1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "ci/ciValueKlass.hpp" 28 #include "gc/g1/g1SATBCardTableModRefBS.hpp" 29 #include "gc/g1/heapRegion.hpp" 30 #include "gc/shared/barrierSet.hpp" 31 #include "gc/shared/cardTableModRefBS.hpp" 32 #include "gc/shared/collectedHeap.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 "opto/valuetypenode.hpp" 46 #include "runtime/deoptimization.hpp" 47 #include "runtime/sharedRuntime.hpp" 48 49 //----------------------------GraphKit----------------------------------------- 50 // Main utility constructor. 51 GraphKit::GraphKit(JVMState* jvms) 52 : Phase(Phase::Parser), 53 _env(C->env()), 54 _gvn(*C->initial_gvn()) 55 { 56 _exceptions = jvms->map()->next_exception(); 57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 58 set_jvms(jvms); 59 } 60 61 // Private constructor for parser. 62 GraphKit::GraphKit() 63 : Phase(Phase::Parser), 64 _env(C->env()), 65 _gvn(*C->initial_gvn()) 66 { 67 _exceptions = NULL; 68 set_map(NULL); 69 debug_only(_sp = -99); 70 debug_only(set_bci(-99)); 71 } 72 73 74 75 //---------------------------clean_stack--------------------------------------- 76 // Clear away rubbish from the stack area of the JVM state. 77 // This destroys any arguments that may be waiting on the stack. 78 void GraphKit::clean_stack(int from_sp) { 79 SafePointNode* map = this->map(); 80 JVMState* jvms = this->jvms(); 81 int stk_size = jvms->stk_size(); 82 int stkoff = jvms->stkoff(); 83 Node* top = this->top(); 84 for (int i = from_sp; i < stk_size; i++) { 85 if (map->in(stkoff + i) != top) { 86 map->set_req(stkoff + i, top); 87 } 88 } 89 } 90 91 92 //--------------------------------sync_jvms----------------------------------- 93 // Make sure our current jvms agrees with our parse state. 94 JVMState* GraphKit::sync_jvms() const { 95 JVMState* jvms = this->jvms(); 96 jvms->set_bci(bci()); // Record the new bci in the JVMState 97 jvms->set_sp(sp()); // Record the new sp in the JVMState 98 assert(jvms_in_sync(), "jvms is now in sync"); 99 return jvms; 100 } 101 102 //--------------------------------sync_jvms_for_reexecute--------------------- 103 // Make sure our current jvms agrees with our parse state. This version 104 // uses the reexecute_sp for reexecuting bytecodes. 105 JVMState* GraphKit::sync_jvms_for_reexecute() { 106 JVMState* jvms = this->jvms(); 107 jvms->set_bci(bci()); // Record the new bci in the JVMState 108 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 109 return jvms; 110 } 111 112 #ifdef ASSERT 113 bool GraphKit::jvms_in_sync() const { 114 Parse* parse = is_Parse(); 115 if (parse == NULL) { 116 if (bci() != jvms()->bci()) return false; 117 if (sp() != (int)jvms()->sp()) return false; 118 return true; 119 } 120 if (jvms()->method() != parse->method()) return false; 121 if (jvms()->bci() != parse->bci()) return false; 122 int jvms_sp = jvms()->sp(); 123 if (jvms_sp != parse->sp()) return false; 124 int jvms_depth = jvms()->depth(); 125 if (jvms_depth != parse->depth()) return false; 126 return true; 127 } 128 129 // Local helper checks for special internal merge points 130 // used to accumulate and merge exception states. 131 // They are marked by the region's in(0) edge being the map itself. 132 // Such merge points must never "escape" into the parser at large, 133 // until they have been handed to gvn.transform. 134 static bool is_hidden_merge(Node* reg) { 135 if (reg == NULL) return false; 136 if (reg->is_Phi()) { 137 reg = reg->in(0); 138 if (reg == NULL) return false; 139 } 140 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 141 } 142 143 void GraphKit::verify_map() const { 144 if (map() == NULL) return; // null map is OK 145 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 146 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 147 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 148 } 149 150 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 151 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 152 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 153 } 154 #endif 155 156 //---------------------------stop_and_kill_map--------------------------------- 157 // Set _map to NULL, signalling a stop to further bytecode execution. 158 // First smash the current map's control to a constant, to mark it dead. 159 void GraphKit::stop_and_kill_map() { 160 SafePointNode* dead_map = stop(); 161 if (dead_map != NULL) { 162 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 163 assert(dead_map->is_killed(), "must be so marked"); 164 } 165 } 166 167 168 //--------------------------------stopped-------------------------------------- 169 // Tell if _map is NULL, or control is top. 170 bool GraphKit::stopped() { 171 if (map() == NULL) return true; 172 else if (control() == top()) return true; 173 else return false; 174 } 175 176 177 //-----------------------------has_ex_handler---------------------------------- 178 // Tell if this method or any caller method has exception handlers. 179 bool GraphKit::has_ex_handler() { 180 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 181 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 182 return true; 183 } 184 } 185 return false; 186 } 187 188 //------------------------------save_ex_oop------------------------------------ 189 // Save an exception without blowing stack contents or other JVM state. 190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 191 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 192 ex_map->add_req(ex_oop); 193 debug_only(verify_exception_state(ex_map)); 194 } 195 196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 197 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 198 Node* ex_oop = ex_map->in(ex_map->req()-1); 199 if (clear_it) ex_map->del_req(ex_map->req()-1); 200 return ex_oop; 201 } 202 203 //-----------------------------saved_ex_oop------------------------------------ 204 // Recover a saved exception from its map. 205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 206 return common_saved_ex_oop(ex_map, false); 207 } 208 209 //--------------------------clear_saved_ex_oop--------------------------------- 210 // Erase a previously saved exception from its map. 211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 212 return common_saved_ex_oop(ex_map, true); 213 } 214 215 #ifdef ASSERT 216 //---------------------------has_saved_ex_oop---------------------------------- 217 // Erase a previously saved exception from its map. 218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 219 return ex_map->req() == ex_map->jvms()->endoff()+1; 220 } 221 #endif 222 223 //-------------------------make_exception_state-------------------------------- 224 // Turn the current JVM state into an exception state, appending the ex_oop. 225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 226 sync_jvms(); 227 SafePointNode* ex_map = stop(); // do not manipulate this map any more 228 set_saved_ex_oop(ex_map, ex_oop); 229 return ex_map; 230 } 231 232 233 //--------------------------add_exception_state-------------------------------- 234 // Add an exception to my list of exceptions. 235 void GraphKit::add_exception_state(SafePointNode* ex_map) { 236 if (ex_map == NULL || ex_map->control() == top()) { 237 return; 238 } 239 #ifdef ASSERT 240 verify_exception_state(ex_map); 241 if (has_exceptions()) { 242 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 243 } 244 #endif 245 246 // If there is already an exception of exactly this type, merge with it. 247 // In particular, null-checks and other low-level exceptions common up here. 248 Node* ex_oop = saved_ex_oop(ex_map); 249 const Type* ex_type = _gvn.type(ex_oop); 250 if (ex_oop == top()) { 251 // No action needed. 252 return; 253 } 254 assert(ex_type->isa_instptr(), "exception must be an instance"); 255 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 256 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 257 // We check sp also because call bytecodes can generate exceptions 258 // both before and after arguments are popped! 259 if (ex_type2 == ex_type 260 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 261 combine_exception_states(ex_map, e2); 262 return; 263 } 264 } 265 266 // No pre-existing exception of the same type. Chain it on the list. 267 push_exception_state(ex_map); 268 } 269 270 //-----------------------add_exception_states_from----------------------------- 271 void GraphKit::add_exception_states_from(JVMState* jvms) { 272 SafePointNode* ex_map = jvms->map()->next_exception(); 273 if (ex_map != NULL) { 274 jvms->map()->set_next_exception(NULL); 275 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 276 next_map = ex_map->next_exception(); 277 ex_map->set_next_exception(NULL); 278 add_exception_state(ex_map); 279 } 280 } 281 } 282 283 //-----------------------transfer_exceptions_into_jvms------------------------- 284 JVMState* GraphKit::transfer_exceptions_into_jvms() { 285 if (map() == NULL) { 286 // We need a JVMS to carry the exceptions, but the map has gone away. 287 // Create a scratch JVMS, cloned from any of the exception states... 288 if (has_exceptions()) { 289 _map = _exceptions; 290 _map = clone_map(); 291 _map->set_next_exception(NULL); 292 clear_saved_ex_oop(_map); 293 debug_only(verify_map()); 294 } else { 295 // ...or created from scratch 296 JVMState* jvms = new (C) JVMState(_method, NULL); 297 jvms->set_bci(_bci); 298 jvms->set_sp(_sp); 299 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 300 set_jvms(jvms); 301 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 302 set_all_memory(top()); 303 while (map()->req() < jvms->endoff()) map()->add_req(top()); 304 } 305 // (This is a kludge, in case you didn't notice.) 306 set_control(top()); 307 } 308 JVMState* jvms = sync_jvms(); 309 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 310 jvms->map()->set_next_exception(_exceptions); 311 _exceptions = NULL; // done with this set of exceptions 312 return jvms; 313 } 314 315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 316 assert(is_hidden_merge(dstphi), "must be a special merge node"); 317 assert(is_hidden_merge(srcphi), "must be a special merge node"); 318 uint limit = srcphi->req(); 319 for (uint i = PhiNode::Input; i < limit; i++) { 320 dstphi->add_req(srcphi->in(i)); 321 } 322 } 323 static inline void add_one_req(Node* dstphi, Node* src) { 324 assert(is_hidden_merge(dstphi), "must be a special merge node"); 325 assert(!is_hidden_merge(src), "must not be a special merge node"); 326 dstphi->add_req(src); 327 } 328 329 //-----------------------combine_exception_states------------------------------ 330 // This helper function combines exception states by building phis on a 331 // specially marked state-merging region. These regions and phis are 332 // untransformed, and can build up gradually. The region is marked by 333 // having a control input of its exception map, rather than NULL. Such 334 // regions do not appear except in this function, and in use_exception_state. 335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 336 if (failing()) return; // dying anyway... 337 JVMState* ex_jvms = ex_map->_jvms; 338 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 339 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 340 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 341 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 342 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 343 assert(ex_map->req() == phi_map->req(), "matching maps"); 344 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 345 Node* hidden_merge_mark = root(); 346 Node* region = phi_map->control(); 347 MergeMemNode* phi_mem = phi_map->merged_memory(); 348 MergeMemNode* ex_mem = ex_map->merged_memory(); 349 if (region->in(0) != hidden_merge_mark) { 350 // The control input is not (yet) a specially-marked region in phi_map. 351 // Make it so, and build some phis. 352 region = new RegionNode(2); 353 _gvn.set_type(region, Type::CONTROL); 354 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 355 region->init_req(1, phi_map->control()); 356 phi_map->set_control(region); 357 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 358 record_for_igvn(io_phi); 359 _gvn.set_type(io_phi, Type::ABIO); 360 phi_map->set_i_o(io_phi); 361 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 362 Node* m = mms.memory(); 363 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 364 record_for_igvn(m_phi); 365 _gvn.set_type(m_phi, Type::MEMORY); 366 mms.set_memory(m_phi); 367 } 368 } 369 370 // Either or both of phi_map and ex_map might already be converted into phis. 371 Node* ex_control = ex_map->control(); 372 // if there is special marking on ex_map also, we add multiple edges from src 373 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 374 // how wide was the destination phi_map, originally? 375 uint orig_width = region->req(); 376 377 if (add_multiple) { 378 add_n_reqs(region, ex_control); 379 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 380 } else { 381 // ex_map has no merges, so we just add single edges everywhere 382 add_one_req(region, ex_control); 383 add_one_req(phi_map->i_o(), ex_map->i_o()); 384 } 385 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 386 if (mms.is_empty()) { 387 // get a copy of the base memory, and patch some inputs into it 388 const TypePtr* adr_type = mms.adr_type(C); 389 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 390 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 391 mms.set_memory(phi); 392 // Prepare to append interesting stuff onto the newly sliced phi: 393 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 394 } 395 // Append stuff from ex_map: 396 if (add_multiple) { 397 add_n_reqs(mms.memory(), mms.memory2()); 398 } else { 399 add_one_req(mms.memory(), mms.memory2()); 400 } 401 } 402 uint limit = ex_map->req(); 403 for (uint i = TypeFunc::Parms; i < limit; i++) { 404 // Skip everything in the JVMS after tos. (The ex_oop follows.) 405 if (i == tos) i = ex_jvms->monoff(); 406 Node* src = ex_map->in(i); 407 Node* dst = phi_map->in(i); 408 if (src != dst) { 409 PhiNode* phi; 410 if (dst->in(0) != region) { 411 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 412 record_for_igvn(phi); 413 _gvn.set_type(phi, phi->type()); 414 phi_map->set_req(i, dst); 415 // Prepare to append interesting stuff onto the new phi: 416 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 417 } else { 418 assert(dst->is_Phi(), "nobody else uses a hidden region"); 419 phi = dst->as_Phi(); 420 } 421 if (add_multiple && src->in(0) == ex_control) { 422 // Both are phis. 423 add_n_reqs(dst, src); 424 } else { 425 while (dst->req() < region->req()) add_one_req(dst, src); 426 } 427 const Type* srctype = _gvn.type(src); 428 if (phi->type() != srctype) { 429 const Type* dsttype = phi->type()->meet_speculative(srctype); 430 if (phi->type() != dsttype) { 431 phi->set_type(dsttype); 432 _gvn.set_type(phi, dsttype); 433 } 434 } 435 } 436 } 437 phi_map->merge_replaced_nodes_with(ex_map); 438 } 439 440 //--------------------------use_exception_state-------------------------------- 441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 442 if (failing()) { stop(); return top(); } 443 Node* region = phi_map->control(); 444 Node* hidden_merge_mark = root(); 445 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 446 Node* ex_oop = clear_saved_ex_oop(phi_map); 447 if (region->in(0) == hidden_merge_mark) { 448 // Special marking for internal ex-states. Process the phis now. 449 region->set_req(0, region); // now it's an ordinary region 450 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 451 // Note: Setting the jvms also sets the bci and sp. 452 set_control(_gvn.transform(region)); 453 uint tos = jvms()->stkoff() + sp(); 454 for (uint i = 1; i < tos; i++) { 455 Node* x = phi_map->in(i); 456 if (x->in(0) == region) { 457 assert(x->is_Phi(), "expected a special phi"); 458 phi_map->set_req(i, _gvn.transform(x)); 459 } 460 } 461 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 462 Node* x = mms.memory(); 463 if (x->in(0) == region) { 464 assert(x->is_Phi(), "nobody else uses a hidden region"); 465 mms.set_memory(_gvn.transform(x)); 466 } 467 } 468 if (ex_oop->in(0) == region) { 469 assert(ex_oop->is_Phi(), "expected a special phi"); 470 ex_oop = _gvn.transform(ex_oop); 471 } 472 } else { 473 set_jvms(phi_map->jvms()); 474 } 475 476 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 477 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 478 return ex_oop; 479 } 480 481 //---------------------------------java_bc------------------------------------- 482 Bytecodes::Code GraphKit::java_bc() const { 483 ciMethod* method = this->method(); 484 int bci = this->bci(); 485 if (method != NULL && bci != InvocationEntryBci) 486 return method->java_code_at_bci(bci); 487 else 488 return Bytecodes::_illegal; 489 } 490 491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 492 bool must_throw) { 493 // if the exception capability is set, then we will generate code 494 // to check the JavaThread.should_post_on_exceptions flag to see 495 // if we actually need to report exception events (for this 496 // thread). If we don't need to report exception events, we will 497 // take the normal fast path provided by add_exception_events. If 498 // exception event reporting is enabled for this thread, we will 499 // take the uncommon_trap in the BuildCutout below. 500 501 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 502 Node* jthread = _gvn.transform(new ThreadLocalNode()); 503 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 504 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 505 506 // Test the should_post_on_exceptions_flag vs. 0 507 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 508 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 509 510 // Branch to slow_path if should_post_on_exceptions_flag was true 511 { BuildCutout unless(this, tst, PROB_MAX); 512 // Do not try anything fancy if we're notifying the VM on every throw. 513 // Cf. case Bytecodes::_athrow in parse2.cpp. 514 uncommon_trap(reason, Deoptimization::Action_none, 515 (ciKlass*)NULL, (char*)NULL, must_throw); 516 } 517 518 } 519 520 //------------------------------builtin_throw---------------------------------- 521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 522 bool must_throw = true; 523 524 if (env()->jvmti_can_post_on_exceptions()) { 525 // check if we must post exception events, take uncommon trap if so 526 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 527 // here if should_post_on_exceptions is false 528 // continue on with the normal codegen 529 } 530 531 // If this particular condition has not yet happened at this 532 // bytecode, then use the uncommon trap mechanism, and allow for 533 // a future recompilation if several traps occur here. 534 // If the throw is hot, try to use a more complicated inline mechanism 535 // which keeps execution inside the compiled code. 536 bool treat_throw_as_hot = false; 537 ciMethodData* md = method()->method_data(); 538 539 if (ProfileTraps) { 540 if (too_many_traps(reason)) { 541 treat_throw_as_hot = true; 542 } 543 // (If there is no MDO at all, assume it is early in 544 // execution, and that any deopts are part of the 545 // startup transient, and don't need to be remembered.) 546 547 // Also, if there is a local exception handler, treat all throws 548 // as hot if there has been at least one in this method. 549 if (C->trap_count(reason) != 0 550 && method()->method_data()->trap_count(reason) != 0 551 && has_ex_handler()) { 552 treat_throw_as_hot = true; 553 } 554 } 555 556 // If this throw happens frequently, an uncommon trap might cause 557 // a performance pothole. If there is a local exception handler, 558 // and if this particular bytecode appears to be deoptimizing often, 559 // let us handle the throw inline, with a preconstructed instance. 560 // Note: If the deopt count has blown up, the uncommon trap 561 // runtime is going to flush this nmethod, not matter what. 562 if (treat_throw_as_hot 563 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 564 // If the throw is local, we use a pre-existing instance and 565 // punt on the backtrace. This would lead to a missing backtrace 566 // (a repeat of 4292742) if the backtrace object is ever asked 567 // for its backtrace. 568 // Fixing this remaining case of 4292742 requires some flavor of 569 // escape analysis. Leave that for the future. 570 ciInstance* ex_obj = NULL; 571 switch (reason) { 572 case Deoptimization::Reason_null_check: 573 ex_obj = env()->NullPointerException_instance(); 574 break; 575 case Deoptimization::Reason_div0_check: 576 ex_obj = env()->ArithmeticException_instance(); 577 break; 578 case Deoptimization::Reason_range_check: 579 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 580 break; 581 case Deoptimization::Reason_class_check: 582 if (java_bc() == Bytecodes::_aastore) { 583 ex_obj = env()->ArrayStoreException_instance(); 584 } else { 585 ex_obj = env()->ClassCastException_instance(); 586 } 587 break; 588 } 589 if (failing()) { stop(); return; } // exception allocation might fail 590 if (ex_obj != NULL) { 591 // Cheat with a preallocated exception object. 592 if (C->log() != NULL) 593 C->log()->elem("hot_throw preallocated='1' reason='%s'", 594 Deoptimization::trap_reason_name(reason)); 595 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 596 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 597 598 // Clear the detail message of the preallocated exception object. 599 // Weblogic sometimes mutates the detail message of exceptions 600 // using reflection. 601 int offset = java_lang_Throwable::get_detailMessage_offset(); 602 const TypePtr* adr_typ = ex_con->add_offset(offset); 603 604 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 605 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 606 // Conservatively release stores of object references. 607 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 608 609 add_exception_state(make_exception_state(ex_node)); 610 return; 611 } 612 } 613 614 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 615 // It won't be much cheaper than bailing to the interp., since we'll 616 // have to pass up all the debug-info, and the runtime will have to 617 // create the stack trace. 618 619 // Usual case: Bail to interpreter. 620 // Reserve the right to recompile if we haven't seen anything yet. 621 622 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 623 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 624 if (treat_throw_as_hot 625 && (method()->method_data()->trap_recompiled_at(bci(), m) 626 || C->too_many_traps(reason))) { 627 // We cannot afford to take more traps here. Suffer in the interpreter. 628 if (C->log() != NULL) 629 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 630 Deoptimization::trap_reason_name(reason), 631 C->trap_count(reason)); 632 action = Deoptimization::Action_none; 633 } 634 635 // "must_throw" prunes the JVM state to include only the stack, if there 636 // are no local exception handlers. This should cut down on register 637 // allocation time and code size, by drastically reducing the number 638 // of in-edges on the call to the uncommon trap. 639 640 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 641 } 642 643 644 //----------------------------PreserveJVMState--------------------------------- 645 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 646 debug_only(kit->verify_map()); 647 _kit = kit; 648 _map = kit->map(); // preserve the map 649 _sp = kit->sp(); 650 kit->set_map(clone_map ? kit->clone_map() : NULL); 651 #ifdef ASSERT 652 _bci = kit->bci(); 653 Parse* parser = kit->is_Parse(); 654 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 655 _block = block; 656 #endif 657 } 658 PreserveJVMState::~PreserveJVMState() { 659 GraphKit* kit = _kit; 660 #ifdef ASSERT 661 assert(kit->bci() == _bci, "bci must not shift"); 662 Parse* parser = kit->is_Parse(); 663 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 664 assert(block == _block, "block must not shift"); 665 #endif 666 kit->set_map(_map); 667 kit->set_sp(_sp); 668 } 669 670 671 //-----------------------------BuildCutout------------------------------------- 672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 673 : PreserveJVMState(kit) 674 { 675 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 676 SafePointNode* outer_map = _map; // preserved map is caller's 677 SafePointNode* inner_map = kit->map(); 678 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 679 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 680 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 681 } 682 BuildCutout::~BuildCutout() { 683 GraphKit* kit = _kit; 684 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 685 } 686 687 //---------------------------PreserveReexecuteState---------------------------- 688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 689 assert(!kit->stopped(), "must call stopped() before"); 690 _kit = kit; 691 _sp = kit->sp(); 692 _reexecute = kit->jvms()->_reexecute; 693 } 694 PreserveReexecuteState::~PreserveReexecuteState() { 695 if (_kit->stopped()) return; 696 _kit->jvms()->_reexecute = _reexecute; 697 _kit->set_sp(_sp); 698 } 699 700 //------------------------------clone_map-------------------------------------- 701 // Implementation of PreserveJVMState 702 // 703 // Only clone_map(...) here. If this function is only used in the 704 // PreserveJVMState class we may want to get rid of this extra 705 // function eventually and do it all there. 706 707 SafePointNode* GraphKit::clone_map() { 708 if (map() == NULL) return NULL; 709 710 // Clone the memory edge first 711 Node* mem = MergeMemNode::make(map()->memory()); 712 gvn().set_type_bottom(mem); 713 714 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 715 JVMState* jvms = this->jvms(); 716 JVMState* clonejvms = jvms->clone_shallow(C); 717 clonemap->set_memory(mem); 718 clonemap->set_jvms(clonejvms); 719 clonejvms->set_map(clonemap); 720 record_for_igvn(clonemap); 721 gvn().set_type_bottom(clonemap); 722 return clonemap; 723 } 724 725 726 //-----------------------------set_map_clone----------------------------------- 727 void GraphKit::set_map_clone(SafePointNode* m) { 728 _map = m; 729 _map = clone_map(); 730 _map->set_next_exception(NULL); 731 debug_only(verify_map()); 732 } 733 734 735 //----------------------------kill_dead_locals--------------------------------- 736 // Detect any locals which are known to be dead, and force them to top. 737 void GraphKit::kill_dead_locals() { 738 // Consult the liveness information for the locals. If any 739 // of them are unused, then they can be replaced by top(). This 740 // should help register allocation time and cut down on the size 741 // of the deoptimization information. 742 743 // This call is made from many of the bytecode handling 744 // subroutines called from the Big Switch in do_one_bytecode. 745 // Every bytecode which might include a slow path is responsible 746 // for killing its dead locals. The more consistent we 747 // are about killing deads, the fewer useless phis will be 748 // constructed for them at various merge points. 749 750 // bci can be -1 (InvocationEntryBci). We return the entry 751 // liveness for the method. 752 753 if (method() == NULL || method()->code_size() == 0) { 754 // We are building a graph for a call to a native method. 755 // All locals are live. 756 return; 757 } 758 759 ResourceMark rm; 760 761 // Consult the liveness information for the locals. If any 762 // of them are unused, then they can be replaced by top(). This 763 // should help register allocation time and cut down on the size 764 // of the deoptimization information. 765 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 766 767 int len = (int)live_locals.size(); 768 assert(len <= jvms()->loc_size(), "too many live locals"); 769 for (int local = 0; local < len; local++) { 770 if (!live_locals.at(local)) { 771 set_local(local, top()); 772 } 773 } 774 } 775 776 #ifdef ASSERT 777 //-------------------------dead_locals_are_killed------------------------------ 778 // Return true if all dead locals are set to top in the map. 779 // Used to assert "clean" debug info at various points. 780 bool GraphKit::dead_locals_are_killed() { 781 if (method() == NULL || method()->code_size() == 0) { 782 // No locals need to be dead, so all is as it should be. 783 return true; 784 } 785 786 // Make sure somebody called kill_dead_locals upstream. 787 ResourceMark rm; 788 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 789 if (jvms->loc_size() == 0) continue; // no locals to consult 790 SafePointNode* map = jvms->map(); 791 ciMethod* method = jvms->method(); 792 int bci = jvms->bci(); 793 if (jvms == this->jvms()) { 794 bci = this->bci(); // it might not yet be synched 795 } 796 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 797 int len = (int)live_locals.size(); 798 if (!live_locals.is_valid() || len == 0) 799 // This method is trivial, or is poisoned by a breakpoint. 800 return true; 801 assert(len == jvms->loc_size(), "live map consistent with locals map"); 802 for (int local = 0; local < len; local++) { 803 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 804 if (PrintMiscellaneous && (Verbose || WizardMode)) { 805 tty->print_cr("Zombie local %d: ", local); 806 jvms->dump(); 807 } 808 return false; 809 } 810 } 811 } 812 return true; 813 } 814 815 #endif //ASSERT 816 817 // Helper function for enforcing certain bytecodes to reexecute if 818 // deoptimization happens 819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 820 ciMethod* cur_method = jvms->method(); 821 int cur_bci = jvms->bci(); 822 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 823 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 824 return Interpreter::bytecode_should_reexecute(code) || 825 is_anewarray && code == Bytecodes::_multianewarray; 826 // Reexecute _multianewarray bytecode which was replaced with 827 // sequence of [a]newarray. See Parse::do_multianewarray(). 828 // 829 // Note: interpreter should not have it set since this optimization 830 // is limited by dimensions and guarded by flag so in some cases 831 // multianewarray() runtime calls will be generated and 832 // the bytecode should not be reexecutes (stack will not be reset). 833 } else 834 return false; 835 } 836 837 // Helper function for adding JVMState and debug information to node 838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 839 // Add the safepoint edges to the call (or other safepoint). 840 841 // Make sure dead locals are set to top. This 842 // should help register allocation time and cut down on the size 843 // of the deoptimization information. 844 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 845 846 // Walk the inline list to fill in the correct set of JVMState's 847 // Also fill in the associated edges for each JVMState. 848 849 // If the bytecode needs to be reexecuted we need to put 850 // the arguments back on the stack. 851 const bool should_reexecute = jvms()->should_reexecute(); 852 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 853 854 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 855 // undefined if the bci is different. This is normal for Parse but it 856 // should not happen for LibraryCallKit because only one bci is processed. 857 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 858 "in LibraryCallKit the reexecute bit should not change"); 859 860 // If we are guaranteed to throw, we can prune everything but the 861 // input to the current bytecode. 862 bool can_prune_locals = false; 863 uint stack_slots_not_pruned = 0; 864 int inputs = 0, depth = 0; 865 if (must_throw) { 866 assert(method() == youngest_jvms->method(), "sanity"); 867 if (compute_stack_effects(inputs, depth)) { 868 can_prune_locals = true; 869 stack_slots_not_pruned = inputs; 870 } 871 } 872 873 if (env()->should_retain_local_variables()) { 874 // At any safepoint, this method can get breakpointed, which would 875 // then require an immediate deoptimization. 876 can_prune_locals = false; // do not prune locals 877 stack_slots_not_pruned = 0; 878 } 879 880 // do not scribble on the input jvms 881 JVMState* out_jvms = youngest_jvms->clone_deep(C); 882 call->set_jvms(out_jvms); // Start jvms list for call node 883 884 // For a known set of bytecodes, the interpreter should reexecute them if 885 // deoptimization happens. We set the reexecute state for them here 886 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 887 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 888 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 889 } 890 891 // Presize the call: 892 DEBUG_ONLY(uint non_debug_edges = call->req()); 893 call->add_req_batch(top(), youngest_jvms->debug_depth()); 894 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 895 896 // Set up edges so that the call looks like this: 897 // Call [state:] ctl io mem fptr retadr 898 // [parms:] parm0 ... parmN 899 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 900 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 901 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 902 // Note that caller debug info precedes callee debug info. 903 904 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 905 uint debug_ptr = call->req(); 906 907 // Loop over the map input edges associated with jvms, add them 908 // to the call node, & reset all offsets to match call node array. 909 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 910 uint debug_end = debug_ptr; 911 uint debug_start = debug_ptr - in_jvms->debug_size(); 912 debug_ptr = debug_start; // back up the ptr 913 914 uint p = debug_start; // walks forward in [debug_start, debug_end) 915 uint j, k, l; 916 SafePointNode* in_map = in_jvms->map(); 917 out_jvms->set_map(call); 918 919 if (can_prune_locals) { 920 assert(in_jvms->method() == out_jvms->method(), "sanity"); 921 // If the current throw can reach an exception handler in this JVMS, 922 // then we must keep everything live that can reach that handler. 923 // As a quick and dirty approximation, we look for any handlers at all. 924 if (in_jvms->method()->has_exception_handlers()) { 925 can_prune_locals = false; 926 } 927 } 928 929 // Add the Locals 930 k = in_jvms->locoff(); 931 l = in_jvms->loc_size(); 932 out_jvms->set_locoff(p); 933 if (!can_prune_locals) { 934 for (j = 0; j < l; j++) 935 call->set_req(p++, in_map->in(k+j)); 936 } else { 937 p += l; // already set to top above by add_req_batch 938 } 939 940 // Add the Expression Stack 941 k = in_jvms->stkoff(); 942 l = in_jvms->sp(); 943 out_jvms->set_stkoff(p); 944 if (!can_prune_locals) { 945 for (j = 0; j < l; j++) 946 call->set_req(p++, in_map->in(k+j)); 947 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 948 // Divide stack into {S0,...,S1}, where S0 is set to top. 949 uint s1 = stack_slots_not_pruned; 950 stack_slots_not_pruned = 0; // for next iteration 951 if (s1 > l) s1 = l; 952 uint s0 = l - s1; 953 p += s0; // skip the tops preinstalled by add_req_batch 954 for (j = s0; j < l; j++) 955 call->set_req(p++, in_map->in(k+j)); 956 } else { 957 p += l; // already set to top above by add_req_batch 958 } 959 960 // Add the Monitors 961 k = in_jvms->monoff(); 962 l = in_jvms->mon_size(); 963 out_jvms->set_monoff(p); 964 for (j = 0; j < l; j++) 965 call->set_req(p++, in_map->in(k+j)); 966 967 // Copy any scalar object fields. 968 k = in_jvms->scloff(); 969 l = in_jvms->scl_size(); 970 out_jvms->set_scloff(p); 971 for (j = 0; j < l; j++) 972 call->set_req(p++, in_map->in(k+j)); 973 974 // Finish the new jvms. 975 out_jvms->set_endoff(p); 976 977 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 978 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 979 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 980 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 981 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 982 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 983 984 // Update the two tail pointers in parallel. 985 out_jvms = out_jvms->caller(); 986 in_jvms = in_jvms->caller(); 987 } 988 989 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 990 991 // Test the correctness of JVMState::debug_xxx accessors: 992 assert(call->jvms()->debug_start() == non_debug_edges, ""); 993 assert(call->jvms()->debug_end() == call->req(), ""); 994 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 995 } 996 997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 998 Bytecodes::Code code = java_bc(); 999 if (code == Bytecodes::_wide) { 1000 code = method()->java_code_at_bci(bci() + 1); 1001 } 1002 1003 BasicType rtype = T_ILLEGAL; 1004 int rsize = 0; 1005 1006 if (code != Bytecodes::_illegal) { 1007 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1008 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1009 if (rtype < T_CONFLICT) 1010 rsize = type2size[rtype]; 1011 } 1012 1013 switch (code) { 1014 case Bytecodes::_illegal: 1015 return false; 1016 1017 case Bytecodes::_ldc: 1018 case Bytecodes::_ldc_w: 1019 case Bytecodes::_ldc2_w: 1020 inputs = 0; 1021 break; 1022 1023 case Bytecodes::_dup: inputs = 1; break; 1024 case Bytecodes::_dup_x1: inputs = 2; break; 1025 case Bytecodes::_dup_x2: inputs = 3; break; 1026 case Bytecodes::_dup2: inputs = 2; break; 1027 case Bytecodes::_dup2_x1: inputs = 3; break; 1028 case Bytecodes::_dup2_x2: inputs = 4; break; 1029 case Bytecodes::_swap: inputs = 2; break; 1030 case Bytecodes::_arraylength: inputs = 1; break; 1031 1032 case Bytecodes::_getstatic: 1033 case Bytecodes::_putstatic: 1034 case Bytecodes::_getfield: 1035 case Bytecodes::_vgetfield: 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::_invokedirect: 1054 case Bytecodes::_invokespecial: 1055 case Bytecodes::_invokestatic: 1056 case Bytecodes::_invokedynamic: 1057 case Bytecodes::_invokeinterface: 1058 { 1059 bool ignored_will_link; 1060 ciSignature* declared_signature = NULL; 1061 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1062 assert(declared_signature != NULL, "cannot be null"); 1063 inputs = declared_signature->arg_size_for_bc(code); 1064 int size = declared_signature->return_type()->size(); 1065 depth = size - inputs; 1066 } 1067 break; 1068 1069 case Bytecodes::_multianewarray: 1070 { 1071 ciBytecodeStream iter(method()); 1072 iter.reset_to_bci(bci()); 1073 iter.next(); 1074 inputs = iter.get_dimensions(); 1075 assert(rsize == 1, ""); 1076 depth = rsize - inputs; 1077 } 1078 break; 1079 1080 case Bytecodes::_vnew: { 1081 // vnew pops the values from the stack 1082 ciValueKlass* vk = method()->holder()->as_value_klass(); 1083 inputs = vk->field_size(); 1084 depth = rsize - inputs; 1085 break; 1086 } 1087 1088 case Bytecodes::_ireturn: 1089 case Bytecodes::_lreturn: 1090 case Bytecodes::_freturn: 1091 case Bytecodes::_dreturn: 1092 case Bytecodes::_areturn: 1093 case Bytecodes::_vreturn: 1094 assert(rsize = -depth, ""); 1095 inputs = rsize; 1096 break; 1097 1098 case Bytecodes::_jsr: 1099 case Bytecodes::_jsr_w: 1100 inputs = 0; 1101 depth = 1; // S.B. depth=1, not zero 1102 break; 1103 1104 default: 1105 // bytecode produces a typed result 1106 inputs = rsize - depth; 1107 assert(inputs >= 0, ""); 1108 break; 1109 } 1110 1111 #ifdef ASSERT 1112 // spot check 1113 int outputs = depth + inputs; 1114 assert(outputs >= 0, "sanity"); 1115 switch (code) { 1116 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1117 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1118 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1119 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1120 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1121 } 1122 #endif //ASSERT 1123 1124 return true; 1125 } 1126 1127 1128 1129 //------------------------------basic_plus_adr--------------------------------- 1130 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1131 // short-circuit a common case 1132 if (offset == intcon(0)) return ptr; 1133 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1134 } 1135 1136 Node* GraphKit::ConvI2L(Node* offset) { 1137 // short-circuit a common case 1138 jint offset_con = find_int_con(offset, Type::OffsetBot); 1139 if (offset_con != Type::OffsetBot) { 1140 return longcon((jlong) offset_con); 1141 } 1142 return _gvn.transform( new ConvI2LNode(offset)); 1143 } 1144 1145 Node* GraphKit::ConvI2UL(Node* offset) { 1146 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1147 if (offset_con != (juint) Type::OffsetBot) { 1148 return longcon((julong) offset_con); 1149 } 1150 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1151 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1152 return _gvn.transform( new AndLNode(conv, mask) ); 1153 } 1154 1155 Node* GraphKit::ConvL2I(Node* offset) { 1156 // short-circuit a common case 1157 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1158 if (offset_con != (jlong)Type::OffsetBot) { 1159 return intcon((int) offset_con); 1160 } 1161 return _gvn.transform( new ConvL2INode(offset)); 1162 } 1163 1164 //-------------------------load_object_klass----------------------------------- 1165 Node* GraphKit::load_object_klass(Node* obj) { 1166 // Special-case a fresh allocation to avoid building nodes: 1167 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1168 if (akls != NULL) return akls; 1169 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1170 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1171 } 1172 1173 //-------------------------load_array_length----------------------------------- 1174 Node* GraphKit::load_array_length(Node* array) { 1175 // Special-case a fresh allocation to avoid building nodes: 1176 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1177 Node *alen; 1178 if (alloc == NULL) { 1179 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1180 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1181 } else { 1182 alen = alloc->Ideal_length(); 1183 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1184 if (ccast != alen) { 1185 alen = _gvn.transform(ccast); 1186 } 1187 } 1188 return alen; 1189 } 1190 1191 //------------------------------do_null_check---------------------------------- 1192 // Helper function to do a NULL pointer check. Returned value is 1193 // the incoming address with NULL casted away. You are allowed to use the 1194 // not-null value only if you are control dependent on the test. 1195 #ifndef PRODUCT 1196 extern int explicit_null_checks_inserted, 1197 explicit_null_checks_elided; 1198 #endif 1199 Node* GraphKit::null_check_common(Node* value, BasicType type, 1200 // optional arguments for variations: 1201 bool assert_null, 1202 Node* *null_control, 1203 bool speculative) { 1204 assert(!assert_null || null_control == NULL, "not both at once"); 1205 if (stopped()) return top(); 1206 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1207 // For some performance testing, we may wish to suppress null checking. 1208 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1209 return value; 1210 } 1211 NOT_PRODUCT(explicit_null_checks_inserted++); 1212 1213 // Construct NULL check 1214 Node *chk = NULL; 1215 switch(type) { 1216 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1217 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1218 case T_ARRAY : // fall through 1219 type = T_OBJECT; // simplify further tests 1220 case T_OBJECT : { 1221 const Type *t = _gvn.type( value ); 1222 1223 const TypeOopPtr* tp = t->isa_oopptr(); 1224 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1225 // Only for do_null_check, not any of its siblings: 1226 && !assert_null && null_control == NULL) { 1227 // Usually, any field access or invocation on an unloaded oop type 1228 // will simply fail to link, since the statically linked class is 1229 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1230 // the static class is loaded but the sharper oop type is not. 1231 // Rather than checking for this obscure case in lots of places, 1232 // we simply observe that a null check on an unloaded class 1233 // will always be followed by a nonsense operation, so we 1234 // can just issue the uncommon trap here. 1235 // Our access to the unloaded class will only be correct 1236 // after it has been loaded and initialized, which requires 1237 // a trip through the interpreter. 1238 #ifndef PRODUCT 1239 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1240 #endif 1241 uncommon_trap(Deoptimization::Reason_unloaded, 1242 Deoptimization::Action_reinterpret, 1243 tp->klass(), "!loaded"); 1244 return top(); 1245 } 1246 1247 if (assert_null) { 1248 // See if the type is contained in NULL_PTR. 1249 // If so, then the value is already null. 1250 if (t->higher_equal(TypePtr::NULL_PTR)) { 1251 NOT_PRODUCT(explicit_null_checks_elided++); 1252 return value; // Elided null assert quickly! 1253 } 1254 } else { 1255 // See if mixing in the NULL pointer changes type. 1256 // If so, then the NULL pointer was not allowed in the original 1257 // type. In other words, "value" was not-null. 1258 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1259 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1260 NOT_PRODUCT(explicit_null_checks_elided++); 1261 return value; // Elided null check quickly! 1262 } 1263 } 1264 chk = new CmpPNode( value, null() ); 1265 break; 1266 } 1267 1268 default: 1269 fatal("unexpected type: %s", type2name(type)); 1270 } 1271 assert(chk != NULL, "sanity check"); 1272 chk = _gvn.transform(chk); 1273 1274 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1275 BoolNode *btst = new BoolNode( chk, btest); 1276 Node *tst = _gvn.transform( btst ); 1277 1278 //----------- 1279 // if peephole optimizations occurred, a prior test existed. 1280 // If a prior test existed, maybe it dominates as we can avoid this test. 1281 if (tst != btst && type == T_OBJECT) { 1282 // At this point we want to scan up the CFG to see if we can 1283 // find an identical test (and so avoid this test altogether). 1284 Node *cfg = control(); 1285 int depth = 0; 1286 while( depth < 16 ) { // Limit search depth for speed 1287 if( cfg->Opcode() == Op_IfTrue && 1288 cfg->in(0)->in(1) == tst ) { 1289 // Found prior test. Use "cast_not_null" to construct an identical 1290 // CastPP (and hence hash to) as already exists for the prior test. 1291 // Return that casted value. 1292 if (assert_null) { 1293 replace_in_map(value, null()); 1294 return null(); // do not issue the redundant test 1295 } 1296 Node *oldcontrol = control(); 1297 set_control(cfg); 1298 Node *res = cast_not_null(value); 1299 set_control(oldcontrol); 1300 NOT_PRODUCT(explicit_null_checks_elided++); 1301 return res; 1302 } 1303 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1304 if (cfg == NULL) break; // Quit at region nodes 1305 depth++; 1306 } 1307 } 1308 1309 //----------- 1310 // Branch to failure if null 1311 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1312 Deoptimization::DeoptReason reason; 1313 if (assert_null) { 1314 reason = Deoptimization::Reason_null_assert; 1315 } else if (type == T_OBJECT) { 1316 reason = Deoptimization::reason_null_check(speculative); 1317 } else { 1318 reason = Deoptimization::Reason_div0_check; 1319 } 1320 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1321 // ciMethodData::has_trap_at will return a conservative -1 if any 1322 // must-be-null assertion has failed. This could cause performance 1323 // problems for a method after its first do_null_assert failure. 1324 // Consider using 'Reason_class_check' instead? 1325 1326 // To cause an implicit null check, we set the not-null probability 1327 // to the maximum (PROB_MAX). For an explicit check the probability 1328 // is set to a smaller value. 1329 if (null_control != NULL || too_many_traps(reason)) { 1330 // probability is less likely 1331 ok_prob = PROB_LIKELY_MAG(3); 1332 } else if (!assert_null && 1333 (ImplicitNullCheckThreshold > 0) && 1334 method() != NULL && 1335 (method()->method_data()->trap_count(reason) 1336 >= (uint)ImplicitNullCheckThreshold)) { 1337 ok_prob = PROB_LIKELY_MAG(3); 1338 } 1339 1340 if (null_control != NULL) { 1341 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1342 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1343 set_control( _gvn.transform( new IfTrueNode(iff))); 1344 #ifndef PRODUCT 1345 if (null_true == top()) { 1346 explicit_null_checks_elided++; 1347 } 1348 #endif 1349 (*null_control) = null_true; 1350 } else { 1351 BuildCutout unless(this, tst, ok_prob); 1352 // Check for optimizer eliding test at parse time 1353 if (stopped()) { 1354 // Failure not possible; do not bother making uncommon trap. 1355 NOT_PRODUCT(explicit_null_checks_elided++); 1356 } else if (assert_null) { 1357 uncommon_trap(reason, 1358 Deoptimization::Action_make_not_entrant, 1359 NULL, "assert_null"); 1360 } else { 1361 replace_in_map(value, zerocon(type)); 1362 builtin_throw(reason); 1363 } 1364 } 1365 1366 // Must throw exception, fall-thru not possible? 1367 if (stopped()) { 1368 return top(); // No result 1369 } 1370 1371 if (assert_null) { 1372 // Cast obj to null on this path. 1373 replace_in_map(value, zerocon(type)); 1374 return zerocon(type); 1375 } 1376 1377 // Cast obj to not-null on this path, if there is no null_control. 1378 // (If there is a null_control, a non-null value may come back to haunt us.) 1379 if (type == T_OBJECT) { 1380 Node* cast = cast_not_null(value, false); 1381 if (null_control == NULL || (*null_control) == top()) 1382 replace_in_map(value, cast); 1383 value = cast; 1384 } 1385 1386 return value; 1387 } 1388 1389 1390 //------------------------------cast_not_null---------------------------------- 1391 // Cast obj to not-null on this path 1392 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1393 const Type *t = _gvn.type(obj); 1394 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1395 // Object is already not-null? 1396 if( t == t_not_null ) return obj; 1397 1398 Node *cast = new CastPPNode(obj,t_not_null); 1399 cast->init_req(0, control()); 1400 cast = _gvn.transform( cast ); 1401 1402 // Scan for instances of 'obj' in the current JVM mapping. 1403 // These instances are known to be not-null after the test. 1404 if (do_replace_in_map) 1405 replace_in_map(obj, cast); 1406 1407 return cast; // Return casted value 1408 } 1409 1410 1411 //--------------------------replace_in_map------------------------------------- 1412 void GraphKit::replace_in_map(Node* old, Node* neww) { 1413 if (old == neww) { 1414 return; 1415 } 1416 1417 map()->replace_edge(old, neww); 1418 1419 // Note: This operation potentially replaces any edge 1420 // on the map. This includes locals, stack, and monitors 1421 // of the current (innermost) JVM state. 1422 1423 // don't let inconsistent types from profiling escape this 1424 // method 1425 1426 const Type* told = _gvn.type(old); 1427 const Type* tnew = _gvn.type(neww); 1428 1429 if (!tnew->higher_equal(told)) { 1430 return; 1431 } 1432 1433 map()->record_replaced_node(old, neww); 1434 } 1435 1436 1437 //============================================================================= 1438 //--------------------------------memory--------------------------------------- 1439 Node* GraphKit::memory(uint alias_idx) { 1440 MergeMemNode* mem = merged_memory(); 1441 Node* p = mem->memory_at(alias_idx); 1442 _gvn.set_type(p, Type::MEMORY); // must be mapped 1443 return p; 1444 } 1445 1446 //-----------------------------reset_memory------------------------------------ 1447 Node* GraphKit::reset_memory() { 1448 Node* mem = map()->memory(); 1449 // do not use this node for any more parsing! 1450 debug_only( map()->set_memory((Node*)NULL) ); 1451 return _gvn.transform( mem ); 1452 } 1453 1454 //------------------------------set_all_memory--------------------------------- 1455 void GraphKit::set_all_memory(Node* newmem) { 1456 Node* mergemem = MergeMemNode::make(newmem); 1457 gvn().set_type_bottom(mergemem); 1458 map()->set_memory(mergemem); 1459 } 1460 1461 //------------------------------set_all_memory_call---------------------------- 1462 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1463 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1464 set_all_memory(newmem); 1465 } 1466 1467 //============================================================================= 1468 // 1469 // parser factory methods for MemNodes 1470 // 1471 // These are layered on top of the factory methods in LoadNode and StoreNode, 1472 // and integrate with the parser's memory state and _gvn engine. 1473 // 1474 1475 // factory methods in "int adr_idx" 1476 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1477 int adr_idx, 1478 MemNode::MemOrd mo, 1479 LoadNode::ControlDependency control_dependency, 1480 bool require_atomic_access, 1481 bool unaligned, 1482 bool mismatched) { 1483 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1484 const TypePtr* adr_type = NULL; // debug-mode-only argument 1485 debug_only(adr_type = C->get_adr_type(adr_idx)); 1486 Node* mem = memory(adr_idx); 1487 Node* ld; 1488 if (require_atomic_access && bt == T_LONG) { 1489 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1490 } else if (require_atomic_access && bt == T_DOUBLE) { 1491 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1492 } else { 1493 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1494 } 1495 ld = _gvn.transform(ld); 1496 1497 if (bt == T_VALUETYPE) { 1498 // Load value type from oop 1499 ld = ValueTypeNode::make(gvn(), map()->memory(), ld); 1500 } else if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1501 // Improve graph before escape analysis and boxing elimination. 1502 record_for_igvn(ld); 1503 } 1504 return ld; 1505 } 1506 1507 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1508 int adr_idx, 1509 MemNode::MemOrd mo, 1510 bool require_atomic_access, 1511 bool unaligned, 1512 bool mismatched) { 1513 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1514 const TypePtr* adr_type = NULL; 1515 debug_only(adr_type = C->get_adr_type(adr_idx)); 1516 Node *mem = memory(adr_idx); 1517 Node* st; 1518 if (require_atomic_access && bt == T_LONG) { 1519 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1520 } else if (require_atomic_access && bt == T_DOUBLE) { 1521 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1522 } else { 1523 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1524 } 1525 if (unaligned) { 1526 st->as_Store()->set_unaligned_access(); 1527 } 1528 if (mismatched) { 1529 st->as_Store()->set_mismatched_access(); 1530 } 1531 st = _gvn.transform(st); 1532 set_memory(st, adr_idx); 1533 // Back-to-back stores can only remove intermediate store with DU info 1534 // so push on worklist for optimizer. 1535 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1536 record_for_igvn(st); 1537 1538 return st; 1539 } 1540 1541 1542 void GraphKit::pre_barrier(bool do_load, 1543 Node* ctl, 1544 Node* obj, 1545 Node* adr, 1546 uint adr_idx, 1547 Node* val, 1548 const TypeOopPtr* val_type, 1549 Node* pre_val, 1550 BasicType bt) { 1551 1552 BarrierSet* bs = Universe::heap()->barrier_set(); 1553 set_control(ctl); 1554 switch (bs->kind()) { 1555 case BarrierSet::G1SATBCTLogging: 1556 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1557 break; 1558 1559 case BarrierSet::CardTableForRS: 1560 case BarrierSet::CardTableExtension: 1561 case BarrierSet::ModRef: 1562 break; 1563 1564 default : 1565 ShouldNotReachHere(); 1566 1567 } 1568 } 1569 1570 bool GraphKit::can_move_pre_barrier() const { 1571 BarrierSet* bs = Universe::heap()->barrier_set(); 1572 switch (bs->kind()) { 1573 case BarrierSet::G1SATBCTLogging: 1574 return true; // Can move it if no safepoint 1575 1576 case BarrierSet::CardTableForRS: 1577 case BarrierSet::CardTableExtension: 1578 case BarrierSet::ModRef: 1579 return true; // There is no pre-barrier 1580 1581 default : 1582 ShouldNotReachHere(); 1583 } 1584 return false; 1585 } 1586 1587 void GraphKit::post_barrier(Node* ctl, 1588 Node* store, 1589 Node* obj, 1590 Node* adr, 1591 uint adr_idx, 1592 Node* val, 1593 BasicType bt, 1594 bool use_precise) { 1595 BarrierSet* bs = Universe::heap()->barrier_set(); 1596 set_control(ctl); 1597 switch (bs->kind()) { 1598 case BarrierSet::G1SATBCTLogging: 1599 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1600 break; 1601 1602 case BarrierSet::CardTableForRS: 1603 case BarrierSet::CardTableExtension: 1604 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1605 break; 1606 1607 case BarrierSet::ModRef: 1608 break; 1609 1610 default : 1611 ShouldNotReachHere(); 1612 1613 } 1614 } 1615 1616 Node* GraphKit::store_oop(Node* ctl, 1617 Node* obj, 1618 Node* adr, 1619 const TypePtr* adr_type, 1620 Node* val, 1621 const TypeOopPtr* val_type, 1622 BasicType bt, 1623 bool use_precise, 1624 MemNode::MemOrd mo, 1625 bool mismatched) { 1626 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1627 // could be delayed during Parse (for example, in adjust_map_after_if()). 1628 // Execute transformation here to avoid barrier generation in such case. 1629 if (_gvn.type(val) == TypePtr::NULL_PTR) 1630 val = _gvn.makecon(TypePtr::NULL_PTR); 1631 1632 set_control(ctl); 1633 if (stopped()) return top(); // Dead path ? 1634 1635 assert(bt == T_OBJECT || bt == T_VALUETYPE, "sanity"); 1636 assert(val != NULL, "not dead path"); 1637 uint adr_idx = C->get_alias_index(adr_type); 1638 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1639 1640 if (bt == T_VALUETYPE) { 1641 // Allocate value type and store oop 1642 val = val->as_ValueType()->store_to_memory(this); 1643 } 1644 1645 pre_barrier(true /* do_load */, 1646 control(), obj, adr, adr_idx, val, val_type, 1647 NULL /* pre_val */, 1648 bt); 1649 1650 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched); 1651 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1652 return store; 1653 } 1654 1655 // Could be an array or object we don't know at compile time (unsafe ref.) 1656 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1657 Node* obj, // containing obj 1658 Node* adr, // actual adress to store val at 1659 const TypePtr* adr_type, 1660 Node* val, 1661 BasicType bt, 1662 MemNode::MemOrd mo, 1663 bool mismatched) { 1664 Compile::AliasType* at = C->alias_type(adr_type); 1665 const TypeOopPtr* val_type = NULL; 1666 if (adr_type->isa_instptr()) { 1667 if (at->field() != NULL) { 1668 // known field. This code is a copy of the do_put_xxx logic. 1669 ciField* field = at->field(); 1670 if (!field->type()->is_loaded()) { 1671 val_type = TypeInstPtr::BOTTOM; 1672 } else { 1673 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1674 } 1675 } 1676 } else if (adr_type->isa_aryptr()) { 1677 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1678 } 1679 if (val_type == NULL) { 1680 val_type = TypeInstPtr::BOTTOM; 1681 } 1682 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched); 1683 } 1684 1685 1686 //-------------------------array_element_address------------------------- 1687 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1688 const TypeInt* sizetype, Node* ctrl) { 1689 uint shift = exact_log2(type2aelembytes(elembt)); 1690 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1691 1692 // short-circuit a common case (saves lots of confusing waste motion) 1693 jint idx_con = find_int_con(idx, -1); 1694 if (idx_con >= 0) { 1695 intptr_t offset = header + ((intptr_t)idx_con << shift); 1696 return basic_plus_adr(ary, offset); 1697 } 1698 1699 // must be correct type for alignment purposes 1700 Node* base = basic_plus_adr(ary, header); 1701 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1702 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1703 return basic_plus_adr(ary, base, scale); 1704 } 1705 1706 //-------------------------load_array_element------------------------- 1707 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1708 const Type* elemtype = arytype->elem(); 1709 BasicType elembt = elemtype->array_element_basic_type(); 1710 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1711 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1712 return ld; 1713 } 1714 1715 //-------------------------set_arguments_for_java_call------------------------- 1716 // Arguments (pre-popped from the stack) are taken from the JVMS. 1717 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1718 // Add the call arguments: 1719 uint nargs = call->method()->arg_size(); 1720 for (uint i = 0; i < nargs; i++) { 1721 Node* arg = argument(i); 1722 if (arg->is_ValueType()) { 1723 // Pass value type argument via oop to callee 1724 arg = arg->as_ValueType()->store_to_memory(this); 1725 } 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 #ifdef ASSERT 2031 case Deoptimization::Action_none: 2032 case Deoptimization::Action_make_not_compilable: 2033 break; 2034 default: 2035 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2036 break; 2037 #endif 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 // Make sure all value types are allocated before calling uncommon_trap 2077 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 2078 SafePointNode* map = jvms->map(); 2079 // Search for value type locals 2080 for (int i = 0; i < jvms->loc_size(); ++i) { 2081 Node* local = map->local(jvms, i); 2082 if (local->is_ValueType()) { 2083 // Allocate value type 2084 local = local->as_ValueType()->store_to_memory(this); 2085 map->set_local(jvms, i, local); 2086 } 2087 } 2088 } 2089 2090 // Now insert the uncommon trap subroutine call 2091 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2092 const TypePtr* no_memory_effects = NULL; 2093 // Pass the index of the class to be loaded 2094 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2095 (must_throw ? RC_MUST_THROW : 0), 2096 OptoRuntime::uncommon_trap_Type(), 2097 call_addr, "uncommon_trap", no_memory_effects, 2098 intcon(trap_request)); 2099 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2100 "must extract request correctly from the graph"); 2101 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2102 2103 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2104 // The debug info is the only real input to this call. 2105 2106 // Halt-and-catch fire here. The above call should never return! 2107 HaltNode* halt = new HaltNode(control(), frameptr()); 2108 _gvn.set_type_bottom(halt); 2109 root()->add_req(halt); 2110 2111 stop_and_kill_map(); 2112 } 2113 2114 2115 //--------------------------just_allocated_object------------------------------ 2116 // Report the object that was just allocated. 2117 // It must be the case that there are no intervening safepoints. 2118 // We use this to determine if an object is so "fresh" that 2119 // it does not require card marks. 2120 Node* GraphKit::just_allocated_object(Node* current_control) { 2121 if (C->recent_alloc_ctl() == current_control) 2122 return C->recent_alloc_obj(); 2123 return NULL; 2124 } 2125 2126 2127 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2128 // (Note: TypeFunc::make has a cache that makes this fast.) 2129 const TypeFunc* tf = TypeFunc::make(dest_method); 2130 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2131 for (int j = 0; j < nargs; j++) { 2132 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2133 if( targ->basic_type() == T_DOUBLE ) { 2134 // If any parameters are doubles, they must be rounded before 2135 // the call, dstore_rounding does gvn.transform 2136 Node *arg = argument(j); 2137 arg = dstore_rounding(arg); 2138 set_argument(j, arg); 2139 } 2140 } 2141 } 2142 2143 /** 2144 * Record profiling data exact_kls for Node n with the type system so 2145 * that it can propagate it (speculation) 2146 * 2147 * @param n node that the type applies to 2148 * @param exact_kls type from profiling 2149 * @param maybe_null did profiling see null? 2150 * 2151 * @return node with improved type 2152 */ 2153 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2154 const Type* current_type = _gvn.type(n); 2155 assert(UseTypeSpeculation, "type speculation must be on"); 2156 2157 const TypePtr* speculative = current_type->speculative(); 2158 2159 // Should the klass from the profile be recorded in the speculative type? 2160 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2161 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2162 const TypeOopPtr* xtype = tklass->as_instance_type(); 2163 assert(xtype->klass_is_exact(), "Should be exact"); 2164 // Any reason to believe n is not null (from this profiling or a previous one)? 2165 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2166 // record the new speculative type's depth 2167 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2168 speculative = speculative->with_inline_depth(jvms()->depth()); 2169 } else if (current_type->would_improve_ptr(maybe_null)) { 2170 // Profiling report that null was never seen so we can change the 2171 // speculative type to non null ptr. 2172 assert(!maybe_null, "nothing to improve"); 2173 if (speculative == NULL) { 2174 speculative = TypePtr::NOTNULL; 2175 } else { 2176 const TypePtr* ptr = TypePtr::NOTNULL; 2177 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2178 } 2179 } 2180 2181 if (speculative != current_type->speculative()) { 2182 // Build a type with a speculative type (what we think we know 2183 // about the type but will need a guard when we use it) 2184 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2185 // We're changing the type, we need a new CheckCast node to carry 2186 // the new type. The new type depends on the control: what 2187 // profiling tells us is only valid from here as far as we can 2188 // tell. 2189 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2190 cast = _gvn.transform(cast); 2191 replace_in_map(n, cast); 2192 n = cast; 2193 } 2194 2195 return n; 2196 } 2197 2198 /** 2199 * Record profiling data from receiver profiling at an invoke with the 2200 * type system so that it can propagate it (speculation) 2201 * 2202 * @param n receiver node 2203 * 2204 * @return node with improved type 2205 */ 2206 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2207 if (!UseTypeSpeculation) { 2208 return n; 2209 } 2210 ciKlass* exact_kls = profile_has_unique_klass(); 2211 bool maybe_null = true; 2212 if (java_bc() == Bytecodes::_checkcast || 2213 java_bc() == Bytecodes::_instanceof || 2214 java_bc() == Bytecodes::_aastore) { 2215 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2216 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2217 } 2218 return record_profile_for_speculation(n, exact_kls, maybe_null); 2219 return n; 2220 } 2221 2222 /** 2223 * Record profiling data from argument profiling at an invoke with the 2224 * type system so that it can propagate it (speculation) 2225 * 2226 * @param dest_method target method for the call 2227 * @param bc what invoke bytecode is this? 2228 */ 2229 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2230 if (!UseTypeSpeculation) { 2231 return; 2232 } 2233 const TypeFunc* tf = TypeFunc::make(dest_method); 2234 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2235 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2236 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2237 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2238 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2239 bool maybe_null = true; 2240 ciKlass* better_type = NULL; 2241 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2242 record_profile_for_speculation(argument(j), better_type, maybe_null); 2243 } 2244 i++; 2245 } 2246 } 2247 } 2248 2249 /** 2250 * Record profiling data from parameter profiling at an invoke with 2251 * the type system so that it can propagate it (speculation) 2252 */ 2253 void GraphKit::record_profiled_parameters_for_speculation() { 2254 if (!UseTypeSpeculation) { 2255 return; 2256 } 2257 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2258 if (_gvn.type(local(i))->isa_oopptr()) { 2259 bool maybe_null = true; 2260 ciKlass* better_type = NULL; 2261 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2262 record_profile_for_speculation(local(i), better_type, maybe_null); 2263 } 2264 j++; 2265 } 2266 } 2267 } 2268 2269 /** 2270 * Record profiling data from return value profiling at an invoke with 2271 * the type system so that it can propagate it (speculation) 2272 */ 2273 void GraphKit::record_profiled_return_for_speculation() { 2274 if (!UseTypeSpeculation) { 2275 return; 2276 } 2277 bool maybe_null = true; 2278 ciKlass* better_type = NULL; 2279 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2280 // If profiling reports a single type for the return value, 2281 // feed it to the type system so it can propagate it as a 2282 // speculative type 2283 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2284 } 2285 } 2286 2287 void GraphKit::round_double_result(ciMethod* dest_method) { 2288 // A non-strict method may return a double value which has an extended 2289 // exponent, but this must not be visible in a caller which is 'strict' 2290 // If a strict caller invokes a non-strict callee, round a double result 2291 2292 BasicType result_type = dest_method->return_type()->basic_type(); 2293 assert( method() != NULL, "must have caller context"); 2294 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2295 // Destination method's return value is on top of stack 2296 // dstore_rounding() does gvn.transform 2297 Node *result = pop_pair(); 2298 result = dstore_rounding(result); 2299 push_pair(result); 2300 } 2301 } 2302 2303 // rounding for strict float precision conformance 2304 Node* GraphKit::precision_rounding(Node* n) { 2305 return UseStrictFP && _method->flags().is_strict() 2306 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2307 ? _gvn.transform( new RoundFloatNode(0, n) ) 2308 : n; 2309 } 2310 2311 // rounding for strict double precision conformance 2312 Node* GraphKit::dprecision_rounding(Node *n) { 2313 return UseStrictFP && _method->flags().is_strict() 2314 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2315 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2316 : n; 2317 } 2318 2319 // rounding for non-strict double stores 2320 Node* GraphKit::dstore_rounding(Node* n) { 2321 return Matcher::strict_fp_requires_explicit_rounding 2322 && UseSSE <= 1 2323 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2324 : n; 2325 } 2326 2327 //============================================================================= 2328 // Generate a fast path/slow path idiom. Graph looks like: 2329 // [foo] indicates that 'foo' is a parameter 2330 // 2331 // [in] NULL 2332 // \ / 2333 // CmpP 2334 // Bool ne 2335 // If 2336 // / \ 2337 // True False-<2> 2338 // / | 2339 // / cast_not_null 2340 // Load | | ^ 2341 // [fast_test] | | 2342 // gvn to opt_test | | 2343 // / \ | <1> 2344 // True False | 2345 // | \\ | 2346 // [slow_call] \[fast_result] 2347 // Ctl Val \ \ 2348 // | \ \ 2349 // Catch <1> \ \ 2350 // / \ ^ \ \ 2351 // Ex No_Ex | \ \ 2352 // | \ \ | \ <2> \ 2353 // ... \ [slow_res] | | \ [null_result] 2354 // \ \--+--+--- | | 2355 // \ | / \ | / 2356 // --------Region Phi 2357 // 2358 //============================================================================= 2359 // Code is structured as a series of driver functions all called 'do_XXX' that 2360 // call a set of helper functions. Helper functions first, then drivers. 2361 2362 //------------------------------null_check_oop--------------------------------- 2363 // Null check oop. Set null-path control into Region in slot 3. 2364 // Make a cast-not-nullness use the other not-null control. Return cast. 2365 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2366 bool never_see_null, 2367 bool safe_for_replace, 2368 bool speculative) { 2369 // Initial NULL check taken path 2370 (*null_control) = top(); 2371 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2372 2373 // Generate uncommon_trap: 2374 if (never_see_null && (*null_control) != top()) { 2375 // If we see an unexpected null at a check-cast we record it and force a 2376 // recompile; the offending check-cast will be compiled to handle NULLs. 2377 // If we see more than one offending BCI, then all checkcasts in the 2378 // method will be compiled to handle NULLs. 2379 PreserveJVMState pjvms(this); 2380 set_control(*null_control); 2381 replace_in_map(value, null()); 2382 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2383 uncommon_trap(reason, 2384 Deoptimization::Action_make_not_entrant); 2385 (*null_control) = top(); // NULL path is dead 2386 } 2387 if ((*null_control) == top() && safe_for_replace) { 2388 replace_in_map(value, cast); 2389 } 2390 2391 // Cast away null-ness on the result 2392 return cast; 2393 } 2394 2395 //------------------------------opt_iff---------------------------------------- 2396 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2397 // Return slow-path control. 2398 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2399 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2400 2401 // Fast path taken; set region slot 2 2402 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2403 region->init_req(2,fast_taken); // Capture fast-control 2404 2405 // Fast path not-taken, i.e. slow path 2406 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2407 return slow_taken; 2408 } 2409 2410 //-----------------------------make_runtime_call------------------------------- 2411 Node* GraphKit::make_runtime_call(int flags, 2412 const TypeFunc* call_type, address call_addr, 2413 const char* call_name, 2414 const TypePtr* adr_type, 2415 // The following parms are all optional. 2416 // The first NULL ends the list. 2417 Node* parm0, Node* parm1, 2418 Node* parm2, Node* parm3, 2419 Node* parm4, Node* parm5, 2420 Node* parm6, Node* parm7) { 2421 // Slow-path call 2422 bool is_leaf = !(flags & RC_NO_LEAF); 2423 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2424 if (call_name == NULL) { 2425 assert(!is_leaf, "must supply name for leaf"); 2426 call_name = OptoRuntime::stub_name(call_addr); 2427 } 2428 CallNode* call; 2429 if (!is_leaf) { 2430 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2431 bci(), adr_type); 2432 } else if (flags & RC_NO_FP) { 2433 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2434 } else { 2435 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2436 } 2437 2438 // The following is similar to set_edges_for_java_call, 2439 // except that the memory effects of the call are restricted to AliasIdxRaw. 2440 2441 // Slow path call has no side-effects, uses few values 2442 bool wide_in = !(flags & RC_NARROW_MEM); 2443 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2444 2445 Node* prev_mem = NULL; 2446 if (wide_in) { 2447 prev_mem = set_predefined_input_for_runtime_call(call); 2448 } else { 2449 assert(!wide_out, "narrow in => narrow out"); 2450 Node* narrow_mem = memory(adr_type); 2451 prev_mem = reset_memory(); 2452 map()->set_memory(narrow_mem); 2453 set_predefined_input_for_runtime_call(call); 2454 } 2455 2456 // Hook each parm in order. Stop looking at the first NULL. 2457 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2458 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2459 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2460 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2461 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2462 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2463 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2464 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2465 /* close each nested if ===> */ } } } } } } } } 2466 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2467 2468 if (!is_leaf) { 2469 // Non-leaves can block and take safepoints: 2470 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2471 } 2472 // Non-leaves can throw exceptions: 2473 if (has_io) { 2474 call->set_req(TypeFunc::I_O, i_o()); 2475 } 2476 2477 if (flags & RC_UNCOMMON) { 2478 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2479 // (An "if" probability corresponds roughly to an unconditional count. 2480 // Sort of.) 2481 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2482 } 2483 2484 Node* c = _gvn.transform(call); 2485 assert(c == call, "cannot disappear"); 2486 2487 if (wide_out) { 2488 // Slow path call has full side-effects. 2489 set_predefined_output_for_runtime_call(call); 2490 } else { 2491 // Slow path call has few side-effects, and/or sets few values. 2492 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2493 } 2494 2495 if (has_io) { 2496 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2497 } 2498 return call; 2499 2500 } 2501 2502 //------------------------------merge_memory----------------------------------- 2503 // Merge memory from one path into the current memory state. 2504 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2505 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2506 Node* old_slice = mms.force_memory(); 2507 Node* new_slice = mms.memory2(); 2508 if (old_slice != new_slice) { 2509 PhiNode* phi; 2510 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2511 if (mms.is_empty()) { 2512 // clone base memory Phi's inputs for this memory slice 2513 assert(old_slice == mms.base_memory(), "sanity"); 2514 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2515 _gvn.set_type(phi, Type::MEMORY); 2516 for (uint i = 1; i < phi->req(); i++) { 2517 phi->init_req(i, old_slice->in(i)); 2518 } 2519 } else { 2520 phi = old_slice->as_Phi(); // Phi was generated already 2521 } 2522 } else { 2523 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2524 _gvn.set_type(phi, Type::MEMORY); 2525 } 2526 phi->set_req(new_path, new_slice); 2527 mms.set_memory(phi); 2528 } 2529 } 2530 } 2531 2532 //------------------------------make_slow_call_ex------------------------------ 2533 // Make the exception handler hookups for the slow call 2534 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2535 if (stopped()) return; 2536 2537 // Make a catch node with just two handlers: fall-through and catch-all 2538 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2539 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2540 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2541 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2542 2543 { PreserveJVMState pjvms(this); 2544 set_control(excp); 2545 set_i_o(i_o); 2546 2547 if (excp != top()) { 2548 if (deoptimize) { 2549 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2550 uncommon_trap(Deoptimization::Reason_unhandled, 2551 Deoptimization::Action_none); 2552 } else { 2553 // Create an exception state also. 2554 // Use an exact type if the caller has specified a specific exception. 2555 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2556 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2557 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2558 } 2559 } 2560 } 2561 2562 // Get the no-exception control from the CatchNode. 2563 set_control(norm); 2564 } 2565 2566 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2567 Node* cmp = NULL; 2568 switch(bt) { 2569 case T_INT: cmp = new CmpINode(in1, in2); break; 2570 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2571 default: fatal("unexpected comparison type %s", type2name(bt)); 2572 } 2573 gvn->transform(cmp); 2574 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2575 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2576 gvn->transform(iff); 2577 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2578 return iff; 2579 } 2580 2581 2582 //-------------------------------gen_subtype_check----------------------------- 2583 // Generate a subtyping check. Takes as input the subtype and supertype. 2584 // Returns 2 values: sets the default control() to the true path and returns 2585 // the false path. Only reads invariant memory; sets no (visible) memory. 2586 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2587 // but that's not exposed to the optimizer. This call also doesn't take in an 2588 // Object; if you wish to check an Object you need to load the Object's class 2589 // prior to coming here. 2590 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2591 Compile* C = gvn->C; 2592 2593 if ((*ctrl)->is_top()) { 2594 return C->top(); 2595 } 2596 2597 // Fast check for identical types, perhaps identical constants. 2598 // The types can even be identical non-constants, in cases 2599 // involving Array.newInstance, Object.clone, etc. 2600 if (subklass == superklass) 2601 return C->top(); // false path is dead; no test needed. 2602 2603 if (gvn->type(superklass)->singleton()) { 2604 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2605 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2606 2607 // In the common case of an exact superklass, try to fold up the 2608 // test before generating code. You may ask, why not just generate 2609 // the code and then let it fold up? The answer is that the generated 2610 // code will necessarily include null checks, which do not always 2611 // completely fold away. If they are also needless, then they turn 2612 // into a performance loss. Example: 2613 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2614 // Here, the type of 'fa' is often exact, so the store check 2615 // of fa[1]=x will fold up, without testing the nullness of x. 2616 switch (C->static_subtype_check(superk, subk)) { 2617 case Compile::SSC_always_false: 2618 { 2619 Node* always_fail = *ctrl; 2620 *ctrl = gvn->C->top(); 2621 return always_fail; 2622 } 2623 case Compile::SSC_always_true: 2624 return C->top(); 2625 case Compile::SSC_easy_test: 2626 { 2627 // Just do a direct pointer compare and be done. 2628 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2629 *ctrl = gvn->transform(new IfTrueNode(iff)); 2630 return gvn->transform(new IfFalseNode(iff)); 2631 } 2632 case Compile::SSC_full_test: 2633 break; 2634 default: 2635 ShouldNotReachHere(); 2636 } 2637 } 2638 2639 // %%% Possible further optimization: Even if the superklass is not exact, 2640 // if the subklass is the unique subtype of the superklass, the check 2641 // will always succeed. We could leave a dependency behind to ensure this. 2642 2643 // First load the super-klass's check-offset 2644 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2645 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2646 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2647 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2648 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2649 2650 // Load from the sub-klass's super-class display list, or a 1-word cache of 2651 // the secondary superclass list, or a failing value with a sentinel offset 2652 // if the super-klass is an interface or exceptionally deep in the Java 2653 // hierarchy and we have to scan the secondary superclass list the hard way. 2654 // Worst-case type is a little odd: NULL is allowed as a result (usually 2655 // klass loads can never produce a NULL). 2656 Node *chk_off_X = chk_off; 2657 #ifdef _LP64 2658 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2659 #endif 2660 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2661 // For some types like interfaces the following loadKlass is from a 1-word 2662 // cache which is mutable so can't use immutable memory. Other 2663 // types load from the super-class display table which is immutable. 2664 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2665 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2666 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2667 2668 // Compile speed common case: ARE a subtype and we canNOT fail 2669 if( superklass == nkls ) 2670 return C->top(); // false path is dead; no test needed. 2671 2672 // See if we get an immediate positive hit. Happens roughly 83% of the 2673 // time. Test to see if the value loaded just previously from the subklass 2674 // is exactly the superklass. 2675 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2676 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2677 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2678 2679 // Compile speed common case: Check for being deterministic right now. If 2680 // chk_off is a constant and not equal to cacheoff then we are NOT a 2681 // subklass. In this case we need exactly the 1 test above and we can 2682 // return those results immediately. 2683 if (!might_be_cache) { 2684 Node* not_subtype_ctrl = *ctrl; 2685 *ctrl = iftrue1; // We need exactly the 1 test above 2686 return not_subtype_ctrl; 2687 } 2688 2689 // Gather the various success & failures here 2690 RegionNode *r_ok_subtype = new RegionNode(4); 2691 gvn->record_for_igvn(r_ok_subtype); 2692 RegionNode *r_not_subtype = new RegionNode(3); 2693 gvn->record_for_igvn(r_not_subtype); 2694 2695 r_ok_subtype->init_req(1, iftrue1); 2696 2697 // Check for immediate negative hit. Happens roughly 11% of the time (which 2698 // is roughly 63% of the remaining cases). Test to see if the loaded 2699 // check-offset points into the subklass display list or the 1-element 2700 // cache. If it points to the display (and NOT the cache) and the display 2701 // missed then it's not a subtype. 2702 Node *cacheoff = gvn->intcon(cacheoff_con); 2703 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2704 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2705 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2706 2707 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2708 // No performance impact (too rare) but allows sharing of secondary arrays 2709 // which has some footprint reduction. 2710 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2711 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2712 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2713 2714 // -- Roads not taken here: -- 2715 // We could also have chosen to perform the self-check at the beginning 2716 // of this code sequence, as the assembler does. This would not pay off 2717 // the same way, since the optimizer, unlike the assembler, can perform 2718 // static type analysis to fold away many successful self-checks. 2719 // Non-foldable self checks work better here in second position, because 2720 // the initial primary superclass check subsumes a self-check for most 2721 // types. An exception would be a secondary type like array-of-interface, 2722 // which does not appear in its own primary supertype display. 2723 // Finally, we could have chosen to move the self-check into the 2724 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2725 // dependent manner. But it is worthwhile to have the check here, 2726 // where it can be perhaps be optimized. The cost in code space is 2727 // small (register compare, branch). 2728 2729 // Now do a linear scan of the secondary super-klass array. Again, no real 2730 // performance impact (too rare) but it's gotta be done. 2731 // Since the code is rarely used, there is no penalty for moving it 2732 // out of line, and it can only improve I-cache density. 2733 // The decision to inline or out-of-line this final check is platform 2734 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2735 Node* psc = gvn->transform( 2736 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2737 2738 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2739 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2740 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2741 2742 // Return false path; set default control to true path. 2743 *ctrl = gvn->transform(r_ok_subtype); 2744 return gvn->transform(r_not_subtype); 2745 } 2746 2747 // Profile-driven exact type check: 2748 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2749 float prob, 2750 Node* *casted_receiver) { 2751 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2752 Node* recv_klass = load_object_klass(receiver); 2753 Node* want_klass = makecon(tklass); 2754 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2755 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2756 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2757 set_control( _gvn.transform( new IfTrueNode (iff) )); 2758 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2759 2760 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2761 assert(recv_xtype->klass_is_exact(), ""); 2762 2763 // Subsume downstream occurrences of receiver with a cast to 2764 // recv_xtype, since now we know what the type will be. 2765 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2766 (*casted_receiver) = _gvn.transform(cast); 2767 // (User must make the replace_in_map call.) 2768 2769 return fail; 2770 } 2771 2772 2773 //------------------------------seems_never_null------------------------------- 2774 // Use null_seen information if it is available from the profile. 2775 // If we see an unexpected null at a type check we record it and force a 2776 // recompile; the offending check will be recompiled to handle NULLs. 2777 // If we see several offending BCIs, then all checks in the 2778 // method will be recompiled. 2779 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2780 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2781 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2782 if (UncommonNullCast // Cutout for this technique 2783 && obj != null() // And not the -Xcomp stupid case? 2784 && !too_many_traps(reason) 2785 ) { 2786 if (speculating) { 2787 return true; 2788 } 2789 if (data == NULL) 2790 // Edge case: no mature data. Be optimistic here. 2791 return true; 2792 // If the profile has not seen a null, assume it won't happen. 2793 assert(java_bc() == Bytecodes::_checkcast || 2794 java_bc() == Bytecodes::_instanceof || 2795 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2796 return !data->as_BitData()->null_seen(); 2797 } 2798 speculating = false; 2799 return false; 2800 } 2801 2802 //------------------------maybe_cast_profiled_receiver------------------------- 2803 // If the profile has seen exactly one type, narrow to exactly that type. 2804 // Subsequent type checks will always fold up. 2805 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2806 ciKlass* require_klass, 2807 ciKlass* spec_klass, 2808 bool safe_for_replace) { 2809 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2810 2811 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2812 2813 // Make sure we haven't already deoptimized from this tactic. 2814 if (too_many_traps(reason) || too_many_recompiles(reason)) 2815 return NULL; 2816 2817 // (No, this isn't a call, but it's enough like a virtual call 2818 // to use the same ciMethod accessor to get the profile info...) 2819 // If we have a speculative type use it instead of profiling (which 2820 // may not help us) 2821 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2822 if (exact_kls != NULL) {// no cast failures here 2823 if (require_klass == NULL || 2824 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2825 // If we narrow the type to match what the type profile sees or 2826 // the speculative type, we can then remove the rest of the 2827 // cast. 2828 // This is a win, even if the exact_kls is very specific, 2829 // because downstream operations, such as method calls, 2830 // will often benefit from the sharper type. 2831 Node* exact_obj = not_null_obj; // will get updated in place... 2832 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2833 &exact_obj); 2834 { PreserveJVMState pjvms(this); 2835 set_control(slow_ctl); 2836 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2837 } 2838 if (safe_for_replace) { 2839 replace_in_map(not_null_obj, exact_obj); 2840 } 2841 return exact_obj; 2842 } 2843 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2844 } 2845 2846 return NULL; 2847 } 2848 2849 /** 2850 * Cast obj to type and emit guard unless we had too many traps here 2851 * already 2852 * 2853 * @param obj node being casted 2854 * @param type type to cast the node to 2855 * @param not_null true if we know node cannot be null 2856 */ 2857 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2858 ciKlass* type, 2859 bool not_null) { 2860 if (stopped()) { 2861 return obj; 2862 } 2863 2864 // type == NULL if profiling tells us this object is always null 2865 if (type != NULL) { 2866 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2867 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2868 2869 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2870 !too_many_traps(class_reason) && 2871 !too_many_recompiles(class_reason)) { 2872 Node* not_null_obj = NULL; 2873 // not_null is true if we know the object is not null and 2874 // there's no need for a null check 2875 if (!not_null) { 2876 Node* null_ctl = top(); 2877 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2878 assert(null_ctl->is_top(), "no null control here"); 2879 } else { 2880 not_null_obj = obj; 2881 } 2882 2883 Node* exact_obj = not_null_obj; 2884 ciKlass* exact_kls = type; 2885 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2886 &exact_obj); 2887 { 2888 PreserveJVMState pjvms(this); 2889 set_control(slow_ctl); 2890 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2891 } 2892 replace_in_map(not_null_obj, exact_obj); 2893 obj = exact_obj; 2894 } 2895 } else { 2896 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2897 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2898 Node* exact_obj = null_assert(obj); 2899 replace_in_map(obj, exact_obj); 2900 obj = exact_obj; 2901 } 2902 } 2903 return obj; 2904 } 2905 2906 //-------------------------------gen_instanceof-------------------------------- 2907 // Generate an instance-of idiom. Used by both the instance-of bytecode 2908 // and the reflective instance-of call. 2909 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2910 kill_dead_locals(); // Benefit all the uncommon traps 2911 assert( !stopped(), "dead parse path should be checked in callers" ); 2912 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2913 "must check for not-null not-dead klass in callers"); 2914 2915 // Make the merge point 2916 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2917 RegionNode* region = new RegionNode(PATH_LIMIT); 2918 Node* phi = new PhiNode(region, TypeInt::BOOL); 2919 C->set_has_split_ifs(true); // Has chance for split-if optimization 2920 2921 ciProfileData* data = NULL; 2922 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2923 data = method()->method_data()->bci_to_data(bci()); 2924 } 2925 bool speculative_not_null = false; 2926 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2927 && seems_never_null(obj, data, speculative_not_null)); 2928 2929 // Null check; get casted pointer; set region slot 3 2930 Node* null_ctl = top(); 2931 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2932 2933 // If not_null_obj is dead, only null-path is taken 2934 if (stopped()) { // Doing instance-of on a NULL? 2935 set_control(null_ctl); 2936 return intcon(0); 2937 } 2938 region->init_req(_null_path, null_ctl); 2939 phi ->init_req(_null_path, intcon(0)); // Set null path value 2940 if (null_ctl == top()) { 2941 // Do this eagerly, so that pattern matches like is_diamond_phi 2942 // will work even during parsing. 2943 assert(_null_path == PATH_LIMIT-1, "delete last"); 2944 region->del_req(_null_path); 2945 phi ->del_req(_null_path); 2946 } 2947 2948 // Do we know the type check always succeed? 2949 bool known_statically = false; 2950 if (_gvn.type(superklass)->singleton()) { 2951 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2952 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2953 if (subk != NULL && subk->is_loaded()) { 2954 int static_res = C->static_subtype_check(superk, subk); 2955 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2956 } 2957 } 2958 2959 if (known_statically && UseTypeSpeculation) { 2960 // If we know the type check always succeeds then we don't use the 2961 // profiling data at this bytecode. Don't lose it, feed it to the 2962 // type system as a speculative type. 2963 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2964 } else { 2965 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2966 // We may not have profiling here or it may not help us. If we 2967 // have a speculative type use it to perform an exact cast. 2968 ciKlass* spec_obj_type = obj_type->speculative_type(); 2969 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2970 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2971 if (stopped()) { // Profile disagrees with this path. 2972 set_control(null_ctl); // Null is the only remaining possibility. 2973 return intcon(0); 2974 } 2975 if (cast_obj != NULL) { 2976 not_null_obj = cast_obj; 2977 } 2978 } 2979 } 2980 2981 // Load the object's klass 2982 Node* obj_klass = load_object_klass(not_null_obj); 2983 2984 // Generate the subtype check 2985 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2986 2987 // Plug in the success path to the general merge in slot 1. 2988 region->init_req(_obj_path, control()); 2989 phi ->init_req(_obj_path, intcon(1)); 2990 2991 // Plug in the failing path to the general merge in slot 2. 2992 region->init_req(_fail_path, not_subtype_ctrl); 2993 phi ->init_req(_fail_path, intcon(0)); 2994 2995 // Return final merged results 2996 set_control( _gvn.transform(region) ); 2997 record_for_igvn(region); 2998 return _gvn.transform(phi); 2999 } 3000 3001 //-------------------------------gen_checkcast--------------------------------- 3002 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3003 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3004 // uncommon-trap paths work. Adjust stack after this call. 3005 // If failure_control is supplied and not null, it is filled in with 3006 // the control edge for the cast failure. Otherwise, an appropriate 3007 // uncommon trap or exception is thrown. 3008 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3009 Node* *failure_control) { 3010 kill_dead_locals(); // Benefit all the uncommon traps 3011 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3012 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3013 3014 // Fast cutout: Check the case that the cast is vacuously true. 3015 // This detects the common cases where the test will short-circuit 3016 // away completely. We do this before we perform the null check, 3017 // because if the test is going to turn into zero code, we don't 3018 // want a residual null check left around. (Causes a slowdown, 3019 // for example, in some objArray manipulations, such as a[i]=a[j].) 3020 if (tk->singleton()) { 3021 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3022 if (objtp != NULL && objtp->klass() != NULL) { 3023 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3024 case Compile::SSC_always_true: 3025 // If we know the type check always succeed then we don't use 3026 // the profiling data at this bytecode. Don't lose it, feed it 3027 // to the type system as a speculative type. 3028 return record_profiled_receiver_for_speculation(obj); 3029 case Compile::SSC_always_false: 3030 // It needs a null check because a null will *pass* the cast check. 3031 // A non-null value will always produce an exception. 3032 return null_assert(obj); 3033 } 3034 } 3035 } 3036 3037 ciProfileData* data = NULL; 3038 bool safe_for_replace = false; 3039 if (failure_control == NULL) { // use MDO in regular case only 3040 assert(java_bc() == Bytecodes::_aastore || 3041 java_bc() == Bytecodes::_checkcast, 3042 "interpreter profiles type checks only for these BCs"); 3043 data = method()->method_data()->bci_to_data(bci()); 3044 safe_for_replace = true; 3045 } 3046 3047 // Make the merge point 3048 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3049 RegionNode* region = new RegionNode(PATH_LIMIT); 3050 Node* phi = new PhiNode(region, toop); 3051 C->set_has_split_ifs(true); // Has chance for split-if optimization 3052 3053 // Use null-cast information if it is available 3054 bool speculative_not_null = false; 3055 bool never_see_null = ((failure_control == NULL) // regular case only 3056 && seems_never_null(obj, data, speculative_not_null)); 3057 3058 // Null check; get casted pointer; set region slot 3 3059 Node* null_ctl = top(); 3060 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3061 3062 // If not_null_obj is dead, only null-path is taken 3063 if (stopped()) { // Doing instance-of on a NULL? 3064 set_control(null_ctl); 3065 return null(); 3066 } 3067 region->init_req(_null_path, null_ctl); 3068 phi ->init_req(_null_path, null()); // Set null path value 3069 if (null_ctl == top()) { 3070 // Do this eagerly, so that pattern matches like is_diamond_phi 3071 // will work even during parsing. 3072 assert(_null_path == PATH_LIMIT-1, "delete last"); 3073 region->del_req(_null_path); 3074 phi ->del_req(_null_path); 3075 } 3076 3077 Node* cast_obj = NULL; 3078 if (tk->klass_is_exact()) { 3079 // The following optimization tries to statically cast the speculative type of the object 3080 // (for example obtained during profiling) to the type of the superklass and then do a 3081 // dynamic check that the type of the object is what we expect. To work correctly 3082 // for checkcast and aastore the type of superklass should be exact. 3083 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3084 // We may not have profiling here or it may not help us. If we have 3085 // a speculative type use it to perform an exact cast. 3086 ciKlass* spec_obj_type = obj_type->speculative_type(); 3087 if (spec_obj_type != NULL || data != NULL) { 3088 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3089 if (cast_obj != NULL) { 3090 if (failure_control != NULL) // failure is now impossible 3091 (*failure_control) = top(); 3092 // adjust the type of the phi to the exact klass: 3093 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3094 } 3095 } 3096 } 3097 3098 if (cast_obj == NULL) { 3099 // Load the object's klass 3100 Node* obj_klass = load_object_klass(not_null_obj); 3101 3102 // Generate the subtype check 3103 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3104 3105 // Plug in success path into the merge 3106 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3107 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3108 if (failure_control == NULL) { 3109 if (not_subtype_ctrl != top()) { // If failure is possible 3110 PreserveJVMState pjvms(this); 3111 set_control(not_subtype_ctrl); 3112 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3113 } 3114 } else { 3115 (*failure_control) = not_subtype_ctrl; 3116 } 3117 } 3118 3119 region->init_req(_obj_path, control()); 3120 phi ->init_req(_obj_path, cast_obj); 3121 3122 // A merge of NULL or Casted-NotNull obj 3123 Node* res = _gvn.transform(phi); 3124 3125 // Note I do NOT always 'replace_in_map(obj,result)' here. 3126 // if( tk->klass()->can_be_primary_super() ) 3127 // This means that if I successfully store an Object into an array-of-String 3128 // I 'forget' that the Object is really now known to be a String. I have to 3129 // do this because we don't have true union types for interfaces - if I store 3130 // a Baz into an array-of-Interface and then tell the optimizer it's an 3131 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3132 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3133 // replace_in_map( obj, res ); 3134 3135 // Return final merged results 3136 set_control( _gvn.transform(region) ); 3137 record_for_igvn(region); 3138 return res; 3139 } 3140 3141 //------------------------------next_monitor----------------------------------- 3142 // What number should be given to the next monitor? 3143 int GraphKit::next_monitor() { 3144 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3145 int next = current + C->sync_stack_slots(); 3146 // Keep the toplevel high water mark current: 3147 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3148 return current; 3149 } 3150 3151 //------------------------------insert_mem_bar--------------------------------- 3152 // Memory barrier to avoid floating things around 3153 // The membar serves as a pinch point between both control and all memory slices. 3154 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3155 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3156 mb->init_req(TypeFunc::Control, control()); 3157 mb->init_req(TypeFunc::Memory, reset_memory()); 3158 Node* membar = _gvn.transform(mb); 3159 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3160 set_all_memory_call(membar); 3161 return membar; 3162 } 3163 3164 //-------------------------insert_mem_bar_volatile---------------------------- 3165 // Memory barrier to avoid floating things around 3166 // The membar serves as a pinch point between both control and memory(alias_idx). 3167 // If you want to make a pinch point on all memory slices, do not use this 3168 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3169 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3170 // When Parse::do_put_xxx updates a volatile field, it appends a series 3171 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3172 // The first membar is on the same memory slice as the field store opcode. 3173 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3174 // All the other membars (for other volatile slices, including AliasIdxBot, 3175 // which stands for all unknown volatile slices) are control-dependent 3176 // on the first membar. This prevents later volatile loads or stores 3177 // from sliding up past the just-emitted store. 3178 3179 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3180 mb->set_req(TypeFunc::Control,control()); 3181 if (alias_idx == Compile::AliasIdxBot) { 3182 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3183 } else { 3184 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3185 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3186 } 3187 Node* membar = _gvn.transform(mb); 3188 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3189 if (alias_idx == Compile::AliasIdxBot) { 3190 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3191 } else { 3192 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3193 } 3194 return membar; 3195 } 3196 3197 void GraphKit::insert_store_load_for_barrier() { 3198 Node* mem = reset_memory(); 3199 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxBot); 3200 mb->init_req(TypeFunc::Control, control()); 3201 mb->init_req(TypeFunc::Memory, mem); 3202 Node* membar = _gvn.transform(mb); 3203 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3204 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory)); 3205 set_all_memory(mem); 3206 set_memory(newmem, Compile::AliasIdxRaw); 3207 } 3208 3209 3210 //------------------------------shared_lock------------------------------------ 3211 // Emit locking code. 3212 FastLockNode* GraphKit::shared_lock(Node* obj) { 3213 // bci is either a monitorenter bc or InvocationEntryBci 3214 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3215 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3216 3217 if( !GenerateSynchronizationCode ) 3218 return NULL; // Not locking things? 3219 if (stopped()) // Dead monitor? 3220 return NULL; 3221 3222 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3223 3224 // Box the stack location 3225 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3226 Node* mem = reset_memory(); 3227 3228 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3229 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3230 // Create the counters for this fast lock. 3231 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3232 } 3233 3234 // Create the rtm counters for this fast lock if needed. 3235 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3236 3237 // Add monitor to debug info for the slow path. If we block inside the 3238 // slow path and de-opt, we need the monitor hanging around 3239 map()->push_monitor( flock ); 3240 3241 const TypeFunc *tf = LockNode::lock_type(); 3242 LockNode *lock = new LockNode(C, tf); 3243 3244 lock->init_req( TypeFunc::Control, control() ); 3245 lock->init_req( TypeFunc::Memory , mem ); 3246 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3247 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3248 lock->init_req( TypeFunc::ReturnAdr, top() ); 3249 3250 lock->init_req(TypeFunc::Parms + 0, obj); 3251 lock->init_req(TypeFunc::Parms + 1, box); 3252 lock->init_req(TypeFunc::Parms + 2, flock); 3253 add_safepoint_edges(lock); 3254 3255 lock = _gvn.transform( lock )->as_Lock(); 3256 3257 // lock has no side-effects, sets few values 3258 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3259 3260 insert_mem_bar(Op_MemBarAcquireLock); 3261 3262 // Add this to the worklist so that the lock can be eliminated 3263 record_for_igvn(lock); 3264 3265 #ifndef PRODUCT 3266 if (PrintLockStatistics) { 3267 // Update the counter for this lock. Don't bother using an atomic 3268 // operation since we don't require absolute accuracy. 3269 lock->create_lock_counter(map()->jvms()); 3270 increment_counter(lock->counter()->addr()); 3271 } 3272 #endif 3273 3274 return flock; 3275 } 3276 3277 3278 //------------------------------shared_unlock---------------------------------- 3279 // Emit unlocking code. 3280 void GraphKit::shared_unlock(Node* box, Node* obj) { 3281 // bci is either a monitorenter bc or InvocationEntryBci 3282 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3283 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3284 3285 if( !GenerateSynchronizationCode ) 3286 return; 3287 if (stopped()) { // Dead monitor? 3288 map()->pop_monitor(); // Kill monitor from debug info 3289 return; 3290 } 3291 3292 // Memory barrier to avoid floating things down past the locked region 3293 insert_mem_bar(Op_MemBarReleaseLock); 3294 3295 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3296 UnlockNode *unlock = new UnlockNode(C, tf); 3297 #ifdef ASSERT 3298 unlock->set_dbg_jvms(sync_jvms()); 3299 #endif 3300 uint raw_idx = Compile::AliasIdxRaw; 3301 unlock->init_req( TypeFunc::Control, control() ); 3302 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3303 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3304 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3305 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3306 3307 unlock->init_req(TypeFunc::Parms + 0, obj); 3308 unlock->init_req(TypeFunc::Parms + 1, box); 3309 unlock = _gvn.transform(unlock)->as_Unlock(); 3310 3311 Node* mem = reset_memory(); 3312 3313 // unlock has no side-effects, sets few values 3314 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3315 3316 // Kill monitor from debug info 3317 map()->pop_monitor( ); 3318 } 3319 3320 //-------------------------------get_layout_helper----------------------------- 3321 // If the given klass is a constant or known to be an array, 3322 // fetch the constant layout helper value into constant_value 3323 // and return (Node*)NULL. Otherwise, load the non-constant 3324 // layout helper value, and return the node which represents it. 3325 // This two-faced routine is useful because allocation sites 3326 // almost always feature constant types. 3327 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3328 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3329 if (!StressReflectiveCode && inst_klass != NULL) { 3330 ciKlass* klass = inst_klass->klass(); 3331 bool xklass = inst_klass->klass_is_exact(); 3332 if (xklass || klass->is_array_klass()) { 3333 jint lhelper = klass->layout_helper(); 3334 if (lhelper != Klass::_lh_neutral_value) { 3335 constant_value = lhelper; 3336 return (Node*) NULL; 3337 } 3338 } 3339 } 3340 constant_value = Klass::_lh_neutral_value; // put in a known value 3341 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3342 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3343 } 3344 3345 // We just put in an allocate/initialize with a big raw-memory effect. 3346 // Hook selected additional alias categories on the initialization. 3347 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3348 MergeMemNode* init_in_merge, 3349 Node* init_out_raw) { 3350 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3351 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3352 3353 Node* prevmem = kit.memory(alias_idx); 3354 init_in_merge->set_memory_at(alias_idx, prevmem); 3355 kit.set_memory(init_out_raw, alias_idx); 3356 } 3357 3358 //---------------------------set_output_for_allocation------------------------- 3359 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3360 const TypeOopPtr* oop_type, 3361 bool deoptimize_on_exception) { 3362 int rawidx = Compile::AliasIdxRaw; 3363 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3364 add_safepoint_edges(alloc); 3365 Node* allocx = _gvn.transform(alloc); 3366 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3367 // create memory projection for i_o 3368 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3369 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3370 3371 // create a memory projection as for the normal control path 3372 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3373 set_memory(malloc, rawidx); 3374 3375 // a normal slow-call doesn't change i_o, but an allocation does 3376 // we create a separate i_o projection for the normal control path 3377 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3378 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3379 3380 // put in an initialization barrier 3381 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3382 rawoop)->as_Initialize(); 3383 assert(alloc->initialization() == init, "2-way macro link must work"); 3384 assert(init ->allocation() == alloc, "2-way macro link must work"); 3385 { 3386 // Extract memory strands which may participate in the new object's 3387 // initialization, and source them from the new InitializeNode. 3388 // This will allow us to observe initializations when they occur, 3389 // and link them properly (as a group) to the InitializeNode. 3390 assert(init->in(InitializeNode::Memory) == malloc, ""); 3391 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3392 init->set_req(InitializeNode::Memory, minit_in); 3393 record_for_igvn(minit_in); // fold it up later, if possible 3394 Node* minit_out = memory(rawidx); 3395 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3396 if (oop_type->isa_aryptr()) { 3397 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3398 int elemidx = C->get_alias_index(telemref); 3399 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3400 } else if (oop_type->isa_instptr()) { 3401 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3402 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3403 ciField* field = ik->nonstatic_field_at(i); 3404 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3405 continue; // do not bother to track really large numbers of fields 3406 // Find (or create) the alias category for this field: 3407 int fieldidx = C->alias_type(field)->index(); 3408 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3409 } 3410 } 3411 } 3412 3413 // Cast raw oop to the real thing... 3414 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3415 javaoop = _gvn.transform(javaoop); 3416 C->set_recent_alloc(control(), javaoop); 3417 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3418 3419 #ifdef ASSERT 3420 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3421 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3422 "Ideal_allocation works"); 3423 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3424 "Ideal_allocation works"); 3425 if (alloc->is_AllocateArray()) { 3426 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3427 "Ideal_allocation works"); 3428 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3429 "Ideal_allocation works"); 3430 } else { 3431 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3432 } 3433 } 3434 #endif //ASSERT 3435 3436 return javaoop; 3437 } 3438 3439 //---------------------------new_instance-------------------------------------- 3440 // This routine takes a klass_node which may be constant (for a static type) 3441 // or may be non-constant (for reflective code). It will work equally well 3442 // for either, and the graph will fold nicely if the optimizer later reduces 3443 // the type to a constant. 3444 // The optional arguments are for specialized use by intrinsics: 3445 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3446 // - If 'return_size_val', report the the total object size to the caller. 3447 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3448 Node* GraphKit::new_instance(Node* klass_node, 3449 Node* extra_slow_test, 3450 Node* *return_size_val, 3451 bool deoptimize_on_exception) { 3452 // Compute size in doublewords 3453 // The size is always an integral number of doublewords, represented 3454 // as a positive bytewise size stored in the klass's layout_helper. 3455 // The layout_helper also encodes (in a low bit) the need for a slow path. 3456 jint layout_con = Klass::_lh_neutral_value; 3457 Node* layout_val = get_layout_helper(klass_node, layout_con); 3458 int layout_is_con = (layout_val == NULL); 3459 3460 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3461 // Generate the initial go-slow test. It's either ALWAYS (return a 3462 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3463 // case) a computed value derived from the layout_helper. 3464 Node* initial_slow_test = NULL; 3465 if (layout_is_con) { 3466 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3467 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3468 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3469 } else { // reflective case 3470 // This reflective path is used by Unsafe.allocateInstance. 3471 // (It may be stress-tested by specifying StressReflectiveCode.) 3472 // Basically, we want to get into the VM is there's an illegal argument. 3473 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3474 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3475 if (extra_slow_test != intcon(0)) { 3476 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3477 } 3478 // (Macro-expander will further convert this to a Bool, if necessary.) 3479 } 3480 3481 // Find the size in bytes. This is easy; it's the layout_helper. 3482 // The size value must be valid even if the slow path is taken. 3483 Node* size = NULL; 3484 if (layout_is_con) { 3485 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3486 } else { // reflective case 3487 // This reflective path is used by clone and Unsafe.allocateInstance. 3488 size = ConvI2X(layout_val); 3489 3490 // Clear the low bits to extract layout_helper_size_in_bytes: 3491 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3492 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3493 size = _gvn.transform( new AndXNode(size, mask) ); 3494 } 3495 if (return_size_val != NULL) { 3496 (*return_size_val) = size; 3497 } 3498 3499 // This is a precise notnull oop of the klass. 3500 // (Actually, it need not be precise if this is a reflective allocation.) 3501 // It's what we cast the result to. 3502 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3503 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3504 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3505 3506 // Now generate allocation code 3507 3508 // The entire memory state is needed for slow path of the allocation 3509 // since GC and deoptimization can happen. 3510 Node *mem = reset_memory(); 3511 set_all_memory(mem); // Create new memory state 3512 3513 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3514 control(), mem, i_o(), 3515 size, klass_node, 3516 initial_slow_test); 3517 3518 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3519 } 3520 3521 //-------------------------------new_array------------------------------------- 3522 // helper for both newarray and anewarray 3523 // The 'length' parameter is (obviously) the length of the array. 3524 // See comments on new_instance for the meaning of the other arguments. 3525 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3526 Node* length, // number of array elements 3527 int nargs, // number of arguments to push back for uncommon trap 3528 Node* *return_size_val, 3529 bool deoptimize_on_exception) { 3530 jint layout_con = Klass::_lh_neutral_value; 3531 Node* layout_val = get_layout_helper(klass_node, layout_con); 3532 int layout_is_con = (layout_val == NULL); 3533 3534 if (!layout_is_con && !StressReflectiveCode && 3535 !too_many_traps(Deoptimization::Reason_class_check)) { 3536 // This is a reflective array creation site. 3537 // Optimistically assume that it is a subtype of Object[], 3538 // so that we can fold up all the address arithmetic. 3539 layout_con = Klass::array_layout_helper(T_OBJECT); 3540 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3541 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3542 { BuildCutout unless(this, bol_lh, PROB_MAX); 3543 inc_sp(nargs); 3544 uncommon_trap(Deoptimization::Reason_class_check, 3545 Deoptimization::Action_maybe_recompile); 3546 } 3547 layout_val = NULL; 3548 layout_is_con = true; 3549 } 3550 3551 // Generate the initial go-slow test. Make sure we do not overflow 3552 // if length is huge (near 2Gig) or negative! We do not need 3553 // exact double-words here, just a close approximation of needed 3554 // double-words. We can't add any offset or rounding bits, lest we 3555 // take a size -1 of bytes and make it positive. Use an unsigned 3556 // compare, so negative sizes look hugely positive. 3557 int fast_size_limit = FastAllocateSizeLimit; 3558 if (layout_is_con) { 3559 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3560 // Increase the size limit if we have exact knowledge of array type. 3561 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3562 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3563 } 3564 3565 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3566 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3567 3568 // --- Size Computation --- 3569 // array_size = round_to_heap(array_header + (length << elem_shift)); 3570 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3571 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3572 // The rounding mask is strength-reduced, if possible. 3573 int round_mask = MinObjAlignmentInBytes - 1; 3574 Node* header_size = NULL; 3575 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3576 // (T_BYTE has the weakest alignment and size restrictions...) 3577 if (layout_is_con) { 3578 int hsize = Klass::layout_helper_header_size(layout_con); 3579 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3580 BasicType etype = Klass::layout_helper_element_type(layout_con); 3581 if ((round_mask & ~right_n_bits(eshift)) == 0) 3582 round_mask = 0; // strength-reduce it if it goes away completely 3583 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3584 assert(header_size_min <= hsize, "generic minimum is smallest"); 3585 header_size_min = hsize; 3586 header_size = intcon(hsize + round_mask); 3587 } else { 3588 Node* hss = intcon(Klass::_lh_header_size_shift); 3589 Node* hsm = intcon(Klass::_lh_header_size_mask); 3590 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3591 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3592 Node* mask = intcon(round_mask); 3593 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3594 } 3595 3596 Node* elem_shift = NULL; 3597 if (layout_is_con) { 3598 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3599 if (eshift != 0) 3600 elem_shift = intcon(eshift); 3601 } else { 3602 // There is no need to mask or shift this value. 3603 // The semantics of LShiftINode include an implicit mask to 0x1F. 3604 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3605 elem_shift = layout_val; 3606 } 3607 3608 // Transition to native address size for all offset calculations: 3609 Node* lengthx = ConvI2X(length); 3610 Node* headerx = ConvI2X(header_size); 3611 #ifdef _LP64 3612 { const TypeInt* tilen = _gvn.find_int_type(length); 3613 if (tilen != NULL && tilen->_lo < 0) { 3614 // Add a manual constraint to a positive range. Cf. array_element_address. 3615 jint size_max = fast_size_limit; 3616 if (size_max > tilen->_hi) size_max = tilen->_hi; 3617 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3618 3619 // Only do a narrow I2L conversion if the range check passed. 3620 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3621 _gvn.transform(iff); 3622 RegionNode* region = new RegionNode(3); 3623 _gvn.set_type(region, Type::CONTROL); 3624 lengthx = new PhiNode(region, TypeLong::LONG); 3625 _gvn.set_type(lengthx, TypeLong::LONG); 3626 3627 // Range check passed. Use ConvI2L node with narrow type. 3628 Node* passed = IfFalse(iff); 3629 region->init_req(1, passed); 3630 // Make I2L conversion control dependent to prevent it from 3631 // floating above the range check during loop optimizations. 3632 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3633 3634 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3635 region->init_req(2, IfTrue(iff)); 3636 lengthx->init_req(2, ConvI2X(length)); 3637 3638 set_control(region); 3639 record_for_igvn(region); 3640 record_for_igvn(lengthx); 3641 } 3642 } 3643 #endif 3644 3645 // Combine header size (plus rounding) and body size. Then round down. 3646 // This computation cannot overflow, because it is used only in two 3647 // places, one where the length is sharply limited, and the other 3648 // after a successful allocation. 3649 Node* abody = lengthx; 3650 if (elem_shift != NULL) 3651 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3652 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3653 if (round_mask != 0) { 3654 Node* mask = MakeConX(~round_mask); 3655 size = _gvn.transform( new AndXNode(size, mask) ); 3656 } 3657 // else if round_mask == 0, the size computation is self-rounding 3658 3659 if (return_size_val != NULL) { 3660 // This is the size 3661 (*return_size_val) = size; 3662 } 3663 3664 // Now generate allocation code 3665 3666 // The entire memory state is needed for slow path of the allocation 3667 // since GC and deoptimization can happened. 3668 Node *mem = reset_memory(); 3669 set_all_memory(mem); // Create new memory state 3670 3671 if (initial_slow_test->is_Bool()) { 3672 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3673 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3674 } 3675 3676 // Create the AllocateArrayNode and its result projections 3677 AllocateArrayNode* alloc 3678 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3679 control(), mem, i_o(), 3680 size, klass_node, 3681 initial_slow_test, 3682 length); 3683 3684 // Cast to correct type. Note that the klass_node may be constant or not, 3685 // and in the latter case the actual array type will be inexact also. 3686 // (This happens via a non-constant argument to inline_native_newArray.) 3687 // In any case, the value of klass_node provides the desired array type. 3688 const TypeInt* length_type = _gvn.find_int_type(length); 3689 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3690 if (ary_type->isa_aryptr() && length_type != NULL) { 3691 // Try to get a better type than POS for the size 3692 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3693 } 3694 3695 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3696 3697 // Cast length on remaining path to be as narrow as possible 3698 if (map()->find_edge(length) >= 0) { 3699 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3700 if (ccast != length) { 3701 _gvn.set_type_bottom(ccast); 3702 record_for_igvn(ccast); 3703 replace_in_map(length, ccast); 3704 } 3705 } 3706 3707 return javaoop; 3708 } 3709 3710 // The following "Ideal_foo" functions are placed here because they recognize 3711 // the graph shapes created by the functions immediately above. 3712 3713 //---------------------------Ideal_allocation---------------------------------- 3714 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3715 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3716 if (ptr == NULL) { // reduce dumb test in callers 3717 return NULL; 3718 } 3719 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3720 ptr = ptr->in(1); 3721 if (ptr == NULL) return NULL; 3722 } 3723 // Return NULL for allocations with several casts: 3724 // j.l.reflect.Array.newInstance(jobject, jint) 3725 // Object.clone() 3726 // to keep more precise type from last cast. 3727 if (ptr->is_Proj()) { 3728 Node* allo = ptr->in(0); 3729 if (allo != NULL && allo->is_Allocate()) { 3730 return allo->as_Allocate(); 3731 } 3732 } 3733 // Report failure to match. 3734 return NULL; 3735 } 3736 3737 // Fancy version which also strips off an offset (and reports it to caller). 3738 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3739 intptr_t& offset) { 3740 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3741 if (base == NULL) return NULL; 3742 return Ideal_allocation(base, phase); 3743 } 3744 3745 // Trace Initialize <- Proj[Parm] <- Allocate 3746 AllocateNode* InitializeNode::allocation() { 3747 Node* rawoop = in(InitializeNode::RawAddress); 3748 if (rawoop->is_Proj()) { 3749 Node* alloc = rawoop->in(0); 3750 if (alloc->is_Allocate()) { 3751 return alloc->as_Allocate(); 3752 } 3753 } 3754 return NULL; 3755 } 3756 3757 // Trace Allocate -> Proj[Parm] -> Initialize 3758 InitializeNode* AllocateNode::initialization() { 3759 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3760 if (rawoop == NULL) return NULL; 3761 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3762 Node* init = rawoop->fast_out(i); 3763 if (init->is_Initialize()) { 3764 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3765 return init->as_Initialize(); 3766 } 3767 } 3768 return NULL; 3769 } 3770 3771 //----------------------------- loop predicates --------------------------- 3772 3773 //------------------------------add_predicate_impl---------------------------- 3774 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3775 // Too many traps seen? 3776 if (too_many_traps(reason)) { 3777 #ifdef ASSERT 3778 if (TraceLoopPredicate) { 3779 int tc = C->trap_count(reason); 3780 tty->print("too many traps=%s tcount=%d in ", 3781 Deoptimization::trap_reason_name(reason), tc); 3782 method()->print(); // which method has too many predicate traps 3783 tty->cr(); 3784 } 3785 #endif 3786 // We cannot afford to take more traps here, 3787 // do not generate predicate. 3788 return; 3789 } 3790 3791 Node *cont = _gvn.intcon(1); 3792 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3793 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3794 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3795 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3796 C->add_predicate_opaq(opq); 3797 { 3798 PreserveJVMState pjvms(this); 3799 set_control(iffalse); 3800 inc_sp(nargs); 3801 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3802 } 3803 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3804 set_control(iftrue); 3805 } 3806 3807 //------------------------------add_predicate--------------------------------- 3808 void GraphKit::add_predicate(int nargs) { 3809 if (UseLoopPredicate) { 3810 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3811 } 3812 // loop's limit check predicate should be near the loop. 3813 if (LoopLimitCheck) { 3814 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3815 } 3816 } 3817 3818 //----------------------------- store barriers ---------------------------- 3819 #define __ ideal. 3820 3821 void GraphKit::sync_kit(IdealKit& ideal) { 3822 set_all_memory(__ merged_memory()); 3823 set_i_o(__ i_o()); 3824 set_control(__ ctrl()); 3825 } 3826 3827 void GraphKit::final_sync(IdealKit& ideal) { 3828 // Final sync IdealKit and graphKit. 3829 sync_kit(ideal); 3830 } 3831 3832 Node* GraphKit::byte_map_base_node() { 3833 // Get base of card map 3834 CardTableModRefBS* ct = 3835 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set()); 3836 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code"); 3837 if (ct->byte_map_base != NULL) { 3838 return makecon(TypeRawPtr::make((address)ct->byte_map_base)); 3839 } else { 3840 return null(); 3841 } 3842 } 3843 3844 // vanilla/CMS post barrier 3845 // Insert a write-barrier store. This is to let generational GC work; we have 3846 // to flag all oop-stores before the next GC point. 3847 void GraphKit::write_barrier_post(Node* oop_store, 3848 Node* obj, 3849 Node* adr, 3850 uint adr_idx, 3851 Node* val, 3852 bool use_precise) { 3853 // No store check needed if we're storing a NULL or an old object 3854 // (latter case is probably a string constant). The concurrent 3855 // mark sweep garbage collector, however, needs to have all nonNull 3856 // oop updates flagged via card-marks. 3857 if (val != NULL && val->is_Con()) { 3858 // must be either an oop or NULL 3859 const Type* t = val->bottom_type(); 3860 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3861 // stores of null never (?) need barriers 3862 return; 3863 } 3864 3865 if (use_ReduceInitialCardMarks() 3866 && obj == just_allocated_object(control())) { 3867 // We can skip marks on a freshly-allocated object in Eden. 3868 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3869 // That routine informs GC to take appropriate compensating steps, 3870 // upon a slow-path allocation, so as to make this card-mark 3871 // elision safe. 3872 return; 3873 } 3874 3875 if (!use_precise) { 3876 // All card marks for a (non-array) instance are in one place: 3877 adr = obj; 3878 } 3879 // (Else it's an array (or unknown), and we want more precise card marks.) 3880 assert(adr != NULL, ""); 3881 3882 IdealKit ideal(this, true); 3883 3884 // Convert the pointer to an int prior to doing math on it 3885 Node* cast = __ CastPX(__ ctrl(), adr); 3886 3887 // Divide by card size 3888 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef), 3889 "Only one we handle so far."); 3890 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3891 3892 // Combine card table base and card offset 3893 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3894 3895 // Get the alias_index for raw card-mark memory 3896 int adr_type = Compile::AliasIdxRaw; 3897 Node* zero = __ ConI(0); // Dirty card value 3898 BasicType bt = T_BYTE; 3899 3900 if (UseConcMarkSweepGC && UseCondCardMark) { 3901 insert_store_load_for_barrier(); 3902 __ sync_kit(this); 3903 } 3904 3905 if (UseCondCardMark) { 3906 // The classic GC reference write barrier is typically implemented 3907 // as a store into the global card mark table. Unfortunately 3908 // unconditional stores can result in false sharing and excessive 3909 // coherence traffic as well as false transactional aborts. 3910 // UseCondCardMark enables MP "polite" conditional card mark 3911 // stores. In theory we could relax the load from ctrl() to 3912 // no_ctrl, but that doesn't buy much latitude. 3913 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3914 __ if_then(card_val, BoolTest::ne, zero); 3915 } 3916 3917 // Smash zero into card 3918 if( !UseConcMarkSweepGC ) { 3919 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered); 3920 } else { 3921 // Specialized path for CM store barrier 3922 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3923 } 3924 3925 if (UseCondCardMark) { 3926 __ end_if(); 3927 } 3928 3929 // Final sync IdealKit and GraphKit. 3930 final_sync(ideal); 3931 } 3932 /* 3933 * Determine if the G1 pre-barrier can be removed. The pre-barrier is 3934 * required by SATB to make sure all objects live at the start of the 3935 * marking are kept alive, all reference updates need to any previous 3936 * reference stored before writing. 3937 * 3938 * If the previous value is NULL there is no need to save the old value. 3939 * References that are NULL are filtered during runtime by the barrier 3940 * code to avoid unnecessary queuing. 3941 * 3942 * However in the case of newly allocated objects it might be possible to 3943 * prove that the reference about to be overwritten is NULL during compile 3944 * time and avoid adding the barrier code completely. 3945 * 3946 * The compiler needs to determine that the object in which a field is about 3947 * to be written is newly allocated, and that no prior store to the same field 3948 * has happened since the allocation. 3949 * 3950 * Returns true if the pre-barrier can be removed 3951 */ 3952 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr, 3953 BasicType bt, uint adr_idx) { 3954 intptr_t offset = 0; 3955 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 3956 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 3957 3958 if (offset == Type::OffsetBot) { 3959 return false; // cannot unalias unless there are precise offsets 3960 } 3961 3962 if (alloc == NULL) { 3963 return false; // No allocation found 3964 } 3965 3966 intptr_t size_in_bytes = type2aelembytes(bt); 3967 3968 Node* mem = memory(adr_idx); // start searching here... 3969 3970 for (int cnt = 0; cnt < 50; cnt++) { 3971 3972 if (mem->is_Store()) { 3973 3974 Node* st_adr = mem->in(MemNode::Address); 3975 intptr_t st_offset = 0; 3976 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); 3977 3978 if (st_base == NULL) { 3979 break; // inscrutable pointer 3980 } 3981 3982 // Break we have found a store with same base and offset as ours so break 3983 if (st_base == base && st_offset == offset) { 3984 break; 3985 } 3986 3987 if (st_offset != offset && st_offset != Type::OffsetBot) { 3988 const int MAX_STORE = BytesPerLong; 3989 if (st_offset >= offset + size_in_bytes || 3990 st_offset <= offset - MAX_STORE || 3991 st_offset <= offset - mem->as_Store()->memory_size()) { 3992 // Success: The offsets are provably independent. 3993 // (You may ask, why not just test st_offset != offset and be done? 3994 // The answer is that stores of different sizes can co-exist 3995 // in the same sequence of RawMem effects. We sometimes initialize 3996 // a whole 'tile' of array elements with a single jint or jlong.) 3997 mem = mem->in(MemNode::Memory); 3998 continue; // advance through independent store memory 3999 } 4000 } 4001 4002 if (st_base != base 4003 && MemNode::detect_ptr_independence(base, alloc, st_base, 4004 AllocateNode::Ideal_allocation(st_base, phase), 4005 phase)) { 4006 // Success: The bases are provably independent. 4007 mem = mem->in(MemNode::Memory); 4008 continue; // advance through independent store memory 4009 } 4010 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4011 4012 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4013 AllocateNode* st_alloc = st_init->allocation(); 4014 4015 // Make sure that we are looking at the same allocation site. 4016 // The alloc variable is guaranteed to not be null here from earlier check. 4017 if (alloc == st_alloc) { 4018 // Check that the initialization is storing NULL so that no previous store 4019 // has been moved up and directly write a reference 4020 Node* captured_store = st_init->find_captured_store(offset, 4021 type2aelembytes(T_OBJECT), 4022 phase); 4023 if (captured_store == NULL || captured_store == st_init->zero_memory()) { 4024 return true; 4025 } 4026 } 4027 } 4028 4029 // Unless there is an explicit 'continue', we must bail out here, 4030 // because 'mem' is an inscrutable memory state (e.g., a call). 4031 break; 4032 } 4033 4034 return false; 4035 } 4036 4037 // G1 pre/post barriers 4038 void GraphKit::g1_write_barrier_pre(bool do_load, 4039 Node* obj, 4040 Node* adr, 4041 uint alias_idx, 4042 Node* val, 4043 const TypeOopPtr* val_type, 4044 Node* pre_val, 4045 BasicType bt) { 4046 4047 // Some sanity checks 4048 // Note: val is unused in this routine. 4049 4050 if (do_load) { 4051 // We need to generate the load of the previous value 4052 assert(obj != NULL, "must have a base"); 4053 assert(adr != NULL, "where are loading from?"); 4054 assert(pre_val == NULL, "loaded already?"); 4055 assert(val_type != NULL, "need a type"); 4056 4057 if (use_ReduceInitialCardMarks() 4058 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) { 4059 return; 4060 } 4061 4062 } else { 4063 // In this case both val_type and alias_idx are unused. 4064 assert(pre_val != NULL, "must be loaded already"); 4065 // Nothing to be done if pre_val is null. 4066 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 4067 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 4068 } 4069 assert(bt == T_OBJECT || bt == T_VALUETYPE, "or we shouldn't be here"); 4070 4071 IdealKit ideal(this, true); 4072 4073 Node* tls = __ thread(); // ThreadLocalStorage 4074 4075 Node* no_ctrl = NULL; 4076 Node* no_base = __ top(); 4077 Node* zero = __ ConI(0); 4078 Node* zeroX = __ ConX(0); 4079 4080 float likely = PROB_LIKELY(0.999); 4081 float unlikely = PROB_UNLIKELY(0.999); 4082 4083 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 4084 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width"); 4085 4086 // Offsets into the thread 4087 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 4088 SATBMarkQueue::byte_offset_of_active()); 4089 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 4090 SATBMarkQueue::byte_offset_of_index()); 4091 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 4092 SATBMarkQueue::byte_offset_of_buf()); 4093 4094 // Now the actual pointers into the thread 4095 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 4096 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4097 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4098 4099 // Now some of the values 4100 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 4101 4102 // if (!marking) 4103 __ if_then(marking, BoolTest::ne, zero, unlikely); { 4104 BasicType index_bt = TypeX_X->basic_type(); 4105 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size."); 4106 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 4107 4108 if (do_load) { 4109 // load original value 4110 // alias_idx correct?? 4111 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 4112 } 4113 4114 // if (pre_val != NULL) 4115 __ if_then(pre_val, BoolTest::ne, null()); { 4116 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4117 4118 // is the queue for this thread full? 4119 __ if_then(index, BoolTest::ne, zeroX, likely); { 4120 4121 // decrement the index 4122 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4123 4124 // Now get the buffer location we will log the previous value into and store it 4125 Node *log_addr = __ AddP(no_base, buffer, next_index); 4126 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 4127 // update the index 4128 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 4129 4130 } __ else_(); { 4131 4132 // logging buffer is full, call the runtime 4133 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 4134 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 4135 } __ end_if(); // (!index) 4136 } __ end_if(); // (pre_val != NULL) 4137 } __ end_if(); // (!marking) 4138 4139 // Final sync IdealKit and GraphKit. 4140 final_sync(ideal); 4141 } 4142 4143 /* 4144 * G1 similar to any GC with a Young Generation requires a way to keep track of 4145 * references from Old Generation to Young Generation to make sure all live 4146 * objects are found. G1 also requires to keep track of object references 4147 * between different regions to enable evacuation of old regions, which is done 4148 * as part of mixed collections. References are tracked in remembered sets and 4149 * is continuously updated as reference are written to with the help of the 4150 * post-barrier. 4151 * 4152 * To reduce the number of updates to the remembered set the post-barrier 4153 * filters updates to fields in objects located in the Young Generation, 4154 * the same region as the reference, when the NULL is being written or 4155 * if the card is already marked as dirty by an earlier write. 4156 * 4157 * Under certain circumstances it is possible to avoid generating the 4158 * post-barrier completely if it is possible during compile time to prove 4159 * the object is newly allocated and that no safepoint exists between the 4160 * allocation and the store. 4161 * 4162 * In the case of slow allocation the allocation code must handle the barrier 4163 * as part of the allocation in the case the allocated object is not located 4164 * in the nursery, this would happen for humongous objects. This is similar to 4165 * how CMS is required to handle this case, see the comments for the method 4166 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier. 4167 * A deferred card mark is required for these objects and handled in the above 4168 * mentioned methods. 4169 * 4170 * Returns true if the post barrier can be removed 4171 */ 4172 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store, 4173 Node* adr) { 4174 intptr_t offset = 0; 4175 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); 4176 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); 4177 4178 if (offset == Type::OffsetBot) { 4179 return false; // cannot unalias unless there are precise offsets 4180 } 4181 4182 if (alloc == NULL) { 4183 return false; // No allocation found 4184 } 4185 4186 // Start search from Store node 4187 Node* mem = store->in(MemNode::Control); 4188 if (mem->is_Proj() && mem->in(0)->is_Initialize()) { 4189 4190 InitializeNode* st_init = mem->in(0)->as_Initialize(); 4191 AllocateNode* st_alloc = st_init->allocation(); 4192 4193 // Make sure we are looking at the same allocation 4194 if (alloc == st_alloc) { 4195 return true; 4196 } 4197 } 4198 4199 return false; 4200 } 4201 4202 // 4203 // Update the card table and add card address to the queue 4204 // 4205 void GraphKit::g1_mark_card(IdealKit& ideal, 4206 Node* card_adr, 4207 Node* oop_store, 4208 uint oop_alias_idx, 4209 Node* index, 4210 Node* index_adr, 4211 Node* buffer, 4212 const TypeFunc* tf) { 4213 4214 Node* zero = __ ConI(0); 4215 Node* zeroX = __ ConX(0); 4216 Node* no_base = __ top(); 4217 BasicType card_bt = T_BYTE; 4218 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4219 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4220 4221 // Now do the queue work 4222 __ if_then(index, BoolTest::ne, zeroX); { 4223 4224 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4225 Node* log_addr = __ AddP(no_base, buffer, next_index); 4226 4227 // Order, see storeCM. 4228 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4229 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4230 4231 } __ else_(); { 4232 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4233 } __ end_if(); 4234 4235 } 4236 4237 void GraphKit::g1_write_barrier_post(Node* oop_store, 4238 Node* obj, 4239 Node* adr, 4240 uint alias_idx, 4241 Node* val, 4242 BasicType bt, 4243 bool use_precise) { 4244 // If we are writing a NULL then we need no post barrier 4245 4246 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4247 // Must be NULL 4248 const Type* t = val->bottom_type(); 4249 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4250 // No post barrier if writing NULLx 4251 return; 4252 } 4253 4254 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) { 4255 // We can skip marks on a freshly-allocated object in Eden. 4256 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 4257 // That routine informs GC to take appropriate compensating steps, 4258 // upon a slow-path allocation, so as to make this card-mark 4259 // elision safe. 4260 return; 4261 } 4262 4263 if (use_ReduceInitialCardMarks() 4264 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) { 4265 return; 4266 } 4267 4268 if (!use_precise) { 4269 // All card marks for a (non-array) instance are in one place: 4270 adr = obj; 4271 } 4272 // (Else it's an array (or unknown), and we want more precise card marks.) 4273 assert(adr != NULL, ""); 4274 4275 IdealKit ideal(this, true); 4276 4277 Node* tls = __ thread(); // ThreadLocalStorage 4278 4279 Node* no_base = __ top(); 4280 float likely = PROB_LIKELY(0.999); 4281 float unlikely = PROB_UNLIKELY(0.999); 4282 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4283 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4284 Node* zeroX = __ ConX(0); 4285 4286 // Get the alias_index for raw card-mark memory 4287 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4288 4289 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4290 4291 // Offsets into the thread 4292 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4293 DirtyCardQueue::byte_offset_of_index()); 4294 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4295 DirtyCardQueue::byte_offset_of_buf()); 4296 4297 // Pointers into the thread 4298 4299 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4300 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4301 4302 // Now some values 4303 // Use ctrl to avoid hoisting these values past a safepoint, which could 4304 // potentially reset these fields in the JavaThread. 4305 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4306 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4307 4308 // Convert the store obj pointer to an int prior to doing math on it 4309 // Must use ctrl to prevent "integerized oop" existing across safepoint 4310 Node* cast = __ CastPX(__ ctrl(), adr); 4311 4312 // Divide pointer by card size 4313 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4314 4315 // Combine card table base and card offset 4316 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4317 4318 // If we know the value being stored does it cross regions? 4319 4320 if (val != NULL) { 4321 // Does the store cause us to cross regions? 4322 4323 // Should be able to do an unsigned compare of region_size instead of 4324 // and extra shift. Do we have an unsigned compare?? 4325 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4326 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4327 4328 // if (xor_res == 0) same region so skip 4329 __ if_then(xor_res, BoolTest::ne, zeroX); { 4330 4331 // No barrier if we are storing a NULL 4332 __ if_then(val, BoolTest::ne, null(), unlikely); { 4333 4334 // Ok must mark the card if not already dirty 4335 4336 // load the original value of the card 4337 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4338 4339 __ if_then(card_val, BoolTest::ne, young_card); { 4340 sync_kit(ideal); 4341 insert_store_load_for_barrier(); 4342 __ sync_kit(this); 4343 4344 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4345 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4346 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4347 } __ end_if(); 4348 } __ end_if(); 4349 } __ end_if(); 4350 } __ end_if(); 4351 } else { 4352 // Object.clone() instrinsic uses this path. 4353 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4354 } 4355 4356 // Final sync IdealKit and GraphKit. 4357 final_sync(ideal); 4358 } 4359 #undef __ 4360 4361 4362 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4363 Node* len = load_array_length(load_String_value(ctrl, str)); 4364 Node* coder = load_String_coder(ctrl, str); 4365 // Divide length by 2 if coder is UTF16 4366 return _gvn.transform(new RShiftINode(len, coder)); 4367 } 4368 4369 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4370 int value_offset = java_lang_String::value_offset_in_bytes(); 4371 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4372 false, NULL, 0); 4373 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4374 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4375 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 4376 ciTypeArrayKlass::make(T_BYTE), true, 0); 4377 int value_field_idx = C->get_alias_index(value_field_type); 4378 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4379 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4380 // String.value field is known to be @Stable. 4381 if (UseImplicitStableValues) { 4382 load = cast_array_to_stable(load, value_type); 4383 } 4384 return load; 4385 } 4386 4387 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 4388 if (java_lang_String::has_coder_field()) { 4389 if (!CompactStrings) { 4390 return intcon(java_lang_String::CODER_UTF16); 4391 } 4392 int coder_offset = java_lang_String::coder_offset_in_bytes(); 4393 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4394 false, NULL, 0); 4395 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 4396 int coder_field_idx = C->get_alias_index(coder_field_type); 4397 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 4398 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 4399 } else { 4400 return intcon(0); // false 4401 } 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::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4514 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4515 // assumption of CCP analysis. 4516 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4517 }