1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "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 default: 588 break; 589 } 590 if (failing()) { stop(); return; } // exception allocation might fail 591 if (ex_obj != NULL) { 592 // Cheat with a preallocated exception object. 593 if (C->log() != NULL) 594 C->log()->elem("hot_throw preallocated='1' reason='%s'", 595 Deoptimization::trap_reason_name(reason)); 596 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 597 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 598 599 // Clear the detail message of the preallocated exception object. 600 // Weblogic sometimes mutates the detail message of exceptions 601 // using reflection. 602 int offset = java_lang_Throwable::get_detailMessage_offset(); 603 const TypePtr* adr_typ = ex_con->add_offset(offset); 604 605 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 606 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 607 // Conservatively release stores of object references. 608 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 609 610 add_exception_state(make_exception_state(ex_node)); 611 return; 612 } 613 } 614 615 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 616 // It won't be much cheaper than bailing to the interp., since we'll 617 // have to pass up all the debug-info, and the runtime will have to 618 // create the stack trace. 619 620 // Usual case: Bail to interpreter. 621 // Reserve the right to recompile if we haven't seen anything yet. 622 623 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 624 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 625 if (treat_throw_as_hot 626 && (method()->method_data()->trap_recompiled_at(bci(), m) 627 || C->too_many_traps(reason))) { 628 // We cannot afford to take more traps here. Suffer in the interpreter. 629 if (C->log() != NULL) 630 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 631 Deoptimization::trap_reason_name(reason), 632 C->trap_count(reason)); 633 action = Deoptimization::Action_none; 634 } 635 636 // "must_throw" prunes the JVM state to include only the stack, if there 637 // are no local exception handlers. This should cut down on register 638 // allocation time and code size, by drastically reducing the number 639 // of in-edges on the call to the uncommon trap. 640 641 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 642 } 643 644 645 //----------------------------PreserveJVMState--------------------------------- 646 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 647 debug_only(kit->verify_map()); 648 _kit = kit; 649 _map = kit->map(); // preserve the map 650 _sp = kit->sp(); 651 kit->set_map(clone_map ? kit->clone_map() : NULL); 652 #ifdef ASSERT 653 _bci = kit->bci(); 654 Parse* parser = kit->is_Parse(); 655 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 656 _block = block; 657 #endif 658 } 659 PreserveJVMState::~PreserveJVMState() { 660 GraphKit* kit = _kit; 661 #ifdef ASSERT 662 assert(kit->bci() == _bci, "bci must not shift"); 663 Parse* parser = kit->is_Parse(); 664 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 665 assert(block == _block, "block must not shift"); 666 #endif 667 kit->set_map(_map); 668 kit->set_sp(_sp); 669 } 670 671 672 //-----------------------------BuildCutout------------------------------------- 673 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 674 : PreserveJVMState(kit) 675 { 676 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 677 SafePointNode* outer_map = _map; // preserved map is caller's 678 SafePointNode* inner_map = kit->map(); 679 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 680 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 681 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 682 } 683 BuildCutout::~BuildCutout() { 684 GraphKit* kit = _kit; 685 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 686 } 687 688 //---------------------------PreserveReexecuteState---------------------------- 689 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 690 assert(!kit->stopped(), "must call stopped() before"); 691 _kit = kit; 692 _sp = kit->sp(); 693 _reexecute = kit->jvms()->_reexecute; 694 } 695 PreserveReexecuteState::~PreserveReexecuteState() { 696 if (_kit->stopped()) return; 697 _kit->jvms()->_reexecute = _reexecute; 698 _kit->set_sp(_sp); 699 } 700 701 //------------------------------clone_map-------------------------------------- 702 // Implementation of PreserveJVMState 703 // 704 // Only clone_map(...) here. If this function is only used in the 705 // PreserveJVMState class we may want to get rid of this extra 706 // function eventually and do it all there. 707 708 SafePointNode* GraphKit::clone_map() { 709 if (map() == NULL) return NULL; 710 711 // Clone the memory edge first 712 Node* mem = MergeMemNode::make(map()->memory()); 713 gvn().set_type_bottom(mem); 714 715 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 716 JVMState* jvms = this->jvms(); 717 JVMState* clonejvms = jvms->clone_shallow(C); 718 clonemap->set_memory(mem); 719 clonemap->set_jvms(clonejvms); 720 clonejvms->set_map(clonemap); 721 record_for_igvn(clonemap); 722 gvn().set_type_bottom(clonemap); 723 return clonemap; 724 } 725 726 727 //-----------------------------set_map_clone----------------------------------- 728 void GraphKit::set_map_clone(SafePointNode* m) { 729 _map = m; 730 _map = clone_map(); 731 _map->set_next_exception(NULL); 732 debug_only(verify_map()); 733 } 734 735 736 //----------------------------kill_dead_locals--------------------------------- 737 // Detect any locals which are known to be dead, and force them to top. 738 void GraphKit::kill_dead_locals() { 739 // Consult the liveness information for the locals. If any 740 // of them are unused, then they can be replaced by top(). This 741 // should help register allocation time and cut down on the size 742 // of the deoptimization information. 743 744 // This call is made from many of the bytecode handling 745 // subroutines called from the Big Switch in do_one_bytecode. 746 // Every bytecode which might include a slow path is responsible 747 // for killing its dead locals. The more consistent we 748 // are about killing deads, the fewer useless phis will be 749 // constructed for them at various merge points. 750 751 // bci can be -1 (InvocationEntryBci). We return the entry 752 // liveness for the method. 753 754 if (method() == NULL || method()->code_size() == 0) { 755 // We are building a graph for a call to a native method. 756 // All locals are live. 757 return; 758 } 759 760 ResourceMark rm; 761 762 // Consult the liveness information for the locals. If any 763 // of them are unused, then they can be replaced by top(). This 764 // should help register allocation time and cut down on the size 765 // of the deoptimization information. 766 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 767 768 int len = (int)live_locals.size(); 769 assert(len <= jvms()->loc_size(), "too many live locals"); 770 for (int local = 0; local < len; local++) { 771 if (!live_locals.at(local)) { 772 set_local(local, top()); 773 } 774 } 775 } 776 777 #ifdef ASSERT 778 //-------------------------dead_locals_are_killed------------------------------ 779 // Return true if all dead locals are set to top in the map. 780 // Used to assert "clean" debug info at various points. 781 bool GraphKit::dead_locals_are_killed() { 782 if (method() == NULL || method()->code_size() == 0) { 783 // No locals need to be dead, so all is as it should be. 784 return true; 785 } 786 787 // Make sure somebody called kill_dead_locals upstream. 788 ResourceMark rm; 789 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 790 if (jvms->loc_size() == 0) continue; // no locals to consult 791 SafePointNode* map = jvms->map(); 792 ciMethod* method = jvms->method(); 793 int bci = jvms->bci(); 794 if (jvms == this->jvms()) { 795 bci = this->bci(); // it might not yet be synched 796 } 797 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 798 int len = (int)live_locals.size(); 799 if (!live_locals.is_valid() || len == 0) 800 // This method is trivial, or is poisoned by a breakpoint. 801 return true; 802 assert(len == jvms->loc_size(), "live map consistent with locals map"); 803 for (int local = 0; local < len; local++) { 804 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 805 if (PrintMiscellaneous && (Verbose || WizardMode)) { 806 tty->print_cr("Zombie local %d: ", local); 807 jvms->dump(); 808 } 809 return false; 810 } 811 } 812 } 813 return true; 814 } 815 816 #endif //ASSERT 817 818 // Helper function for enforcing certain bytecodes to reexecute if 819 // deoptimization happens 820 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 821 ciMethod* cur_method = jvms->method(); 822 int cur_bci = jvms->bci(); 823 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 824 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 825 return Interpreter::bytecode_should_reexecute(code) || 826 (is_anewarray && code == Bytecodes::_multianewarray); 827 // Reexecute _multianewarray bytecode which was replaced with 828 // sequence of [a]newarray. See Parse::do_multianewarray(). 829 // 830 // Note: interpreter should not have it set since this optimization 831 // is limited by dimensions and guarded by flag so in some cases 832 // multianewarray() runtime calls will be generated and 833 // the bytecode should not be reexecutes (stack will not be reset). 834 } else 835 return false; 836 } 837 838 // Helper function for adding JVMState and debug information to node 839 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 840 // Add the safepoint edges to the call (or other safepoint). 841 842 // Make sure dead locals are set to top. This 843 // should help register allocation time and cut down on the size 844 // of the deoptimization information. 845 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 846 847 // Walk the inline list to fill in the correct set of JVMState's 848 // Also fill in the associated edges for each JVMState. 849 850 // If the bytecode needs to be reexecuted we need to put 851 // the arguments back on the stack. 852 const bool should_reexecute = jvms()->should_reexecute(); 853 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 854 855 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 856 // undefined if the bci is different. This is normal for Parse but it 857 // should not happen for LibraryCallKit because only one bci is processed. 858 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 859 "in LibraryCallKit the reexecute bit should not change"); 860 861 // If we are guaranteed to throw, we can prune everything but the 862 // input to the current bytecode. 863 bool can_prune_locals = false; 864 uint stack_slots_not_pruned = 0; 865 int inputs = 0, depth = 0; 866 if (must_throw) { 867 assert(method() == youngest_jvms->method(), "sanity"); 868 if (compute_stack_effects(inputs, depth)) { 869 can_prune_locals = true; 870 stack_slots_not_pruned = inputs; 871 } 872 } 873 874 if (env()->should_retain_local_variables()) { 875 // At any safepoint, this method can get breakpointed, which would 876 // then require an immediate deoptimization. 877 can_prune_locals = false; // do not prune locals 878 stack_slots_not_pruned = 0; 879 } 880 881 // do not scribble on the input jvms 882 JVMState* out_jvms = youngest_jvms->clone_deep(C); 883 call->set_jvms(out_jvms); // Start jvms list for call node 884 885 // For a known set of bytecodes, the interpreter should reexecute them if 886 // deoptimization happens. We set the reexecute state for them here 887 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 888 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 889 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 890 } 891 892 // Presize the call: 893 DEBUG_ONLY(uint non_debug_edges = call->req()); 894 call->add_req_batch(top(), youngest_jvms->debug_depth()); 895 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 896 897 // Set up edges so that the call looks like this: 898 // Call [state:] ctl io mem fptr retadr 899 // [parms:] parm0 ... parmN 900 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 901 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 902 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 903 // Note that caller debug info precedes callee debug info. 904 905 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 906 uint debug_ptr = call->req(); 907 908 // Loop over the map input edges associated with jvms, add them 909 // to the call node, & reset all offsets to match call node array. 910 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 911 uint debug_end = debug_ptr; 912 uint debug_start = debug_ptr - in_jvms->debug_size(); 913 debug_ptr = debug_start; // back up the ptr 914 915 uint p = debug_start; // walks forward in [debug_start, debug_end) 916 uint j, k, l; 917 SafePointNode* in_map = in_jvms->map(); 918 out_jvms->set_map(call); 919 920 if (can_prune_locals) { 921 assert(in_jvms->method() == out_jvms->method(), "sanity"); 922 // If the current throw can reach an exception handler in this JVMS, 923 // then we must keep everything live that can reach that handler. 924 // As a quick and dirty approximation, we look for any handlers at all. 925 if (in_jvms->method()->has_exception_handlers()) { 926 can_prune_locals = false; 927 } 928 } 929 930 // Add the Locals 931 k = in_jvms->locoff(); 932 l = in_jvms->loc_size(); 933 out_jvms->set_locoff(p); 934 if (!can_prune_locals) { 935 for (j = 0; j < l; j++) 936 call->set_req(p++, in_map->in(k+j)); 937 } else { 938 p += l; // already set to top above by add_req_batch 939 } 940 941 // Add the Expression Stack 942 k = in_jvms->stkoff(); 943 l = in_jvms->sp(); 944 out_jvms->set_stkoff(p); 945 if (!can_prune_locals) { 946 for (j = 0; j < l; j++) 947 call->set_req(p++, in_map->in(k+j)); 948 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 949 // Divide stack into {S0,...,S1}, where S0 is set to top. 950 uint s1 = stack_slots_not_pruned; 951 stack_slots_not_pruned = 0; // for next iteration 952 if (s1 > l) s1 = l; 953 uint s0 = l - s1; 954 p += s0; // skip the tops preinstalled by add_req_batch 955 for (j = s0; j < l; j++) 956 call->set_req(p++, in_map->in(k+j)); 957 } else { 958 p += l; // already set to top above by add_req_batch 959 } 960 961 // Add the Monitors 962 k = in_jvms->monoff(); 963 l = in_jvms->mon_size(); 964 out_jvms->set_monoff(p); 965 for (j = 0; j < l; j++) 966 call->set_req(p++, in_map->in(k+j)); 967 968 // Copy any scalar object fields. 969 k = in_jvms->scloff(); 970 l = in_jvms->scl_size(); 971 out_jvms->set_scloff(p); 972 for (j = 0; j < l; j++) 973 call->set_req(p++, in_map->in(k+j)); 974 975 // Finish the new jvms. 976 out_jvms->set_endoff(p); 977 978 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 979 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 980 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 981 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 982 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 983 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 984 985 // Update the two tail pointers in parallel. 986 out_jvms = out_jvms->caller(); 987 in_jvms = in_jvms->caller(); 988 } 989 990 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 991 992 // Test the correctness of JVMState::debug_xxx accessors: 993 assert(call->jvms()->debug_start() == non_debug_edges, ""); 994 assert(call->jvms()->debug_end() == call->req(), ""); 995 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 996 } 997 998 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 999 Bytecodes::Code code = java_bc(); 1000 if (code == Bytecodes::_wide) { 1001 code = method()->java_code_at_bci(bci() + 1); 1002 } 1003 1004 BasicType rtype = T_ILLEGAL; 1005 int rsize = 0; 1006 1007 if (code != Bytecodes::_illegal) { 1008 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1009 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1010 if (rtype < T_CONFLICT) 1011 rsize = type2size[rtype]; 1012 } 1013 1014 switch (code) { 1015 case Bytecodes::_illegal: 1016 return false; 1017 1018 case Bytecodes::_ldc: 1019 case Bytecodes::_ldc_w: 1020 case Bytecodes::_ldc2_w: 1021 inputs = 0; 1022 break; 1023 1024 case Bytecodes::_dup: inputs = 1; break; 1025 case Bytecodes::_dup_x1: inputs = 2; break; 1026 case Bytecodes::_dup_x2: inputs = 3; break; 1027 case Bytecodes::_dup2: inputs = 2; break; 1028 case Bytecodes::_dup2_x1: inputs = 3; break; 1029 case Bytecodes::_dup2_x2: inputs = 4; break; 1030 case Bytecodes::_swap: inputs = 2; break; 1031 case Bytecodes::_arraylength: inputs = 1; break; 1032 1033 case Bytecodes::_getstatic: 1034 case Bytecodes::_putstatic: 1035 case Bytecodes::_getfield: 1036 case Bytecodes::_putfield: 1037 { 1038 bool ignored_will_link; 1039 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1040 int size = field->type()->size(); 1041 bool is_get = (depth >= 0), is_static = (depth & 1); 1042 inputs = (is_static ? 0 : 1); 1043 if (is_get) { 1044 depth = size - inputs; 1045 } else { 1046 inputs += size; // putxxx pops the value from the stack 1047 depth = - inputs; 1048 } 1049 } 1050 break; 1051 1052 case Bytecodes::_invokevirtual: 1053 case Bytecodes::_invokespecial: 1054 case Bytecodes::_invokestatic: 1055 case Bytecodes::_invokedynamic: 1056 case Bytecodes::_invokeinterface: 1057 { 1058 bool ignored_will_link; 1059 ciSignature* declared_signature = NULL; 1060 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1061 assert(declared_signature != NULL, "cannot be null"); 1062 inputs = declared_signature->arg_size_for_bc(code); 1063 int size = declared_signature->return_type()->size(); 1064 depth = size - inputs; 1065 } 1066 break; 1067 1068 case Bytecodes::_multianewarray: 1069 { 1070 ciBytecodeStream iter(method()); 1071 iter.reset_to_bci(bci()); 1072 iter.next(); 1073 inputs = iter.get_dimensions(); 1074 assert(rsize == 1, ""); 1075 depth = rsize - inputs; 1076 } 1077 break; 1078 1079 case Bytecodes::_ireturn: 1080 case Bytecodes::_lreturn: 1081 case Bytecodes::_freturn: 1082 case Bytecodes::_dreturn: 1083 case Bytecodes::_areturn: 1084 assert(rsize == -depth, ""); 1085 inputs = rsize; 1086 break; 1087 1088 case Bytecodes::_jsr: 1089 case Bytecodes::_jsr_w: 1090 inputs = 0; 1091 depth = 1; // S.B. depth=1, not zero 1092 break; 1093 1094 default: 1095 // bytecode produces a typed result 1096 inputs = rsize - depth; 1097 assert(inputs >= 0, ""); 1098 break; 1099 } 1100 1101 #ifdef ASSERT 1102 // spot check 1103 int outputs = depth + inputs; 1104 assert(outputs >= 0, "sanity"); 1105 switch (code) { 1106 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1107 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1108 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1109 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1110 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1111 default: break; 1112 } 1113 #endif //ASSERT 1114 1115 return true; 1116 } 1117 1118 1119 1120 //------------------------------basic_plus_adr--------------------------------- 1121 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1122 // short-circuit a common case 1123 if (offset == intcon(0)) return ptr; 1124 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1125 } 1126 1127 Node* GraphKit::ConvI2L(Node* offset) { 1128 // short-circuit a common case 1129 jint offset_con = find_int_con(offset, Type::OffsetBot); 1130 if (offset_con != Type::OffsetBot) { 1131 return longcon((jlong) offset_con); 1132 } 1133 return _gvn.transform( new ConvI2LNode(offset)); 1134 } 1135 1136 Node* GraphKit::ConvI2UL(Node* offset) { 1137 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1138 if (offset_con != (juint) Type::OffsetBot) { 1139 return longcon((julong) offset_con); 1140 } 1141 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1142 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1143 return _gvn.transform( new AndLNode(conv, mask) ); 1144 } 1145 1146 Node* GraphKit::ConvL2I(Node* offset) { 1147 // short-circuit a common case 1148 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1149 if (offset_con != (jlong)Type::OffsetBot) { 1150 return intcon((int) offset_con); 1151 } 1152 return _gvn.transform( new ConvL2INode(offset)); 1153 } 1154 1155 //-------------------------load_object_klass----------------------------------- 1156 Node* GraphKit::load_object_klass(Node* obj) { 1157 // Special-case a fresh allocation to avoid building nodes: 1158 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1159 if (akls != NULL) return akls; 1160 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1161 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1162 } 1163 1164 //-------------------------load_array_length----------------------------------- 1165 Node* GraphKit::load_array_length(Node* array) { 1166 // Special-case a fresh allocation to avoid building nodes: 1167 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1168 Node *alen; 1169 if (alloc == NULL) { 1170 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1171 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1172 } else { 1173 alen = alloc->Ideal_length(); 1174 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1175 if (ccast != alen) { 1176 alen = _gvn.transform(ccast); 1177 } 1178 } 1179 return alen; 1180 } 1181 1182 //------------------------------do_null_check---------------------------------- 1183 // Helper function to do a NULL pointer check. Returned value is 1184 // the incoming address with NULL casted away. You are allowed to use the 1185 // not-null value only if you are control dependent on the test. 1186 #ifndef PRODUCT 1187 extern int explicit_null_checks_inserted, 1188 explicit_null_checks_elided; 1189 #endif 1190 Node* GraphKit::null_check_common(Node* value, BasicType type, 1191 // optional arguments for variations: 1192 bool assert_null, 1193 Node* *null_control, 1194 bool speculative) { 1195 assert(!assert_null || null_control == NULL, "not both at once"); 1196 if (stopped()) return top(); 1197 NOT_PRODUCT(explicit_null_checks_inserted++); 1198 1199 // Construct NULL check 1200 Node *chk = NULL; 1201 switch(type) { 1202 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1203 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1204 case T_ARRAY : // fall through 1205 type = T_OBJECT; // simplify further tests 1206 case T_OBJECT : { 1207 const Type *t = _gvn.type( value ); 1208 1209 const TypeOopPtr* tp = t->isa_oopptr(); 1210 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1211 // Only for do_null_check, not any of its siblings: 1212 && !assert_null && null_control == NULL) { 1213 // Usually, any field access or invocation on an unloaded oop type 1214 // will simply fail to link, since the statically linked class is 1215 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1216 // the static class is loaded but the sharper oop type is not. 1217 // Rather than checking for this obscure case in lots of places, 1218 // we simply observe that a null check on an unloaded class 1219 // will always be followed by a nonsense operation, so we 1220 // can just issue the uncommon trap here. 1221 // Our access to the unloaded class will only be correct 1222 // after it has been loaded and initialized, which requires 1223 // a trip through the interpreter. 1224 #ifndef PRODUCT 1225 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1226 #endif 1227 uncommon_trap(Deoptimization::Reason_unloaded, 1228 Deoptimization::Action_reinterpret, 1229 tp->klass(), "!loaded"); 1230 return top(); 1231 } 1232 1233 if (assert_null) { 1234 // See if the type is contained in NULL_PTR. 1235 // If so, then the value is already null. 1236 if (t->higher_equal(TypePtr::NULL_PTR)) { 1237 NOT_PRODUCT(explicit_null_checks_elided++); 1238 return value; // Elided null assert quickly! 1239 } 1240 } else { 1241 // See if mixing in the NULL pointer changes type. 1242 // If so, then the NULL pointer was not allowed in the original 1243 // type. In other words, "value" was not-null. 1244 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1245 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1246 NOT_PRODUCT(explicit_null_checks_elided++); 1247 return value; // Elided null check quickly! 1248 } 1249 } 1250 chk = new CmpPNode( value, null() ); 1251 break; 1252 } 1253 1254 default: 1255 fatal("unexpected type: %s", type2name(type)); 1256 } 1257 assert(chk != NULL, "sanity check"); 1258 chk = _gvn.transform(chk); 1259 1260 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1261 BoolNode *btst = new BoolNode( chk, btest); 1262 Node *tst = _gvn.transform( btst ); 1263 1264 //----------- 1265 // if peephole optimizations occurred, a prior test existed. 1266 // If a prior test existed, maybe it dominates as we can avoid this test. 1267 if (tst != btst && type == T_OBJECT) { 1268 // At this point we want to scan up the CFG to see if we can 1269 // find an identical test (and so avoid this test altogether). 1270 Node *cfg = control(); 1271 int depth = 0; 1272 while( depth < 16 ) { // Limit search depth for speed 1273 if( cfg->Opcode() == Op_IfTrue && 1274 cfg->in(0)->in(1) == tst ) { 1275 // Found prior test. Use "cast_not_null" to construct an identical 1276 // CastPP (and hence hash to) as already exists for the prior test. 1277 // Return that casted value. 1278 if (assert_null) { 1279 replace_in_map(value, null()); 1280 return null(); // do not issue the redundant test 1281 } 1282 Node *oldcontrol = control(); 1283 set_control(cfg); 1284 Node *res = cast_not_null(value); 1285 set_control(oldcontrol); 1286 NOT_PRODUCT(explicit_null_checks_elided++); 1287 return res; 1288 } 1289 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1290 if (cfg == NULL) break; // Quit at region nodes 1291 depth++; 1292 } 1293 } 1294 1295 //----------- 1296 // Branch to failure if null 1297 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1298 Deoptimization::DeoptReason reason; 1299 if (assert_null) { 1300 reason = Deoptimization::reason_null_assert(speculative); 1301 } else if (type == T_OBJECT) { 1302 reason = Deoptimization::reason_null_check(speculative); 1303 } else { 1304 reason = Deoptimization::Reason_div0_check; 1305 } 1306 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1307 // ciMethodData::has_trap_at will return a conservative -1 if any 1308 // must-be-null assertion has failed. This could cause performance 1309 // problems for a method after its first do_null_assert failure. 1310 // Consider using 'Reason_class_check' instead? 1311 1312 // To cause an implicit null check, we set the not-null probability 1313 // to the maximum (PROB_MAX). For an explicit check the probability 1314 // is set to a smaller value. 1315 if (null_control != NULL || too_many_traps(reason)) { 1316 // probability is less likely 1317 ok_prob = PROB_LIKELY_MAG(3); 1318 } else if (!assert_null && 1319 (ImplicitNullCheckThreshold > 0) && 1320 method() != NULL && 1321 (method()->method_data()->trap_count(reason) 1322 >= (uint)ImplicitNullCheckThreshold)) { 1323 ok_prob = PROB_LIKELY_MAG(3); 1324 } 1325 1326 if (null_control != NULL) { 1327 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1328 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1329 set_control( _gvn.transform( new IfTrueNode(iff))); 1330 #ifndef PRODUCT 1331 if (null_true == top()) { 1332 explicit_null_checks_elided++; 1333 } 1334 #endif 1335 (*null_control) = null_true; 1336 } else { 1337 BuildCutout unless(this, tst, ok_prob); 1338 // Check for optimizer eliding test at parse time 1339 if (stopped()) { 1340 // Failure not possible; do not bother making uncommon trap. 1341 NOT_PRODUCT(explicit_null_checks_elided++); 1342 } else if (assert_null) { 1343 uncommon_trap(reason, 1344 Deoptimization::Action_make_not_entrant, 1345 NULL, "assert_null"); 1346 } else { 1347 replace_in_map(value, zerocon(type)); 1348 builtin_throw(reason); 1349 } 1350 } 1351 1352 // Must throw exception, fall-thru not possible? 1353 if (stopped()) { 1354 return top(); // No result 1355 } 1356 1357 if (assert_null) { 1358 // Cast obj to null on this path. 1359 replace_in_map(value, zerocon(type)); 1360 return zerocon(type); 1361 } 1362 1363 // Cast obj to not-null on this path, if there is no null_control. 1364 // (If there is a null_control, a non-null value may come back to haunt us.) 1365 if (type == T_OBJECT) { 1366 Node* cast = cast_not_null(value, false); 1367 if (null_control == NULL || (*null_control) == top()) 1368 replace_in_map(value, cast); 1369 value = cast; 1370 } 1371 1372 return value; 1373 } 1374 1375 1376 //------------------------------cast_not_null---------------------------------- 1377 // Cast obj to not-null on this path 1378 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1379 const Type *t = _gvn.type(obj); 1380 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1381 // Object is already not-null? 1382 if( t == t_not_null ) return obj; 1383 1384 Node *cast = new CastPPNode(obj,t_not_null); 1385 cast->init_req(0, control()); 1386 cast = _gvn.transform( cast ); 1387 1388 // Scan for instances of 'obj' in the current JVM mapping. 1389 // These instances are known to be not-null after the test. 1390 if (do_replace_in_map) 1391 replace_in_map(obj, cast); 1392 1393 return cast; // Return casted value 1394 } 1395 1396 // Sometimes in intrinsics, we implicitly know an object is not null 1397 // (there's no actual null check) so we can cast it to not null. In 1398 // the course of optimizations, the input to the cast can become null. 1399 // In that case that data path will die and we need the control path 1400 // to become dead as well to keep the graph consistent. So we have to 1401 // add a check for null for which one branch can't be taken. It uses 1402 // an Opaque4 node that will cause the check to be removed after loop 1403 // opts so the test goes away and the compiled code doesn't execute a 1404 // useless check. 1405 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1406 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1407 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1408 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1409 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1410 _gvn.set_type(iff, iff->Value(&_gvn)); 1411 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1412 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1413 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1414 C->root()->add_req(halt); 1415 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1416 set_control(if_t); 1417 return cast_not_null(value, do_replace_in_map); 1418 } 1419 1420 1421 //--------------------------replace_in_map------------------------------------- 1422 void GraphKit::replace_in_map(Node* old, Node* neww) { 1423 if (old == neww) { 1424 return; 1425 } 1426 1427 map()->replace_edge(old, neww); 1428 1429 // Note: This operation potentially replaces any edge 1430 // on the map. This includes locals, stack, and monitors 1431 // of the current (innermost) JVM state. 1432 1433 // don't let inconsistent types from profiling escape this 1434 // method 1435 1436 const Type* told = _gvn.type(old); 1437 const Type* tnew = _gvn.type(neww); 1438 1439 if (!tnew->higher_equal(told)) { 1440 return; 1441 } 1442 1443 map()->record_replaced_node(old, neww); 1444 } 1445 1446 1447 //============================================================================= 1448 //--------------------------------memory--------------------------------------- 1449 Node* GraphKit::memory(uint alias_idx) { 1450 MergeMemNode* mem = merged_memory(); 1451 Node* p = mem->memory_at(alias_idx); 1452 _gvn.set_type(p, Type::MEMORY); // must be mapped 1453 return p; 1454 } 1455 1456 //-----------------------------reset_memory------------------------------------ 1457 Node* GraphKit::reset_memory() { 1458 Node* mem = map()->memory(); 1459 // do not use this node for any more parsing! 1460 debug_only( map()->set_memory((Node*)NULL) ); 1461 return _gvn.transform( mem ); 1462 } 1463 1464 //------------------------------set_all_memory--------------------------------- 1465 void GraphKit::set_all_memory(Node* newmem) { 1466 Node* mergemem = MergeMemNode::make(newmem); 1467 gvn().set_type_bottom(mergemem); 1468 map()->set_memory(mergemem); 1469 } 1470 1471 //------------------------------set_all_memory_call---------------------------- 1472 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1473 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1474 set_all_memory(newmem); 1475 } 1476 1477 //============================================================================= 1478 // 1479 // parser factory methods for MemNodes 1480 // 1481 // These are layered on top of the factory methods in LoadNode and StoreNode, 1482 // and integrate with the parser's memory state and _gvn engine. 1483 // 1484 1485 // factory methods in "int adr_idx" 1486 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1487 int adr_idx, 1488 MemNode::MemOrd mo, 1489 LoadNode::ControlDependency control_dependency, 1490 bool require_atomic_access, 1491 bool unaligned, 1492 bool mismatched) { 1493 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1494 const TypePtr* adr_type = NULL; // debug-mode-only argument 1495 debug_only(adr_type = C->get_adr_type(adr_idx)); 1496 Node* mem = memory(adr_idx); 1497 Node* ld; 1498 if (require_atomic_access && bt == T_LONG) { 1499 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1500 } else if (require_atomic_access && bt == T_DOUBLE) { 1501 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1502 } else { 1503 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1504 } 1505 ld = _gvn.transform(ld); 1506 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1507 // Improve graph before escape analysis and boxing elimination. 1508 record_for_igvn(ld); 1509 } 1510 return ld; 1511 } 1512 1513 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1514 int adr_idx, 1515 MemNode::MemOrd mo, 1516 bool require_atomic_access, 1517 bool unaligned, 1518 bool mismatched) { 1519 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1520 const TypePtr* adr_type = NULL; 1521 debug_only(adr_type = C->get_adr_type(adr_idx)); 1522 Node *mem = memory(adr_idx); 1523 Node* st; 1524 if (require_atomic_access && bt == T_LONG) { 1525 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1526 } else if (require_atomic_access && bt == T_DOUBLE) { 1527 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1528 } else { 1529 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1530 } 1531 if (unaligned) { 1532 st->as_Store()->set_unaligned_access(); 1533 } 1534 if (mismatched) { 1535 st->as_Store()->set_mismatched_access(); 1536 } 1537 st = _gvn.transform(st); 1538 set_memory(st, adr_idx); 1539 // Back-to-back stores can only remove intermediate store with DU info 1540 // so push on worklist for optimizer. 1541 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1542 record_for_igvn(st); 1543 1544 return st; 1545 } 1546 1547 1548 void GraphKit::pre_barrier(bool do_load, 1549 Node* ctl, 1550 Node* obj, 1551 Node* adr, 1552 uint adr_idx, 1553 Node* val, 1554 const TypeOopPtr* val_type, 1555 Node* pre_val, 1556 BasicType bt) { 1557 1558 BarrierSet* bs = Universe::heap()->barrier_set(); 1559 set_control(ctl); 1560 switch (bs->kind()) { 1561 case BarrierSet::G1SATBCTLogging: 1562 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1563 break; 1564 1565 case BarrierSet::CardTableForRS: 1566 case BarrierSet::CardTableExtension: 1567 case BarrierSet::ModRef: 1568 break; 1569 1570 default : 1571 ShouldNotReachHere(); 1572 1573 } 1574 } 1575 1576 bool GraphKit::can_move_pre_barrier() const { 1577 BarrierSet* bs = Universe::heap()->barrier_set(); 1578 switch (bs->kind()) { 1579 case BarrierSet::G1SATBCTLogging: 1580 return true; // Can move it if no safepoint 1581 1582 case BarrierSet::CardTableForRS: 1583 case BarrierSet::CardTableExtension: 1584 case BarrierSet::ModRef: 1585 return true; // There is no pre-barrier 1586 1587 default : 1588 ShouldNotReachHere(); 1589 } 1590 return false; 1591 } 1592 1593 void GraphKit::post_barrier(Node* ctl, 1594 Node* store, 1595 Node* obj, 1596 Node* adr, 1597 uint adr_idx, 1598 Node* val, 1599 BasicType bt, 1600 bool use_precise) { 1601 BarrierSet* bs = Universe::heap()->barrier_set(); 1602 set_control(ctl); 1603 switch (bs->kind()) { 1604 case BarrierSet::G1SATBCTLogging: 1605 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1606 break; 1607 1608 case BarrierSet::CardTableForRS: 1609 case BarrierSet::CardTableExtension: 1610 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1611 break; 1612 1613 case BarrierSet::ModRef: 1614 break; 1615 1616 default : 1617 ShouldNotReachHere(); 1618 1619 } 1620 } 1621 1622 Node* GraphKit::store_oop(Node* ctl, 1623 Node* obj, 1624 Node* adr, 1625 const TypePtr* adr_type, 1626 Node* val, 1627 const TypeOopPtr* val_type, 1628 BasicType bt, 1629 bool use_precise, 1630 MemNode::MemOrd mo, 1631 bool mismatched) { 1632 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1633 // could be delayed during Parse (for example, in adjust_map_after_if()). 1634 // Execute transformation here to avoid barrier generation in such case. 1635 if (_gvn.type(val) == TypePtr::NULL_PTR) 1636 val = _gvn.makecon(TypePtr::NULL_PTR); 1637 1638 set_control(ctl); 1639 if (stopped()) return top(); // Dead path ? 1640 1641 assert(bt == T_OBJECT, "sanity"); 1642 assert(val != NULL, "not dead path"); 1643 uint adr_idx = C->get_alias_index(adr_type); 1644 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1645 1646 pre_barrier(true /* do_load */, 1647 control(), obj, adr, adr_idx, val, val_type, 1648 NULL /* pre_val */, 1649 bt); 1650 1651 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1652 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1653 return store; 1654 } 1655 1656 // Could be an array or object we don't know at compile time (unsafe ref.) 1657 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1658 Node* obj, // containing obj 1659 Node* adr, // actual adress to store val at 1660 const TypePtr* adr_type, 1661 Node* val, 1662 BasicType bt, 1663 MemNode::MemOrd mo, 1664 bool mismatched) { 1665 Compile::AliasType* at = C->alias_type(adr_type); 1666 const TypeOopPtr* val_type = NULL; 1667 if (adr_type->isa_instptr()) { 1668 if (at->field() != NULL) { 1669 // known field. This code is a copy of the do_put_xxx logic. 1670 ciField* field = at->field(); 1671 if (!field->type()->is_loaded()) { 1672 val_type = TypeInstPtr::BOTTOM; 1673 } else { 1674 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1675 } 1676 } 1677 } else if (adr_type->isa_aryptr()) { 1678 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1679 } 1680 if (val_type == NULL) { 1681 val_type = TypeInstPtr::BOTTOM; 1682 } 1683 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1684 } 1685 1686 1687 //-------------------------array_element_address------------------------- 1688 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1689 const TypeInt* sizetype, Node* ctrl) { 1690 uint shift = exact_log2(type2aelembytes(elembt)); 1691 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1692 1693 // short-circuit a common case (saves lots of confusing waste motion) 1694 jint idx_con = find_int_con(idx, -1); 1695 if (idx_con >= 0) { 1696 intptr_t offset = header + ((intptr_t)idx_con << shift); 1697 return basic_plus_adr(ary, offset); 1698 } 1699 1700 // must be correct type for alignment purposes 1701 Node* base = basic_plus_adr(ary, header); 1702 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1703 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1704 return basic_plus_adr(ary, base, scale); 1705 } 1706 1707 //-------------------------load_array_element------------------------- 1708 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1709 const Type* elemtype = arytype->elem(); 1710 BasicType elembt = elemtype->array_element_basic_type(); 1711 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1712 if (elembt == T_NARROWOOP) { 1713 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1714 } 1715 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1716 return ld; 1717 } 1718 1719 //-------------------------set_arguments_for_java_call------------------------- 1720 // Arguments (pre-popped from the stack) are taken from the JVMS. 1721 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1722 // Add the call arguments: 1723 uint nargs = call->method()->arg_size(); 1724 for (uint i = 0; i < nargs; i++) { 1725 Node* arg = argument(i); 1726 call->init_req(i + TypeFunc::Parms, arg); 1727 } 1728 } 1729 1730 //---------------------------set_edges_for_java_call--------------------------- 1731 // Connect a newly created call into the current JVMS. 1732 // A return value node (if any) is returned from set_edges_for_java_call. 1733 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1734 1735 // Add the predefined inputs: 1736 call->init_req( TypeFunc::Control, control() ); 1737 call->init_req( TypeFunc::I_O , i_o() ); 1738 call->init_req( TypeFunc::Memory , reset_memory() ); 1739 call->init_req( TypeFunc::FramePtr, frameptr() ); 1740 call->init_req( TypeFunc::ReturnAdr, top() ); 1741 1742 add_safepoint_edges(call, must_throw); 1743 1744 Node* xcall = _gvn.transform(call); 1745 1746 if (xcall == top()) { 1747 set_control(top()); 1748 return; 1749 } 1750 assert(xcall == call, "call identity is stable"); 1751 1752 // Re-use the current map to produce the result. 1753 1754 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1755 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1756 set_all_memory_call(xcall, separate_io_proj); 1757 1758 //return xcall; // no need, caller already has it 1759 } 1760 1761 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1762 if (stopped()) return top(); // maybe the call folded up? 1763 1764 // Capture the return value, if any. 1765 Node* ret; 1766 if (call->method() == NULL || 1767 call->method()->return_type()->basic_type() == T_VOID) 1768 ret = top(); 1769 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1770 1771 // Note: Since any out-of-line call can produce an exception, 1772 // we always insert an I_O projection from the call into the result. 1773 1774 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1775 1776 if (separate_io_proj) { 1777 // The caller requested separate projections be used by the fall 1778 // through and exceptional paths, so replace the projections for 1779 // the fall through path. 1780 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1781 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1782 } 1783 return ret; 1784 } 1785 1786 //--------------------set_predefined_input_for_runtime_call-------------------- 1787 // Reading and setting the memory state is way conservative here. 1788 // The real problem is that I am not doing real Type analysis on memory, 1789 // so I cannot distinguish card mark stores from other stores. Across a GC 1790 // point the Store Barrier and the card mark memory has to agree. I cannot 1791 // have a card mark store and its barrier split across the GC point from 1792 // either above or below. Here I get that to happen by reading ALL of memory. 1793 // A better answer would be to separate out card marks from other memory. 1794 // For now, return the input memory state, so that it can be reused 1795 // after the call, if this call has restricted memory effects. 1796 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1797 // Set fixed predefined input arguments 1798 Node* memory = reset_memory(); 1799 call->init_req( TypeFunc::Control, control() ); 1800 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1801 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1802 call->init_req( TypeFunc::FramePtr, frameptr() ); 1803 call->init_req( TypeFunc::ReturnAdr, top() ); 1804 return memory; 1805 } 1806 1807 //-------------------set_predefined_output_for_runtime_call-------------------- 1808 // Set control and memory (not i_o) from the call. 1809 // If keep_mem is not NULL, use it for the output state, 1810 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1811 // If hook_mem is NULL, this call produces no memory effects at all. 1812 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1813 // then only that memory slice is taken from the call. 1814 // In the last case, we must put an appropriate memory barrier before 1815 // the call, so as to create the correct anti-dependencies on loads 1816 // preceding the call. 1817 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1818 Node* keep_mem, 1819 const TypePtr* hook_mem) { 1820 // no i/o 1821 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1822 if (keep_mem) { 1823 // First clone the existing memory state 1824 set_all_memory(keep_mem); 1825 if (hook_mem != NULL) { 1826 // Make memory for the call 1827 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1828 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1829 // We also use hook_mem to extract specific effects from arraycopy stubs. 1830 set_memory(mem, hook_mem); 1831 } 1832 // ...else the call has NO memory effects. 1833 1834 // Make sure the call advertises its memory effects precisely. 1835 // This lets us build accurate anti-dependences in gcm.cpp. 1836 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1837 "call node must be constructed correctly"); 1838 } else { 1839 assert(hook_mem == NULL, ""); 1840 // This is not a "slow path" call; all memory comes from the call. 1841 set_all_memory_call(call); 1842 } 1843 } 1844 1845 1846 // Replace the call with the current state of the kit. 1847 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1848 JVMState* ejvms = NULL; 1849 if (has_exceptions()) { 1850 ejvms = transfer_exceptions_into_jvms(); 1851 } 1852 1853 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1854 ReplacedNodes replaced_nodes_exception; 1855 Node* ex_ctl = top(); 1856 1857 SafePointNode* final_state = stop(); 1858 1859 // Find all the needed outputs of this call 1860 CallProjections callprojs; 1861 call->extract_projections(&callprojs, true); 1862 1863 Node* init_mem = call->in(TypeFunc::Memory); 1864 Node* final_mem = final_state->in(TypeFunc::Memory); 1865 Node* final_ctl = final_state->in(TypeFunc::Control); 1866 Node* final_io = final_state->in(TypeFunc::I_O); 1867 1868 // Replace all the old call edges with the edges from the inlining result 1869 if (callprojs.fallthrough_catchproj != NULL) { 1870 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1871 } 1872 if (callprojs.fallthrough_memproj != NULL) { 1873 if (final_mem->is_MergeMem()) { 1874 // Parser's exits MergeMem was not transformed but may be optimized 1875 final_mem = _gvn.transform(final_mem); 1876 } 1877 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1878 } 1879 if (callprojs.fallthrough_ioproj != NULL) { 1880 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1881 } 1882 1883 // Replace the result with the new result if it exists and is used 1884 if (callprojs.resproj != NULL && result != NULL) { 1885 C->gvn_replace_by(callprojs.resproj, result); 1886 } 1887 1888 if (ejvms == NULL) { 1889 // No exception edges to simply kill off those paths 1890 if (callprojs.catchall_catchproj != NULL) { 1891 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1892 } 1893 if (callprojs.catchall_memproj != NULL) { 1894 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1895 } 1896 if (callprojs.catchall_ioproj != NULL) { 1897 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1898 } 1899 // Replace the old exception object with top 1900 if (callprojs.exobj != NULL) { 1901 C->gvn_replace_by(callprojs.exobj, C->top()); 1902 } 1903 } else { 1904 GraphKit ekit(ejvms); 1905 1906 // Load my combined exception state into the kit, with all phis transformed: 1907 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1908 replaced_nodes_exception = ex_map->replaced_nodes(); 1909 1910 Node* ex_oop = ekit.use_exception_state(ex_map); 1911 1912 if (callprojs.catchall_catchproj != NULL) { 1913 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1914 ex_ctl = ekit.control(); 1915 } 1916 if (callprojs.catchall_memproj != NULL) { 1917 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1918 } 1919 if (callprojs.catchall_ioproj != NULL) { 1920 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1921 } 1922 1923 // Replace the old exception object with the newly created one 1924 if (callprojs.exobj != NULL) { 1925 C->gvn_replace_by(callprojs.exobj, ex_oop); 1926 } 1927 } 1928 1929 // Disconnect the call from the graph 1930 call->disconnect_inputs(NULL, C); 1931 C->gvn_replace_by(call, C->top()); 1932 1933 // Clean up any MergeMems that feed other MergeMems since the 1934 // optimizer doesn't like that. 1935 if (final_mem->is_MergeMem()) { 1936 Node_List wl; 1937 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1938 Node* m = i.get(); 1939 if (m->is_MergeMem() && !wl.contains(m)) { 1940 wl.push(m); 1941 } 1942 } 1943 while (wl.size() > 0) { 1944 _gvn.transform(wl.pop()); 1945 } 1946 } 1947 1948 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1949 replaced_nodes.apply(C, final_ctl); 1950 } 1951 if (!ex_ctl->is_top() && do_replaced_nodes) { 1952 replaced_nodes_exception.apply(C, ex_ctl); 1953 } 1954 } 1955 1956 1957 //------------------------------increment_counter------------------------------ 1958 // for statistics: increment a VM counter by 1 1959 1960 void GraphKit::increment_counter(address counter_addr) { 1961 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1962 increment_counter(adr1); 1963 } 1964 1965 void GraphKit::increment_counter(Node* counter_addr) { 1966 int adr_type = Compile::AliasIdxRaw; 1967 Node* ctrl = control(); 1968 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1969 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1970 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1971 } 1972 1973 1974 //------------------------------uncommon_trap---------------------------------- 1975 // Bail out to the interpreter in mid-method. Implemented by calling the 1976 // uncommon_trap blob. This helper function inserts a runtime call with the 1977 // right debug info. 1978 void GraphKit::uncommon_trap(int trap_request, 1979 ciKlass* klass, const char* comment, 1980 bool must_throw, 1981 bool keep_exact_action) { 1982 if (failing()) stop(); 1983 if (stopped()) return; // trap reachable? 1984 1985 // Note: If ProfileTraps is true, and if a deopt. actually 1986 // occurs here, the runtime will make sure an MDO exists. There is 1987 // no need to call method()->ensure_method_data() at this point. 1988 1989 // Set the stack pointer to the right value for reexecution: 1990 set_sp(reexecute_sp()); 1991 1992 #ifdef ASSERT 1993 if (!must_throw) { 1994 // Make sure the stack has at least enough depth to execute 1995 // the current bytecode. 1996 int inputs, ignored_depth; 1997 if (compute_stack_effects(inputs, ignored_depth)) { 1998 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1999 Bytecodes::name(java_bc()), sp(), inputs); 2000 } 2001 } 2002 #endif 2003 2004 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2005 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2006 2007 switch (action) { 2008 case Deoptimization::Action_maybe_recompile: 2009 case Deoptimization::Action_reinterpret: 2010 // Temporary fix for 6529811 to allow virtual calls to be sure they 2011 // get the chance to go from mono->bi->mega 2012 if (!keep_exact_action && 2013 Deoptimization::trap_request_index(trap_request) < 0 && 2014 too_many_recompiles(reason)) { 2015 // This BCI is causing too many recompilations. 2016 if (C->log() != NULL) { 2017 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2018 Deoptimization::trap_reason_name(reason), 2019 Deoptimization::trap_action_name(action)); 2020 } 2021 action = Deoptimization::Action_none; 2022 trap_request = Deoptimization::make_trap_request(reason, action); 2023 } else { 2024 C->set_trap_can_recompile(true); 2025 } 2026 break; 2027 case Deoptimization::Action_make_not_entrant: 2028 C->set_trap_can_recompile(true); 2029 break; 2030 case Deoptimization::Action_none: 2031 case Deoptimization::Action_make_not_compilable: 2032 break; 2033 default: 2034 #ifdef ASSERT 2035 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2036 #endif 2037 break; 2038 } 2039 2040 if (TraceOptoParse) { 2041 char buf[100]; 2042 tty->print_cr("Uncommon trap %s at bci:%d", 2043 Deoptimization::format_trap_request(buf, sizeof(buf), 2044 trap_request), bci()); 2045 } 2046 2047 CompileLog* log = C->log(); 2048 if (log != NULL) { 2049 int kid = (klass == NULL)? -1: log->identify(klass); 2050 log->begin_elem("uncommon_trap bci='%d'", bci()); 2051 char buf[100]; 2052 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2053 trap_request)); 2054 if (kid >= 0) log->print(" klass='%d'", kid); 2055 if (comment != NULL) log->print(" comment='%s'", comment); 2056 log->end_elem(); 2057 } 2058 2059 // Make sure any guarding test views this path as very unlikely 2060 Node *i0 = control()->in(0); 2061 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2062 IfNode *iff = i0->as_If(); 2063 float f = iff->_prob; // Get prob 2064 if (control()->Opcode() == Op_IfTrue) { 2065 if (f > PROB_UNLIKELY_MAG(4)) 2066 iff->_prob = PROB_MIN; 2067 } else { 2068 if (f < PROB_LIKELY_MAG(4)) 2069 iff->_prob = PROB_MAX; 2070 } 2071 } 2072 2073 // Clear out dead values from the debug info. 2074 kill_dead_locals(); 2075 2076 // Now insert the uncommon trap subroutine call 2077 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2078 const TypePtr* no_memory_effects = NULL; 2079 // Pass the index of the class to be loaded 2080 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2081 (must_throw ? RC_MUST_THROW : 0), 2082 OptoRuntime::uncommon_trap_Type(), 2083 call_addr, "uncommon_trap", no_memory_effects, 2084 intcon(trap_request)); 2085 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2086 "must extract request correctly from the graph"); 2087 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2088 2089 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2090 // The debug info is the only real input to this call. 2091 2092 // Halt-and-catch fire here. The above call should never return! 2093 HaltNode* halt = new HaltNode(control(), frameptr()); 2094 _gvn.set_type_bottom(halt); 2095 root()->add_req(halt); 2096 2097 stop_and_kill_map(); 2098 } 2099 2100 2101 //--------------------------just_allocated_object------------------------------ 2102 // Report the object that was just allocated. 2103 // It must be the case that there are no intervening safepoints. 2104 // We use this to determine if an object is so "fresh" that 2105 // it does not require card marks. 2106 Node* GraphKit::just_allocated_object(Node* current_control) { 2107 if (C->recent_alloc_ctl() == current_control) 2108 return C->recent_alloc_obj(); 2109 return NULL; 2110 } 2111 2112 2113 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2114 // (Note: TypeFunc::make has a cache that makes this fast.) 2115 const TypeFunc* tf = TypeFunc::make(dest_method); 2116 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2117 for (int j = 0; j < nargs; j++) { 2118 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2119 if( targ->basic_type() == T_DOUBLE ) { 2120 // If any parameters are doubles, they must be rounded before 2121 // the call, dstore_rounding does gvn.transform 2122 Node *arg = argument(j); 2123 arg = dstore_rounding(arg); 2124 set_argument(j, arg); 2125 } 2126 } 2127 } 2128 2129 /** 2130 * Record profiling data exact_kls for Node n with the type system so 2131 * that it can propagate it (speculation) 2132 * 2133 * @param n node that the type applies to 2134 * @param exact_kls type from profiling 2135 * @param maybe_null did profiling see null? 2136 * 2137 * @return node with improved type 2138 */ 2139 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2140 const Type* current_type = _gvn.type(n); 2141 assert(UseTypeSpeculation, "type speculation must be on"); 2142 2143 const TypePtr* speculative = current_type->speculative(); 2144 2145 // Should the klass from the profile be recorded in the speculative type? 2146 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2147 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2148 const TypeOopPtr* xtype = tklass->as_instance_type(); 2149 assert(xtype->klass_is_exact(), "Should be exact"); 2150 // Any reason to believe n is not null (from this profiling or a previous one)? 2151 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2152 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2153 // record the new speculative type's depth 2154 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2155 speculative = speculative->with_inline_depth(jvms()->depth()); 2156 } else if (current_type->would_improve_ptr(ptr_kind)) { 2157 // Profiling report that null was never seen so we can change the 2158 // speculative type to non null ptr. 2159 if (ptr_kind == ProfileAlwaysNull) { 2160 speculative = TypePtr::NULL_PTR; 2161 } else { 2162 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2163 const TypePtr* ptr = TypePtr::NOTNULL; 2164 if (speculative != NULL) { 2165 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2166 } else { 2167 speculative = ptr; 2168 } 2169 } 2170 } 2171 2172 if (speculative != current_type->speculative()) { 2173 // Build a type with a speculative type (what we think we know 2174 // about the type but will need a guard when we use it) 2175 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2176 // We're changing the type, we need a new CheckCast node to carry 2177 // the new type. The new type depends on the control: what 2178 // profiling tells us is only valid from here as far as we can 2179 // tell. 2180 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2181 cast = _gvn.transform(cast); 2182 replace_in_map(n, cast); 2183 n = cast; 2184 } 2185 2186 return n; 2187 } 2188 2189 /** 2190 * Record profiling data from receiver profiling at an invoke with the 2191 * type system so that it can propagate it (speculation) 2192 * 2193 * @param n receiver node 2194 * 2195 * @return node with improved type 2196 */ 2197 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2198 if (!UseTypeSpeculation) { 2199 return n; 2200 } 2201 ciKlass* exact_kls = profile_has_unique_klass(); 2202 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2203 if ((java_bc() == Bytecodes::_checkcast || 2204 java_bc() == Bytecodes::_instanceof || 2205 java_bc() == Bytecodes::_aastore) && 2206 method()->method_data()->is_mature()) { 2207 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2208 if (data != NULL) { 2209 if (!data->as_BitData()->null_seen()) { 2210 ptr_kind = ProfileNeverNull; 2211 } else { 2212 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2213 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2214 uint i = 0; 2215 for (; i < call->row_limit(); i++) { 2216 ciKlass* receiver = call->receiver(i); 2217 if (receiver != NULL) { 2218 break; 2219 } 2220 } 2221 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2222 } 2223 } 2224 } 2225 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2226 } 2227 2228 /** 2229 * Record profiling data from argument profiling at an invoke with the 2230 * type system so that it can propagate it (speculation) 2231 * 2232 * @param dest_method target method for the call 2233 * @param bc what invoke bytecode is this? 2234 */ 2235 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2236 if (!UseTypeSpeculation) { 2237 return; 2238 } 2239 const TypeFunc* tf = TypeFunc::make(dest_method); 2240 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2241 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2242 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2243 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2244 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2245 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2246 ciKlass* better_type = NULL; 2247 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2248 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2249 } 2250 i++; 2251 } 2252 } 2253 } 2254 2255 /** 2256 * Record profiling data from parameter profiling at an invoke with 2257 * the type system so that it can propagate it (speculation) 2258 */ 2259 void GraphKit::record_profiled_parameters_for_speculation() { 2260 if (!UseTypeSpeculation) { 2261 return; 2262 } 2263 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2264 if (_gvn.type(local(i))->isa_oopptr()) { 2265 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2266 ciKlass* better_type = NULL; 2267 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2268 record_profile_for_speculation(local(i), better_type, ptr_kind); 2269 } 2270 j++; 2271 } 2272 } 2273 } 2274 2275 /** 2276 * Record profiling data from return value profiling at an invoke with 2277 * the type system so that it can propagate it (speculation) 2278 */ 2279 void GraphKit::record_profiled_return_for_speculation() { 2280 if (!UseTypeSpeculation) { 2281 return; 2282 } 2283 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2284 ciKlass* better_type = NULL; 2285 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2286 // If profiling reports a single type for the return value, 2287 // feed it to the type system so it can propagate it as a 2288 // speculative type 2289 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2290 } 2291 } 2292 2293 void GraphKit::round_double_result(ciMethod* dest_method) { 2294 // A non-strict method may return a double value which has an extended 2295 // exponent, but this must not be visible in a caller which is 'strict' 2296 // If a strict caller invokes a non-strict callee, round a double result 2297 2298 BasicType result_type = dest_method->return_type()->basic_type(); 2299 assert( method() != NULL, "must have caller context"); 2300 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2301 // Destination method's return value is on top of stack 2302 // dstore_rounding() does gvn.transform 2303 Node *result = pop_pair(); 2304 result = dstore_rounding(result); 2305 push_pair(result); 2306 } 2307 } 2308 2309 // rounding for strict float precision conformance 2310 Node* GraphKit::precision_rounding(Node* n) { 2311 return UseStrictFP && _method->flags().is_strict() 2312 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2313 ? _gvn.transform( new RoundFloatNode(0, n) ) 2314 : n; 2315 } 2316 2317 // rounding for strict double precision conformance 2318 Node* GraphKit::dprecision_rounding(Node *n) { 2319 return UseStrictFP && _method->flags().is_strict() 2320 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2321 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2322 : n; 2323 } 2324 2325 // rounding for non-strict double stores 2326 Node* GraphKit::dstore_rounding(Node* n) { 2327 return Matcher::strict_fp_requires_explicit_rounding 2328 && UseSSE <= 1 2329 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2330 : n; 2331 } 2332 2333 //============================================================================= 2334 // Generate a fast path/slow path idiom. Graph looks like: 2335 // [foo] indicates that 'foo' is a parameter 2336 // 2337 // [in] NULL 2338 // \ / 2339 // CmpP 2340 // Bool ne 2341 // If 2342 // / \ 2343 // True False-<2> 2344 // / | 2345 // / cast_not_null 2346 // Load | | ^ 2347 // [fast_test] | | 2348 // gvn to opt_test | | 2349 // / \ | <1> 2350 // True False | 2351 // | \\ | 2352 // [slow_call] \[fast_result] 2353 // Ctl Val \ \ 2354 // | \ \ 2355 // Catch <1> \ \ 2356 // / \ ^ \ \ 2357 // Ex No_Ex | \ \ 2358 // | \ \ | \ <2> \ 2359 // ... \ [slow_res] | | \ [null_result] 2360 // \ \--+--+--- | | 2361 // \ | / \ | / 2362 // --------Region Phi 2363 // 2364 //============================================================================= 2365 // Code is structured as a series of driver functions all called 'do_XXX' that 2366 // call a set of helper functions. Helper functions first, then drivers. 2367 2368 //------------------------------null_check_oop--------------------------------- 2369 // Null check oop. Set null-path control into Region in slot 3. 2370 // Make a cast-not-nullness use the other not-null control. Return cast. 2371 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2372 bool never_see_null, 2373 bool safe_for_replace, 2374 bool speculative) { 2375 // Initial NULL check taken path 2376 (*null_control) = top(); 2377 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2378 2379 // Generate uncommon_trap: 2380 if (never_see_null && (*null_control) != top()) { 2381 // If we see an unexpected null at a check-cast we record it and force a 2382 // recompile; the offending check-cast will be compiled to handle NULLs. 2383 // If we see more than one offending BCI, then all checkcasts in the 2384 // method will be compiled to handle NULLs. 2385 PreserveJVMState pjvms(this); 2386 set_control(*null_control); 2387 replace_in_map(value, null()); 2388 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2389 uncommon_trap(reason, 2390 Deoptimization::Action_make_not_entrant); 2391 (*null_control) = top(); // NULL path is dead 2392 } 2393 if ((*null_control) == top() && safe_for_replace) { 2394 replace_in_map(value, cast); 2395 } 2396 2397 // Cast away null-ness on the result 2398 return cast; 2399 } 2400 2401 //------------------------------opt_iff---------------------------------------- 2402 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2403 // Return slow-path control. 2404 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2405 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2406 2407 // Fast path taken; set region slot 2 2408 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2409 region->init_req(2,fast_taken); // Capture fast-control 2410 2411 // Fast path not-taken, i.e. slow path 2412 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2413 return slow_taken; 2414 } 2415 2416 //-----------------------------make_runtime_call------------------------------- 2417 Node* GraphKit::make_runtime_call(int flags, 2418 const TypeFunc* call_type, address call_addr, 2419 const char* call_name, 2420 const TypePtr* adr_type, 2421 // The following parms are all optional. 2422 // The first NULL ends the list. 2423 Node* parm0, Node* parm1, 2424 Node* parm2, Node* parm3, 2425 Node* parm4, Node* parm5, 2426 Node* parm6, Node* parm7) { 2427 // Slow-path call 2428 bool is_leaf = !(flags & RC_NO_LEAF); 2429 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2430 if (call_name == NULL) { 2431 assert(!is_leaf, "must supply name for leaf"); 2432 call_name = OptoRuntime::stub_name(call_addr); 2433 } 2434 CallNode* call; 2435 if (!is_leaf) { 2436 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2437 bci(), adr_type); 2438 } else if (flags & RC_NO_FP) { 2439 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2440 } else { 2441 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2442 } 2443 2444 // The following is similar to set_edges_for_java_call, 2445 // except that the memory effects of the call are restricted to AliasIdxRaw. 2446 2447 // Slow path call has no side-effects, uses few values 2448 bool wide_in = !(flags & RC_NARROW_MEM); 2449 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2450 2451 Node* prev_mem = NULL; 2452 if (wide_in) { 2453 prev_mem = set_predefined_input_for_runtime_call(call); 2454 } else { 2455 assert(!wide_out, "narrow in => narrow out"); 2456 Node* narrow_mem = memory(adr_type); 2457 prev_mem = reset_memory(); 2458 map()->set_memory(narrow_mem); 2459 set_predefined_input_for_runtime_call(call); 2460 } 2461 2462 // Hook each parm in order. Stop looking at the first NULL. 2463 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2464 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2465 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2466 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2467 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2468 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2469 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2470 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2471 /* close each nested if ===> */ } } } } } } } } 2472 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2473 2474 if (!is_leaf) { 2475 // Non-leaves can block and take safepoints: 2476 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2477 } 2478 // Non-leaves can throw exceptions: 2479 if (has_io) { 2480 call->set_req(TypeFunc::I_O, i_o()); 2481 } 2482 2483 if (flags & RC_UNCOMMON) { 2484 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2485 // (An "if" probability corresponds roughly to an unconditional count. 2486 // Sort of.) 2487 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2488 } 2489 2490 Node* c = _gvn.transform(call); 2491 assert(c == call, "cannot disappear"); 2492 2493 if (wide_out) { 2494 // Slow path call has full side-effects. 2495 set_predefined_output_for_runtime_call(call); 2496 } else { 2497 // Slow path call has few side-effects, and/or sets few values. 2498 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2499 } 2500 2501 if (has_io) { 2502 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2503 } 2504 return call; 2505 2506 } 2507 2508 //------------------------------merge_memory----------------------------------- 2509 // Merge memory from one path into the current memory state. 2510 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2511 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2512 Node* old_slice = mms.force_memory(); 2513 Node* new_slice = mms.memory2(); 2514 if (old_slice != new_slice) { 2515 PhiNode* phi; 2516 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2517 if (mms.is_empty()) { 2518 // clone base memory Phi's inputs for this memory slice 2519 assert(old_slice == mms.base_memory(), "sanity"); 2520 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2521 _gvn.set_type(phi, Type::MEMORY); 2522 for (uint i = 1; i < phi->req(); i++) { 2523 phi->init_req(i, old_slice->in(i)); 2524 } 2525 } else { 2526 phi = old_slice->as_Phi(); // Phi was generated already 2527 } 2528 } else { 2529 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2530 _gvn.set_type(phi, Type::MEMORY); 2531 } 2532 phi->set_req(new_path, new_slice); 2533 mms.set_memory(phi); 2534 } 2535 } 2536 } 2537 2538 //------------------------------make_slow_call_ex------------------------------ 2539 // Make the exception handler hookups for the slow call 2540 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2541 if (stopped()) return; 2542 2543 // Make a catch node with just two handlers: fall-through and catch-all 2544 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2545 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2546 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2547 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2548 2549 { PreserveJVMState pjvms(this); 2550 set_control(excp); 2551 set_i_o(i_o); 2552 2553 if (excp != top()) { 2554 if (deoptimize) { 2555 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2556 uncommon_trap(Deoptimization::Reason_unhandled, 2557 Deoptimization::Action_none); 2558 } else { 2559 // Create an exception state also. 2560 // Use an exact type if the caller has specified a specific exception. 2561 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2562 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2563 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2564 } 2565 } 2566 } 2567 2568 // Get the no-exception control from the CatchNode. 2569 set_control(norm); 2570 } 2571 2572 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2573 Node* cmp = NULL; 2574 switch(bt) { 2575 case T_INT: cmp = new CmpINode(in1, in2); break; 2576 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2577 default: fatal("unexpected comparison type %s", type2name(bt)); 2578 } 2579 gvn->transform(cmp); 2580 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2581 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2582 gvn->transform(iff); 2583 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2584 return iff; 2585 } 2586 2587 2588 //-------------------------------gen_subtype_check----------------------------- 2589 // Generate a subtyping check. Takes as input the subtype and supertype. 2590 // Returns 2 values: sets the default control() to the true path and returns 2591 // the false path. Only reads invariant memory; sets no (visible) memory. 2592 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2593 // but that's not exposed to the optimizer. This call also doesn't take in an 2594 // Object; if you wish to check an Object you need to load the Object's class 2595 // prior to coming here. 2596 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2597 Compile* C = gvn->C; 2598 2599 if ((*ctrl)->is_top()) { 2600 return C->top(); 2601 } 2602 2603 // Fast check for identical types, perhaps identical constants. 2604 // The types can even be identical non-constants, in cases 2605 // involving Array.newInstance, Object.clone, etc. 2606 if (subklass == superklass) 2607 return C->top(); // false path is dead; no test needed. 2608 2609 if (gvn->type(superklass)->singleton()) { 2610 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2611 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2612 2613 // In the common case of an exact superklass, try to fold up the 2614 // test before generating code. You may ask, why not just generate 2615 // the code and then let it fold up? The answer is that the generated 2616 // code will necessarily include null checks, which do not always 2617 // completely fold away. If they are also needless, then they turn 2618 // into a performance loss. Example: 2619 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2620 // Here, the type of 'fa' is often exact, so the store check 2621 // of fa[1]=x will fold up, without testing the nullness of x. 2622 switch (C->static_subtype_check(superk, subk)) { 2623 case Compile::SSC_always_false: 2624 { 2625 Node* always_fail = *ctrl; 2626 *ctrl = gvn->C->top(); 2627 return always_fail; 2628 } 2629 case Compile::SSC_always_true: 2630 return C->top(); 2631 case Compile::SSC_easy_test: 2632 { 2633 // Just do a direct pointer compare and be done. 2634 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2635 *ctrl = gvn->transform(new IfTrueNode(iff)); 2636 return gvn->transform(new IfFalseNode(iff)); 2637 } 2638 case Compile::SSC_full_test: 2639 break; 2640 default: 2641 ShouldNotReachHere(); 2642 } 2643 } 2644 2645 // %%% Possible further optimization: Even if the superklass is not exact, 2646 // if the subklass is the unique subtype of the superklass, the check 2647 // will always succeed. We could leave a dependency behind to ensure this. 2648 2649 // First load the super-klass's check-offset 2650 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2651 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2652 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2653 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2654 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2655 2656 // Load from the sub-klass's super-class display list, or a 1-word cache of 2657 // the secondary superclass list, or a failing value with a sentinel offset 2658 // if the super-klass is an interface or exceptionally deep in the Java 2659 // hierarchy and we have to scan the secondary superclass list the hard way. 2660 // Worst-case type is a little odd: NULL is allowed as a result (usually 2661 // klass loads can never produce a NULL). 2662 Node *chk_off_X = chk_off; 2663 #ifdef _LP64 2664 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2665 #endif 2666 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2667 // For some types like interfaces the following loadKlass is from a 1-word 2668 // cache which is mutable so can't use immutable memory. Other 2669 // types load from the super-class display table which is immutable. 2670 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2671 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2672 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2673 2674 // Compile speed common case: ARE a subtype and we canNOT fail 2675 if( superklass == nkls ) 2676 return C->top(); // false path is dead; no test needed. 2677 2678 // See if we get an immediate positive hit. Happens roughly 83% of the 2679 // time. Test to see if the value loaded just previously from the subklass 2680 // is exactly the superklass. 2681 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2682 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2683 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2684 2685 // Compile speed common case: Check for being deterministic right now. If 2686 // chk_off is a constant and not equal to cacheoff then we are NOT a 2687 // subklass. In this case we need exactly the 1 test above and we can 2688 // return those results immediately. 2689 if (!might_be_cache) { 2690 Node* not_subtype_ctrl = *ctrl; 2691 *ctrl = iftrue1; // We need exactly the 1 test above 2692 return not_subtype_ctrl; 2693 } 2694 2695 // Gather the various success & failures here 2696 RegionNode *r_ok_subtype = new RegionNode(4); 2697 gvn->record_for_igvn(r_ok_subtype); 2698 RegionNode *r_not_subtype = new RegionNode(3); 2699 gvn->record_for_igvn(r_not_subtype); 2700 2701 r_ok_subtype->init_req(1, iftrue1); 2702 2703 // Check for immediate negative hit. Happens roughly 11% of the time (which 2704 // is roughly 63% of the remaining cases). Test to see if the loaded 2705 // check-offset points into the subklass display list or the 1-element 2706 // cache. If it points to the display (and NOT the cache) and the display 2707 // missed then it's not a subtype. 2708 Node *cacheoff = gvn->intcon(cacheoff_con); 2709 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2710 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2711 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2712 2713 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2714 // No performance impact (too rare) but allows sharing of secondary arrays 2715 // which has some footprint reduction. 2716 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2717 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2718 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2719 2720 // -- Roads not taken here: -- 2721 // We could also have chosen to perform the self-check at the beginning 2722 // of this code sequence, as the assembler does. This would not pay off 2723 // the same way, since the optimizer, unlike the assembler, can perform 2724 // static type analysis to fold away many successful self-checks. 2725 // Non-foldable self checks work better here in second position, because 2726 // the initial primary superclass check subsumes a self-check for most 2727 // types. An exception would be a secondary type like array-of-interface, 2728 // which does not appear in its own primary supertype display. 2729 // Finally, we could have chosen to move the self-check into the 2730 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2731 // dependent manner. But it is worthwhile to have the check here, 2732 // where it can be perhaps be optimized. The cost in code space is 2733 // small (register compare, branch). 2734 2735 // Now do a linear scan of the secondary super-klass array. Again, no real 2736 // performance impact (too rare) but it's gotta be done. 2737 // Since the code is rarely used, there is no penalty for moving it 2738 // out of line, and it can only improve I-cache density. 2739 // The decision to inline or out-of-line this final check is platform 2740 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2741 Node* psc = gvn->transform( 2742 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2743 2744 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2745 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2746 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2747 2748 // Return false path; set default control to true path. 2749 *ctrl = gvn->transform(r_ok_subtype); 2750 return gvn->transform(r_not_subtype); 2751 } 2752 2753 // Profile-driven exact type check: 2754 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2755 float prob, 2756 Node* *casted_receiver) { 2757 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2758 Node* recv_klass = load_object_klass(receiver); 2759 Node* want_klass = makecon(tklass); 2760 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2761 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2762 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2763 set_control( _gvn.transform( new IfTrueNode (iff) )); 2764 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2765 2766 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2767 assert(recv_xtype->klass_is_exact(), ""); 2768 2769 // Subsume downstream occurrences of receiver with a cast to 2770 // recv_xtype, since now we know what the type will be. 2771 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2772 (*casted_receiver) = _gvn.transform(cast); 2773 // (User must make the replace_in_map call.) 2774 2775 return fail; 2776 } 2777 2778 2779 //------------------------------seems_never_null------------------------------- 2780 // Use null_seen information if it is available from the profile. 2781 // If we see an unexpected null at a type check we record it and force a 2782 // recompile; the offending check will be recompiled to handle NULLs. 2783 // If we see several offending BCIs, then all checks in the 2784 // method will be recompiled. 2785 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2786 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2787 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2788 if (UncommonNullCast // Cutout for this technique 2789 && obj != null() // And not the -Xcomp stupid case? 2790 && !too_many_traps(reason) 2791 ) { 2792 if (speculating) { 2793 return true; 2794 } 2795 if (data == NULL) 2796 // Edge case: no mature data. Be optimistic here. 2797 return true; 2798 // If the profile has not seen a null, assume it won't happen. 2799 assert(java_bc() == Bytecodes::_checkcast || 2800 java_bc() == Bytecodes::_instanceof || 2801 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2802 return !data->as_BitData()->null_seen(); 2803 } 2804 speculating = false; 2805 return false; 2806 } 2807 2808 //------------------------maybe_cast_profiled_receiver------------------------- 2809 // If the profile has seen exactly one type, narrow to exactly that type. 2810 // Subsequent type checks will always fold up. 2811 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2812 ciKlass* require_klass, 2813 ciKlass* spec_klass, 2814 bool safe_for_replace) { 2815 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2816 2817 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2818 2819 // Make sure we haven't already deoptimized from this tactic. 2820 if (too_many_traps(reason) || too_many_recompiles(reason)) 2821 return NULL; 2822 2823 // (No, this isn't a call, but it's enough like a virtual call 2824 // to use the same ciMethod accessor to get the profile info...) 2825 // If we have a speculative type use it instead of profiling (which 2826 // may not help us) 2827 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2828 if (exact_kls != NULL) {// no cast failures here 2829 if (require_klass == NULL || 2830 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2831 // If we narrow the type to match what the type profile sees or 2832 // the speculative type, we can then remove the rest of the 2833 // cast. 2834 // This is a win, even if the exact_kls is very specific, 2835 // because downstream operations, such as method calls, 2836 // will often benefit from the sharper type. 2837 Node* exact_obj = not_null_obj; // will get updated in place... 2838 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2839 &exact_obj); 2840 { PreserveJVMState pjvms(this); 2841 set_control(slow_ctl); 2842 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2843 } 2844 if (safe_for_replace) { 2845 replace_in_map(not_null_obj, exact_obj); 2846 } 2847 return exact_obj; 2848 } 2849 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2850 } 2851 2852 return NULL; 2853 } 2854 2855 /** 2856 * Cast obj to type and emit guard unless we had too many traps here 2857 * already 2858 * 2859 * @param obj node being casted 2860 * @param type type to cast the node to 2861 * @param not_null true if we know node cannot be null 2862 */ 2863 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2864 ciKlass* type, 2865 bool not_null) { 2866 if (stopped()) { 2867 return obj; 2868 } 2869 2870 // type == NULL if profiling tells us this object is always null 2871 if (type != NULL) { 2872 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2873 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2874 2875 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2876 !too_many_traps(class_reason) && 2877 !too_many_recompiles(class_reason)) { 2878 Node* not_null_obj = NULL; 2879 // not_null is true if we know the object is not null and 2880 // there's no need for a null check 2881 if (!not_null) { 2882 Node* null_ctl = top(); 2883 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2884 assert(null_ctl->is_top(), "no null control here"); 2885 } else { 2886 not_null_obj = obj; 2887 } 2888 2889 Node* exact_obj = not_null_obj; 2890 ciKlass* exact_kls = type; 2891 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2892 &exact_obj); 2893 { 2894 PreserveJVMState pjvms(this); 2895 set_control(slow_ctl); 2896 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2897 } 2898 replace_in_map(not_null_obj, exact_obj); 2899 obj = exact_obj; 2900 } 2901 } else { 2902 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2903 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2904 Node* exact_obj = null_assert(obj); 2905 replace_in_map(obj, exact_obj); 2906 obj = exact_obj; 2907 } 2908 } 2909 return obj; 2910 } 2911 2912 //-------------------------------gen_instanceof-------------------------------- 2913 // Generate an instance-of idiom. Used by both the instance-of bytecode 2914 // and the reflective instance-of call. 2915 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2916 kill_dead_locals(); // Benefit all the uncommon traps 2917 assert( !stopped(), "dead parse path should be checked in callers" ); 2918 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2919 "must check for not-null not-dead klass in callers"); 2920 2921 // Make the merge point 2922 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2923 RegionNode* region = new RegionNode(PATH_LIMIT); 2924 Node* phi = new PhiNode(region, TypeInt::BOOL); 2925 C->set_has_split_ifs(true); // Has chance for split-if optimization 2926 2927 ciProfileData* data = NULL; 2928 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2929 data = method()->method_data()->bci_to_data(bci()); 2930 } 2931 bool speculative_not_null = false; 2932 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2933 && seems_never_null(obj, data, speculative_not_null)); 2934 2935 // Null check; get casted pointer; set region slot 3 2936 Node* null_ctl = top(); 2937 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2938 2939 // If not_null_obj is dead, only null-path is taken 2940 if (stopped()) { // Doing instance-of on a NULL? 2941 set_control(null_ctl); 2942 return intcon(0); 2943 } 2944 region->init_req(_null_path, null_ctl); 2945 phi ->init_req(_null_path, intcon(0)); // Set null path value 2946 if (null_ctl == top()) { 2947 // Do this eagerly, so that pattern matches like is_diamond_phi 2948 // will work even during parsing. 2949 assert(_null_path == PATH_LIMIT-1, "delete last"); 2950 region->del_req(_null_path); 2951 phi ->del_req(_null_path); 2952 } 2953 2954 // Do we know the type check always succeed? 2955 bool known_statically = false; 2956 if (_gvn.type(superklass)->singleton()) { 2957 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2958 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2959 if (subk != NULL && subk->is_loaded()) { 2960 int static_res = C->static_subtype_check(superk, subk); 2961 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2962 } 2963 } 2964 2965 if (!known_statically) { 2966 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2967 // We may not have profiling here or it may not help us. If we 2968 // have a speculative type use it to perform an exact cast. 2969 ciKlass* spec_obj_type = obj_type->speculative_type(); 2970 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2971 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2972 if (stopped()) { // Profile disagrees with this path. 2973 set_control(null_ctl); // Null is the only remaining possibility. 2974 return intcon(0); 2975 } 2976 if (cast_obj != NULL) { 2977 not_null_obj = cast_obj; 2978 } 2979 } 2980 } 2981 2982 // Load the object's klass 2983 Node* obj_klass = load_object_klass(not_null_obj); 2984 2985 // Generate the subtype check 2986 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2987 2988 // Plug in the success path to the general merge in slot 1. 2989 region->init_req(_obj_path, control()); 2990 phi ->init_req(_obj_path, intcon(1)); 2991 2992 // Plug in the failing path to the general merge in slot 2. 2993 region->init_req(_fail_path, not_subtype_ctrl); 2994 phi ->init_req(_fail_path, intcon(0)); 2995 2996 // Return final merged results 2997 set_control( _gvn.transform(region) ); 2998 record_for_igvn(region); 2999 3000 // If we know the type check always succeeds then we don't use the 3001 // profiling data at this bytecode. Don't lose it, feed it to the 3002 // type system as a speculative type. 3003 if (safe_for_replace) { 3004 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3005 replace_in_map(obj, casted_obj); 3006 } 3007 3008 return _gvn.transform(phi); 3009 } 3010 3011 //-------------------------------gen_checkcast--------------------------------- 3012 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3013 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3014 // uncommon-trap paths work. Adjust stack after this call. 3015 // If failure_control is supplied and not null, it is filled in with 3016 // the control edge for the cast failure. Otherwise, an appropriate 3017 // uncommon trap or exception is thrown. 3018 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3019 Node* *failure_control) { 3020 kill_dead_locals(); // Benefit all the uncommon traps 3021 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3022 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3023 3024 // Fast cutout: Check the case that the cast is vacuously true. 3025 // This detects the common cases where the test will short-circuit 3026 // away completely. We do this before we perform the null check, 3027 // because if the test is going to turn into zero code, we don't 3028 // want a residual null check left around. (Causes a slowdown, 3029 // for example, in some objArray manipulations, such as a[i]=a[j].) 3030 if (tk->singleton()) { 3031 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3032 if (objtp != NULL && objtp->klass() != NULL) { 3033 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3034 case Compile::SSC_always_true: 3035 // If we know the type check always succeed then we don't use 3036 // the profiling data at this bytecode. Don't lose it, feed it 3037 // to the type system as a speculative type. 3038 return record_profiled_receiver_for_speculation(obj); 3039 case Compile::SSC_always_false: 3040 // It needs a null check because a null will *pass* the cast check. 3041 // A non-null value will always produce an exception. 3042 return null_assert(obj); 3043 } 3044 } 3045 } 3046 3047 ciProfileData* data = NULL; 3048 bool safe_for_replace = false; 3049 if (failure_control == NULL) { // use MDO in regular case only 3050 assert(java_bc() == Bytecodes::_aastore || 3051 java_bc() == Bytecodes::_checkcast, 3052 "interpreter profiles type checks only for these BCs"); 3053 data = method()->method_data()->bci_to_data(bci()); 3054 safe_for_replace = true; 3055 } 3056 3057 // Make the merge point 3058 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3059 RegionNode* region = new RegionNode(PATH_LIMIT); 3060 Node* phi = new PhiNode(region, toop); 3061 C->set_has_split_ifs(true); // Has chance for split-if optimization 3062 3063 // Use null-cast information if it is available 3064 bool speculative_not_null = false; 3065 bool never_see_null = ((failure_control == NULL) // regular case only 3066 && seems_never_null(obj, data, speculative_not_null)); 3067 3068 // Null check; get casted pointer; set region slot 3 3069 Node* null_ctl = top(); 3070 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3071 3072 // If not_null_obj is dead, only null-path is taken 3073 if (stopped()) { // Doing instance-of on a NULL? 3074 set_control(null_ctl); 3075 return null(); 3076 } 3077 region->init_req(_null_path, null_ctl); 3078 phi ->init_req(_null_path, null()); // Set null path value 3079 if (null_ctl == top()) { 3080 // Do this eagerly, so that pattern matches like is_diamond_phi 3081 // will work even during parsing. 3082 assert(_null_path == PATH_LIMIT-1, "delete last"); 3083 region->del_req(_null_path); 3084 phi ->del_req(_null_path); 3085 } 3086 3087 Node* cast_obj = NULL; 3088 if (tk->klass_is_exact()) { 3089 // The following optimization tries to statically cast the speculative type of the object 3090 // (for example obtained during profiling) to the type of the superklass and then do a 3091 // dynamic check that the type of the object is what we expect. To work correctly 3092 // for checkcast and aastore the type of superklass should be exact. 3093 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3094 // We may not have profiling here or it may not help us. If we have 3095 // a speculative type use it to perform an exact cast. 3096 ciKlass* spec_obj_type = obj_type->speculative_type(); 3097 if (spec_obj_type != NULL || data != NULL) { 3098 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3099 if (cast_obj != NULL) { 3100 if (failure_control != NULL) // failure is now impossible 3101 (*failure_control) = top(); 3102 // adjust the type of the phi to the exact klass: 3103 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3104 } 3105 } 3106 } 3107 3108 if (cast_obj == NULL) { 3109 // Load the object's klass 3110 Node* obj_klass = load_object_klass(not_null_obj); 3111 3112 // Generate the subtype check 3113 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3114 3115 // Plug in success path into the merge 3116 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3117 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3118 if (failure_control == NULL) { 3119 if (not_subtype_ctrl != top()) { // If failure is possible 3120 PreserveJVMState pjvms(this); 3121 set_control(not_subtype_ctrl); 3122 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3123 } 3124 } else { 3125 (*failure_control) = not_subtype_ctrl; 3126 } 3127 } 3128 3129 region->init_req(_obj_path, control()); 3130 phi ->init_req(_obj_path, cast_obj); 3131 3132 // A merge of NULL or Casted-NotNull obj 3133 Node* res = _gvn.transform(phi); 3134 3135 // Note I do NOT always 'replace_in_map(obj,result)' here. 3136 // if( tk->klass()->can_be_primary_super() ) 3137 // This means that if I successfully store an Object into an array-of-String 3138 // I 'forget' that the Object is really now known to be a String. I have to 3139 // do this because we don't have true union types for interfaces - if I store 3140 // a Baz into an array-of-Interface and then tell the optimizer it's an 3141 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3142 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3143 // replace_in_map( obj, res ); 3144 3145 // Return final merged results 3146 set_control( _gvn.transform(region) ); 3147 record_for_igvn(region); 3148 3149 return record_profiled_receiver_for_speculation(res); 3150 } 3151 3152 //------------------------------next_monitor----------------------------------- 3153 // What number should be given to the next monitor? 3154 int GraphKit::next_monitor() { 3155 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3156 int next = current + C->sync_stack_slots(); 3157 // Keep the toplevel high water mark current: 3158 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3159 return current; 3160 } 3161 3162 //------------------------------insert_mem_bar--------------------------------- 3163 // Memory barrier to avoid floating things around 3164 // The membar serves as a pinch point between both control and all memory slices. 3165 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3166 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3167 mb->init_req(TypeFunc::Control, control()); 3168 mb->init_req(TypeFunc::Memory, reset_memory()); 3169 Node* membar = _gvn.transform(mb); 3170 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3171 set_all_memory_call(membar); 3172 return membar; 3173 } 3174 3175 //-------------------------insert_mem_bar_volatile---------------------------- 3176 // Memory barrier to avoid floating things around 3177 // The membar serves as a pinch point between both control and memory(alias_idx). 3178 // If you want to make a pinch point on all memory slices, do not use this 3179 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3180 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3181 // When Parse::do_put_xxx updates a volatile field, it appends a series 3182 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3183 // The first membar is on the same memory slice as the field store opcode. 3184 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3185 // All the other membars (for other volatile slices, including AliasIdxBot, 3186 // which stands for all unknown volatile slices) are control-dependent 3187 // on the first membar. This prevents later volatile loads or stores 3188 // from sliding up past the just-emitted store. 3189 3190 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3191 mb->set_req(TypeFunc::Control,control()); 3192 if (alias_idx == Compile::AliasIdxBot) { 3193 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3194 } else { 3195 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3196 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3197 } 3198 Node* membar = _gvn.transform(mb); 3199 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3200 if (alias_idx == Compile::AliasIdxBot) { 3201 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3202 } else { 3203 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3204 } 3205 return membar; 3206 } 3207 3208 //------------------------------shared_lock------------------------------------ 3209 // Emit locking code. 3210 FastLockNode* GraphKit::shared_lock(Node* obj) { 3211 // bci is either a monitorenter bc or InvocationEntryBci 3212 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3213 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3214 3215 if( !GenerateSynchronizationCode ) 3216 return NULL; // Not locking things? 3217 if (stopped()) // Dead monitor? 3218 return NULL; 3219 3220 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3221 3222 // Box the stack location 3223 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3224 Node* mem = reset_memory(); 3225 3226 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3227 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3228 // Create the counters for this fast lock. 3229 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3230 } 3231 3232 // Create the rtm counters for this fast lock if needed. 3233 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3234 3235 // Add monitor to debug info for the slow path. If we block inside the 3236 // slow path and de-opt, we need the monitor hanging around 3237 map()->push_monitor( flock ); 3238 3239 const TypeFunc *tf = LockNode::lock_type(); 3240 LockNode *lock = new LockNode(C, tf); 3241 3242 lock->init_req( TypeFunc::Control, control() ); 3243 lock->init_req( TypeFunc::Memory , mem ); 3244 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3245 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3246 lock->init_req( TypeFunc::ReturnAdr, top() ); 3247 3248 lock->init_req(TypeFunc::Parms + 0, obj); 3249 lock->init_req(TypeFunc::Parms + 1, box); 3250 lock->init_req(TypeFunc::Parms + 2, flock); 3251 add_safepoint_edges(lock); 3252 3253 lock = _gvn.transform( lock )->as_Lock(); 3254 3255 // lock has no side-effects, sets few values 3256 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3257 3258 insert_mem_bar(Op_MemBarAcquireLock); 3259 3260 // Add this to the worklist so that the lock can be eliminated 3261 record_for_igvn(lock); 3262 3263 #ifndef PRODUCT 3264 if (PrintLockStatistics) { 3265 // Update the counter for this lock. Don't bother using an atomic 3266 // operation since we don't require absolute accuracy. 3267 lock->create_lock_counter(map()->jvms()); 3268 increment_counter(lock->counter()->addr()); 3269 } 3270 #endif 3271 3272 return flock; 3273 } 3274 3275 3276 //------------------------------shared_unlock---------------------------------- 3277 // Emit unlocking code. 3278 void GraphKit::shared_unlock(Node* box, Node* obj) { 3279 // bci is either a monitorenter bc or InvocationEntryBci 3280 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3281 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3282 3283 if( !GenerateSynchronizationCode ) 3284 return; 3285 if (stopped()) { // Dead monitor? 3286 map()->pop_monitor(); // Kill monitor from debug info 3287 return; 3288 } 3289 3290 // Memory barrier to avoid floating things down past the locked region 3291 insert_mem_bar(Op_MemBarReleaseLock); 3292 3293 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3294 UnlockNode *unlock = new UnlockNode(C, tf); 3295 #ifdef ASSERT 3296 unlock->set_dbg_jvms(sync_jvms()); 3297 #endif 3298 uint raw_idx = Compile::AliasIdxRaw; 3299 unlock->init_req( TypeFunc::Control, control() ); 3300 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3301 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3302 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3303 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3304 3305 unlock->init_req(TypeFunc::Parms + 0, obj); 3306 unlock->init_req(TypeFunc::Parms + 1, box); 3307 unlock = _gvn.transform(unlock)->as_Unlock(); 3308 3309 Node* mem = reset_memory(); 3310 3311 // unlock has no side-effects, sets few values 3312 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3313 3314 // Kill monitor from debug info 3315 map()->pop_monitor( ); 3316 } 3317 3318 //-------------------------------get_layout_helper----------------------------- 3319 // If the given klass is a constant or known to be an array, 3320 // fetch the constant layout helper value into constant_value 3321 // and return (Node*)NULL. Otherwise, load the non-constant 3322 // layout helper value, and return the node which represents it. 3323 // This two-faced routine is useful because allocation sites 3324 // almost always feature constant types. 3325 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3326 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3327 if (!StressReflectiveCode && inst_klass != NULL) { 3328 ciKlass* klass = inst_klass->klass(); 3329 bool xklass = inst_klass->klass_is_exact(); 3330 if (xklass || klass->is_array_klass()) { 3331 jint lhelper = klass->layout_helper(); 3332 if (lhelper != Klass::_lh_neutral_value) { 3333 constant_value = lhelper; 3334 return (Node*) NULL; 3335 } 3336 } 3337 } 3338 constant_value = Klass::_lh_neutral_value; // put in a known value 3339 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3340 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3341 } 3342 3343 // We just put in an allocate/initialize with a big raw-memory effect. 3344 // Hook selected additional alias categories on the initialization. 3345 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3346 MergeMemNode* init_in_merge, 3347 Node* init_out_raw) { 3348 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3349 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3350 3351 Node* prevmem = kit.memory(alias_idx); 3352 init_in_merge->set_memory_at(alias_idx, prevmem); 3353 kit.set_memory(init_out_raw, alias_idx); 3354 } 3355 3356 //---------------------------set_output_for_allocation------------------------- 3357 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3358 const TypeOopPtr* oop_type, 3359 bool deoptimize_on_exception) { 3360 int rawidx = Compile::AliasIdxRaw; 3361 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3362 add_safepoint_edges(alloc); 3363 Node* allocx = _gvn.transform(alloc); 3364 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3365 // create memory projection for i_o 3366 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3367 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3368 3369 // create a memory projection as for the normal control path 3370 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3371 set_memory(malloc, rawidx); 3372 3373 // a normal slow-call doesn't change i_o, but an allocation does 3374 // we create a separate i_o projection for the normal control path 3375 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3376 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3377 3378 // put in an initialization barrier 3379 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3380 rawoop)->as_Initialize(); 3381 assert(alloc->initialization() == init, "2-way macro link must work"); 3382 assert(init ->allocation() == alloc, "2-way macro link must work"); 3383 { 3384 // Extract memory strands which may participate in the new object's 3385 // initialization, and source them from the new InitializeNode. 3386 // This will allow us to observe initializations when they occur, 3387 // and link them properly (as a group) to the InitializeNode. 3388 assert(init->in(InitializeNode::Memory) == malloc, ""); 3389 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3390 init->set_req(InitializeNode::Memory, minit_in); 3391 record_for_igvn(minit_in); // fold it up later, if possible 3392 Node* minit_out = memory(rawidx); 3393 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3394 if (oop_type->isa_aryptr()) { 3395 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3396 int elemidx = C->get_alias_index(telemref); 3397 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3398 } else if (oop_type->isa_instptr()) { 3399 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3400 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3401 ciField* field = ik->nonstatic_field_at(i); 3402 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3403 continue; // do not bother to track really large numbers of fields 3404 // Find (or create) the alias category for this field: 3405 int fieldidx = C->alias_type(field)->index(); 3406 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3407 } 3408 } 3409 } 3410 3411 // Cast raw oop to the real thing... 3412 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3413 javaoop = _gvn.transform(javaoop); 3414 C->set_recent_alloc(control(), javaoop); 3415 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3416 3417 #ifdef ASSERT 3418 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3419 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3420 "Ideal_allocation works"); 3421 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3422 "Ideal_allocation works"); 3423 if (alloc->is_AllocateArray()) { 3424 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3425 "Ideal_allocation works"); 3426 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3427 "Ideal_allocation works"); 3428 } else { 3429 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3430 } 3431 } 3432 #endif //ASSERT 3433 3434 return javaoop; 3435 } 3436 3437 //---------------------------new_instance-------------------------------------- 3438 // This routine takes a klass_node which may be constant (for a static type) 3439 // or may be non-constant (for reflective code). It will work equally well 3440 // for either, and the graph will fold nicely if the optimizer later reduces 3441 // the type to a constant. 3442 // The optional arguments are for specialized use by intrinsics: 3443 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3444 // - If 'return_size_val', report the the total object size to the caller. 3445 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3446 Node* GraphKit::new_instance(Node* klass_node, 3447 Node* extra_slow_test, 3448 Node* *return_size_val, 3449 bool deoptimize_on_exception) { 3450 // Compute size in doublewords 3451 // The size is always an integral number of doublewords, represented 3452 // as a positive bytewise size stored in the klass's layout_helper. 3453 // The layout_helper also encodes (in a low bit) the need for a slow path. 3454 jint layout_con = Klass::_lh_neutral_value; 3455 Node* layout_val = get_layout_helper(klass_node, layout_con); 3456 int layout_is_con = (layout_val == NULL); 3457 3458 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3459 // Generate the initial go-slow test. It's either ALWAYS (return a 3460 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3461 // case) a computed value derived from the layout_helper. 3462 Node* initial_slow_test = NULL; 3463 if (layout_is_con) { 3464 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3465 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3466 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3467 } else { // reflective case 3468 // This reflective path is used by Unsafe.allocateInstance. 3469 // (It may be stress-tested by specifying StressReflectiveCode.) 3470 // Basically, we want to get into the VM is there's an illegal argument. 3471 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3472 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3473 if (extra_slow_test != intcon(0)) { 3474 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3475 } 3476 // (Macro-expander will further convert this to a Bool, if necessary.) 3477 } 3478 3479 // Find the size in bytes. This is easy; it's the layout_helper. 3480 // The size value must be valid even if the slow path is taken. 3481 Node* size = NULL; 3482 if (layout_is_con) { 3483 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3484 } else { // reflective case 3485 // This reflective path is used by clone and Unsafe.allocateInstance. 3486 size = ConvI2X(layout_val); 3487 3488 // Clear the low bits to extract layout_helper_size_in_bytes: 3489 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3490 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3491 size = _gvn.transform( new AndXNode(size, mask) ); 3492 } 3493 if (return_size_val != NULL) { 3494 (*return_size_val) = size; 3495 } 3496 3497 // This is a precise notnull oop of the klass. 3498 // (Actually, it need not be precise if this is a reflective allocation.) 3499 // It's what we cast the result to. 3500 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3501 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3502 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3503 3504 // Now generate allocation code 3505 3506 // The entire memory state is needed for slow path of the allocation 3507 // since GC and deoptimization can happened. 3508 Node *mem = reset_memory(); 3509 set_all_memory(mem); // Create new memory state 3510 3511 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3512 control(), mem, i_o(), 3513 size, klass_node, 3514 initial_slow_test); 3515 3516 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3517 } 3518 3519 //-------------------------------new_array------------------------------------- 3520 // helper for both newarray and anewarray 3521 // The 'length' parameter is (obviously) the length of the array. 3522 // See comments on new_instance for the meaning of the other arguments. 3523 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3524 Node* length, // number of array elements 3525 int nargs, // number of arguments to push back for uncommon trap 3526 Node* *return_size_val, 3527 bool deoptimize_on_exception) { 3528 jint layout_con = Klass::_lh_neutral_value; 3529 Node* layout_val = get_layout_helper(klass_node, layout_con); 3530 int layout_is_con = (layout_val == NULL); 3531 3532 if (!layout_is_con && !StressReflectiveCode && 3533 !too_many_traps(Deoptimization::Reason_class_check)) { 3534 // This is a reflective array creation site. 3535 // Optimistically assume that it is a subtype of Object[], 3536 // so that we can fold up all the address arithmetic. 3537 layout_con = Klass::array_layout_helper(T_OBJECT); 3538 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3539 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3540 { BuildCutout unless(this, bol_lh, PROB_MAX); 3541 inc_sp(nargs); 3542 uncommon_trap(Deoptimization::Reason_class_check, 3543 Deoptimization::Action_maybe_recompile); 3544 } 3545 layout_val = NULL; 3546 layout_is_con = true; 3547 } 3548 3549 // Generate the initial go-slow test. Make sure we do not overflow 3550 // if length is huge (near 2Gig) or negative! We do not need 3551 // exact double-words here, just a close approximation of needed 3552 // double-words. We can't add any offset or rounding bits, lest we 3553 // take a size -1 of bytes and make it positive. Use an unsigned 3554 // compare, so negative sizes look hugely positive. 3555 int fast_size_limit = FastAllocateSizeLimit; 3556 if (layout_is_con) { 3557 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3558 // Increase the size limit if we have exact knowledge of array type. 3559 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3560 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3561 } 3562 3563 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3564 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3565 3566 // --- Size Computation --- 3567 // array_size = round_to_heap(array_header + (length << elem_shift)); 3568 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3569 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3570 // The rounding mask is strength-reduced, if possible. 3571 int round_mask = MinObjAlignmentInBytes - 1; 3572 Node* header_size = NULL; 3573 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3574 // (T_BYTE has the weakest alignment and size restrictions...) 3575 if (layout_is_con) { 3576 int hsize = Klass::layout_helper_header_size(layout_con); 3577 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3578 BasicType etype = Klass::layout_helper_element_type(layout_con); 3579 if ((round_mask & ~right_n_bits(eshift)) == 0) 3580 round_mask = 0; // strength-reduce it if it goes away completely 3581 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3582 assert(header_size_min <= hsize, "generic minimum is smallest"); 3583 header_size_min = hsize; 3584 header_size = intcon(hsize + round_mask); 3585 } else { 3586 Node* hss = intcon(Klass::_lh_header_size_shift); 3587 Node* hsm = intcon(Klass::_lh_header_size_mask); 3588 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3589 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3590 Node* mask = intcon(round_mask); 3591 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3592 } 3593 3594 Node* elem_shift = NULL; 3595 if (layout_is_con) { 3596 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3597 if (eshift != 0) 3598 elem_shift = intcon(eshift); 3599 } else { 3600 // There is no need to mask or shift this value. 3601 // The semantics of LShiftINode include an implicit mask to 0x1F. 3602 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3603 elem_shift = layout_val; 3604 } 3605 3606 // Transition to native address size for all offset calculations: 3607 Node* lengthx = ConvI2X(length); 3608 Node* headerx = ConvI2X(header_size); 3609 #ifdef _LP64 3610 { const TypeInt* tilen = _gvn.find_int_type(length); 3611 if (tilen != NULL && tilen->_lo < 0) { 3612 // Add a manual constraint to a positive range. Cf. array_element_address. 3613 jint size_max = fast_size_limit; 3614 if (size_max > tilen->_hi) size_max = tilen->_hi; 3615 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3616 3617 // Only do a narrow I2L conversion if the range check passed. 3618 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3619 _gvn.transform(iff); 3620 RegionNode* region = new RegionNode(3); 3621 _gvn.set_type(region, Type::CONTROL); 3622 lengthx = new PhiNode(region, TypeLong::LONG); 3623 _gvn.set_type(lengthx, TypeLong::LONG); 3624 3625 // Range check passed. Use ConvI2L node with narrow type. 3626 Node* passed = IfFalse(iff); 3627 region->init_req(1, passed); 3628 // Make I2L conversion control dependent to prevent it from 3629 // floating above the range check during loop optimizations. 3630 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3631 3632 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3633 region->init_req(2, IfTrue(iff)); 3634 lengthx->init_req(2, ConvI2X(length)); 3635 3636 set_control(region); 3637 record_for_igvn(region); 3638 record_for_igvn(lengthx); 3639 } 3640 } 3641 #endif 3642 3643 // Combine header size (plus rounding) and body size. Then round down. 3644 // This computation cannot overflow, because it is used only in two 3645 // places, one where the length is sharply limited, and the other 3646 // after a successful allocation. 3647 Node* abody = lengthx; 3648 if (elem_shift != NULL) 3649 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3650 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3651 if (round_mask != 0) { 3652 Node* mask = MakeConX(~round_mask); 3653 size = _gvn.transform( new AndXNode(size, mask) ); 3654 } 3655 // else if round_mask == 0, the size computation is self-rounding 3656 3657 if (return_size_val != NULL) { 3658 // This is the size 3659 (*return_size_val) = size; 3660 } 3661 3662 // Now generate allocation code 3663 3664 // The entire memory state is needed for slow path of the allocation 3665 // since GC and deoptimization can happened. 3666 Node *mem = reset_memory(); 3667 set_all_memory(mem); // Create new memory state 3668 3669 if (initial_slow_test->is_Bool()) { 3670 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3671 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3672 } 3673 3674 // Create the AllocateArrayNode and its result projections 3675 AllocateArrayNode* alloc 3676 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3677 control(), mem, i_o(), 3678 size, klass_node, 3679 initial_slow_test, 3680 length); 3681 3682 // Cast to correct type. Note that the klass_node may be constant or not, 3683 // and in the latter case the actual array type will be inexact also. 3684 // (This happens via a non-constant argument to inline_native_newArray.) 3685 // In any case, the value of klass_node provides the desired array type. 3686 const TypeInt* length_type = _gvn.find_int_type(length); 3687 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3688 if (ary_type->isa_aryptr() && length_type != NULL) { 3689 // Try to get a better type than POS for the size 3690 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3691 } 3692 3693 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3694 3695 // Cast length on remaining path to be as narrow as possible 3696 if (map()->find_edge(length) >= 0) { 3697 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3698 if (ccast != length) { 3699 _gvn.set_type_bottom(ccast); 3700 record_for_igvn(ccast); 3701 replace_in_map(length, ccast); 3702 } 3703 } 3704 3705 return javaoop; 3706 } 3707 3708 // The following "Ideal_foo" functions are placed here because they recognize 3709 // the graph shapes created by the functions immediately above. 3710 3711 //---------------------------Ideal_allocation---------------------------------- 3712 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3713 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3714 if (ptr == NULL) { // reduce dumb test in callers 3715 return NULL; 3716 } 3717 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3718 ptr = ptr->in(1); 3719 if (ptr == NULL) return NULL; 3720 } 3721 // Return NULL for allocations with several casts: 3722 // j.l.reflect.Array.newInstance(jobject, jint) 3723 // Object.clone() 3724 // to keep more precise type from last cast. 3725 if (ptr->is_Proj()) { 3726 Node* allo = ptr->in(0); 3727 if (allo != NULL && allo->is_Allocate()) { 3728 return allo->as_Allocate(); 3729 } 3730 } 3731 // Report failure to match. 3732 return NULL; 3733 } 3734 3735 // Fancy version which also strips off an offset (and reports it to caller). 3736 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3737 intptr_t& offset) { 3738 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3739 if (base == NULL) return NULL; 3740 return Ideal_allocation(base, phase); 3741 } 3742 3743 // Trace Initialize <- Proj[Parm] <- Allocate 3744 AllocateNode* InitializeNode::allocation() { 3745 Node* rawoop = in(InitializeNode::RawAddress); 3746 if (rawoop->is_Proj()) { 3747 Node* alloc = rawoop->in(0); 3748 if (alloc->is_Allocate()) { 3749 return alloc->as_Allocate(); 3750 } 3751 } 3752 return NULL; 3753 } 3754 3755 // Trace Allocate -> Proj[Parm] -> Initialize 3756 InitializeNode* AllocateNode::initialization() { 3757 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3758 if (rawoop == NULL) return NULL; 3759 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3760 Node* init = rawoop->fast_out(i); 3761 if (init->is_Initialize()) { 3762 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3763 return init->as_Initialize(); 3764 } 3765 } 3766 return NULL; 3767 } 3768 3769 //----------------------------- loop predicates --------------------------- 3770 3771 //------------------------------add_predicate_impl---------------------------- 3772 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3773 // Too many traps seen? 3774 if (too_many_traps(reason)) { 3775 #ifdef ASSERT 3776 if (TraceLoopPredicate) { 3777 int tc = C->trap_count(reason); 3778 tty->print("too many traps=%s tcount=%d in ", 3779 Deoptimization::trap_reason_name(reason), tc); 3780 method()->print(); // which method has too many predicate traps 3781 tty->cr(); 3782 } 3783 #endif 3784 // We cannot afford to take more traps here, 3785 // do not generate predicate. 3786 return; 3787 } 3788 3789 Node *cont = _gvn.intcon(1); 3790 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3791 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3792 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3793 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3794 C->add_predicate_opaq(opq); 3795 { 3796 PreserveJVMState pjvms(this); 3797 set_control(iffalse); 3798 inc_sp(nargs); 3799 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3800 } 3801 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3802 set_control(iftrue); 3803 } 3804 3805 //------------------------------add_predicate--------------------------------- 3806 void GraphKit::add_predicate(int nargs) { 3807 if (UseLoopPredicate) { 3808 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3809 } 3810 // loop's limit check predicate should be near the loop. 3811 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3812 } 3813 3814 //----------------------------- store barriers ---------------------------- 3815 #define __ ideal. 3816 3817 void GraphKit::sync_kit(IdealKit& ideal) { 3818 set_all_memory(__ merged_memory()); 3819 set_i_o(__ i_o()); 3820 set_control(__ ctrl()); 3821 } 3822 3823 void GraphKit::final_sync(IdealKit& ideal) { 3824 // Final sync IdealKit and graphKit. 3825 sync_kit(ideal); 3826 } 3827 3828 Node* GraphKit::byte_map_base_node() { 3829 // Get base of card map 3830 CardTableModRefBS* ct = 3831 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3832 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3833 if (ct->byte_map_base != NULL) { 3834 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3835 } else { 3836 return null(); 3837 } 3838 } 3839 3840 // vanilla/CMS post barrier 3841 // Insert a write-barrier store. This is to let generational GC work; we have 3842 // to flag all oop-stores before the next GC point. 3843 void GraphKit::write_barrier_post(Node* oop_store, 3844 Node* obj, 3845 Node* adr, 3846 uint adr_idx, 3847 Node* val, 3848 bool use_precise) { 3849 // No store check needed if we're storing a NULL or an old object 3850 // (latter case is probably a string constant). The concurrent 3851 // mark sweep garbage collector, however, needs to have all nonNull 3852 // oop updates flagged via card-marks. 3853 if (val != NULL && val->is_Con()) { 3854 // must be either an oop or NULL 3855 const Type* t = val->bottom_type(); 3856 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3857 // stores of null never (?) need barriers 3858 return; 3859 } 3860 3861 if (use_ReduceInitialCardMarks() 3862 && obj == just_allocated_object(control())) { 3863 // We can skip marks on a freshly-allocated object in Eden. 3864 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3865 // That routine informs GC to take appropriate compensating steps, 3866 // upon a slow-path allocation, so as to make this card-mark 3867 // elision safe. 3868 return; 3869 } 3870 3871 if (!use_precise) { 3872 // All card marks for a (non-array) instance are in one place: 3873 adr = obj; 3874 } 3875 // (Else it's an array (or unknown), and we want more precise card marks.) 3876 assert(adr != NULL, ""); 3877 3878 IdealKit ideal(this, true); 3879 3880 // Convert the pointer to an int prior to doing math on it 3881 Node* cast = __ CastPX(__ ctrl(), adr); 3882 3883 // Divide by card size 3884 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3885 "Only one we handle so far."); 3886 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3887 3888 // Combine card table base and card offset 3889 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3890 3891 // Get the alias_index for raw card-mark memory 3892 int adr_type = Compile::AliasIdxRaw; 3893 Node* zero = __ ConI(0); // Dirty card value 3894 BasicType bt = T_BYTE; 3895 3896 if (UseConcMarkSweepGC && UseCondCardMark) { 3897 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 3898 __ sync_kit(this); 3899 } 3900 3901 if (UseCondCardMark) { 3902 // The classic GC reference write barrier is typically implemented 3903 // as a store into the global card mark table. Unfortunately 3904 // unconditional stores can result in false sharing and excessive 3905 // coherence traffic as well as false transactional aborts. 3906 // UseCondCardMark enables MP "polite" conditional card mark 3907 // stores. In theory we could relax the load from ctrl() to 3908 // no_ctrl, but that doesn't buy much latitude. 3909 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3910 __ if_then(card_val, BoolTest::ne, zero); 3911 } 3912 3913 // Smash zero into card 3914 if( !UseConcMarkSweepGC ) { 3915 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3916 } else { 3917 // Specialized path for CM store barrier 3918 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3919 } 3920 3921 if (UseCondCardMark) { 3922 __ end_if(); 3923 } 3924 3925 // Final sync IdealKit and GraphKit. 3926 final_sync(ideal); 3927 } 3928 /* 3929 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3930 * required by SATB to make sure all objects live at the start of the 3931 * marking are kept alive, all reference updates need to any previous 3932 * reference stored before writing. 3933 * 3934 * If the previous value is NULL there is no need to save the old value. 3935 * References that are NULL are filtered during runtime by the barrier 3936 * code to avoid unnecessary queuing. 3937 * 3938 * However in the case of newly allocated objects it might be possible to 3939 * prove that the reference about to be overwritten is NULL during compile 3940 * time and avoid adding the barrier code completely. 3941 * 3942 * The compiler needs to determine that the object in which a field is about 3943 * to be written is newly allocated, and that no prior store to the same field 3944 * has happened since the allocation. 3945 * 3946 * Returns true if the pre-barrier can be removed 3947 */ 3948 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3949 BasicType bt, uint adr_idx) { 3950 intptr_t offset = 0; 3951 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3952 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3953 3954 if (offset == Type::OffsetBot) { 3955 return false; // cannot unalias unless there are precise offsets 3956 } 3957 3958 if (alloc == NULL) { 3959 return false; // No allocation found 3960 } 3961 3962 intptr_t size_in_bytes = type2aelembytes(bt); 3963 3964 Node* mem = memory(adr_idx); // start searching here... 3965 3966 for (int cnt = 0; cnt < 50; cnt++) { 3967 3968 if (mem->is_Store()) { 3969 3970 Node* st_adr = mem->in(MemNode::Address); 3971 intptr_t st_offset = 0; 3972 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3973 3974 if (st_base == NULL) { 3975 break; // inscrutable pointer 3976 } 3977 3978 // Break we have found a store with same base and offset as ours so break 3979 if (st_base == base && st_offset == offset) { 3980 break; 3981 } 3982 3983 if (st_offset != offset && st_offset != Type::OffsetBot) { 3984 const int MAX_STORE = BytesPerLong; 3985 if (st_offset >= offset + size_in_bytes || 3986 st_offset <= offset - MAX_STORE || 3987 st_offset <= offset - mem->as_Store()->memory_size()) { 3988 // Success: The offsets are provably independent. 3989 // (You may ask, why not just test st_offset != offset and be done? 3990 // The answer is that stores of different sizes can co-exist 3991 // in the same sequence of RawMem effects. We sometimes initialize 3992 // a whole 'tile' of array elements with a single jint or jlong.) 3993 mem = mem->in(MemNode::Memory); 3994 continue; // advance through independent store memory 3995 } 3996 } 3997 3998 if (st_base != base 3999 && MemNode::detect_ptr_independence(base, alloc, st_base, 4000 AllocateNode::Ideal_allocation(st_base, phase), 4001 phase)) { 4002 // Success: The bases are provably independent. 4003 mem = mem->in(MemNode::Memory); 4004 continue; // advance through independent store memory 4005 } 4006 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4007 4008 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4009 AllocateNode* st_alloc = st_init->allocation(); 4010 4011 // Make sure that we are looking at the same allocation site. 4012 // The alloc variable is guaranteed to not be null here from earlier check. 4013 if (alloc == st_alloc) { 4014 // Check that the initialization is storing NULL so that no previous store 4015 // has been moved up and directly write a reference 4016 Node* captured_store = st_init->find_captured_store(offset, 4017 type2aelembytes(T_OBJECT), 4018 phase); 4019 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4020 return true; 4021 } 4022 } 4023 } 4024 4025 // Unless there is an explicit 'continue', we must bail out here, 4026 // because 'mem' is an inscrutable memory state (e.g., a call). 4027 break; 4028 } 4029 4030 return false; 4031 } 4032 4033 // G1 pre/post barriers 4034 void GraphKit::g1_write_barrier_pre(bool do_load, 4035 Node* obj, 4036 Node* adr, 4037 uint alias_idx, 4038 Node* val, 4039 const TypeOopPtr* val_type, 4040 Node* pre_val, 4041 BasicType bt) { 4042 4043 // Some sanity checks 4044 // Note: val is unused in this routine. 4045 4046 if (do_load) { 4047 // We need to generate the load of the previous value 4048 assert(obj != NULL, "must have a base"); 4049 assert(adr != NULL, "where are loading from?"); 4050 assert(pre_val == NULL, "loaded already?"); 4051 assert(val_type != NULL, "need a type"); 4052 4053 if (use_ReduceInitialCardMarks() 4054 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4055 return; 4056 } 4057 4058 } else { 4059 // In this case both val_type and alias_idx are unused. 4060 assert(pre_val != NULL, "must be loaded already"); 4061 // Nothing to be done if pre_val is null. 4062 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4063 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4064 } 4065 assert(bt == T_OBJECT, "or we shouldn't be here"); 4066 4067 IdealKit ideal(this, true); 4068 4069 Node* tls = __ thread(); // ThreadLocalStorage 4070 4071 Node* no_ctrl = NULL; 4072 Node* no_base = __ top(); 4073 Node* zero = __ ConI(0); 4074 Node* zeroX = __ ConX(0); 4075 4076 float likely = PROB_LIKELY(0.999); 4077 float unlikely = PROB_UNLIKELY(0.999); 4078 4079 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4080 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4081 4082 // Offsets into the thread 4083 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4084 SATBMarkQueue::byte_offset_of_active()); 4085 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4086 SATBMarkQueue::byte_offset_of_index()); 4087 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4088 SATBMarkQueue::byte_offset_of_buf()); 4089 4090 // Now the actual pointers into the thread 4091 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4092 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4093 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4094 4095 // Now some of the values 4096 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4097 4098 // if (!marking) 4099 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4100 BasicType index_bt = TypeX_X->basic_type(); 4101 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4102 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4103 4104 if (do_load) { 4105 // load original value 4106 // alias_idx correct?? 4107 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4108 } 4109 4110 // if (pre_val != NULL) 4111 __ if_then(pre_val, BoolTest::ne, null()); { 4112 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4113 4114 // is the queue for this thread full? 4115 __ if_then(index, BoolTest::ne, zeroX, likely); { 4116 4117 // decrement the index 4118 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4119 4120 // Now get the buffer location we will log the previous value into and store it 4121 Node *log_addr = __ AddP(no_base, buffer, next_index); 4122 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4123 // update the index 4124 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4125 4126 } __ else_(); { 4127 4128 // logging buffer is full, call the runtime 4129 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4130 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4131 } __ end_if(); // (!index) 4132 } __ end_if(); // (pre_val != NULL) 4133 } __ end_if(); // (!marking) 4134 4135 // Final sync IdealKit and GraphKit. 4136 final_sync(ideal); 4137 } 4138 4139 /* 4140 * G1 similar to any GC with a Young Generation requires a way to keep track of 4141 * references from Old Generation to Young Generation to make sure all live 4142 * objects are found. G1 also requires to keep track of object references 4143 * between different regions to enable evacuation of old regions, which is done 4144 * as part of mixed collections. References are tracked in remembered sets and 4145 * is continuously updated as reference are written to with the help of the 4146 * post-barrier. 4147 * 4148 * To reduce the number of updates to the remembered set the post-barrier 4149 * filters updates to fields in objects located in the Young Generation, 4150 * the same region as the reference, when the NULL is being written or 4151 * if the card is already marked as dirty by an earlier write. 4152 * 4153 * Under certain circumstances it is possible to avoid generating the 4154 * post-barrier completely if it is possible during compile time to prove 4155 * the object is newly allocated and that no safepoint exists between the 4156 * allocation and the store. 4157 * 4158 * In the case of slow allocation the allocation code must handle the barrier 4159 * as part of the allocation in the case the allocated object is not located 4160 * in the nursery, this would happen for humongous objects. This is similar to 4161 * how CMS is required to handle this case, see the comments for the method 4162 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4163 * A deferred card mark is required for these objects and handled in the above 4164 * mentioned methods. 4165 * 4166 * Returns true if the post barrier can be removed 4167 */ 4168 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4169 Node* adr) { 4170 intptr_t offset = 0; 4171 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4172 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4173 4174 if (offset == Type::OffsetBot) { 4175 return false; // cannot unalias unless there are precise offsets 4176 } 4177 4178 if (alloc == NULL) { 4179 return false; // No allocation found 4180 } 4181 4182 // Start search from Store node 4183 Node* mem = store->in(MemNode::Control); 4184 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4185 4186 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4187 AllocateNode* st_alloc = st_init->allocation(); 4188 4189 // Make sure we are looking at the same allocation 4190 if (alloc == st_alloc) { 4191 return true; 4192 } 4193 } 4194 4195 return false; 4196 } 4197 4198 // 4199 // Update the card table and add card address to the queue 4200 // 4201 void GraphKit::g1_mark_card(IdealKit& ideal, 4202 Node* card_adr, 4203 Node* oop_store, 4204 uint oop_alias_idx, 4205 Node* index, 4206 Node* index_adr, 4207 Node* buffer, 4208 const TypeFunc* tf) { 4209 4210 Node* zero = __ ConI(0); 4211 Node* zeroX = __ ConX(0); 4212 Node* no_base = __ top(); 4213 BasicType card_bt = T_BYTE; 4214 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4215 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4216 4217 // Now do the queue work 4218 __ if_then(index, BoolTest::ne, zeroX); { 4219 4220 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4221 Node* log_addr = __ AddP(no_base, buffer, next_index); 4222 4223 // Order, see storeCM. 4224 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4225 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4226 4227 } __ else_(); { 4228 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4229 } __ end_if(); 4230 4231 } 4232 4233 void GraphKit::g1_write_barrier_post(Node* oop_store, 4234 Node* obj, 4235 Node* adr, 4236 uint alias_idx, 4237 Node* val, 4238 BasicType bt, 4239 bool use_precise) { 4240 // If we are writing a NULL then we need no post barrier 4241 4242 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4243 // Must be NULL 4244 const Type* t = val->bottom_type(); 4245 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4246 // No post barrier if writing NULLx 4247 return; 4248 } 4249 4250 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4251 // We can skip marks on a freshly-allocated object in Eden. 4252 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4253 // That routine informs GC to take appropriate compensating steps, 4254 // upon a slow-path allocation, so as to make this card-mark 4255 // elision safe. 4256 return; 4257 } 4258 4259 if (use_ReduceInitialCardMarks() 4260 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4261 return; 4262 } 4263 4264 if (!use_precise) { 4265 // All card marks for a (non-array) instance are in one place: 4266 adr = obj; 4267 } 4268 // (Else it's an array (or unknown), and we want more precise card marks.) 4269 assert(adr != NULL, ""); 4270 4271 IdealKit ideal(this, true); 4272 4273 Node* tls = __ thread(); // ThreadLocalStorage 4274 4275 Node* no_base = __ top(); 4276 float likely = PROB_LIKELY(0.999); 4277 float unlikely = PROB_UNLIKELY(0.999); 4278 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4279 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4280 Node* zeroX = __ ConX(0); 4281 4282 // Get the alias_index for raw card-mark memory 4283 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4284 4285 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4286 4287 // Offsets into the thread 4288 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4289 DirtyCardQueue::byte_offset_of_index()); 4290 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4291 DirtyCardQueue::byte_offset_of_buf()); 4292 4293 // Pointers into the thread 4294 4295 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4296 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4297 4298 // Now some values 4299 // Use ctrl to avoid hoisting these values past a safepoint, which could 4300 // potentially reset these fields in the JavaThread. 4301 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4302 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4303 4304 // Convert the store obj pointer to an int prior to doing math on it 4305 // Must use ctrl to prevent "integerized oop" existing across safepoint 4306 Node* cast = __ CastPX(__ ctrl(), adr); 4307 4308 // Divide pointer by card size 4309 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4310 4311 // Combine card table base and card offset 4312 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4313 4314 // If we know the value being stored does it cross regions? 4315 4316 if (val != NULL) { 4317 // Does the store cause us to cross regions? 4318 4319 // Should be able to do an unsigned compare of region_size instead of 4320 // and extra shift. Do we have an unsigned compare?? 4321 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4322 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4323 4324 // if (xor_res == 0) same region so skip 4325 __ if_then(xor_res, BoolTest::ne, zeroX); { 4326 4327 // No barrier if we are storing a NULL 4328 __ if_then(val, BoolTest::ne, null(), unlikely); { 4329 4330 // Ok must mark the card if not already dirty 4331 4332 // load the original value of the card 4333 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4334 4335 __ if_then(card_val, BoolTest::ne, young_card); { 4336 sync_kit(ideal); 4337 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4338 insert_mem_bar(Op_MemBarVolatile, oop_store); 4339 __ sync_kit(this); 4340 4341 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4342 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4343 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4344 } __ end_if(); 4345 } __ end_if(); 4346 } __ end_if(); 4347 } __ end_if(); 4348 } else { 4349 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks. 4350 // We don't need a barrier here if the destination is a newly allocated object 4351 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden 4352 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()). 4353 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking"); 4354 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4355 __ if_then(card_val, BoolTest::ne, young_card); { 4356 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4357 } __ end_if(); 4358 } 4359 4360 // Final sync IdealKit and GraphKit. 4361 final_sync(ideal); 4362 } 4363 #undef __ 4364 4365 4366 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4367 Node* len = load_array_length(load_String_value(ctrl, str)); 4368 Node* coder = load_String_coder(ctrl, str); 4369 // Divide length by 2 if coder is UTF16 4370 return _gvn.transform(new RShiftINode(len, coder)); 4371 } 4372 4373 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4374 int value_offset = java_lang_String::value_offset_in_bytes(); 4375 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4376 false, NULL, 0); 4377 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4378 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4379 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4380 ciTypeArrayKlass::make(T_BYTE), true, 0); 4381 int value_field_idx = C->get_alias_index(value_field_type); 4382 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4383 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4384 // String.value field is known to be @Stable. 4385 if (UseImplicitStableValues) { 4386 load = cast_array_to_stable(load, value_type); 4387 } 4388 return load; 4389 } 4390 4391 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4392 if (!CompactStrings) { 4393 return intcon(java_lang_String::CODER_UTF16); 4394 } 4395 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4396 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4397 false, NULL, 0); 4398 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4399 int coder_field_idx = C->get_alias_index(coder_field_type); 4400 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4401 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4402 } 4403 4404 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4405 int value_offset = java_lang_String::value_offset_in_bytes(); 4406 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4407 false, NULL, 0); 4408 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4409 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4410 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered); 4411 } 4412 4413 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 4414 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4415 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4416 false, NULL, 0); 4417 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4418 int coder_field_idx = C->get_alias_index(coder_field_type); 4419 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 4420 value, T_BYTE, coder_field_idx, MemNode::unordered); 4421 } 4422 4423 // Capture src and dst memory state with a MergeMemNode 4424 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 4425 if (src_type == dst_type) { 4426 // Types are equal, we don't need a MergeMemNode 4427 return memory(src_type); 4428 } 4429 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 4430 record_for_igvn(merge); // fold it up later, if possible 4431 int src_idx = C->get_alias_index(src_type); 4432 int dst_idx = C->get_alias_index(dst_type); 4433 merge->set_memory_at(src_idx, memory(src_idx)); 4434 merge->set_memory_at(dst_idx, memory(dst_idx)); 4435 return merge; 4436 } 4437 4438 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 4439 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 4440 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 4441 // If input and output memory types differ, capture both states to preserve 4442 // the dependency between preceding and subsequent loads/stores. 4443 // For example, the following program: 4444 // StoreB 4445 // compress_string 4446 // LoadB 4447 // has this memory graph (use->def): 4448 // LoadB -> compress_string -> CharMem 4449 // ... -> StoreB -> ByteMem 4450 // The intrinsic hides the dependency between LoadB and StoreB, causing 4451 // the load to read from memory not containing the result of the StoreB. 4452 // The correct memory graph should look like this: 4453 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 4454 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 4455 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 4456 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 4457 set_memory(res_mem, TypeAryPtr::BYTES); 4458 return str; 4459 } 4460 4461 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 4462 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 4463 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 4464 // Capture src and dst memory (see comment in 'compress_string'). 4465 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 4466 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 4467 set_memory(_gvn.transform(str), dst_type); 4468 } 4469 4470 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 4471 /** 4472 * int i_char = start; 4473 * for (int i_byte = 0; i_byte < count; i_byte++) { 4474 * dst[i_char++] = (char)(src[i_byte] & 0xff); 4475 * } 4476 */ 4477 add_predicate(); 4478 RegionNode* head = new RegionNode(3); 4479 head->init_req(1, control()); 4480 gvn().set_type(head, Type::CONTROL); 4481 record_for_igvn(head); 4482 4483 Node* i_byte = new PhiNode(head, TypeInt::INT); 4484 i_byte->init_req(1, intcon(0)); 4485 gvn().set_type(i_byte, TypeInt::INT); 4486 record_for_igvn(i_byte); 4487 4488 Node* i_char = new PhiNode(head, TypeInt::INT); 4489 i_char->init_req(1, start); 4490 gvn().set_type(i_char, TypeInt::INT); 4491 record_for_igvn(i_char); 4492 4493 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 4494 gvn().set_type(mem, Type::MEMORY); 4495 record_for_igvn(mem); 4496 set_control(head); 4497 set_memory(mem, TypeAryPtr::BYTES); 4498 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 4499 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 4500 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 4501 false, false, true /* mismatched */); 4502 4503 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 4504 head->init_req(2, IfTrue(iff)); 4505 mem->init_req(2, st); 4506 i_byte->init_req(2, AddI(i_byte, intcon(1))); 4507 i_char->init_req(2, AddI(i_char, intcon(2))); 4508 4509 set_control(IfFalse(iff)); 4510 set_memory(st, TypeAryPtr::BYTES); 4511 } 4512 4513 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 4514 if (!field->is_constant()) { 4515 return NULL; // Field not marked as constant. 4516 } 4517 ciInstance* holder = NULL; 4518 if (!field->is_static()) { 4519 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 4520 if (const_oop != NULL && const_oop->is_instance()) { 4521 holder = const_oop->as_instance(); 4522 } 4523 } 4524 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 4525 /*is_unsigned_load=*/false); 4526 if (con_type != NULL) { 4527 return makecon(con_type); 4528 } 4529 return NULL; 4530 } 4531 4532 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4533 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4534 // assumption of CCP analysis. 4535 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4536 }