1826 1827 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1828 ReplacedNodes replaced_nodes_exception; 1829 Node* ex_ctl = top(); 1830 1831 SafePointNode* final_state = stop(); 1832 1833 // Find all the needed outputs of this call 1834 CallProjections callprojs; 1835 call->extract_projections(&callprojs, true); 1836 1837 Node* init_mem = call->in(TypeFunc::Memory); 1838 Node* final_mem = final_state->in(TypeFunc::Memory); 1839 Node* final_ctl = final_state->in(TypeFunc::Control); 1840 Node* final_io = final_state->in(TypeFunc::I_O); 1841 1842 // Replace all the old call edges with the edges from the inlining result 1843 if (callprojs.fallthrough_catchproj != NULL) { 1844 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1845 } 1846 if (callprojs.fallthrough_memproj != NULL) { 1847 if (final_mem->is_MergeMem()) { 1848 // Parser's exits MergeMem was not transformed but may be optimized 1849 final_mem = _gvn.transform(final_mem); 1850 } 1851 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1852 } 1853 if (callprojs.fallthrough_ioproj != NULL) { 1854 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1855 } 1856 1857 // Replace the result with the new result if it exists and is used 1858 if (callprojs.resproj != NULL && result != NULL) { 1859 C->gvn_replace_by(callprojs.resproj, result); 1860 } 1861 1862 if (ejvms == NULL) { 1863 // No exception edges to simply kill off those paths 1864 if (callprojs.catchall_catchproj != NULL) { 1865 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1866 } 1867 if (callprojs.catchall_memproj != NULL) { 1868 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1869 } 1870 if (callprojs.catchall_ioproj != NULL) { 1871 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1872 } 1873 // Replace the old exception object with top 1874 if (callprojs.exobj != NULL) { 1875 C->gvn_replace_by(callprojs.exobj, C->top()); 1876 } 1877 } else { 1878 GraphKit ekit(ejvms); 1879 1880 // Load my combined exception state into the kit, with all phis transformed: 1881 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1882 replaced_nodes_exception = ex_map->replaced_nodes(); 1883 1884 Node* ex_oop = ekit.use_exception_state(ex_map); 1885 1886 if (callprojs.catchall_catchproj != NULL) { 1887 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1888 ex_ctl = ekit.control(); 1889 } 1890 if (callprojs.catchall_memproj != NULL) { 1891 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1892 } 1893 if (callprojs.catchall_ioproj != NULL) { 1894 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1895 } 1896 1897 // Replace the old exception object with the newly created one 1898 if (callprojs.exobj != NULL) { 1899 C->gvn_replace_by(callprojs.exobj, ex_oop); 1900 } 1901 } 1902 1903 // Disconnect the call from the graph 1904 call->disconnect_inputs(NULL, C); 1905 C->gvn_replace_by(call, C->top()); 1906 1907 // Clean up any MergeMems that feed other MergeMems since the 1908 // optimizer doesn't like that. 1909 if (final_mem->is_MergeMem()) { 1910 Node_List wl; 1911 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1912 Node* m = i.get(); 1913 if (m->is_MergeMem() && !wl.contains(m)) { 1914 wl.push(m); 1915 } 1916 } 1917 while (wl.size() > 0) { 1918 _gvn.transform(wl.pop()); 1919 } 1920 } 1921 1922 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1923 replaced_nodes.apply(C, final_ctl); 1924 } 1925 if (!ex_ctl->is_top() && do_replaced_nodes) { 1926 replaced_nodes_exception.apply(C, ex_ctl); 1927 } 1928 } 1929 1930 1931 //------------------------------increment_counter------------------------------ 1932 // for statistics: increment a VM counter by 1 1933 1934 void GraphKit::increment_counter(address counter_addr) { 1935 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1936 increment_counter(adr1); 1937 } 1938 1939 void GraphKit::increment_counter(Node* counter_addr) { 1940 int adr_type = Compile::AliasIdxRaw; 1941 Node* ctrl = control(); 1942 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1943 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1944 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1945 } 1946 1947 1948 //------------------------------uncommon_trap---------------------------------- 1949 // Bail out to the interpreter in mid-method. Implemented by calling the 1950 // uncommon_trap blob. This helper function inserts a runtime call with the 1951 // right debug info. 1952 void GraphKit::uncommon_trap(int trap_request, 1953 ciKlass* klass, const char* comment, 1954 bool must_throw, 1955 bool keep_exact_action) { 1956 if (failing()) stop(); 1957 if (stopped()) return; // trap reachable? 1958 1959 // Note: If ProfileTraps is true, and if a deopt. actually 1960 // occurs here, the runtime will make sure an MDO exists. There is 1961 // no need to call method()->ensure_method_data() at this point. 1962 1963 // Set the stack pointer to the right value for reexecution: 1964 set_sp(reexecute_sp()); 1965 1966 #ifdef ASSERT 1967 if (!must_throw) { 1968 // Make sure the stack has at least enough depth to execute 1969 // the current bytecode. 1970 int inputs, ignored_depth; 1971 if (compute_stack_effects(inputs, ignored_depth)) { 1972 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1973 Bytecodes::name(java_bc()), sp(), inputs); 1974 } 1975 } 1976 #endif 1977 1978 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1979 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1980 1981 switch (action) { 1982 case Deoptimization::Action_maybe_recompile: 1983 case Deoptimization::Action_reinterpret: 1984 // Temporary fix for 6529811 to allow virtual calls to be sure they 1985 // get the chance to go from mono->bi->mega 1986 if (!keep_exact_action && 1987 Deoptimization::trap_request_index(trap_request) < 0 && 1988 too_many_recompiles(reason)) { 1989 // This BCI is causing too many recompilations. 1990 if (C->log() != NULL) { 1991 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1992 Deoptimization::trap_reason_name(reason), 1993 Deoptimization::trap_action_name(action)); 1994 } 1995 action = Deoptimization::Action_none; 1996 trap_request = Deoptimization::make_trap_request(reason, action); 1997 } else { 1998 C->set_trap_can_recompile(true); 1999 } 2000 break; 2001 case Deoptimization::Action_make_not_entrant: 2002 C->set_trap_can_recompile(true); 2003 break; 2004 #ifdef ASSERT 2005 case Deoptimization::Action_none: 2006 case Deoptimization::Action_make_not_compilable: 2007 break; 2008 default: 2009 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2010 break; 2011 #endif 2012 } 2013 2014 if (TraceOptoParse) { 2015 char buf[100]; 2016 tty->print_cr("Uncommon trap %s at bci:%d", 2017 Deoptimization::format_trap_request(buf, sizeof(buf), 2018 trap_request), bci()); 2019 } 2020 2021 CompileLog* log = C->log(); 2022 if (log != NULL) { 2023 int kid = (klass == NULL)? -1: log->identify(klass); 2024 log->begin_elem("uncommon_trap bci='%d'", bci()); 2025 char buf[100]; 2026 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2027 trap_request)); 2028 if (kid >= 0) log->print(" klass='%d'", kid); 2029 if (comment != NULL) log->print(" comment='%s'", comment); 2030 log->end_elem(); 2031 } 2032 2033 // Make sure any guarding test views this path as very unlikely 2034 Node *i0 = control()->in(0); 2035 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2036 IfNode *iff = i0->as_If(); 2037 float f = iff->_prob; // Get prob 2038 if (control()->Opcode() == Op_IfTrue) { 2039 if (f > PROB_UNLIKELY_MAG(4)) 2040 iff->_prob = PROB_MIN; 2041 } else { 2042 if (f < PROB_LIKELY_MAG(4)) 2043 iff->_prob = PROB_MAX; 2044 } 2045 } 2046 2047 // Clear out dead values from the debug info. 2048 kill_dead_locals(); 2049 2050 // Now insert the uncommon trap subroutine call 2051 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2052 const TypePtr* no_memory_effects = NULL; 2053 // Pass the index of the class to be loaded 2054 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2055 (must_throw ? RC_MUST_THROW : 0), 2056 OptoRuntime::uncommon_trap_Type(), 2057 call_addr, "uncommon_trap", no_memory_effects, 2058 intcon(trap_request)); 2059 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2060 "must extract request correctly from the graph"); 2061 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2062 2063 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2064 // The debug info is the only real input to this call. 2065 2066 // Halt-and-catch fire here. The above call should never return! 2067 HaltNode* halt = new HaltNode(control(), frameptr()); 2068 _gvn.set_type_bottom(halt); 2069 root()->add_req(halt); 2070 2071 stop_and_kill_map(); 2072 } 2073 2074 2075 //--------------------------just_allocated_object------------------------------ 2076 // Report the object that was just allocated. 2077 // It must be the case that there are no intervening safepoints. 2078 // We use this to determine if an object is so "fresh" that 2079 // it does not require card marks. 2080 Node* GraphKit::just_allocated_object(Node* current_control) { 2081 if (C->recent_alloc_ctl() == current_control) 2082 return C->recent_alloc_obj(); 2083 return NULL; 2084 } 2085 2086 2087 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2088 // (Note: TypeFunc::make has a cache that makes this fast.) 2089 const TypeFunc* tf = TypeFunc::make(dest_method); 2090 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2091 for (int j = 0; j < nargs; j++) { 2092 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2093 if( targ->basic_type() == T_DOUBLE ) { 2094 // If any parameters are doubles, they must be rounded before 2095 // the call, dstore_rounding does gvn.transform 2096 Node *arg = argument(j); 2097 arg = dstore_rounding(arg); 2098 set_argument(j, arg); 2099 } 2100 } 2101 } 2102 2103 /** 2104 * Record profiling data exact_kls for Node n with the type system so 2105 * that it can propagate it (speculation) 2106 * 2107 * @param n node that the type applies to 2108 * @param exact_kls type from profiling 2109 * @param maybe_null did profiling see null? 2110 * 2111 * @return node with improved type 2112 */ 2113 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2114 const Type* current_type = _gvn.type(n); 2115 assert(UseTypeSpeculation, "type speculation must be on"); 2116 2117 const TypePtr* speculative = current_type->speculative(); 2118 2119 // Should the klass from the profile be recorded in the speculative type? 2120 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2121 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2122 const TypeOopPtr* xtype = tklass->as_instance_type(); 2123 assert(xtype->klass_is_exact(), "Should be exact"); 2124 // Any reason to believe n is not null (from this profiling or a previous one)? 2125 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2126 // record the new speculative type's depth 2127 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2128 speculative = speculative->with_inline_depth(jvms()->depth()); 2129 } else if (current_type->would_improve_ptr(maybe_null)) { 2130 // Profiling report that null was never seen so we can change the 2131 // speculative type to non null ptr. 2132 assert(!maybe_null, "nothing to improve"); 2133 if (speculative == NULL) { 2134 speculative = TypePtr::NOTNULL; 2135 } else { 2136 const TypePtr* ptr = TypePtr::NOTNULL; 2137 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2138 } 2139 } 2140 2141 if (speculative != current_type->speculative()) { 2142 // Build a type with a speculative type (what we think we know 2143 // about the type but will need a guard when we use it) 2144 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2145 // We're changing the type, we need a new CheckCast node to carry 2146 // the new type. The new type depends on the control: what 2147 // profiling tells us is only valid from here as far as we can 2148 // tell. 2149 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2150 cast = _gvn.transform(cast); 2151 replace_in_map(n, cast); 2152 n = cast; 2153 } 2154 2155 return n; 2156 } 2157 2158 /** 2159 * Record profiling data from receiver profiling at an invoke with the 2160 * type system so that it can propagate it (speculation) 2161 * 2162 * @param n receiver node 2163 * 2164 * @return node with improved type 2165 */ 2166 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2167 if (!UseTypeSpeculation) { 2168 return n; 2169 } 2170 ciKlass* exact_kls = profile_has_unique_klass(); 2171 bool maybe_null = true; 2172 if (java_bc() == Bytecodes::_checkcast || 2173 java_bc() == Bytecodes::_instanceof || 2174 java_bc() == Bytecodes::_aastore) { 2175 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2176 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2177 } 2178 return record_profile_for_speculation(n, exact_kls, maybe_null); 2179 return n; 2180 } 2181 2182 /** 2183 * Record profiling data from argument profiling at an invoke with the 2184 * type system so that it can propagate it (speculation) 2185 * 2186 * @param dest_method target method for the call 2187 * @param bc what invoke bytecode is this? 2188 */ 2189 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2190 if (!UseTypeSpeculation) { 2191 return; 2192 } 2193 const TypeFunc* tf = TypeFunc::make(dest_method); 2194 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2195 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2196 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2197 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2198 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2199 bool maybe_null = true; 2200 ciKlass* better_type = NULL; 2201 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2202 record_profile_for_speculation(argument(j), better_type, maybe_null); 2203 } 2204 i++; 2205 } 2206 } 2207 } 2208 2209 /** 2210 * Record profiling data from parameter profiling at an invoke with 2211 * the type system so that it can propagate it (speculation) 2212 */ 2213 void GraphKit::record_profiled_parameters_for_speculation() { 2214 if (!UseTypeSpeculation) { 2215 return; 2216 } 2217 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2218 if (_gvn.type(local(i))->isa_oopptr()) { 2219 bool maybe_null = true; 2220 ciKlass* better_type = NULL; 2221 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2222 record_profile_for_speculation(local(i), better_type, maybe_null); 2223 } 2224 j++; 2225 } 2226 } 2227 } 2228 2229 /** 2230 * Record profiling data from return value profiling at an invoke with 2231 * the type system so that it can propagate it (speculation) 2232 */ 2233 void GraphKit::record_profiled_return_for_speculation() { 2234 if (!UseTypeSpeculation) { 2235 return; 2236 } 2237 bool maybe_null = true; 2238 ciKlass* better_type = NULL; 2239 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2240 // If profiling reports a single type for the return value, 2241 // feed it to the type system so it can propagate it as a 2242 // speculative type 2243 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2244 } 2245 } 2246 2247 void GraphKit::round_double_result(ciMethod* dest_method) { 2248 // A non-strict method may return a double value which has an extended 2249 // exponent, but this must not be visible in a caller which is 'strict' 2250 // If a strict caller invokes a non-strict callee, round a double result 2251 2252 BasicType result_type = dest_method->return_type()->basic_type(); 2253 assert( method() != NULL, "must have caller context"); 2254 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2255 // Destination method's return value is on top of stack 2256 // dstore_rounding() does gvn.transform 2257 Node *result = pop_pair(); 2258 result = dstore_rounding(result); 2259 push_pair(result); 2260 } 2261 } 2262 2263 // rounding for strict float precision conformance 2264 Node* GraphKit::precision_rounding(Node* n) { 2265 return UseStrictFP && _method->flags().is_strict() 2266 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2267 ? _gvn.transform( new RoundFloatNode(0, n) ) 2268 : n; 2269 } 2270 2271 // rounding for strict double precision conformance 2272 Node* GraphKit::dprecision_rounding(Node *n) { 2273 return UseStrictFP && _method->flags().is_strict() 2274 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2275 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2276 : n; 2277 } 2278 2279 // rounding for non-strict double stores 2280 Node* GraphKit::dstore_rounding(Node* n) { 2281 return Matcher::strict_fp_requires_explicit_rounding 2282 && UseSSE <= 1 2283 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2284 : n; 2285 } 2286 2287 //============================================================================= 2288 // Generate a fast path/slow path idiom. Graph looks like: 2289 // [foo] indicates that 'foo' is a parameter 2290 // 2291 // [in] NULL 2292 // \ / 2293 // CmpP 2294 // Bool ne 2295 // If 2296 // / \ 2297 // True False-<2> 2298 // / | 2299 // / cast_not_null 2300 // Load | | ^ 2301 // [fast_test] | | 2302 // gvn to opt_test | | 2303 // / \ | <1> 2304 // True False | 2305 // | \\ | 2306 // [slow_call] \[fast_result] 2307 // Ctl Val \ \ 2308 // | \ \ 2309 // Catch <1> \ \ 2310 // / \ ^ \ \ 2311 // Ex No_Ex | \ \ 2312 // | \ \ | \ <2> \ 2313 // ... \ [slow_res] | | \ [null_result] 2314 // \ \--+--+--- | | 2315 // \ | / \ | / 2316 // --------Region Phi 2317 // 2318 //============================================================================= 2319 // Code is structured as a series of driver functions all called 'do_XXX' that 2320 // call a set of helper functions. Helper functions first, then drivers. 2321 2322 //------------------------------null_check_oop--------------------------------- 2323 // Null check oop. Set null-path control into Region in slot 3. 2324 // Make a cast-not-nullness use the other not-null control. Return cast. 2325 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2326 bool never_see_null, 2327 bool safe_for_replace, 2328 bool speculative) { 2329 // Initial NULL check taken path 2330 (*null_control) = top(); 2331 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2332 2333 // Generate uncommon_trap: 2334 if (never_see_null && (*null_control) != top()) { 2335 // If we see an unexpected null at a check-cast we record it and force a 2336 // recompile; the offending check-cast will be compiled to handle NULLs. 2337 // If we see more than one offending BCI, then all checkcasts in the 2338 // method will be compiled to handle NULLs. 2339 PreserveJVMState pjvms(this); 2340 set_control(*null_control); 2341 replace_in_map(value, null()); 2342 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2343 uncommon_trap(reason, 2344 Deoptimization::Action_make_not_entrant); 2345 (*null_control) = top(); // NULL path is dead 2346 } 2347 if ((*null_control) == top() && safe_for_replace) { 2348 replace_in_map(value, cast); 2349 } 2350 2351 // Cast away null-ness on the result 2352 return cast; 2353 } 2354 2355 //------------------------------opt_iff---------------------------------------- 2356 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2357 // Return slow-path control. 2358 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2359 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2360 2361 // Fast path taken; set region slot 2 2362 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2363 region->init_req(2,fast_taken); // Capture fast-control 2364 2365 // Fast path not-taken, i.e. slow path 2366 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2367 return slow_taken; 2368 } 2369 2370 //-----------------------------make_runtime_call------------------------------- 2371 Node* GraphKit::make_runtime_call(int flags, 2372 const TypeFunc* call_type, address call_addr, 2373 const char* call_name, 2374 const TypePtr* adr_type, 2375 // The following parms are all optional. 2376 // The first NULL ends the list. 2377 Node* parm0, Node* parm1, 2378 Node* parm2, Node* parm3, 2379 Node* parm4, Node* parm5, 2380 Node* parm6, Node* parm7) { 2381 // Slow-path call 2382 bool is_leaf = !(flags & RC_NO_LEAF); 2383 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2384 if (call_name == NULL) { 2385 assert(!is_leaf, "must supply name for leaf"); 2386 call_name = OptoRuntime::stub_name(call_addr); 2387 } 2388 CallNode* call; 2389 if (!is_leaf) { 2390 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2391 bci(), adr_type); 2392 } else if (flags & RC_NO_FP) { 2393 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2394 } else { 2395 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2396 } 2397 2398 // The following is similar to set_edges_for_java_call, 2399 // except that the memory effects of the call are restricted to AliasIdxRaw. 2400 2401 // Slow path call has no side-effects, uses few values 2402 bool wide_in = !(flags & RC_NARROW_MEM); 2403 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2404 2405 Node* prev_mem = NULL; 2406 if (wide_in) { 2407 prev_mem = set_predefined_input_for_runtime_call(call); 2408 } else { 2409 assert(!wide_out, "narrow in => narrow out"); 2410 Node* narrow_mem = memory(adr_type); 2411 prev_mem = reset_memory(); 2412 map()->set_memory(narrow_mem); 2413 set_predefined_input_for_runtime_call(call); 2414 } 2415 2416 // Hook each parm in order. Stop looking at the first NULL. 2417 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2418 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2419 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2420 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2421 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2422 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2423 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2424 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2425 /* close each nested if ===> */ } } } } } } } } 2426 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2427 2428 if (!is_leaf) { 2429 // Non-leaves can block and take safepoints: 2430 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2431 } 2432 // Non-leaves can throw exceptions: 2433 if (has_io) { 2434 call->set_req(TypeFunc::I_O, i_o()); 2435 } 2436 2437 if (flags & RC_UNCOMMON) { 2438 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2439 // (An "if" probability corresponds roughly to an unconditional count. 2440 // Sort of.) 2441 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2442 } 2443 2444 Node* c = _gvn.transform(call); 2445 assert(c == call, "cannot disappear"); 2446 2447 if (wide_out) { 2448 // Slow path call has full side-effects. 2449 set_predefined_output_for_runtime_call(call); 2450 } else { 2451 // Slow path call has few side-effects, and/or sets few values. 2452 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2453 } 2454 2455 if (has_io) { 2456 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2457 } 2458 return call; 2459 2460 } 2461 2462 //------------------------------merge_memory----------------------------------- 2463 // Merge memory from one path into the current memory state. 2464 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2465 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2466 Node* old_slice = mms.force_memory(); 2467 Node* new_slice = mms.memory2(); 2468 if (old_slice != new_slice) { 2469 PhiNode* phi; 2470 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2471 if (mms.is_empty()) { 2472 // clone base memory Phi's inputs for this memory slice 2473 assert(old_slice == mms.base_memory(), "sanity"); 2474 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2475 _gvn.set_type(phi, Type::MEMORY); 2476 for (uint i = 1; i < phi->req(); i++) { 2477 phi->init_req(i, old_slice->in(i)); 2478 } 2479 } else { 2480 phi = old_slice->as_Phi(); // Phi was generated already 2481 } 2482 } else { 2483 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2484 _gvn.set_type(phi, Type::MEMORY); 2485 } 2486 phi->set_req(new_path, new_slice); 2487 mms.set_memory(phi); 2488 } 2489 } 2490 } 2491 2492 //------------------------------make_slow_call_ex------------------------------ 2493 // Make the exception handler hookups for the slow call 2494 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2495 if (stopped()) return; 2496 2497 // Make a catch node with just two handlers: fall-through and catch-all 2498 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2499 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2500 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2501 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2502 2503 { PreserveJVMState pjvms(this); 2504 set_control(excp); 2505 set_i_o(i_o); 2506 2507 if (excp != top()) { 2508 if (deoptimize) { 2509 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2510 uncommon_trap(Deoptimization::Reason_unhandled, 2511 Deoptimization::Action_none); 2512 } else { 2513 // Create an exception state also. 2514 // Use an exact type if the caller has specified a specific exception. 2515 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2516 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2517 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2518 } 2519 } 2520 } 2521 2522 // Get the no-exception control from the CatchNode. 2523 set_control(norm); 2524 } 2525 2526 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2527 Node* cmp = NULL; 2528 switch(bt) { 2529 case T_INT: cmp = new CmpINode(in1, in2); break; 2530 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2531 default: fatal("unexpected comparison type %s", type2name(bt)); 2532 } 2533 gvn->transform(cmp); 2534 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2535 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2536 gvn->transform(iff); 2537 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2538 return iff; 2539 } 2540 2541 2542 //-------------------------------gen_subtype_check----------------------------- 2543 // Generate a subtyping check. Takes as input the subtype and supertype. 2544 // Returns 2 values: sets the default control() to the true path and returns 2545 // the false path. Only reads invariant memory; sets no (visible) memory. 2546 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2547 // but that's not exposed to the optimizer. This call also doesn't take in an 2548 // Object; if you wish to check an Object you need to load the Object's class 2549 // prior to coming here. 2550 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2551 Compile* C = gvn->C; 2552 2553 if ((*ctrl)->is_top()) { 2554 return C->top(); 2555 } 2556 2557 // Fast check for identical types, perhaps identical constants. 2558 // The types can even be identical non-constants, in cases 2559 // involving Array.newInstance, Object.clone, etc. 2560 if (subklass == superklass) 2561 return C->top(); // false path is dead; no test needed. 2562 2563 if (gvn->type(superklass)->singleton()) { 2564 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2565 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2566 2567 // In the common case of an exact superklass, try to fold up the 2568 // test before generating code. You may ask, why not just generate 2569 // the code and then let it fold up? The answer is that the generated 2570 // code will necessarily include null checks, which do not always 2571 // completely fold away. If they are also needless, then they turn 2572 // into a performance loss. Example: 2573 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2574 // Here, the type of 'fa' is often exact, so the store check 2575 // of fa[1]=x will fold up, without testing the nullness of x. 2576 switch (C->static_subtype_check(superk, subk)) { 2577 case Compile::SSC_always_false: 2578 { 2579 Node* always_fail = *ctrl; 2580 *ctrl = gvn->C->top(); 2581 return always_fail; 2582 } 2583 case Compile::SSC_always_true: 2584 return C->top(); 2585 case Compile::SSC_easy_test: 2586 { 2587 // Just do a direct pointer compare and be done. 2588 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2589 *ctrl = gvn->transform(new IfTrueNode(iff)); 2590 return gvn->transform(new IfFalseNode(iff)); 2591 } 2592 case Compile::SSC_full_test: 2593 break; 2594 default: 2595 ShouldNotReachHere(); 2596 } 2597 } 2598 2599 // %%% Possible further optimization: Even if the superklass is not exact, 2600 // if the subklass is the unique subtype of the superklass, the check 2601 // will always succeed. We could leave a dependency behind to ensure this. 2602 2603 // First load the super-klass's check-offset 2604 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2605 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2606 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2607 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2608 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2609 2610 // Load from the sub-klass's super-class display list, or a 1-word cache of 2611 // the secondary superclass list, or a failing value with a sentinel offset 2612 // if the super-klass is an interface or exceptionally deep in the Java 2613 // hierarchy and we have to scan the secondary superclass list the hard way. 2614 // Worst-case type is a little odd: NULL is allowed as a result (usually 2615 // klass loads can never produce a NULL). 2616 Node *chk_off_X = chk_off; 2617 #ifdef _LP64 2618 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2619 #endif 2620 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2621 // For some types like interfaces the following loadKlass is from a 1-word 2622 // cache which is mutable so can't use immutable memory. Other 2623 // types load from the super-class display table which is immutable. 2624 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2625 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2626 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2627 2628 // Compile speed common case: ARE a subtype and we canNOT fail 2629 if( superklass == nkls ) 2630 return C->top(); // false path is dead; no test needed. 2631 2632 // See if we get an immediate positive hit. Happens roughly 83% of the 2633 // time. Test to see if the value loaded just previously from the subklass 2634 // is exactly the superklass. 2635 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2636 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2637 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2638 2639 // Compile speed common case: Check for being deterministic right now. If 2640 // chk_off is a constant and not equal to cacheoff then we are NOT a 2641 // subklass. In this case we need exactly the 1 test above and we can 2642 // return those results immediately. 2643 if (!might_be_cache) { 2644 Node* not_subtype_ctrl = *ctrl; 2645 *ctrl = iftrue1; // We need exactly the 1 test above 2646 return not_subtype_ctrl; 2647 } 2648 2649 // Gather the various success & failures here 2650 RegionNode *r_ok_subtype = new RegionNode(4); 2651 gvn->record_for_igvn(r_ok_subtype); 2652 RegionNode *r_not_subtype = new RegionNode(3); 2653 gvn->record_for_igvn(r_not_subtype); 2654 2655 r_ok_subtype->init_req(1, iftrue1); 2656 2657 // Check for immediate negative hit. Happens roughly 11% of the time (which 2658 // is roughly 63% of the remaining cases). Test to see if the loaded 2659 // check-offset points into the subklass display list or the 1-element 2660 // cache. If it points to the display (and NOT the cache) and the display 2661 // missed then it's not a subtype. 2662 Node *cacheoff = gvn->intcon(cacheoff_con); 2663 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2664 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2665 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2666 2667 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2668 // No performance impact (too rare) but allows sharing of secondary arrays 2669 // which has some footprint reduction. 2670 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2671 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2672 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2673 2674 // -- Roads not taken here: -- 2675 // We could also have chosen to perform the self-check at the beginning 2676 // of this code sequence, as the assembler does. This would not pay off 2677 // the same way, since the optimizer, unlike the assembler, can perform 2678 // static type analysis to fold away many successful self-checks. 2679 // Non-foldable self checks work better here in second position, because 2680 // the initial primary superclass check subsumes a self-check for most 2681 // types. An exception would be a secondary type like array-of-interface, 2682 // which does not appear in its own primary supertype display. 2683 // Finally, we could have chosen to move the self-check into the 2684 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2685 // dependent manner. But it is worthwhile to have the check here, 2686 // where it can be perhaps be optimized. The cost in code space is 2687 // small (register compare, branch). 2688 2689 // Now do a linear scan of the secondary super-klass array. Again, no real 2690 // performance impact (too rare) but it's gotta be done. 2691 // Since the code is rarely used, there is no penalty for moving it 2692 // out of line, and it can only improve I-cache density. 2693 // The decision to inline or out-of-line this final check is platform 2694 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2695 Node* psc = gvn->transform( 2696 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2697 2698 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2699 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2700 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2701 2702 // Return false path; set default control to true path. 2703 *ctrl = gvn->transform(r_ok_subtype); 2704 return gvn->transform(r_not_subtype); 2705 } 2706 2707 // Profile-driven exact type check: 2708 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2709 float prob, 2710 Node* *casted_receiver) { 2711 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2712 Node* recv_klass = load_object_klass(receiver); 2713 Node* want_klass = makecon(tklass); 2714 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2715 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2716 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2717 set_control( _gvn.transform( new IfTrueNode (iff) )); 2718 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2719 2720 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2721 assert(recv_xtype->klass_is_exact(), ""); 2722 2723 // Subsume downstream occurrences of receiver with a cast to 2724 // recv_xtype, since now we know what the type will be. 2725 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2726 (*casted_receiver) = _gvn.transform(cast); 2727 // (User must make the replace_in_map call.) 2728 2729 return fail; 2730 } 2731 2732 2733 //------------------------------seems_never_null------------------------------- 2734 // Use null_seen information if it is available from the profile. 2735 // If we see an unexpected null at a type check we record it and force a 2736 // recompile; the offending check will be recompiled to handle NULLs. 2737 // If we see several offending BCIs, then all checks in the 2738 // method will be recompiled. 2739 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2740 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2741 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2742 if (UncommonNullCast // Cutout for this technique 2743 && obj != null() // And not the -Xcomp stupid case? 2744 && !too_many_traps(reason) 2745 ) { 2746 if (speculating) { 2747 return true; 2748 } 2749 if (data == NULL) 2750 // Edge case: no mature data. Be optimistic here. 2751 return true; 2752 // If the profile has not seen a null, assume it won't happen. 2753 assert(java_bc() == Bytecodes::_checkcast || 2754 java_bc() == Bytecodes::_instanceof || 2755 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2756 return !data->as_BitData()->null_seen(); 2757 } 2758 speculating = false; 2759 return false; 2760 } 2761 2762 //------------------------maybe_cast_profiled_receiver------------------------- 2763 // If the profile has seen exactly one type, narrow to exactly that type. 2764 // Subsequent type checks will always fold up. 2765 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2766 ciKlass* require_klass, 2767 ciKlass* spec_klass, 2768 bool safe_for_replace) { 2769 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2770 2771 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2772 2773 // Make sure we haven't already deoptimized from this tactic. 2774 if (too_many_traps(reason) || too_many_recompiles(reason)) 2775 return NULL; 2776 2777 // (No, this isn't a call, but it's enough like a virtual call 2778 // to use the same ciMethod accessor to get the profile info...) 2779 // If we have a speculative type use it instead of profiling (which 2780 // may not help us) 2781 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2782 if (exact_kls != NULL) {// no cast failures here 2783 if (require_klass == NULL || 2784 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2785 // If we narrow the type to match what the type profile sees or 2786 // the speculative type, we can then remove the rest of the 2787 // cast. 2788 // This is a win, even if the exact_kls is very specific, 2789 // because downstream operations, such as method calls, 2790 // will often benefit from the sharper type. 2791 Node* exact_obj = not_null_obj; // will get updated in place... 2792 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2793 &exact_obj); 2794 { PreserveJVMState pjvms(this); 2795 set_control(slow_ctl); 2796 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2797 } 2798 if (safe_for_replace) { 2799 replace_in_map(not_null_obj, exact_obj); 2800 } 2801 return exact_obj; 2802 } 2803 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2804 } 2805 2806 return NULL; 2807 } 2808 2809 /** 2810 * Cast obj to type and emit guard unless we had too many traps here 2811 * already 2812 * 2813 * @param obj node being casted 2814 * @param type type to cast the node to 2815 * @param not_null true if we know node cannot be null 2816 */ 2817 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2818 ciKlass* type, 2819 bool not_null) { 2820 if (stopped()) { 2821 return obj; 2822 } 2823 2824 // type == NULL if profiling tells us this object is always null 2825 if (type != NULL) { 2826 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2827 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2828 2829 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2830 !too_many_traps(class_reason) && 2831 !too_many_recompiles(class_reason)) { 2832 Node* not_null_obj = NULL; 2833 // not_null is true if we know the object is not null and 2834 // there's no need for a null check 2835 if (!not_null) { 2836 Node* null_ctl = top(); 2837 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2838 assert(null_ctl->is_top(), "no null control here"); 2839 } else { 2840 not_null_obj = obj; 2841 } 2842 2843 Node* exact_obj = not_null_obj; 2844 ciKlass* exact_kls = type; 2845 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2846 &exact_obj); 2847 { 2848 PreserveJVMState pjvms(this); 2849 set_control(slow_ctl); 2850 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2851 } 2852 replace_in_map(not_null_obj, exact_obj); 2853 obj = exact_obj; 2854 } 2855 } else { 2856 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2857 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2858 Node* exact_obj = null_assert(obj); 2859 replace_in_map(obj, exact_obj); 2860 obj = exact_obj; 2861 } 2862 } 2863 return obj; 2864 } 2865 2866 //-------------------------------gen_instanceof-------------------------------- 2867 // Generate an instance-of idiom. Used by both the instance-of bytecode 2868 // and the reflective instance-of call. 2869 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2870 kill_dead_locals(); // Benefit all the uncommon traps 2871 assert( !stopped(), "dead parse path should be checked in callers" ); 2872 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2873 "must check for not-null not-dead klass in callers"); 2874 2875 // Make the merge point 2876 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2877 RegionNode* region = new RegionNode(PATH_LIMIT); 2878 Node* phi = new PhiNode(region, TypeInt::BOOL); 2879 C->set_has_split_ifs(true); // Has chance for split-if optimization 2880 2881 ciProfileData* data = NULL; 2882 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2883 data = method()->method_data()->bci_to_data(bci()); 2884 } 2885 bool speculative_not_null = false; 2886 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2887 && seems_never_null(obj, data, speculative_not_null)); 2888 2889 // Null check; get casted pointer; set region slot 3 2890 Node* null_ctl = top(); 2891 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2892 2893 // If not_null_obj is dead, only null-path is taken 2894 if (stopped()) { // Doing instance-of on a NULL? 2895 set_control(null_ctl); 2896 return intcon(0); 2897 } 2898 region->init_req(_null_path, null_ctl); 2899 phi ->init_req(_null_path, intcon(0)); // Set null path value 2900 if (null_ctl == top()) { 2901 // Do this eagerly, so that pattern matches like is_diamond_phi 2902 // will work even during parsing. 2903 assert(_null_path == PATH_LIMIT-1, "delete last"); 2904 region->del_req(_null_path); 2905 phi ->del_req(_null_path); 2906 } 2907 2908 // Do we know the type check always succeed? 2909 bool known_statically = false; 2910 if (_gvn.type(superklass)->singleton()) { 2911 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2912 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2913 if (subk != NULL && subk->is_loaded()) { 2914 int static_res = C->static_subtype_check(superk, subk); 2915 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2916 } 2917 } 2918 2919 if (known_statically && UseTypeSpeculation) { 2920 // If we know the type check always succeeds then we don't use the 2921 // profiling data at this bytecode. Don't lose it, feed it to the 2922 // type system as a speculative type. 2923 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2924 } else { 2925 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2926 // We may not have profiling here or it may not help us. If we 2927 // have a speculative type use it to perform an exact cast. 2928 ciKlass* spec_obj_type = obj_type->speculative_type(); 2929 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2930 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2931 if (stopped()) { // Profile disagrees with this path. 2932 set_control(null_ctl); // Null is the only remaining possibility. 2933 return intcon(0); 2934 } 2935 if (cast_obj != NULL) { 2936 not_null_obj = cast_obj; 2937 } 2938 } 2939 } 2940 2941 // Load the object's klass 2942 Node* obj_klass = load_object_klass(not_null_obj); 2943 2944 // Generate the subtype check 2945 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2946 2947 // Plug in the success path to the general merge in slot 1. 2948 region->init_req(_obj_path, control()); 2949 phi ->init_req(_obj_path, intcon(1)); 2950 2951 // Plug in the failing path to the general merge in slot 2. 2952 region->init_req(_fail_path, not_subtype_ctrl); 2953 phi ->init_req(_fail_path, intcon(0)); 2954 2955 // Return final merged results 2956 set_control( _gvn.transform(region) ); 2957 record_for_igvn(region); 2958 return _gvn.transform(phi); 2959 } 2960 2961 //-------------------------------gen_checkcast--------------------------------- 2962 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2963 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2964 // uncommon-trap paths work. Adjust stack after this call. 2965 // If failure_control is supplied and not null, it is filled in with 2966 // the control edge for the cast failure. Otherwise, an appropriate 2967 // uncommon trap or exception is thrown. 2968 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2969 Node* *failure_control) { 2970 kill_dead_locals(); // Benefit all the uncommon traps 2971 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2972 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2973 2974 // Fast cutout: Check the case that the cast is vacuously true. 2975 // This detects the common cases where the test will short-circuit 2976 // away completely. We do this before we perform the null check, 2977 // because if the test is going to turn into zero code, we don't 2978 // want a residual null check left around. (Causes a slowdown, 2979 // for example, in some objArray manipulations, such as a[i]=a[j].) 2980 if (tk->singleton()) { 2981 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2982 if (objtp != NULL && objtp->klass() != NULL) { 2983 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2984 case Compile::SSC_always_true: 2985 // If we know the type check always succeed then we don't use 2986 // the profiling data at this bytecode. Don't lose it, feed it 2987 // to the type system as a speculative type. 2988 return record_profiled_receiver_for_speculation(obj); 2989 case Compile::SSC_always_false: 2990 // It needs a null check because a null will *pass* the cast check. 2991 // A non-null value will always produce an exception. 2992 return null_assert(obj); 2993 } 2994 } 2995 } 2996 2997 ciProfileData* data = NULL; 2998 bool safe_for_replace = false; 2999 if (failure_control == NULL) { // use MDO in regular case only 3000 assert(java_bc() == Bytecodes::_aastore || 3001 java_bc() == Bytecodes::_checkcast, 3002 "interpreter profiles type checks only for these BCs"); 3003 data = method()->method_data()->bci_to_data(bci()); 3004 safe_for_replace = true; 3005 } 3006 3007 // Make the merge point 3008 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3009 RegionNode* region = new RegionNode(PATH_LIMIT); 3010 Node* phi = new PhiNode(region, toop); 3011 C->set_has_split_ifs(true); // Has chance for split-if optimization 3012 3013 // Use null-cast information if it is available 3014 bool speculative_not_null = false; 3015 bool never_see_null = ((failure_control == NULL) // regular case only 3016 && seems_never_null(obj, data, speculative_not_null)); 3017 3018 // Null check; get casted pointer; set region slot 3 3019 Node* null_ctl = top(); 3020 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3021 3022 // If not_null_obj is dead, only null-path is taken 3023 if (stopped()) { // Doing instance-of on a NULL? 3024 set_control(null_ctl); 3025 return null(); 3026 } 3027 region->init_req(_null_path, null_ctl); 3028 phi ->init_req(_null_path, null()); // Set null path value 3029 if (null_ctl == top()) { 3030 // Do this eagerly, so that pattern matches like is_diamond_phi 3031 // will work even during parsing. 3032 assert(_null_path == PATH_LIMIT-1, "delete last"); 3033 region->del_req(_null_path); 3034 phi ->del_req(_null_path); 3035 } 3036 3037 Node* cast_obj = NULL; 3038 if (tk->klass_is_exact()) { 3039 // The following optimization tries to statically cast the speculative type of the object 3040 // (for example obtained during profiling) to the type of the superklass and then do a 3041 // dynamic check that the type of the object is what we expect. To work correctly 3042 // for checkcast and aastore the type of superklass should be exact. 3043 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3044 // We may not have profiling here or it may not help us. If we have 3045 // a speculative type use it to perform an exact cast. 3046 ciKlass* spec_obj_type = obj_type->speculative_type(); 3047 if (spec_obj_type != NULL || data != NULL) { 3048 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3049 if (cast_obj != NULL) { 3050 if (failure_control != NULL) // failure is now impossible 3051 (*failure_control) = top(); 3052 // adjust the type of the phi to the exact klass: 3053 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3054 } 3055 } 3056 } 3057 3058 if (cast_obj == NULL) { 3059 // Load the object's klass 3060 Node* obj_klass = load_object_klass(not_null_obj); 3061 3062 // Generate the subtype check 3063 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3064 3065 // Plug in success path into the merge 3066 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3067 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3068 if (failure_control == NULL) { 3069 if (not_subtype_ctrl != top()) { // If failure is possible 3070 PreserveJVMState pjvms(this); 3071 set_control(not_subtype_ctrl); 3072 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3073 } 3074 } else { 3075 (*failure_control) = not_subtype_ctrl; 3076 } 3077 } 3078 3079 region->init_req(_obj_path, control()); 3080 phi ->init_req(_obj_path, cast_obj); 3081 3082 // A merge of NULL or Casted-NotNull obj 3083 Node* res = _gvn.transform(phi); 3084 3085 // Note I do NOT always 'replace_in_map(obj,result)' here. 3086 // if( tk->klass()->can_be_primary_super() ) 3087 // This means that if I successfully store an Object into an array-of-String 3088 // I 'forget' that the Object is really now known to be a String. I have to 3089 // do this because we don't have true union types for interfaces - if I store 3090 // a Baz into an array-of-Interface and then tell the optimizer it's an 3091 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3092 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3093 // replace_in_map( obj, res ); 3094 3095 // Return final merged results 3096 set_control( _gvn.transform(region) ); 3097 record_for_igvn(region); 3098 return res; 3099 } 3100 3101 //------------------------------next_monitor----------------------------------- 3102 // What number should be given to the next monitor? 3103 int GraphKit::next_monitor() { 3104 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3105 int next = current + C->sync_stack_slots(); 3106 // Keep the toplevel high water mark current: 3107 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3108 return current; 3109 } 3110 3111 //------------------------------insert_mem_bar--------------------------------- 3112 // Memory barrier to avoid floating things around 3113 // The membar serves as a pinch point between both control and all memory slices. 3114 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3115 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3116 mb->init_req(TypeFunc::Control, control()); 3117 mb->init_req(TypeFunc::Memory, reset_memory()); 3118 Node* membar = _gvn.transform(mb); 3119 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3120 set_all_memory_call(membar); 3121 return membar; 3122 } 3123 3124 //-------------------------insert_mem_bar_volatile---------------------------- 3125 // Memory barrier to avoid floating things around 3126 // The membar serves as a pinch point between both control and memory(alias_idx). 3127 // If you want to make a pinch point on all memory slices, do not use this 3128 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3129 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3130 // When Parse::do_put_xxx updates a volatile field, it appends a series 3131 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3132 // The first membar is on the same memory slice as the field store opcode. 3133 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3134 // All the other membars (for other volatile slices, including AliasIdxBot, 3135 // which stands for all unknown volatile slices) are control-dependent 3136 // on the first membar. This prevents later volatile loads or stores 3137 // from sliding up past the just-emitted store. 3138 3139 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3140 mb->set_req(TypeFunc::Control,control()); 3141 if (alias_idx == Compile::AliasIdxBot) { 3142 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3143 } else { 3144 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3145 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3146 } 3147 Node* membar = _gvn.transform(mb); 3148 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3149 if (alias_idx == Compile::AliasIdxBot) { 3150 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3151 } else { 3152 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3153 } 3154 return membar; 3155 } 3156 3157 void GraphKit::insert_store_load_for_barrier() { 3158 Node* mem = reset_memory(); 3159 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxBot); 3160 mb->init_req(TypeFunc::Control, control()); 3161 mb->init_req(TypeFunc::Memory, mem); 3162 Node* membar = _gvn.transform(mb); 3163 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3164 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory)); 3165 set_all_memory(mem); 3166 set_memory(newmem, Compile::AliasIdxRaw); 3167 } 3168 3169 3170 //------------------------------shared_lock------------------------------------ 3171 // Emit locking code. 3172 FastLockNode* GraphKit::shared_lock(Node* obj) { 3173 // bci is either a monitorenter bc or InvocationEntryBci 3174 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3175 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3176 3177 if( !GenerateSynchronizationCode ) 3178 return NULL; // Not locking things? 3179 if (stopped()) // Dead monitor? 3180 return NULL; 3181 3182 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3183 3184 // Box the stack location 3185 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3186 Node* mem = reset_memory(); 3187 3188 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3189 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3190 // Create the counters for this fast lock. 3191 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3192 } 3193 3194 // Create the rtm counters for this fast lock if needed. 3195 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3196 3197 // Add monitor to debug info for the slow path. If we block inside the 3198 // slow path and de-opt, we need the monitor hanging around 3199 map()->push_monitor( flock ); 3200 3201 const TypeFunc *tf = LockNode::lock_type(); 3202 LockNode *lock = new LockNode(C, tf); 3203 3204 lock->init_req( TypeFunc::Control, control() ); 3205 lock->init_req( TypeFunc::Memory , mem ); 3206 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3207 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3208 lock->init_req( TypeFunc::ReturnAdr, top() ); 3209 3210 lock->init_req(TypeFunc::Parms + 0, obj); 3211 lock->init_req(TypeFunc::Parms + 1, box); 3212 lock->init_req(TypeFunc::Parms + 2, flock); 3213 add_safepoint_edges(lock); 3214 3215 lock = _gvn.transform( lock )->as_Lock(); 3216 3217 // lock has no side-effects, sets few values 3218 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3219 3220 insert_mem_bar(Op_MemBarAcquireLock); 3221 3222 // Add this to the worklist so that the lock can be eliminated 3223 record_for_igvn(lock); 3224 3225 #ifndef PRODUCT 3226 if (PrintLockStatistics) { 3227 // Update the counter for this lock. Don't bother using an atomic 3228 // operation since we don't require absolute accuracy. 3229 lock->create_lock_counter(map()->jvms()); 3230 increment_counter(lock->counter()->addr()); 3231 } 3232 #endif 3233 3234 return flock; 3235 } 3236 3237 3238 //------------------------------shared_unlock---------------------------------- 3239 // Emit unlocking code. 3240 void GraphKit::shared_unlock(Node* box, Node* obj) { 3241 // bci is either a monitorenter bc or InvocationEntryBci 3242 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3243 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3244 3245 if( !GenerateSynchronizationCode ) 3246 return; 3247 if (stopped()) { // Dead monitor? 3248 map()->pop_monitor(); // Kill monitor from debug info 3249 return; 3250 } 3251 3252 // Memory barrier to avoid floating things down past the locked region 3253 insert_mem_bar(Op_MemBarReleaseLock); 3254 3255 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3256 UnlockNode *unlock = new UnlockNode(C, tf); 3257 #ifdef ASSERT 3258 unlock->set_dbg_jvms(sync_jvms()); 3259 #endif 3260 uint raw_idx = Compile::AliasIdxRaw; 3261 unlock->init_req( TypeFunc::Control, control() ); 3262 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3263 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3264 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3265 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3266 3267 unlock->init_req(TypeFunc::Parms + 0, obj); 3268 unlock->init_req(TypeFunc::Parms + 1, box); 3269 unlock = _gvn.transform(unlock)->as_Unlock(); 3270 3271 Node* mem = reset_memory(); 3272 3273 // unlock has no side-effects, sets few values 3274 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3275 3276 // Kill monitor from debug info 3277 map()->pop_monitor( ); 3278 } 3279 3280 //-------------------------------get_layout_helper----------------------------- 3281 // If the given klass is a constant or known to be an array, 3282 // fetch the constant layout helper value into constant_value 3283 // and return (Node*)NULL. Otherwise, load the non-constant 3284 // layout helper value, and return the node which represents it. 3285 // This two-faced routine is useful because allocation sites 3286 // almost always feature constant types. 3287 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3288 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3289 if (!StressReflectiveCode && inst_klass != NULL) { 3290 ciKlass* klass = inst_klass->klass(); 3291 bool xklass = inst_klass->klass_is_exact(); 3292 if (xklass || klass->is_array_klass()) { 3293 jint lhelper = klass->layout_helper(); 3294 if (lhelper != Klass::_lh_neutral_value) { 3295 constant_value = lhelper; 3296 return (Node*) NULL; 3297 } 3298 } 3299 } 3300 constant_value = Klass::_lh_neutral_value; // put in a known value 3301 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3302 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3303 } 3304 3305 // We just put in an allocate/initialize with a big raw-memory effect. 3306 // Hook selected additional alias categories on the initialization. 3307 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3308 MergeMemNode* init_in_merge, 3309 Node* init_out_raw) { 3310 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3311 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3312 3313 Node* prevmem = kit.memory(alias_idx); 3314 init_in_merge->set_memory_at(alias_idx, prevmem); 3315 kit.set_memory(init_out_raw, alias_idx); 3316 } 3317 3318 //---------------------------set_output_for_allocation------------------------- 3319 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3320 const TypeOopPtr* oop_type, 3321 bool deoptimize_on_exception) { 3322 int rawidx = Compile::AliasIdxRaw; 3323 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3324 add_safepoint_edges(alloc); 3325 Node* allocx = _gvn.transform(alloc); 3326 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3327 // create memory projection for i_o 3328 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3329 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3330 3331 // create a memory projection as for the normal control path 3332 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3333 set_memory(malloc, rawidx); 3334 3335 // a normal slow-call doesn't change i_o, but an allocation does 3336 // we create a separate i_o projection for the normal control path 3337 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3338 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3339 3340 // put in an initialization barrier 3341 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3342 rawoop)->as_Initialize(); 3343 assert(alloc->initialization() == init, "2-way macro link must work"); 3344 assert(init ->allocation() == alloc, "2-way macro link must work"); 3345 { 3346 // Extract memory strands which may participate in the new object's 3347 // initialization, and source them from the new InitializeNode. 3348 // This will allow us to observe initializations when they occur, 3349 // and link them properly (as a group) to the InitializeNode. 3350 assert(init->in(InitializeNode::Memory) == malloc, ""); 3351 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3352 init->set_req(InitializeNode::Memory, minit_in); 3353 record_for_igvn(minit_in); // fold it up later, if possible 3354 Node* minit_out = memory(rawidx); 3355 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3356 if (oop_type->isa_aryptr()) { 3357 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3358 int elemidx = C->get_alias_index(telemref); 3359 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3360 } else if (oop_type->isa_instptr()) { 3361 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3362 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3363 ciField* field = ik->nonstatic_field_at(i); 3364 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3365 continue; // do not bother to track really large numbers of fields 3366 // Find (or create) the alias category for this field: 3367 int fieldidx = C->alias_type(field)->index(); 3368 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3369 } 3370 } 3371 } 3372 3373 // Cast raw oop to the real thing... 3374 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3375 javaoop = _gvn.transform(javaoop); 3376 C->set_recent_alloc(control(), javaoop); 3377 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3378 3379 #ifdef ASSERT 3380 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3381 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3382 "Ideal_allocation works"); 3383 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3384 "Ideal_allocation works"); 3385 if (alloc->is_AllocateArray()) { 3386 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3387 "Ideal_allocation works"); 3388 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3389 "Ideal_allocation works"); 3390 } else { 3391 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3392 } 3393 } 3394 #endif //ASSERT 3395 3396 return javaoop; 3397 } 3398 3399 //---------------------------new_instance-------------------------------------- 3400 // This routine takes a klass_node which may be constant (for a static type) 3401 // or may be non-constant (for reflective code). It will work equally well 3402 // for either, and the graph will fold nicely if the optimizer later reduces 3403 // the type to a constant. 3404 // The optional arguments are for specialized use by intrinsics: 3405 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3406 // - If 'return_size_val', report the the total object size to the caller. 3407 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3408 Node* GraphKit::new_instance(Node* klass_node, 3409 Node* extra_slow_test, 3410 Node* *return_size_val, 3411 bool deoptimize_on_exception) { 3412 // Compute size in doublewords 3413 // The size is always an integral number of doublewords, represented 3414 // as a positive bytewise size stored in the klass's layout_helper. 3415 // The layout_helper also encodes (in a low bit) the need for a slow path. 3416 jint layout_con = Klass::_lh_neutral_value; 3417 Node* layout_val = get_layout_helper(klass_node, layout_con); 3418 int layout_is_con = (layout_val == NULL); 3419 3420 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3421 // Generate the initial go-slow test. It's either ALWAYS (return a 3422 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3423 // case) a computed value derived from the layout_helper. 3424 Node* initial_slow_test = NULL; 3425 if (layout_is_con) { 3426 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3427 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3428 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3429 } else { // reflective case 3430 // This reflective path is used by Unsafe.allocateInstance. 3431 // (It may be stress-tested by specifying StressReflectiveCode.) 3432 // Basically, we want to get into the VM is there's an illegal argument. 3433 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3434 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3435 if (extra_slow_test != intcon(0)) { 3436 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3437 } 3438 // (Macro-expander will further convert this to a Bool, if necessary.) 3439 } 3440 3441 // Find the size in bytes. This is easy; it's the layout_helper. 3442 // The size value must be valid even if the slow path is taken. 3443 Node* size = NULL; 3444 if (layout_is_con) { 3445 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3446 } else { // reflective case 3447 // This reflective path is used by clone and Unsafe.allocateInstance. 3448 size = ConvI2X(layout_val); 3449 3450 // Clear the low bits to extract layout_helper_size_in_bytes: 3451 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3452 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3453 size = _gvn.transform( new AndXNode(size, mask) ); 3454 } 3455 if (return_size_val != NULL) { 3456 (*return_size_val) = size; 3457 } 3458 3459 // This is a precise notnull oop of the klass. 3460 // (Actually, it need not be precise if this is a reflective allocation.) 3461 // It's what we cast the result to. 3462 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3463 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3464 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3465 3466 // Now generate allocation code 3467 3468 // The entire memory state is needed for slow path of the allocation 3469 // since GC and deoptimization can happened. 3470 Node *mem = reset_memory(); 3471 set_all_memory(mem); // Create new memory state 3472 3473 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3474 control(), mem, i_o(), 3475 size, klass_node, 3476 initial_slow_test); 3477 3478 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3479 } 3480 3481 //-------------------------------new_array------------------------------------- 3482 // helper for both newarray and anewarray 3483 // The 'length' parameter is (obviously) the length of the array. 3484 // See comments on new_instance for the meaning of the other arguments. 3485 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3486 Node* length, // number of array elements 3487 int nargs, // number of arguments to push back for uncommon trap 3488 Node* *return_size_val, 3489 bool deoptimize_on_exception) { 3490 jint layout_con = Klass::_lh_neutral_value; 3491 Node* layout_val = get_layout_helper(klass_node, layout_con); 3492 int layout_is_con = (layout_val == NULL); 3493 3494 if (!layout_is_con && !StressReflectiveCode && 3495 !too_many_traps(Deoptimization::Reason_class_check)) { 3496 // This is a reflective array creation site. 3497 // Optimistically assume that it is a subtype of Object[], 3498 // so that we can fold up all the address arithmetic. 3499 layout_con = Klass::array_layout_helper(T_OBJECT); 3500 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3501 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3502 { BuildCutout unless(this, bol_lh, PROB_MAX); 3503 inc_sp(nargs); 3504 uncommon_trap(Deoptimization::Reason_class_check, 3505 Deoptimization::Action_maybe_recompile); 3506 } 3507 layout_val = NULL; 3508 layout_is_con = true; 3509 } 3510 3511 // Generate the initial go-slow test. Make sure we do not overflow 3512 // if length is huge (near 2Gig) or negative! We do not need 3513 // exact double-words here, just a close approximation of needed 3514 // double-words. We can't add any offset or rounding bits, lest we 3515 // take a size -1 of bytes and make it positive. Use an unsigned 3516 // compare, so negative sizes look hugely positive. 3517 int fast_size_limit = FastAllocateSizeLimit; 3518 if (layout_is_con) { 3519 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3520 // Increase the size limit if we have exact knowledge of array type. 3521 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3522 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3523 } 3524 3525 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3526 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3527 3528 // --- Size Computation --- 3529 // array_size = round_to_heap(array_header + (length << elem_shift)); 3530 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3531 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3532 // The rounding mask is strength-reduced, if possible. 3533 int round_mask = MinObjAlignmentInBytes - 1; 3534 Node* header_size = NULL; 3535 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3536 // (T_BYTE has the weakest alignment and size restrictions...) 3537 if (layout_is_con) { 3538 int hsize = Klass::layout_helper_header_size(layout_con); 3539 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3540 BasicType etype = Klass::layout_helper_element_type(layout_con); 3541 if ((round_mask & ~right_n_bits(eshift)) == 0) 3542 round_mask = 0; // strength-reduce it if it goes away completely 3543 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3544 assert(header_size_min <= hsize, "generic minimum is smallest"); 3545 header_size_min = hsize; 3546 header_size = intcon(hsize + round_mask); 3547 } else { 3548 Node* hss = intcon(Klass::_lh_header_size_shift); 3549 Node* hsm = intcon(Klass::_lh_header_size_mask); 3550 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3551 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3552 Node* mask = intcon(round_mask); 3553 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3554 } 3555 3556 Node* elem_shift = NULL; 3557 if (layout_is_con) { 3558 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3559 if (eshift != 0) 3560 elem_shift = intcon(eshift); 3561 } else { 3562 // There is no need to mask or shift this value. 3563 // The semantics of LShiftINode include an implicit mask to 0x1F. 3564 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3565 elem_shift = layout_val; 3566 } 3567 3568 // Transition to native address size for all offset calculations: 3569 Node* lengthx = ConvI2X(length); 3570 Node* headerx = ConvI2X(header_size); 3571 #ifdef _LP64 3572 { const TypeInt* tilen = _gvn.find_int_type(length); 3573 if (tilen != NULL && tilen->_lo < 0) { 3574 // Add a manual constraint to a positive range. Cf. array_element_address. 3575 jint size_max = fast_size_limit; 3576 if (size_max > tilen->_hi) size_max = tilen->_hi; 3577 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3578 3579 // Only do a narrow I2L conversion if the range check passed. 3580 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3581 _gvn.transform(iff); 3582 RegionNode* region = new RegionNode(3); 3583 _gvn.set_type(region, Type::CONTROL); 3584 lengthx = new PhiNode(region, TypeLong::LONG); 3585 _gvn.set_type(lengthx, TypeLong::LONG); 3586 3587 // Range check passed. Use ConvI2L node with narrow type. 3588 Node* passed = IfFalse(iff); 3589 region->init_req(1, passed); 3590 // Make I2L conversion control dependent to prevent it from 3591 // floating above the range check during loop optimizations. 3592 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3593 3594 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3595 region->init_req(2, IfTrue(iff)); 3596 lengthx->init_req(2, ConvI2X(length)); 3597 3598 set_control(region); 3599 record_for_igvn(region); 3600 record_for_igvn(lengthx); 3601 } 3602 } 3603 #endif 3604 3605 // Combine header size (plus rounding) and body size. Then round down. 3606 // This computation cannot overflow, because it is used only in two 3607 // places, one where the length is sharply limited, and the other 3608 // after a successful allocation. 3609 Node* abody = lengthx; 3610 if (elem_shift != NULL) 3611 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3612 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3613 if (round_mask != 0) { 3614 Node* mask = MakeConX(~round_mask); 3615 size = _gvn.transform( new AndXNode(size, mask) ); 3616 } 3617 // else if round_mask == 0, the size computation is self-rounding 3618 3619 if (return_size_val != NULL) { 3620 // This is the size 3621 (*return_size_val) = size; 3622 } 3623 3624 // Now generate allocation code 3625 3626 // The entire memory state is needed for slow path of the allocation 3627 // since GC and deoptimization can happened. 3628 Node *mem = reset_memory(); 3629 set_all_memory(mem); // Create new memory state 3630 3631 if (initial_slow_test->is_Bool()) { 3632 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3633 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3634 } 3635 3636 // Create the AllocateArrayNode and its result projections 3637 AllocateArrayNode* alloc 3638 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3639 control(), mem, i_o(), 3640 size, klass_node, 3641 initial_slow_test, 3642 length); 3643 3644 // Cast to correct type. Note that the klass_node may be constant or not, 3645 // and in the latter case the actual array type will be inexact also. 3646 // (This happens via a non-constant argument to inline_native_newArray.) 3647 // In any case, the value of klass_node provides the desired array type. 3648 const TypeInt* length_type = _gvn.find_int_type(length); 3649 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3650 if (ary_type->isa_aryptr() && length_type != NULL) { 3651 // Try to get a better type than POS for the size 3652 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3653 } 3654 3655 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3656 3657 // Cast length on remaining path to be as narrow as possible 3658 if (map()->find_edge(length) >= 0) { 3659 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3660 if (ccast != length) { 3661 _gvn.set_type_bottom(ccast); 3662 record_for_igvn(ccast); 3663 replace_in_map(length, ccast); 3664 } 3665 } 3666 3667 return javaoop; 3668 } 3669 3670 // The following "Ideal_foo" functions are placed here because they recognize 3671 // the graph shapes created by the functions immediately above. 3672 3673 //---------------------------Ideal_allocation---------------------------------- 3674 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3675 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3676 if (ptr == NULL) { // reduce dumb test in callers 3677 return NULL; 3678 } 3679 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3680 ptr = ptr->in(1); 3681 if (ptr == NULL) return NULL; 3682 } 3683 // Return NULL for allocations with several casts: 3684 // j.l.reflect.Array.newInstance(jobject, jint) 3685 // Object.clone() 3686 // to keep more precise type from last cast. 3687 if (ptr->is_Proj()) { 3688 Node* allo = ptr->in(0); 3689 if (allo != NULL && allo->is_Allocate()) { 3690 return allo->as_Allocate(); 3691 } 3692 } 3693 // Report failure to match. 3694 return NULL; 3695 } 3696 3697 // Fancy version which also strips off an offset (and reports it to caller). 3698 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3699 intptr_t& offset) { 3700 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3701 if (base == NULL) return NULL; 3702 return Ideal_allocation(base, phase); 3703 } 3704 3705 // Trace Initialize <- Proj[Parm] <- Allocate 3706 AllocateNode* InitializeNode::allocation() { 3707 Node* rawoop = in(InitializeNode::RawAddress); 3708 if (rawoop->is_Proj()) { 3709 Node* alloc = rawoop->in(0); 3710 if (alloc->is_Allocate()) { 3711 return alloc->as_Allocate(); 3712 } 3713 } 3714 return NULL; 3715 } 3716 3717 // Trace Allocate -> Proj[Parm] -> Initialize 3718 InitializeNode* AllocateNode::initialization() { 3719 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3720 if (rawoop == NULL) return NULL; 3721 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3722 Node* init = rawoop->fast_out(i); 3723 if (init->is_Initialize()) { 3724 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3725 return init->as_Initialize(); 3726 } 3727 } 3728 return NULL; 3729 } 3730 3731 //----------------------------- loop predicates --------------------------- 3732 3733 //------------------------------add_predicate_impl---------------------------- 3734 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3735 // Too many traps seen? 3736 if (too_many_traps(reason)) { 3737 #ifdef ASSERT 3738 if (TraceLoopPredicate) { 3739 int tc = C->trap_count(reason); 3740 tty->print("too many traps=%s tcount=%d in ", 3741 Deoptimization::trap_reason_name(reason), tc); 3742 method()->print(); // which method has too many predicate traps 3743 tty->cr(); 3744 } 3745 #endif 3746 // We cannot afford to take more traps here, 3747 // do not generate predicate. 3748 return; 3749 } 3750 3751 Node *cont = _gvn.intcon(1); 3752 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3753 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3754 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3755 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3756 C->add_predicate_opaq(opq); 3757 { 3758 PreserveJVMState pjvms(this); 3759 set_control(iffalse); 3760 inc_sp(nargs); 3761 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3762 } 3763 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3764 set_control(iftrue); 3765 } 3766 3767 //------------------------------add_predicate--------------------------------- 3768 void GraphKit::add_predicate(int nargs) { 3769 if (UseLoopPredicate) { 3770 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3771 } 3772 // loop's limit check predicate should be near the loop. 3773 if (LoopLimitCheck) { 3774 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3775 } 3776 } 3777 3778 //----------------------------- store barriers ---------------------------- 3779 #define __ ideal. 3780 3781 void GraphKit::sync_kit(IdealKit& ideal) { 3782 set_all_memory(__ merged_memory()); 3783 set_i_o(__ i_o()); 3784 set_control(__ ctrl()); 3785 } 3786 3787 void GraphKit::final_sync(IdealKit& ideal) { 3788 // Final sync IdealKit and graphKit. 3789 sync_kit(ideal); 3790 } 3791 3792 Node* GraphKit::byte_map_base_node() { 3793 // Get base of card map 3794 CardTableModRefBS* ct = 3795 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3796 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3797 if (ct->byte_map_base != NULL) { 3798 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3799 } else { 3800 return null(); 3801 } 3802 } 3803 3804 // vanilla/CMS post barrier 3805 // Insert a write-barrier store. This is to let generational GC work; we have 3806 // to flag all oop-stores before the next GC point. 3807 void GraphKit::write_barrier_post(Node* oop_store, 3808 Node* obj, 3809 Node* adr, 3810 uint adr_idx, 3811 Node* val, 3812 bool use_precise) { 3813 // No store check needed if we're storing a NULL or an old object 3814 // (latter case is probably a string constant). The concurrent 3815 // mark sweep garbage collector, however, needs to have all nonNull 3816 // oop updates flagged via card-marks. 3817 if (val != NULL && val->is_Con()) { 3818 // must be either an oop or NULL 3819 const Type* t = val->bottom_type(); 3820 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3821 // stores of null never (?) need barriers 3822 return; 3823 } 3824 3825 if (use_ReduceInitialCardMarks() 3826 && obj == just_allocated_object(control())) { 3827 // We can skip marks on a freshly-allocated object in Eden. 3828 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3829 // That routine informs GC to take appropriate compensating steps, 3830 // upon a slow-path allocation, so as to make this card-mark 3831 // elision safe. 3832 return; 3833 } 3834 3835 if (!use_precise) { 3836 // All card marks for a (non-array) instance are in one place: 3837 adr = obj; 3838 } 3839 // (Else it's an array (or unknown), and we want more precise card marks.) 3840 assert(adr != NULL, ""); 3841 3842 IdealKit ideal(this, true); 3843 3844 // Convert the pointer to an int prior to doing math on it 3845 Node* cast = __ CastPX(__ ctrl(), adr); 3846 3847 // Divide by card size 3848 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3849 "Only one we handle so far."); 3850 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3851 3852 // Combine card table base and card offset 3853 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3854 3855 // Get the alias_index for raw card-mark memory 3856 int adr_type = Compile::AliasIdxRaw; 3857 Node* zero = __ ConI(0); // Dirty card value 3858 BasicType bt = T_BYTE; 3859 3860 if (UseConcMarkSweepGC && UseCondCardMark) { 3861 insert_store_load_for_barrier(); 3862 __ sync_kit(this); 3863 } 3864 3865 if (UseCondCardMark) { 3866 // The classic GC reference write barrier is typically implemented 3867 // as a store into the global card mark table. Unfortunately 3868 // unconditional stores can result in false sharing and excessive 3869 // coherence traffic as well as false transactional aborts. 3870 // UseCondCardMark enables MP "polite" conditional card mark 3871 // stores. In theory we could relax the load from ctrl() to 3872 // no_ctrl, but that doesn't buy much latitude. 3873 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3874 __ if_then(card_val, BoolTest::ne, zero); 3875 } 3876 3877 // Smash zero into card 3878 if( !UseConcMarkSweepGC ) { 3879 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3880 } else { 3881 // Specialized path for CM store barrier 3882 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3883 } 3884 3885 if (UseCondCardMark) { 3886 __ end_if(); 3887 } 3888 3889 // Final sync IdealKit and GraphKit. 3890 final_sync(ideal); 3891 } 3892 /* 3893 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3894 * required by SATB to make sure all objects live at the start of the 3895 * marking are kept alive, all reference updates need to any previous 3896 * reference stored before writing. 3897 * 3898 * If the previous value is NULL there is no need to save the old value. 3899 * References that are NULL are filtered during runtime by the barrier 3900 * code to avoid unnecessary queuing. 3901 * 3902 * However in the case of newly allocated objects it might be possible to 3903 * prove that the reference about to be overwritten is NULL during compile 3904 * time and avoid adding the barrier code completely. 3905 * 3906 * The compiler needs to determine that the object in which a field is about 3907 * to be written is newly allocated, and that no prior store to the same field 3908 * has happened since the allocation. 3909 * 3910 * Returns true if the pre-barrier can be removed 3911 */ 3912 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3913 BasicType bt, uint adr_idx) { 3914 intptr_t offset = 0; 3915 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3916 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3917 3918 if (offset == Type::OffsetBot) { 3919 return false; // cannot unalias unless there are precise offsets 3920 } 3921 3922 if (alloc == NULL) { 3923 return false; // No allocation found 3924 } 3925 3926 intptr_t size_in_bytes = type2aelembytes(bt); 3927 3928 Node* mem = memory(adr_idx); // start searching here... 3929 3930 for (int cnt = 0; cnt < 50; cnt++) { 3931 3932 if (mem->is_Store()) { 3933 3934 Node* st_adr = mem->in(MemNode::Address); 3935 intptr_t st_offset = 0; 3936 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3937 3938 if (st_base == NULL) { 3939 break; // inscrutable pointer 3940 } 3941 3942 // Break we have found a store with same base and offset as ours so break 3943 if (st_base == base && st_offset == offset) { 3944 break; 3945 } 3946 3947 if (st_offset != offset && st_offset != Type::OffsetBot) { 3948 const int MAX_STORE = BytesPerLong; 3949 if (st_offset >= offset + size_in_bytes || 3950 st_offset <= offset - MAX_STORE || 3951 st_offset <= offset - mem->as_Store()->memory_size()) { 3952 // Success: The offsets are provably independent. 3953 // (You may ask, why not just test st_offset != offset and be done? 3954 // The answer is that stores of different sizes can co-exist 3955 // in the same sequence of RawMem effects. We sometimes initialize 3956 // a whole 'tile' of array elements with a single jint or jlong.) 3957 mem = mem->in(MemNode::Memory); 3958 continue; // advance through independent store memory 3959 } 3960 } 3961 3962 if (st_base != base 3963 && MemNode::detect_ptr_independence(base, alloc, st_base, 3964 AllocateNode::Ideal_allocation(st_base, phase), 3965 phase)) { 3966 // Success: The bases are provably independent. 3967 mem = mem->in(MemNode::Memory); 3968 continue; // advance through independent store memory 3969 } 3970 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3971 3972 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3973 AllocateNode* st_alloc = st_init->allocation(); 3974 3975 // Make sure that we are looking at the same allocation site. 3976 // The alloc variable is guaranteed to not be null here from earlier check. 3977 if (alloc == st_alloc) { 3978 // Check that the initialization is storing NULL so that no previous store 3979 // has been moved up and directly write a reference 3980 Node* captured_store = st_init->find_captured_store(offset, 3981 type2aelembytes(T_OBJECT), 3982 phase); 3983 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3984 return true; 3985 } 3986 } 3987 } 3988 3989 // Unless there is an explicit 'continue', we must bail out here, 3990 // because 'mem' is an inscrutable memory state (e.g., a call). 3991 break; 3992 } 3993 3994 return false; 3995 } 3996 3997 // G1 pre/post barriers 3998 void GraphKit::g1_write_barrier_pre(bool do_load, 3999 Node* obj, 4000 Node* adr, 4001 uint alias_idx, 4002 Node* val, 4003 const TypeOopPtr* val_type, 4004 Node* pre_val, 4005 BasicType bt) { 4006 4007 // Some sanity checks 4008 // Note: val is unused in this routine. 4009 4010 if (do_load) { 4011 // We need to generate the load of the previous value 4012 assert(obj != NULL, "must have a base"); 4013 assert(adr != NULL, "where are loading from?"); 4014 assert(pre_val == NULL, "loaded already?"); 4015 assert(val_type != NULL, "need a type"); 4016 4017 if (use_ReduceInitialCardMarks() 4018 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4019 return; 4020 } 4021 4022 } else { 4023 // In this case both val_type and alias_idx are unused. 4024 assert(pre_val != NULL, "must be loaded already"); 4025 // Nothing to be done if pre_val is null. 4026 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4027 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4028 } 4029 assert(bt == T_OBJECT, "or we shouldn't be here"); 4030 4031 IdealKit ideal(this, true); 4032 4033 Node* tls = __ thread(); // ThreadLocalStorage 4034 4035 Node* no_ctrl = NULL; 4036 Node* no_base = __ top(); 4037 Node* zero = __ ConI(0); 4038 Node* zeroX = __ ConX(0); 4039 4040 float likely = PROB_LIKELY(0.999); 4041 float unlikely = PROB_UNLIKELY(0.999); 4042 4043 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4044 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4045 4046 // Offsets into the thread 4047 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4048 SATBMarkQueue::byte_offset_of_active()); 4049 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4050 SATBMarkQueue::byte_offset_of_index()); 4051 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4052 SATBMarkQueue::byte_offset_of_buf()); 4053 4054 // Now the actual pointers into the thread 4055 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4056 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4057 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4058 4059 // Now some of the values 4060 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4061 4062 // if (!marking) 4063 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4064 BasicType index_bt = TypeX_X->basic_type(); 4065 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4066 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4067 4068 if (do_load) { 4069 // load original value 4070 // alias_idx correct?? 4071 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4072 } 4073 4074 // if (pre_val != NULL) 4075 __ if_then(pre_val, BoolTest::ne, null()); { 4076 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4077 4078 // is the queue for this thread full? 4079 __ if_then(index, BoolTest::ne, zeroX, likely); { 4080 4081 // decrement the index 4082 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4083 4084 // Now get the buffer location we will log the previous value into and store it 4085 Node *log_addr = __ AddP(no_base, buffer, next_index); 4086 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4087 // update the index 4088 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4089 4090 } __ else_(); { 4091 4092 // logging buffer is full, call the runtime 4093 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4094 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4095 } __ end_if(); // (!index) 4096 } __ end_if(); // (pre_val != NULL) 4097 } __ end_if(); // (!marking) 4098 4099 // Final sync IdealKit and GraphKit. 4100 final_sync(ideal); 4101 } 4102 4103 /* 4104 * G1 similar to any GC with a Young Generation requires a way to keep track of 4105 * references from Old Generation to Young Generation to make sure all live 4106 * objects are found. G1 also requires to keep track of object references 4107 * between different regions to enable evacuation of old regions, which is done 4108 * as part of mixed collections. References are tracked in remembered sets and 4109 * is continuously updated as reference are written to with the help of the 4110 * post-barrier. 4111 * 4112 * To reduce the number of updates to the remembered set the post-barrier 4113 * filters updates to fields in objects located in the Young Generation, 4114 * the same region as the reference, when the NULL is being written or 4115 * if the card is already marked as dirty by an earlier write. 4116 * 4117 * Under certain circumstances it is possible to avoid generating the 4118 * post-barrier completely if it is possible during compile time to prove 4119 * the object is newly allocated and that no safepoint exists between the 4120 * allocation and the store. 4121 * 4122 * In the case of slow allocation the allocation code must handle the barrier 4123 * as part of the allocation in the case the allocated object is not located 4124 * in the nursery, this would happen for humongous objects. This is similar to 4125 * how CMS is required to handle this case, see the comments for the method 4126 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4127 * A deferred card mark is required for these objects and handled in the above 4128 * mentioned methods. 4129 * 4130 * Returns true if the post barrier can be removed 4131 */ 4132 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4133 Node* adr) { 4134 intptr_t offset = 0; 4135 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4136 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4137 4138 if (offset == Type::OffsetBot) { 4139 return false; // cannot unalias unless there are precise offsets 4140 } 4141 4142 if (alloc == NULL) { 4143 return false; // No allocation found 4144 } 4145 4146 // Start search from Store node 4147 Node* mem = store->in(MemNode::Control); 4148 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4149 4150 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4151 AllocateNode* st_alloc = st_init->allocation(); 4152 4153 // Make sure we are looking at the same allocation 4154 if (alloc == st_alloc) { 4155 return true; 4156 } 4157 } 4158 4159 return false; 4160 } 4161 4162 // 4163 // Update the card table and add card address to the queue 4164 // 4165 void GraphKit::g1_mark_card(IdealKit& ideal, 4166 Node* card_adr, 4167 Node* oop_store, 4168 uint oop_alias_idx, 4169 Node* index, 4170 Node* index_adr, 4171 Node* buffer, 4172 const TypeFunc* tf) { 4173 4174 Node* zero = __ ConI(0); 4175 Node* zeroX = __ ConX(0); 4176 Node* no_base = __ top(); 4177 BasicType card_bt = T_BYTE; 4178 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4179 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4180 4181 // Now do the queue work 4182 __ if_then(index, BoolTest::ne, zeroX); { 4183 4184 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4185 Node* log_addr = __ AddP(no_base, buffer, next_index); 4186 4187 // Order, see storeCM. 4188 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4189 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4190 4191 } __ else_(); { 4192 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4193 } __ end_if(); 4194 4195 } 4196 4197 void GraphKit::g1_write_barrier_post(Node* oop_store, 4198 Node* obj, 4199 Node* adr, 4200 uint alias_idx, 4201 Node* val, 4202 BasicType bt, 4203 bool use_precise) { 4204 // If we are writing a NULL then we need no post barrier 4205 4206 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4207 // Must be NULL 4208 const Type* t = val->bottom_type(); 4209 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4210 // No post barrier if writing NULLx 4211 return; 4212 } 4213 4214 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4215 // We can skip marks on a freshly-allocated object in Eden. 4216 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4217 // That routine informs GC to take appropriate compensating steps, 4218 // upon a slow-path allocation, so as to make this card-mark 4219 // elision safe. 4220 return; 4221 } 4222 4223 if (use_ReduceInitialCardMarks() 4224 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4225 return; 4226 } 4227 4228 if (!use_precise) { 4229 // All card marks for a (non-array) instance are in one place: 4230 adr = obj; 4231 } 4232 // (Else it's an array (or unknown), and we want more precise card marks.) 4233 assert(adr != NULL, ""); 4234 4235 IdealKit ideal(this, true); 4236 4237 Node* tls = __ thread(); // ThreadLocalStorage 4238 4239 Node* no_base = __ top(); 4240 float likely = PROB_LIKELY(0.999); 4241 float unlikely = PROB_UNLIKELY(0.999); 4242 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4243 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4244 Node* zeroX = __ ConX(0); 4245 4246 // Get the alias_index for raw card-mark memory 4247 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4248 4249 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4250 4251 // Offsets into the thread 4252 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4253 DirtyCardQueue::byte_offset_of_index()); 4254 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4255 DirtyCardQueue::byte_offset_of_buf()); 4256 4257 // Pointers into the thread 4258 4259 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4260 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4261 4262 // Now some values 4263 // Use ctrl to avoid hoisting these values past a safepoint, which could 4264 // potentially reset these fields in the JavaThread. 4265 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4266 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4267 4268 // Convert the store obj pointer to an int prior to doing math on it 4269 // Must use ctrl to prevent "integerized oop" existing across safepoint 4270 Node* cast = __ CastPX(__ ctrl(), adr); 4271 4272 // Divide pointer by card size 4273 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4274 4275 // Combine card table base and card offset 4276 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4277 4278 // If we know the value being stored does it cross regions? 4279 4280 if (val != NULL) { 4281 // Does the store cause us to cross regions? 4282 4283 // Should be able to do an unsigned compare of region_size instead of 4284 // and extra shift. Do we have an unsigned compare?? 4285 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4286 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4287 4288 // if (xor_res == 0) same region so skip 4289 __ if_then(xor_res, BoolTest::ne, zeroX); { 4290 4291 // No barrier if we are storing a NULL 4292 __ if_then(val, BoolTest::ne, null(), unlikely); { 4293 4294 // Ok must mark the card if not already dirty 4295 4296 // load the original value of the card 4297 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4298 4299 __ if_then(card_val, BoolTest::ne, young_card); { 4300 sync_kit(ideal); 4301 insert_store_load_for_barrier(); 4302 __ sync_kit(this); 4303 4304 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4305 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4306 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4307 } __ end_if(); 4308 } __ end_if(); 4309 } __ end_if(); 4310 } __ end_if(); 4311 } else { 4312 // Object.clone() instrinsic uses this path. 4313 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4314 } 4315 4316 // Final sync IdealKit and GraphKit. 4317 final_sync(ideal); 4318 } 4319 #undef __ 4320 4321 4322 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4323 Node* len = load_array_length(load_String_value(ctrl, str)); 4324 Node* coder = load_String_coder(ctrl, str); 4325 // Divide length by 2 if coder is UTF16 4326 return _gvn.transform(new RShiftINode(len, coder)); 4327 } 4328 4329 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4330 int value_offset = java_lang_String::value_offset_in_bytes(); 4331 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4332 false, NULL, 0); 4333 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4334 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4335 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4336 ciTypeArrayKlass::make(T_BYTE), true, 0); 4337 int value_field_idx = C->get_alias_index(value_field_type); 4338 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4339 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4340 // String.value field is known to be @Stable. 4341 if (UseImplicitStableValues) { 4342 load = cast_array_to_stable(load, value_type); 4343 } 4344 return load; 4345 } 4346 4347 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4348 if (java_lang_String::has_coder_field()) { 4349 if (!CompactStrings) { 4350 return intcon(java_lang_String::CODER_UTF16); 4351 } 4352 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4353 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4354 false, NULL, 0); 4355 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4356 int coder_field_idx = C->get_alias_index(coder_field_type); 4357 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4358 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4359 } else { 4360 return intcon(0); // false 4361 } 4362 } 4363 4364 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4365 int value_offset = java_lang_String::value_offset_in_bytes(); 4366 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4367 false, NULL, 0); 4368 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4369 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4370 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4371 } 4372 4373 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4374 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4375 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4376 false, NULL, 0); 4377 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4378 int coder_field_idx = C->get_alias_index(coder_field_type); 4379 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4380 value, T_BYTE, coder_field_idx, MemNode::unordered); 4381 } 4382 4383 // Capture src and dst memory state with a MergeMemNode 4384 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4385 if (src_type == dst_type) { 4386 // Types are equal, we don't need a MergeMemNode 4387 return memory(src_type); 4388 } 4389 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4390 record_for_igvn(merge); // fold it up later, if possible 4391 int src_idx = C->get_alias_index(src_type); 4392 int dst_idx = C->get_alias_index(dst_type); 4393 merge->set_memory_at(src_idx, memory(src_idx)); 4394 merge->set_memory_at(dst_idx, memory(dst_idx)); 4395 return merge; 4396 } 4397 4398 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4399 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4400 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4401 // If input and output memory types differ, capture both states to preserve 4402 // the dependency between preceding and subsequent loads/stores. 4403 // For example, the following program: 4404 // StoreB 4405 // compress_string 4406 // LoadB 4407 // has this memory graph (use->def): 4408 // LoadB -> compress_string -> CharMem 4409 // ... -> StoreB -> ByteMem 4410 // The intrinsic hides the dependency between LoadB and StoreB, causing 4411 // the load to read from memory not containing the result of the StoreB. 4412 // The correct memory graph should look like this: 4413 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4414 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4415 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4416 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4417 set_memory(res_mem, TypeAryPtr::BYTES); 4418 return str; 4419 } 4420 4421 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4422 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4423 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4424 // Capture src and dst memory (see comment in 'compress_string'). 4425 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4426 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4427 set_memory(_gvn.transform(str), dst_type); 4428 } 4429 4430 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4431 /** 4432 * int i_char = start; 4433 * for (int i_byte = 0; i_byte < count; i_byte++) { 4434 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4435 * } 4436 */ 4437 add_predicate(); 4438 RegionNode* head = new RegionNode(3); 4439 head->init_req(1, control()); 4440 gvn().set_type(head, Type::CONTROL); 4441 record_for_igvn(head); 4442 4443 Node* i_byte = new PhiNode(head, TypeInt::INT); 4444 i_byte->init_req(1, intcon(0)); 4445 gvn().set_type(i_byte, TypeInt::INT); 4446 record_for_igvn(i_byte); 4447 4448 Node* i_char = new PhiNode(head, TypeInt::INT); 4449 i_char->init_req(1, start); 4450 gvn().set_type(i_char, TypeInt::INT); 4451 record_for_igvn(i_char); 4452 4453 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4454 gvn().set_type(mem, Type::MEMORY); 4455 record_for_igvn(mem); 4456 set_control(head); 4457 set_memory(mem, TypeAryPtr::BYTES); 4458 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4459 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4460 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4461 false, false, true /* mismatched */); 4462 4463 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4464 head->init_req(2, IfTrue(iff)); 4465 mem->init_req(2, st); 4466 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4467 i_char->init_req(2, AddI(i_char, intcon(2))); 4468 4469 set_control(IfFalse(iff)); 4470 set_memory(st, TypeAryPtr::BYTES); 4471 } 4472 4473 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4474 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4475 // assumption of CCP analysis. 4476 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4477 }