1 /* 2 * Copyright (c) 2001, 2016, 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 "memory/resourceArea.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/graphKit.hpp" 37 #include "opto/idealKit.hpp" 38 #include "opto/intrinsicnode.hpp" 39 #include "opto/locknode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/opaquenode.hpp" 42 #include "opto/parse.hpp" 43 #include "opto/rootnode.hpp" 44 #include "opto/runtime.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 //----------------------------GraphKit----------------------------------------- 49 // Main utility constructor. 50 GraphKit::GraphKit(JVMState* jvms) 51 : Phase(Phase::Parser), 52 _env(C->env()), 53 _gvn(*C->initial_gvn()) 54 { 55 _exceptions = jvms->map()->next_exception(); 56 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 57 set_jvms(jvms); 58 } 59 60 // Private constructor for parser. 61 GraphKit::GraphKit() 62 : Phase(Phase::Parser), 63 _env(C->env()), 64 _gvn(*C->initial_gvn()) 65 { 66 _exceptions = NULL; 67 set_map(NULL); 68 debug_only(_sp = -99); 69 debug_only(set_bci(-99)); 70 } 71 72 73 74 //---------------------------clean_stack--------------------------------------- 75 // Clear away rubbish from the stack area of the JVM state. 76 // This destroys any arguments that may be waiting on the stack. 77 void GraphKit::clean_stack(int from_sp) { 78 SafePointNode* map = this->map(); 79 JVMState* jvms = this->jvms(); 80 int stk_size = jvms->stk_size(); 81 int stkoff = jvms->stkoff(); 82 Node* top = this->top(); 83 for (int i = from_sp; i < stk_size; i++) { 84 if (map->in(stkoff + i) != top) { 85 map->set_req(stkoff + i, top); 86 } 87 } 88 } 89 90 91 //--------------------------------sync_jvms----------------------------------- 92 // Make sure our current jvms agrees with our parse state. 93 JVMState* GraphKit::sync_jvms() const { 94 JVMState* jvms = this->jvms(); 95 jvms->set_bci(bci()); // Record the new bci in the JVMState 96 jvms->set_sp(sp()); // Record the new sp in the JVMState 97 assert(jvms_in_sync(), "jvms is now in sync"); 98 return jvms; 99 } 100 101 //--------------------------------sync_jvms_for_reexecute--------------------- 102 // Make sure our current jvms agrees with our parse state. This version 103 // uses the reexecute_sp for reexecuting bytecodes. 104 JVMState* GraphKit::sync_jvms_for_reexecute() { 105 JVMState* jvms = this->jvms(); 106 jvms->set_bci(bci()); // Record the new bci in the JVMState 107 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 108 return jvms; 109 } 110 111 #ifdef ASSERT 112 bool GraphKit::jvms_in_sync() const { 113 Parse* parse = is_Parse(); 114 if (parse == NULL) { 115 if (bci() != jvms()->bci()) return false; 116 if (sp() != (int)jvms()->sp()) return false; 117 return true; 118 } 119 if (jvms()->method() != parse->method()) return false; 120 if (jvms()->bci() != parse->bci()) return false; 121 int jvms_sp = jvms()->sp(); 122 if (jvms_sp != parse->sp()) return false; 123 int jvms_depth = jvms()->depth(); 124 if (jvms_depth != parse->depth()) return false; 125 return true; 126 } 127 128 // Local helper checks for special internal merge points 129 // used to accumulate and merge exception states. 130 // They are marked by the region's in(0) edge being the map itself. 131 // Such merge points must never "escape" into the parser at large, 132 // until they have been handed to gvn.transform. 133 static bool is_hidden_merge(Node* reg) { 134 if (reg == NULL) return false; 135 if (reg->is_Phi()) { 136 reg = reg->in(0); 137 if (reg == NULL) return false; 138 } 139 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 140 } 141 142 void GraphKit::verify_map() const { 143 if (map() == NULL) return; // null map is OK 144 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 145 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 146 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 147 } 148 149 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 150 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 151 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 152 } 153 #endif 154 155 //---------------------------stop_and_kill_map--------------------------------- 156 // Set _map to NULL, signalling a stop to further bytecode execution. 157 // First smash the current map's control to a constant, to mark it dead. 158 void GraphKit::stop_and_kill_map() { 159 SafePointNode* dead_map = stop(); 160 if (dead_map != NULL) { 161 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 162 assert(dead_map->is_killed(), "must be so marked"); 163 } 164 } 165 166 167 //--------------------------------stopped-------------------------------------- 168 // Tell if _map is NULL, or control is top. 169 bool GraphKit::stopped() { 170 if (map() == NULL) return true; 171 else if (control() == top()) return true; 172 else return false; 173 } 174 175 176 //-----------------------------has_ex_handler---------------------------------- 177 // Tell if this method or any caller method has exception handlers. 178 bool GraphKit::has_ex_handler() { 179 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 180 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 181 return true; 182 } 183 } 184 return false; 185 } 186 187 //------------------------------save_ex_oop------------------------------------ 188 // Save an exception without blowing stack contents or other JVM state. 189 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 190 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 191 ex_map->add_req(ex_oop); 192 debug_only(verify_exception_state(ex_map)); 193 } 194 195 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 196 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 197 Node* ex_oop = ex_map->in(ex_map->req()-1); 198 if (clear_it) ex_map->del_req(ex_map->req()-1); 199 return ex_oop; 200 } 201 202 //-----------------------------saved_ex_oop------------------------------------ 203 // Recover a saved exception from its map. 204 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 205 return common_saved_ex_oop(ex_map, false); 206 } 207 208 //--------------------------clear_saved_ex_oop--------------------------------- 209 // Erase a previously saved exception from its map. 210 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 211 return common_saved_ex_oop(ex_map, true); 212 } 213 214 #ifdef ASSERT 215 //---------------------------has_saved_ex_oop---------------------------------- 216 // Erase a previously saved exception from its map. 217 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 218 return ex_map->req() == ex_map->jvms()->endoff()+1; 219 } 220 #endif 221 222 //-------------------------make_exception_state-------------------------------- 223 // Turn the current JVM state into an exception state, appending the ex_oop. 224 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 225 sync_jvms(); 226 SafePointNode* ex_map = stop(); // do not manipulate this map any more 227 set_saved_ex_oop(ex_map, ex_oop); 228 return ex_map; 229 } 230 231 232 //--------------------------add_exception_state-------------------------------- 233 // Add an exception to my list of exceptions. 234 void GraphKit::add_exception_state(SafePointNode* ex_map) { 235 if (ex_map == NULL || ex_map->control() == top()) { 236 return; 237 } 238 #ifdef ASSERT 239 verify_exception_state(ex_map); 240 if (has_exceptions()) { 241 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 242 } 243 #endif 244 245 // If there is already an exception of exactly this type, merge with it. 246 // In particular, null-checks and other low-level exceptions common up here. 247 Node* ex_oop = saved_ex_oop(ex_map); 248 const Type* ex_type = _gvn.type(ex_oop); 249 if (ex_oop == top()) { 250 // No action needed. 251 return; 252 } 253 assert(ex_type->isa_instptr(), "exception must be an instance"); 254 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 255 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 256 // We check sp also because call bytecodes can generate exceptions 257 // both before and after arguments are popped! 258 if (ex_type2 == ex_type 259 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 260 combine_exception_states(ex_map, e2); 261 return; 262 } 263 } 264 265 // No pre-existing exception of the same type. Chain it on the list. 266 push_exception_state(ex_map); 267 } 268 269 //-----------------------add_exception_states_from----------------------------- 270 void GraphKit::add_exception_states_from(JVMState* jvms) { 271 SafePointNode* ex_map = jvms->map()->next_exception(); 272 if (ex_map != NULL) { 273 jvms->map()->set_next_exception(NULL); 274 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 275 next_map = ex_map->next_exception(); 276 ex_map->set_next_exception(NULL); 277 add_exception_state(ex_map); 278 } 279 } 280 } 281 282 //-----------------------transfer_exceptions_into_jvms------------------------- 283 JVMState* GraphKit::transfer_exceptions_into_jvms() { 284 if (map() == NULL) { 285 // We need a JVMS to carry the exceptions, but the map has gone away. 286 // Create a scratch JVMS, cloned from any of the exception states... 287 if (has_exceptions()) { 288 _map = _exceptions; 289 _map = clone_map(); 290 _map->set_next_exception(NULL); 291 clear_saved_ex_oop(_map); 292 debug_only(verify_map()); 293 } else { 294 // ...or created from scratch 295 JVMState* jvms = new (C) JVMState(_method, NULL); 296 jvms->set_bci(_bci); 297 jvms->set_sp(_sp); 298 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 299 set_jvms(jvms); 300 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 301 set_all_memory(top()); 302 while (map()->req() < jvms->endoff()) map()->add_req(top()); 303 } 304 // (This is a kludge, in case you didn't notice.) 305 set_control(top()); 306 } 307 JVMState* jvms = sync_jvms(); 308 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 309 jvms->map()->set_next_exception(_exceptions); 310 _exceptions = NULL; // done with this set of exceptions 311 return jvms; 312 } 313 314 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 315 assert(is_hidden_merge(dstphi), "must be a special merge node"); 316 assert(is_hidden_merge(srcphi), "must be a special merge node"); 317 uint limit = srcphi->req(); 318 for (uint i = PhiNode::Input; i < limit; i++) { 319 dstphi->add_req(srcphi->in(i)); 320 } 321 } 322 static inline void add_one_req(Node* dstphi, Node* src) { 323 assert(is_hidden_merge(dstphi), "must be a special merge node"); 324 assert(!is_hidden_merge(src), "must not be a special merge node"); 325 dstphi->add_req(src); 326 } 327 328 //-----------------------combine_exception_states------------------------------ 329 // This helper function combines exception states by building phis on a 330 // specially marked state-merging region. These regions and phis are 331 // untransformed, and can build up gradually. The region is marked by 332 // having a control input of its exception map, rather than NULL. Such 333 // regions do not appear except in this function, and in use_exception_state. 334 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 335 if (failing()) return; // dying anyway... 336 JVMState* ex_jvms = ex_map->_jvms; 337 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 338 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 339 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 340 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 341 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 342 assert(ex_map->req() == phi_map->req(), "matching maps"); 343 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 344 Node* hidden_merge_mark = root(); 345 Node* region = phi_map->control(); 346 MergeMemNode* phi_mem = phi_map->merged_memory(); 347 MergeMemNode* ex_mem = ex_map->merged_memory(); 348 if (region->in(0) != hidden_merge_mark) { 349 // The control input is not (yet) a specially-marked region in phi_map. 350 // Make it so, and build some phis. 351 region = new RegionNode(2); 352 _gvn.set_type(region, Type::CONTROL); 353 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 354 region->init_req(1, phi_map->control()); 355 phi_map->set_control(region); 356 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 357 record_for_igvn(io_phi); 358 _gvn.set_type(io_phi, Type::ABIO); 359 phi_map->set_i_o(io_phi); 360 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 361 Node* m = mms.memory(); 362 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 363 record_for_igvn(m_phi); 364 _gvn.set_type(m_phi, Type::MEMORY); 365 mms.set_memory(m_phi); 366 } 367 } 368 369 // Either or both of phi_map and ex_map might already be converted into phis. 370 Node* ex_control = ex_map->control(); 371 // if there is special marking on ex_map also, we add multiple edges from src 372 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 373 // how wide was the destination phi_map, originally? 374 uint orig_width = region->req(); 375 376 if (add_multiple) { 377 add_n_reqs(region, ex_control); 378 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 379 } else { 380 // ex_map has no merges, so we just add single edges everywhere 381 add_one_req(region, ex_control); 382 add_one_req(phi_map->i_o(), ex_map->i_o()); 383 } 384 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 385 if (mms.is_empty()) { 386 // get a copy of the base memory, and patch some inputs into it 387 const TypePtr* adr_type = mms.adr_type(C); 388 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 389 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 390 mms.set_memory(phi); 391 // Prepare to append interesting stuff onto the newly sliced phi: 392 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 393 } 394 // Append stuff from ex_map: 395 if (add_multiple) { 396 add_n_reqs(mms.memory(), mms.memory2()); 397 } else { 398 add_one_req(mms.memory(), mms.memory2()); 399 } 400 } 401 uint limit = ex_map->req(); 402 for (uint i = TypeFunc::Parms; i < limit; i++) { 403 // Skip everything in the JVMS after tos. (The ex_oop follows.) 404 if (i == tos) i = ex_jvms->monoff(); 405 Node* src = ex_map->in(i); 406 Node* dst = phi_map->in(i); 407 if (src != dst) { 408 PhiNode* phi; 409 if (dst->in(0) != region) { 410 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 411 record_for_igvn(phi); 412 _gvn.set_type(phi, phi->type()); 413 phi_map->set_req(i, dst); 414 // Prepare to append interesting stuff onto the new phi: 415 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 416 } else { 417 assert(dst->is_Phi(), "nobody else uses a hidden region"); 418 phi = dst->as_Phi(); 419 } 420 if (add_multiple && src->in(0) == ex_control) { 421 // Both are phis. 422 add_n_reqs(dst, src); 423 } else { 424 while (dst->req() < region->req()) add_one_req(dst, src); 425 } 426 const Type* srctype = _gvn.type(src); 427 if (phi->type() != srctype) { 428 const Type* dsttype = phi->type()->meet_speculative(srctype); 429 if (phi->type() != dsttype) { 430 phi->set_type(dsttype); 431 _gvn.set_type(phi, dsttype); 432 } 433 } 434 } 435 } 436 phi_map->merge_replaced_nodes_with(ex_map); 437 } 438 439 //--------------------------use_exception_state-------------------------------- 440 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 441 if (failing()) { stop(); return top(); } 442 Node* region = phi_map->control(); 443 Node* hidden_merge_mark = root(); 444 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 445 Node* ex_oop = clear_saved_ex_oop(phi_map); 446 if (region->in(0) == hidden_merge_mark) { 447 // Special marking for internal ex-states. Process the phis now. 448 region->set_req(0, region); // now it's an ordinary region 449 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 450 // Note: Setting the jvms also sets the bci and sp. 451 set_control(_gvn.transform(region)); 452 uint tos = jvms()->stkoff() + sp(); 453 for (uint i = 1; i < tos; i++) { 454 Node* x = phi_map->in(i); 455 if (x->in(0) == region) { 456 assert(x->is_Phi(), "expected a special phi"); 457 phi_map->set_req(i, _gvn.transform(x)); 458 } 459 } 460 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 461 Node* x = mms.memory(); 462 if (x->in(0) == region) { 463 assert(x->is_Phi(), "nobody else uses a hidden region"); 464 mms.set_memory(_gvn.transform(x)); 465 } 466 } 467 if (ex_oop->in(0) == region) { 468 assert(ex_oop->is_Phi(), "expected a special phi"); 469 ex_oop = _gvn.transform(ex_oop); 470 } 471 } else { 472 set_jvms(phi_map->jvms()); 473 } 474 475 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 476 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 477 return ex_oop; 478 } 479 480 //---------------------------------java_bc------------------------------------- 481 Bytecodes::Code GraphKit::java_bc() const { 482 ciMethod* method = this->method(); 483 int bci = this->bci(); 484 if (method != NULL && bci != InvocationEntryBci) 485 return method->java_code_at_bci(bci); 486 else 487 return Bytecodes::_illegal; 488 } 489 490 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 491 bool must_throw) { 492 // if the exception capability is set, then we will generate code 493 // to check the JavaThread.should_post_on_exceptions flag to see 494 // if we actually need to report exception events (for this 495 // thread). If we don't need to report exception events, we will 496 // take the normal fast path provided by add_exception_events. If 497 // exception event reporting is enabled for this thread, we will 498 // take the uncommon_trap in the BuildCutout below. 499 500 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 501 Node* jthread = _gvn.transform(new ThreadLocalNode()); 502 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 503 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 504 505 // Test the should_post_on_exceptions_flag vs. 0 506 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 507 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 508 509 // Branch to slow_path if should_post_on_exceptions_flag was true 510 { BuildCutout unless(this, tst, PROB_MAX); 511 // Do not try anything fancy if we're notifying the VM on every throw. 512 // Cf. case Bytecodes::_athrow in parse2.cpp. 513 uncommon_trap(reason, Deoptimization::Action_none, 514 (ciKlass*)NULL, (char*)NULL, must_throw); 515 } 516 517 } 518 519 //------------------------------builtin_throw---------------------------------- 520 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 521 bool must_throw = true; 522 523 if (env()->jvmti_can_post_on_exceptions()) { 524 // check if we must post exception events, take uncommon trap if so 525 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 526 // here if should_post_on_exceptions is false 527 // continue on with the normal codegen 528 } 529 530 // If this particular condition has not yet happened at this 531 // bytecode, then use the uncommon trap mechanism, and allow for 532 // a future recompilation if several traps occur here. 533 // If the throw is hot, try to use a more complicated inline mechanism 534 // which keeps execution inside the compiled code. 535 bool treat_throw_as_hot = false; 536 ciMethodData* md = method()->method_data(); 537 538 if (ProfileTraps) { 539 if (too_many_traps(reason)) { 540 treat_throw_as_hot = true; 541 } 542 // (If there is no MDO at all, assume it is early in 543 // execution, and that any deopts are part of the 544 // startup transient, and don't need to be remembered.) 545 546 // Also, if there is a local exception handler, treat all throws 547 // as hot if there has been at least one in this method. 548 if (C->trap_count(reason) != 0 549 && method()->method_data()->trap_count(reason) != 0 550 && has_ex_handler()) { 551 treat_throw_as_hot = true; 552 } 553 } 554 555 // If this throw happens frequently, an uncommon trap might cause 556 // a performance pothole. If there is a local exception handler, 557 // and if this particular bytecode appears to be deoptimizing often, 558 // let us handle the throw inline, with a preconstructed instance. 559 // Note: If the deopt count has blown up, the uncommon trap 560 // runtime is going to flush this nmethod, not matter what. 561 if (treat_throw_as_hot 562 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 563 // If the throw is local, we use a pre-existing instance and 564 // punt on the backtrace. This would lead to a missing backtrace 565 // (a repeat of 4292742) if the backtrace object is ever asked 566 // for its backtrace. 567 // Fixing this remaining case of 4292742 requires some flavor of 568 // escape analysis. Leave that for the future. 569 ciInstance* ex_obj = NULL; 570 switch (reason) { 571 case Deoptimization::Reason_null_check: 572 ex_obj = env()->NullPointerException_instance(); 573 break; 574 case Deoptimization::Reason_div0_check: 575 ex_obj = env()->ArithmeticException_instance(); 576 break; 577 case Deoptimization::Reason_range_check: 578 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 579 break; 580 case Deoptimization::Reason_class_check: 581 if (java_bc() == Bytecodes::_aastore) { 582 ex_obj = env()->ArrayStoreException_instance(); 583 } else { 584 ex_obj = env()->ClassCastException_instance(); 585 } 586 break; 587 } 588 if (failing()) { stop(); return; } // exception allocation might fail 589 if (ex_obj != NULL) { 590 // Cheat with a preallocated exception object. 591 if (C->log() != NULL) 592 C->log()->elem("hot_throw preallocated='1' reason='%s'", 593 Deoptimization::trap_reason_name(reason)); 594 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 595 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 596 597 // Clear the detail message of the preallocated exception object. 598 // Weblogic sometimes mutates the detail message of exceptions 599 // using reflection. 600 int offset = java_lang_Throwable::get_detailMessage_offset(); 601 const TypePtr* adr_typ = ex_con->add_offset(offset); 602 603 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 604 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 605 // Conservatively release stores of object references. 606 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 607 608 add_exception_state(make_exception_state(ex_node)); 609 return; 610 } 611 } 612 613 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 614 // It won't be much cheaper than bailing to the interp., since we'll 615 // have to pass up all the debug-info, and the runtime will have to 616 // create the stack trace. 617 618 // Usual case: Bail to interpreter. 619 // Reserve the right to recompile if we haven't seen anything yet. 620 621 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 622 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 623 if (treat_throw_as_hot 624 && (method()->method_data()->trap_recompiled_at(bci(), m) 625 || C->too_many_traps(reason))) { 626 // We cannot afford to take more traps here. Suffer in the interpreter. 627 if (C->log() != NULL) 628 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 629 Deoptimization::trap_reason_name(reason), 630 C->trap_count(reason)); 631 action = Deoptimization::Action_none; 632 } 633 634 // "must_throw" prunes the JVM state to include only the stack, if there 635 // are no local exception handlers. This should cut down on register 636 // allocation time and code size, by drastically reducing the number 637 // of in-edges on the call to the uncommon trap. 638 639 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 640 } 641 642 643 //----------------------------PreserveJVMState--------------------------------- 644 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 645 debug_only(kit->verify_map()); 646 _kit = kit; 647 _map = kit->map(); // preserve the map 648 _sp = kit->sp(); 649 kit->set_map(clone_map ? kit->clone_map() : NULL); 650 #ifdef ASSERT 651 _bci = kit->bci(); 652 Parse* parser = kit->is_Parse(); 653 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 654 _block = block; 655 #endif 656 } 657 PreserveJVMState::~PreserveJVMState() { 658 GraphKit* kit = _kit; 659 #ifdef ASSERT 660 assert(kit->bci() == _bci, "bci must not shift"); 661 Parse* parser = kit->is_Parse(); 662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 663 assert(block == _block, "block must not shift"); 664 #endif 665 kit->set_map(_map); 666 kit->set_sp(_sp); 667 } 668 669 670 //-----------------------------BuildCutout------------------------------------- 671 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 672 : PreserveJVMState(kit) 673 { 674 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 675 SafePointNode* outer_map = _map; // preserved map is caller's 676 SafePointNode* inner_map = kit->map(); 677 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 678 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 679 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 680 } 681 BuildCutout::~BuildCutout() { 682 GraphKit* kit = _kit; 683 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 684 } 685 686 //---------------------------PreserveReexecuteState---------------------------- 687 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 688 assert(!kit->stopped(), "must call stopped() before"); 689 _kit = kit; 690 _sp = kit->sp(); 691 _reexecute = kit->jvms()->_reexecute; 692 } 693 PreserveReexecuteState::~PreserveReexecuteState() { 694 if (_kit->stopped()) return; 695 _kit->jvms()->_reexecute = _reexecute; 696 _kit->set_sp(_sp); 697 } 698 699 //------------------------------clone_map-------------------------------------- 700 // Implementation of PreserveJVMState 701 // 702 // Only clone_map(...) here. If this function is only used in the 703 // PreserveJVMState class we may want to get rid of this extra 704 // function eventually and do it all there. 705 706 SafePointNode* GraphKit::clone_map() { 707 if (map() == NULL) return NULL; 708 709 // Clone the memory edge first 710 Node* mem = MergeMemNode::make(map()->memory()); 711 gvn().set_type_bottom(mem); 712 713 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 714 JVMState* jvms = this->jvms(); 715 JVMState* clonejvms = jvms->clone_shallow(C); 716 clonemap->set_memory(mem); 717 clonemap->set_jvms(clonejvms); 718 clonejvms->set_map(clonemap); 719 record_for_igvn(clonemap); 720 gvn().set_type_bottom(clonemap); 721 return clonemap; 722 } 723 724 725 //-----------------------------set_map_clone----------------------------------- 726 void GraphKit::set_map_clone(SafePointNode* m) { 727 _map = m; 728 _map = clone_map(); 729 _map->set_next_exception(NULL); 730 debug_only(verify_map()); 731 } 732 733 734 //----------------------------kill_dead_locals--------------------------------- 735 // Detect any locals which are known to be dead, and force them to top. 736 void GraphKit::kill_dead_locals() { 737 // Consult the liveness information for the locals. If any 738 // of them are unused, then they can be replaced by top(). This 739 // should help register allocation time and cut down on the size 740 // of the deoptimization information. 741 742 // This call is made from many of the bytecode handling 743 // subroutines called from the Big Switch in do_one_bytecode. 744 // Every bytecode which might include a slow path is responsible 745 // for killing its dead locals. The more consistent we 746 // are about killing deads, the fewer useless phis will be 747 // constructed for them at various merge points. 748 749 // bci can be -1 (InvocationEntryBci). We return the entry 750 // liveness for the method. 751 752 if (method() == NULL || method()->code_size() == 0) { 753 // We are building a graph for a call to a native method. 754 // All locals are live. 755 return; 756 } 757 758 ResourceMark rm; 759 760 // Consult the liveness information for the locals. If any 761 // of them are unused, then they can be replaced by top(). This 762 // should help register allocation time and cut down on the size 763 // of the deoptimization information. 764 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 765 766 int len = (int)live_locals.size(); 767 assert(len <= jvms()->loc_size(), "too many live locals"); 768 for (int local = 0; local < len; local++) { 769 if (!live_locals.at(local)) { 770 set_local(local, top()); 771 } 772 } 773 } 774 775 #ifdef ASSERT 776 //-------------------------dead_locals_are_killed------------------------------ 777 // Return true if all dead locals are set to top in the map. 778 // Used to assert "clean" debug info at various points. 779 bool GraphKit::dead_locals_are_killed() { 780 if (method() == NULL || method()->code_size() == 0) { 781 // No locals need to be dead, so all is as it should be. 782 return true; 783 } 784 785 // Make sure somebody called kill_dead_locals upstream. 786 ResourceMark rm; 787 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 788 if (jvms->loc_size() == 0) continue; // no locals to consult 789 SafePointNode* map = jvms->map(); 790 ciMethod* method = jvms->method(); 791 int bci = jvms->bci(); 792 if (jvms == this->jvms()) { 793 bci = this->bci(); // it might not yet be synched 794 } 795 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 796 int len = (int)live_locals.size(); 797 if (!live_locals.is_valid() || len == 0) 798 // This method is trivial, or is poisoned by a breakpoint. 799 return true; 800 assert(len == jvms->loc_size(), "live map consistent with locals map"); 801 for (int local = 0; local < len; local++) { 802 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 803 if (PrintMiscellaneous && (Verbose || WizardMode)) { 804 tty->print_cr("Zombie local %d: ", local); 805 jvms->dump(); 806 } 807 return false; 808 } 809 } 810 } 811 return true; 812 } 813 814 #endif //ASSERT 815 816 // Helper function for enforcing certain bytecodes to reexecute if 817 // deoptimization happens 818 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 819 ciMethod* cur_method = jvms->method(); 820 int cur_bci = jvms->bci(); 821 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 822 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 823 return Interpreter::bytecode_should_reexecute(code) || 824 is_anewarray && code == Bytecodes::_multianewarray; 825 // Reexecute _multianewarray bytecode which was replaced with 826 // sequence of [a]newarray. See Parse::do_multianewarray(). 827 // 828 // Note: interpreter should not have it set since this optimization 829 // is limited by dimensions and guarded by flag so in some cases 830 // multianewarray() runtime calls will be generated and 831 // the bytecode should not be reexecutes (stack will not be reset). 832 } else 833 return false; 834 } 835 836 // Helper function for adding JVMState and debug information to node 837 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 838 // Add the safepoint edges to the call (or other safepoint). 839 840 // Make sure dead locals are set to top. This 841 // should help register allocation time and cut down on the size 842 // of the deoptimization information. 843 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 844 845 // Walk the inline list to fill in the correct set of JVMState's 846 // Also fill in the associated edges for each JVMState. 847 848 // If the bytecode needs to be reexecuted we need to put 849 // the arguments back on the stack. 850 const bool should_reexecute = jvms()->should_reexecute(); 851 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 852 853 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 854 // undefined if the bci is different. This is normal for Parse but it 855 // should not happen for LibraryCallKit because only one bci is processed. 856 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 857 "in LibraryCallKit the reexecute bit should not change"); 858 859 // If we are guaranteed to throw, we can prune everything but the 860 // input to the current bytecode. 861 bool can_prune_locals = false; 862 uint stack_slots_not_pruned = 0; 863 int inputs = 0, depth = 0; 864 if (must_throw) { 865 assert(method() == youngest_jvms->method(), "sanity"); 866 if (compute_stack_effects(inputs, depth)) { 867 can_prune_locals = true; 868 stack_slots_not_pruned = inputs; 869 } 870 } 871 872 if (env()->should_retain_local_variables()) { 873 // At any safepoint, this method can get breakpointed, which would 874 // then require an immediate deoptimization. 875 can_prune_locals = false; // do not prune locals 876 stack_slots_not_pruned = 0; 877 } 878 879 // do not scribble on the input jvms 880 JVMState* out_jvms = youngest_jvms->clone_deep(C); 881 call->set_jvms(out_jvms); // Start jvms list for call node 882 883 // For a known set of bytecodes, the interpreter should reexecute them if 884 // deoptimization happens. We set the reexecute state for them here 885 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 886 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 887 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 888 } 889 890 // Presize the call: 891 DEBUG_ONLY(uint non_debug_edges = call->req()); 892 call->add_req_batch(top(), youngest_jvms->debug_depth()); 893 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 894 895 // Set up edges so that the call looks like this: 896 // Call [state:] ctl io mem fptr retadr 897 // [parms:] parm0 ... parmN 898 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 899 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 900 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 901 // Note that caller debug info precedes callee debug info. 902 903 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 904 uint debug_ptr = call->req(); 905 906 // Loop over the map input edges associated with jvms, add them 907 // to the call node, & reset all offsets to match call node array. 908 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 909 uint debug_end = debug_ptr; 910 uint debug_start = debug_ptr - in_jvms->debug_size(); 911 debug_ptr = debug_start; // back up the ptr 912 913 uint p = debug_start; // walks forward in [debug_start, debug_end) 914 uint j, k, l; 915 SafePointNode* in_map = in_jvms->map(); 916 out_jvms->set_map(call); 917 918 if (can_prune_locals) { 919 assert(in_jvms->method() == out_jvms->method(), "sanity"); 920 // If the current throw can reach an exception handler in this JVMS, 921 // then we must keep everything live that can reach that handler. 922 // As a quick and dirty approximation, we look for any handlers at all. 923 if (in_jvms->method()->has_exception_handlers()) { 924 can_prune_locals = false; 925 } 926 } 927 928 // Add the Locals 929 k = in_jvms->locoff(); 930 l = in_jvms->loc_size(); 931 out_jvms->set_locoff(p); 932 if (!can_prune_locals) { 933 for (j = 0; j < l; j++) 934 call->set_req(p++, in_map->in(k+j)); 935 } else { 936 p += l; // already set to top above by add_req_batch 937 } 938 939 // Add the Expression Stack 940 k = in_jvms->stkoff(); 941 l = in_jvms->sp(); 942 out_jvms->set_stkoff(p); 943 if (!can_prune_locals) { 944 for (j = 0; j < l; j++) 945 call->set_req(p++, in_map->in(k+j)); 946 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 947 // Divide stack into {S0,...,S1}, where S0 is set to top. 948 uint s1 = stack_slots_not_pruned; 949 stack_slots_not_pruned = 0; // for next iteration 950 if (s1 > l) s1 = l; 951 uint s0 = l - s1; 952 p += s0; // skip the tops preinstalled by add_req_batch 953 for (j = s0; j < l; j++) 954 call->set_req(p++, in_map->in(k+j)); 955 } else { 956 p += l; // already set to top above by add_req_batch 957 } 958 959 // Add the Monitors 960 k = in_jvms->monoff(); 961 l = in_jvms->mon_size(); 962 out_jvms->set_monoff(p); 963 for (j = 0; j < l; j++) 964 call->set_req(p++, in_map->in(k+j)); 965 966 // Copy any scalar object fields. 967 k = in_jvms->scloff(); 968 l = in_jvms->scl_size(); 969 out_jvms->set_scloff(p); 970 for (j = 0; j < l; j++) 971 call->set_req(p++, in_map->in(k+j)); 972 973 // Finish the new jvms. 974 out_jvms->set_endoff(p); 975 976 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 977 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 978 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 979 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 980 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 981 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 982 983 // Update the two tail pointers in parallel. 984 out_jvms = out_jvms->caller(); 985 in_jvms = in_jvms->caller(); 986 } 987 988 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 989 990 // Test the correctness of JVMState::debug_xxx accessors: 991 assert(call->jvms()->debug_start() == non_debug_edges, ""); 992 assert(call->jvms()->debug_end() == call->req(), ""); 993 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 994 } 995 996 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 997 Bytecodes::Code code = java_bc(); 998 if (code == Bytecodes::_wide) { 999 code = method()->java_code_at_bci(bci() + 1); 1000 } 1001 1002 BasicType rtype = T_ILLEGAL; 1003 int rsize = 0; 1004 1005 if (code != Bytecodes::_illegal) { 1006 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1007 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1008 if (rtype < T_CONFLICT) 1009 rsize = type2size[rtype]; 1010 } 1011 1012 switch (code) { 1013 case Bytecodes::_illegal: 1014 return false; 1015 1016 case Bytecodes::_ldc: 1017 case Bytecodes::_ldc_w: 1018 case Bytecodes::_ldc2_w: 1019 inputs = 0; 1020 break; 1021 1022 case Bytecodes::_dup: inputs = 1; break; 1023 case Bytecodes::_dup_x1: inputs = 2; break; 1024 case Bytecodes::_dup_x2: inputs = 3; break; 1025 case Bytecodes::_dup2: inputs = 2; break; 1026 case Bytecodes::_dup2_x1: inputs = 3; break; 1027 case Bytecodes::_dup2_x2: inputs = 4; break; 1028 case Bytecodes::_swap: inputs = 2; break; 1029 case Bytecodes::_arraylength: inputs = 1; break; 1030 1031 case Bytecodes::_getstatic: 1032 case Bytecodes::_putstatic: 1033 case Bytecodes::_getfield: 1034 case Bytecodes::_putfield: 1035 { 1036 bool ignored_will_link; 1037 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1038 int size = field->type()->size(); 1039 bool is_get = (depth >= 0), is_static = (depth & 1); 1040 inputs = (is_static ? 0 : 1); 1041 if (is_get) { 1042 depth = size - inputs; 1043 } else { 1044 inputs += size; // putxxx pops the value from the stack 1045 depth = - inputs; 1046 } 1047 } 1048 break; 1049 1050 case Bytecodes::_invokevirtual: 1051 case Bytecodes::_invokespecial: 1052 case Bytecodes::_invokestatic: 1053 case Bytecodes::_invokedynamic: 1054 case Bytecodes::_invokeinterface: 1055 { 1056 bool ignored_will_link; 1057 ciSignature* declared_signature = NULL; 1058 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1059 assert(declared_signature != NULL, "cannot be null"); 1060 inputs = declared_signature->arg_size_for_bc(code); 1061 int size = declared_signature->return_type()->size(); 1062 depth = size - inputs; 1063 } 1064 break; 1065 1066 case Bytecodes::_multianewarray: 1067 { 1068 ciBytecodeStream iter(method()); 1069 iter.reset_to_bci(bci()); 1070 iter.next(); 1071 inputs = iter.get_dimensions(); 1072 assert(rsize == 1, ""); 1073 depth = rsize - inputs; 1074 } 1075 break; 1076 1077 case Bytecodes::_ireturn: 1078 case Bytecodes::_lreturn: 1079 case Bytecodes::_freturn: 1080 case Bytecodes::_dreturn: 1081 case Bytecodes::_areturn: 1082 assert(rsize == -depth, ""); 1083 inputs = rsize; 1084 break; 1085 1086 case Bytecodes::_jsr: 1087 case Bytecodes::_jsr_w: 1088 inputs = 0; 1089 depth = 1; // S.B. depth=1, not zero 1090 break; 1091 1092 default: 1093 // bytecode produces a typed result 1094 inputs = rsize - depth; 1095 assert(inputs >= 0, ""); 1096 break; 1097 } 1098 1099 #ifdef ASSERT 1100 // spot check 1101 int outputs = depth + inputs; 1102 assert(outputs >= 0, "sanity"); 1103 switch (code) { 1104 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1105 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1106 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1107 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1108 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1109 } 1110 #endif //ASSERT 1111 1112 return true; 1113 } 1114 1115 1116 1117 //------------------------------basic_plus_adr--------------------------------- 1118 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1119 // short-circuit a common case 1120 if (offset == intcon(0)) return ptr; 1121 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1122 } 1123 1124 Node* GraphKit::ConvI2L(Node* offset) { 1125 // short-circuit a common case 1126 jint offset_con = find_int_con(offset, Type::OffsetBot); 1127 if (offset_con != Type::OffsetBot) { 1128 return longcon((jlong) offset_con); 1129 } 1130 return _gvn.transform( new ConvI2LNode(offset)); 1131 } 1132 1133 Node* GraphKit::ConvI2UL(Node* offset) { 1134 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1135 if (offset_con != (juint) Type::OffsetBot) { 1136 return longcon((julong) offset_con); 1137 } 1138 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1139 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1140 return _gvn.transform( new AndLNode(conv, mask) ); 1141 } 1142 1143 Node* GraphKit::ConvL2I(Node* offset) { 1144 // short-circuit a common case 1145 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1146 if (offset_con != (jlong)Type::OffsetBot) { 1147 return intcon((int) offset_con); 1148 } 1149 return _gvn.transform( new ConvL2INode(offset)); 1150 } 1151 1152 //-------------------------load_object_klass----------------------------------- 1153 Node* GraphKit::load_object_klass(Node* obj) { 1154 // Special-case a fresh allocation to avoid building nodes: 1155 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1156 if (akls != NULL) return akls; 1157 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1158 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1159 } 1160 1161 //-------------------------load_array_length----------------------------------- 1162 Node* GraphKit::load_array_length(Node* array) { 1163 // Special-case a fresh allocation to avoid building nodes: 1164 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1165 Node *alen; 1166 if (alloc == NULL) { 1167 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1168 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1169 } else { 1170 alen = alloc->Ideal_length(); 1171 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1172 if (ccast != alen) { 1173 alen = _gvn.transform(ccast); 1174 } 1175 } 1176 return alen; 1177 } 1178 1179 //------------------------------do_null_check---------------------------------- 1180 // Helper function to do a NULL pointer check. Returned value is 1181 // the incoming address with NULL casted away. You are allowed to use the 1182 // not-null value only if you are control dependent on the test. 1183 #ifndef PRODUCT 1184 extern int explicit_null_checks_inserted, 1185 explicit_null_checks_elided; 1186 #endif 1187 Node* GraphKit::null_check_common(Node* value, BasicType type, 1188 // optional arguments for variations: 1189 bool assert_null, 1190 Node* *null_control, 1191 bool speculative) { 1192 assert(!assert_null || null_control == NULL, "not both at once"); 1193 if (stopped()) return top(); 1194 NOT_PRODUCT(explicit_null_checks_inserted++); 1195 1196 // Construct NULL check 1197 Node *chk = NULL; 1198 switch(type) { 1199 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1200 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1201 case T_ARRAY : // fall through 1202 type = T_OBJECT; // simplify further tests 1203 case T_OBJECT : { 1204 const Type *t = _gvn.type( value ); 1205 1206 const TypeOopPtr* tp = t->isa_oopptr(); 1207 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1208 // Only for do_null_check, not any of its siblings: 1209 && !assert_null && null_control == NULL) { 1210 // Usually, any field access or invocation on an unloaded oop type 1211 // will simply fail to link, since the statically linked class is 1212 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1213 // the static class is loaded but the sharper oop type is not. 1214 // Rather than checking for this obscure case in lots of places, 1215 // we simply observe that a null check on an unloaded class 1216 // will always be followed by a nonsense operation, so we 1217 // can just issue the uncommon trap here. 1218 // Our access to the unloaded class will only be correct 1219 // after it has been loaded and initialized, which requires 1220 // a trip through the interpreter. 1221 #ifndef PRODUCT 1222 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1223 #endif 1224 uncommon_trap(Deoptimization::Reason_unloaded, 1225 Deoptimization::Action_reinterpret, 1226 tp->klass(), "!loaded"); 1227 return top(); 1228 } 1229 1230 if (assert_null) { 1231 // See if the type is contained in NULL_PTR. 1232 // If so, then the value is already null. 1233 if (t->higher_equal(TypePtr::NULL_PTR)) { 1234 NOT_PRODUCT(explicit_null_checks_elided++); 1235 return value; // Elided null assert quickly! 1236 } 1237 } else { 1238 // See if mixing in the NULL pointer changes type. 1239 // If so, then the NULL pointer was not allowed in the original 1240 // type. In other words, "value" was not-null. 1241 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1242 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1243 NOT_PRODUCT(explicit_null_checks_elided++); 1244 return value; // Elided null check quickly! 1245 } 1246 } 1247 chk = new CmpPNode( value, null() ); 1248 break; 1249 } 1250 1251 default: 1252 fatal("unexpected type: %s", type2name(type)); 1253 } 1254 assert(chk != NULL, "sanity check"); 1255 chk = _gvn.transform(chk); 1256 1257 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1258 BoolNode *btst = new BoolNode( chk, btest); 1259 Node *tst = _gvn.transform( btst ); 1260 1261 //----------- 1262 // if peephole optimizations occurred, a prior test existed. 1263 // If a prior test existed, maybe it dominates as we can avoid this test. 1264 if (tst != btst && type == T_OBJECT) { 1265 // At this point we want to scan up the CFG to see if we can 1266 // find an identical test (and so avoid this test altogether). 1267 Node *cfg = control(); 1268 int depth = 0; 1269 while( depth < 16 ) { // Limit search depth for speed 1270 if( cfg->Opcode() == Op_IfTrue && 1271 cfg->in(0)->in(1) == tst ) { 1272 // Found prior test. Use "cast_not_null" to construct an identical 1273 // CastPP (and hence hash to) as already exists for the prior test. 1274 // Return that casted value. 1275 if (assert_null) { 1276 replace_in_map(value, null()); 1277 return null(); // do not issue the redundant test 1278 } 1279 Node *oldcontrol = control(); 1280 set_control(cfg); 1281 Node *res = cast_not_null(value); 1282 set_control(oldcontrol); 1283 NOT_PRODUCT(explicit_null_checks_elided++); 1284 return res; 1285 } 1286 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1287 if (cfg == NULL) break; // Quit at region nodes 1288 depth++; 1289 } 1290 } 1291 1292 //----------- 1293 // Branch to failure if null 1294 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1295 Deoptimization::DeoptReason reason; 1296 if (assert_null) { 1297 reason = Deoptimization::Reason_null_assert; 1298 } else if (type == T_OBJECT) { 1299 reason = Deoptimization::reason_null_check(speculative); 1300 } else { 1301 reason = Deoptimization::Reason_div0_check; 1302 } 1303 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1304 // ciMethodData::has_trap_at will return a conservative -1 if any 1305 // must-be-null assertion has failed. This could cause performance 1306 // problems for a method after its first do_null_assert failure. 1307 // Consider using 'Reason_class_check' instead? 1308 1309 // To cause an implicit null check, we set the not-null probability 1310 // to the maximum (PROB_MAX). For an explicit check the probability 1311 // is set to a smaller value. 1312 if (null_control != NULL || too_many_traps(reason)) { 1313 // probability is less likely 1314 ok_prob = PROB_LIKELY_MAG(3); 1315 } else if (!assert_null && 1316 (ImplicitNullCheckThreshold > 0) && 1317 method() != NULL && 1318 (method()->method_data()->trap_count(reason) 1319 >= (uint)ImplicitNullCheckThreshold)) { 1320 ok_prob = PROB_LIKELY_MAG(3); 1321 } 1322 1323 if (null_control != NULL) { 1324 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1325 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1326 set_control( _gvn.transform( new IfTrueNode(iff))); 1327 #ifndef PRODUCT 1328 if (null_true == top()) { 1329 explicit_null_checks_elided++; 1330 } 1331 #endif 1332 (*null_control) = null_true; 1333 } else { 1334 BuildCutout unless(this, tst, ok_prob); 1335 // Check for optimizer eliding test at parse time 1336 if (stopped()) { 1337 // Failure not possible; do not bother making uncommon trap. 1338 NOT_PRODUCT(explicit_null_checks_elided++); 1339 } else if (assert_null) { 1340 uncommon_trap(reason, 1341 Deoptimization::Action_make_not_entrant, 1342 NULL, "assert_null"); 1343 } else { 1344 replace_in_map(value, zerocon(type)); 1345 builtin_throw(reason); 1346 } 1347 } 1348 1349 // Must throw exception, fall-thru not possible? 1350 if (stopped()) { 1351 return top(); // No result 1352 } 1353 1354 if (assert_null) { 1355 // Cast obj to null on this path. 1356 replace_in_map(value, zerocon(type)); 1357 return zerocon(type); 1358 } 1359 1360 // Cast obj to not-null on this path, if there is no null_control. 1361 // (If there is a null_control, a non-null value may come back to haunt us.) 1362 if (type == T_OBJECT) { 1363 Node* cast = cast_not_null(value, false); 1364 if (null_control == NULL || (*null_control) == top()) 1365 replace_in_map(value, cast); 1366 value = cast; 1367 } 1368 1369 return value; 1370 } 1371 1372 1373 //------------------------------cast_not_null---------------------------------- 1374 // Cast obj to not-null on this path 1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1376 const Type *t = _gvn.type(obj); 1377 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1378 // Object is already not-null? 1379 if( t == t_not_null ) return obj; 1380 1381 Node *cast = new CastPPNode(obj,t_not_null); 1382 cast->init_req(0, control()); 1383 cast = _gvn.transform( cast ); 1384 1385 // Scan for instances of 'obj' in the current JVM mapping. 1386 // These instances are known to be not-null after the test. 1387 if (do_replace_in_map) 1388 replace_in_map(obj, cast); 1389 1390 return cast; // Return casted value 1391 } 1392 1393 // Sometimes in intrinsics, we implicitly know an object is not null 1394 // (there's no actual null check) so we can cast it to not null. In 1395 // the course of optimizations, the input to the cast can become null. 1396 // In that case that data path will die and we need the control path 1397 // to become dead as well to keep the graph consistent. So we have to 1398 // add a check for null for which one branch can't be taken. It uses 1399 // an Opaque4 node that will cause the check to be removed after loop 1400 // opts so the test goes away and the compiled code doesn't execute a 1401 // useless check. 1402 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1403 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1404 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1405 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1406 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1407 _gvn.set_type(iff, iff->Value(&_gvn)); 1408 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1409 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1410 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1411 C->root()->add_req(halt); 1412 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1413 set_control(if_t); 1414 return cast_not_null(value, do_replace_in_map); 1415 } 1416 1417 1418 //--------------------------replace_in_map------------------------------------- 1419 void GraphKit::replace_in_map(Node* old, Node* neww) { 1420 if (old == neww) { 1421 return; 1422 } 1423 1424 map()->replace_edge(old, neww); 1425 1426 // Note: This operation potentially replaces any edge 1427 // on the map. This includes locals, stack, and monitors 1428 // of the current (innermost) JVM state. 1429 1430 // don't let inconsistent types from profiling escape this 1431 // method 1432 1433 const Type* told = _gvn.type(old); 1434 const Type* tnew = _gvn.type(neww); 1435 1436 if (!tnew->higher_equal(told)) { 1437 return; 1438 } 1439 1440 map()->record_replaced_node(old, neww); 1441 } 1442 1443 1444 //============================================================================= 1445 //--------------------------------memory--------------------------------------- 1446 Node* GraphKit::memory(uint alias_idx) { 1447 MergeMemNode* mem = merged_memory(); 1448 Node* p = mem->memory_at(alias_idx); 1449 _gvn.set_type(p, Type::MEMORY); // must be mapped 1450 return p; 1451 } 1452 1453 //-----------------------------reset_memory------------------------------------ 1454 Node* GraphKit::reset_memory() { 1455 Node* mem = map()->memory(); 1456 // do not use this node for any more parsing! 1457 debug_only( map()->set_memory((Node*)NULL) ); 1458 return _gvn.transform( mem ); 1459 } 1460 1461 //------------------------------set_all_memory--------------------------------- 1462 void GraphKit::set_all_memory(Node* newmem) { 1463 Node* mergemem = MergeMemNode::make(newmem); 1464 gvn().set_type_bottom(mergemem); 1465 map()->set_memory(mergemem); 1466 } 1467 1468 //------------------------------set_all_memory_call---------------------------- 1469 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1470 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1471 set_all_memory(newmem); 1472 } 1473 1474 //============================================================================= 1475 // 1476 // parser factory methods for MemNodes 1477 // 1478 // These are layered on top of the factory methods in LoadNode and StoreNode, 1479 // and integrate with the parser's memory state and _gvn engine. 1480 // 1481 1482 // factory methods in "int adr_idx" 1483 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1484 int adr_idx, 1485 MemNode::MemOrd mo, 1486 LoadNode::ControlDependency control_dependency, 1487 bool require_atomic_access, 1488 bool unaligned, 1489 bool mismatched) { 1490 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1491 const TypePtr* adr_type = NULL; // debug-mode-only argument 1492 debug_only(adr_type = C->get_adr_type(adr_idx)); 1493 Node* mem = memory(adr_idx); 1494 Node* ld; 1495 if (require_atomic_access && bt == T_LONG) { 1496 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1497 } else if (require_atomic_access && bt == T_DOUBLE) { 1498 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1499 } else { 1500 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1501 } 1502 ld = _gvn.transform(ld); 1503 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1504 // Improve graph before escape analysis and boxing elimination. 1505 record_for_igvn(ld); 1506 } 1507 return ld; 1508 } 1509 1510 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1511 int adr_idx, 1512 MemNode::MemOrd mo, 1513 bool require_atomic_access, 1514 bool unaligned, 1515 bool mismatched) { 1516 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1517 const TypePtr* adr_type = NULL; 1518 debug_only(adr_type = C->get_adr_type(adr_idx)); 1519 Node *mem = memory(adr_idx); 1520 Node* st; 1521 if (require_atomic_access && bt == T_LONG) { 1522 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1523 } else if (require_atomic_access && bt == T_DOUBLE) { 1524 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1525 } else { 1526 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1527 } 1528 if (unaligned) { 1529 st->as_Store()->set_unaligned_access(); 1530 } 1531 if (mismatched) { 1532 st->as_Store()->set_mismatched_access(); 1533 } 1534 st = _gvn.transform(st); 1535 set_memory(st, adr_idx); 1536 // Back-to-back stores can only remove intermediate store with DU info 1537 // so push on worklist for optimizer. 1538 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1539 record_for_igvn(st); 1540 1541 return st; 1542 } 1543 1544 1545 void GraphKit::pre_barrier(bool do_load, 1546 Node* ctl, 1547 Node* obj, 1548 Node* adr, 1549 uint adr_idx, 1550 Node* val, 1551 const TypeOopPtr* val_type, 1552 Node* pre_val, 1553 BasicType bt) { 1554 1555 BarrierSet* bs = Universe::heap()->barrier_set(); 1556 set_control(ctl); 1557 switch (bs->kind()) { 1558 case BarrierSet::G1SATBCTLogging: 1559 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1560 break; 1561 1562 case BarrierSet::CardTableForRS: 1563 case BarrierSet::CardTableExtension: 1564 case BarrierSet::ModRef: 1565 break; 1566 1567 default : 1568 ShouldNotReachHere(); 1569 1570 } 1571 } 1572 1573 bool GraphKit::can_move_pre_barrier() const { 1574 BarrierSet* bs = Universe::heap()->barrier_set(); 1575 switch (bs->kind()) { 1576 case BarrierSet::G1SATBCTLogging: 1577 return true; // Can move it if no safepoint 1578 1579 case BarrierSet::CardTableForRS: 1580 case BarrierSet::CardTableExtension: 1581 case BarrierSet::ModRef: 1582 return true; // There is no pre-barrier 1583 1584 default : 1585 ShouldNotReachHere(); 1586 } 1587 return false; 1588 } 1589 1590 void GraphKit::post_barrier(Node* ctl, 1591 Node* store, 1592 Node* obj, 1593 Node* adr, 1594 uint adr_idx, 1595 Node* val, 1596 BasicType bt, 1597 bool use_precise) { 1598 BarrierSet* bs = Universe::heap()->barrier_set(); 1599 set_control(ctl); 1600 switch (bs->kind()) { 1601 case BarrierSet::G1SATBCTLogging: 1602 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1603 break; 1604 1605 case BarrierSet::CardTableForRS: 1606 case BarrierSet::CardTableExtension: 1607 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1608 break; 1609 1610 case BarrierSet::ModRef: 1611 break; 1612 1613 default : 1614 ShouldNotReachHere(); 1615 1616 } 1617 } 1618 1619 Node* GraphKit::store_oop(Node* ctl, 1620 Node* obj, 1621 Node* adr, 1622 const TypePtr* adr_type, 1623 Node* val, 1624 const TypeOopPtr* val_type, 1625 BasicType bt, 1626 bool use_precise, 1627 MemNode::MemOrd mo, 1628 bool mismatched) { 1629 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1630 // could be delayed during Parse (for example, in adjust_map_after_if()). 1631 // Execute transformation here to avoid barrier generation in such case. 1632 if (_gvn.type(val) == TypePtr::NULL_PTR) 1633 val = _gvn.makecon(TypePtr::NULL_PTR); 1634 1635 set_control(ctl); 1636 if (stopped()) return top(); // Dead path ? 1637 1638 assert(bt == T_OBJECT, "sanity"); 1639 assert(val != NULL, "not dead path"); 1640 uint adr_idx = C->get_alias_index(adr_type); 1641 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1642 1643 pre_barrier(true /* do_load */, 1644 control(), obj, adr, adr_idx, val, val_type, 1645 NULL /* pre_val */, 1646 bt); 1647 1648 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1649 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1650 return store; 1651 } 1652 1653 // Could be an array or object we don't know at compile time (unsafe ref.) 1654 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1655 Node* obj, // containing obj 1656 Node* adr, // actual adress to store val at 1657 const TypePtr* adr_type, 1658 Node* val, 1659 BasicType bt, 1660 MemNode::MemOrd mo, 1661 bool mismatched) { 1662 Compile::AliasType* at = C->alias_type(adr_type); 1663 const TypeOopPtr* val_type = NULL; 1664 if (adr_type->isa_instptr()) { 1665 if (at->field() != NULL) { 1666 // known field. This code is a copy of the do_put_xxx logic. 1667 ciField* field = at->field(); 1668 if (!field->type()->is_loaded()) { 1669 val_type = TypeInstPtr::BOTTOM; 1670 } else { 1671 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1672 } 1673 } 1674 } else if (adr_type->isa_aryptr()) { 1675 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1676 } 1677 if (val_type == NULL) { 1678 val_type = TypeInstPtr::BOTTOM; 1679 } 1680 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1681 } 1682 1683 1684 //-------------------------array_element_address------------------------- 1685 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1686 const TypeInt* sizetype, Node* ctrl) { 1687 uint shift = exact_log2(type2aelembytes(elembt)); 1688 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1689 1690 // short-circuit a common case (saves lots of confusing waste motion) 1691 jint idx_con = find_int_con(idx, -1); 1692 if (idx_con >= 0) { 1693 intptr_t offset = header + ((intptr_t)idx_con << shift); 1694 return basic_plus_adr(ary, offset); 1695 } 1696 1697 // must be correct type for alignment purposes 1698 Node* base = basic_plus_adr(ary, header); 1699 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1700 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1701 return basic_plus_adr(ary, base, scale); 1702 } 1703 1704 //-------------------------load_array_element------------------------- 1705 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1706 const Type* elemtype = arytype->elem(); 1707 BasicType elembt = elemtype->array_element_basic_type(); 1708 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1709 if (elembt == T_NARROWOOP) { 1710 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1711 } 1712 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1713 return ld; 1714 } 1715 1716 //-------------------------set_arguments_for_java_call------------------------- 1717 // Arguments (pre-popped from the stack) are taken from the JVMS. 1718 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1719 // Add the call arguments: 1720 uint nargs = call->method()->arg_size(); 1721 for (uint i = 0; i < nargs; i++) { 1722 Node* arg = argument(i); 1723 call->init_req(i + TypeFunc::Parms, arg); 1724 } 1725 } 1726 1727 //---------------------------set_edges_for_java_call--------------------------- 1728 // Connect a newly created call into the current JVMS. 1729 // A return value node (if any) is returned from set_edges_for_java_call. 1730 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1731 1732 // Add the predefined inputs: 1733 call->init_req( TypeFunc::Control, control() ); 1734 call->init_req( TypeFunc::I_O , i_o() ); 1735 call->init_req( TypeFunc::Memory , reset_memory() ); 1736 call->init_req( TypeFunc::FramePtr, frameptr() ); 1737 call->init_req( TypeFunc::ReturnAdr, top() ); 1738 1739 add_safepoint_edges(call, must_throw); 1740 1741 Node* xcall = _gvn.transform(call); 1742 1743 if (xcall == top()) { 1744 set_control(top()); 1745 return; 1746 } 1747 assert(xcall == call, "call identity is stable"); 1748 1749 // Re-use the current map to produce the result. 1750 1751 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1752 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1753 set_all_memory_call(xcall, separate_io_proj); 1754 1755 //return xcall; // no need, caller already has it 1756 } 1757 1758 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1759 if (stopped()) return top(); // maybe the call folded up? 1760 1761 // Capture the return value, if any. 1762 Node* ret; 1763 if (call->method() == NULL || 1764 call->method()->return_type()->basic_type() == T_VOID) 1765 ret = top(); 1766 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1767 1768 // Note: Since any out-of-line call can produce an exception, 1769 // we always insert an I_O projection from the call into the result. 1770 1771 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1772 1773 if (separate_io_proj) { 1774 // The caller requested separate projections be used by the fall 1775 // through and exceptional paths, so replace the projections for 1776 // the fall through path. 1777 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1778 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1779 } 1780 return ret; 1781 } 1782 1783 //--------------------set_predefined_input_for_runtime_call-------------------- 1784 // Reading and setting the memory state is way conservative here. 1785 // The real problem is that I am not doing real Type analysis on memory, 1786 // so I cannot distinguish card mark stores from other stores. Across a GC 1787 // point the Store Barrier and the card mark memory has to agree. I cannot 1788 // have a card mark store and its barrier split across the GC point from 1789 // either above or below. Here I get that to happen by reading ALL of memory. 1790 // A better answer would be to separate out card marks from other memory. 1791 // For now, return the input memory state, so that it can be reused 1792 // after the call, if this call has restricted memory effects. 1793 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1794 // Set fixed predefined input arguments 1795 Node* memory = reset_memory(); 1796 call->init_req( TypeFunc::Control, control() ); 1797 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1798 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1799 call->init_req( TypeFunc::FramePtr, frameptr() ); 1800 call->init_req( TypeFunc::ReturnAdr, top() ); 1801 return memory; 1802 } 1803 1804 //-------------------set_predefined_output_for_runtime_call-------------------- 1805 // Set control and memory (not i_o) from the call. 1806 // If keep_mem is not NULL, use it for the output state, 1807 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1808 // If hook_mem is NULL, this call produces no memory effects at all. 1809 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1810 // then only that memory slice is taken from the call. 1811 // In the last case, we must put an appropriate memory barrier before 1812 // the call, so as to create the correct anti-dependencies on loads 1813 // preceding the call. 1814 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1815 Node* keep_mem, 1816 const TypePtr* hook_mem) { 1817 // no i/o 1818 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1819 if (keep_mem) { 1820 // First clone the existing memory state 1821 set_all_memory(keep_mem); 1822 if (hook_mem != NULL) { 1823 // Make memory for the call 1824 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1825 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1826 // We also use hook_mem to extract specific effects from arraycopy stubs. 1827 set_memory(mem, hook_mem); 1828 } 1829 // ...else the call has NO memory effects. 1830 1831 // Make sure the call advertises its memory effects precisely. 1832 // This lets us build accurate anti-dependences in gcm.cpp. 1833 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1834 "call node must be constructed correctly"); 1835 } else { 1836 assert(hook_mem == NULL, ""); 1837 // This is not a "slow path" call; all memory comes from the call. 1838 set_all_memory_call(call); 1839 } 1840 } 1841 1842 1843 // Replace the call with the current state of the kit. 1844 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1845 JVMState* ejvms = NULL; 1846 if (has_exceptions()) { 1847 ejvms = transfer_exceptions_into_jvms(); 1848 } 1849 1850 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1851 ReplacedNodes replaced_nodes_exception; 1852 Node* ex_ctl = top(); 1853 1854 SafePointNode* final_state = stop(); 1855 1856 // Find all the needed outputs of this call 1857 CallProjections callprojs; 1858 call->extract_projections(&callprojs, true); 1859 1860 Node* init_mem = call->in(TypeFunc::Memory); 1861 Node* final_mem = final_state->in(TypeFunc::Memory); 1862 Node* final_ctl = final_state->in(TypeFunc::Control); 1863 Node* final_io = final_state->in(TypeFunc::I_O); 1864 1865 // Replace all the old call edges with the edges from the inlining result 1866 if (callprojs.fallthrough_catchproj != NULL) { 1867 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1868 } 1869 if (callprojs.fallthrough_memproj != NULL) { 1870 if (final_mem->is_MergeMem()) { 1871 // Parser's exits MergeMem was not transformed but may be optimized 1872 final_mem = _gvn.transform(final_mem); 1873 } 1874 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1875 } 1876 if (callprojs.fallthrough_ioproj != NULL) { 1877 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1878 } 1879 1880 // Replace the result with the new result if it exists and is used 1881 if (callprojs.resproj != NULL && result != NULL) { 1882 C->gvn_replace_by(callprojs.resproj, result); 1883 } 1884 1885 if (ejvms == NULL) { 1886 // No exception edges to simply kill off those paths 1887 if (callprojs.catchall_catchproj != NULL) { 1888 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1889 } 1890 if (callprojs.catchall_memproj != NULL) { 1891 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1892 } 1893 if (callprojs.catchall_ioproj != NULL) { 1894 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1895 } 1896 // Replace the old exception object with top 1897 if (callprojs.exobj != NULL) { 1898 C->gvn_replace_by(callprojs.exobj, C->top()); 1899 } 1900 } else { 1901 GraphKit ekit(ejvms); 1902 1903 // Load my combined exception state into the kit, with all phis transformed: 1904 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1905 replaced_nodes_exception = ex_map->replaced_nodes(); 1906 1907 Node* ex_oop = ekit.use_exception_state(ex_map); 1908 1909 if (callprojs.catchall_catchproj != NULL) { 1910 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1911 ex_ctl = ekit.control(); 1912 } 1913 if (callprojs.catchall_memproj != NULL) { 1914 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1915 } 1916 if (callprojs.catchall_ioproj != NULL) { 1917 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1918 } 1919 1920 // Replace the old exception object with the newly created one 1921 if (callprojs.exobj != NULL) { 1922 C->gvn_replace_by(callprojs.exobj, ex_oop); 1923 } 1924 } 1925 1926 // Disconnect the call from the graph 1927 call->disconnect_inputs(NULL, C); 1928 C->gvn_replace_by(call, C->top()); 1929 1930 // Clean up any MergeMems that feed other MergeMems since the 1931 // optimizer doesn't like that. 1932 if (final_mem->is_MergeMem()) { 1933 Node_List wl; 1934 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1935 Node* m = i.get(); 1936 if (m->is_MergeMem() && !wl.contains(m)) { 1937 wl.push(m); 1938 } 1939 } 1940 while (wl.size() > 0) { 1941 _gvn.transform(wl.pop()); 1942 } 1943 } 1944 1945 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1946 replaced_nodes.apply(C, final_ctl); 1947 } 1948 if (!ex_ctl->is_top() && do_replaced_nodes) { 1949 replaced_nodes_exception.apply(C, ex_ctl); 1950 } 1951 } 1952 1953 1954 //------------------------------increment_counter------------------------------ 1955 // for statistics: increment a VM counter by 1 1956 1957 void GraphKit::increment_counter(address counter_addr) { 1958 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1959 increment_counter(adr1); 1960 } 1961 1962 void GraphKit::increment_counter(Node* counter_addr) { 1963 int adr_type = Compile::AliasIdxRaw; 1964 Node* ctrl = control(); 1965 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1966 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1967 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1968 } 1969 1970 1971 //------------------------------uncommon_trap---------------------------------- 1972 // Bail out to the interpreter in mid-method. Implemented by calling the 1973 // uncommon_trap blob. This helper function inserts a runtime call with the 1974 // right debug info. 1975 void GraphKit::uncommon_trap(int trap_request, 1976 ciKlass* klass, const char* comment, 1977 bool must_throw, 1978 bool keep_exact_action) { 1979 if (failing()) stop(); 1980 if (stopped()) return; // trap reachable? 1981 1982 // Note: If ProfileTraps is true, and if a deopt. actually 1983 // occurs here, the runtime will make sure an MDO exists. There is 1984 // no need to call method()->ensure_method_data() at this point. 1985 1986 // Set the stack pointer to the right value for reexecution: 1987 set_sp(reexecute_sp()); 1988 1989 #ifdef ASSERT 1990 if (!must_throw) { 1991 // Make sure the stack has at least enough depth to execute 1992 // the current bytecode. 1993 int inputs, ignored_depth; 1994 if (compute_stack_effects(inputs, ignored_depth)) { 1995 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1996 Bytecodes::name(java_bc()), sp(), inputs); 1997 } 1998 } 1999 #endif 2000 2001 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2002 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2003 2004 switch (action) { 2005 case Deoptimization::Action_maybe_recompile: 2006 case Deoptimization::Action_reinterpret: 2007 // Temporary fix for 6529811 to allow virtual calls to be sure they 2008 // get the chance to go from mono->bi->mega 2009 if (!keep_exact_action && 2010 Deoptimization::trap_request_index(trap_request) < 0 && 2011 too_many_recompiles(reason)) { 2012 // This BCI is causing too many recompilations. 2013 if (C->log() != NULL) { 2014 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2015 Deoptimization::trap_reason_name(reason), 2016 Deoptimization::trap_action_name(action)); 2017 } 2018 action = Deoptimization::Action_none; 2019 trap_request = Deoptimization::make_trap_request(reason, action); 2020 } else { 2021 C->set_trap_can_recompile(true); 2022 } 2023 break; 2024 case Deoptimization::Action_make_not_entrant: 2025 C->set_trap_can_recompile(true); 2026 break; 2027 #ifdef ASSERT 2028 case Deoptimization::Action_none: 2029 case Deoptimization::Action_make_not_compilable: 2030 break; 2031 default: 2032 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2033 break; 2034 #endif 2035 } 2036 2037 if (TraceOptoParse) { 2038 char buf[100]; 2039 tty->print_cr("Uncommon trap %s at bci:%d", 2040 Deoptimization::format_trap_request(buf, sizeof(buf), 2041 trap_request), bci()); 2042 } 2043 2044 CompileLog* log = C->log(); 2045 if (log != NULL) { 2046 int kid = (klass == NULL)? -1: log->identify(klass); 2047 log->begin_elem("uncommon_trap bci='%d'", bci()); 2048 char buf[100]; 2049 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2050 trap_request)); 2051 if (kid >= 0) log->print(" klass='%d'", kid); 2052 if (comment != NULL) log->print(" comment='%s'", comment); 2053 log->end_elem(); 2054 } 2055 2056 // Make sure any guarding test views this path as very unlikely 2057 Node *i0 = control()->in(0); 2058 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2059 IfNode *iff = i0->as_If(); 2060 float f = iff->_prob; // Get prob 2061 if (control()->Opcode() == Op_IfTrue) { 2062 if (f > PROB_UNLIKELY_MAG(4)) 2063 iff->_prob = PROB_MIN; 2064 } else { 2065 if (f < PROB_LIKELY_MAG(4)) 2066 iff->_prob = PROB_MAX; 2067 } 2068 } 2069 2070 // Clear out dead values from the debug info. 2071 kill_dead_locals(); 2072 2073 // Now insert the uncommon trap subroutine call 2074 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2075 const TypePtr* no_memory_effects = NULL; 2076 // Pass the index of the class to be loaded 2077 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2078 (must_throw ? RC_MUST_THROW : 0), 2079 OptoRuntime::uncommon_trap_Type(), 2080 call_addr, "uncommon_trap", no_memory_effects, 2081 intcon(trap_request)); 2082 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2083 "must extract request correctly from the graph"); 2084 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2085 2086 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2087 // The debug info is the only real input to this call. 2088 2089 // Halt-and-catch fire here. The above call should never return! 2090 HaltNode* halt = new HaltNode(control(), frameptr()); 2091 _gvn.set_type_bottom(halt); 2092 root()->add_req(halt); 2093 2094 stop_and_kill_map(); 2095 } 2096 2097 2098 //--------------------------just_allocated_object------------------------------ 2099 // Report the object that was just allocated. 2100 // It must be the case that there are no intervening safepoints. 2101 // We use this to determine if an object is so "fresh" that 2102 // it does not require card marks. 2103 Node* GraphKit::just_allocated_object(Node* current_control) { 2104 if (C->recent_alloc_ctl() == current_control) 2105 return C->recent_alloc_obj(); 2106 return NULL; 2107 } 2108 2109 2110 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2111 // (Note: TypeFunc::make has a cache that makes this fast.) 2112 const TypeFunc* tf = TypeFunc::make(dest_method); 2113 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2114 for (int j = 0; j < nargs; j++) { 2115 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2116 if( targ->basic_type() == T_DOUBLE ) { 2117 // If any parameters are doubles, they must be rounded before 2118 // the call, dstore_rounding does gvn.transform 2119 Node *arg = argument(j); 2120 arg = dstore_rounding(arg); 2121 set_argument(j, arg); 2122 } 2123 } 2124 } 2125 2126 /** 2127 * Record profiling data exact_kls for Node n with the type system so 2128 * that it can propagate it (speculation) 2129 * 2130 * @param n node that the type applies to 2131 * @param exact_kls type from profiling 2132 * @param maybe_null did profiling see null? 2133 * 2134 * @return node with improved type 2135 */ 2136 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2137 const Type* current_type = _gvn.type(n); 2138 assert(UseTypeSpeculation, "type speculation must be on"); 2139 2140 const TypePtr* speculative = current_type->speculative(); 2141 2142 // Should the klass from the profile be recorded in the speculative type? 2143 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2144 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2145 const TypeOopPtr* xtype = tklass->as_instance_type(); 2146 assert(xtype->klass_is_exact(), "Should be exact"); 2147 // Any reason to believe n is not null (from this profiling or a previous one)? 2148 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2149 // record the new speculative type's depth 2150 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2151 speculative = speculative->with_inline_depth(jvms()->depth()); 2152 } else if (current_type->would_improve_ptr(maybe_null)) { 2153 // Profiling report that null was never seen so we can change the 2154 // speculative type to non null ptr. 2155 assert(!maybe_null, "nothing to improve"); 2156 if (speculative == NULL) { 2157 speculative = TypePtr::NOTNULL; 2158 } else { 2159 const TypePtr* ptr = TypePtr::NOTNULL; 2160 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2161 } 2162 } 2163 2164 if (speculative != current_type->speculative()) { 2165 // Build a type with a speculative type (what we think we know 2166 // about the type but will need a guard when we use it) 2167 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2168 // We're changing the type, we need a new CheckCast node to carry 2169 // the new type. The new type depends on the control: what 2170 // profiling tells us is only valid from here as far as we can 2171 // tell. 2172 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2173 cast = _gvn.transform(cast); 2174 replace_in_map(n, cast); 2175 n = cast; 2176 } 2177 2178 return n; 2179 } 2180 2181 /** 2182 * Record profiling data from receiver profiling at an invoke with the 2183 * type system so that it can propagate it (speculation) 2184 * 2185 * @param n receiver node 2186 * 2187 * @return node with improved type 2188 */ 2189 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2190 if (!UseTypeSpeculation) { 2191 return n; 2192 } 2193 ciKlass* exact_kls = profile_has_unique_klass(); 2194 bool maybe_null = true; 2195 if (java_bc() == Bytecodes::_checkcast || 2196 java_bc() == Bytecodes::_instanceof || 2197 java_bc() == Bytecodes::_aastore) { 2198 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2199 maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2200 } 2201 return record_profile_for_speculation(n, exact_kls, maybe_null); 2202 } 2203 2204 /** 2205 * Record profiling data from argument profiling at an invoke with the 2206 * type system so that it can propagate it (speculation) 2207 * 2208 * @param dest_method target method for the call 2209 * @param bc what invoke bytecode is this? 2210 */ 2211 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2212 if (!UseTypeSpeculation) { 2213 return; 2214 } 2215 const TypeFunc* tf = TypeFunc::make(dest_method); 2216 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2217 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2218 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2219 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2220 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2221 bool maybe_null = true; 2222 ciKlass* better_type = NULL; 2223 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2224 record_profile_for_speculation(argument(j), better_type, maybe_null); 2225 } 2226 i++; 2227 } 2228 } 2229 } 2230 2231 /** 2232 * Record profiling data from parameter profiling at an invoke with 2233 * the type system so that it can propagate it (speculation) 2234 */ 2235 void GraphKit::record_profiled_parameters_for_speculation() { 2236 if (!UseTypeSpeculation) { 2237 return; 2238 } 2239 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2240 if (_gvn.type(local(i))->isa_oopptr()) { 2241 bool maybe_null = true; 2242 ciKlass* better_type = NULL; 2243 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2244 record_profile_for_speculation(local(i), better_type, maybe_null); 2245 } 2246 j++; 2247 } 2248 } 2249 } 2250 2251 /** 2252 * Record profiling data from return value profiling at an invoke with 2253 * the type system so that it can propagate it (speculation) 2254 */ 2255 void GraphKit::record_profiled_return_for_speculation() { 2256 if (!UseTypeSpeculation) { 2257 return; 2258 } 2259 bool maybe_null = true; 2260 ciKlass* better_type = NULL; 2261 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2262 // If profiling reports a single type for the return value, 2263 // feed it to the type system so it can propagate it as a 2264 // speculative type 2265 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2266 } 2267 } 2268 2269 void GraphKit::round_double_result(ciMethod* dest_method) { 2270 // A non-strict method may return a double value which has an extended 2271 // exponent, but this must not be visible in a caller which is 'strict' 2272 // If a strict caller invokes a non-strict callee, round a double result 2273 2274 BasicType result_type = dest_method->return_type()->basic_type(); 2275 assert( method() != NULL, "must have caller context"); 2276 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2277 // Destination method's return value is on top of stack 2278 // dstore_rounding() does gvn.transform 2279 Node *result = pop_pair(); 2280 result = dstore_rounding(result); 2281 push_pair(result); 2282 } 2283 } 2284 2285 // rounding for strict float precision conformance 2286 Node* GraphKit::precision_rounding(Node* n) { 2287 return UseStrictFP && _method->flags().is_strict() 2288 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2289 ? _gvn.transform( new RoundFloatNode(0, n) ) 2290 : n; 2291 } 2292 2293 // rounding for strict double precision conformance 2294 Node* GraphKit::dprecision_rounding(Node *n) { 2295 return UseStrictFP && _method->flags().is_strict() 2296 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2297 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2298 : n; 2299 } 2300 2301 // rounding for non-strict double stores 2302 Node* GraphKit::dstore_rounding(Node* n) { 2303 return Matcher::strict_fp_requires_explicit_rounding 2304 && UseSSE <= 1 2305 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2306 : n; 2307 } 2308 2309 //============================================================================= 2310 // Generate a fast path/slow path idiom. Graph looks like: 2311 // [foo] indicates that 'foo' is a parameter 2312 // 2313 // [in] NULL 2314 // \ / 2315 // CmpP 2316 // Bool ne 2317 // If 2318 // / \ 2319 // True False-<2> 2320 // / | 2321 // / cast_not_null 2322 // Load | | ^ 2323 // [fast_test] | | 2324 // gvn to opt_test | | 2325 // / \ | <1> 2326 // True False | 2327 // | \\ | 2328 // [slow_call] \[fast_result] 2329 // Ctl Val \ \ 2330 // | \ \ 2331 // Catch <1> \ \ 2332 // / \ ^ \ \ 2333 // Ex No_Ex | \ \ 2334 // | \ \ | \ <2> \ 2335 // ... \ [slow_res] | | \ [null_result] 2336 // \ \--+--+--- | | 2337 // \ | / \ | / 2338 // --------Region Phi 2339 // 2340 //============================================================================= 2341 // Code is structured as a series of driver functions all called 'do_XXX' that 2342 // call a set of helper functions. Helper functions first, then drivers. 2343 2344 //------------------------------null_check_oop--------------------------------- 2345 // Null check oop. Set null-path control into Region in slot 3. 2346 // Make a cast-not-nullness use the other not-null control. Return cast. 2347 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2348 bool never_see_null, 2349 bool safe_for_replace, 2350 bool speculative) { 2351 // Initial NULL check taken path 2352 (*null_control) = top(); 2353 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2354 2355 // Generate uncommon_trap: 2356 if (never_see_null && (*null_control) != top()) { 2357 // If we see an unexpected null at a check-cast we record it and force a 2358 // recompile; the offending check-cast will be compiled to handle NULLs. 2359 // If we see more than one offending BCI, then all checkcasts in the 2360 // method will be compiled to handle NULLs. 2361 PreserveJVMState pjvms(this); 2362 set_control(*null_control); 2363 replace_in_map(value, null()); 2364 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2365 uncommon_trap(reason, 2366 Deoptimization::Action_make_not_entrant); 2367 (*null_control) = top(); // NULL path is dead 2368 } 2369 if ((*null_control) == top() && safe_for_replace) { 2370 replace_in_map(value, cast); 2371 } 2372 2373 // Cast away null-ness on the result 2374 return cast; 2375 } 2376 2377 //------------------------------opt_iff---------------------------------------- 2378 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2379 // Return slow-path control. 2380 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2381 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2382 2383 // Fast path taken; set region slot 2 2384 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2385 region->init_req(2,fast_taken); // Capture fast-control 2386 2387 // Fast path not-taken, i.e. slow path 2388 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2389 return slow_taken; 2390 } 2391 2392 //-----------------------------make_runtime_call------------------------------- 2393 Node* GraphKit::make_runtime_call(int flags, 2394 const TypeFunc* call_type, address call_addr, 2395 const char* call_name, 2396 const TypePtr* adr_type, 2397 // The following parms are all optional. 2398 // The first NULL ends the list. 2399 Node* parm0, Node* parm1, 2400 Node* parm2, Node* parm3, 2401 Node* parm4, Node* parm5, 2402 Node* parm6, Node* parm7) { 2403 // Slow-path call 2404 bool is_leaf = !(flags & RC_NO_LEAF); 2405 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2406 if (call_name == NULL) { 2407 assert(!is_leaf, "must supply name for leaf"); 2408 call_name = OptoRuntime::stub_name(call_addr); 2409 } 2410 CallNode* call; 2411 if (!is_leaf) { 2412 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2413 bci(), adr_type); 2414 } else if (flags & RC_NO_FP) { 2415 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2416 } else { 2417 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2418 } 2419 2420 // The following is similar to set_edges_for_java_call, 2421 // except that the memory effects of the call are restricted to AliasIdxRaw. 2422 2423 // Slow path call has no side-effects, uses few values 2424 bool wide_in = !(flags & RC_NARROW_MEM); 2425 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2426 2427 Node* prev_mem = NULL; 2428 if (wide_in) { 2429 prev_mem = set_predefined_input_for_runtime_call(call); 2430 } else { 2431 assert(!wide_out, "narrow in => narrow out"); 2432 Node* narrow_mem = memory(adr_type); 2433 prev_mem = reset_memory(); 2434 map()->set_memory(narrow_mem); 2435 set_predefined_input_for_runtime_call(call); 2436 } 2437 2438 // Hook each parm in order. Stop looking at the first NULL. 2439 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2440 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2441 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2442 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2443 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2444 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2445 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2446 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2447 /* close each nested if ===> */ } } } } } } } } 2448 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2449 2450 if (!is_leaf) { 2451 // Non-leaves can block and take safepoints: 2452 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2453 } 2454 // Non-leaves can throw exceptions: 2455 if (has_io) { 2456 call->set_req(TypeFunc::I_O, i_o()); 2457 } 2458 2459 if (flags & RC_UNCOMMON) { 2460 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2461 // (An "if" probability corresponds roughly to an unconditional count. 2462 // Sort of.) 2463 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2464 } 2465 2466 Node* c = _gvn.transform(call); 2467 assert(c == call, "cannot disappear"); 2468 2469 if (wide_out) { 2470 // Slow path call has full side-effects. 2471 set_predefined_output_for_runtime_call(call); 2472 } else { 2473 // Slow path call has few side-effects, and/or sets few values. 2474 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2475 } 2476 2477 if (has_io) { 2478 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2479 } 2480 return call; 2481 2482 } 2483 2484 //------------------------------merge_memory----------------------------------- 2485 // Merge memory from one path into the current memory state. 2486 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2487 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2488 Node* old_slice = mms.force_memory(); 2489 Node* new_slice = mms.memory2(); 2490 if (old_slice != new_slice) { 2491 PhiNode* phi; 2492 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2493 if (mms.is_empty()) { 2494 // clone base memory Phi's inputs for this memory slice 2495 assert(old_slice == mms.base_memory(), "sanity"); 2496 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2497 _gvn.set_type(phi, Type::MEMORY); 2498 for (uint i = 1; i < phi->req(); i++) { 2499 phi->init_req(i, old_slice->in(i)); 2500 } 2501 } else { 2502 phi = old_slice->as_Phi(); // Phi was generated already 2503 } 2504 } else { 2505 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2506 _gvn.set_type(phi, Type::MEMORY); 2507 } 2508 phi->set_req(new_path, new_slice); 2509 mms.set_memory(phi); 2510 } 2511 } 2512 } 2513 2514 //------------------------------make_slow_call_ex------------------------------ 2515 // Make the exception handler hookups for the slow call 2516 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2517 if (stopped()) return; 2518 2519 // Make a catch node with just two handlers: fall-through and catch-all 2520 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2521 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2522 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2523 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2524 2525 { PreserveJVMState pjvms(this); 2526 set_control(excp); 2527 set_i_o(i_o); 2528 2529 if (excp != top()) { 2530 if (deoptimize) { 2531 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2532 uncommon_trap(Deoptimization::Reason_unhandled, 2533 Deoptimization::Action_none); 2534 } else { 2535 // Create an exception state also. 2536 // Use an exact type if the caller has specified a specific exception. 2537 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2538 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2539 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2540 } 2541 } 2542 } 2543 2544 // Get the no-exception control from the CatchNode. 2545 set_control(norm); 2546 } 2547 2548 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2549 Node* cmp = NULL; 2550 switch(bt) { 2551 case T_INT: cmp = new CmpINode(in1, in2); break; 2552 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2553 default: fatal("unexpected comparison type %s", type2name(bt)); 2554 } 2555 gvn->transform(cmp); 2556 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2557 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2558 gvn->transform(iff); 2559 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2560 return iff; 2561 } 2562 2563 2564 //-------------------------------gen_subtype_check----------------------------- 2565 // Generate a subtyping check. Takes as input the subtype and supertype. 2566 // Returns 2 values: sets the default control() to the true path and returns 2567 // the false path. Only reads invariant memory; sets no (visible) memory. 2568 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2569 // but that's not exposed to the optimizer. This call also doesn't take in an 2570 // Object; if you wish to check an Object you need to load the Object's class 2571 // prior to coming here. 2572 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2573 Compile* C = gvn->C; 2574 2575 if ((*ctrl)->is_top()) { 2576 return C->top(); 2577 } 2578 2579 // Fast check for identical types, perhaps identical constants. 2580 // The types can even be identical non-constants, in cases 2581 // involving Array.newInstance, Object.clone, etc. 2582 if (subklass == superklass) 2583 return C->top(); // false path is dead; no test needed. 2584 2585 if (gvn->type(superklass)->singleton()) { 2586 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2587 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2588 2589 // In the common case of an exact superklass, try to fold up the 2590 // test before generating code. You may ask, why not just generate 2591 // the code and then let it fold up? The answer is that the generated 2592 // code will necessarily include null checks, which do not always 2593 // completely fold away. If they are also needless, then they turn 2594 // into a performance loss. Example: 2595 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2596 // Here, the type of 'fa' is often exact, so the store check 2597 // of fa[1]=x will fold up, without testing the nullness of x. 2598 switch (C->static_subtype_check(superk, subk)) { 2599 case Compile::SSC_always_false: 2600 { 2601 Node* always_fail = *ctrl; 2602 *ctrl = gvn->C->top(); 2603 return always_fail; 2604 } 2605 case Compile::SSC_always_true: 2606 return C->top(); 2607 case Compile::SSC_easy_test: 2608 { 2609 // Just do a direct pointer compare and be done. 2610 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2611 *ctrl = gvn->transform(new IfTrueNode(iff)); 2612 return gvn->transform(new IfFalseNode(iff)); 2613 } 2614 case Compile::SSC_full_test: 2615 break; 2616 default: 2617 ShouldNotReachHere(); 2618 } 2619 } 2620 2621 // %%% Possible further optimization: Even if the superklass is not exact, 2622 // if the subklass is the unique subtype of the superklass, the check 2623 // will always succeed. We could leave a dependency behind to ensure this. 2624 2625 // First load the super-klass's check-offset 2626 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2627 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2628 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2629 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2630 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2631 2632 // Load from the sub-klass's super-class display list, or a 1-word cache of 2633 // the secondary superclass list, or a failing value with a sentinel offset 2634 // if the super-klass is an interface or exceptionally deep in the Java 2635 // hierarchy and we have to scan the secondary superclass list the hard way. 2636 // Worst-case type is a little odd: NULL is allowed as a result (usually 2637 // klass loads can never produce a NULL). 2638 Node *chk_off_X = chk_off; 2639 #ifdef _LP64 2640 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2641 #endif 2642 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2643 // For some types like interfaces the following loadKlass is from a 1-word 2644 // cache which is mutable so can't use immutable memory. Other 2645 // types load from the super-class display table which is immutable. 2646 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2647 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2648 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2649 2650 // Compile speed common case: ARE a subtype and we canNOT fail 2651 if( superklass == nkls ) 2652 return C->top(); // false path is dead; no test needed. 2653 2654 // See if we get an immediate positive hit. Happens roughly 83% of the 2655 // time. Test to see if the value loaded just previously from the subklass 2656 // is exactly the superklass. 2657 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2658 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2659 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2660 2661 // Compile speed common case: Check for being deterministic right now. If 2662 // chk_off is a constant and not equal to cacheoff then we are NOT a 2663 // subklass. In this case we need exactly the 1 test above and we can 2664 // return those results immediately. 2665 if (!might_be_cache) { 2666 Node* not_subtype_ctrl = *ctrl; 2667 *ctrl = iftrue1; // We need exactly the 1 test above 2668 return not_subtype_ctrl; 2669 } 2670 2671 // Gather the various success & failures here 2672 RegionNode *r_ok_subtype = new RegionNode(4); 2673 gvn->record_for_igvn(r_ok_subtype); 2674 RegionNode *r_not_subtype = new RegionNode(3); 2675 gvn->record_for_igvn(r_not_subtype); 2676 2677 r_ok_subtype->init_req(1, iftrue1); 2678 2679 // Check for immediate negative hit. Happens roughly 11% of the time (which 2680 // is roughly 63% of the remaining cases). Test to see if the loaded 2681 // check-offset points into the subklass display list or the 1-element 2682 // cache. If it points to the display (and NOT the cache) and the display 2683 // missed then it's not a subtype. 2684 Node *cacheoff = gvn->intcon(cacheoff_con); 2685 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2686 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2687 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2688 2689 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2690 // No performance impact (too rare) but allows sharing of secondary arrays 2691 // which has some footprint reduction. 2692 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2693 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2694 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2695 2696 // -- Roads not taken here: -- 2697 // We could also have chosen to perform the self-check at the beginning 2698 // of this code sequence, as the assembler does. This would not pay off 2699 // the same way, since the optimizer, unlike the assembler, can perform 2700 // static type analysis to fold away many successful self-checks. 2701 // Non-foldable self checks work better here in second position, because 2702 // the initial primary superclass check subsumes a self-check for most 2703 // types. An exception would be a secondary type like array-of-interface, 2704 // which does not appear in its own primary supertype display. 2705 // Finally, we could have chosen to move the self-check into the 2706 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2707 // dependent manner. But it is worthwhile to have the check here, 2708 // where it can be perhaps be optimized. The cost in code space is 2709 // small (register compare, branch). 2710 2711 // Now do a linear scan of the secondary super-klass array. Again, no real 2712 // performance impact (too rare) but it's gotta be done. 2713 // Since the code is rarely used, there is no penalty for moving it 2714 // out of line, and it can only improve I-cache density. 2715 // The decision to inline or out-of-line this final check is platform 2716 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2717 Node* psc = gvn->transform( 2718 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2719 2720 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2721 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2722 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2723 2724 // Return false path; set default control to true path. 2725 *ctrl = gvn->transform(r_ok_subtype); 2726 return gvn->transform(r_not_subtype); 2727 } 2728 2729 // Profile-driven exact type check: 2730 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2731 float prob, 2732 Node* *casted_receiver) { 2733 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2734 Node* recv_klass = load_object_klass(receiver); 2735 Node* want_klass = makecon(tklass); 2736 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2737 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2738 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2739 set_control( _gvn.transform( new IfTrueNode (iff) )); 2740 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2741 2742 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2743 assert(recv_xtype->klass_is_exact(), ""); 2744 2745 // Subsume downstream occurrences of receiver with a cast to 2746 // recv_xtype, since now we know what the type will be. 2747 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2748 (*casted_receiver) = _gvn.transform(cast); 2749 // (User must make the replace_in_map call.) 2750 2751 return fail; 2752 } 2753 2754 2755 //------------------------------seems_never_null------------------------------- 2756 // Use null_seen information if it is available from the profile. 2757 // If we see an unexpected null at a type check we record it and force a 2758 // recompile; the offending check will be recompiled to handle NULLs. 2759 // If we see several offending BCIs, then all checks in the 2760 // method will be recompiled. 2761 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2762 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2763 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2764 if (UncommonNullCast // Cutout for this technique 2765 && obj != null() // And not the -Xcomp stupid case? 2766 && !too_many_traps(reason) 2767 ) { 2768 if (speculating) { 2769 return true; 2770 } 2771 if (data == NULL) 2772 // Edge case: no mature data. Be optimistic here. 2773 return true; 2774 // If the profile has not seen a null, assume it won't happen. 2775 assert(java_bc() == Bytecodes::_checkcast || 2776 java_bc() == Bytecodes::_instanceof || 2777 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2778 return !data->as_BitData()->null_seen(); 2779 } 2780 speculating = false; 2781 return false; 2782 } 2783 2784 //------------------------maybe_cast_profiled_receiver------------------------- 2785 // If the profile has seen exactly one type, narrow to exactly that type. 2786 // Subsequent type checks will always fold up. 2787 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2788 ciKlass* require_klass, 2789 ciKlass* spec_klass, 2790 bool safe_for_replace) { 2791 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2792 2793 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2794 2795 // Make sure we haven't already deoptimized from this tactic. 2796 if (too_many_traps(reason) || too_many_recompiles(reason)) 2797 return NULL; 2798 2799 // (No, this isn't a call, but it's enough like a virtual call 2800 // to use the same ciMethod accessor to get the profile info...) 2801 // If we have a speculative type use it instead of profiling (which 2802 // may not help us) 2803 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2804 if (exact_kls != NULL) {// no cast failures here 2805 if (require_klass == NULL || 2806 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2807 // If we narrow the type to match what the type profile sees or 2808 // the speculative type, we can then remove the rest of the 2809 // cast. 2810 // This is a win, even if the exact_kls is very specific, 2811 // because downstream operations, such as method calls, 2812 // will often benefit from the sharper type. 2813 Node* exact_obj = not_null_obj; // will get updated in place... 2814 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2815 &exact_obj); 2816 { PreserveJVMState pjvms(this); 2817 set_control(slow_ctl); 2818 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2819 } 2820 if (safe_for_replace) { 2821 replace_in_map(not_null_obj, exact_obj); 2822 } 2823 return exact_obj; 2824 } 2825 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2826 } 2827 2828 return NULL; 2829 } 2830 2831 /** 2832 * Cast obj to type and emit guard unless we had too many traps here 2833 * already 2834 * 2835 * @param obj node being casted 2836 * @param type type to cast the node to 2837 * @param not_null true if we know node cannot be null 2838 */ 2839 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2840 ciKlass* type, 2841 bool not_null) { 2842 if (stopped()) { 2843 return obj; 2844 } 2845 2846 // type == NULL if profiling tells us this object is always null 2847 if (type != NULL) { 2848 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2849 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2850 2851 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2852 !too_many_traps(class_reason) && 2853 !too_many_recompiles(class_reason)) { 2854 Node* not_null_obj = NULL; 2855 // not_null is true if we know the object is not null and 2856 // there's no need for a null check 2857 if (!not_null) { 2858 Node* null_ctl = top(); 2859 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2860 assert(null_ctl->is_top(), "no null control here"); 2861 } else { 2862 not_null_obj = obj; 2863 } 2864 2865 Node* exact_obj = not_null_obj; 2866 ciKlass* exact_kls = type; 2867 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2868 &exact_obj); 2869 { 2870 PreserveJVMState pjvms(this); 2871 set_control(slow_ctl); 2872 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2873 } 2874 replace_in_map(not_null_obj, exact_obj); 2875 obj = exact_obj; 2876 } 2877 } else { 2878 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2879 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2880 Node* exact_obj = null_assert(obj); 2881 replace_in_map(obj, exact_obj); 2882 obj = exact_obj; 2883 } 2884 } 2885 return obj; 2886 } 2887 2888 //-------------------------------gen_instanceof-------------------------------- 2889 // Generate an instance-of idiom. Used by both the instance-of bytecode 2890 // and the reflective instance-of call. 2891 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2892 kill_dead_locals(); // Benefit all the uncommon traps 2893 assert( !stopped(), "dead parse path should be checked in callers" ); 2894 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2895 "must check for not-null not-dead klass in callers"); 2896 2897 // Make the merge point 2898 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2899 RegionNode* region = new RegionNode(PATH_LIMIT); 2900 Node* phi = new PhiNode(region, TypeInt::BOOL); 2901 C->set_has_split_ifs(true); // Has chance for split-if optimization 2902 2903 ciProfileData* data = NULL; 2904 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2905 data = method()->method_data()->bci_to_data(bci()); 2906 } 2907 bool speculative_not_null = false; 2908 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2909 && seems_never_null(obj, data, speculative_not_null)); 2910 2911 // Null check; get casted pointer; set region slot 3 2912 Node* null_ctl = top(); 2913 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2914 2915 // If not_null_obj is dead, only null-path is taken 2916 if (stopped()) { // Doing instance-of on a NULL? 2917 set_control(null_ctl); 2918 return intcon(0); 2919 } 2920 region->init_req(_null_path, null_ctl); 2921 phi ->init_req(_null_path, intcon(0)); // Set null path value 2922 if (null_ctl == top()) { 2923 // Do this eagerly, so that pattern matches like is_diamond_phi 2924 // will work even during parsing. 2925 assert(_null_path == PATH_LIMIT-1, "delete last"); 2926 region->del_req(_null_path); 2927 phi ->del_req(_null_path); 2928 } 2929 2930 // Do we know the type check always succeed? 2931 bool known_statically = false; 2932 if (_gvn.type(superklass)->singleton()) { 2933 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2934 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2935 if (subk != NULL && subk->is_loaded()) { 2936 int static_res = C->static_subtype_check(superk, subk); 2937 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2938 } 2939 } 2940 2941 if (known_statically && UseTypeSpeculation) { 2942 // If we know the type check always succeeds then we don't use the 2943 // profiling data at this bytecode. Don't lose it, feed it to the 2944 // type system as a speculative type. 2945 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2946 } else { 2947 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2948 // We may not have profiling here or it may not help us. If we 2949 // have a speculative type use it to perform an exact cast. 2950 ciKlass* spec_obj_type = obj_type->speculative_type(); 2951 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2952 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2953 if (stopped()) { // Profile disagrees with this path. 2954 set_control(null_ctl); // Null is the only remaining possibility. 2955 return intcon(0); 2956 } 2957 if (cast_obj != NULL) { 2958 not_null_obj = cast_obj; 2959 } 2960 } 2961 } 2962 2963 // Load the object's klass 2964 Node* obj_klass = load_object_klass(not_null_obj); 2965 2966 // Generate the subtype check 2967 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2968 2969 // Plug in the success path to the general merge in slot 1. 2970 region->init_req(_obj_path, control()); 2971 phi ->init_req(_obj_path, intcon(1)); 2972 2973 // Plug in the failing path to the general merge in slot 2. 2974 region->init_req(_fail_path, not_subtype_ctrl); 2975 phi ->init_req(_fail_path, intcon(0)); 2976 2977 // Return final merged results 2978 set_control( _gvn.transform(region) ); 2979 record_for_igvn(region); 2980 return _gvn.transform(phi); 2981 } 2982 2983 //-------------------------------gen_checkcast--------------------------------- 2984 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2985 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2986 // uncommon-trap paths work. Adjust stack after this call. 2987 // If failure_control is supplied and not null, it is filled in with 2988 // the control edge for the cast failure. Otherwise, an appropriate 2989 // uncommon trap or exception is thrown. 2990 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2991 Node* *failure_control) { 2992 kill_dead_locals(); // Benefit all the uncommon traps 2993 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2994 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2995 2996 // Fast cutout: Check the case that the cast is vacuously true. 2997 // This detects the common cases where the test will short-circuit 2998 // away completely. We do this before we perform the null check, 2999 // because if the test is going to turn into zero code, we don't 3000 // want a residual null check left around. (Causes a slowdown, 3001 // for example, in some objArray manipulations, such as a[i]=a[j].) 3002 if (tk->singleton()) { 3003 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3004 if (objtp != NULL && objtp->klass() != NULL) { 3005 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3006 case Compile::SSC_always_true: 3007 // If we know the type check always succeed then we don't use 3008 // the profiling data at this bytecode. Don't lose it, feed it 3009 // to the type system as a speculative type. 3010 return record_profiled_receiver_for_speculation(obj); 3011 case Compile::SSC_always_false: 3012 // It needs a null check because a null will *pass* the cast check. 3013 // A non-null value will always produce an exception. 3014 return null_assert(obj); 3015 } 3016 } 3017 } 3018 3019 ciProfileData* data = NULL; 3020 bool safe_for_replace = false; 3021 if (failure_control == NULL) { // use MDO in regular case only 3022 assert(java_bc() == Bytecodes::_aastore || 3023 java_bc() == Bytecodes::_checkcast, 3024 "interpreter profiles type checks only for these BCs"); 3025 data = method()->method_data()->bci_to_data(bci()); 3026 safe_for_replace = true; 3027 } 3028 3029 // Make the merge point 3030 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3031 RegionNode* region = new RegionNode(PATH_LIMIT); 3032 Node* phi = new PhiNode(region, toop); 3033 C->set_has_split_ifs(true); // Has chance for split-if optimization 3034 3035 // Use null-cast information if it is available 3036 bool speculative_not_null = false; 3037 bool never_see_null = ((failure_control == NULL) // regular case only 3038 && seems_never_null(obj, data, speculative_not_null)); 3039 3040 // Null check; get casted pointer; set region slot 3 3041 Node* null_ctl = top(); 3042 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3043 3044 // If not_null_obj is dead, only null-path is taken 3045 if (stopped()) { // Doing instance-of on a NULL? 3046 set_control(null_ctl); 3047 return null(); 3048 } 3049 region->init_req(_null_path, null_ctl); 3050 phi ->init_req(_null_path, null()); // Set null path value 3051 if (null_ctl == top()) { 3052 // Do this eagerly, so that pattern matches like is_diamond_phi 3053 // will work even during parsing. 3054 assert(_null_path == PATH_LIMIT-1, "delete last"); 3055 region->del_req(_null_path); 3056 phi ->del_req(_null_path); 3057 } 3058 3059 Node* cast_obj = NULL; 3060 if (tk->klass_is_exact()) { 3061 // The following optimization tries to statically cast the speculative type of the object 3062 // (for example obtained during profiling) to the type of the superklass and then do a 3063 // dynamic check that the type of the object is what we expect. To work correctly 3064 // for checkcast and aastore the type of superklass should be exact. 3065 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3066 // We may not have profiling here or it may not help us. If we have 3067 // a speculative type use it to perform an exact cast. 3068 ciKlass* spec_obj_type = obj_type->speculative_type(); 3069 if (spec_obj_type != NULL || data != NULL) { 3070 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3071 if (cast_obj != NULL) { 3072 if (failure_control != NULL) // failure is now impossible 3073 (*failure_control) = top(); 3074 // adjust the type of the phi to the exact klass: 3075 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3076 } 3077 } 3078 } 3079 3080 if (cast_obj == NULL) { 3081 // Load the object's klass 3082 Node* obj_klass = load_object_klass(not_null_obj); 3083 3084 // Generate the subtype check 3085 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3086 3087 // Plug in success path into the merge 3088 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3089 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3090 if (failure_control == NULL) { 3091 if (not_subtype_ctrl != top()) { // If failure is possible 3092 PreserveJVMState pjvms(this); 3093 set_control(not_subtype_ctrl); 3094 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3095 } 3096 } else { 3097 (*failure_control) = not_subtype_ctrl; 3098 } 3099 } 3100 3101 region->init_req(_obj_path, control()); 3102 phi ->init_req(_obj_path, cast_obj); 3103 3104 // A merge of NULL or Casted-NotNull obj 3105 Node* res = _gvn.transform(phi); 3106 3107 // Note I do NOT always 'replace_in_map(obj,result)' here. 3108 // if( tk->klass()->can_be_primary_super() ) 3109 // This means that if I successfully store an Object into an array-of-String 3110 // I 'forget' that the Object is really now known to be a String. I have to 3111 // do this because we don't have true union types for interfaces - if I store 3112 // a Baz into an array-of-Interface and then tell the optimizer it's an 3113 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3114 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3115 // replace_in_map( obj, res ); 3116 3117 // Return final merged results 3118 set_control( _gvn.transform(region) ); 3119 record_for_igvn(region); 3120 return res; 3121 } 3122 3123 //------------------------------next_monitor----------------------------------- 3124 // What number should be given to the next monitor? 3125 int GraphKit::next_monitor() { 3126 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3127 int next = current + C->sync_stack_slots(); 3128 // Keep the toplevel high water mark current: 3129 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3130 return current; 3131 } 3132 3133 //------------------------------insert_mem_bar--------------------------------- 3134 // Memory barrier to avoid floating things around 3135 // The membar serves as a pinch point between both control and all memory slices. 3136 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3137 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3138 mb->init_req(TypeFunc::Control, control()); 3139 mb->init_req(TypeFunc::Memory, reset_memory()); 3140 Node* membar = _gvn.transform(mb); 3141 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3142 set_all_memory_call(membar); 3143 return membar; 3144 } 3145 3146 //-------------------------insert_mem_bar_volatile---------------------------- 3147 // Memory barrier to avoid floating things around 3148 // The membar serves as a pinch point between both control and memory(alias_idx). 3149 // If you want to make a pinch point on all memory slices, do not use this 3150 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3151 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3152 // When Parse::do_put_xxx updates a volatile field, it appends a series 3153 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3154 // The first membar is on the same memory slice as the field store opcode. 3155 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3156 // All the other membars (for other volatile slices, including AliasIdxBot, 3157 // which stands for all unknown volatile slices) are control-dependent 3158 // on the first membar. This prevents later volatile loads or stores 3159 // from sliding up past the just-emitted store. 3160 3161 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3162 mb->set_req(TypeFunc::Control,control()); 3163 if (alias_idx == Compile::AliasIdxBot) { 3164 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3165 } else { 3166 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3167 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3168 } 3169 Node* membar = _gvn.transform(mb); 3170 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3171 if (alias_idx == Compile::AliasIdxBot) { 3172 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3173 } else { 3174 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3175 } 3176 return membar; 3177 } 3178 3179 //------------------------------shared_lock------------------------------------ 3180 // Emit locking code. 3181 FastLockNode* GraphKit::shared_lock(Node* obj) { 3182 // bci is either a monitorenter bc or InvocationEntryBci 3183 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3184 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3185 3186 if( !GenerateSynchronizationCode ) 3187 return NULL; // Not locking things? 3188 if (stopped()) // Dead monitor? 3189 return NULL; 3190 3191 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3192 3193 // Box the stack location 3194 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3195 Node* mem = reset_memory(); 3196 3197 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3198 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3199 // Create the counters for this fast lock. 3200 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3201 } 3202 3203 // Create the rtm counters for this fast lock if needed. 3204 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3205 3206 // Add monitor to debug info for the slow path. If we block inside the 3207 // slow path and de-opt, we need the monitor hanging around 3208 map()->push_monitor( flock ); 3209 3210 const TypeFunc *tf = LockNode::lock_type(); 3211 LockNode *lock = new LockNode(C, tf); 3212 3213 lock->init_req( TypeFunc::Control, control() ); 3214 lock->init_req( TypeFunc::Memory , mem ); 3215 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3216 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3217 lock->init_req( TypeFunc::ReturnAdr, top() ); 3218 3219 lock->init_req(TypeFunc::Parms + 0, obj); 3220 lock->init_req(TypeFunc::Parms + 1, box); 3221 lock->init_req(TypeFunc::Parms + 2, flock); 3222 add_safepoint_edges(lock); 3223 3224 lock = _gvn.transform( lock )->as_Lock(); 3225 3226 // lock has no side-effects, sets few values 3227 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3228 3229 insert_mem_bar(Op_MemBarAcquireLock); 3230 3231 // Add this to the worklist so that the lock can be eliminated 3232 record_for_igvn(lock); 3233 3234 #ifndef PRODUCT 3235 if (PrintLockStatistics) { 3236 // Update the counter for this lock. Don't bother using an atomic 3237 // operation since we don't require absolute accuracy. 3238 lock->create_lock_counter(map()->jvms()); 3239 increment_counter(lock->counter()->addr()); 3240 } 3241 #endif 3242 3243 return flock; 3244 } 3245 3246 3247 //------------------------------shared_unlock---------------------------------- 3248 // Emit unlocking code. 3249 void GraphKit::shared_unlock(Node* box, Node* obj) { 3250 // bci is either a monitorenter bc or InvocationEntryBci 3251 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3252 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3253 3254 if( !GenerateSynchronizationCode ) 3255 return; 3256 if (stopped()) { // Dead monitor? 3257 map()->pop_monitor(); // Kill monitor from debug info 3258 return; 3259 } 3260 3261 // Memory barrier to avoid floating things down past the locked region 3262 insert_mem_bar(Op_MemBarReleaseLock); 3263 3264 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3265 UnlockNode *unlock = new UnlockNode(C, tf); 3266 #ifdef ASSERT 3267 unlock->set_dbg_jvms(sync_jvms()); 3268 #endif 3269 uint raw_idx = Compile::AliasIdxRaw; 3270 unlock->init_req( TypeFunc::Control, control() ); 3271 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3272 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3273 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3274 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3275 3276 unlock->init_req(TypeFunc::Parms + 0, obj); 3277 unlock->init_req(TypeFunc::Parms + 1, box); 3278 unlock = _gvn.transform(unlock)->as_Unlock(); 3279 3280 Node* mem = reset_memory(); 3281 3282 // unlock has no side-effects, sets few values 3283 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3284 3285 // Kill monitor from debug info 3286 map()->pop_monitor( ); 3287 } 3288 3289 //-------------------------------get_layout_helper----------------------------- 3290 // If the given klass is a constant or known to be an array, 3291 // fetch the constant layout helper value into constant_value 3292 // and return (Node*)NULL. Otherwise, load the non-constant 3293 // layout helper value, and return the node which represents it. 3294 // This two-faced routine is useful because allocation sites 3295 // almost always feature constant types. 3296 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3297 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3298 if (!StressReflectiveCode && inst_klass != NULL) { 3299 ciKlass* klass = inst_klass->klass(); 3300 bool xklass = inst_klass->klass_is_exact(); 3301 if (xklass || klass->is_array_klass()) { 3302 jint lhelper = klass->layout_helper(); 3303 if (lhelper != Klass::_lh_neutral_value) { 3304 constant_value = lhelper; 3305 return (Node*) NULL; 3306 } 3307 } 3308 } 3309 constant_value = Klass::_lh_neutral_value; // put in a known value 3310 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3311 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3312 } 3313 3314 // We just put in an allocate/initialize with a big raw-memory effect. 3315 // Hook selected additional alias categories on the initialization. 3316 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3317 MergeMemNode* init_in_merge, 3318 Node* init_out_raw) { 3319 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3320 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3321 3322 Node* prevmem = kit.memory(alias_idx); 3323 init_in_merge->set_memory_at(alias_idx, prevmem); 3324 kit.set_memory(init_out_raw, alias_idx); 3325 } 3326 3327 //---------------------------set_output_for_allocation------------------------- 3328 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3329 const TypeOopPtr* oop_type, 3330 bool deoptimize_on_exception) { 3331 int rawidx = Compile::AliasIdxRaw; 3332 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3333 add_safepoint_edges(alloc); 3334 Node* allocx = _gvn.transform(alloc); 3335 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3336 // create memory projection for i_o 3337 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3338 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3339 3340 // create a memory projection as for the normal control path 3341 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3342 set_memory(malloc, rawidx); 3343 3344 // a normal slow-call doesn't change i_o, but an allocation does 3345 // we create a separate i_o projection for the normal control path 3346 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3347 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3348 3349 // put in an initialization barrier 3350 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3351 rawoop)->as_Initialize(); 3352 assert(alloc->initialization() == init, "2-way macro link must work"); 3353 assert(init ->allocation() == alloc, "2-way macro link must work"); 3354 { 3355 // Extract memory strands which may participate in the new object's 3356 // initialization, and source them from the new InitializeNode. 3357 // This will allow us to observe initializations when they occur, 3358 // and link them properly (as a group) to the InitializeNode. 3359 assert(init->in(InitializeNode::Memory) == malloc, ""); 3360 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3361 init->set_req(InitializeNode::Memory, minit_in); 3362 record_for_igvn(minit_in); // fold it up later, if possible 3363 Node* minit_out = memory(rawidx); 3364 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3365 if (oop_type->isa_aryptr()) { 3366 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3367 int elemidx = C->get_alias_index(telemref); 3368 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3369 } else if (oop_type->isa_instptr()) { 3370 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3371 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3372 ciField* field = ik->nonstatic_field_at(i); 3373 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3374 continue; // do not bother to track really large numbers of fields 3375 // Find (or create) the alias category for this field: 3376 int fieldidx = C->alias_type(field)->index(); 3377 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3378 } 3379 } 3380 } 3381 3382 // Cast raw oop to the real thing... 3383 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3384 javaoop = _gvn.transform(javaoop); 3385 C->set_recent_alloc(control(), javaoop); 3386 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3387 3388 #ifdef ASSERT 3389 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3390 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3391 "Ideal_allocation works"); 3392 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3393 "Ideal_allocation works"); 3394 if (alloc->is_AllocateArray()) { 3395 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3396 "Ideal_allocation works"); 3397 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3398 "Ideal_allocation works"); 3399 } else { 3400 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3401 } 3402 } 3403 #endif //ASSERT 3404 3405 return javaoop; 3406 } 3407 3408 //---------------------------new_instance-------------------------------------- 3409 // This routine takes a klass_node which may be constant (for a static type) 3410 // or may be non-constant (for reflective code). It will work equally well 3411 // for either, and the graph will fold nicely if the optimizer later reduces 3412 // the type to a constant. 3413 // The optional arguments are for specialized use by intrinsics: 3414 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3415 // - If 'return_size_val', report the the total object size to the caller. 3416 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3417 Node* GraphKit::new_instance(Node* klass_node, 3418 Node* extra_slow_test, 3419 Node* *return_size_val, 3420 bool deoptimize_on_exception) { 3421 // Compute size in doublewords 3422 // The size is always an integral number of doublewords, represented 3423 // as a positive bytewise size stored in the klass's layout_helper. 3424 // The layout_helper also encodes (in a low bit) the need for a slow path. 3425 jint layout_con = Klass::_lh_neutral_value; 3426 Node* layout_val = get_layout_helper(klass_node, layout_con); 3427 int layout_is_con = (layout_val == NULL); 3428 3429 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3430 // Generate the initial go-slow test. It's either ALWAYS (return a 3431 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3432 // case) a computed value derived from the layout_helper. 3433 Node* initial_slow_test = NULL; 3434 if (layout_is_con) { 3435 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3436 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3437 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3438 } else { // reflective case 3439 // This reflective path is used by Unsafe.allocateInstance. 3440 // (It may be stress-tested by specifying StressReflectiveCode.) 3441 // Basically, we want to get into the VM is there's an illegal argument. 3442 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3443 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3444 if (extra_slow_test != intcon(0)) { 3445 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3446 } 3447 // (Macro-expander will further convert this to a Bool, if necessary.) 3448 } 3449 3450 // Find the size in bytes. This is easy; it's the layout_helper. 3451 // The size value must be valid even if the slow path is taken. 3452 Node* size = NULL; 3453 if (layout_is_con) { 3454 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3455 } else { // reflective case 3456 // This reflective path is used by clone and Unsafe.allocateInstance. 3457 size = ConvI2X(layout_val); 3458 3459 // Clear the low bits to extract layout_helper_size_in_bytes: 3460 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3461 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3462 size = _gvn.transform( new AndXNode(size, mask) ); 3463 } 3464 if (return_size_val != NULL) { 3465 (*return_size_val) = size; 3466 } 3467 3468 // This is a precise notnull oop of the klass. 3469 // (Actually, it need not be precise if this is a reflective allocation.) 3470 // It's what we cast the result to. 3471 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3472 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3473 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3474 3475 // Now generate allocation code 3476 3477 // The entire memory state is needed for slow path of the allocation 3478 // since GC and deoptimization can happened. 3479 Node *mem = reset_memory(); 3480 set_all_memory(mem); // Create new memory state 3481 3482 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3483 control(), mem, i_o(), 3484 size, klass_node, 3485 initial_slow_test); 3486 3487 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3488 } 3489 3490 //-------------------------------new_array------------------------------------- 3491 // helper for both newarray and anewarray 3492 // The 'length' parameter is (obviously) the length of the array. 3493 // See comments on new_instance for the meaning of the other arguments. 3494 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3495 Node* length, // number of array elements 3496 int nargs, // number of arguments to push back for uncommon trap 3497 Node* *return_size_val, 3498 bool deoptimize_on_exception) { 3499 jint layout_con = Klass::_lh_neutral_value; 3500 Node* layout_val = get_layout_helper(klass_node, layout_con); 3501 int layout_is_con = (layout_val == NULL); 3502 3503 if (!layout_is_con && !StressReflectiveCode && 3504 !too_many_traps(Deoptimization::Reason_class_check)) { 3505 // This is a reflective array creation site. 3506 // Optimistically assume that it is a subtype of Object[], 3507 // so that we can fold up all the address arithmetic. 3508 layout_con = Klass::array_layout_helper(T_OBJECT); 3509 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3510 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3511 { BuildCutout unless(this, bol_lh, PROB_MAX); 3512 inc_sp(nargs); 3513 uncommon_trap(Deoptimization::Reason_class_check, 3514 Deoptimization::Action_maybe_recompile); 3515 } 3516 layout_val = NULL; 3517 layout_is_con = true; 3518 } 3519 3520 // Generate the initial go-slow test. Make sure we do not overflow 3521 // if length is huge (near 2Gig) or negative! We do not need 3522 // exact double-words here, just a close approximation of needed 3523 // double-words. We can't add any offset or rounding bits, lest we 3524 // take a size -1 of bytes and make it positive. Use an unsigned 3525 // compare, so negative sizes look hugely positive. 3526 int fast_size_limit = FastAllocateSizeLimit; 3527 if (layout_is_con) { 3528 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3529 // Increase the size limit if we have exact knowledge of array type. 3530 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3531 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3532 } 3533 3534 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3535 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3536 3537 // --- Size Computation --- 3538 // array_size = round_to_heap(array_header + (length << elem_shift)); 3539 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3540 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3541 // The rounding mask is strength-reduced, if possible. 3542 int round_mask = MinObjAlignmentInBytes - 1; 3543 Node* header_size = NULL; 3544 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3545 // (T_BYTE has the weakest alignment and size restrictions...) 3546 if (layout_is_con) { 3547 int hsize = Klass::layout_helper_header_size(layout_con); 3548 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3549 BasicType etype = Klass::layout_helper_element_type(layout_con); 3550 if ((round_mask & ~right_n_bits(eshift)) == 0) 3551 round_mask = 0; // strength-reduce it if it goes away completely 3552 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3553 assert(header_size_min <= hsize, "generic minimum is smallest"); 3554 header_size_min = hsize; 3555 header_size = intcon(hsize + round_mask); 3556 } else { 3557 Node* hss = intcon(Klass::_lh_header_size_shift); 3558 Node* hsm = intcon(Klass::_lh_header_size_mask); 3559 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3560 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3561 Node* mask = intcon(round_mask); 3562 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3563 } 3564 3565 Node* elem_shift = NULL; 3566 if (layout_is_con) { 3567 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3568 if (eshift != 0) 3569 elem_shift = intcon(eshift); 3570 } else { 3571 // There is no need to mask or shift this value. 3572 // The semantics of LShiftINode include an implicit mask to 0x1F. 3573 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3574 elem_shift = layout_val; 3575 } 3576 3577 // Transition to native address size for all offset calculations: 3578 Node* lengthx = ConvI2X(length); 3579 Node* headerx = ConvI2X(header_size); 3580 #ifdef _LP64 3581 { const TypeInt* tilen = _gvn.find_int_type(length); 3582 if (tilen != NULL && tilen->_lo < 0) { 3583 // Add a manual constraint to a positive range. Cf. array_element_address. 3584 jint size_max = fast_size_limit; 3585 if (size_max > tilen->_hi) size_max = tilen->_hi; 3586 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3587 3588 // Only do a narrow I2L conversion if the range check passed. 3589 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3590 _gvn.transform(iff); 3591 RegionNode* region = new RegionNode(3); 3592 _gvn.set_type(region, Type::CONTROL); 3593 lengthx = new PhiNode(region, TypeLong::LONG); 3594 _gvn.set_type(lengthx, TypeLong::LONG); 3595 3596 // Range check passed. Use ConvI2L node with narrow type. 3597 Node* passed = IfFalse(iff); 3598 region->init_req(1, passed); 3599 // Make I2L conversion control dependent to prevent it from 3600 // floating above the range check during loop optimizations. 3601 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3602 3603 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3604 region->init_req(2, IfTrue(iff)); 3605 lengthx->init_req(2, ConvI2X(length)); 3606 3607 set_control(region); 3608 record_for_igvn(region); 3609 record_for_igvn(lengthx); 3610 } 3611 } 3612 #endif 3613 3614 // Combine header size (plus rounding) and body size. Then round down. 3615 // This computation cannot overflow, because it is used only in two 3616 // places, one where the length is sharply limited, and the other 3617 // after a successful allocation. 3618 Node* abody = lengthx; 3619 if (elem_shift != NULL) 3620 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3621 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3622 if (round_mask != 0) { 3623 Node* mask = MakeConX(~round_mask); 3624 size = _gvn.transform( new AndXNode(size, mask) ); 3625 } 3626 // else if round_mask == 0, the size computation is self-rounding 3627 3628 if (return_size_val != NULL) { 3629 // This is the size 3630 (*return_size_val) = size; 3631 } 3632 3633 // Now generate allocation code 3634 3635 // The entire memory state is needed for slow path of the allocation 3636 // since GC and deoptimization can happened. 3637 Node *mem = reset_memory(); 3638 set_all_memory(mem); // Create new memory state 3639 3640 if (initial_slow_test->is_Bool()) { 3641 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3642 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3643 } 3644 3645 // Create the AllocateArrayNode and its result projections 3646 AllocateArrayNode* alloc 3647 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3648 control(), mem, i_o(), 3649 size, klass_node, 3650 initial_slow_test, 3651 length); 3652 3653 // Cast to correct type. Note that the klass_node may be constant or not, 3654 // and in the latter case the actual array type will be inexact also. 3655 // (This happens via a non-constant argument to inline_native_newArray.) 3656 // In any case, the value of klass_node provides the desired array type. 3657 const TypeInt* length_type = _gvn.find_int_type(length); 3658 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3659 if (ary_type->isa_aryptr() && length_type != NULL) { 3660 // Try to get a better type than POS for the size 3661 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3662 } 3663 3664 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3665 3666 // Cast length on remaining path to be as narrow as possible 3667 if (map()->find_edge(length) >= 0) { 3668 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3669 if (ccast != length) { 3670 _gvn.set_type_bottom(ccast); 3671 record_for_igvn(ccast); 3672 replace_in_map(length, ccast); 3673 } 3674 } 3675 3676 return javaoop; 3677 } 3678 3679 // The following "Ideal_foo" functions are placed here because they recognize 3680 // the graph shapes created by the functions immediately above. 3681 3682 //---------------------------Ideal_allocation---------------------------------- 3683 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3684 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3685 if (ptr == NULL) { // reduce dumb test in callers 3686 return NULL; 3687 } 3688 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3689 ptr = ptr->in(1); 3690 if (ptr == NULL) return NULL; 3691 } 3692 // Return NULL for allocations with several casts: 3693 // j.l.reflect.Array.newInstance(jobject, jint) 3694 // Object.clone() 3695 // to keep more precise type from last cast. 3696 if (ptr->is_Proj()) { 3697 Node* allo = ptr->in(0); 3698 if (allo != NULL && allo->is_Allocate()) { 3699 return allo->as_Allocate(); 3700 } 3701 } 3702 // Report failure to match. 3703 return NULL; 3704 } 3705 3706 // Fancy version which also strips off an offset (and reports it to caller). 3707 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3708 intptr_t& offset) { 3709 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3710 if (base == NULL) return NULL; 3711 return Ideal_allocation(base, phase); 3712 } 3713 3714 // Trace Initialize <- Proj[Parm] <- Allocate 3715 AllocateNode* InitializeNode::allocation() { 3716 Node* rawoop = in(InitializeNode::RawAddress); 3717 if (rawoop->is_Proj()) { 3718 Node* alloc = rawoop->in(0); 3719 if (alloc->is_Allocate()) { 3720 return alloc->as_Allocate(); 3721 } 3722 } 3723 return NULL; 3724 } 3725 3726 // Trace Allocate -> Proj[Parm] -> Initialize 3727 InitializeNode* AllocateNode::initialization() { 3728 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3729 if (rawoop == NULL) return NULL; 3730 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3731 Node* init = rawoop->fast_out(i); 3732 if (init->is_Initialize()) { 3733 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3734 return init->as_Initialize(); 3735 } 3736 } 3737 return NULL; 3738 } 3739 3740 //----------------------------- loop predicates --------------------------- 3741 3742 //------------------------------add_predicate_impl---------------------------- 3743 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3744 // Too many traps seen? 3745 if (too_many_traps(reason)) { 3746 #ifdef ASSERT 3747 if (TraceLoopPredicate) { 3748 int tc = C->trap_count(reason); 3749 tty->print("too many traps=%s tcount=%d in ", 3750 Deoptimization::trap_reason_name(reason), tc); 3751 method()->print(); // which method has too many predicate traps 3752 tty->cr(); 3753 } 3754 #endif 3755 // We cannot afford to take more traps here, 3756 // do not generate predicate. 3757 return; 3758 } 3759 3760 Node *cont = _gvn.intcon(1); 3761 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3762 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3763 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3764 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3765 C->add_predicate_opaq(opq); 3766 { 3767 PreserveJVMState pjvms(this); 3768 set_control(iffalse); 3769 inc_sp(nargs); 3770 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3771 } 3772 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3773 set_control(iftrue); 3774 } 3775 3776 //------------------------------add_predicate--------------------------------- 3777 void GraphKit::add_predicate(int nargs) { 3778 if (UseLoopPredicate) { 3779 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3780 } 3781 // loop's limit check predicate should be near the loop. 3782 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3783 } 3784 3785 //----------------------------- store barriers ---------------------------- 3786 #define __ ideal. 3787 3788 void GraphKit::sync_kit(IdealKit& ideal) { 3789 set_all_memory(__ merged_memory()); 3790 set_i_o(__ i_o()); 3791 set_control(__ ctrl()); 3792 } 3793 3794 void GraphKit::final_sync(IdealKit& ideal) { 3795 // Final sync IdealKit and graphKit. 3796 sync_kit(ideal); 3797 } 3798 3799 Node* GraphKit::byte_map_base_node() { 3800 // Get base of card map 3801 CardTableModRefBS* ct = 3802 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3803 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3804 if (ct->byte_map_base != NULL) { 3805 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3806 } else { 3807 return null(); 3808 } 3809 } 3810 3811 // vanilla/CMS post barrier 3812 // Insert a write-barrier store. This is to let generational GC work; we have 3813 // to flag all oop-stores before the next GC point. 3814 void GraphKit::write_barrier_post(Node* oop_store, 3815 Node* obj, 3816 Node* adr, 3817 uint adr_idx, 3818 Node* val, 3819 bool use_precise) { 3820 // No store check needed if we're storing a NULL or an old object 3821 // (latter case is probably a string constant). The concurrent 3822 // mark sweep garbage collector, however, needs to have all nonNull 3823 // oop updates flagged via card-marks. 3824 if (val != NULL && val->is_Con()) { 3825 // must be either an oop or NULL 3826 const Type* t = val->bottom_type(); 3827 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3828 // stores of null never (?) need barriers 3829 return; 3830 } 3831 3832 if (use_ReduceInitialCardMarks() 3833 && obj == just_allocated_object(control())) { 3834 // We can skip marks on a freshly-allocated object in Eden. 3835 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3836 // That routine informs GC to take appropriate compensating steps, 3837 // upon a slow-path allocation, so as to make this card-mark 3838 // elision safe. 3839 return; 3840 } 3841 3842 if (!use_precise) { 3843 // All card marks for a (non-array) instance are in one place: 3844 adr = obj; 3845 } 3846 // (Else it's an array (or unknown), and we want more precise card marks.) 3847 assert(adr != NULL, ""); 3848 3849 IdealKit ideal(this, true); 3850 3851 // Convert the pointer to an int prior to doing math on it 3852 Node* cast = __ CastPX(__ ctrl(), adr); 3853 3854 // Divide by card size 3855 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3856 "Only one we handle so far."); 3857 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3858 3859 // Combine card table base and card offset 3860 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3861 3862 // Get the alias_index for raw card-mark memory 3863 int adr_type = Compile::AliasIdxRaw; 3864 Node* zero = __ ConI(0); // Dirty card value 3865 BasicType bt = T_BYTE; 3866 3867 if (UseConcMarkSweepGC && UseCondCardMark) { 3868 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3869 __ sync_kit(this); 3870 } 3871 3872 if (UseCondCardMark) { 3873 // The classic GC reference write barrier is typically implemented 3874 // as a store into the global card mark table. Unfortunately 3875 // unconditional stores can result in false sharing and excessive 3876 // coherence traffic as well as false transactional aborts. 3877 // UseCondCardMark enables MP "polite" conditional card mark 3878 // stores. In theory we could relax the load from ctrl() to 3879 // no_ctrl, but that doesn't buy much latitude. 3880 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3881 __ if_then(card_val, BoolTest::ne, zero); 3882 } 3883 3884 // Smash zero into card 3885 if( !UseConcMarkSweepGC ) { 3886 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3887 } else { 3888 // Specialized path for CM store barrier 3889 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3890 } 3891 3892 if (UseCondCardMark) { 3893 __ end_if(); 3894 } 3895 3896 // Final sync IdealKit and GraphKit. 3897 final_sync(ideal); 3898 } 3899 /* 3900 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3901 * required by SATB to make sure all objects live at the start of the 3902 * marking are kept alive, all reference updates need to any previous 3903 * reference stored before writing. 3904 * 3905 * If the previous value is NULL there is no need to save the old value. 3906 * References that are NULL are filtered during runtime by the barrier 3907 * code to avoid unnecessary queuing. 3908 * 3909 * However in the case of newly allocated objects it might be possible to 3910 * prove that the reference about to be overwritten is NULL during compile 3911 * time and avoid adding the barrier code completely. 3912 * 3913 * The compiler needs to determine that the object in which a field is about 3914 * to be written is newly allocated, and that no prior store to the same field 3915 * has happened since the allocation. 3916 * 3917 * Returns true if the pre-barrier can be removed 3918 */ 3919 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3920 BasicType bt, uint adr_idx) { 3921 intptr_t offset = 0; 3922 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3923 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3924 3925 if (offset == Type::OffsetBot) { 3926 return false; // cannot unalias unless there are precise offsets 3927 } 3928 3929 if (alloc == NULL) { 3930 return false; // No allocation found 3931 } 3932 3933 intptr_t size_in_bytes = type2aelembytes(bt); 3934 3935 Node* mem = memory(adr_idx); // start searching here... 3936 3937 for (int cnt = 0; cnt < 50; cnt++) { 3938 3939 if (mem->is_Store()) { 3940 3941 Node* st_adr = mem->in(MemNode::Address); 3942 intptr_t st_offset = 0; 3943 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3944 3945 if (st_base == NULL) { 3946 break; // inscrutable pointer 3947 } 3948 3949 // Break we have found a store with same base and offset as ours so break 3950 if (st_base == base && st_offset == offset) { 3951 break; 3952 } 3953 3954 if (st_offset != offset && st_offset != Type::OffsetBot) { 3955 const int MAX_STORE = BytesPerLong; 3956 if (st_offset >= offset + size_in_bytes || 3957 st_offset <= offset - MAX_STORE || 3958 st_offset <= offset - mem->as_Store()->memory_size()) { 3959 // Success: The offsets are provably independent. 3960 // (You may ask, why not just test st_offset != offset and be done? 3961 // The answer is that stores of different sizes can co-exist 3962 // in the same sequence of RawMem effects. We sometimes initialize 3963 // a whole 'tile' of array elements with a single jint or jlong.) 3964 mem = mem->in(MemNode::Memory); 3965 continue; // advance through independent store memory 3966 } 3967 } 3968 3969 if (st_base != base 3970 && MemNode::detect_ptr_independence(base, alloc, st_base, 3971 AllocateNode::Ideal_allocation(st_base, phase), 3972 phase)) { 3973 // Success: The bases are provably independent. 3974 mem = mem->in(MemNode::Memory); 3975 continue; // advance through independent store memory 3976 } 3977 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 3978 3979 InitializeNode* st_init = mem->in(0)->as_Initialize(); 3980 AllocateNode* st_alloc = st_init->allocation(); 3981 3982 // Make sure that we are looking at the same allocation site. 3983 // The alloc variable is guaranteed to not be null here from earlier check. 3984 if (alloc == st_alloc) { 3985 // Check that the initialization is storing NULL so that no previous store 3986 // has been moved up and directly write a reference 3987 Node* captured_store = st_init->find_captured_store(offset, 3988 type2aelembytes(T_OBJECT), 3989 phase); 3990 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 3991 return true; 3992 } 3993 } 3994 } 3995 3996 // Unless there is an explicit 'continue', we must bail out here, 3997 // because 'mem' is an inscrutable memory state (e.g., a call). 3998 break; 3999 } 4000 4001 return false; 4002 } 4003 4004 // G1 pre/post barriers 4005 void GraphKit::g1_write_barrier_pre(bool do_load, 4006 Node* obj, 4007 Node* adr, 4008 uint alias_idx, 4009 Node* val, 4010 const TypeOopPtr* val_type, 4011 Node* pre_val, 4012 BasicType bt) { 4013 4014 // Some sanity checks 4015 // Note: val is unused in this routine. 4016 4017 if (do_load) { 4018 // We need to generate the load of the previous value 4019 assert(obj != NULL, "must have a base"); 4020 assert(adr != NULL, "where are loading from?"); 4021 assert(pre_val == NULL, "loaded already?"); 4022 assert(val_type != NULL, "need a type"); 4023 4024 if (use_ReduceInitialCardMarks() 4025 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4026 return; 4027 } 4028 4029 } else { 4030 // In this case both val_type and alias_idx are unused. 4031 assert(pre_val != NULL, "must be loaded already"); 4032 // Nothing to be done if pre_val is null. 4033 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4034 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4035 } 4036 assert(bt == T_OBJECT, "or we shouldn't be here"); 4037 4038 IdealKit ideal(this, true); 4039 4040 Node* tls = __ thread(); // ThreadLocalStorage 4041 4042 Node* no_ctrl = NULL; 4043 Node* no_base = __ top(); 4044 Node* zero = __ ConI(0); 4045 Node* zeroX = __ ConX(0); 4046 4047 float likely = PROB_LIKELY(0.999); 4048 float unlikely = PROB_UNLIKELY(0.999); 4049 4050 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4051 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4052 4053 // Offsets into the thread 4054 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4055 SATBMarkQueue::byte_offset_of_active()); 4056 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4057 SATBMarkQueue::byte_offset_of_index()); 4058 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4059 SATBMarkQueue::byte_offset_of_buf()); 4060 4061 // Now the actual pointers into the thread 4062 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4063 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4064 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4065 4066 // Now some of the values 4067 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4068 4069 // if (!marking) 4070 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4071 BasicType index_bt = TypeX_X->basic_type(); 4072 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4073 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4074 4075 if (do_load) { 4076 // load original value 4077 // alias_idx correct?? 4078 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4079 } 4080 4081 // if (pre_val != NULL) 4082 __ if_then(pre_val, BoolTest::ne, null()); { 4083 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4084 4085 // is the queue for this thread full? 4086 __ if_then(index, BoolTest::ne, zeroX, likely); { 4087 4088 // decrement the index 4089 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4090 4091 // Now get the buffer location we will log the previous value into and store it 4092 Node *log_addr = __ AddP(no_base, buffer, next_index); 4093 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4094 // update the index 4095 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4096 4097 } __ else_(); { 4098 4099 // logging buffer is full, call the runtime 4100 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4101 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4102 } __ end_if(); // (!index) 4103 } __ end_if(); // (pre_val != NULL) 4104 } __ end_if(); // (!marking) 4105 4106 // Final sync IdealKit and GraphKit. 4107 final_sync(ideal); 4108 } 4109 4110 /* 4111 * G1 similar to any GC with a Young Generation requires a way to keep track of 4112 * references from Old Generation to Young Generation to make sure all live 4113 * objects are found. G1 also requires to keep track of object references 4114 * between different regions to enable evacuation of old regions, which is done 4115 * as part of mixed collections. References are tracked in remembered sets and 4116 * is continuously updated as reference are written to with the help of the 4117 * post-barrier. 4118 * 4119 * To reduce the number of updates to the remembered set the post-barrier 4120 * filters updates to fields in objects located in the Young Generation, 4121 * the same region as the reference, when the NULL is being written or 4122 * if the card is already marked as dirty by an earlier write. 4123 * 4124 * Under certain circumstances it is possible to avoid generating the 4125 * post-barrier completely if it is possible during compile time to prove 4126 * the object is newly allocated and that no safepoint exists between the 4127 * allocation and the store. 4128 * 4129 * In the case of slow allocation the allocation code must handle the barrier 4130 * as part of the allocation in the case the allocated object is not located 4131 * in the nursery, this would happen for humongous objects. This is similar to 4132 * how CMS is required to handle this case, see the comments for the method 4133 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4134 * A deferred card mark is required for these objects and handled in the above 4135 * mentioned methods. 4136 * 4137 * Returns true if the post barrier can be removed 4138 */ 4139 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4140 Node* adr) { 4141 intptr_t offset = 0; 4142 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4143 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4144 4145 if (offset == Type::OffsetBot) { 4146 return false; // cannot unalias unless there are precise offsets 4147 } 4148 4149 if (alloc == NULL) { 4150 return false; // No allocation found 4151 } 4152 4153 // Start search from Store node 4154 Node* mem = store->in(MemNode::Control); 4155 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4156 4157 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4158 AllocateNode* st_alloc = st_init->allocation(); 4159 4160 // Make sure we are looking at the same allocation 4161 if (alloc == st_alloc) { 4162 return true; 4163 } 4164 } 4165 4166 return false; 4167 } 4168 4169 // 4170 // Update the card table and add card address to the queue 4171 // 4172 void GraphKit::g1_mark_card(IdealKit& ideal, 4173 Node* card_adr, 4174 Node* oop_store, 4175 uint oop_alias_idx, 4176 Node* index, 4177 Node* index_adr, 4178 Node* buffer, 4179 const TypeFunc* tf) { 4180 4181 Node* zero = __ ConI(0); 4182 Node* zeroX = __ ConX(0); 4183 Node* no_base = __ top(); 4184 BasicType card_bt = T_BYTE; 4185 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4186 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4187 4188 // Now do the queue work 4189 __ if_then(index, BoolTest::ne, zeroX); { 4190 4191 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4192 Node* log_addr = __ AddP(no_base, buffer, next_index); 4193 4194 // Order, see storeCM. 4195 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4196 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4197 4198 } __ else_(); { 4199 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4200 } __ end_if(); 4201 4202 } 4203 4204 void GraphKit::g1_write_barrier_post(Node* oop_store, 4205 Node* obj, 4206 Node* adr, 4207 uint alias_idx, 4208 Node* val, 4209 BasicType bt, 4210 bool use_precise) { 4211 // If we are writing a NULL then we need no post barrier 4212 4213 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4214 // Must be NULL 4215 const Type* t = val->bottom_type(); 4216 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4217 // No post barrier if writing NULLx 4218 return; 4219 } 4220 4221 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4222 // We can skip marks on a freshly-allocated object in Eden. 4223 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4224 // That routine informs GC to take appropriate compensating steps, 4225 // upon a slow-path allocation, so as to make this card-mark 4226 // elision safe. 4227 return; 4228 } 4229 4230 if (use_ReduceInitialCardMarks() 4231 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4232 return; 4233 } 4234 4235 if (!use_precise) { 4236 // All card marks for a (non-array) instance are in one place: 4237 adr = obj; 4238 } 4239 // (Else it's an array (or unknown), and we want more precise card marks.) 4240 assert(adr != NULL, ""); 4241 4242 IdealKit ideal(this, true); 4243 4244 Node* tls = __ thread(); // ThreadLocalStorage 4245 4246 Node* no_base = __ top(); 4247 float likely = PROB_LIKELY(0.999); 4248 float unlikely = PROB_UNLIKELY(0.999); 4249 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4250 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4251 Node* zeroX = __ ConX(0); 4252 4253 // Get the alias_index for raw card-mark memory 4254 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4255 4256 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4257 4258 // Offsets into the thread 4259 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4260 DirtyCardQueue::byte_offset_of_index()); 4261 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4262 DirtyCardQueue::byte_offset_of_buf()); 4263 4264 // Pointers into the thread 4265 4266 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4267 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4268 4269 // Now some values 4270 // Use ctrl to avoid hoisting these values past a safepoint, which could 4271 // potentially reset these fields in the JavaThread. 4272 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4273 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4274 4275 // Convert the store obj pointer to an int prior to doing math on it 4276 // Must use ctrl to prevent "integerized oop" existing across safepoint 4277 Node* cast = __ CastPX(__ ctrl(), adr); 4278 4279 // Divide pointer by card size 4280 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4281 4282 // Combine card table base and card offset 4283 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4284 4285 // If we know the value being stored does it cross regions? 4286 4287 if (val != NULL) { 4288 // Does the store cause us to cross regions? 4289 4290 // Should be able to do an unsigned compare of region_size instead of 4291 // and extra shift. Do we have an unsigned compare?? 4292 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4293 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4294 4295 // if (xor_res == 0) same region so skip 4296 __ if_then(xor_res, BoolTest::ne, zeroX); { 4297 4298 // No barrier if we are storing a NULL 4299 __ if_then(val, BoolTest::ne, null(), unlikely); { 4300 4301 // Ok must mark the card if not already dirty 4302 4303 // load the original value of the card 4304 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4305 4306 __ if_then(card_val, BoolTest::ne, young_card); { 4307 sync_kit(ideal); 4308 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4309 insert_mem_bar(Op_MemBarVolatile, oop_store); 4310 __ sync_kit(this); 4311 4312 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4313 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4314 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4315 } __ end_if(); 4316 } __ end_if(); 4317 } __ end_if(); 4318 } __ end_if(); 4319 } else { 4320 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4321 // We don't need a barrier here if the destination is a newly allocated object 4322 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4323 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4324 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4325 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4326 __ if_then(card_val, BoolTest::ne, young_card); { 4327 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4328 } __ end_if(); 4329 } 4330 4331 // Final sync IdealKit and GraphKit. 4332 final_sync(ideal); 4333 } 4334 #undef __ 4335 4336 4337 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4338 Node* len = load_array_length(load_String_value(ctrl, str)); 4339 Node* coder = load_String_coder(ctrl, str); 4340 // Divide length by 2 if coder is UTF16 4341 return _gvn.transform(new RShiftINode(len, coder)); 4342 } 4343 4344 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4345 int value_offset = java_lang_String::value_offset_in_bytes(); 4346 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4347 false, NULL, 0); 4348 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4349 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4350 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4351 ciTypeArrayKlass::make(T_BYTE), true, 0); 4352 int value_field_idx = C->get_alias_index(value_field_type); 4353 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4354 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4355 // String.value field is known to be @Stable. 4356 if (UseImplicitStableValues) { 4357 load = cast_array_to_stable(load, value_type); 4358 } 4359 return load; 4360 } 4361 4362 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4363 if (!CompactStrings) { 4364 return intcon(java_lang_String::CODER_UTF16); 4365 } 4366 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4367 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4368 false, NULL, 0); 4369 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4370 int coder_field_idx = C->get_alias_index(coder_field_type); 4371 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4372 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4373 } 4374 4375 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4376 int value_offset = java_lang_String::value_offset_in_bytes(); 4377 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4378 false, NULL, 0); 4379 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4380 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4381 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4382 } 4383 4384 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4385 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4386 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4387 false, NULL, 0); 4388 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4389 int coder_field_idx = C->get_alias_index(coder_field_type); 4390 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4391 value, T_BYTE, coder_field_idx, MemNode::unordered); 4392 } 4393 4394 // Capture src and dst memory state with a MergeMemNode 4395 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4396 if (src_type == dst_type) { 4397 // Types are equal, we don't need a MergeMemNode 4398 return memory(src_type); 4399 } 4400 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4401 record_for_igvn(merge); // fold it up later, if possible 4402 int src_idx = C->get_alias_index(src_type); 4403 int dst_idx = C->get_alias_index(dst_type); 4404 merge->set_memory_at(src_idx, memory(src_idx)); 4405 merge->set_memory_at(dst_idx, memory(dst_idx)); 4406 return merge; 4407 } 4408 4409 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4410 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4411 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4412 // If input and output memory types differ, capture both states to preserve 4413 // the dependency between preceding and subsequent loads/stores. 4414 // For example, the following program: 4415 // StoreB 4416 // compress_string 4417 // LoadB 4418 // has this memory graph (use->def): 4419 // LoadB -> compress_string -> CharMem 4420 // ... -> StoreB -> ByteMem 4421 // The intrinsic hides the dependency between LoadB and StoreB, causing 4422 // the load to read from memory not containing the result of the StoreB. 4423 // The correct memory graph should look like this: 4424 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4425 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4426 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4427 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4428 set_memory(res_mem, TypeAryPtr::BYTES); 4429 return str; 4430 } 4431 4432 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4433 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4434 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4435 // Capture src and dst memory (see comment in 'compress_string'). 4436 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4437 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4438 set_memory(_gvn.transform(str), dst_type); 4439 } 4440 4441 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4442 /** 4443 * int i_char = start; 4444 * for (int i_byte = 0; i_byte < count; i_byte++) { 4445 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4446 * } 4447 */ 4448 add_predicate(); 4449 RegionNode* head = new RegionNode(3); 4450 head->init_req(1, control()); 4451 gvn().set_type(head, Type::CONTROL); 4452 record_for_igvn(head); 4453 4454 Node* i_byte = new PhiNode(head, TypeInt::INT); 4455 i_byte->init_req(1, intcon(0)); 4456 gvn().set_type(i_byte, TypeInt::INT); 4457 record_for_igvn(i_byte); 4458 4459 Node* i_char = new PhiNode(head, TypeInt::INT); 4460 i_char->init_req(1, start); 4461 gvn().set_type(i_char, TypeInt::INT); 4462 record_for_igvn(i_char); 4463 4464 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4465 gvn().set_type(mem, Type::MEMORY); 4466 record_for_igvn(mem); 4467 set_control(head); 4468 set_memory(mem, TypeAryPtr::BYTES); 4469 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4470 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4471 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4472 false, false, true /* mismatched */); 4473 4474 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4475 head->init_req(2, IfTrue(iff)); 4476 mem->init_req(2, st); 4477 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4478 i_char->init_req(2, AddI(i_char, intcon(2))); 4479 4480 set_control(IfFalse(iff)); 4481 set_memory(st, TypeAryPtr::BYTES); 4482 } 4483 4484 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4485 if (!field->is_constant()) { 4486 return NULL; // Field not marked as constant. 4487 } 4488 ciInstance* holder = NULL; 4489 if (!field->is_static()) { 4490 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4491 if (const_oop != NULL && const_oop->is_instance()) { 4492 holder = const_oop->as_instance(); 4493 } 4494 } 4495 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4496 /*is_unsigned_load=*/false); 4497 if (con_type != NULL) { 4498 return makecon(con_type); 4499 } 4500 return NULL; 4501 } 4502 4503 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4504 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4505 // assumption of CCP analysis. 4506 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4507 }