1 /* 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "memory/barrierSet.hpp" 31 #include "memory/cardTableModRefBS.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/graphKit.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/locknode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/parse.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "runtime/deoptimization.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 43 //----------------------------GraphKit----------------------------------------- 44 // Main utility constructor. 45 GraphKit::GraphKit(JVMState* jvms) 46 : Phase(Phase::Parser), 47 _env(C->env()), 48 _gvn(*C->initial_gvn()) 49 { 50 _exceptions = jvms->map()->next_exception(); 51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 52 set_jvms(jvms); 53 } 54 55 // Private constructor for parser. 56 GraphKit::GraphKit() 57 : Phase(Phase::Parser), 58 _env(C->env()), 59 _gvn(*C->initial_gvn()) 60 { 61 _exceptions = NULL; 62 set_map(NULL); 63 debug_only(_sp = -99); 64 debug_only(set_bci(-99)); 65 } 66 67 68 69 //---------------------------clean_stack--------------------------------------- 70 // Clear away rubbish from the stack area of the JVM state. 71 // This destroys any arguments that may be waiting on the stack. 72 void GraphKit::clean_stack(int from_sp) { 73 SafePointNode* map = this->map(); 74 JVMState* jvms = this->jvms(); 75 int stk_size = jvms->stk_size(); 76 int stkoff = jvms->stkoff(); 77 Node* top = this->top(); 78 for (int i = from_sp; i < stk_size; i++) { 79 if (map->in(stkoff + i) != top) { 80 map->set_req(stkoff + i, top); 81 } 82 } 83 } 84 85 86 //--------------------------------sync_jvms----------------------------------- 87 // Make sure our current jvms agrees with our parse state. 88 JVMState* GraphKit::sync_jvms() const { 89 JVMState* jvms = this->jvms(); 90 jvms->set_bci(bci()); // Record the new bci in the JVMState 91 jvms->set_sp(sp()); // Record the new sp in the JVMState 92 assert(jvms_in_sync(), "jvms is now in sync"); 93 return jvms; 94 } 95 96 //--------------------------------sync_jvms_for_reexecute--------------------- 97 // Make sure our current jvms agrees with our parse state. This version 98 // uses the reexecute_sp for reexecuting bytecodes. 99 JVMState* GraphKit::sync_jvms_for_reexecute() { 100 JVMState* jvms = this->jvms(); 101 jvms->set_bci(bci()); // Record the new bci in the JVMState 102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 103 return jvms; 104 } 105 106 #ifdef ASSERT 107 bool GraphKit::jvms_in_sync() const { 108 Parse* parse = is_Parse(); 109 if (parse == NULL) { 110 if (bci() != jvms()->bci()) return false; 111 if (sp() != (int)jvms()->sp()) return false; 112 return true; 113 } 114 if (jvms()->method() != parse->method()) return false; 115 if (jvms()->bci() != parse->bci()) return false; 116 int jvms_sp = jvms()->sp(); 117 if (jvms_sp != parse->sp()) return false; 118 int jvms_depth = jvms()->depth(); 119 if (jvms_depth != parse->depth()) return false; 120 return true; 121 } 122 123 // Local helper checks for special internal merge points 124 // used to accumulate and merge exception states. 125 // They are marked by the region's in(0) edge being the map itself. 126 // Such merge points must never "escape" into the parser at large, 127 // until they have been handed to gvn.transform. 128 static bool is_hidden_merge(Node* reg) { 129 if (reg == NULL) return false; 130 if (reg->is_Phi()) { 131 reg = reg->in(0); 132 if (reg == NULL) return false; 133 } 134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 135 } 136 137 void GraphKit::verify_map() const { 138 if (map() == NULL) return; // null map is OK 139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 142 } 143 144 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 145 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 147 } 148 #endif 149 150 //---------------------------stop_and_kill_map--------------------------------- 151 // Set _map to NULL, signalling a stop to further bytecode execution. 152 // First smash the current map's control to a constant, to mark it dead. 153 void GraphKit::stop_and_kill_map() { 154 SafePointNode* dead_map = stop(); 155 if (dead_map != NULL) { 156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 157 assert(dead_map->is_killed(), "must be so marked"); 158 } 159 } 160 161 162 //--------------------------------stopped-------------------------------------- 163 // Tell if _map is NULL, or control is top. 164 bool GraphKit::stopped() { 165 if (map() == NULL) return true; 166 else if (control() == top()) return true; 167 else return false; 168 } 169 170 171 //-----------------------------has_ex_handler---------------------------------- 172 // Tell if this method or any caller method has exception handlers. 173 bool GraphKit::has_ex_handler() { 174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 176 return true; 177 } 178 } 179 return false; 180 } 181 182 //------------------------------save_ex_oop------------------------------------ 183 // Save an exception without blowing stack contents or other JVM state. 184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 186 ex_map->add_req(ex_oop); 187 debug_only(verify_exception_state(ex_map)); 188 } 189 190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 192 Node* ex_oop = ex_map->in(ex_map->req()-1); 193 if (clear_it) ex_map->del_req(ex_map->req()-1); 194 return ex_oop; 195 } 196 197 //-----------------------------saved_ex_oop------------------------------------ 198 // Recover a saved exception from its map. 199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 200 return common_saved_ex_oop(ex_map, false); 201 } 202 203 //--------------------------clear_saved_ex_oop--------------------------------- 204 // Erase a previously saved exception from its map. 205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 206 return common_saved_ex_oop(ex_map, true); 207 } 208 209 #ifdef ASSERT 210 //---------------------------has_saved_ex_oop---------------------------------- 211 // Erase a previously saved exception from its map. 212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 213 return ex_map->req() == ex_map->jvms()->endoff()+1; 214 } 215 #endif 216 217 //-------------------------make_exception_state-------------------------------- 218 // Turn the current JVM state into an exception state, appending the ex_oop. 219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 220 sync_jvms(); 221 SafePointNode* ex_map = stop(); // do not manipulate this map any more 222 set_saved_ex_oop(ex_map, ex_oop); 223 return ex_map; 224 } 225 226 227 //--------------------------add_exception_state-------------------------------- 228 // Add an exception to my list of exceptions. 229 void GraphKit::add_exception_state(SafePointNode* ex_map) { 230 if (ex_map == NULL || ex_map->control() == top()) { 231 return; 232 } 233 #ifdef ASSERT 234 verify_exception_state(ex_map); 235 if (has_exceptions()) { 236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 237 } 238 #endif 239 240 // If there is already an exception of exactly this type, merge with it. 241 // In particular, null-checks and other low-level exceptions common up here. 242 Node* ex_oop = saved_ex_oop(ex_map); 243 const Type* ex_type = _gvn.type(ex_oop); 244 if (ex_oop == top()) { 245 // No action needed. 246 return; 247 } 248 assert(ex_type->isa_instptr(), "exception must be an instance"); 249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 251 // We check sp also because call bytecodes can generate exceptions 252 // both before and after arguments are popped! 253 if (ex_type2 == ex_type 254 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 255 combine_exception_states(ex_map, e2); 256 return; 257 } 258 } 259 260 // No pre-existing exception of the same type. Chain it on the list. 261 push_exception_state(ex_map); 262 } 263 264 //-----------------------add_exception_states_from----------------------------- 265 void GraphKit::add_exception_states_from(JVMState* jvms) { 266 SafePointNode* ex_map = jvms->map()->next_exception(); 267 if (ex_map != NULL) { 268 jvms->map()->set_next_exception(NULL); 269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 270 next_map = ex_map->next_exception(); 271 ex_map->set_next_exception(NULL); 272 add_exception_state(ex_map); 273 } 274 } 275 } 276 277 //-----------------------transfer_exceptions_into_jvms------------------------- 278 JVMState* GraphKit::transfer_exceptions_into_jvms() { 279 if (map() == NULL) { 280 // We need a JVMS to carry the exceptions, but the map has gone away. 281 // Create a scratch JVMS, cloned from any of the exception states... 282 if (has_exceptions()) { 283 _map = _exceptions; 284 _map = clone_map(); 285 _map->set_next_exception(NULL); 286 clear_saved_ex_oop(_map); 287 debug_only(verify_map()); 288 } else { 289 // ...or created from scratch 290 JVMState* jvms = new (C) JVMState(_method, NULL); 291 jvms->set_bci(_bci); 292 jvms->set_sp(_sp); 293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); 294 set_jvms(jvms); 295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 296 set_all_memory(top()); 297 while (map()->req() < jvms->endoff()) map()->add_req(top()); 298 } 299 // (This is a kludge, in case you didn't notice.) 300 set_control(top()); 301 } 302 JVMState* jvms = sync_jvms(); 303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 304 jvms->map()->set_next_exception(_exceptions); 305 _exceptions = NULL; // done with this set of exceptions 306 return jvms; 307 } 308 309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 310 assert(is_hidden_merge(dstphi), "must be a special merge node"); 311 assert(is_hidden_merge(srcphi), "must be a special merge node"); 312 uint limit = srcphi->req(); 313 for (uint i = PhiNode::Input; i < limit; i++) { 314 dstphi->add_req(srcphi->in(i)); 315 } 316 } 317 static inline void add_one_req(Node* dstphi, Node* src) { 318 assert(is_hidden_merge(dstphi), "must be a special merge node"); 319 assert(!is_hidden_merge(src), "must not be a special merge node"); 320 dstphi->add_req(src); 321 } 322 323 //-----------------------combine_exception_states------------------------------ 324 // This helper function combines exception states by building phis on a 325 // specially marked state-merging region. These regions and phis are 326 // untransformed, and can build up gradually. The region is marked by 327 // having a control input of its exception map, rather than NULL. Such 328 // regions do not appear except in this function, and in use_exception_state. 329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 330 if (failing()) return; // dying anyway... 331 JVMState* ex_jvms = ex_map->_jvms; 332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 337 assert(ex_map->req() == phi_map->req(), "matching maps"); 338 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 339 Node* hidden_merge_mark = root(); 340 Node* region = phi_map->control(); 341 MergeMemNode* phi_mem = phi_map->merged_memory(); 342 MergeMemNode* ex_mem = ex_map->merged_memory(); 343 if (region->in(0) != hidden_merge_mark) { 344 // The control input is not (yet) a specially-marked region in phi_map. 345 // Make it so, and build some phis. 346 region = new (C) RegionNode(2); 347 _gvn.set_type(region, Type::CONTROL); 348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 349 region->init_req(1, phi_map->control()); 350 phi_map->set_control(region); 351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 352 record_for_igvn(io_phi); 353 _gvn.set_type(io_phi, Type::ABIO); 354 phi_map->set_i_o(io_phi); 355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 356 Node* m = mms.memory(); 357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 358 record_for_igvn(m_phi); 359 _gvn.set_type(m_phi, Type::MEMORY); 360 mms.set_memory(m_phi); 361 } 362 } 363 364 // Either or both of phi_map and ex_map might already be converted into phis. 365 Node* ex_control = ex_map->control(); 366 // if there is special marking on ex_map also, we add multiple edges from src 367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 368 // how wide was the destination phi_map, originally? 369 uint orig_width = region->req(); 370 371 if (add_multiple) { 372 add_n_reqs(region, ex_control); 373 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 374 } else { 375 // ex_map has no merges, so we just add single edges everywhere 376 add_one_req(region, ex_control); 377 add_one_req(phi_map->i_o(), ex_map->i_o()); 378 } 379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 380 if (mms.is_empty()) { 381 // get a copy of the base memory, and patch some inputs into it 382 const TypePtr* adr_type = mms.adr_type(C); 383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 385 mms.set_memory(phi); 386 // Prepare to append interesting stuff onto the newly sliced phi: 387 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 388 } 389 // Append stuff from ex_map: 390 if (add_multiple) { 391 add_n_reqs(mms.memory(), mms.memory2()); 392 } else { 393 add_one_req(mms.memory(), mms.memory2()); 394 } 395 } 396 uint limit = ex_map->req(); 397 for (uint i = TypeFunc::Parms; i < limit; i++) { 398 // Skip everything in the JVMS after tos. (The ex_oop follows.) 399 if (i == tos) i = ex_jvms->monoff(); 400 Node* src = ex_map->in(i); 401 Node* dst = phi_map->in(i); 402 if (src != dst) { 403 PhiNode* phi; 404 if (dst->in(0) != region) { 405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 406 record_for_igvn(phi); 407 _gvn.set_type(phi, phi->type()); 408 phi_map->set_req(i, dst); 409 // Prepare to append interesting stuff onto the new phi: 410 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 411 } else { 412 assert(dst->is_Phi(), "nobody else uses a hidden region"); 413 phi = dst->as_Phi(); 414 } 415 if (add_multiple && src->in(0) == ex_control) { 416 // Both are phis. 417 add_n_reqs(dst, src); 418 } else { 419 while (dst->req() < region->req()) add_one_req(dst, src); 420 } 421 const Type* srctype = _gvn.type(src); 422 if (phi->type() != srctype) { 423 const Type* dsttype = phi->type()->meet(srctype); 424 if (phi->type() != dsttype) { 425 phi->set_type(dsttype); 426 _gvn.set_type(phi, dsttype); 427 } 428 } 429 } 430 } 431 } 432 433 //--------------------------use_exception_state-------------------------------- 434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 435 if (failing()) { stop(); return top(); } 436 Node* region = phi_map->control(); 437 Node* hidden_merge_mark = root(); 438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 439 Node* ex_oop = clear_saved_ex_oop(phi_map); 440 if (region->in(0) == hidden_merge_mark) { 441 // Special marking for internal ex-states. Process the phis now. 442 region->set_req(0, region); // now it's an ordinary region 443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 444 // Note: Setting the jvms also sets the bci and sp. 445 set_control(_gvn.transform(region)); 446 uint tos = jvms()->stkoff() + sp(); 447 for (uint i = 1; i < tos; i++) { 448 Node* x = phi_map->in(i); 449 if (x->in(0) == region) { 450 assert(x->is_Phi(), "expected a special phi"); 451 phi_map->set_req(i, _gvn.transform(x)); 452 } 453 } 454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 455 Node* x = mms.memory(); 456 if (x->in(0) == region) { 457 assert(x->is_Phi(), "nobody else uses a hidden region"); 458 mms.set_memory(_gvn.transform(x)); 459 } 460 } 461 if (ex_oop->in(0) == region) { 462 assert(ex_oop->is_Phi(), "expected a special phi"); 463 ex_oop = _gvn.transform(ex_oop); 464 } 465 } else { 466 set_jvms(phi_map->jvms()); 467 } 468 469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 471 return ex_oop; 472 } 473 474 //---------------------------------java_bc------------------------------------- 475 Bytecodes::Code GraphKit::java_bc() const { 476 ciMethod* method = this->method(); 477 int bci = this->bci(); 478 if (method != NULL && bci != InvocationEntryBci) 479 return method->java_code_at_bci(bci); 480 else 481 return Bytecodes::_illegal; 482 } 483 484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 485 bool must_throw) { 486 // if the exception capability is set, then we will generate code 487 // to check the JavaThread.should_post_on_exceptions flag to see 488 // if we actually need to report exception events (for this 489 // thread). If we don't need to report exception events, we will 490 // take the normal fast path provided by add_exception_events. If 491 // exception event reporting is enabled for this thread, we will 492 // take the uncommon_trap in the BuildCutout below. 493 494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false); 498 499 // Test the should_post_on_exceptions_flag vs. 0 500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 502 503 // Branch to slow_path if should_post_on_exceptions_flag was true 504 { BuildCutout unless(this, tst, PROB_MAX); 505 // Do not try anything fancy if we're notifying the VM on every throw. 506 // Cf. case Bytecodes::_athrow in parse2.cpp. 507 uncommon_trap(reason, Deoptimization::Action_none, 508 (ciKlass*)NULL, (char*)NULL, must_throw); 509 } 510 511 } 512 513 //------------------------------builtin_throw---------------------------------- 514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 515 bool must_throw = true; 516 517 if (env()->jvmti_can_post_on_exceptions()) { 518 // check if we must post exception events, take uncommon trap if so 519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 520 // here if should_post_on_exceptions is false 521 // continue on with the normal codegen 522 } 523 524 // If this particular condition has not yet happened at this 525 // bytecode, then use the uncommon trap mechanism, and allow for 526 // a future recompilation if several traps occur here. 527 // If the throw is hot, try to use a more complicated inline mechanism 528 // which keeps execution inside the compiled code. 529 bool treat_throw_as_hot = false; 530 ciMethodData* md = method()->method_data(); 531 532 if (ProfileTraps) { 533 if (too_many_traps(reason)) { 534 treat_throw_as_hot = true; 535 } 536 // (If there is no MDO at all, assume it is early in 537 // execution, and that any deopts are part of the 538 // startup transient, and don't need to be remembered.) 539 540 // Also, if there is a local exception handler, treat all throws 541 // as hot if there has been at least one in this method. 542 if (C->trap_count(reason) != 0 543 && method()->method_data()->trap_count(reason) != 0 544 && has_ex_handler()) { 545 treat_throw_as_hot = true; 546 } 547 } 548 549 // If this throw happens frequently, an uncommon trap might cause 550 // a performance pothole. If there is a local exception handler, 551 // and if this particular bytecode appears to be deoptimizing often, 552 // let us handle the throw inline, with a preconstructed instance. 553 // Note: If the deopt count has blown up, the uncommon trap 554 // runtime is going to flush this nmethod, not matter what. 555 if (treat_throw_as_hot 556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 557 // If the throw is local, we use a pre-existing instance and 558 // punt on the backtrace. This would lead to a missing backtrace 559 // (a repeat of 4292742) if the backtrace object is ever asked 560 // for its backtrace. 561 // Fixing this remaining case of 4292742 requires some flavor of 562 // escape analysis. Leave that for the future. 563 ciInstance* ex_obj = NULL; 564 switch (reason) { 565 case Deoptimization::Reason_null_check: 566 ex_obj = env()->NullPointerException_instance(); 567 break; 568 case Deoptimization::Reason_div0_check: 569 ex_obj = env()->ArithmeticException_instance(); 570 break; 571 case Deoptimization::Reason_range_check: 572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 573 break; 574 case Deoptimization::Reason_class_check: 575 if (java_bc() == Bytecodes::_aastore) { 576 ex_obj = env()->ArrayStoreException_instance(); 577 } else { 578 ex_obj = env()->ClassCastException_instance(); 579 } 580 break; 581 } 582 if (failing()) { stop(); return; } // exception allocation might fail 583 if (ex_obj != NULL) { 584 // Cheat with a preallocated exception object. 585 if (C->log() != NULL) 586 C->log()->elem("hot_throw preallocated='1' reason='%s'", 587 Deoptimization::trap_reason_name(reason)); 588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 590 591 // Clear the detail message of the preallocated exception object. 592 // Weblogic sometimes mutates the detail message of exceptions 593 // using reflection. 594 int offset = java_lang_Throwable::get_detailMessage_offset(); 595 const TypePtr* adr_typ = ex_con->add_offset(offset); 596 597 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 599 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT); 600 601 add_exception_state(make_exception_state(ex_node)); 602 return; 603 } 604 } 605 606 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 607 // It won't be much cheaper than bailing to the interp., since we'll 608 // have to pass up all the debug-info, and the runtime will have to 609 // create the stack trace. 610 611 // Usual case: Bail to interpreter. 612 // Reserve the right to recompile if we haven't seen anything yet. 613 614 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 615 if (treat_throw_as_hot 616 && (method()->method_data()->trap_recompiled_at(bci()) 617 || C->too_many_traps(reason))) { 618 // We cannot afford to take more traps here. Suffer in the interpreter. 619 if (C->log() != NULL) 620 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 621 Deoptimization::trap_reason_name(reason), 622 C->trap_count(reason)); 623 action = Deoptimization::Action_none; 624 } 625 626 // "must_throw" prunes the JVM state to include only the stack, if there 627 // are no local exception handlers. This should cut down on register 628 // allocation time and code size, by drastically reducing the number 629 // of in-edges on the call to the uncommon trap. 630 631 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 632 } 633 634 635 //----------------------------PreserveJVMState--------------------------------- 636 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 637 debug_only(kit->verify_map()); 638 _kit = kit; 639 _map = kit->map(); // preserve the map 640 _sp = kit->sp(); 641 kit->set_map(clone_map ? kit->clone_map() : NULL); 642 Compile::current()->inc_preserve_jvm_state(); 643 #ifdef ASSERT 644 _bci = kit->bci(); 645 Parse* parser = kit->is_Parse(); 646 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 647 _block = block; 648 #endif 649 } 650 PreserveJVMState::~PreserveJVMState() { 651 GraphKit* kit = _kit; 652 #ifdef ASSERT 653 assert(kit->bci() == _bci, "bci must not shift"); 654 Parse* parser = kit->is_Parse(); 655 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 656 assert(block == _block, "block must not shift"); 657 #endif 658 kit->set_map(_map); 659 kit->set_sp(_sp); 660 Compile::current()->dec_preserve_jvm_state(); 661 } 662 663 664 //-----------------------------BuildCutout------------------------------------- 665 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 666 : PreserveJVMState(kit) 667 { 668 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 669 SafePointNode* outer_map = _map; // preserved map is caller's 670 SafePointNode* inner_map = kit->map(); 671 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 672 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 673 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 674 } 675 BuildCutout::~BuildCutout() { 676 GraphKit* kit = _kit; 677 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 678 } 679 680 //---------------------------PreserveReexecuteState---------------------------- 681 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 682 assert(!kit->stopped(), "must call stopped() before"); 683 _kit = kit; 684 _sp = kit->sp(); 685 _reexecute = kit->jvms()->_reexecute; 686 } 687 PreserveReexecuteState::~PreserveReexecuteState() { 688 if (_kit->stopped()) return; 689 _kit->jvms()->_reexecute = _reexecute; 690 _kit->set_sp(_sp); 691 } 692 693 //------------------------------clone_map-------------------------------------- 694 // Implementation of PreserveJVMState 695 // 696 // Only clone_map(...) here. If this function is only used in the 697 // PreserveJVMState class we may want to get rid of this extra 698 // function eventually and do it all there. 699 700 SafePointNode* GraphKit::clone_map() { 701 if (map() == NULL) return NULL; 702 703 // Clone the memory edge first 704 Node* mem = MergeMemNode::make(C, map()->memory()); 705 gvn().set_type_bottom(mem); 706 707 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 708 JVMState* jvms = this->jvms(); 709 JVMState* clonejvms = jvms->clone_shallow(C); 710 clonemap->set_memory(mem); 711 clonemap->set_jvms(clonejvms); 712 clonejvms->set_map(clonemap); 713 record_for_igvn(clonemap); 714 gvn().set_type_bottom(clonemap); 715 return clonemap; 716 } 717 718 719 //-----------------------------set_map_clone----------------------------------- 720 void GraphKit::set_map_clone(SafePointNode* m) { 721 _map = m; 722 _map = clone_map(); 723 _map->set_next_exception(NULL); 724 debug_only(verify_map()); 725 } 726 727 728 //----------------------------kill_dead_locals--------------------------------- 729 // Detect any locals which are known to be dead, and force them to top. 730 void GraphKit::kill_dead_locals() { 731 // Consult the liveness information for the locals. If any 732 // of them are unused, then they can be replaced by top(). This 733 // should help register allocation time and cut down on the size 734 // of the deoptimization information. 735 736 // This call is made from many of the bytecode handling 737 // subroutines called from the Big Switch in do_one_bytecode. 738 // Every bytecode which might include a slow path is responsible 739 // for killing its dead locals. The more consistent we 740 // are about killing deads, the fewer useless phis will be 741 // constructed for them at various merge points. 742 743 // bci can be -1 (InvocationEntryBci). We return the entry 744 // liveness for the method. 745 746 if (method() == NULL || method()->code_size() == 0) { 747 // We are building a graph for a call to a native method. 748 // All locals are live. 749 return; 750 } 751 752 ResourceMark rm; 753 754 // Consult the liveness information for the locals. If any 755 // of them are unused, then they can be replaced by top(). This 756 // should help register allocation time and cut down on the size 757 // of the deoptimization information. 758 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 759 760 int len = (int)live_locals.size(); 761 assert(len <= jvms()->loc_size(), "too many live locals"); 762 for (int local = 0; local < len; local++) { 763 if (!live_locals.at(local)) { 764 set_local(local, top()); 765 } 766 } 767 } 768 769 #ifdef ASSERT 770 //-------------------------dead_locals_are_killed------------------------------ 771 // Return true if all dead locals are set to top in the map. 772 // Used to assert "clean" debug info at various points. 773 bool GraphKit::dead_locals_are_killed() { 774 if (method() == NULL || method()->code_size() == 0) { 775 // No locals need to be dead, so all is as it should be. 776 return true; 777 } 778 779 // Make sure somebody called kill_dead_locals upstream. 780 ResourceMark rm; 781 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 782 if (jvms->loc_size() == 0) continue; // no locals to consult 783 SafePointNode* map = jvms->map(); 784 ciMethod* method = jvms->method(); 785 int bci = jvms->bci(); 786 if (jvms == this->jvms()) { 787 bci = this->bci(); // it might not yet be synched 788 } 789 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 790 int len = (int)live_locals.size(); 791 if (!live_locals.is_valid() || len == 0) 792 // This method is trivial, or is poisoned by a breakpoint. 793 return true; 794 assert(len == jvms->loc_size(), "live map consistent with locals map"); 795 for (int local = 0; local < len; local++) { 796 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 797 if (PrintMiscellaneous && (Verbose || WizardMode)) { 798 tty->print_cr("Zombie local %d: ", local); 799 jvms->dump(); 800 } 801 return false; 802 } 803 } 804 } 805 return true; 806 } 807 808 #endif //ASSERT 809 810 // Helper function for enforcing certain bytecodes to reexecute if 811 // deoptimization happens 812 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 813 ciMethod* cur_method = jvms->method(); 814 int cur_bci = jvms->bci(); 815 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 816 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 817 return Interpreter::bytecode_should_reexecute(code) || 818 is_anewarray && code == Bytecodes::_multianewarray; 819 // Reexecute _multianewarray bytecode which was replaced with 820 // sequence of [a]newarray. See Parse::do_multianewarray(). 821 // 822 // Note: interpreter should not have it set since this optimization 823 // is limited by dimensions and guarded by flag so in some cases 824 // multianewarray() runtime calls will be generated and 825 // the bytecode should not be reexecutes (stack will not be reset). 826 } else 827 return false; 828 } 829 830 // Helper function for adding JVMState and debug information to node 831 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 832 // Add the safepoint edges to the call (or other safepoint). 833 834 // Make sure dead locals are set to top. This 835 // should help register allocation time and cut down on the size 836 // of the deoptimization information. 837 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 838 839 // Walk the inline list to fill in the correct set of JVMState's 840 // Also fill in the associated edges for each JVMState. 841 842 // If the bytecode needs to be reexecuted we need to put 843 // the arguments back on the stack. 844 const bool should_reexecute = jvms()->should_reexecute(); 845 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 846 847 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 848 // undefined if the bci is different. This is normal for Parse but it 849 // should not happen for LibraryCallKit because only one bci is processed. 850 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 851 "in LibraryCallKit the reexecute bit should not change"); 852 853 // If we are guaranteed to throw, we can prune everything but the 854 // input to the current bytecode. 855 bool can_prune_locals = false; 856 uint stack_slots_not_pruned = 0; 857 int inputs = 0, depth = 0; 858 if (must_throw) { 859 assert(method() == youngest_jvms->method(), "sanity"); 860 if (compute_stack_effects(inputs, depth)) { 861 can_prune_locals = true; 862 stack_slots_not_pruned = inputs; 863 } 864 } 865 866 if (env()->jvmti_can_access_local_variables()) { 867 // At any safepoint, this method can get breakpointed, which would 868 // then require an immediate deoptimization. 869 can_prune_locals = false; // do not prune locals 870 stack_slots_not_pruned = 0; 871 } 872 873 // do not scribble on the input jvms 874 JVMState* out_jvms = youngest_jvms->clone_deep(C); 875 call->set_jvms(out_jvms); // Start jvms list for call node 876 877 // For a known set of bytecodes, the interpreter should reexecute them if 878 // deoptimization happens. We set the reexecute state for them here 879 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 880 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 881 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 882 } 883 884 // Presize the call: 885 DEBUG_ONLY(uint non_debug_edges = call->req()); 886 call->add_req_batch(top(), youngest_jvms->debug_depth()); 887 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 888 889 // Set up edges so that the call looks like this: 890 // Call [state:] ctl io mem fptr retadr 891 // [parms:] parm0 ... parmN 892 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 893 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 894 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 895 // Note that caller debug info precedes callee debug info. 896 897 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 898 uint debug_ptr = call->req(); 899 900 // Loop over the map input edges associated with jvms, add them 901 // to the call node, & reset all offsets to match call node array. 902 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 903 uint debug_end = debug_ptr; 904 uint debug_start = debug_ptr - in_jvms->debug_size(); 905 debug_ptr = debug_start; // back up the ptr 906 907 uint p = debug_start; // walks forward in [debug_start, debug_end) 908 uint j, k, l; 909 SafePointNode* in_map = in_jvms->map(); 910 out_jvms->set_map(call); 911 912 if (can_prune_locals) { 913 assert(in_jvms->method() == out_jvms->method(), "sanity"); 914 // If the current throw can reach an exception handler in this JVMS, 915 // then we must keep everything live that can reach that handler. 916 // As a quick and dirty approximation, we look for any handlers at all. 917 if (in_jvms->method()->has_exception_handlers()) { 918 can_prune_locals = false; 919 } 920 } 921 922 // Add the Locals 923 k = in_jvms->locoff(); 924 l = in_jvms->loc_size(); 925 out_jvms->set_locoff(p); 926 if (!can_prune_locals) { 927 for (j = 0; j < l; j++) 928 call->set_req(p++, in_map->in(k+j)); 929 } else { 930 p += l; // already set to top above by add_req_batch 931 } 932 933 // Add the Expression Stack 934 k = in_jvms->stkoff(); 935 l = in_jvms->sp(); 936 out_jvms->set_stkoff(p); 937 if (!can_prune_locals) { 938 for (j = 0; j < l; j++) 939 call->set_req(p++, in_map->in(k+j)); 940 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 941 // Divide stack into {S0,...,S1}, where S0 is set to top. 942 uint s1 = stack_slots_not_pruned; 943 stack_slots_not_pruned = 0; // for next iteration 944 if (s1 > l) s1 = l; 945 uint s0 = l - s1; 946 p += s0; // skip the tops preinstalled by add_req_batch 947 for (j = s0; j < l; j++) 948 call->set_req(p++, in_map->in(k+j)); 949 } else { 950 p += l; // already set to top above by add_req_batch 951 } 952 953 // Add the Monitors 954 k = in_jvms->monoff(); 955 l = in_jvms->mon_size(); 956 out_jvms->set_monoff(p); 957 for (j = 0; j < l; j++) 958 call->set_req(p++, in_map->in(k+j)); 959 960 // Copy any scalar object fields. 961 k = in_jvms->scloff(); 962 l = in_jvms->scl_size(); 963 out_jvms->set_scloff(p); 964 for (j = 0; j < l; j++) 965 call->set_req(p++, in_map->in(k+j)); 966 967 // Finish the new jvms. 968 out_jvms->set_endoff(p); 969 970 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 971 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 972 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 973 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 974 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 975 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 976 977 // Update the two tail pointers in parallel. 978 out_jvms = out_jvms->caller(); 979 in_jvms = in_jvms->caller(); 980 } 981 982 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 983 984 // Test the correctness of JVMState::debug_xxx accessors: 985 assert(call->jvms()->debug_start() == non_debug_edges, ""); 986 assert(call->jvms()->debug_end() == call->req(), ""); 987 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 988 } 989 990 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 991 Bytecodes::Code code = java_bc(); 992 if (code == Bytecodes::_wide) { 993 code = method()->java_code_at_bci(bci() + 1); 994 } 995 996 BasicType rtype = T_ILLEGAL; 997 int rsize = 0; 998 999 if (code != Bytecodes::_illegal) { 1000 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1001 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1002 if (rtype < T_CONFLICT) 1003 rsize = type2size[rtype]; 1004 } 1005 1006 switch (code) { 1007 case Bytecodes::_illegal: 1008 return false; 1009 1010 case Bytecodes::_ldc: 1011 case Bytecodes::_ldc_w: 1012 case Bytecodes::_ldc2_w: 1013 inputs = 0; 1014 break; 1015 1016 case Bytecodes::_dup: inputs = 1; break; 1017 case Bytecodes::_dup_x1: inputs = 2; break; 1018 case Bytecodes::_dup_x2: inputs = 3; break; 1019 case Bytecodes::_dup2: inputs = 2; break; 1020 case Bytecodes::_dup2_x1: inputs = 3; break; 1021 case Bytecodes::_dup2_x2: inputs = 4; break; 1022 case Bytecodes::_swap: inputs = 2; break; 1023 case Bytecodes::_arraylength: inputs = 1; break; 1024 1025 case Bytecodes::_getstatic: 1026 case Bytecodes::_putstatic: 1027 case Bytecodes::_getfield: 1028 case Bytecodes::_putfield: 1029 { 1030 bool ignored_will_link; 1031 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1032 int size = field->type()->size(); 1033 bool is_get = (depth >= 0), is_static = (depth & 1); 1034 inputs = (is_static ? 0 : 1); 1035 if (is_get) { 1036 depth = size - inputs; 1037 } else { 1038 inputs += size; // putxxx pops the value from the stack 1039 depth = - inputs; 1040 } 1041 } 1042 break; 1043 1044 case Bytecodes::_invokevirtual: 1045 case Bytecodes::_invokespecial: 1046 case Bytecodes::_invokestatic: 1047 case Bytecodes::_invokedynamic: 1048 case Bytecodes::_invokeinterface: 1049 { 1050 bool ignored_will_link; 1051 ciSignature* declared_signature = NULL; 1052 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1053 assert(declared_signature != NULL, "cannot be null"); 1054 inputs = declared_signature->arg_size_for_bc(code); 1055 int size = declared_signature->return_type()->size(); 1056 depth = size - inputs; 1057 } 1058 break; 1059 1060 case Bytecodes::_multianewarray: 1061 { 1062 ciBytecodeStream iter(method()); 1063 iter.reset_to_bci(bci()); 1064 iter.next(); 1065 inputs = iter.get_dimensions(); 1066 assert(rsize == 1, ""); 1067 depth = rsize - inputs; 1068 } 1069 break; 1070 1071 case Bytecodes::_ireturn: 1072 case Bytecodes::_lreturn: 1073 case Bytecodes::_freturn: 1074 case Bytecodes::_dreturn: 1075 case Bytecodes::_areturn: 1076 assert(rsize = -depth, ""); 1077 inputs = rsize; 1078 break; 1079 1080 case Bytecodes::_jsr: 1081 case Bytecodes::_jsr_w: 1082 inputs = 0; 1083 depth = 1; // S.B. depth=1, not zero 1084 break; 1085 1086 default: 1087 // bytecode produces a typed result 1088 inputs = rsize - depth; 1089 assert(inputs >= 0, ""); 1090 break; 1091 } 1092 1093 #ifdef ASSERT 1094 // spot check 1095 int outputs = depth + inputs; 1096 assert(outputs >= 0, "sanity"); 1097 switch (code) { 1098 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1099 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1100 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1101 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1102 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1103 } 1104 #endif //ASSERT 1105 1106 return true; 1107 } 1108 1109 1110 1111 //------------------------------basic_plus_adr--------------------------------- 1112 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1113 // short-circuit a common case 1114 if (offset == intcon(0)) return ptr; 1115 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1116 } 1117 1118 Node* GraphKit::ConvI2L(Node* offset) { 1119 // short-circuit a common case 1120 jint offset_con = find_int_con(offset, Type::OffsetBot); 1121 if (offset_con != Type::OffsetBot) { 1122 return longcon((jlong) offset_con); 1123 } 1124 return _gvn.transform( new (C) ConvI2LNode(offset)); 1125 } 1126 Node* GraphKit::ConvL2I(Node* offset) { 1127 // short-circuit a common case 1128 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1129 if (offset_con != (jlong)Type::OffsetBot) { 1130 return intcon((int) offset_con); 1131 } 1132 return _gvn.transform( new (C) ConvL2INode(offset)); 1133 } 1134 1135 //-------------------------load_object_klass----------------------------------- 1136 Node* GraphKit::load_object_klass(Node* obj) { 1137 // Special-case a fresh allocation to avoid building nodes: 1138 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1139 if (akls != NULL) return akls; 1140 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1141 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1142 } 1143 1144 //-------------------------load_array_length----------------------------------- 1145 Node* GraphKit::load_array_length(Node* array) { 1146 // Special-case a fresh allocation to avoid building nodes: 1147 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1148 Node *alen; 1149 if (alloc == NULL) { 1150 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1151 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1152 } else { 1153 alen = alloc->Ideal_length(); 1154 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1155 if (ccast != alen) { 1156 alen = _gvn.transform(ccast); 1157 } 1158 } 1159 return alen; 1160 } 1161 1162 //------------------------------do_null_check---------------------------------- 1163 // Helper function to do a NULL pointer check. Returned value is 1164 // the incoming address with NULL casted away. You are allowed to use the 1165 // not-null value only if you are control dependent on the test. 1166 extern int explicit_null_checks_inserted, 1167 explicit_null_checks_elided; 1168 Node* GraphKit::null_check_common(Node* value, BasicType type, 1169 // optional arguments for variations: 1170 bool assert_null, 1171 Node* *null_control) { 1172 assert(!assert_null || null_control == NULL, "not both at once"); 1173 if (stopped()) return top(); 1174 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1175 // For some performance testing, we may wish to suppress null checking. 1176 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1177 return value; 1178 } 1179 explicit_null_checks_inserted++; 1180 1181 // Construct NULL check 1182 Node *chk = NULL; 1183 switch(type) { 1184 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1185 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1186 case T_ARRAY : // fall through 1187 type = T_OBJECT; // simplify further tests 1188 case T_OBJECT : { 1189 const Type *t = _gvn.type( value ); 1190 1191 const TypeOopPtr* tp = t->isa_oopptr(); 1192 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1193 // Only for do_null_check, not any of its siblings: 1194 && !assert_null && null_control == NULL) { 1195 // Usually, any field access or invocation on an unloaded oop type 1196 // will simply fail to link, since the statically linked class is 1197 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1198 // the static class is loaded but the sharper oop type is not. 1199 // Rather than checking for this obscure case in lots of places, 1200 // we simply observe that a null check on an unloaded class 1201 // will always be followed by a nonsense operation, so we 1202 // can just issue the uncommon trap here. 1203 // Our access to the unloaded class will only be correct 1204 // after it has been loaded and initialized, which requires 1205 // a trip through the interpreter. 1206 #ifndef PRODUCT 1207 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1208 #endif 1209 uncommon_trap(Deoptimization::Reason_unloaded, 1210 Deoptimization::Action_reinterpret, 1211 tp->klass(), "!loaded"); 1212 return top(); 1213 } 1214 1215 if (assert_null) { 1216 // See if the type is contained in NULL_PTR. 1217 // If so, then the value is already null. 1218 if (t->higher_equal(TypePtr::NULL_PTR)) { 1219 explicit_null_checks_elided++; 1220 return value; // Elided null assert quickly! 1221 } 1222 } else { 1223 // See if mixing in the NULL pointer changes type. 1224 // If so, then the NULL pointer was not allowed in the original 1225 // type. In other words, "value" was not-null. 1226 if (t->meet(TypePtr::NULL_PTR) != t) { 1227 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1228 explicit_null_checks_elided++; 1229 return value; // Elided null check quickly! 1230 } 1231 } 1232 chk = new (C) CmpPNode( value, null() ); 1233 break; 1234 } 1235 1236 default: 1237 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1238 } 1239 assert(chk != NULL, "sanity check"); 1240 chk = _gvn.transform(chk); 1241 1242 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1243 BoolNode *btst = new (C) BoolNode( chk, btest); 1244 Node *tst = _gvn.transform( btst ); 1245 1246 //----------- 1247 // if peephole optimizations occurred, a prior test existed. 1248 // If a prior test existed, maybe it dominates as we can avoid this test. 1249 if (tst != btst && type == T_OBJECT) { 1250 // At this point we want to scan up the CFG to see if we can 1251 // find an identical test (and so avoid this test altogether). 1252 Node *cfg = control(); 1253 int depth = 0; 1254 while( depth < 16 ) { // Limit search depth for speed 1255 if( cfg->Opcode() == Op_IfTrue && 1256 cfg->in(0)->in(1) == tst ) { 1257 // Found prior test. Use "cast_not_null" to construct an identical 1258 // CastPP (and hence hash to) as already exists for the prior test. 1259 // Return that casted value. 1260 if (assert_null) { 1261 replace_in_map(value, null()); 1262 return null(); // do not issue the redundant test 1263 } 1264 Node *oldcontrol = control(); 1265 set_control(cfg); 1266 Node *res = cast_not_null(value); 1267 set_control(oldcontrol); 1268 explicit_null_checks_elided++; 1269 return res; 1270 } 1271 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1272 if (cfg == NULL) break; // Quit at region nodes 1273 depth++; 1274 } 1275 } 1276 1277 //----------- 1278 // Branch to failure if null 1279 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1280 Deoptimization::DeoptReason reason; 1281 if (assert_null) 1282 reason = Deoptimization::Reason_null_assert; 1283 else if (type == T_OBJECT) 1284 reason = Deoptimization::Reason_null_check; 1285 else 1286 reason = Deoptimization::Reason_div0_check; 1287 1288 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1289 // ciMethodData::has_trap_at will return a conservative -1 if any 1290 // must-be-null assertion has failed. This could cause performance 1291 // problems for a method after its first do_null_assert failure. 1292 // Consider using 'Reason_class_check' instead? 1293 1294 // To cause an implicit null check, we set the not-null probability 1295 // to the maximum (PROB_MAX). For an explicit check the probability 1296 // is set to a smaller value. 1297 if (null_control != NULL || too_many_traps(reason)) { 1298 // probability is less likely 1299 ok_prob = PROB_LIKELY_MAG(3); 1300 } else if (!assert_null && 1301 (ImplicitNullCheckThreshold > 0) && 1302 method() != NULL && 1303 (method()->method_data()->trap_count(reason) 1304 >= (uint)ImplicitNullCheckThreshold)) { 1305 ok_prob = PROB_LIKELY_MAG(3); 1306 } 1307 1308 if (null_control != NULL) { 1309 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1310 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1311 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1312 if (null_true == top()) 1313 explicit_null_checks_elided++; 1314 (*null_control) = null_true; 1315 } else { 1316 BuildCutout unless(this, tst, ok_prob); 1317 // Check for optimizer eliding test at parse time 1318 if (stopped()) { 1319 // Failure not possible; do not bother making uncommon trap. 1320 explicit_null_checks_elided++; 1321 } else if (assert_null) { 1322 uncommon_trap(reason, 1323 Deoptimization::Action_make_not_entrant, 1324 NULL, "assert_null"); 1325 } else { 1326 replace_in_map(value, zerocon(type)); 1327 builtin_throw(reason); 1328 } 1329 } 1330 1331 // Must throw exception, fall-thru not possible? 1332 if (stopped()) { 1333 return top(); // No result 1334 } 1335 1336 if (assert_null) { 1337 // Cast obj to null on this path. 1338 replace_in_map(value, zerocon(type)); 1339 return zerocon(type); 1340 } 1341 1342 // Cast obj to not-null on this path, if there is no null_control. 1343 // (If there is a null_control, a non-null value may come back to haunt us.) 1344 if (type == T_OBJECT) { 1345 Node* cast = cast_not_null(value, false); 1346 if (null_control == NULL || (*null_control) == top()) 1347 replace_in_map(value, cast); 1348 value = cast; 1349 } 1350 1351 return value; 1352 } 1353 1354 1355 //------------------------------cast_not_null---------------------------------- 1356 // Cast obj to not-null on this path 1357 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1358 const Type *t = _gvn.type(obj); 1359 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1360 // Object is already not-null? 1361 if( t == t_not_null ) return obj; 1362 1363 Node *cast = new (C) CastPPNode(obj,t_not_null); 1364 cast->init_req(0, control()); 1365 cast = _gvn.transform( cast ); 1366 1367 // Scan for instances of 'obj' in the current JVM mapping. 1368 // These instances are known to be not-null after the test. 1369 if (do_replace_in_map) 1370 replace_in_map(obj, cast); 1371 1372 return cast; // Return casted value 1373 } 1374 1375 1376 //--------------------------replace_in_map------------------------------------- 1377 void GraphKit::replace_in_map(Node* old, Node* neww) { 1378 if (old == neww) { 1379 return; 1380 } 1381 1382 map()->replace_edge(old, neww); 1383 1384 // Note: This operation potentially replaces any edge 1385 // on the map. This includes locals, stack, and monitors 1386 // of the current (innermost) JVM state. 1387 1388 if (!ReplaceInParentMaps) { 1389 return; 1390 } 1391 1392 // PreserveJVMState doesn't do a deep copy so we can't modify 1393 // parents 1394 if (Compile::current()->has_preserve_jvm_state()) { 1395 return; 1396 } 1397 1398 Parse* parser = is_Parse(); 1399 bool progress = true; 1400 Node* ctrl = map()->in(0); 1401 // Follow the chain of parsers and see whether the update can be 1402 // done in the map of callers. We can do the replace for a caller if 1403 // the current control post dominates the control of a caller. 1404 while (parser != NULL && parser->caller() != NULL && progress) { 1405 progress = false; 1406 Node* parent_map = parser->caller()->map(); 1407 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch"); 1408 1409 Node* parent_ctrl = parent_map->in(0); 1410 1411 while (parent_ctrl->is_Region()) { 1412 Node* n = parent_ctrl->as_Region()->is_copy(); 1413 if (n == NULL) { 1414 break; 1415 } 1416 parent_ctrl = n; 1417 } 1418 1419 for (;;) { 1420 if (ctrl == parent_ctrl) { 1421 // update the map of the exits which is the one that will be 1422 // used when compilation resume after inlining 1423 parser->exits().map()->replace_edge(old, neww); 1424 progress = true; 1425 break; 1426 } 1427 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 1428 ctrl = ctrl->in(0)->in(0); 1429 } else if (ctrl->is_Region()) { 1430 Node* n = ctrl->as_Region()->is_copy(); 1431 if (n == NULL) { 1432 break; 1433 } 1434 ctrl = n; 1435 } else { 1436 break; 1437 } 1438 } 1439 1440 parser = parser->parent_parser(); 1441 } 1442 } 1443 1444 1445 //============================================================================= 1446 //--------------------------------memory--------------------------------------- 1447 Node* GraphKit::memory(uint alias_idx) { 1448 MergeMemNode* mem = merged_memory(); 1449 Node* p = mem->memory_at(alias_idx); 1450 _gvn.set_type(p, Type::MEMORY); // must be mapped 1451 return p; 1452 } 1453 1454 //-----------------------------reset_memory------------------------------------ 1455 Node* GraphKit::reset_memory() { 1456 Node* mem = map()->memory(); 1457 // do not use this node for any more parsing! 1458 debug_only( map()->set_memory((Node*)NULL) ); 1459 return _gvn.transform( mem ); 1460 } 1461 1462 //------------------------------set_all_memory--------------------------------- 1463 void GraphKit::set_all_memory(Node* newmem) { 1464 Node* mergemem = MergeMemNode::make(C, newmem); 1465 gvn().set_type_bottom(mergemem); 1466 map()->set_memory(mergemem); 1467 } 1468 1469 //------------------------------set_all_memory_call---------------------------- 1470 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1471 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1472 set_all_memory(newmem); 1473 } 1474 1475 //============================================================================= 1476 // 1477 // parser factory methods for MemNodes 1478 // 1479 // These are layered on top of the factory methods in LoadNode and StoreNode, 1480 // and integrate with the parser's memory state and _gvn engine. 1481 // 1482 1483 // factory methods in "int adr_idx" 1484 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1485 int adr_idx, 1486 bool require_atomic_access) { 1487 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1488 const TypePtr* adr_type = NULL; // debug-mode-only argument 1489 debug_only(adr_type = C->get_adr_type(adr_idx)); 1490 Node* mem = memory(adr_idx); 1491 Node* ld; 1492 if (require_atomic_access && bt == T_LONG) { 1493 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); 1494 } else { 1495 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); 1496 } 1497 ld = _gvn.transform(ld); 1498 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1499 // Improve graph before escape analysis and boxing elimination. 1500 record_for_igvn(ld); 1501 } 1502 return ld; 1503 } 1504 1505 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1506 int adr_idx, 1507 bool require_atomic_access) { 1508 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1509 const TypePtr* adr_type = NULL; 1510 debug_only(adr_type = C->get_adr_type(adr_idx)); 1511 Node *mem = memory(adr_idx); 1512 Node* st; 1513 if (require_atomic_access && bt == T_LONG) { 1514 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); 1515 } else { 1516 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); 1517 } 1518 st = _gvn.transform(st); 1519 set_memory(st, adr_idx); 1520 // Back-to-back stores can only remove intermediate store with DU info 1521 // so push on worklist for optimizer. 1522 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1523 record_for_igvn(st); 1524 1525 return st; 1526 } 1527 1528 1529 void GraphKit::pre_barrier(bool do_load, 1530 Node* ctl, 1531 Node* obj, 1532 Node* adr, 1533 uint adr_idx, 1534 Node* val, 1535 const TypeOopPtr* val_type, 1536 Node* pre_val, 1537 BasicType bt) { 1538 1539 BarrierSet* bs = Universe::heap()->barrier_set(); 1540 set_control(ctl); 1541 switch (bs->kind()) { 1542 case BarrierSet::G1SATBCT: 1543 case BarrierSet::G1SATBCTLogging: 1544 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1545 break; 1546 1547 case BarrierSet::CardTableModRef: 1548 case BarrierSet::CardTableExtension: 1549 case BarrierSet::ModRef: 1550 break; 1551 1552 case BarrierSet::Other: 1553 default : 1554 ShouldNotReachHere(); 1555 1556 } 1557 } 1558 1559 bool GraphKit::can_move_pre_barrier() const { 1560 BarrierSet* bs = Universe::heap()->barrier_set(); 1561 switch (bs->kind()) { 1562 case BarrierSet::G1SATBCT: 1563 case BarrierSet::G1SATBCTLogging: 1564 return true; // Can move it if no safepoint 1565 1566 case BarrierSet::CardTableModRef: 1567 case BarrierSet::CardTableExtension: 1568 case BarrierSet::ModRef: 1569 return true; // There is no pre-barrier 1570 1571 case BarrierSet::Other: 1572 default : 1573 ShouldNotReachHere(); 1574 } 1575 return false; 1576 } 1577 1578 void GraphKit::post_barrier(Node* ctl, 1579 Node* store, 1580 Node* obj, 1581 Node* adr, 1582 uint adr_idx, 1583 Node* val, 1584 BasicType bt, 1585 bool use_precise) { 1586 BarrierSet* bs = Universe::heap()->barrier_set(); 1587 set_control(ctl); 1588 switch (bs->kind()) { 1589 case BarrierSet::G1SATBCT: 1590 case BarrierSet::G1SATBCTLogging: 1591 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1592 break; 1593 1594 case BarrierSet::CardTableModRef: 1595 case BarrierSet::CardTableExtension: 1596 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1597 break; 1598 1599 case BarrierSet::ModRef: 1600 break; 1601 1602 case BarrierSet::Other: 1603 default : 1604 ShouldNotReachHere(); 1605 1606 } 1607 } 1608 1609 Node* GraphKit::store_oop(Node* ctl, 1610 Node* obj, 1611 Node* adr, 1612 const TypePtr* adr_type, 1613 Node* val, 1614 const TypeOopPtr* val_type, 1615 BasicType bt, 1616 bool use_precise) { 1617 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1618 // could be delayed during Parse (for example, in adjust_map_after_if()). 1619 // Execute transformation here to avoid barrier generation in such case. 1620 if (_gvn.type(val) == TypePtr::NULL_PTR) 1621 val = _gvn.makecon(TypePtr::NULL_PTR); 1622 1623 set_control(ctl); 1624 if (stopped()) return top(); // Dead path ? 1625 1626 assert(bt == T_OBJECT, "sanity"); 1627 assert(val != NULL, "not dead path"); 1628 uint adr_idx = C->get_alias_index(adr_type); 1629 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1630 1631 pre_barrier(true /* do_load */, 1632 control(), obj, adr, adr_idx, val, val_type, 1633 NULL /* pre_val */, 1634 bt); 1635 1636 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1637 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1638 return store; 1639 } 1640 1641 // Could be an array or object we don't know at compile time (unsafe ref.) 1642 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1643 Node* obj, // containing obj 1644 Node* adr, // actual adress to store val at 1645 const TypePtr* adr_type, 1646 Node* val, 1647 BasicType bt) { 1648 Compile::AliasType* at = C->alias_type(adr_type); 1649 const TypeOopPtr* val_type = NULL; 1650 if (adr_type->isa_instptr()) { 1651 if (at->field() != NULL) { 1652 // known field. This code is a copy of the do_put_xxx logic. 1653 ciField* field = at->field(); 1654 if (!field->type()->is_loaded()) { 1655 val_type = TypeInstPtr::BOTTOM; 1656 } else { 1657 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1658 } 1659 } 1660 } else if (adr_type->isa_aryptr()) { 1661 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1662 } 1663 if (val_type == NULL) { 1664 val_type = TypeInstPtr::BOTTOM; 1665 } 1666 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); 1667 } 1668 1669 1670 //-------------------------array_element_address------------------------- 1671 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1672 const TypeInt* sizetype) { 1673 uint shift = exact_log2(type2aelembytes(elembt)); 1674 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1675 1676 // short-circuit a common case (saves lots of confusing waste motion) 1677 jint idx_con = find_int_con(idx, -1); 1678 if (idx_con >= 0) { 1679 intptr_t offset = header + ((intptr_t)idx_con << shift); 1680 return basic_plus_adr(ary, offset); 1681 } 1682 1683 // must be correct type for alignment purposes 1684 Node* base = basic_plus_adr(ary, header); 1685 #ifdef _LP64 1686 // The scaled index operand to AddP must be a clean 64-bit value. 1687 // Java allows a 32-bit int to be incremented to a negative 1688 // value, which appears in a 64-bit register as a large 1689 // positive number. Using that large positive number as an 1690 // operand in pointer arithmetic has bad consequences. 1691 // On the other hand, 32-bit overflow is rare, and the possibility 1692 // can often be excluded, if we annotate the ConvI2L node with 1693 // a type assertion that its value is known to be a small positive 1694 // number. (The prior range check has ensured this.) 1695 // This assertion is used by ConvI2LNode::Ideal. 1696 int index_max = max_jint - 1; // array size is max_jint, index is one less 1697 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1698 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1699 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) ); 1700 #endif 1701 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1702 return basic_plus_adr(ary, base, scale); 1703 } 1704 1705 //-------------------------load_array_element------------------------- 1706 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1707 const Type* elemtype = arytype->elem(); 1708 BasicType elembt = elemtype->array_element_basic_type(); 1709 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1710 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1711 return ld; 1712 } 1713 1714 //-------------------------set_arguments_for_java_call------------------------- 1715 // Arguments (pre-popped from the stack) are taken from the JVMS. 1716 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1717 // Add the call arguments: 1718 uint nargs = call->method()->arg_size(); 1719 for (uint i = 0; i < nargs; i++) { 1720 Node* arg = argument(i); 1721 call->init_req(i + TypeFunc::Parms, arg); 1722 } 1723 } 1724 1725 //---------------------------set_edges_for_java_call--------------------------- 1726 // Connect a newly created call into the current JVMS. 1727 // A return value node (if any) is returned from set_edges_for_java_call. 1728 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1729 1730 // Add the predefined inputs: 1731 call->init_req( TypeFunc::Control, control() ); 1732 call->init_req( TypeFunc::I_O , i_o() ); 1733 call->init_req( TypeFunc::Memory , reset_memory() ); 1734 call->init_req( TypeFunc::FramePtr, frameptr() ); 1735 call->init_req( TypeFunc::ReturnAdr, top() ); 1736 1737 add_safepoint_edges(call, must_throw); 1738 1739 Node* xcall = _gvn.transform(call); 1740 1741 if (xcall == top()) { 1742 set_control(top()); 1743 return; 1744 } 1745 assert(xcall == call, "call identity is stable"); 1746 1747 // Re-use the current map to produce the result. 1748 1749 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1750 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1751 set_all_memory_call(xcall, separate_io_proj); 1752 1753 //return xcall; // no need, caller already has it 1754 } 1755 1756 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1757 if (stopped()) return top(); // maybe the call folded up? 1758 1759 // Capture the return value, if any. 1760 Node* ret; 1761 if (call->method() == NULL || 1762 call->method()->return_type()->basic_type() == T_VOID) 1763 ret = top(); 1764 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1765 1766 // Note: Since any out-of-line call can produce an exception, 1767 // we always insert an I_O projection from the call into the result. 1768 1769 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1770 1771 if (separate_io_proj) { 1772 // The caller requested separate projections be used by the fall 1773 // through and exceptional paths, so replace the projections for 1774 // the fall through path. 1775 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1776 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1777 } 1778 return ret; 1779 } 1780 1781 //--------------------set_predefined_input_for_runtime_call-------------------- 1782 // Reading and setting the memory state is way conservative here. 1783 // The real problem is that I am not doing real Type analysis on memory, 1784 // so I cannot distinguish card mark stores from other stores. Across a GC 1785 // point the Store Barrier and the card mark memory has to agree. I cannot 1786 // have a card mark store and its barrier split across the GC point from 1787 // either above or below. Here I get that to happen by reading ALL of memory. 1788 // A better answer would be to separate out card marks from other memory. 1789 // For now, return the input memory state, so that it can be reused 1790 // after the call, if this call has restricted memory effects. 1791 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1792 // Set fixed predefined input arguments 1793 Node* memory = reset_memory(); 1794 call->init_req( TypeFunc::Control, control() ); 1795 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1796 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1797 call->init_req( TypeFunc::FramePtr, frameptr() ); 1798 call->init_req( TypeFunc::ReturnAdr, top() ); 1799 return memory; 1800 } 1801 1802 //-------------------set_predefined_output_for_runtime_call-------------------- 1803 // Set control and memory (not i_o) from the call. 1804 // If keep_mem is not NULL, use it for the output state, 1805 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1806 // If hook_mem is NULL, this call produces no memory effects at all. 1807 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1808 // then only that memory slice is taken from the call. 1809 // In the last case, we must put an appropriate memory barrier before 1810 // the call, so as to create the correct anti-dependencies on loads 1811 // preceding the call. 1812 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1813 Node* keep_mem, 1814 const TypePtr* hook_mem) { 1815 // no i/o 1816 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1817 if (keep_mem) { 1818 // First clone the existing memory state 1819 set_all_memory(keep_mem); 1820 if (hook_mem != NULL) { 1821 // Make memory for the call 1822 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1823 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1824 // We also use hook_mem to extract specific effects from arraycopy stubs. 1825 set_memory(mem, hook_mem); 1826 } 1827 // ...else the call has NO memory effects. 1828 1829 // Make sure the call advertises its memory effects precisely. 1830 // This lets us build accurate anti-dependences in gcm.cpp. 1831 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1832 "call node must be constructed correctly"); 1833 } else { 1834 assert(hook_mem == NULL, ""); 1835 // This is not a "slow path" call; all memory comes from the call. 1836 set_all_memory_call(call); 1837 } 1838 } 1839 1840 1841 // Replace the call with the current state of the kit. 1842 void GraphKit::replace_call(CallNode* call, Node* result) { 1843 JVMState* ejvms = NULL; 1844 if (has_exceptions()) { 1845 ejvms = transfer_exceptions_into_jvms(); 1846 } 1847 1848 SafePointNode* final_state = stop(); 1849 1850 // Find all the needed outputs of this call 1851 CallProjections callprojs; 1852 call->extract_projections(&callprojs, true); 1853 1854 Node* init_mem = call->in(TypeFunc::Memory); 1855 Node* final_mem = final_state->in(TypeFunc::Memory); 1856 Node* final_ctl = final_state->in(TypeFunc::Control); 1857 Node* final_io = final_state->in(TypeFunc::I_O); 1858 1859 // Replace all the old call edges with the edges from the inlining result 1860 if (callprojs.fallthrough_catchproj != NULL) { 1861 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1862 } 1863 if (callprojs.fallthrough_memproj != NULL) { 1864 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1865 } 1866 if (callprojs.fallthrough_ioproj != NULL) { 1867 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1868 } 1869 1870 // Replace the result with the new result if it exists and is used 1871 if (callprojs.resproj != NULL && result != NULL) { 1872 C->gvn_replace_by(callprojs.resproj, result); 1873 } 1874 1875 if (ejvms == NULL) { 1876 // No exception edges to simply kill off those paths 1877 if (callprojs.catchall_catchproj != NULL) { 1878 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1879 } 1880 if (callprojs.catchall_memproj != NULL) { 1881 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1882 } 1883 if (callprojs.catchall_ioproj != NULL) { 1884 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1885 } 1886 // Replace the old exception object with top 1887 if (callprojs.exobj != NULL) { 1888 C->gvn_replace_by(callprojs.exobj, C->top()); 1889 } 1890 } else { 1891 GraphKit ekit(ejvms); 1892 1893 // Load my combined exception state into the kit, with all phis transformed: 1894 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1895 1896 Node* ex_oop = ekit.use_exception_state(ex_map); 1897 if (callprojs.catchall_catchproj != NULL) { 1898 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1899 } 1900 if (callprojs.catchall_memproj != NULL) { 1901 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1902 } 1903 if (callprojs.catchall_ioproj != NULL) { 1904 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1905 } 1906 1907 // Replace the old exception object with the newly created one 1908 if (callprojs.exobj != NULL) { 1909 C->gvn_replace_by(callprojs.exobj, ex_oop); 1910 } 1911 } 1912 1913 // Disconnect the call from the graph 1914 call->disconnect_inputs(NULL, C); 1915 C->gvn_replace_by(call, C->top()); 1916 1917 // Clean up any MergeMems that feed other MergeMems since the 1918 // optimizer doesn't like that. 1919 if (final_mem->is_MergeMem()) { 1920 Node_List wl; 1921 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1922 Node* m = i.get(); 1923 if (m->is_MergeMem() && !wl.contains(m)) { 1924 wl.push(m); 1925 } 1926 } 1927 while (wl.size() > 0) { 1928 _gvn.transform(wl.pop()); 1929 } 1930 } 1931 } 1932 1933 1934 //------------------------------increment_counter------------------------------ 1935 // for statistics: increment a VM counter by 1 1936 1937 void GraphKit::increment_counter(address counter_addr) { 1938 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1939 increment_counter(adr1); 1940 } 1941 1942 void GraphKit::increment_counter(Node* counter_addr) { 1943 int adr_type = Compile::AliasIdxRaw; 1944 Node* ctrl = control(); 1945 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type); 1946 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1947 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type ); 1948 } 1949 1950 1951 //------------------------------uncommon_trap---------------------------------- 1952 // Bail out to the interpreter in mid-method. Implemented by calling the 1953 // uncommon_trap blob. This helper function inserts a runtime call with the 1954 // right debug info. 1955 void GraphKit::uncommon_trap(int trap_request, 1956 ciKlass* klass, const char* comment, 1957 bool must_throw, 1958 bool keep_exact_action) { 1959 if (failing()) stop(); 1960 if (stopped()) return; // trap reachable? 1961 1962 // Note: If ProfileTraps is true, and if a deopt. actually 1963 // occurs here, the runtime will make sure an MDO exists. There is 1964 // no need to call method()->ensure_method_data() at this point. 1965 1966 // Set the stack pointer to the right value for reexecution: 1967 set_sp(reexecute_sp()); 1968 1969 #ifdef ASSERT 1970 if (!must_throw) { 1971 // Make sure the stack has at least enough depth to execute 1972 // the current bytecode. 1973 int inputs, ignored_depth; 1974 if (compute_stack_effects(inputs, ignored_depth)) { 1975 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1976 Bytecodes::name(java_bc()), sp(), inputs)); 1977 } 1978 } 1979 #endif 1980 1981 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1982 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1983 1984 switch (action) { 1985 case Deoptimization::Action_maybe_recompile: 1986 case Deoptimization::Action_reinterpret: 1987 // Temporary fix for 6529811 to allow virtual calls to be sure they 1988 // get the chance to go from mono->bi->mega 1989 if (!keep_exact_action && 1990 Deoptimization::trap_request_index(trap_request) < 0 && 1991 too_many_recompiles(reason)) { 1992 // This BCI is causing too many recompilations. 1993 action = Deoptimization::Action_none; 1994 trap_request = Deoptimization::make_trap_request(reason, action); 1995 } else { 1996 C->set_trap_can_recompile(true); 1997 } 1998 break; 1999 case Deoptimization::Action_make_not_entrant: 2000 C->set_trap_can_recompile(true); 2001 break; 2002 #ifdef ASSERT 2003 case Deoptimization::Action_none: 2004 case Deoptimization::Action_make_not_compilable: 2005 break; 2006 default: 2007 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2008 break; 2009 #endif 2010 } 2011 2012 if (TraceOptoParse) { 2013 char buf[100]; 2014 tty->print_cr("Uncommon trap %s at bci:%d", 2015 Deoptimization::format_trap_request(buf, sizeof(buf), 2016 trap_request), bci()); 2017 } 2018 2019 CompileLog* log = C->log(); 2020 if (log != NULL) { 2021 int kid = (klass == NULL)? -1: log->identify(klass); 2022 log->begin_elem("uncommon_trap bci='%d'", bci()); 2023 char buf[100]; 2024 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2025 trap_request)); 2026 if (kid >= 0) log->print(" klass='%d'", kid); 2027 if (comment != NULL) log->print(" comment='%s'", comment); 2028 log->end_elem(); 2029 } 2030 2031 // Make sure any guarding test views this path as very unlikely 2032 Node *i0 = control()->in(0); 2033 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2034 IfNode *iff = i0->as_If(); 2035 float f = iff->_prob; // Get prob 2036 if (control()->Opcode() == Op_IfTrue) { 2037 if (f > PROB_UNLIKELY_MAG(4)) 2038 iff->_prob = PROB_MIN; 2039 } else { 2040 if (f < PROB_LIKELY_MAG(4)) 2041 iff->_prob = PROB_MAX; 2042 } 2043 } 2044 2045 // Clear out dead values from the debug info. 2046 kill_dead_locals(); 2047 2048 // Now insert the uncommon trap subroutine call 2049 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2050 const TypePtr* no_memory_effects = NULL; 2051 // Pass the index of the class to be loaded 2052 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2053 (must_throw ? RC_MUST_THROW : 0), 2054 OptoRuntime::uncommon_trap_Type(), 2055 call_addr, "uncommon_trap", no_memory_effects, 2056 intcon(trap_request)); 2057 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2058 "must extract request correctly from the graph"); 2059 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2060 2061 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2062 // The debug info is the only real input to this call. 2063 2064 // Halt-and-catch fire here. The above call should never return! 2065 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2066 _gvn.set_type_bottom(halt); 2067 root()->add_req(halt); 2068 2069 stop_and_kill_map(); 2070 } 2071 2072 2073 //--------------------------just_allocated_object------------------------------ 2074 // Report the object that was just allocated. 2075 // It must be the case that there are no intervening safepoints. 2076 // We use this to determine if an object is so "fresh" that 2077 // it does not require card marks. 2078 Node* GraphKit::just_allocated_object(Node* current_control) { 2079 if (C->recent_alloc_ctl() == current_control) 2080 return C->recent_alloc_obj(); 2081 return NULL; 2082 } 2083 2084 2085 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2086 // (Note: TypeFunc::make has a cache that makes this fast.) 2087 const TypeFunc* tf = TypeFunc::make(dest_method); 2088 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2089 for (int j = 0; j < nargs; j++) { 2090 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2091 if( targ->basic_type() == T_DOUBLE ) { 2092 // If any parameters are doubles, they must be rounded before 2093 // the call, dstore_rounding does gvn.transform 2094 Node *arg = argument(j); 2095 arg = dstore_rounding(arg); 2096 set_argument(j, arg); 2097 } 2098 } 2099 } 2100 2101 /** 2102 * Record profiling data exact_kls for Node n with the type system so 2103 * that it can propagate it (speculation) 2104 * 2105 * @param n node that the type applies to 2106 * @param exact_kls type from profiling 2107 * 2108 * @return node with improved type 2109 */ 2110 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) { 2111 const TypeOopPtr* current_type = _gvn.type(n)->isa_oopptr(); 2112 assert(UseTypeSpeculation, "type speculation must be on"); 2113 if (exact_kls != NULL && 2114 // nothing to improve if type is already exact 2115 (current_type == NULL || 2116 (!current_type->klass_is_exact() && 2117 (current_type->speculative() == NULL || 2118 !current_type->speculative()->klass_is_exact())))) { 2119 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2120 const TypeOopPtr* xtype = tklass->as_instance_type(); 2121 assert(xtype->klass_is_exact(), "Should be exact"); 2122 2123 // Build a type with a speculative type (what we think we know 2124 // about the type but will need a guard when we use it) 2125 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, xtype); 2126 // We're changing the type, we need a new cast node to carry the 2127 // new type. The new type depends on the control: what profiling 2128 // tells us is only valid from here as far as we can tell. 2129 Node* cast = new(C) CastPPNode(n, spec_type); 2130 cast->init_req(0, control()); 2131 cast = _gvn.transform(cast); 2132 replace_in_map(n, cast); 2133 n = cast; 2134 } 2135 return n; 2136 } 2137 2138 /** 2139 * Record profiling data from receiver profiling at an invoke with the 2140 * type system so that it can propagate it (speculation) 2141 * 2142 * @param n receiver node 2143 * 2144 * @return node with improved type 2145 */ 2146 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2147 if (!UseTypeSpeculation) { 2148 return n; 2149 } 2150 ciKlass* exact_kls = profile_has_unique_klass(); 2151 return record_profile_for_speculation(n, exact_kls); 2152 } 2153 2154 /** 2155 * Record profiling data from argument profiling at an invoke with the 2156 * type system so that it can propagate it (speculation) 2157 * 2158 * @param dest_method target method for the call 2159 * @param bc what invoke bytecode is this? 2160 */ 2161 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2162 if (!UseTypeSpeculation) { 2163 return; 2164 } 2165 const TypeFunc* tf = TypeFunc::make(dest_method); 2166 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2167 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2168 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2169 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2170 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2171 ciKlass* better_type = method()->argument_profiled_type(bci(), i); 2172 if (better_type != NULL) { 2173 record_profile_for_speculation(argument(j), better_type); 2174 } 2175 i++; 2176 } 2177 } 2178 } 2179 2180 /** 2181 * Record profiling data from parameter profiling at an invoke with 2182 * the type system so that it can propagate it (speculation) 2183 */ 2184 void GraphKit::record_profiled_parameters_for_speculation() { 2185 if (!UseTypeSpeculation) { 2186 return; 2187 } 2188 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2189 if (_gvn.type(local(i))->isa_oopptr()) { 2190 ciKlass* better_type = method()->parameter_profiled_type(j); 2191 if (better_type != NULL) { 2192 record_profile_for_speculation(local(i), better_type); 2193 } 2194 j++; 2195 } 2196 } 2197 } 2198 2199 void GraphKit::round_double_result(ciMethod* dest_method) { 2200 // A non-strict method may return a double value which has an extended 2201 // exponent, but this must not be visible in a caller which is 'strict' 2202 // If a strict caller invokes a non-strict callee, round a double result 2203 2204 BasicType result_type = dest_method->return_type()->basic_type(); 2205 assert( method() != NULL, "must have caller context"); 2206 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2207 // Destination method's return value is on top of stack 2208 // dstore_rounding() does gvn.transform 2209 Node *result = pop_pair(); 2210 result = dstore_rounding(result); 2211 push_pair(result); 2212 } 2213 } 2214 2215 // rounding for strict float precision conformance 2216 Node* GraphKit::precision_rounding(Node* n) { 2217 return UseStrictFP && _method->flags().is_strict() 2218 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2219 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2220 : n; 2221 } 2222 2223 // rounding for strict double precision conformance 2224 Node* GraphKit::dprecision_rounding(Node *n) { 2225 return UseStrictFP && _method->flags().is_strict() 2226 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2227 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2228 : n; 2229 } 2230 2231 // rounding for non-strict double stores 2232 Node* GraphKit::dstore_rounding(Node* n) { 2233 return Matcher::strict_fp_requires_explicit_rounding 2234 && UseSSE <= 1 2235 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2236 : n; 2237 } 2238 2239 //============================================================================= 2240 // Generate a fast path/slow path idiom. Graph looks like: 2241 // [foo] indicates that 'foo' is a parameter 2242 // 2243 // [in] NULL 2244 // \ / 2245 // CmpP 2246 // Bool ne 2247 // If 2248 // / \ 2249 // True False-<2> 2250 // / | 2251 // / cast_not_null 2252 // Load | | ^ 2253 // [fast_test] | | 2254 // gvn to opt_test | | 2255 // / \ | <1> 2256 // True False | 2257 // | \\ | 2258 // [slow_call] \[fast_result] 2259 // Ctl Val \ \ 2260 // | \ \ 2261 // Catch <1> \ \ 2262 // / \ ^ \ \ 2263 // Ex No_Ex | \ \ 2264 // | \ \ | \ <2> \ 2265 // ... \ [slow_res] | | \ [null_result] 2266 // \ \--+--+--- | | 2267 // \ | / \ | / 2268 // --------Region Phi 2269 // 2270 //============================================================================= 2271 // Code is structured as a series of driver functions all called 'do_XXX' that 2272 // call a set of helper functions. Helper functions first, then drivers. 2273 2274 //------------------------------null_check_oop--------------------------------- 2275 // Null check oop. Set null-path control into Region in slot 3. 2276 // Make a cast-not-nullness use the other not-null control. Return cast. 2277 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2278 bool never_see_null, bool safe_for_replace) { 2279 // Initial NULL check taken path 2280 (*null_control) = top(); 2281 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2282 2283 // Generate uncommon_trap: 2284 if (never_see_null && (*null_control) != top()) { 2285 // If we see an unexpected null at a check-cast we record it and force a 2286 // recompile; the offending check-cast will be compiled to handle NULLs. 2287 // If we see more than one offending BCI, then all checkcasts in the 2288 // method will be compiled to handle NULLs. 2289 PreserveJVMState pjvms(this); 2290 set_control(*null_control); 2291 replace_in_map(value, null()); 2292 uncommon_trap(Deoptimization::Reason_null_check, 2293 Deoptimization::Action_make_not_entrant); 2294 (*null_control) = top(); // NULL path is dead 2295 } 2296 if ((*null_control) == top() && safe_for_replace) { 2297 replace_in_map(value, cast); 2298 } 2299 2300 // Cast away null-ness on the result 2301 return cast; 2302 } 2303 2304 //------------------------------opt_iff---------------------------------------- 2305 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2306 // Return slow-path control. 2307 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2308 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2309 2310 // Fast path taken; set region slot 2 2311 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2312 region->init_req(2,fast_taken); // Capture fast-control 2313 2314 // Fast path not-taken, i.e. slow path 2315 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2316 return slow_taken; 2317 } 2318 2319 //-----------------------------make_runtime_call------------------------------- 2320 Node* GraphKit::make_runtime_call(int flags, 2321 const TypeFunc* call_type, address call_addr, 2322 const char* call_name, 2323 const TypePtr* adr_type, 2324 // The following parms are all optional. 2325 // The first NULL ends the list. 2326 Node* parm0, Node* parm1, 2327 Node* parm2, Node* parm3, 2328 Node* parm4, Node* parm5, 2329 Node* parm6, Node* parm7) { 2330 // Slow-path call 2331 bool is_leaf = !(flags & RC_NO_LEAF); 2332 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2333 if (call_name == NULL) { 2334 assert(!is_leaf, "must supply name for leaf"); 2335 call_name = OptoRuntime::stub_name(call_addr); 2336 } 2337 CallNode* call; 2338 if (!is_leaf) { 2339 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2340 bci(), adr_type); 2341 } else if (flags & RC_NO_FP) { 2342 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2343 } else { 2344 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2345 } 2346 2347 // The following is similar to set_edges_for_java_call, 2348 // except that the memory effects of the call are restricted to AliasIdxRaw. 2349 2350 // Slow path call has no side-effects, uses few values 2351 bool wide_in = !(flags & RC_NARROW_MEM); 2352 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2353 2354 Node* prev_mem = NULL; 2355 if (wide_in) { 2356 prev_mem = set_predefined_input_for_runtime_call(call); 2357 } else { 2358 assert(!wide_out, "narrow in => narrow out"); 2359 Node* narrow_mem = memory(adr_type); 2360 prev_mem = reset_memory(); 2361 map()->set_memory(narrow_mem); 2362 set_predefined_input_for_runtime_call(call); 2363 } 2364 2365 // Hook each parm in order. Stop looking at the first NULL. 2366 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2367 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2368 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2369 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2370 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2371 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2372 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2373 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2374 /* close each nested if ===> */ } } } } } } } } 2375 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2376 2377 if (!is_leaf) { 2378 // Non-leaves can block and take safepoints: 2379 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2380 } 2381 // Non-leaves can throw exceptions: 2382 if (has_io) { 2383 call->set_req(TypeFunc::I_O, i_o()); 2384 } 2385 2386 if (flags & RC_UNCOMMON) { 2387 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2388 // (An "if" probability corresponds roughly to an unconditional count. 2389 // Sort of.) 2390 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2391 } 2392 2393 Node* c = _gvn.transform(call); 2394 assert(c == call, "cannot disappear"); 2395 2396 if (wide_out) { 2397 // Slow path call has full side-effects. 2398 set_predefined_output_for_runtime_call(call); 2399 } else { 2400 // Slow path call has few side-effects, and/or sets few values. 2401 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2402 } 2403 2404 if (has_io) { 2405 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2406 } 2407 return call; 2408 2409 } 2410 2411 //------------------------------merge_memory----------------------------------- 2412 // Merge memory from one path into the current memory state. 2413 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2414 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2415 Node* old_slice = mms.force_memory(); 2416 Node* new_slice = mms.memory2(); 2417 if (old_slice != new_slice) { 2418 PhiNode* phi; 2419 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2420 phi = new_slice->as_Phi(); 2421 #ifdef ASSERT 2422 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2423 old_slice = old_slice->in(new_path); 2424 // Caller is responsible for ensuring that any pre-existing 2425 // phis are already aware of old memory. 2426 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2427 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2428 #endif 2429 mms.set_memory(phi); 2430 } else { 2431 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2432 _gvn.set_type(phi, Type::MEMORY); 2433 phi->set_req(new_path, new_slice); 2434 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2435 } 2436 } 2437 } 2438 } 2439 2440 //------------------------------make_slow_call_ex------------------------------ 2441 // Make the exception handler hookups for the slow call 2442 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2443 if (stopped()) return; 2444 2445 // Make a catch node with just two handlers: fall-through and catch-all 2446 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2447 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2448 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2449 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2450 2451 { PreserveJVMState pjvms(this); 2452 set_control(excp); 2453 set_i_o(i_o); 2454 2455 if (excp != top()) { 2456 // Create an exception state also. 2457 // Use an exact type if the caller has specified a specific exception. 2458 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2459 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2460 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2461 } 2462 } 2463 2464 // Get the no-exception control from the CatchNode. 2465 set_control(norm); 2466 } 2467 2468 2469 //-------------------------------gen_subtype_check----------------------------- 2470 // Generate a subtyping check. Takes as input the subtype and supertype. 2471 // Returns 2 values: sets the default control() to the true path and returns 2472 // the false path. Only reads invariant memory; sets no (visible) memory. 2473 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2474 // but that's not exposed to the optimizer. This call also doesn't take in an 2475 // Object; if you wish to check an Object you need to load the Object's class 2476 // prior to coming here. 2477 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2478 // Fast check for identical types, perhaps identical constants. 2479 // The types can even be identical non-constants, in cases 2480 // involving Array.newInstance, Object.clone, etc. 2481 if (subklass == superklass) 2482 return top(); // false path is dead; no test needed. 2483 2484 if (_gvn.type(superklass)->singleton()) { 2485 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2486 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2487 2488 // In the common case of an exact superklass, try to fold up the 2489 // test before generating code. You may ask, why not just generate 2490 // the code and then let it fold up? The answer is that the generated 2491 // code will necessarily include null checks, which do not always 2492 // completely fold away. If they are also needless, then they turn 2493 // into a performance loss. Example: 2494 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2495 // Here, the type of 'fa' is often exact, so the store check 2496 // of fa[1]=x will fold up, without testing the nullness of x. 2497 switch (static_subtype_check(superk, subk)) { 2498 case SSC_always_false: 2499 { 2500 Node* always_fail = control(); 2501 set_control(top()); 2502 return always_fail; 2503 } 2504 case SSC_always_true: 2505 return top(); 2506 case SSC_easy_test: 2507 { 2508 // Just do a direct pointer compare and be done. 2509 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2510 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2511 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2512 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2513 return _gvn.transform( new(C) IfFalseNode(iff) ); 2514 } 2515 case SSC_full_test: 2516 break; 2517 default: 2518 ShouldNotReachHere(); 2519 } 2520 } 2521 2522 // %%% Possible further optimization: Even if the superklass is not exact, 2523 // if the subklass is the unique subtype of the superklass, the check 2524 // will always succeed. We could leave a dependency behind to ensure this. 2525 2526 // First load the super-klass's check-offset 2527 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2528 Node *chk_off = _gvn.transform( new (C) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2529 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2530 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2531 2532 // Load from the sub-klass's super-class display list, or a 1-word cache of 2533 // the secondary superclass list, or a failing value with a sentinel offset 2534 // if the super-klass is an interface or exceptionally deep in the Java 2535 // hierarchy and we have to scan the secondary superclass list the hard way. 2536 // Worst-case type is a little odd: NULL is allowed as a result (usually 2537 // klass loads can never produce a NULL). 2538 Node *chk_off_X = ConvI2X(chk_off); 2539 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2540 // For some types like interfaces the following loadKlass is from a 1-word 2541 // cache which is mutable so can't use immutable memory. Other 2542 // types load from the super-class display table which is immutable. 2543 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2544 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2545 2546 // Compile speed common case: ARE a subtype and we canNOT fail 2547 if( superklass == nkls ) 2548 return top(); // false path is dead; no test needed. 2549 2550 // See if we get an immediate positive hit. Happens roughly 83% of the 2551 // time. Test to see if the value loaded just previously from the subklass 2552 // is exactly the superklass. 2553 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2554 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2555 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2556 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2557 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2558 2559 // Compile speed common case: Check for being deterministic right now. If 2560 // chk_off is a constant and not equal to cacheoff then we are NOT a 2561 // subklass. In this case we need exactly the 1 test above and we can 2562 // return those results immediately. 2563 if (!might_be_cache) { 2564 Node* not_subtype_ctrl = control(); 2565 set_control(iftrue1); // We need exactly the 1 test above 2566 return not_subtype_ctrl; 2567 } 2568 2569 // Gather the various success & failures here 2570 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2571 record_for_igvn(r_ok_subtype); 2572 RegionNode *r_not_subtype = new (C) RegionNode(3); 2573 record_for_igvn(r_not_subtype); 2574 2575 r_ok_subtype->init_req(1, iftrue1); 2576 2577 // Check for immediate negative hit. Happens roughly 11% of the time (which 2578 // is roughly 63% of the remaining cases). Test to see if the loaded 2579 // check-offset points into the subklass display list or the 1-element 2580 // cache. If it points to the display (and NOT the cache) and the display 2581 // missed then it's not a subtype. 2582 Node *cacheoff = _gvn.intcon(cacheoff_con); 2583 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2584 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2585 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2586 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2587 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2588 2589 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2590 // No performance impact (too rare) but allows sharing of secondary arrays 2591 // which has some footprint reduction. 2592 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2593 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2594 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2595 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2596 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2597 2598 // -- Roads not taken here: -- 2599 // We could also have chosen to perform the self-check at the beginning 2600 // of this code sequence, as the assembler does. This would not pay off 2601 // the same way, since the optimizer, unlike the assembler, can perform 2602 // static type analysis to fold away many successful self-checks. 2603 // Non-foldable self checks work better here in second position, because 2604 // the initial primary superclass check subsumes a self-check for most 2605 // types. An exception would be a secondary type like array-of-interface, 2606 // which does not appear in its own primary supertype display. 2607 // Finally, we could have chosen to move the self-check into the 2608 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2609 // dependent manner. But it is worthwhile to have the check here, 2610 // where it can be perhaps be optimized. The cost in code space is 2611 // small (register compare, branch). 2612 2613 // Now do a linear scan of the secondary super-klass array. Again, no real 2614 // performance impact (too rare) but it's gotta be done. 2615 // Since the code is rarely used, there is no penalty for moving it 2616 // out of line, and it can only improve I-cache density. 2617 // The decision to inline or out-of-line this final check is platform 2618 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2619 Node* psc = _gvn.transform( 2620 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2621 2622 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2623 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2624 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2625 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2626 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2627 2628 // Return false path; set default control to true path. 2629 set_control( _gvn.transform(r_ok_subtype) ); 2630 return _gvn.transform(r_not_subtype); 2631 } 2632 2633 //----------------------------static_subtype_check----------------------------- 2634 // Shortcut important common cases when superklass is exact: 2635 // (0) superklass is java.lang.Object (can occur in reflective code) 2636 // (1) subklass is already limited to a subtype of superklass => always ok 2637 // (2) subklass does not overlap with superklass => always fail 2638 // (3) superklass has NO subtypes and we can check with a simple compare. 2639 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2640 if (StressReflectiveCode) { 2641 return SSC_full_test; // Let caller generate the general case. 2642 } 2643 2644 if (superk == env()->Object_klass()) { 2645 return SSC_always_true; // (0) this test cannot fail 2646 } 2647 2648 ciType* superelem = superk; 2649 if (superelem->is_array_klass()) 2650 superelem = superelem->as_array_klass()->base_element_type(); 2651 2652 if (!subk->is_interface()) { // cannot trust static interface types yet 2653 if (subk->is_subtype_of(superk)) { 2654 return SSC_always_true; // (1) false path dead; no dynamic test needed 2655 } 2656 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2657 !superk->is_subtype_of(subk)) { 2658 return SSC_always_false; 2659 } 2660 } 2661 2662 // If casting to an instance klass, it must have no subtypes 2663 if (superk->is_interface()) { 2664 // Cannot trust interfaces yet. 2665 // %%% S.B. superk->nof_implementors() == 1 2666 } else if (superelem->is_instance_klass()) { 2667 ciInstanceKlass* ik = superelem->as_instance_klass(); 2668 if (!ik->has_subklass() && !ik->is_interface()) { 2669 if (!ik->is_final()) { 2670 // Add a dependency if there is a chance of a later subclass. 2671 C->dependencies()->assert_leaf_type(ik); 2672 } 2673 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2674 } 2675 } else { 2676 // A primitive array type has no subtypes. 2677 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2678 } 2679 2680 return SSC_full_test; 2681 } 2682 2683 // Profile-driven exact type check: 2684 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2685 float prob, 2686 Node* *casted_receiver) { 2687 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2688 Node* recv_klass = load_object_klass(receiver); 2689 Node* want_klass = makecon(tklass); 2690 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2691 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2692 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2693 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2694 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2695 2696 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2697 assert(recv_xtype->klass_is_exact(), ""); 2698 2699 // Subsume downstream occurrences of receiver with a cast to 2700 // recv_xtype, since now we know what the type will be. 2701 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2702 (*casted_receiver) = _gvn.transform(cast); 2703 // (User must make the replace_in_map call.) 2704 2705 return fail; 2706 } 2707 2708 2709 //------------------------------seems_never_null------------------------------- 2710 // Use null_seen information if it is available from the profile. 2711 // If we see an unexpected null at a type check we record it and force a 2712 // recompile; the offending check will be recompiled to handle NULLs. 2713 // If we see several offending BCIs, then all checks in the 2714 // method will be recompiled. 2715 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2716 if (UncommonNullCast // Cutout for this technique 2717 && obj != null() // And not the -Xcomp stupid case? 2718 && !too_many_traps(Deoptimization::Reason_null_check) 2719 ) { 2720 if (data == NULL) 2721 // Edge case: no mature data. Be optimistic here. 2722 return true; 2723 // If the profile has not seen a null, assume it won't happen. 2724 assert(java_bc() == Bytecodes::_checkcast || 2725 java_bc() == Bytecodes::_instanceof || 2726 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2727 return !data->as_BitData()->null_seen(); 2728 } 2729 return false; 2730 } 2731 2732 //------------------------maybe_cast_profiled_receiver------------------------- 2733 // If the profile has seen exactly one type, narrow to exactly that type. 2734 // Subsequent type checks will always fold up. 2735 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2736 ciKlass* require_klass, 2737 ciKlass* spec_klass, 2738 bool safe_for_replace) { 2739 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2740 2741 // Make sure we haven't already deoptimized from this tactic. 2742 if (too_many_traps(Deoptimization::Reason_class_check)) 2743 return NULL; 2744 2745 // (No, this isn't a call, but it's enough like a virtual call 2746 // to use the same ciMethod accessor to get the profile info...) 2747 // If we have a speculative type use it instead of profiling (which 2748 // may not help us) 2749 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2750 if (exact_kls != NULL) {// no cast failures here 2751 if (require_klass == NULL || 2752 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2753 // If we narrow the type to match what the type profile sees or 2754 // the speculative type, we can then remove the rest of the 2755 // cast. 2756 // This is a win, even if the exact_kls is very specific, 2757 // because downstream operations, such as method calls, 2758 // will often benefit from the sharper type. 2759 Node* exact_obj = not_null_obj; // will get updated in place... 2760 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2761 &exact_obj); 2762 { PreserveJVMState pjvms(this); 2763 set_control(slow_ctl); 2764 uncommon_trap(Deoptimization::Reason_class_check, 2765 Deoptimization::Action_maybe_recompile); 2766 } 2767 if (safe_for_replace) { 2768 replace_in_map(not_null_obj, exact_obj); 2769 } 2770 return exact_obj; 2771 } 2772 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2773 } 2774 2775 return NULL; 2776 } 2777 2778 /** 2779 * Cast obj to type and emit guard unless we had too many traps here 2780 * already 2781 * 2782 * @param obj node being casted 2783 * @param type type to cast the node to 2784 * @param not_null true if we know node cannot be null 2785 */ 2786 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2787 ciKlass* type, 2788 bool not_null) { 2789 // type == NULL if profiling tells us this object is always null 2790 if (type != NULL) { 2791 if (!too_many_traps(Deoptimization::Reason_null_check) && 2792 !too_many_traps(Deoptimization::Reason_class_check)) { 2793 Node* not_null_obj = NULL; 2794 // not_null is true if we know the object is not null and 2795 // there's no need for a null check 2796 if (!not_null) { 2797 Node* null_ctl = top(); 2798 not_null_obj = null_check_oop(obj, &null_ctl, true, true); 2799 assert(null_ctl->is_top(), "no null control here"); 2800 } else { 2801 not_null_obj = obj; 2802 } 2803 2804 Node* exact_obj = not_null_obj; 2805 ciKlass* exact_kls = type; 2806 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2807 &exact_obj); 2808 { 2809 PreserveJVMState pjvms(this); 2810 set_control(slow_ctl); 2811 uncommon_trap(Deoptimization::Reason_class_check, 2812 Deoptimization::Action_maybe_recompile); 2813 } 2814 replace_in_map(not_null_obj, exact_obj); 2815 obj = exact_obj; 2816 } 2817 } else { 2818 if (!too_many_traps(Deoptimization::Reason_null_assert)) { 2819 Node* exact_obj = null_assert(obj); 2820 replace_in_map(obj, exact_obj); 2821 obj = exact_obj; 2822 } 2823 } 2824 return obj; 2825 } 2826 2827 //-------------------------------gen_instanceof-------------------------------- 2828 // Generate an instance-of idiom. Used by both the instance-of bytecode 2829 // and the reflective instance-of call. 2830 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2831 kill_dead_locals(); // Benefit all the uncommon traps 2832 assert( !stopped(), "dead parse path should be checked in callers" ); 2833 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2834 "must check for not-null not-dead klass in callers"); 2835 2836 // Make the merge point 2837 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2838 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2839 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2840 C->set_has_split_ifs(true); // Has chance for split-if optimization 2841 2842 ciProfileData* data = NULL; 2843 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2844 data = method()->method_data()->bci_to_data(bci()); 2845 } 2846 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2847 && seems_never_null(obj, data)); 2848 2849 // Null check; get casted pointer; set region slot 3 2850 Node* null_ctl = top(); 2851 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2852 2853 // If not_null_obj is dead, only null-path is taken 2854 if (stopped()) { // Doing instance-of on a NULL? 2855 set_control(null_ctl); 2856 return intcon(0); 2857 } 2858 region->init_req(_null_path, null_ctl); 2859 phi ->init_req(_null_path, intcon(0)); // Set null path value 2860 if (null_ctl == top()) { 2861 // Do this eagerly, so that pattern matches like is_diamond_phi 2862 // will work even during parsing. 2863 assert(_null_path == PATH_LIMIT-1, "delete last"); 2864 region->del_req(_null_path); 2865 phi ->del_req(_null_path); 2866 } 2867 2868 // Do we know the type check always succeed? 2869 bool known_statically = false; 2870 if (_gvn.type(superklass)->singleton()) { 2871 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2872 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2873 if (subk != NULL && subk->is_loaded()) { 2874 int static_res = static_subtype_check(superk, subk); 2875 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2876 } 2877 } 2878 2879 if (known_statically && UseTypeSpeculation) { 2880 // If we know the type check always succeed then we don't use the 2881 // profiling data at this bytecode. Don't lose it, feed it to the 2882 // type system as a speculative type. 2883 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2884 } else { 2885 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2886 // We may not have profiling here or it may not help us. If we 2887 // have a speculative type use it to perform an exact cast. 2888 ciKlass* spec_obj_type = obj_type->speculative_type(); 2889 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2890 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2891 if (stopped()) { // Profile disagrees with this path. 2892 set_control(null_ctl); // Null is the only remaining possibility. 2893 return intcon(0); 2894 } 2895 if (cast_obj != NULL) { 2896 not_null_obj = cast_obj; 2897 } 2898 } 2899 } 2900 2901 // Load the object's klass 2902 Node* obj_klass = load_object_klass(not_null_obj); 2903 2904 // Generate the subtype check 2905 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2906 2907 // Plug in the success path to the general merge in slot 1. 2908 region->init_req(_obj_path, control()); 2909 phi ->init_req(_obj_path, intcon(1)); 2910 2911 // Plug in the failing path to the general merge in slot 2. 2912 region->init_req(_fail_path, not_subtype_ctrl); 2913 phi ->init_req(_fail_path, intcon(0)); 2914 2915 // Return final merged results 2916 set_control( _gvn.transform(region) ); 2917 record_for_igvn(region); 2918 return _gvn.transform(phi); 2919 } 2920 2921 //-------------------------------gen_checkcast--------------------------------- 2922 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2923 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2924 // uncommon-trap paths work. Adjust stack after this call. 2925 // If failure_control is supplied and not null, it is filled in with 2926 // the control edge for the cast failure. Otherwise, an appropriate 2927 // uncommon trap or exception is thrown. 2928 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2929 Node* *failure_control) { 2930 kill_dead_locals(); // Benefit all the uncommon traps 2931 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2932 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2933 2934 // Fast cutout: Check the case that the cast is vacuously true. 2935 // This detects the common cases where the test will short-circuit 2936 // away completely. We do this before we perform the null check, 2937 // because if the test is going to turn into zero code, we don't 2938 // want a residual null check left around. (Causes a slowdown, 2939 // for example, in some objArray manipulations, such as a[i]=a[j].) 2940 if (tk->singleton()) { 2941 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2942 if (objtp != NULL && objtp->klass() != NULL) { 2943 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2944 case SSC_always_true: 2945 // If we know the type check always succeed then we don't use 2946 // the profiling data at this bytecode. Don't lose it, feed it 2947 // to the type system as a speculative type. 2948 return record_profiled_receiver_for_speculation(obj); 2949 case SSC_always_false: 2950 // It needs a null check because a null will *pass* the cast check. 2951 // A non-null value will always produce an exception. 2952 return null_assert(obj); 2953 } 2954 } 2955 } 2956 2957 ciProfileData* data = NULL; 2958 bool safe_for_replace = false; 2959 if (failure_control == NULL) { // use MDO in regular case only 2960 assert(java_bc() == Bytecodes::_aastore || 2961 java_bc() == Bytecodes::_checkcast, 2962 "interpreter profiles type checks only for these BCs"); 2963 data = method()->method_data()->bci_to_data(bci()); 2964 safe_for_replace = true; 2965 } 2966 2967 // Make the merge point 2968 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2969 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 2970 Node* phi = new (C) PhiNode(region, toop); 2971 C->set_has_split_ifs(true); // Has chance for split-if optimization 2972 2973 // Use null-cast information if it is available 2974 bool never_see_null = ((failure_control == NULL) // regular case only 2975 && seems_never_null(obj, data)); 2976 2977 // Null check; get casted pointer; set region slot 3 2978 Node* null_ctl = top(); 2979 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2980 2981 // If not_null_obj is dead, only null-path is taken 2982 if (stopped()) { // Doing instance-of on a NULL? 2983 set_control(null_ctl); 2984 return null(); 2985 } 2986 region->init_req(_null_path, null_ctl); 2987 phi ->init_req(_null_path, null()); // Set null path value 2988 if (null_ctl == top()) { 2989 // Do this eagerly, so that pattern matches like is_diamond_phi 2990 // will work even during parsing. 2991 assert(_null_path == PATH_LIMIT-1, "delete last"); 2992 region->del_req(_null_path); 2993 phi ->del_req(_null_path); 2994 } 2995 2996 Node* cast_obj = NULL; 2997 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2998 // We may not have profiling here or it may not help us. If we have 2999 // a speculative type use it to perform an exact cast. 3000 ciKlass* spec_obj_type = obj_type->speculative_type(); 3001 if (spec_obj_type != NULL || 3002 (data != NULL && 3003 // Counter has never been decremented (due to cast failure). 3004 // ...This is a reasonable thing to expect. It is true of 3005 // all casts inserted by javac to implement generic types. 3006 data->as_CounterData()->count() >= 0)) { 3007 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3008 if (cast_obj != NULL) { 3009 if (failure_control != NULL) // failure is now impossible 3010 (*failure_control) = top(); 3011 // adjust the type of the phi to the exact klass: 3012 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR)); 3013 } 3014 } 3015 3016 if (cast_obj == NULL) { 3017 // Load the object's klass 3018 Node* obj_klass = load_object_klass(not_null_obj); 3019 3020 // Generate the subtype check 3021 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3022 3023 // Plug in success path into the merge 3024 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 3025 not_null_obj, toop)); 3026 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3027 if (failure_control == NULL) { 3028 if (not_subtype_ctrl != top()) { // If failure is possible 3029 PreserveJVMState pjvms(this); 3030 set_control(not_subtype_ctrl); 3031 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3032 } 3033 } else { 3034 (*failure_control) = not_subtype_ctrl; 3035 } 3036 } 3037 3038 region->init_req(_obj_path, control()); 3039 phi ->init_req(_obj_path, cast_obj); 3040 3041 // A merge of NULL or Casted-NotNull obj 3042 Node* res = _gvn.transform(phi); 3043 3044 // Note I do NOT always 'replace_in_map(obj,result)' here. 3045 // if( tk->klass()->can_be_primary_super() ) 3046 // This means that if I successfully store an Object into an array-of-String 3047 // I 'forget' that the Object is really now known to be a String. I have to 3048 // do this because we don't have true union types for interfaces - if I store 3049 // a Baz into an array-of-Interface and then tell the optimizer it's an 3050 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3051 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3052 // replace_in_map( obj, res ); 3053 3054 // Return final merged results 3055 set_control( _gvn.transform(region) ); 3056 record_for_igvn(region); 3057 return res; 3058 } 3059 3060 //------------------------------next_monitor----------------------------------- 3061 // What number should be given to the next monitor? 3062 int GraphKit::next_monitor() { 3063 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3064 int next = current + C->sync_stack_slots(); 3065 // Keep the toplevel high water mark current: 3066 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3067 return current; 3068 } 3069 3070 //------------------------------insert_mem_bar--------------------------------- 3071 // Memory barrier to avoid floating things around 3072 // The membar serves as a pinch point between both control and all memory slices. 3073 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3074 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3075 mb->init_req(TypeFunc::Control, control()); 3076 mb->init_req(TypeFunc::Memory, reset_memory()); 3077 Node* membar = _gvn.transform(mb); 3078 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3079 set_all_memory_call(membar); 3080 return membar; 3081 } 3082 3083 //-------------------------insert_mem_bar_volatile---------------------------- 3084 // Memory barrier to avoid floating things around 3085 // The membar serves as a pinch point between both control and memory(alias_idx). 3086 // If you want to make a pinch point on all memory slices, do not use this 3087 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3088 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3089 // When Parse::do_put_xxx updates a volatile field, it appends a series 3090 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3091 // The first membar is on the same memory slice as the field store opcode. 3092 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3093 // All the other membars (for other volatile slices, including AliasIdxBot, 3094 // which stands for all unknown volatile slices) are control-dependent 3095 // on the first membar. This prevents later volatile loads or stores 3096 // from sliding up past the just-emitted store. 3097 3098 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3099 mb->set_req(TypeFunc::Control,control()); 3100 if (alias_idx == Compile::AliasIdxBot) { 3101 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3102 } else { 3103 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3104 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3105 } 3106 Node* membar = _gvn.transform(mb); 3107 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3108 if (alias_idx == Compile::AliasIdxBot) { 3109 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 3110 } else { 3111 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 3112 } 3113 return membar; 3114 } 3115 3116 //------------------------------shared_lock------------------------------------ 3117 // Emit locking code. 3118 FastLockNode* GraphKit::shared_lock(Node* obj) { 3119 // bci is either a monitorenter bc or InvocationEntryBci 3120 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3121 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3122 3123 if( !GenerateSynchronizationCode ) 3124 return NULL; // Not locking things? 3125 if (stopped()) // Dead monitor? 3126 return NULL; 3127 3128 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3129 3130 // Box the stack location 3131 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 3132 Node* mem = reset_memory(); 3133 3134 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 3135 if (PrintPreciseBiasedLockingStatistics) { 3136 // Create the counters for this fast lock. 3137 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3138 } 3139 // Add monitor to debug info for the slow path. If we block inside the 3140 // slow path and de-opt, we need the monitor hanging around 3141 map()->push_monitor( flock ); 3142 3143 const TypeFunc *tf = LockNode::lock_type(); 3144 LockNode *lock = new (C) LockNode(C, tf); 3145 3146 lock->init_req( TypeFunc::Control, control() ); 3147 lock->init_req( TypeFunc::Memory , mem ); 3148 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3149 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3150 lock->init_req( TypeFunc::ReturnAdr, top() ); 3151 3152 lock->init_req(TypeFunc::Parms + 0, obj); 3153 lock->init_req(TypeFunc::Parms + 1, box); 3154 lock->init_req(TypeFunc::Parms + 2, flock); 3155 add_safepoint_edges(lock); 3156 3157 lock = _gvn.transform( lock )->as_Lock(); 3158 3159 // lock has no side-effects, sets few values 3160 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3161 3162 insert_mem_bar(Op_MemBarAcquireLock); 3163 3164 // Add this to the worklist so that the lock can be eliminated 3165 record_for_igvn(lock); 3166 3167 #ifndef PRODUCT 3168 if (PrintLockStatistics) { 3169 // Update the counter for this lock. Don't bother using an atomic 3170 // operation since we don't require absolute accuracy. 3171 lock->create_lock_counter(map()->jvms()); 3172 increment_counter(lock->counter()->addr()); 3173 } 3174 #endif 3175 3176 return flock; 3177 } 3178 3179 3180 //------------------------------shared_unlock---------------------------------- 3181 // Emit unlocking code. 3182 void GraphKit::shared_unlock(Node* box, Node* obj) { 3183 // bci is either a monitorenter bc or InvocationEntryBci 3184 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3185 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3186 3187 if( !GenerateSynchronizationCode ) 3188 return; 3189 if (stopped()) { // Dead monitor? 3190 map()->pop_monitor(); // Kill monitor from debug info 3191 return; 3192 } 3193 3194 // Memory barrier to avoid floating things down past the locked region 3195 insert_mem_bar(Op_MemBarReleaseLock); 3196 3197 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3198 UnlockNode *unlock = new (C) UnlockNode(C, tf); 3199 uint raw_idx = Compile::AliasIdxRaw; 3200 unlock->init_req( TypeFunc::Control, control() ); 3201 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3202 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3203 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3204 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3205 3206 unlock->init_req(TypeFunc::Parms + 0, obj); 3207 unlock->init_req(TypeFunc::Parms + 1, box); 3208 unlock = _gvn.transform(unlock)->as_Unlock(); 3209 3210 Node* mem = reset_memory(); 3211 3212 // unlock has no side-effects, sets few values 3213 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3214 3215 // Kill monitor from debug info 3216 map()->pop_monitor( ); 3217 } 3218 3219 //-------------------------------get_layout_helper----------------------------- 3220 // If the given klass is a constant or known to be an array, 3221 // fetch the constant layout helper value into constant_value 3222 // and return (Node*)NULL. Otherwise, load the non-constant 3223 // layout helper value, and return the node which represents it. 3224 // This two-faced routine is useful because allocation sites 3225 // almost always feature constant types. 3226 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3227 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3228 if (!StressReflectiveCode && inst_klass != NULL) { 3229 ciKlass* klass = inst_klass->klass(); 3230 bool xklass = inst_klass->klass_is_exact(); 3231 if (xklass || klass->is_array_klass()) { 3232 jint lhelper = klass->layout_helper(); 3233 if (lhelper != Klass::_lh_neutral_value) { 3234 constant_value = lhelper; 3235 return (Node*) NULL; 3236 } 3237 } 3238 } 3239 constant_value = Klass::_lh_neutral_value; // put in a known value 3240 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3241 return make_load(NULL, lhp, TypeInt::INT, T_INT); 3242 } 3243 3244 // We just put in an allocate/initialize with a big raw-memory effect. 3245 // Hook selected additional alias categories on the initialization. 3246 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3247 MergeMemNode* init_in_merge, 3248 Node* init_out_raw) { 3249 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3250 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3251 3252 Node* prevmem = kit.memory(alias_idx); 3253 init_in_merge->set_memory_at(alias_idx, prevmem); 3254 kit.set_memory(init_out_raw, alias_idx); 3255 } 3256 3257 //---------------------------set_output_for_allocation------------------------- 3258 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3259 const TypeOopPtr* oop_type) { 3260 int rawidx = Compile::AliasIdxRaw; 3261 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3262 add_safepoint_edges(alloc); 3263 Node* allocx = _gvn.transform(alloc); 3264 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3265 // create memory projection for i_o 3266 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3267 make_slow_call_ex(allocx, env()->Throwable_klass(), true); 3268 3269 // create a memory projection as for the normal control path 3270 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3271 set_memory(malloc, rawidx); 3272 3273 // a normal slow-call doesn't change i_o, but an allocation does 3274 // we create a separate i_o projection for the normal control path 3275 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3276 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3277 3278 // put in an initialization barrier 3279 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3280 rawoop)->as_Initialize(); 3281 assert(alloc->initialization() == init, "2-way macro link must work"); 3282 assert(init ->allocation() == alloc, "2-way macro link must work"); 3283 { 3284 // Extract memory strands which may participate in the new object's 3285 // initialization, and source them from the new InitializeNode. 3286 // This will allow us to observe initializations when they occur, 3287 // and link them properly (as a group) to the InitializeNode. 3288 assert(init->in(InitializeNode::Memory) == malloc, ""); 3289 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3290 init->set_req(InitializeNode::Memory, minit_in); 3291 record_for_igvn(minit_in); // fold it up later, if possible 3292 Node* minit_out = memory(rawidx); 3293 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3294 if (oop_type->isa_aryptr()) { 3295 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3296 int elemidx = C->get_alias_index(telemref); 3297 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3298 } else if (oop_type->isa_instptr()) { 3299 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3300 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3301 ciField* field = ik->nonstatic_field_at(i); 3302 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3303 continue; // do not bother to track really large numbers of fields 3304 // Find (or create) the alias category for this field: 3305 int fieldidx = C->alias_type(field)->index(); 3306 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3307 } 3308 } 3309 } 3310 3311 // Cast raw oop to the real thing... 3312 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3313 javaoop = _gvn.transform(javaoop); 3314 C->set_recent_alloc(control(), javaoop); 3315 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3316 3317 #ifdef ASSERT 3318 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3319 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3320 "Ideal_allocation works"); 3321 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3322 "Ideal_allocation works"); 3323 if (alloc->is_AllocateArray()) { 3324 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3325 "Ideal_allocation works"); 3326 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3327 "Ideal_allocation works"); 3328 } else { 3329 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3330 } 3331 } 3332 #endif //ASSERT 3333 3334 return javaoop; 3335 } 3336 3337 //---------------------------new_instance-------------------------------------- 3338 // This routine takes a klass_node which may be constant (for a static type) 3339 // or may be non-constant (for reflective code). It will work equally well 3340 // for either, and the graph will fold nicely if the optimizer later reduces 3341 // the type to a constant. 3342 // The optional arguments are for specialized use by intrinsics: 3343 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3344 // - If 'return_size_val', report the the total object size to the caller. 3345 Node* GraphKit::new_instance(Node* klass_node, 3346 Node* extra_slow_test, 3347 Node* *return_size_val) { 3348 // Compute size in doublewords 3349 // The size is always an integral number of doublewords, represented 3350 // as a positive bytewise size stored in the klass's layout_helper. 3351 // The layout_helper also encodes (in a low bit) the need for a slow path. 3352 jint layout_con = Klass::_lh_neutral_value; 3353 Node* layout_val = get_layout_helper(klass_node, layout_con); 3354 int layout_is_con = (layout_val == NULL); 3355 3356 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3357 // Generate the initial go-slow test. It's either ALWAYS (return a 3358 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3359 // case) a computed value derived from the layout_helper. 3360 Node* initial_slow_test = NULL; 3361 if (layout_is_con) { 3362 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3363 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3364 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3365 3366 } else { // reflective case 3367 // This reflective path is used by Unsafe.allocateInstance. 3368 // (It may be stress-tested by specifying StressReflectiveCode.) 3369 // Basically, we want to get into the VM is there's an illegal argument. 3370 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3371 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3372 if (extra_slow_test != intcon(0)) { 3373 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3374 } 3375 // (Macro-expander will further convert this to a Bool, if necessary.) 3376 } 3377 3378 // Find the size in bytes. This is easy; it's the layout_helper. 3379 // The size value must be valid even if the slow path is taken. 3380 Node* size = NULL; 3381 if (layout_is_con) { 3382 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3383 } else { // reflective case 3384 // This reflective path is used by clone and Unsafe.allocateInstance. 3385 size = ConvI2X(layout_val); 3386 3387 // Clear the low bits to extract layout_helper_size_in_bytes: 3388 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3389 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3390 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3391 } 3392 if (return_size_val != NULL) { 3393 (*return_size_val) = size; 3394 } 3395 3396 // This is a precise notnull oop of the klass. 3397 // (Actually, it need not be precise if this is a reflective allocation.) 3398 // It's what we cast the result to. 3399 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3400 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3401 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3402 3403 // Now generate allocation code 3404 3405 // The entire memory state is needed for slow path of the allocation 3406 // since GC and deoptimization can happened. 3407 Node *mem = reset_memory(); 3408 set_all_memory(mem); // Create new memory state 3409 3410 AllocateNode* alloc 3411 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3412 control(), mem, i_o(), 3413 size, klass_node, 3414 initial_slow_test); 3415 3416 return set_output_for_allocation(alloc, oop_type); 3417 } 3418 3419 //-------------------------------new_array------------------------------------- 3420 // helper for both newarray and anewarray 3421 // The 'length' parameter is (obviously) the length of the array. 3422 // See comments on new_instance for the meaning of the other arguments. 3423 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3424 Node* length, // number of array elements 3425 int nargs, // number of arguments to push back for uncommon trap 3426 Node* *return_size_val) { 3427 jint layout_con = Klass::_lh_neutral_value; 3428 Node* layout_val = get_layout_helper(klass_node, layout_con); 3429 int layout_is_con = (layout_val == NULL); 3430 3431 if (!layout_is_con && !StressReflectiveCode && 3432 !too_many_traps(Deoptimization::Reason_class_check)) { 3433 // This is a reflective array creation site. 3434 // Optimistically assume that it is a subtype of Object[], 3435 // so that we can fold up all the address arithmetic. 3436 layout_con = Klass::array_layout_helper(T_OBJECT); 3437 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3438 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3439 { BuildCutout unless(this, bol_lh, PROB_MAX); 3440 inc_sp(nargs); 3441 uncommon_trap(Deoptimization::Reason_class_check, 3442 Deoptimization::Action_maybe_recompile); 3443 } 3444 layout_val = NULL; 3445 layout_is_con = true; 3446 } 3447 3448 // Generate the initial go-slow test. Make sure we do not overflow 3449 // if length is huge (near 2Gig) or negative! We do not need 3450 // exact double-words here, just a close approximation of needed 3451 // double-words. We can't add any offset or rounding bits, lest we 3452 // take a size -1 of bytes and make it positive. Use an unsigned 3453 // compare, so negative sizes look hugely positive. 3454 int fast_size_limit = FastAllocateSizeLimit; 3455 if (layout_is_con) { 3456 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3457 // Increase the size limit if we have exact knowledge of array type. 3458 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3459 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3460 } 3461 3462 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3463 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3464 if (initial_slow_test->is_Bool()) { 3465 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3466 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3467 } 3468 3469 // --- Size Computation --- 3470 // array_size = round_to_heap(array_header + (length << elem_shift)); 3471 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3472 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3473 // The rounding mask is strength-reduced, if possible. 3474 int round_mask = MinObjAlignmentInBytes - 1; 3475 Node* header_size = NULL; 3476 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3477 // (T_BYTE has the weakest alignment and size restrictions...) 3478 if (layout_is_con) { 3479 int hsize = Klass::layout_helper_header_size(layout_con); 3480 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3481 BasicType etype = Klass::layout_helper_element_type(layout_con); 3482 if ((round_mask & ~right_n_bits(eshift)) == 0) 3483 round_mask = 0; // strength-reduce it if it goes away completely 3484 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3485 assert(header_size_min <= hsize, "generic minimum is smallest"); 3486 header_size_min = hsize; 3487 header_size = intcon(hsize + round_mask); 3488 } else { 3489 Node* hss = intcon(Klass::_lh_header_size_shift); 3490 Node* hsm = intcon(Klass::_lh_header_size_mask); 3491 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3492 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3493 Node* mask = intcon(round_mask); 3494 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3495 } 3496 3497 Node* elem_shift = NULL; 3498 if (layout_is_con) { 3499 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3500 if (eshift != 0) 3501 elem_shift = intcon(eshift); 3502 } else { 3503 // There is no need to mask or shift this value. 3504 // The semantics of LShiftINode include an implicit mask to 0x1F. 3505 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3506 elem_shift = layout_val; 3507 } 3508 3509 // Transition to native address size for all offset calculations: 3510 Node* lengthx = ConvI2X(length); 3511 Node* headerx = ConvI2X(header_size); 3512 #ifdef _LP64 3513 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3514 if (tllen != NULL && tllen->_lo < 0) { 3515 // Add a manual constraint to a positive range. Cf. array_element_address. 3516 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3517 if (size_max > tllen->_hi) size_max = tllen->_hi; 3518 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3519 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon)); 3520 } 3521 } 3522 #endif 3523 3524 // Combine header size (plus rounding) and body size. Then round down. 3525 // This computation cannot overflow, because it is used only in two 3526 // places, one where the length is sharply limited, and the other 3527 // after a successful allocation. 3528 Node* abody = lengthx; 3529 if (elem_shift != NULL) 3530 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3531 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3532 if (round_mask != 0) { 3533 Node* mask = MakeConX(~round_mask); 3534 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3535 } 3536 // else if round_mask == 0, the size computation is self-rounding 3537 3538 if (return_size_val != NULL) { 3539 // This is the size 3540 (*return_size_val) = size; 3541 } 3542 3543 // Now generate allocation code 3544 3545 // The entire memory state is needed for slow path of the allocation 3546 // since GC and deoptimization can happened. 3547 Node *mem = reset_memory(); 3548 set_all_memory(mem); // Create new memory state 3549 3550 // Create the AllocateArrayNode and its result projections 3551 AllocateArrayNode* alloc 3552 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3553 control(), mem, i_o(), 3554 size, klass_node, 3555 initial_slow_test, 3556 length); 3557 3558 // Cast to correct type. Note that the klass_node may be constant or not, 3559 // and in the latter case the actual array type will be inexact also. 3560 // (This happens via a non-constant argument to inline_native_newArray.) 3561 // In any case, the value of klass_node provides the desired array type. 3562 const TypeInt* length_type = _gvn.find_int_type(length); 3563 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3564 if (ary_type->isa_aryptr() && length_type != NULL) { 3565 // Try to get a better type than POS for the size 3566 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3567 } 3568 3569 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3570 3571 // Cast length on remaining path to be as narrow as possible 3572 if (map()->find_edge(length) >= 0) { 3573 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3574 if (ccast != length) { 3575 _gvn.set_type_bottom(ccast); 3576 record_for_igvn(ccast); 3577 replace_in_map(length, ccast); 3578 } 3579 } 3580 3581 return javaoop; 3582 } 3583 3584 // The following "Ideal_foo" functions are placed here because they recognize 3585 // the graph shapes created by the functions immediately above. 3586 3587 //---------------------------Ideal_allocation---------------------------------- 3588 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3589 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3590 if (ptr == NULL) { // reduce dumb test in callers 3591 return NULL; 3592 } 3593 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3594 ptr = ptr->in(1); 3595 if (ptr == NULL) return NULL; 3596 } 3597 // Return NULL for allocations with several casts: 3598 // j.l.reflect.Array.newInstance(jobject, jint) 3599 // Object.clone() 3600 // to keep more precise type from last cast. 3601 if (ptr->is_Proj()) { 3602 Node* allo = ptr->in(0); 3603 if (allo != NULL && allo->is_Allocate()) { 3604 return allo->as_Allocate(); 3605 } 3606 } 3607 // Report failure to match. 3608 return NULL; 3609 } 3610 3611 // Fancy version which also strips off an offset (and reports it to caller). 3612 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3613 intptr_t& offset) { 3614 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3615 if (base == NULL) return NULL; 3616 return Ideal_allocation(base, phase); 3617 } 3618 3619 // Trace Initialize <- Proj[Parm] <- Allocate 3620 AllocateNode* InitializeNode::allocation() { 3621 Node* rawoop = in(InitializeNode::RawAddress); 3622 if (rawoop->is_Proj()) { 3623 Node* alloc = rawoop->in(0); 3624 if (alloc->is_Allocate()) { 3625 return alloc->as_Allocate(); 3626 } 3627 } 3628 return NULL; 3629 } 3630 3631 // Trace Allocate -> Proj[Parm] -> Initialize 3632 InitializeNode* AllocateNode::initialization() { 3633 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3634 if (rawoop == NULL) return NULL; 3635 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3636 Node* init = rawoop->fast_out(i); 3637 if (init->is_Initialize()) { 3638 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3639 return init->as_Initialize(); 3640 } 3641 } 3642 return NULL; 3643 } 3644 3645 //----------------------------- loop predicates --------------------------- 3646 3647 //------------------------------add_predicate_impl---------------------------- 3648 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3649 // Too many traps seen? 3650 if (too_many_traps(reason)) { 3651 #ifdef ASSERT 3652 if (TraceLoopPredicate) { 3653 int tc = C->trap_count(reason); 3654 tty->print("too many traps=%s tcount=%d in ", 3655 Deoptimization::trap_reason_name(reason), tc); 3656 method()->print(); // which method has too many predicate traps 3657 tty->cr(); 3658 } 3659 #endif 3660 // We cannot afford to take more traps here, 3661 // do not generate predicate. 3662 return; 3663 } 3664 3665 Node *cont = _gvn.intcon(1); 3666 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3667 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3668 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3669 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3670 C->add_predicate_opaq(opq); 3671 { 3672 PreserveJVMState pjvms(this); 3673 set_control(iffalse); 3674 inc_sp(nargs); 3675 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3676 } 3677 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3678 set_control(iftrue); 3679 } 3680 3681 //------------------------------add_predicate--------------------------------- 3682 void GraphKit::add_predicate(int nargs) { 3683 if (UseLoopPredicate) { 3684 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3685 } 3686 // loop's limit check predicate should be near the loop. 3687 if (LoopLimitCheck) { 3688 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3689 } 3690 } 3691 3692 //----------------------------- store barriers ---------------------------- 3693 #define __ ideal. 3694 3695 void GraphKit::sync_kit(IdealKit& ideal) { 3696 set_all_memory(__ merged_memory()); 3697 set_i_o(__ i_o()); 3698 set_control(__ ctrl()); 3699 } 3700 3701 void GraphKit::final_sync(IdealKit& ideal) { 3702 // Final sync IdealKit and graphKit. 3703 sync_kit(ideal); 3704 } 3705 3706 // vanilla/CMS post barrier 3707 // Insert a write-barrier store. This is to let generational GC work; we have 3708 // to flag all oop-stores before the next GC point. 3709 void GraphKit::write_barrier_post(Node* oop_store, 3710 Node* obj, 3711 Node* adr, 3712 uint adr_idx, 3713 Node* val, 3714 bool use_precise) { 3715 // No store check needed if we're storing a NULL or an old object 3716 // (latter case is probably a string constant). The concurrent 3717 // mark sweep garbage collector, however, needs to have all nonNull 3718 // oop updates flagged via card-marks. 3719 if (val != NULL && val->is_Con()) { 3720 // must be either an oop or NULL 3721 const Type* t = val->bottom_type(); 3722 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3723 // stores of null never (?) need barriers 3724 return; 3725 } 3726 3727 if (use_ReduceInitialCardMarks() 3728 && obj == just_allocated_object(control())) { 3729 // We can skip marks on a freshly-allocated object in Eden. 3730 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3731 // That routine informs GC to take appropriate compensating steps, 3732 // upon a slow-path allocation, so as to make this card-mark 3733 // elision safe. 3734 return; 3735 } 3736 3737 if (!use_precise) { 3738 // All card marks for a (non-array) instance are in one place: 3739 adr = obj; 3740 } 3741 // (Else it's an array (or unknown), and we want more precise card marks.) 3742 assert(adr != NULL, ""); 3743 3744 IdealKit ideal(this, true); 3745 3746 // Convert the pointer to an int prior to doing math on it 3747 Node* cast = __ CastPX(__ ctrl(), adr); 3748 3749 // Divide by card size 3750 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3751 "Only one we handle so far."); 3752 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3753 3754 // Combine card table base and card offset 3755 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3756 3757 // Get the alias_index for raw card-mark memory 3758 int adr_type = Compile::AliasIdxRaw; 3759 Node* zero = __ ConI(0); // Dirty card value 3760 BasicType bt = T_BYTE; 3761 3762 if (UseCondCardMark) { 3763 // The classic GC reference write barrier is typically implemented 3764 // as a store into the global card mark table. Unfortunately 3765 // unconditional stores can result in false sharing and excessive 3766 // coherence traffic as well as false transactional aborts. 3767 // UseCondCardMark enables MP "polite" conditional card mark 3768 // stores. In theory we could relax the load from ctrl() to 3769 // no_ctrl, but that doesn't buy much latitude. 3770 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3771 __ if_then(card_val, BoolTest::ne, zero); 3772 } 3773 3774 // Smash zero into card 3775 if( !UseConcMarkSweepGC ) { 3776 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3777 } else { 3778 // Specialized path for CM store barrier 3779 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3780 } 3781 3782 if (UseCondCardMark) { 3783 __ end_if(); 3784 } 3785 3786 // Final sync IdealKit and GraphKit. 3787 final_sync(ideal); 3788 } 3789 3790 // G1 pre/post barriers 3791 void GraphKit::g1_write_barrier_pre(bool do_load, 3792 Node* obj, 3793 Node* adr, 3794 uint alias_idx, 3795 Node* val, 3796 const TypeOopPtr* val_type, 3797 Node* pre_val, 3798 BasicType bt) { 3799 3800 // Some sanity checks 3801 // Note: val is unused in this routine. 3802 3803 if (do_load) { 3804 // We need to generate the load of the previous value 3805 assert(obj != NULL, "must have a base"); 3806 assert(adr != NULL, "where are loading from?"); 3807 assert(pre_val == NULL, "loaded already?"); 3808 assert(val_type != NULL, "need a type"); 3809 } else { 3810 // In this case both val_type and alias_idx are unused. 3811 assert(pre_val != NULL, "must be loaded already"); 3812 // Nothing to be done if pre_val is null. 3813 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3814 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3815 } 3816 assert(bt == T_OBJECT, "or we shouldn't be here"); 3817 3818 IdealKit ideal(this, true); 3819 3820 Node* tls = __ thread(); // ThreadLocalStorage 3821 3822 Node* no_ctrl = NULL; 3823 Node* no_base = __ top(); 3824 Node* zero = __ ConI(0); 3825 Node* zeroX = __ ConX(0); 3826 3827 float likely = PROB_LIKELY(0.999); 3828 float unlikely = PROB_UNLIKELY(0.999); 3829 3830 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3831 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3832 3833 // Offsets into the thread 3834 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3835 PtrQueue::byte_offset_of_active()); 3836 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3837 PtrQueue::byte_offset_of_index()); 3838 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3839 PtrQueue::byte_offset_of_buf()); 3840 3841 // Now the actual pointers into the thread 3842 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3843 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3844 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3845 3846 // Now some of the values 3847 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3848 3849 // if (!marking) 3850 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3851 BasicType index_bt = TypeX_X->basic_type(); 3852 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3853 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3854 3855 if (do_load) { 3856 // load original value 3857 // alias_idx correct?? 3858 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3859 } 3860 3861 // if (pre_val != NULL) 3862 __ if_then(pre_val, BoolTest::ne, null()); { 3863 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3864 3865 // is the queue for this thread full? 3866 __ if_then(index, BoolTest::ne, zeroX, likely); { 3867 3868 // decrement the index 3869 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3870 3871 // Now get the buffer location we will log the previous value into and store it 3872 Node *log_addr = __ AddP(no_base, buffer, next_index); 3873 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw); 3874 // update the index 3875 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw); 3876 3877 } __ else_(); { 3878 3879 // logging buffer is full, call the runtime 3880 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3881 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3882 } __ end_if(); // (!index) 3883 } __ end_if(); // (pre_val != NULL) 3884 } __ end_if(); // (!marking) 3885 3886 // Final sync IdealKit and GraphKit. 3887 final_sync(ideal); 3888 } 3889 3890 // 3891 // Update the card table and add card address to the queue 3892 // 3893 void GraphKit::g1_mark_card(IdealKit& ideal, 3894 Node* card_adr, 3895 Node* oop_store, 3896 uint oop_alias_idx, 3897 Node* index, 3898 Node* index_adr, 3899 Node* buffer, 3900 const TypeFunc* tf) { 3901 3902 Node* zero = __ ConI(0); 3903 Node* zeroX = __ ConX(0); 3904 Node* no_base = __ top(); 3905 BasicType card_bt = T_BYTE; 3906 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3907 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3908 3909 // Now do the queue work 3910 __ if_then(index, BoolTest::ne, zeroX); { 3911 3912 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3913 Node* log_addr = __ AddP(no_base, buffer, next_index); 3914 3915 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3916 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw); 3917 3918 } __ else_(); { 3919 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3920 } __ end_if(); 3921 3922 } 3923 3924 void GraphKit::g1_write_barrier_post(Node* oop_store, 3925 Node* obj, 3926 Node* adr, 3927 uint alias_idx, 3928 Node* val, 3929 BasicType bt, 3930 bool use_precise) { 3931 // If we are writing a NULL then we need no post barrier 3932 3933 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3934 // Must be NULL 3935 const Type* t = val->bottom_type(); 3936 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3937 // No post barrier if writing NULLx 3938 return; 3939 } 3940 3941 if (!use_precise) { 3942 // All card marks for a (non-array) instance are in one place: 3943 adr = obj; 3944 } 3945 // (Else it's an array (or unknown), and we want more precise card marks.) 3946 assert(adr != NULL, ""); 3947 3948 IdealKit ideal(this, true); 3949 3950 Node* tls = __ thread(); // ThreadLocalStorage 3951 3952 Node* no_base = __ top(); 3953 float likely = PROB_LIKELY(0.999); 3954 float unlikely = PROB_UNLIKELY(0.999); 3955 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 3956 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 3957 Node* zeroX = __ ConX(0); 3958 3959 // Get the alias_index for raw card-mark memory 3960 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3961 3962 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3963 3964 // Offsets into the thread 3965 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3966 PtrQueue::byte_offset_of_index()); 3967 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3968 PtrQueue::byte_offset_of_buf()); 3969 3970 // Pointers into the thread 3971 3972 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3973 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3974 3975 // Now some values 3976 // Use ctrl to avoid hoisting these values past a safepoint, which could 3977 // potentially reset these fields in the JavaThread. 3978 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 3979 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3980 3981 // Convert the store obj pointer to an int prior to doing math on it 3982 // Must use ctrl to prevent "integerized oop" existing across safepoint 3983 Node* cast = __ CastPX(__ ctrl(), adr); 3984 3985 // Divide pointer by card size 3986 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3987 3988 // Combine card table base and card offset 3989 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3990 3991 // If we know the value being stored does it cross regions? 3992 3993 if (val != NULL) { 3994 // Does the store cause us to cross regions? 3995 3996 // Should be able to do an unsigned compare of region_size instead of 3997 // and extra shift. Do we have an unsigned compare?? 3998 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3999 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4000 4001 // if (xor_res == 0) same region so skip 4002 __ if_then(xor_res, BoolTest::ne, zeroX); { 4003 4004 // No barrier if we are storing a NULL 4005 __ if_then(val, BoolTest::ne, null(), unlikely); { 4006 4007 // Ok must mark the card if not already dirty 4008 4009 // load the original value of the card 4010 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4011 4012 __ if_then(card_val, BoolTest::ne, young_card); { 4013 sync_kit(ideal); 4014 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4015 insert_mem_bar(Op_MemBarVolatile, oop_store); 4016 __ sync_kit(this); 4017 4018 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4019 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4020 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4021 } __ end_if(); 4022 } __ end_if(); 4023 } __ end_if(); 4024 } __ end_if(); 4025 } else { 4026 // Object.clone() instrinsic uses this path. 4027 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4028 } 4029 4030 // Final sync IdealKit and GraphKit. 4031 final_sync(ideal); 4032 } 4033 #undef __ 4034 4035 4036 4037 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4038 if (java_lang_String::has_offset_field()) { 4039 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4040 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4041 false, NULL, 0); 4042 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4043 int offset_field_idx = C->get_alias_index(offset_field_type); 4044 return make_load(ctrl, 4045 basic_plus_adr(str, str, offset_offset), 4046 TypeInt::INT, T_INT, offset_field_idx); 4047 } else { 4048 return intcon(0); 4049 } 4050 } 4051 4052 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4053 if (java_lang_String::has_count_field()) { 4054 int count_offset = java_lang_String::count_offset_in_bytes(); 4055 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4056 false, NULL, 0); 4057 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4058 int count_field_idx = C->get_alias_index(count_field_type); 4059 return make_load(ctrl, 4060 basic_plus_adr(str, str, count_offset), 4061 TypeInt::INT, T_INT, count_field_idx); 4062 } else { 4063 return load_array_length(load_String_value(ctrl, str)); 4064 } 4065 } 4066 4067 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4068 int value_offset = java_lang_String::value_offset_in_bytes(); 4069 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4070 false, NULL, 0); 4071 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4072 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4073 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4074 ciTypeArrayKlass::make(T_CHAR), true, 0); 4075 int value_field_idx = C->get_alias_index(value_field_type); 4076 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4077 value_type, T_OBJECT, value_field_idx); 4078 // String.value field is known to be @Stable. 4079 if (UseImplicitStableValues) { 4080 load = cast_array_to_stable(load, value_type); 4081 } 4082 return load; 4083 } 4084 4085 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4086 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4087 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4088 false, NULL, 0); 4089 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4090 int offset_field_idx = C->get_alias_index(offset_field_type); 4091 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4092 value, T_INT, offset_field_idx); 4093 } 4094 4095 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4096 int value_offset = java_lang_String::value_offset_in_bytes(); 4097 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4098 false, NULL, 0); 4099 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4100 4101 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4102 value, TypeAryPtr::CHARS, T_OBJECT); 4103 } 4104 4105 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4106 int count_offset = java_lang_String::count_offset_in_bytes(); 4107 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4108 false, NULL, 0); 4109 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4110 int count_field_idx = C->get_alias_index(count_field_type); 4111 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4112 value, T_INT, count_field_idx); 4113 } 4114 4115 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4116 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4117 // assumption of CCP analysis. 4118 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); 4119 }