1 /* 2 * Copyright (c) 2001, 2013, 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/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/locknode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/optonode.hpp" 40 #include "opto/parse.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/runtime.hpp" 43 #include "opto/threadnode.hpp" 44 #include "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 47 //----------------------------GraphKit----------------------------------------- 48 // Main utility constructor. 49 GraphKit::GraphKit(JVMState* jvms) 50 : Phase(Phase::Parser), 51 _env(C->env()), 52 _gvn(*C->initial_gvn()) 53 { 54 _exceptions = jvms->map()->next_exception(); 55 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 56 set_jvms(jvms); 57 } 58 59 // Private constructor for parser. 60 GraphKit::GraphKit() 61 : Phase(Phase::Parser), 62 _env(C->env()), 63 _gvn(*C->initial_gvn()) 64 { 65 _exceptions = NULL; 66 set_map(NULL); 67 debug_only(_sp = -99); 68 debug_only(set_bci(-99)); 69 } 70 71 72 73 //---------------------------clean_stack--------------------------------------- 74 // Clear away rubbish from the stack area of the JVM state. 75 // This destroys any arguments that may be waiting on the stack. 76 void GraphKit::clean_stack(int from_sp) { 77 SafePointNode* map = this->map(); 78 JVMState* jvms = this->jvms(); 79 int stk_size = jvms->stk_size(); 80 int stkoff = jvms->stkoff(); 81 Node* top = this->top(); 82 for (int i = from_sp; i < stk_size; i++) { 83 if (map->in(stkoff + i) != top) { 84 map->set_req(stkoff + i, top); 85 } 86 } 87 } 88 89 90 //--------------------------------sync_jvms----------------------------------- 91 // Make sure our current jvms agrees with our parse state. 92 JVMState* GraphKit::sync_jvms() const { 93 JVMState* jvms = this->jvms(); 94 jvms->set_bci(bci()); // Record the new bci in the JVMState 95 jvms->set_sp(sp()); // Record the new sp in the JVMState 96 assert(jvms_in_sync(), "jvms is now in sync"); 97 return jvms; 98 } 99 100 //--------------------------------sync_jvms_for_reexecute--------------------- 101 // Make sure our current jvms agrees with our parse state. This version 102 // uses the reexecute_sp for reexecuting bytecodes. 103 JVMState* GraphKit::sync_jvms_for_reexecute() { 104 JVMState* jvms = this->jvms(); 105 jvms->set_bci(bci()); // Record the new bci in the JVMState 106 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 107 return jvms; 108 } 109 110 #ifdef ASSERT 111 bool GraphKit::jvms_in_sync() const { 112 Parse* parse = is_Parse(); 113 if (parse == NULL) { 114 if (bci() != jvms()->bci()) return false; 115 if (sp() != (int)jvms()->sp()) return false; 116 return true; 117 } 118 if (jvms()->method() != parse->method()) return false; 119 if (jvms()->bci() != parse->bci()) return false; 120 int jvms_sp = jvms()->sp(); 121 if (jvms_sp != parse->sp()) return false; 122 int jvms_depth = jvms()->depth(); 123 if (jvms_depth != parse->depth()) return false; 124 return true; 125 } 126 127 // Local helper checks for special internal merge points 128 // used to accumulate and merge exception states. 129 // They are marked by the region's in(0) edge being the map itself. 130 // Such merge points must never "escape" into the parser at large, 131 // until they have been handed to gvn.transform. 132 static bool is_hidden_merge(Node* reg) { 133 if (reg == NULL) return false; 134 if (reg->is_Phi()) { 135 reg = reg->in(0); 136 if (reg == NULL) return false; 137 } 138 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 139 } 140 141 void GraphKit::verify_map() const { 142 if (map() == NULL) return; // null map is OK 143 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 144 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 145 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 146 } 147 148 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 149 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 150 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 151 } 152 #endif 153 154 //---------------------------stop_and_kill_map--------------------------------- 155 // Set _map to NULL, signalling a stop to further bytecode execution. 156 // First smash the current map's control to a constant, to mark it dead. 157 void GraphKit::stop_and_kill_map() { 158 SafePointNode* dead_map = stop(); 159 if (dead_map != NULL) { 160 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 161 assert(dead_map->is_killed(), "must be so marked"); 162 } 163 } 164 165 166 //--------------------------------stopped-------------------------------------- 167 // Tell if _map is NULL, or control is top. 168 bool GraphKit::stopped() { 169 if (map() == NULL) return true; 170 else if (control() == top()) return true; 171 else return false; 172 } 173 174 175 //-----------------------------has_ex_handler---------------------------------- 176 // Tell if this method or any caller method has exception handlers. 177 bool GraphKit::has_ex_handler() { 178 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 179 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 180 return true; 181 } 182 } 183 return false; 184 } 185 186 //------------------------------save_ex_oop------------------------------------ 187 // Save an exception without blowing stack contents or other JVM state. 188 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 189 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 190 ex_map->add_req(ex_oop); 191 debug_only(verify_exception_state(ex_map)); 192 } 193 194 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 195 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 196 Node* ex_oop = ex_map->in(ex_map->req()-1); 197 if (clear_it) ex_map->del_req(ex_map->req()-1); 198 return ex_oop; 199 } 200 201 //-----------------------------saved_ex_oop------------------------------------ 202 // Recover a saved exception from its map. 203 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 204 return common_saved_ex_oop(ex_map, false); 205 } 206 207 //--------------------------clear_saved_ex_oop--------------------------------- 208 // Erase a previously saved exception from its map. 209 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, true); 211 } 212 213 #ifdef ASSERT 214 //---------------------------has_saved_ex_oop---------------------------------- 215 // Erase a previously saved exception from its map. 216 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 217 return ex_map->req() == ex_map->jvms()->endoff()+1; 218 } 219 #endif 220 221 //-------------------------make_exception_state-------------------------------- 222 // Turn the current JVM state into an exception state, appending the ex_oop. 223 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 224 sync_jvms(); 225 SafePointNode* ex_map = stop(); // do not manipulate this map any more 226 set_saved_ex_oop(ex_map, ex_oop); 227 return ex_map; 228 } 229 230 231 //--------------------------add_exception_state-------------------------------- 232 // Add an exception to my list of exceptions. 233 void GraphKit::add_exception_state(SafePointNode* ex_map) { 234 if (ex_map == NULL || ex_map->control() == top()) { 235 return; 236 } 237 #ifdef ASSERT 238 verify_exception_state(ex_map); 239 if (has_exceptions()) { 240 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 241 } 242 #endif 243 244 // If there is already an exception of exactly this type, merge with it. 245 // In particular, null-checks and other low-level exceptions common up here. 246 Node* ex_oop = saved_ex_oop(ex_map); 247 const Type* ex_type = _gvn.type(ex_oop); 248 if (ex_oop == top()) { 249 // No action needed. 250 return; 251 } 252 assert(ex_type->isa_instptr(), "exception must be an instance"); 253 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 254 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 255 // We check sp also because call bytecodes can generate exceptions 256 // both before and after arguments are popped! 257 if (ex_type2 == ex_type 258 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 259 combine_exception_states(ex_map, e2); 260 return; 261 } 262 } 263 264 // No pre-existing exception of the same type. Chain it on the list. 265 push_exception_state(ex_map); 266 } 267 268 //-----------------------add_exception_states_from----------------------------- 269 void GraphKit::add_exception_states_from(JVMState* jvms) { 270 SafePointNode* ex_map = jvms->map()->next_exception(); 271 if (ex_map != NULL) { 272 jvms->map()->set_next_exception(NULL); 273 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 274 next_map = ex_map->next_exception(); 275 ex_map->set_next_exception(NULL); 276 add_exception_state(ex_map); 277 } 278 } 279 } 280 281 //-----------------------transfer_exceptions_into_jvms------------------------- 282 JVMState* GraphKit::transfer_exceptions_into_jvms() { 283 if (map() == NULL) { 284 // We need a JVMS to carry the exceptions, but the map has gone away. 285 // Create a scratch JVMS, cloned from any of the exception states... 286 if (has_exceptions()) { 287 _map = _exceptions; 288 _map = clone_map(); 289 _map->set_next_exception(NULL); 290 clear_saved_ex_oop(_map); 291 debug_only(verify_map()); 292 } else { 293 // ...or created from scratch 294 JVMState* jvms = new (C) JVMState(_method, NULL); 295 jvms->set_bci(_bci); 296 jvms->set_sp(_sp); 297 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); 298 set_jvms(jvms); 299 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 300 set_all_memory(top()); 301 while (map()->req() < jvms->endoff()) map()->add_req(top()); 302 } 303 // (This is a kludge, in case you didn't notice.) 304 set_control(top()); 305 } 306 JVMState* jvms = sync_jvms(); 307 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 308 jvms->map()->set_next_exception(_exceptions); 309 _exceptions = NULL; // done with this set of exceptions 310 return jvms; 311 } 312 313 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 314 assert(is_hidden_merge(dstphi), "must be a special merge node"); 315 assert(is_hidden_merge(srcphi), "must be a special merge node"); 316 uint limit = srcphi->req(); 317 for (uint i = PhiNode::Input; i < limit; i++) { 318 dstphi->add_req(srcphi->in(i)); 319 } 320 } 321 static inline void add_one_req(Node* dstphi, Node* src) { 322 assert(is_hidden_merge(dstphi), "must be a special merge node"); 323 assert(!is_hidden_merge(src), "must not be a special merge node"); 324 dstphi->add_req(src); 325 } 326 327 //-----------------------combine_exception_states------------------------------ 328 // This helper function combines exception states by building phis on a 329 // specially marked state-merging region. These regions and phis are 330 // untransformed, and can build up gradually. The region is marked by 331 // having a control input of its exception map, rather than NULL. Such 332 // regions do not appear except in this function, and in use_exception_state. 333 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 334 if (failing()) return; // dying anyway... 335 JVMState* ex_jvms = ex_map->_jvms; 336 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 337 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 338 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 339 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 340 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 341 assert(ex_map->req() == phi_map->req(), "matching maps"); 342 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 343 Node* hidden_merge_mark = root(); 344 Node* region = phi_map->control(); 345 MergeMemNode* phi_mem = phi_map->merged_memory(); 346 MergeMemNode* ex_mem = ex_map->merged_memory(); 347 if (region->in(0) != hidden_merge_mark) { 348 // The control input is not (yet) a specially-marked region in phi_map. 349 // Make it so, and build some phis. 350 region = new (C) RegionNode(2); 351 _gvn.set_type(region, Type::CONTROL); 352 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 353 region->init_req(1, phi_map->control()); 354 phi_map->set_control(region); 355 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 356 record_for_igvn(io_phi); 357 _gvn.set_type(io_phi, Type::ABIO); 358 phi_map->set_i_o(io_phi); 359 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 360 Node* m = mms.memory(); 361 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 362 record_for_igvn(m_phi); 363 _gvn.set_type(m_phi, Type::MEMORY); 364 mms.set_memory(m_phi); 365 } 366 } 367 368 // Either or both of phi_map and ex_map might already be converted into phis. 369 Node* ex_control = ex_map->control(); 370 // if there is special marking on ex_map also, we add multiple edges from src 371 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 372 // how wide was the destination phi_map, originally? 373 uint orig_width = region->req(); 374 375 if (add_multiple) { 376 add_n_reqs(region, ex_control); 377 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 378 } else { 379 // ex_map has no merges, so we just add single edges everywhere 380 add_one_req(region, ex_control); 381 add_one_req(phi_map->i_o(), ex_map->i_o()); 382 } 383 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 384 if (mms.is_empty()) { 385 // get a copy of the base memory, and patch some inputs into it 386 const TypePtr* adr_type = mms.adr_type(C); 387 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 388 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 389 mms.set_memory(phi); 390 // Prepare to append interesting stuff onto the newly sliced phi: 391 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 392 } 393 // Append stuff from ex_map: 394 if (add_multiple) { 395 add_n_reqs(mms.memory(), mms.memory2()); 396 } else { 397 add_one_req(mms.memory(), mms.memory2()); 398 } 399 } 400 uint limit = ex_map->req(); 401 for (uint i = TypeFunc::Parms; i < limit; i++) { 402 // Skip everything in the JVMS after tos. (The ex_oop follows.) 403 if (i == tos) i = ex_jvms->monoff(); 404 Node* src = ex_map->in(i); 405 Node* dst = phi_map->in(i); 406 if (src != dst) { 407 PhiNode* phi; 408 if (dst->in(0) != region) { 409 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 410 record_for_igvn(phi); 411 _gvn.set_type(phi, phi->type()); 412 phi_map->set_req(i, dst); 413 // Prepare to append interesting stuff onto the new phi: 414 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 415 } else { 416 assert(dst->is_Phi(), "nobody else uses a hidden region"); 417 phi = dst->as_Phi(); 418 } 419 if (add_multiple && src->in(0) == ex_control) { 420 // Both are phis. 421 add_n_reqs(dst, src); 422 } else { 423 while (dst->req() < region->req()) add_one_req(dst, src); 424 } 425 const Type* srctype = _gvn.type(src); 426 if (phi->type() != srctype) { 427 const Type* dsttype = phi->type()->meet_speculative(srctype); 428 if (phi->type() != dsttype) { 429 phi->set_type(dsttype); 430 _gvn.set_type(phi, dsttype); 431 } 432 } 433 } 434 } 435 } 436 437 //--------------------------use_exception_state-------------------------------- 438 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 439 if (failing()) { stop(); return top(); } 440 Node* region = phi_map->control(); 441 Node* hidden_merge_mark = root(); 442 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 443 Node* ex_oop = clear_saved_ex_oop(phi_map); 444 if (region->in(0) == hidden_merge_mark) { 445 // Special marking for internal ex-states. Process the phis now. 446 region->set_req(0, region); // now it's an ordinary region 447 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 448 // Note: Setting the jvms also sets the bci and sp. 449 set_control(_gvn.transform(region)); 450 uint tos = jvms()->stkoff() + sp(); 451 for (uint i = 1; i < tos; i++) { 452 Node* x = phi_map->in(i); 453 if (x->in(0) == region) { 454 assert(x->is_Phi(), "expected a special phi"); 455 phi_map->set_req(i, _gvn.transform(x)); 456 } 457 } 458 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 459 Node* x = mms.memory(); 460 if (x->in(0) == region) { 461 assert(x->is_Phi(), "nobody else uses a hidden region"); 462 mms.set_memory(_gvn.transform(x)); 463 } 464 } 465 if (ex_oop->in(0) == region) { 466 assert(ex_oop->is_Phi(), "expected a special phi"); 467 ex_oop = _gvn.transform(ex_oop); 468 } 469 } else { 470 set_jvms(phi_map->jvms()); 471 } 472 473 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 474 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 475 return ex_oop; 476 } 477 478 //---------------------------------java_bc------------------------------------- 479 Bytecodes::Code GraphKit::java_bc() const { 480 ciMethod* method = this->method(); 481 int bci = this->bci(); 482 if (method != NULL && bci != InvocationEntryBci) 483 return method->java_code_at_bci(bci); 484 else 485 return Bytecodes::_illegal; 486 } 487 488 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 489 bool must_throw) { 490 // if the exception capability is set, then we will generate code 491 // to check the JavaThread.should_post_on_exceptions flag to see 492 // if we actually need to report exception events (for this 493 // thread). If we don't need to report exception events, we will 494 // take the normal fast path provided by add_exception_events. If 495 // exception event reporting is enabled for this thread, we will 496 // take the uncommon_trap in the BuildCutout below. 497 498 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 499 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 500 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 501 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 502 503 // Test the should_post_on_exceptions_flag vs. 0 504 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 505 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 506 507 // Branch to slow_path if should_post_on_exceptions_flag was true 508 { BuildCutout unless(this, tst, PROB_MAX); 509 // Do not try anything fancy if we're notifying the VM on every throw. 510 // Cf. case Bytecodes::_athrow in parse2.cpp. 511 uncommon_trap(reason, Deoptimization::Action_none, 512 (ciKlass*)NULL, (char*)NULL, must_throw); 513 } 514 515 } 516 517 //------------------------------builtin_throw---------------------------------- 518 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 519 bool must_throw = true; 520 521 if (env()->jvmti_can_post_on_exceptions()) { 522 // check if we must post exception events, take uncommon trap if so 523 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 524 // here if should_post_on_exceptions is false 525 // continue on with the normal codegen 526 } 527 528 // If this particular condition has not yet happened at this 529 // bytecode, then use the uncommon trap mechanism, and allow for 530 // a future recompilation if several traps occur here. 531 // If the throw is hot, try to use a more complicated inline mechanism 532 // which keeps execution inside the compiled code. 533 bool treat_throw_as_hot = false; 534 ciMethodData* md = method()->method_data(); 535 536 if (ProfileTraps) { 537 if (too_many_traps(reason)) { 538 treat_throw_as_hot = true; 539 } 540 // (If there is no MDO at all, assume it is early in 541 // execution, and that any deopts are part of the 542 // startup transient, and don't need to be remembered.) 543 544 // Also, if there is a local exception handler, treat all throws 545 // as hot if there has been at least one in this method. 546 if (C->trap_count(reason) != 0 547 && method()->method_data()->trap_count(reason) != 0 548 && has_ex_handler()) { 549 treat_throw_as_hot = true; 550 } 551 } 552 553 // If this throw happens frequently, an uncommon trap might cause 554 // a performance pothole. If there is a local exception handler, 555 // and if this particular bytecode appears to be deoptimizing often, 556 // let us handle the throw inline, with a preconstructed instance. 557 // Note: If the deopt count has blown up, the uncommon trap 558 // runtime is going to flush this nmethod, not matter what. 559 if (treat_throw_as_hot 560 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 561 // If the throw is local, we use a pre-existing instance and 562 // punt on the backtrace. This would lead to a missing backtrace 563 // (a repeat of 4292742) if the backtrace object is ever asked 564 // for its backtrace. 565 // Fixing this remaining case of 4292742 requires some flavor of 566 // escape analysis. Leave that for the future. 567 ciInstance* ex_obj = NULL; 568 switch (reason) { 569 case Deoptimization::Reason_null_check: 570 ex_obj = env()->NullPointerException_instance(); 571 break; 572 case Deoptimization::Reason_div0_check: 573 ex_obj = env()->ArithmeticException_instance(); 574 break; 575 case Deoptimization::Reason_range_check: 576 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 577 break; 578 case Deoptimization::Reason_class_check: 579 if (java_bc() == Bytecodes::_aastore) { 580 ex_obj = env()->ArrayStoreException_instance(); 581 } else { 582 ex_obj = env()->ClassCastException_instance(); 583 } 584 break; 585 } 586 if (failing()) { stop(); return; } // exception allocation might fail 587 if (ex_obj != NULL) { 588 // Cheat with a preallocated exception object. 589 if (C->log() != NULL) 590 C->log()->elem("hot_throw preallocated='1' reason='%s'", 591 Deoptimization::trap_reason_name(reason)); 592 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 593 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 594 595 // Clear the detail message of the preallocated exception object. 596 // Weblogic sometimes mutates the detail message of exceptions 597 // using reflection. 598 int offset = java_lang_Throwable::get_detailMessage_offset(); 599 const TypePtr* adr_typ = ex_con->add_offset(offset); 600 601 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 602 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 603 // Conservatively release stores of object references. 604 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 605 606 add_exception_state(make_exception_state(ex_node)); 607 return; 608 } 609 } 610 611 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 612 // It won't be much cheaper than bailing to the interp., since we'll 613 // have to pass up all the debug-info, and the runtime will have to 614 // create the stack trace. 615 616 // Usual case: Bail to interpreter. 617 // Reserve the right to recompile if we haven't seen anything yet. 618 619 assert(!Deoptimization::reason_is_speculate(reason), "unsupported"); 620 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 621 if (treat_throw_as_hot 622 && (method()->method_data()->trap_recompiled_at(bci(), NULL) 623 || C->too_many_traps(reason))) { 624 // We cannot afford to take more traps here. Suffer in the interpreter. 625 if (C->log() != NULL) 626 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 627 Deoptimization::trap_reason_name(reason), 628 C->trap_count(reason)); 629 action = Deoptimization::Action_none; 630 } 631 632 // "must_throw" prunes the JVM state to include only the stack, if there 633 // are no local exception handlers. This should cut down on register 634 // allocation time and code size, by drastically reducing the number 635 // of in-edges on the call to the uncommon trap. 636 637 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 638 } 639 640 641 //----------------------------PreserveJVMState--------------------------------- 642 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 643 debug_only(kit->verify_map()); 644 _kit = kit; 645 _map = kit->map(); // preserve the map 646 _sp = kit->sp(); 647 kit->set_map(clone_map ? kit->clone_map() : NULL); 648 Compile::current()->inc_preserve_jvm_state(); 649 #ifdef ASSERT 650 _bci = kit->bci(); 651 Parse* parser = kit->is_Parse(); 652 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 653 _block = block; 654 #endif 655 } 656 PreserveJVMState::~PreserveJVMState() { 657 GraphKit* kit = _kit; 658 #ifdef ASSERT 659 assert(kit->bci() == _bci, "bci must not shift"); 660 Parse* parser = kit->is_Parse(); 661 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 662 assert(block == _block, "block must not shift"); 663 #endif 664 kit->set_map(_map); 665 kit->set_sp(_sp); 666 Compile::current()->dec_preserve_jvm_state(); 667 } 668 669 670 //-----------------------------BuildCutout------------------------------------- 671 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 672 : PreserveJVMState(kit) 673 { 674 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 675 SafePointNode* outer_map = _map; // preserved map is caller's 676 SafePointNode* inner_map = kit->map(); 677 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 678 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 679 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 680 } 681 BuildCutout::~BuildCutout() { 682 GraphKit* kit = _kit; 683 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 684 } 685 686 //---------------------------PreserveReexecuteState---------------------------- 687 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 688 assert(!kit->stopped(), "must call stopped() before"); 689 _kit = kit; 690 _sp = kit->sp(); 691 _reexecute = kit->jvms()->_reexecute; 692 } 693 PreserveReexecuteState::~PreserveReexecuteState() { 694 if (_kit->stopped()) return; 695 _kit->jvms()->_reexecute = _reexecute; 696 _kit->set_sp(_sp); 697 } 698 699 //------------------------------clone_map-------------------------------------- 700 // Implementation of PreserveJVMState 701 // 702 // Only clone_map(...) here. If this function is only used in the 703 // PreserveJVMState class we may want to get rid of this extra 704 // function eventually and do it all there. 705 706 SafePointNode* GraphKit::clone_map() { 707 if (map() == NULL) return NULL; 708 709 // Clone the memory edge first 710 Node* mem = MergeMemNode::make(C, map()->memory()); 711 gvn().set_type_bottom(mem); 712 713 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 714 JVMState* jvms = this->jvms(); 715 JVMState* clonejvms = jvms->clone_shallow(C); 716 clonemap->set_memory(mem); 717 clonemap->set_jvms(clonejvms); 718 clonejvms->set_map(clonemap); 719 record_for_igvn(clonemap); 720 gvn().set_type_bottom(clonemap); 721 return clonemap; 722 } 723 724 725 //-----------------------------set_map_clone----------------------------------- 726 void GraphKit::set_map_clone(SafePointNode* m) { 727 _map = m; 728 _map = clone_map(); 729 _map->set_next_exception(NULL); 730 debug_only(verify_map()); 731 } 732 733 734 //----------------------------kill_dead_locals--------------------------------- 735 // Detect any locals which are known to be dead, and force them to top. 736 void GraphKit::kill_dead_locals() { 737 // Consult the liveness information for the locals. If any 738 // of them are unused, then they can be replaced by top(). This 739 // should help register allocation time and cut down on the size 740 // of the deoptimization information. 741 742 // This call is made from many of the bytecode handling 743 // subroutines called from the Big Switch in do_one_bytecode. 744 // Every bytecode which might include a slow path is responsible 745 // for killing its dead locals. The more consistent we 746 // are about killing deads, the fewer useless phis will be 747 // constructed for them at various merge points. 748 749 // bci can be -1 (InvocationEntryBci). We return the entry 750 // liveness for the method. 751 752 if (method() == NULL || method()->code_size() == 0) { 753 // We are building a graph for a call to a native method. 754 // All locals are live. 755 return; 756 } 757 758 ResourceMark rm; 759 760 // Consult the liveness information for the locals. If any 761 // of them are unused, then they can be replaced by top(). This 762 // should help register allocation time and cut down on the size 763 // of the deoptimization information. 764 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 765 766 int len = (int)live_locals.size(); 767 assert(len <= jvms()->loc_size(), "too many live locals"); 768 for (int local = 0; local < len; local++) { 769 if (!live_locals.at(local)) { 770 set_local(local, top()); 771 } 772 } 773 } 774 775 #ifdef ASSERT 776 //-------------------------dead_locals_are_killed------------------------------ 777 // Return true if all dead locals are set to top in the map. 778 // Used to assert "clean" debug info at various points. 779 bool GraphKit::dead_locals_are_killed() { 780 if (method() == NULL || method()->code_size() == 0) { 781 // No locals need to be dead, so all is as it should be. 782 return true; 783 } 784 785 // Make sure somebody called kill_dead_locals upstream. 786 ResourceMark rm; 787 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 788 if (jvms->loc_size() == 0) continue; // no locals to consult 789 SafePointNode* map = jvms->map(); 790 ciMethod* method = jvms->method(); 791 int bci = jvms->bci(); 792 if (jvms == this->jvms()) { 793 bci = this->bci(); // it might not yet be synched 794 } 795 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 796 int len = (int)live_locals.size(); 797 if (!live_locals.is_valid() || len == 0) 798 // This method is trivial, or is poisoned by a breakpoint. 799 return true; 800 assert(len == jvms->loc_size(), "live map consistent with locals map"); 801 for (int local = 0; local < len; local++) { 802 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 803 if (PrintMiscellaneous && (Verbose || WizardMode)) { 804 tty->print_cr("Zombie local %d: ", local); 805 jvms->dump(); 806 } 807 return false; 808 } 809 } 810 } 811 return true; 812 } 813 814 #endif //ASSERT 815 816 // Helper function for enforcing certain bytecodes to reexecute if 817 // deoptimization happens 818 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 819 ciMethod* cur_method = jvms->method(); 820 int cur_bci = jvms->bci(); 821 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 822 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 823 return Interpreter::bytecode_should_reexecute(code) || 824 is_anewarray && code == Bytecodes::_multianewarray; 825 // Reexecute _multianewarray bytecode which was replaced with 826 // sequence of [a]newarray. See Parse::do_multianewarray(). 827 // 828 // Note: interpreter should not have it set since this optimization 829 // is limited by dimensions and guarded by flag so in some cases 830 // multianewarray() runtime calls will be generated and 831 // the bytecode should not be reexecutes (stack will not be reset). 832 } else 833 return false; 834 } 835 836 // Helper function for adding JVMState and debug information to node 837 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 838 // Add the safepoint edges to the call (or other safepoint). 839 840 // Make sure dead locals are set to top. This 841 // should help register allocation time and cut down on the size 842 // of the deoptimization information. 843 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 844 845 // Walk the inline list to fill in the correct set of JVMState's 846 // Also fill in the associated edges for each JVMState. 847 848 // If the bytecode needs to be reexecuted we need to put 849 // the arguments back on the stack. 850 const bool should_reexecute = jvms()->should_reexecute(); 851 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 852 853 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 854 // undefined if the bci is different. This is normal for Parse but it 855 // should not happen for LibraryCallKit because only one bci is processed. 856 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 857 "in LibraryCallKit the reexecute bit should not change"); 858 859 // If we are guaranteed to throw, we can prune everything but the 860 // input to the current bytecode. 861 bool can_prune_locals = false; 862 uint stack_slots_not_pruned = 0; 863 int inputs = 0, depth = 0; 864 if (must_throw) { 865 assert(method() == youngest_jvms->method(), "sanity"); 866 if (compute_stack_effects(inputs, depth)) { 867 can_prune_locals = true; 868 stack_slots_not_pruned = inputs; 869 } 870 } 871 872 if (env()->should_retain_local_variables()) { 873 // At any safepoint, this method can get breakpointed, which would 874 // then require an immediate deoptimization. 875 can_prune_locals = false; // do not prune locals 876 stack_slots_not_pruned = 0; 877 } 878 879 // do not scribble on the input jvms 880 JVMState* out_jvms = youngest_jvms->clone_deep(C); 881 call->set_jvms(out_jvms); // Start jvms list for call node 882 883 // For a known set of bytecodes, the interpreter should reexecute them if 884 // deoptimization happens. We set the reexecute state for them here 885 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 886 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 887 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 888 } 889 890 // Presize the call: 891 DEBUG_ONLY(uint non_debug_edges = call->req()); 892 call->add_req_batch(top(), youngest_jvms->debug_depth()); 893 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 894 895 // Set up edges so that the call looks like this: 896 // Call [state:] ctl io mem fptr retadr 897 // [parms:] parm0 ... parmN 898 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 899 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 900 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 901 // Note that caller debug info precedes callee debug info. 902 903 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 904 uint debug_ptr = call->req(); 905 906 // Loop over the map input edges associated with jvms, add them 907 // to the call node, & reset all offsets to match call node array. 908 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 909 uint debug_end = debug_ptr; 910 uint debug_start = debug_ptr - in_jvms->debug_size(); 911 debug_ptr = debug_start; // back up the ptr 912 913 uint p = debug_start; // walks forward in [debug_start, debug_end) 914 uint j, k, l; 915 SafePointNode* in_map = in_jvms->map(); 916 out_jvms->set_map(call); 917 918 if (can_prune_locals) { 919 assert(in_jvms->method() == out_jvms->method(), "sanity"); 920 // If the current throw can reach an exception handler in this JVMS, 921 // then we must keep everything live that can reach that handler. 922 // As a quick and dirty approximation, we look for any handlers at all. 923 if (in_jvms->method()->has_exception_handlers()) { 924 can_prune_locals = false; 925 } 926 } 927 928 // Add the Locals 929 k = in_jvms->locoff(); 930 l = in_jvms->loc_size(); 931 out_jvms->set_locoff(p); 932 if (!can_prune_locals) { 933 for (j = 0; j < l; j++) 934 call->set_req(p++, in_map->in(k+j)); 935 } else { 936 p += l; // already set to top above by add_req_batch 937 } 938 939 // Add the Expression Stack 940 k = in_jvms->stkoff(); 941 l = in_jvms->sp(); 942 out_jvms->set_stkoff(p); 943 if (!can_prune_locals) { 944 for (j = 0; j < l; j++) 945 call->set_req(p++, in_map->in(k+j)); 946 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 947 // Divide stack into {S0,...,S1}, where S0 is set to top. 948 uint s1 = stack_slots_not_pruned; 949 stack_slots_not_pruned = 0; // for next iteration 950 if (s1 > l) s1 = l; 951 uint s0 = l - s1; 952 p += s0; // skip the tops preinstalled by add_req_batch 953 for (j = s0; j < l; j++) 954 call->set_req(p++, in_map->in(k+j)); 955 } else { 956 p += l; // already set to top above by add_req_batch 957 } 958 959 // Add the Monitors 960 k = in_jvms->monoff(); 961 l = in_jvms->mon_size(); 962 out_jvms->set_monoff(p); 963 for (j = 0; j < l; j++) 964 call->set_req(p++, in_map->in(k+j)); 965 966 // Copy any scalar object fields. 967 k = in_jvms->scloff(); 968 l = in_jvms->scl_size(); 969 out_jvms->set_scloff(p); 970 for (j = 0; j < l; j++) 971 call->set_req(p++, in_map->in(k+j)); 972 973 // Finish the new jvms. 974 out_jvms->set_endoff(p); 975 976 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 977 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 978 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 979 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 980 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 981 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 982 983 // Update the two tail pointers in parallel. 984 out_jvms = out_jvms->caller(); 985 in_jvms = in_jvms->caller(); 986 } 987 988 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 989 990 // Test the correctness of JVMState::debug_xxx accessors: 991 assert(call->jvms()->debug_start() == non_debug_edges, ""); 992 assert(call->jvms()->debug_end() == call->req(), ""); 993 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 994 } 995 996 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 997 Bytecodes::Code code = java_bc(); 998 if (code == Bytecodes::_wide) { 999 code = method()->java_code_at_bci(bci() + 1); 1000 } 1001 1002 BasicType rtype = T_ILLEGAL; 1003 int rsize = 0; 1004 1005 if (code != Bytecodes::_illegal) { 1006 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1007 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1008 if (rtype < T_CONFLICT) 1009 rsize = type2size[rtype]; 1010 } 1011 1012 switch (code) { 1013 case Bytecodes::_illegal: 1014 return false; 1015 1016 case Bytecodes::_ldc: 1017 case Bytecodes::_ldc_w: 1018 case Bytecodes::_ldc2_w: 1019 inputs = 0; 1020 break; 1021 1022 case Bytecodes::_dup: inputs = 1; break; 1023 case Bytecodes::_dup_x1: inputs = 2; break; 1024 case Bytecodes::_dup_x2: inputs = 3; break; 1025 case Bytecodes::_dup2: inputs = 2; break; 1026 case Bytecodes::_dup2_x1: inputs = 3; break; 1027 case Bytecodes::_dup2_x2: inputs = 4; break; 1028 case Bytecodes::_swap: inputs = 2; break; 1029 case Bytecodes::_arraylength: inputs = 1; break; 1030 1031 case Bytecodes::_getstatic: 1032 case Bytecodes::_putstatic: 1033 case Bytecodes::_getfield: 1034 case Bytecodes::_putfield: 1035 { 1036 bool ignored_will_link; 1037 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1038 int size = field->type()->size(); 1039 bool is_get = (depth >= 0), is_static = (depth & 1); 1040 inputs = (is_static ? 0 : 1); 1041 if (is_get) { 1042 depth = size - inputs; 1043 } else { 1044 inputs += size; // putxxx pops the value from the stack 1045 depth = - inputs; 1046 } 1047 } 1048 break; 1049 1050 case Bytecodes::_invokevirtual: 1051 case Bytecodes::_invokespecial: 1052 case Bytecodes::_invokestatic: 1053 case Bytecodes::_invokedynamic: 1054 case Bytecodes::_invokeinterface: 1055 { 1056 bool ignored_will_link; 1057 ciSignature* declared_signature = NULL; 1058 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1059 assert(declared_signature != NULL, "cannot be null"); 1060 inputs = declared_signature->arg_size_for_bc(code); 1061 int size = declared_signature->return_type()->size(); 1062 depth = size - inputs; 1063 } 1064 break; 1065 1066 case Bytecodes::_multianewarray: 1067 { 1068 ciBytecodeStream iter(method()); 1069 iter.reset_to_bci(bci()); 1070 iter.next(); 1071 inputs = iter.get_dimensions(); 1072 assert(rsize == 1, ""); 1073 depth = rsize - inputs; 1074 } 1075 break; 1076 1077 case Bytecodes::_ireturn: 1078 case Bytecodes::_lreturn: 1079 case Bytecodes::_freturn: 1080 case Bytecodes::_dreturn: 1081 case Bytecodes::_areturn: 1082 assert(rsize = -depth, ""); 1083 inputs = rsize; 1084 break; 1085 1086 case Bytecodes::_jsr: 1087 case Bytecodes::_jsr_w: 1088 inputs = 0; 1089 depth = 1; // S.B. depth=1, not zero 1090 break; 1091 1092 default: 1093 // bytecode produces a typed result 1094 inputs = rsize - depth; 1095 assert(inputs >= 0, ""); 1096 break; 1097 } 1098 1099 #ifdef ASSERT 1100 // spot check 1101 int outputs = depth + inputs; 1102 assert(outputs >= 0, "sanity"); 1103 switch (code) { 1104 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1105 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1106 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1107 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1108 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1109 } 1110 #endif //ASSERT 1111 1112 return true; 1113 } 1114 1115 1116 1117 //------------------------------basic_plus_adr--------------------------------- 1118 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1119 // short-circuit a common case 1120 if (offset == intcon(0)) return ptr; 1121 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1122 } 1123 1124 Node* GraphKit::ConvI2L(Node* offset) { 1125 // short-circuit a common case 1126 jint offset_con = find_int_con(offset, Type::OffsetBot); 1127 if (offset_con != Type::OffsetBot) { 1128 return longcon((jlong) offset_con); 1129 } 1130 return _gvn.transform( new (C) ConvI2LNode(offset)); 1131 } 1132 Node* GraphKit::ConvL2I(Node* offset) { 1133 // short-circuit a common case 1134 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1135 if (offset_con != (jlong)Type::OffsetBot) { 1136 return intcon((int) offset_con); 1137 } 1138 return _gvn.transform( new (C) ConvL2INode(offset)); 1139 } 1140 1141 //-------------------------load_object_klass----------------------------------- 1142 Node* GraphKit::load_object_klass(Node* obj) { 1143 // Special-case a fresh allocation to avoid building nodes: 1144 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1145 if (akls != NULL) return akls; 1146 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1147 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1148 } 1149 1150 //-------------------------load_array_length----------------------------------- 1151 Node* GraphKit::load_array_length(Node* array) { 1152 // Special-case a fresh allocation to avoid building nodes: 1153 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1154 Node *alen; 1155 if (alloc == NULL) { 1156 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1157 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1158 } else { 1159 alen = alloc->Ideal_length(); 1160 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1161 if (ccast != alen) { 1162 alen = _gvn.transform(ccast); 1163 } 1164 } 1165 return alen; 1166 } 1167 1168 //------------------------------do_null_check---------------------------------- 1169 // Helper function to do a NULL pointer check. Returned value is 1170 // the incoming address with NULL casted away. You are allowed to use the 1171 // not-null value only if you are control dependent on the test. 1172 extern int explicit_null_checks_inserted, 1173 explicit_null_checks_elided; 1174 Node* GraphKit::null_check_common(Node* value, BasicType type, 1175 // optional arguments for variations: 1176 bool assert_null, 1177 Node* *null_control) { 1178 assert(!assert_null || null_control == NULL, "not both at once"); 1179 if (stopped()) return top(); 1180 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1181 // For some performance testing, we may wish to suppress null checking. 1182 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1183 return value; 1184 } 1185 explicit_null_checks_inserted++; 1186 1187 // Construct NULL check 1188 Node *chk = NULL; 1189 switch(type) { 1190 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1191 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1192 case T_ARRAY : // fall through 1193 type = T_OBJECT; // simplify further tests 1194 case T_OBJECT : { 1195 const Type *t = _gvn.type( value ); 1196 1197 const TypeOopPtr* tp = t->isa_oopptr(); 1198 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1199 // Only for do_null_check, not any of its siblings: 1200 && !assert_null && null_control == NULL) { 1201 // Usually, any field access or invocation on an unloaded oop type 1202 // will simply fail to link, since the statically linked class is 1203 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1204 // the static class is loaded but the sharper oop type is not. 1205 // Rather than checking for this obscure case in lots of places, 1206 // we simply observe that a null check on an unloaded class 1207 // will always be followed by a nonsense operation, so we 1208 // can just issue the uncommon trap here. 1209 // Our access to the unloaded class will only be correct 1210 // after it has been loaded and initialized, which requires 1211 // a trip through the interpreter. 1212 #ifndef PRODUCT 1213 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1214 #endif 1215 uncommon_trap(Deoptimization::Reason_unloaded, 1216 Deoptimization::Action_reinterpret, 1217 tp->klass(), "!loaded"); 1218 return top(); 1219 } 1220 1221 if (assert_null) { 1222 // See if the type is contained in NULL_PTR. 1223 // If so, then the value is already null. 1224 if (t->higher_equal(TypePtr::NULL_PTR)) { 1225 explicit_null_checks_elided++; 1226 return value; // Elided null assert quickly! 1227 } 1228 } else { 1229 // See if mixing in the NULL pointer changes type. 1230 // If so, then the NULL pointer was not allowed in the original 1231 // type. In other words, "value" was not-null. 1232 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1233 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1234 explicit_null_checks_elided++; 1235 return value; // Elided null check quickly! 1236 } 1237 } 1238 chk = new (C) CmpPNode( value, null() ); 1239 break; 1240 } 1241 1242 default: 1243 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1244 } 1245 assert(chk != NULL, "sanity check"); 1246 chk = _gvn.transform(chk); 1247 1248 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1249 BoolNode *btst = new (C) BoolNode( chk, btest); 1250 Node *tst = _gvn.transform( btst ); 1251 1252 //----------- 1253 // if peephole optimizations occurred, a prior test existed. 1254 // If a prior test existed, maybe it dominates as we can avoid this test. 1255 if (tst != btst && type == T_OBJECT) { 1256 // At this point we want to scan up the CFG to see if we can 1257 // find an identical test (and so avoid this test altogether). 1258 Node *cfg = control(); 1259 int depth = 0; 1260 while( depth < 16 ) { // Limit search depth for speed 1261 if( cfg->Opcode() == Op_IfTrue && 1262 cfg->in(0)->in(1) == tst ) { 1263 // Found prior test. Use "cast_not_null" to construct an identical 1264 // CastPP (and hence hash to) as already exists for the prior test. 1265 // Return that casted value. 1266 if (assert_null) { 1267 replace_in_map(value, null()); 1268 return null(); // do not issue the redundant test 1269 } 1270 Node *oldcontrol = control(); 1271 set_control(cfg); 1272 Node *res = cast_not_null(value); 1273 set_control(oldcontrol); 1274 explicit_null_checks_elided++; 1275 return res; 1276 } 1277 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1278 if (cfg == NULL) break; // Quit at region nodes 1279 depth++; 1280 } 1281 } 1282 1283 //----------- 1284 // Branch to failure if null 1285 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1286 Deoptimization::DeoptReason reason; 1287 if (assert_null) 1288 reason = Deoptimization::Reason_null_assert; 1289 else if (type == T_OBJECT) 1290 reason = Deoptimization::Reason_null_check; 1291 else 1292 reason = Deoptimization::Reason_div0_check; 1293 1294 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1295 // ciMethodData::has_trap_at will return a conservative -1 if any 1296 // must-be-null assertion has failed. This could cause performance 1297 // problems for a method after its first do_null_assert failure. 1298 // Consider using 'Reason_class_check' instead? 1299 1300 // To cause an implicit null check, we set the not-null probability 1301 // to the maximum (PROB_MAX). For an explicit check the probability 1302 // is set to a smaller value. 1303 if (null_control != NULL || too_many_traps(reason)) { 1304 // probability is less likely 1305 ok_prob = PROB_LIKELY_MAG(3); 1306 } else if (!assert_null && 1307 (ImplicitNullCheckThreshold > 0) && 1308 method() != NULL && 1309 (method()->method_data()->trap_count(reason) 1310 >= (uint)ImplicitNullCheckThreshold)) { 1311 ok_prob = PROB_LIKELY_MAG(3); 1312 } 1313 1314 if (null_control != NULL) { 1315 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1316 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1317 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1318 if (null_true == top()) 1319 explicit_null_checks_elided++; 1320 (*null_control) = null_true; 1321 } else { 1322 BuildCutout unless(this, tst, ok_prob); 1323 // Check for optimizer eliding test at parse time 1324 if (stopped()) { 1325 // Failure not possible; do not bother making uncommon trap. 1326 explicit_null_checks_elided++; 1327 } else if (assert_null) { 1328 uncommon_trap(reason, 1329 Deoptimization::Action_make_not_entrant, 1330 NULL, "assert_null"); 1331 } else { 1332 replace_in_map(value, zerocon(type)); 1333 builtin_throw(reason); 1334 } 1335 } 1336 1337 // Must throw exception, fall-thru not possible? 1338 if (stopped()) { 1339 return top(); // No result 1340 } 1341 1342 if (assert_null) { 1343 // Cast obj to null on this path. 1344 replace_in_map(value, zerocon(type)); 1345 return zerocon(type); 1346 } 1347 1348 // Cast obj to not-null on this path, if there is no null_control. 1349 // (If there is a null_control, a non-null value may come back to haunt us.) 1350 if (type == T_OBJECT) { 1351 Node* cast = cast_not_null(value, false); 1352 if (null_control == NULL || (*null_control) == top()) 1353 replace_in_map(value, cast); 1354 value = cast; 1355 } 1356 1357 return value; 1358 } 1359 1360 1361 //------------------------------cast_not_null---------------------------------- 1362 // Cast obj to not-null on this path 1363 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1364 const Type *t = _gvn.type(obj); 1365 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1366 // Object is already not-null? 1367 if( t == t_not_null ) return obj; 1368 1369 Node *cast = new (C) CastPPNode(obj,t_not_null); 1370 cast->init_req(0, control()); 1371 cast = _gvn.transform( cast ); 1372 1373 // Scan for instances of 'obj' in the current JVM mapping. 1374 // These instances are known to be not-null after the test. 1375 if (do_replace_in_map) 1376 replace_in_map(obj, cast); 1377 1378 return cast; // Return casted value 1379 } 1380 1381 1382 //--------------------------replace_in_map------------------------------------- 1383 void GraphKit::replace_in_map(Node* old, Node* neww) { 1384 if (old == neww) { 1385 return; 1386 } 1387 1388 map()->replace_edge(old, neww); 1389 1390 // Note: This operation potentially replaces any edge 1391 // on the map. This includes locals, stack, and monitors 1392 // of the current (innermost) JVM state. 1393 1394 if (!ReplaceInParentMaps) { 1395 return; 1396 } 1397 1398 // PreserveJVMState doesn't do a deep copy so we can't modify 1399 // parents 1400 if (Compile::current()->has_preserve_jvm_state()) { 1401 return; 1402 } 1403 1404 Parse* parser = is_Parse(); 1405 bool progress = true; 1406 Node* ctrl = map()->in(0); 1407 // Follow the chain of parsers and see whether the update can be 1408 // done in the map of callers. We can do the replace for a caller if 1409 // the current control post dominates the control of a caller. 1410 while (parser != NULL && parser->caller() != NULL && progress) { 1411 progress = false; 1412 Node* parent_map = parser->caller()->map(); 1413 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch"); 1414 1415 Node* parent_ctrl = parent_map->in(0); 1416 1417 while (parent_ctrl->is_Region()) { 1418 Node* n = parent_ctrl->as_Region()->is_copy(); 1419 if (n == NULL) { 1420 break; 1421 } 1422 parent_ctrl = n; 1423 } 1424 1425 for (;;) { 1426 if (ctrl == parent_ctrl) { 1427 // update the map of the exits which is the one that will be 1428 // used when compilation resume after inlining 1429 parser->exits().map()->replace_edge(old, neww); 1430 progress = true; 1431 break; 1432 } 1433 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 1434 ctrl = ctrl->in(0)->in(0); 1435 } else if (ctrl->is_Region()) { 1436 Node* n = ctrl->as_Region()->is_copy(); 1437 if (n == NULL) { 1438 break; 1439 } 1440 ctrl = n; 1441 } else { 1442 break; 1443 } 1444 } 1445 1446 parser = parser->parent_parser(); 1447 } 1448 } 1449 1450 1451 //============================================================================= 1452 //--------------------------------memory--------------------------------------- 1453 Node* GraphKit::memory(uint alias_idx) { 1454 MergeMemNode* mem = merged_memory(); 1455 Node* p = mem->memory_at(alias_idx); 1456 _gvn.set_type(p, Type::MEMORY); // must be mapped 1457 return p; 1458 } 1459 1460 //-----------------------------reset_memory------------------------------------ 1461 Node* GraphKit::reset_memory() { 1462 Node* mem = map()->memory(); 1463 // do not use this node for any more parsing! 1464 debug_only( map()->set_memory((Node*)NULL) ); 1465 return _gvn.transform( mem ); 1466 } 1467 1468 //------------------------------set_all_memory--------------------------------- 1469 void GraphKit::set_all_memory(Node* newmem) { 1470 Node* mergemem = MergeMemNode::make(C, newmem); 1471 gvn().set_type_bottom(mergemem); 1472 map()->set_memory(mergemem); 1473 } 1474 1475 //------------------------------set_all_memory_call---------------------------- 1476 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1477 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1478 set_all_memory(newmem); 1479 } 1480 1481 //============================================================================= 1482 // 1483 // parser factory methods for MemNodes 1484 // 1485 // These are layered on top of the factory methods in LoadNode and StoreNode, 1486 // and integrate with the parser's memory state and _gvn engine. 1487 // 1488 1489 // factory methods in "int adr_idx" 1490 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1491 int adr_idx, 1492 MemNode::MemOrd mo, bool require_atomic_access) { 1493 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1494 const TypePtr* adr_type = NULL; // debug-mode-only argument 1495 debug_only(adr_type = C->get_adr_type(adr_idx)); 1496 Node* mem = memory(adr_idx); 1497 Node* ld; 1498 if (require_atomic_access && bt == T_LONG) { 1499 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo); 1500 } else { 1501 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); 1502 } 1503 ld = _gvn.transform(ld); 1504 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1505 // Improve graph before escape analysis and boxing elimination. 1506 record_for_igvn(ld); 1507 } 1508 return ld; 1509 } 1510 1511 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1512 int adr_idx, 1513 MemNode::MemOrd mo, 1514 bool require_atomic_access) { 1515 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1516 const TypePtr* adr_type = NULL; 1517 debug_only(adr_type = C->get_adr_type(adr_idx)); 1518 Node *mem = memory(adr_idx); 1519 Node* st; 1520 if (require_atomic_access && bt == T_LONG) { 1521 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1522 } else { 1523 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1524 } 1525 st = _gvn.transform(st); 1526 set_memory(st, adr_idx); 1527 // Back-to-back stores can only remove intermediate store with DU info 1528 // so push on worklist for optimizer. 1529 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1530 record_for_igvn(st); 1531 1532 return st; 1533 } 1534 1535 1536 void GraphKit::pre_barrier(bool do_load, 1537 Node* ctl, 1538 Node* obj, 1539 Node* adr, 1540 uint adr_idx, 1541 Node* val, 1542 const TypeOopPtr* val_type, 1543 Node* pre_val, 1544 BasicType bt) { 1545 1546 BarrierSet* bs = Universe::heap()->barrier_set(); 1547 set_control(ctl); 1548 switch (bs->kind()) { 1549 case BarrierSet::G1SATBCT: 1550 case BarrierSet::G1SATBCTLogging: 1551 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1552 break; 1553 1554 case BarrierSet::CardTableModRef: 1555 case BarrierSet::CardTableExtension: 1556 case BarrierSet::ModRef: 1557 break; 1558 1559 case BarrierSet::Other: 1560 default : 1561 ShouldNotReachHere(); 1562 1563 } 1564 } 1565 1566 bool GraphKit::can_move_pre_barrier() const { 1567 BarrierSet* bs = Universe::heap()->barrier_set(); 1568 switch (bs->kind()) { 1569 case BarrierSet::G1SATBCT: 1570 case BarrierSet::G1SATBCTLogging: 1571 return true; // Can move it if no safepoint 1572 1573 case BarrierSet::CardTableModRef: 1574 case BarrierSet::CardTableExtension: 1575 case BarrierSet::ModRef: 1576 return true; // There is no pre-barrier 1577 1578 case BarrierSet::Other: 1579 default : 1580 ShouldNotReachHere(); 1581 } 1582 return false; 1583 } 1584 1585 void GraphKit::post_barrier(Node* ctl, 1586 Node* store, 1587 Node* obj, 1588 Node* adr, 1589 uint adr_idx, 1590 Node* val, 1591 BasicType bt, 1592 bool use_precise) { 1593 BarrierSet* bs = Universe::heap()->barrier_set(); 1594 set_control(ctl); 1595 switch (bs->kind()) { 1596 case BarrierSet::G1SATBCT: 1597 case BarrierSet::G1SATBCTLogging: 1598 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1599 break; 1600 1601 case BarrierSet::CardTableModRef: 1602 case BarrierSet::CardTableExtension: 1603 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1604 break; 1605 1606 case BarrierSet::ModRef: 1607 break; 1608 1609 case BarrierSet::Other: 1610 default : 1611 ShouldNotReachHere(); 1612 1613 } 1614 } 1615 1616 Node* GraphKit::store_oop(Node* ctl, 1617 Node* obj, 1618 Node* adr, 1619 const TypePtr* adr_type, 1620 Node* val, 1621 const TypeOopPtr* val_type, 1622 BasicType bt, 1623 bool use_precise, 1624 MemNode::MemOrd mo) { 1625 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1626 // could be delayed during Parse (for example, in adjust_map_after_if()). 1627 // Execute transformation here to avoid barrier generation in such case. 1628 if (_gvn.type(val) == TypePtr::NULL_PTR) 1629 val = _gvn.makecon(TypePtr::NULL_PTR); 1630 1631 set_control(ctl); 1632 if (stopped()) return top(); // Dead path ? 1633 1634 assert(bt == T_OBJECT, "sanity"); 1635 assert(val != NULL, "not dead path"); 1636 uint adr_idx = C->get_alias_index(adr_type); 1637 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1638 1639 pre_barrier(true /* do_load */, 1640 control(), obj, adr, adr_idx, val, val_type, 1641 NULL /* pre_val */, 1642 bt); 1643 1644 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1645 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1646 return store; 1647 } 1648 1649 // Could be an array or object we don't know at compile time (unsafe ref.) 1650 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1651 Node* obj, // containing obj 1652 Node* adr, // actual adress to store val at 1653 const TypePtr* adr_type, 1654 Node* val, 1655 BasicType bt, 1656 MemNode::MemOrd mo) { 1657 Compile::AliasType* at = C->alias_type(adr_type); 1658 const TypeOopPtr* val_type = NULL; 1659 if (adr_type->isa_instptr()) { 1660 if (at->field() != NULL) { 1661 // known field. This code is a copy of the do_put_xxx logic. 1662 ciField* field = at->field(); 1663 if (!field->type()->is_loaded()) { 1664 val_type = TypeInstPtr::BOTTOM; 1665 } else { 1666 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1667 } 1668 } 1669 } else if (adr_type->isa_aryptr()) { 1670 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1671 } 1672 if (val_type == NULL) { 1673 val_type = TypeInstPtr::BOTTOM; 1674 } 1675 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1676 } 1677 1678 1679 //-------------------------array_element_address------------------------- 1680 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1681 const TypeInt* sizetype) { 1682 uint shift = exact_log2(type2aelembytes(elembt)); 1683 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1684 1685 // short-circuit a common case (saves lots of confusing waste motion) 1686 jint idx_con = find_int_con(idx, -1); 1687 if (idx_con >= 0) { 1688 intptr_t offset = header + ((intptr_t)idx_con << shift); 1689 return basic_plus_adr(ary, offset); 1690 } 1691 1692 // must be correct type for alignment purposes 1693 Node* base = basic_plus_adr(ary, header); 1694 #ifdef _LP64 1695 // The scaled index operand to AddP must be a clean 64-bit value. 1696 // Java allows a 32-bit int to be incremented to a negative 1697 // value, which appears in a 64-bit register as a large 1698 // positive number. Using that large positive number as an 1699 // operand in pointer arithmetic has bad consequences. 1700 // On the other hand, 32-bit overflow is rare, and the possibility 1701 // can often be excluded, if we annotate the ConvI2L node with 1702 // a type assertion that its value is known to be a small positive 1703 // number. (The prior range check has ensured this.) 1704 // This assertion is used by ConvI2LNode::Ideal. 1705 int index_max = max_jint - 1; // array size is max_jint, index is one less 1706 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1707 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1708 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) ); 1709 #endif 1710 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1711 return basic_plus_adr(ary, base, scale); 1712 } 1713 1714 //-------------------------load_array_element------------------------- 1715 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1716 const Type* elemtype = arytype->elem(); 1717 BasicType elembt = elemtype->array_element_basic_type(); 1718 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1719 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1720 return ld; 1721 } 1722 1723 //-------------------------set_arguments_for_java_call------------------------- 1724 // Arguments (pre-popped from the stack) are taken from the JVMS. 1725 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1726 // Add the call arguments: 1727 uint nargs = call->method()->arg_size(); 1728 for (uint i = 0; i < nargs; i++) { 1729 Node* arg = argument(i); 1730 call->init_req(i + TypeFunc::Parms, arg); 1731 } 1732 } 1733 1734 //---------------------------set_edges_for_java_call--------------------------- 1735 // Connect a newly created call into the current JVMS. 1736 // A return value node (if any) is returned from set_edges_for_java_call. 1737 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1738 1739 // Add the predefined inputs: 1740 call->init_req( TypeFunc::Control, control() ); 1741 call->init_req( TypeFunc::I_O , i_o() ); 1742 call->init_req( TypeFunc::Memory , reset_memory() ); 1743 call->init_req( TypeFunc::FramePtr, frameptr() ); 1744 call->init_req( TypeFunc::ReturnAdr, top() ); 1745 1746 add_safepoint_edges(call, must_throw); 1747 1748 Node* xcall = _gvn.transform(call); 1749 1750 if (xcall == top()) { 1751 set_control(top()); 1752 return; 1753 } 1754 assert(xcall == call, "call identity is stable"); 1755 1756 // Re-use the current map to produce the result. 1757 1758 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1759 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1760 set_all_memory_call(xcall, separate_io_proj); 1761 1762 //return xcall; // no need, caller already has it 1763 } 1764 1765 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1766 if (stopped()) return top(); // maybe the call folded up? 1767 1768 // Capture the return value, if any. 1769 Node* ret; 1770 if (call->method() == NULL || 1771 call->method()->return_type()->basic_type() == T_VOID) 1772 ret = top(); 1773 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1774 1775 // Note: Since any out-of-line call can produce an exception, 1776 // we always insert an I_O projection from the call into the result. 1777 1778 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1779 1780 if (separate_io_proj) { 1781 // The caller requested separate projections be used by the fall 1782 // through and exceptional paths, so replace the projections for 1783 // the fall through path. 1784 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1785 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1786 } 1787 return ret; 1788 } 1789 1790 //--------------------set_predefined_input_for_runtime_call-------------------- 1791 // Reading and setting the memory state is way conservative here. 1792 // The real problem is that I am not doing real Type analysis on memory, 1793 // so I cannot distinguish card mark stores from other stores. Across a GC 1794 // point the Store Barrier and the card mark memory has to agree. I cannot 1795 // have a card mark store and its barrier split across the GC point from 1796 // either above or below. Here I get that to happen by reading ALL of memory. 1797 // A better answer would be to separate out card marks from other memory. 1798 // For now, return the input memory state, so that it can be reused 1799 // after the call, if this call has restricted memory effects. 1800 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1801 // Set fixed predefined input arguments 1802 Node* memory = reset_memory(); 1803 call->init_req( TypeFunc::Control, control() ); 1804 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1805 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1806 call->init_req( TypeFunc::FramePtr, frameptr() ); 1807 call->init_req( TypeFunc::ReturnAdr, top() ); 1808 return memory; 1809 } 1810 1811 //-------------------set_predefined_output_for_runtime_call-------------------- 1812 // Set control and memory (not i_o) from the call. 1813 // If keep_mem is not NULL, use it for the output state, 1814 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1815 // If hook_mem is NULL, this call produces no memory effects at all. 1816 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1817 // then only that memory slice is taken from the call. 1818 // In the last case, we must put an appropriate memory barrier before 1819 // the call, so as to create the correct anti-dependencies on loads 1820 // preceding the call. 1821 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1822 Node* keep_mem, 1823 const TypePtr* hook_mem) { 1824 // no i/o 1825 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1826 if (keep_mem) { 1827 // First clone the existing memory state 1828 set_all_memory(keep_mem); 1829 if (hook_mem != NULL) { 1830 // Make memory for the call 1831 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1832 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1833 // We also use hook_mem to extract specific effects from arraycopy stubs. 1834 set_memory(mem, hook_mem); 1835 } 1836 // ...else the call has NO memory effects. 1837 1838 // Make sure the call advertises its memory effects precisely. 1839 // This lets us build accurate anti-dependences in gcm.cpp. 1840 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1841 "call node must be constructed correctly"); 1842 } else { 1843 assert(hook_mem == NULL, ""); 1844 // This is not a "slow path" call; all memory comes from the call. 1845 set_all_memory_call(call); 1846 } 1847 } 1848 1849 1850 // Replace the call with the current state of the kit. 1851 void GraphKit::replace_call(CallNode* call, Node* result) { 1852 JVMState* ejvms = NULL; 1853 if (has_exceptions()) { 1854 ejvms = transfer_exceptions_into_jvms(); 1855 } 1856 1857 SafePointNode* final_state = stop(); 1858 1859 // Find all the needed outputs of this call 1860 CallProjections callprojs; 1861 call->extract_projections(&callprojs, true); 1862 1863 Node* init_mem = call->in(TypeFunc::Memory); 1864 Node* final_mem = final_state->in(TypeFunc::Memory); 1865 Node* final_ctl = final_state->in(TypeFunc::Control); 1866 Node* final_io = final_state->in(TypeFunc::I_O); 1867 1868 // Replace all the old call edges with the edges from the inlining result 1869 if (callprojs.fallthrough_catchproj != NULL) { 1870 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1871 } 1872 if (callprojs.fallthrough_memproj != NULL) { 1873 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1874 } 1875 if (callprojs.fallthrough_ioproj != NULL) { 1876 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1877 } 1878 1879 // Replace the result with the new result if it exists and is used 1880 if (callprojs.resproj != NULL && result != NULL) { 1881 C->gvn_replace_by(callprojs.resproj, result); 1882 } 1883 1884 if (ejvms == NULL) { 1885 // No exception edges to simply kill off those paths 1886 if (callprojs.catchall_catchproj != NULL) { 1887 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1888 } 1889 if (callprojs.catchall_memproj != NULL) { 1890 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1891 } 1892 if (callprojs.catchall_ioproj != NULL) { 1893 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1894 } 1895 // Replace the old exception object with top 1896 if (callprojs.exobj != NULL) { 1897 C->gvn_replace_by(callprojs.exobj, C->top()); 1898 } 1899 } else { 1900 GraphKit ekit(ejvms); 1901 1902 // Load my combined exception state into the kit, with all phis transformed: 1903 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1904 1905 Node* ex_oop = ekit.use_exception_state(ex_map); 1906 if (callprojs.catchall_catchproj != NULL) { 1907 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1908 } 1909 if (callprojs.catchall_memproj != NULL) { 1910 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1911 } 1912 if (callprojs.catchall_ioproj != NULL) { 1913 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1914 } 1915 1916 // Replace the old exception object with the newly created one 1917 if (callprojs.exobj != NULL) { 1918 C->gvn_replace_by(callprojs.exobj, ex_oop); 1919 } 1920 } 1921 1922 // Disconnect the call from the graph 1923 call->disconnect_inputs(NULL, C); 1924 C->gvn_replace_by(call, C->top()); 1925 1926 // Clean up any MergeMems that feed other MergeMems since the 1927 // optimizer doesn't like that. 1928 if (final_mem->is_MergeMem()) { 1929 Node_List wl; 1930 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1931 Node* m = i.get(); 1932 if (m->is_MergeMem() && !wl.contains(m)) { 1933 wl.push(m); 1934 } 1935 } 1936 while (wl.size() > 0) { 1937 _gvn.transform(wl.pop()); 1938 } 1939 } 1940 } 1941 1942 1943 //------------------------------increment_counter------------------------------ 1944 // for statistics: increment a VM counter by 1 1945 1946 void GraphKit::increment_counter(address counter_addr) { 1947 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1948 increment_counter(adr1); 1949 } 1950 1951 void GraphKit::increment_counter(Node* counter_addr) { 1952 int adr_type = Compile::AliasIdxRaw; 1953 Node* ctrl = control(); 1954 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1955 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1956 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1957 } 1958 1959 1960 //------------------------------uncommon_trap---------------------------------- 1961 // Bail out to the interpreter in mid-method. Implemented by calling the 1962 // uncommon_trap blob. This helper function inserts a runtime call with the 1963 // right debug info. 1964 void GraphKit::uncommon_trap(int trap_request, 1965 ciKlass* klass, const char* comment, 1966 bool must_throw, 1967 bool keep_exact_action) { 1968 if (failing()) stop(); 1969 if (stopped()) return; // trap reachable? 1970 1971 // Note: If ProfileTraps is true, and if a deopt. actually 1972 // occurs here, the runtime will make sure an MDO exists. There is 1973 // no need to call method()->ensure_method_data() at this point. 1974 1975 // Set the stack pointer to the right value for reexecution: 1976 set_sp(reexecute_sp()); 1977 1978 #ifdef ASSERT 1979 if (!must_throw) { 1980 // Make sure the stack has at least enough depth to execute 1981 // the current bytecode. 1982 int inputs, ignored_depth; 1983 if (compute_stack_effects(inputs, ignored_depth)) { 1984 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1985 Bytecodes::name(java_bc()), sp(), inputs)); 1986 } 1987 } 1988 #endif 1989 1990 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1991 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1992 1993 switch (action) { 1994 case Deoptimization::Action_maybe_recompile: 1995 case Deoptimization::Action_reinterpret: 1996 // Temporary fix for 6529811 to allow virtual calls to be sure they 1997 // get the chance to go from mono->bi->mega 1998 if (!keep_exact_action && 1999 Deoptimization::trap_request_index(trap_request) < 0 && 2000 too_many_recompiles(reason)) { 2001 // This BCI is causing too many recompilations. 2002 action = Deoptimization::Action_none; 2003 trap_request = Deoptimization::make_trap_request(reason, action); 2004 } else { 2005 C->set_trap_can_recompile(true); 2006 } 2007 break; 2008 case Deoptimization::Action_make_not_entrant: 2009 C->set_trap_can_recompile(true); 2010 break; 2011 #ifdef ASSERT 2012 case Deoptimization::Action_none: 2013 case Deoptimization::Action_make_not_compilable: 2014 break; 2015 default: 2016 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2017 break; 2018 #endif 2019 } 2020 2021 if (TraceOptoParse) { 2022 char buf[100]; 2023 tty->print_cr("Uncommon trap %s at bci:%d", 2024 Deoptimization::format_trap_request(buf, sizeof(buf), 2025 trap_request), bci()); 2026 } 2027 2028 CompileLog* log = C->log(); 2029 if (log != NULL) { 2030 int kid = (klass == NULL)? -1: log->identify(klass); 2031 log->begin_elem("uncommon_trap bci='%d'", bci()); 2032 char buf[100]; 2033 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2034 trap_request)); 2035 if (kid >= 0) log->print(" klass='%d'", kid); 2036 if (comment != NULL) log->print(" comment='%s'", comment); 2037 log->end_elem(); 2038 } 2039 2040 // Make sure any guarding test views this path as very unlikely 2041 Node *i0 = control()->in(0); 2042 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2043 IfNode *iff = i0->as_If(); 2044 float f = iff->_prob; // Get prob 2045 if (control()->Opcode() == Op_IfTrue) { 2046 if (f > PROB_UNLIKELY_MAG(4)) 2047 iff->_prob = PROB_MIN; 2048 } else { 2049 if (f < PROB_LIKELY_MAG(4)) 2050 iff->_prob = PROB_MAX; 2051 } 2052 } 2053 2054 // Clear out dead values from the debug info. 2055 kill_dead_locals(); 2056 2057 // Now insert the uncommon trap subroutine call 2058 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2059 const TypePtr* no_memory_effects = NULL; 2060 // Pass the index of the class to be loaded 2061 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2062 (must_throw ? RC_MUST_THROW : 0), 2063 OptoRuntime::uncommon_trap_Type(), 2064 call_addr, "uncommon_trap", no_memory_effects, 2065 intcon(trap_request)); 2066 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2067 "must extract request correctly from the graph"); 2068 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2069 2070 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2071 // The debug info is the only real input to this call. 2072 2073 // Halt-and-catch fire here. The above call should never return! 2074 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2075 _gvn.set_type_bottom(halt); 2076 root()->add_req(halt); 2077 2078 stop_and_kill_map(); 2079 } 2080 2081 2082 //--------------------------just_allocated_object------------------------------ 2083 // Report the object that was just allocated. 2084 // It must be the case that there are no intervening safepoints. 2085 // We use this to determine if an object is so "fresh" that 2086 // it does not require card marks. 2087 Node* GraphKit::just_allocated_object(Node* current_control) { 2088 if (C->recent_alloc_ctl() == current_control) 2089 return C->recent_alloc_obj(); 2090 return NULL; 2091 } 2092 2093 2094 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2095 // (Note: TypeFunc::make has a cache that makes this fast.) 2096 const TypeFunc* tf = TypeFunc::make(dest_method); 2097 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2098 for (int j = 0; j < nargs; j++) { 2099 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2100 if( targ->basic_type() == T_DOUBLE ) { 2101 // If any parameters are doubles, they must be rounded before 2102 // the call, dstore_rounding does gvn.transform 2103 Node *arg = argument(j); 2104 arg = dstore_rounding(arg); 2105 set_argument(j, arg); 2106 } 2107 } 2108 } 2109 2110 /** 2111 * Record profiling data exact_kls for Node n with the type system so 2112 * that it can propagate it (speculation) 2113 * 2114 * @param n node that the type applies to 2115 * @param exact_kls type from profiling 2116 * 2117 * @return node with improved type 2118 */ 2119 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) { 2120 const Type* current_type = _gvn.type(n); 2121 assert(UseTypeSpeculation, "type speculation must be on"); 2122 2123 const TypeOopPtr* speculative = current_type->speculative(); 2124 2125 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2126 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2127 const TypeOopPtr* xtype = tklass->as_instance_type(); 2128 assert(xtype->klass_is_exact(), "Should be exact"); 2129 // record the new speculative type's depth 2130 speculative = xtype->with_inline_depth(jvms()->depth()); 2131 } 2132 2133 if (speculative != current_type->speculative()) { 2134 // Build a type with a speculative type (what we think we know 2135 // about the type but will need a guard when we use it) 2136 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2137 // We're changing the type, we need a new CheckCast node to carry 2138 // the new type. The new type depends on the control: what 2139 // profiling tells us is only valid from here as far as we can 2140 // tell. 2141 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2142 cast = _gvn.transform(cast); 2143 replace_in_map(n, cast); 2144 n = cast; 2145 } 2146 2147 return n; 2148 } 2149 2150 /** 2151 * Record profiling data from receiver profiling at an invoke with the 2152 * type system so that it can propagate it (speculation) 2153 * 2154 * @param n receiver node 2155 * 2156 * @return node with improved type 2157 */ 2158 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2159 if (!UseTypeSpeculation) { 2160 return n; 2161 } 2162 ciKlass* exact_kls = profile_has_unique_klass(); 2163 return record_profile_for_speculation(n, exact_kls); 2164 } 2165 2166 /** 2167 * Record profiling data from argument profiling at an invoke with the 2168 * type system so that it can propagate it (speculation) 2169 * 2170 * @param dest_method target method for the call 2171 * @param bc what invoke bytecode is this? 2172 */ 2173 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2174 if (!UseTypeSpeculation) { 2175 return; 2176 } 2177 const TypeFunc* tf = TypeFunc::make(dest_method); 2178 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2179 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2180 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2181 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2182 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2183 ciKlass* better_type = method()->argument_profiled_type(bci(), i); 2184 if (better_type != NULL) { 2185 record_profile_for_speculation(argument(j), better_type); 2186 } 2187 i++; 2188 } 2189 } 2190 } 2191 2192 /** 2193 * Record profiling data from parameter profiling at an invoke with 2194 * the type system so that it can propagate it (speculation) 2195 */ 2196 void GraphKit::record_profiled_parameters_for_speculation() { 2197 if (!UseTypeSpeculation) { 2198 return; 2199 } 2200 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2201 if (_gvn.type(local(i))->isa_oopptr()) { 2202 ciKlass* better_type = method()->parameter_profiled_type(j); 2203 if (better_type != NULL) { 2204 record_profile_for_speculation(local(i), better_type); 2205 } 2206 j++; 2207 } 2208 } 2209 } 2210 2211 void GraphKit::round_double_result(ciMethod* dest_method) { 2212 // A non-strict method may return a double value which has an extended 2213 // exponent, but this must not be visible in a caller which is 'strict' 2214 // If a strict caller invokes a non-strict callee, round a double result 2215 2216 BasicType result_type = dest_method->return_type()->basic_type(); 2217 assert( method() != NULL, "must have caller context"); 2218 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2219 // Destination method's return value is on top of stack 2220 // dstore_rounding() does gvn.transform 2221 Node *result = pop_pair(); 2222 result = dstore_rounding(result); 2223 push_pair(result); 2224 } 2225 } 2226 2227 // rounding for strict float precision conformance 2228 Node* GraphKit::precision_rounding(Node* n) { 2229 return UseStrictFP && _method->flags().is_strict() 2230 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2231 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2232 : n; 2233 } 2234 2235 // rounding for strict double precision conformance 2236 Node* GraphKit::dprecision_rounding(Node *n) { 2237 return UseStrictFP && _method->flags().is_strict() 2238 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2239 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2240 : n; 2241 } 2242 2243 // rounding for non-strict double stores 2244 Node* GraphKit::dstore_rounding(Node* n) { 2245 return Matcher::strict_fp_requires_explicit_rounding 2246 && UseSSE <= 1 2247 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2248 : n; 2249 } 2250 2251 //============================================================================= 2252 // Generate a fast path/slow path idiom. Graph looks like: 2253 // [foo] indicates that 'foo' is a parameter 2254 // 2255 // [in] NULL 2256 // \ / 2257 // CmpP 2258 // Bool ne 2259 // If 2260 // / \ 2261 // True False-<2> 2262 // / | 2263 // / cast_not_null 2264 // Load | | ^ 2265 // [fast_test] | | 2266 // gvn to opt_test | | 2267 // / \ | <1> 2268 // True False | 2269 // | \\ | 2270 // [slow_call] \[fast_result] 2271 // Ctl Val \ \ 2272 // | \ \ 2273 // Catch <1> \ \ 2274 // / \ ^ \ \ 2275 // Ex No_Ex | \ \ 2276 // | \ \ | \ <2> \ 2277 // ... \ [slow_res] | | \ [null_result] 2278 // \ \--+--+--- | | 2279 // \ | / \ | / 2280 // --------Region Phi 2281 // 2282 //============================================================================= 2283 // Code is structured as a series of driver functions all called 'do_XXX' that 2284 // call a set of helper functions. Helper functions first, then drivers. 2285 2286 //------------------------------null_check_oop--------------------------------- 2287 // Null check oop. Set null-path control into Region in slot 3. 2288 // Make a cast-not-nullness use the other not-null control. Return cast. 2289 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2290 bool never_see_null, bool safe_for_replace) { 2291 // Initial NULL check taken path 2292 (*null_control) = top(); 2293 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2294 2295 // Generate uncommon_trap: 2296 if (never_see_null && (*null_control) != top()) { 2297 // If we see an unexpected null at a check-cast we record it and force a 2298 // recompile; the offending check-cast will be compiled to handle NULLs. 2299 // If we see more than one offending BCI, then all checkcasts in the 2300 // method will be compiled to handle NULLs. 2301 PreserveJVMState pjvms(this); 2302 set_control(*null_control); 2303 replace_in_map(value, null()); 2304 uncommon_trap(Deoptimization::Reason_null_check, 2305 Deoptimization::Action_make_not_entrant); 2306 (*null_control) = top(); // NULL path is dead 2307 } 2308 if ((*null_control) == top() && safe_for_replace) { 2309 replace_in_map(value, cast); 2310 } 2311 2312 // Cast away null-ness on the result 2313 return cast; 2314 } 2315 2316 //------------------------------opt_iff---------------------------------------- 2317 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2318 // Return slow-path control. 2319 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2320 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2321 2322 // Fast path taken; set region slot 2 2323 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2324 region->init_req(2,fast_taken); // Capture fast-control 2325 2326 // Fast path not-taken, i.e. slow path 2327 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2328 return slow_taken; 2329 } 2330 2331 //-----------------------------make_runtime_call------------------------------- 2332 Node* GraphKit::make_runtime_call(int flags, 2333 const TypeFunc* call_type, address call_addr, 2334 const char* call_name, 2335 const TypePtr* adr_type, 2336 // The following parms are all optional. 2337 // The first NULL ends the list. 2338 Node* parm0, Node* parm1, 2339 Node* parm2, Node* parm3, 2340 Node* parm4, Node* parm5, 2341 Node* parm6, Node* parm7) { 2342 // Slow-path call 2343 bool is_leaf = !(flags & RC_NO_LEAF); 2344 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2345 if (call_name == NULL) { 2346 assert(!is_leaf, "must supply name for leaf"); 2347 call_name = OptoRuntime::stub_name(call_addr); 2348 } 2349 CallNode* call; 2350 if (!is_leaf) { 2351 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2352 bci(), adr_type); 2353 } else if (flags & RC_NO_FP) { 2354 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2355 } else { 2356 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2357 } 2358 2359 // The following is similar to set_edges_for_java_call, 2360 // except that the memory effects of the call are restricted to AliasIdxRaw. 2361 2362 // Slow path call has no side-effects, uses few values 2363 bool wide_in = !(flags & RC_NARROW_MEM); 2364 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2365 2366 Node* prev_mem = NULL; 2367 if (wide_in) { 2368 prev_mem = set_predefined_input_for_runtime_call(call); 2369 } else { 2370 assert(!wide_out, "narrow in => narrow out"); 2371 Node* narrow_mem = memory(adr_type); 2372 prev_mem = reset_memory(); 2373 map()->set_memory(narrow_mem); 2374 set_predefined_input_for_runtime_call(call); 2375 } 2376 2377 // Hook each parm in order. Stop looking at the first NULL. 2378 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2379 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2380 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2381 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2382 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2383 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2384 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2385 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2386 /* close each nested if ===> */ } } } } } } } } 2387 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2388 2389 if (!is_leaf) { 2390 // Non-leaves can block and take safepoints: 2391 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2392 } 2393 // Non-leaves can throw exceptions: 2394 if (has_io) { 2395 call->set_req(TypeFunc::I_O, i_o()); 2396 } 2397 2398 if (flags & RC_UNCOMMON) { 2399 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2400 // (An "if" probability corresponds roughly to an unconditional count. 2401 // Sort of.) 2402 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2403 } 2404 2405 Node* c = _gvn.transform(call); 2406 assert(c == call, "cannot disappear"); 2407 2408 if (wide_out) { 2409 // Slow path call has full side-effects. 2410 set_predefined_output_for_runtime_call(call); 2411 } else { 2412 // Slow path call has few side-effects, and/or sets few values. 2413 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2414 } 2415 2416 if (has_io) { 2417 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2418 } 2419 return call; 2420 2421 } 2422 2423 //------------------------------merge_memory----------------------------------- 2424 // Merge memory from one path into the current memory state. 2425 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2426 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2427 Node* old_slice = mms.force_memory(); 2428 Node* new_slice = mms.memory2(); 2429 if (old_slice != new_slice) { 2430 PhiNode* phi; 2431 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2432 phi = new_slice->as_Phi(); 2433 #ifdef ASSERT 2434 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2435 old_slice = old_slice->in(new_path); 2436 // Caller is responsible for ensuring that any pre-existing 2437 // phis are already aware of old memory. 2438 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2439 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2440 #endif 2441 mms.set_memory(phi); 2442 } else { 2443 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2444 _gvn.set_type(phi, Type::MEMORY); 2445 phi->set_req(new_path, new_slice); 2446 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2447 } 2448 } 2449 } 2450 } 2451 2452 //------------------------------make_slow_call_ex------------------------------ 2453 // Make the exception handler hookups for the slow call 2454 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2455 if (stopped()) return; 2456 2457 // Make a catch node with just two handlers: fall-through and catch-all 2458 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2459 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2460 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2461 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2462 2463 { PreserveJVMState pjvms(this); 2464 set_control(excp); 2465 set_i_o(i_o); 2466 2467 if (excp != top()) { 2468 // Create an exception state also. 2469 // Use an exact type if the caller has specified a specific exception. 2470 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2471 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2472 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2473 } 2474 } 2475 2476 // Get the no-exception control from the CatchNode. 2477 set_control(norm); 2478 } 2479 2480 2481 //-------------------------------gen_subtype_check----------------------------- 2482 // Generate a subtyping check. Takes as input the subtype and supertype. 2483 // Returns 2 values: sets the default control() to the true path and returns 2484 // the false path. Only reads invariant memory; sets no (visible) memory. 2485 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2486 // but that's not exposed to the optimizer. This call also doesn't take in an 2487 // Object; if you wish to check an Object you need to load the Object's class 2488 // prior to coming here. 2489 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2490 // Fast check for identical types, perhaps identical constants. 2491 // The types can even be identical non-constants, in cases 2492 // involving Array.newInstance, Object.clone, etc. 2493 if (subklass == superklass) 2494 return top(); // false path is dead; no test needed. 2495 2496 if (_gvn.type(superklass)->singleton()) { 2497 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2498 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2499 2500 // In the common case of an exact superklass, try to fold up the 2501 // test before generating code. You may ask, why not just generate 2502 // the code and then let it fold up? The answer is that the generated 2503 // code will necessarily include null checks, which do not always 2504 // completely fold away. If they are also needless, then they turn 2505 // into a performance loss. Example: 2506 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2507 // Here, the type of 'fa' is often exact, so the store check 2508 // of fa[1]=x will fold up, without testing the nullness of x. 2509 switch (static_subtype_check(superk, subk)) { 2510 case SSC_always_false: 2511 { 2512 Node* always_fail = control(); 2513 set_control(top()); 2514 return always_fail; 2515 } 2516 case SSC_always_true: 2517 return top(); 2518 case SSC_easy_test: 2519 { 2520 // Just do a direct pointer compare and be done. 2521 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2522 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2523 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2524 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2525 return _gvn.transform( new(C) IfFalseNode(iff) ); 2526 } 2527 case SSC_full_test: 2528 break; 2529 default: 2530 ShouldNotReachHere(); 2531 } 2532 } 2533 2534 // %%% Possible further optimization: Even if the superklass is not exact, 2535 // if the subklass is the unique subtype of the superklass, the check 2536 // will always succeed. We could leave a dependency behind to ensure this. 2537 2538 // First load the super-klass's check-offset 2539 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2540 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), 2541 TypeInt::INT, MemNode::unordered)); 2542 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2543 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2544 2545 // Load from the sub-klass's super-class display list, or a 1-word cache of 2546 // the secondary superclass list, or a failing value with a sentinel offset 2547 // if the super-klass is an interface or exceptionally deep in the Java 2548 // hierarchy and we have to scan the secondary superclass list the hard way. 2549 // Worst-case type is a little odd: NULL is allowed as a result (usually 2550 // klass loads can never produce a NULL). 2551 Node *chk_off_X = ConvI2X(chk_off); 2552 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2553 // For some types like interfaces the following loadKlass is from a 1-word 2554 // cache which is mutable so can't use immutable memory. Other 2555 // types load from the super-class display table which is immutable. 2556 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2557 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2558 2559 // Compile speed common case: ARE a subtype and we canNOT fail 2560 if( superklass == nkls ) 2561 return top(); // false path is dead; no test needed. 2562 2563 // See if we get an immediate positive hit. Happens roughly 83% of the 2564 // time. Test to see if the value loaded just previously from the subklass 2565 // is exactly the superklass. 2566 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2567 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2568 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2569 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2570 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2571 2572 // Compile speed common case: Check for being deterministic right now. If 2573 // chk_off is a constant and not equal to cacheoff then we are NOT a 2574 // subklass. In this case we need exactly the 1 test above and we can 2575 // return those results immediately. 2576 if (!might_be_cache) { 2577 Node* not_subtype_ctrl = control(); 2578 set_control(iftrue1); // We need exactly the 1 test above 2579 return not_subtype_ctrl; 2580 } 2581 2582 // Gather the various success & failures here 2583 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2584 record_for_igvn(r_ok_subtype); 2585 RegionNode *r_not_subtype = new (C) RegionNode(3); 2586 record_for_igvn(r_not_subtype); 2587 2588 r_ok_subtype->init_req(1, iftrue1); 2589 2590 // Check for immediate negative hit. Happens roughly 11% of the time (which 2591 // is roughly 63% of the remaining cases). Test to see if the loaded 2592 // check-offset points into the subklass display list or the 1-element 2593 // cache. If it points to the display (and NOT the cache) and the display 2594 // missed then it's not a subtype. 2595 Node *cacheoff = _gvn.intcon(cacheoff_con); 2596 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2597 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2598 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2599 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2600 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2601 2602 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2603 // No performance impact (too rare) but allows sharing of secondary arrays 2604 // which has some footprint reduction. 2605 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2606 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2607 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2608 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2609 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2610 2611 // -- Roads not taken here: -- 2612 // We could also have chosen to perform the self-check at the beginning 2613 // of this code sequence, as the assembler does. This would not pay off 2614 // the same way, since the optimizer, unlike the assembler, can perform 2615 // static type analysis to fold away many successful self-checks. 2616 // Non-foldable self checks work better here in second position, because 2617 // the initial primary superclass check subsumes a self-check for most 2618 // types. An exception would be a secondary type like array-of-interface, 2619 // which does not appear in its own primary supertype display. 2620 // Finally, we could have chosen to move the self-check into the 2621 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2622 // dependent manner. But it is worthwhile to have the check here, 2623 // where it can be perhaps be optimized. The cost in code space is 2624 // small (register compare, branch). 2625 2626 // Now do a linear scan of the secondary super-klass array. Again, no real 2627 // performance impact (too rare) but it's gotta be done. 2628 // Since the code is rarely used, there is no penalty for moving it 2629 // out of line, and it can only improve I-cache density. 2630 // The decision to inline or out-of-line this final check is platform 2631 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2632 Node* psc = _gvn.transform( 2633 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2634 2635 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2636 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2637 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2638 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2639 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2640 2641 // Return false path; set default control to true path. 2642 set_control( _gvn.transform(r_ok_subtype) ); 2643 return _gvn.transform(r_not_subtype); 2644 } 2645 2646 //----------------------------static_subtype_check----------------------------- 2647 // Shortcut important common cases when superklass is exact: 2648 // (0) superklass is java.lang.Object (can occur in reflective code) 2649 // (1) subklass is already limited to a subtype of superklass => always ok 2650 // (2) subklass does not overlap with superklass => always fail 2651 // (3) superklass has NO subtypes and we can check with a simple compare. 2652 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2653 if (StressReflectiveCode) { 2654 return SSC_full_test; // Let caller generate the general case. 2655 } 2656 2657 if (superk == env()->Object_klass()) { 2658 return SSC_always_true; // (0) this test cannot fail 2659 } 2660 2661 ciType* superelem = superk; 2662 if (superelem->is_array_klass()) 2663 superelem = superelem->as_array_klass()->base_element_type(); 2664 2665 if (!subk->is_interface()) { // cannot trust static interface types yet 2666 if (subk->is_subtype_of(superk)) { 2667 return SSC_always_true; // (1) false path dead; no dynamic test needed 2668 } 2669 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2670 !superk->is_subtype_of(subk)) { 2671 return SSC_always_false; 2672 } 2673 } 2674 2675 // If casting to an instance klass, it must have no subtypes 2676 if (superk->is_interface()) { 2677 // Cannot trust interfaces yet. 2678 // %%% S.B. superk->nof_implementors() == 1 2679 } else if (superelem->is_instance_klass()) { 2680 ciInstanceKlass* ik = superelem->as_instance_klass(); 2681 if (!ik->has_subklass() && !ik->is_interface()) { 2682 if (!ik->is_final()) { 2683 // Add a dependency if there is a chance of a later subclass. 2684 C->dependencies()->assert_leaf_type(ik); 2685 } 2686 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2687 } 2688 } else { 2689 // A primitive array type has no subtypes. 2690 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2691 } 2692 2693 return SSC_full_test; 2694 } 2695 2696 // Profile-driven exact type check: 2697 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2698 float prob, 2699 Node* *casted_receiver) { 2700 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2701 Node* recv_klass = load_object_klass(receiver); 2702 Node* want_klass = makecon(tklass); 2703 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2704 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2705 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2706 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2707 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2708 2709 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2710 assert(recv_xtype->klass_is_exact(), ""); 2711 2712 // Subsume downstream occurrences of receiver with a cast to 2713 // recv_xtype, since now we know what the type will be. 2714 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2715 (*casted_receiver) = _gvn.transform(cast); 2716 // (User must make the replace_in_map call.) 2717 2718 return fail; 2719 } 2720 2721 2722 //------------------------------seems_never_null------------------------------- 2723 // Use null_seen information if it is available from the profile. 2724 // If we see an unexpected null at a type check we record it and force a 2725 // recompile; the offending check will be recompiled to handle NULLs. 2726 // If we see several offending BCIs, then all checks in the 2727 // method will be recompiled. 2728 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2729 if (UncommonNullCast // Cutout for this technique 2730 && obj != null() // And not the -Xcomp stupid case? 2731 && !too_many_traps(Deoptimization::Reason_null_check) 2732 ) { 2733 if (data == NULL) 2734 // Edge case: no mature data. Be optimistic here. 2735 return true; 2736 // If the profile has not seen a null, assume it won't happen. 2737 assert(java_bc() == Bytecodes::_checkcast || 2738 java_bc() == Bytecodes::_instanceof || 2739 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2740 return !data->as_BitData()->null_seen(); 2741 } 2742 return false; 2743 } 2744 2745 //------------------------maybe_cast_profiled_receiver------------------------- 2746 // If the profile has seen exactly one type, narrow to exactly that type. 2747 // Subsequent type checks will always fold up. 2748 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2749 ciKlass* require_klass, 2750 ciKlass* spec_klass, 2751 bool safe_for_replace) { 2752 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2753 2754 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check; 2755 2756 // Make sure we haven't already deoptimized from this tactic. 2757 if (too_many_traps(reason)) 2758 return NULL; 2759 2760 // (No, this isn't a call, but it's enough like a virtual call 2761 // to use the same ciMethod accessor to get the profile info...) 2762 // If we have a speculative type use it instead of profiling (which 2763 // may not help us) 2764 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2765 if (exact_kls != NULL) {// no cast failures here 2766 if (require_klass == NULL || 2767 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2768 // If we narrow the type to match what the type profile sees or 2769 // the speculative type, we can then remove the rest of the 2770 // cast. 2771 // This is a win, even if the exact_kls is very specific, 2772 // because downstream operations, such as method calls, 2773 // will often benefit from the sharper type. 2774 Node* exact_obj = not_null_obj; // will get updated in place... 2775 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2776 &exact_obj); 2777 { PreserveJVMState pjvms(this); 2778 set_control(slow_ctl); 2779 uncommon_trap(reason, 2780 Deoptimization::Action_maybe_recompile); 2781 } 2782 if (safe_for_replace) { 2783 replace_in_map(not_null_obj, exact_obj); 2784 } 2785 return exact_obj; 2786 } 2787 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2788 } 2789 2790 return NULL; 2791 } 2792 2793 /** 2794 * Cast obj to type and emit guard unless we had too many traps here 2795 * already 2796 * 2797 * @param obj node being casted 2798 * @param type type to cast the node to 2799 * @param not_null true if we know node cannot be null 2800 */ 2801 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2802 ciKlass* type, 2803 bool not_null) { 2804 // type == NULL if profiling tells us this object is always null 2805 if (type != NULL) { 2806 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2807 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check; 2808 if (!too_many_traps(null_reason) && 2809 !too_many_traps(class_reason)) { 2810 Node* not_null_obj = NULL; 2811 // not_null is true if we know the object is not null and 2812 // there's no need for a null check 2813 if (!not_null) { 2814 Node* null_ctl = top(); 2815 not_null_obj = null_check_oop(obj, &null_ctl, true, true); 2816 assert(null_ctl->is_top(), "no null control here"); 2817 } else { 2818 not_null_obj = obj; 2819 } 2820 2821 Node* exact_obj = not_null_obj; 2822 ciKlass* exact_kls = type; 2823 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2824 &exact_obj); 2825 { 2826 PreserveJVMState pjvms(this); 2827 set_control(slow_ctl); 2828 uncommon_trap(class_reason, 2829 Deoptimization::Action_maybe_recompile); 2830 } 2831 replace_in_map(not_null_obj, exact_obj); 2832 obj = exact_obj; 2833 } 2834 } else { 2835 if (!too_many_traps(Deoptimization::Reason_null_assert)) { 2836 Node* exact_obj = null_assert(obj); 2837 replace_in_map(obj, exact_obj); 2838 obj = exact_obj; 2839 } 2840 } 2841 return obj; 2842 } 2843 2844 //-------------------------------gen_instanceof-------------------------------- 2845 // Generate an instance-of idiom. Used by both the instance-of bytecode 2846 // and the reflective instance-of call. 2847 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2848 kill_dead_locals(); // Benefit all the uncommon traps 2849 assert( !stopped(), "dead parse path should be checked in callers" ); 2850 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2851 "must check for not-null not-dead klass in callers"); 2852 2853 // Make the merge point 2854 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2855 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2856 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2857 C->set_has_split_ifs(true); // Has chance for split-if optimization 2858 2859 ciProfileData* data = NULL; 2860 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2861 data = method()->method_data()->bci_to_data(bci()); 2862 } 2863 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2864 && seems_never_null(obj, data)); 2865 2866 // Null check; get casted pointer; set region slot 3 2867 Node* null_ctl = top(); 2868 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2869 2870 // If not_null_obj is dead, only null-path is taken 2871 if (stopped()) { // Doing instance-of on a NULL? 2872 set_control(null_ctl); 2873 return intcon(0); 2874 } 2875 region->init_req(_null_path, null_ctl); 2876 phi ->init_req(_null_path, intcon(0)); // Set null path value 2877 if (null_ctl == top()) { 2878 // Do this eagerly, so that pattern matches like is_diamond_phi 2879 // will work even during parsing. 2880 assert(_null_path == PATH_LIMIT-1, "delete last"); 2881 region->del_req(_null_path); 2882 phi ->del_req(_null_path); 2883 } 2884 2885 // Do we know the type check always succeed? 2886 bool known_statically = false; 2887 if (_gvn.type(superklass)->singleton()) { 2888 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2889 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2890 if (subk != NULL && subk->is_loaded()) { 2891 int static_res = static_subtype_check(superk, subk); 2892 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2893 } 2894 } 2895 2896 if (known_statically && UseTypeSpeculation) { 2897 // If we know the type check always succeeds then we don't use the 2898 // profiling data at this bytecode. Don't lose it, feed it to the 2899 // type system as a speculative type. 2900 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2901 } else { 2902 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2903 // We may not have profiling here or it may not help us. If we 2904 // have a speculative type use it to perform an exact cast. 2905 ciKlass* spec_obj_type = obj_type->speculative_type(); 2906 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2907 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2908 if (stopped()) { // Profile disagrees with this path. 2909 set_control(null_ctl); // Null is the only remaining possibility. 2910 return intcon(0); 2911 } 2912 if (cast_obj != NULL) { 2913 not_null_obj = cast_obj; 2914 } 2915 } 2916 } 2917 2918 // Load the object's klass 2919 Node* obj_klass = load_object_klass(not_null_obj); 2920 2921 // Generate the subtype check 2922 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2923 2924 // Plug in the success path to the general merge in slot 1. 2925 region->init_req(_obj_path, control()); 2926 phi ->init_req(_obj_path, intcon(1)); 2927 2928 // Plug in the failing path to the general merge in slot 2. 2929 region->init_req(_fail_path, not_subtype_ctrl); 2930 phi ->init_req(_fail_path, intcon(0)); 2931 2932 // Return final merged results 2933 set_control( _gvn.transform(region) ); 2934 record_for_igvn(region); 2935 return _gvn.transform(phi); 2936 } 2937 2938 //-------------------------------gen_checkcast--------------------------------- 2939 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2940 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2941 // uncommon-trap paths work. Adjust stack after this call. 2942 // If failure_control is supplied and not null, it is filled in with 2943 // the control edge for the cast failure. Otherwise, an appropriate 2944 // uncommon trap or exception is thrown. 2945 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2946 Node* *failure_control) { 2947 kill_dead_locals(); // Benefit all the uncommon traps 2948 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2949 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2950 2951 // Fast cutout: Check the case that the cast is vacuously true. 2952 // This detects the common cases where the test will short-circuit 2953 // away completely. We do this before we perform the null check, 2954 // because if the test is going to turn into zero code, we don't 2955 // want a residual null check left around. (Causes a slowdown, 2956 // for example, in some objArray manipulations, such as a[i]=a[j].) 2957 if (tk->singleton()) { 2958 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2959 if (objtp != NULL && objtp->klass() != NULL) { 2960 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2961 case SSC_always_true: 2962 // If we know the type check always succeed then we don't use 2963 // the profiling data at this bytecode. Don't lose it, feed it 2964 // to the type system as a speculative type. 2965 return record_profiled_receiver_for_speculation(obj); 2966 case SSC_always_false: 2967 // It needs a null check because a null will *pass* the cast check. 2968 // A non-null value will always produce an exception. 2969 return null_assert(obj); 2970 } 2971 } 2972 } 2973 2974 ciProfileData* data = NULL; 2975 bool safe_for_replace = false; 2976 if (failure_control == NULL) { // use MDO in regular case only 2977 assert(java_bc() == Bytecodes::_aastore || 2978 java_bc() == Bytecodes::_checkcast, 2979 "interpreter profiles type checks only for these BCs"); 2980 data = method()->method_data()->bci_to_data(bci()); 2981 safe_for_replace = true; 2982 } 2983 2984 // Make the merge point 2985 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2986 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 2987 Node* phi = new (C) PhiNode(region, toop); 2988 C->set_has_split_ifs(true); // Has chance for split-if optimization 2989 2990 // Use null-cast information if it is available 2991 bool never_see_null = ((failure_control == NULL) // regular case only 2992 && seems_never_null(obj, data)); 2993 2994 // Null check; get casted pointer; set region slot 3 2995 Node* null_ctl = top(); 2996 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2997 2998 // If not_null_obj is dead, only null-path is taken 2999 if (stopped()) { // Doing instance-of on a NULL? 3000 set_control(null_ctl); 3001 return null(); 3002 } 3003 region->init_req(_null_path, null_ctl); 3004 phi ->init_req(_null_path, null()); // Set null path value 3005 if (null_ctl == top()) { 3006 // Do this eagerly, so that pattern matches like is_diamond_phi 3007 // will work even during parsing. 3008 assert(_null_path == PATH_LIMIT-1, "delete last"); 3009 region->del_req(_null_path); 3010 phi ->del_req(_null_path); 3011 } 3012 3013 Node* cast_obj = NULL; 3014 if (tk->klass_is_exact()) { 3015 // The following optimization tries to statically cast the speculative type of the object 3016 // (for example obtained during profiling) to the type of the superklass and then do a 3017 // dynamic check that the type of the object is what we expect. To work correctly 3018 // for checkcast and aastore the type of superklass should be exact. 3019 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3020 // We may not have profiling here or it may not help us. If we have 3021 // a speculative type use it to perform an exact cast. 3022 ciKlass* spec_obj_type = obj_type->speculative_type(); 3023 if (spec_obj_type != NULL || 3024 (data != NULL && 3025 // Counter has never been decremented (due to cast failure). 3026 // ...This is a reasonable thing to expect. It is true of 3027 // all casts inserted by javac to implement generic types. 3028 data->as_CounterData()->count() >= 0)) { 3029 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3030 if (cast_obj != NULL) { 3031 if (failure_control != NULL) // failure is now impossible 3032 (*failure_control) = top(); 3033 // adjust the type of the phi to the exact klass: 3034 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3035 } 3036 } 3037 } 3038 3039 if (cast_obj == NULL) { 3040 // Load the object's klass 3041 Node* obj_klass = load_object_klass(not_null_obj); 3042 3043 // Generate the subtype check 3044 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3045 3046 // Plug in success path into the merge 3047 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 3048 not_null_obj, toop)); 3049 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3050 if (failure_control == NULL) { 3051 if (not_subtype_ctrl != top()) { // If failure is possible 3052 PreserveJVMState pjvms(this); 3053 set_control(not_subtype_ctrl); 3054 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3055 } 3056 } else { 3057 (*failure_control) = not_subtype_ctrl; 3058 } 3059 } 3060 3061 region->init_req(_obj_path, control()); 3062 phi ->init_req(_obj_path, cast_obj); 3063 3064 // A merge of NULL or Casted-NotNull obj 3065 Node* res = _gvn.transform(phi); 3066 3067 // Note I do NOT always 'replace_in_map(obj,result)' here. 3068 // if( tk->klass()->can_be_primary_super() ) 3069 // This means that if I successfully store an Object into an array-of-String 3070 // I 'forget' that the Object is really now known to be a String. I have to 3071 // do this because we don't have true union types for interfaces - if I store 3072 // a Baz into an array-of-Interface and then tell the optimizer it's an 3073 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3074 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3075 // replace_in_map( obj, res ); 3076 3077 // Return final merged results 3078 set_control( _gvn.transform(region) ); 3079 record_for_igvn(region); 3080 return res; 3081 } 3082 3083 //------------------------------next_monitor----------------------------------- 3084 // What number should be given to the next monitor? 3085 int GraphKit::next_monitor() { 3086 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3087 int next = current + C->sync_stack_slots(); 3088 // Keep the toplevel high water mark current: 3089 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3090 return current; 3091 } 3092 3093 //------------------------------insert_mem_bar--------------------------------- 3094 // Memory barrier to avoid floating things around 3095 // The membar serves as a pinch point between both control and all memory slices. 3096 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3097 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3098 mb->init_req(TypeFunc::Control, control()); 3099 mb->init_req(TypeFunc::Memory, reset_memory()); 3100 Node* membar = _gvn.transform(mb); 3101 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3102 set_all_memory_call(membar); 3103 return membar; 3104 } 3105 3106 //-------------------------insert_mem_bar_volatile---------------------------- 3107 // Memory barrier to avoid floating things around 3108 // The membar serves as a pinch point between both control and memory(alias_idx). 3109 // If you want to make a pinch point on all memory slices, do not use this 3110 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3111 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3112 // When Parse::do_put_xxx updates a volatile field, it appends a series 3113 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3114 // The first membar is on the same memory slice as the field store opcode. 3115 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3116 // All the other membars (for other volatile slices, including AliasIdxBot, 3117 // which stands for all unknown volatile slices) are control-dependent 3118 // on the first membar. This prevents later volatile loads or stores 3119 // from sliding up past the just-emitted store. 3120 3121 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3122 mb->set_req(TypeFunc::Control,control()); 3123 if (alias_idx == Compile::AliasIdxBot) { 3124 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3125 } else { 3126 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3127 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3128 } 3129 Node* membar = _gvn.transform(mb); 3130 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3131 if (alias_idx == Compile::AliasIdxBot) { 3132 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 3133 } else { 3134 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 3135 } 3136 return membar; 3137 } 3138 3139 //------------------------------shared_lock------------------------------------ 3140 // Emit locking code. 3141 FastLockNode* GraphKit::shared_lock(Node* obj) { 3142 // bci is either a monitorenter bc or InvocationEntryBci 3143 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3144 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3145 3146 if( !GenerateSynchronizationCode ) 3147 return NULL; // Not locking things? 3148 if (stopped()) // Dead monitor? 3149 return NULL; 3150 3151 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3152 3153 // Box the stack location 3154 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 3155 Node* mem = reset_memory(); 3156 3157 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 3158 if (PrintPreciseBiasedLockingStatistics) { 3159 // Create the counters for this fast lock. 3160 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3161 } 3162 // Add monitor to debug info for the slow path. If we block inside the 3163 // slow path and de-opt, we need the monitor hanging around 3164 map()->push_monitor( flock ); 3165 3166 const TypeFunc *tf = LockNode::lock_type(); 3167 LockNode *lock = new (C) LockNode(C, tf); 3168 3169 lock->init_req( TypeFunc::Control, control() ); 3170 lock->init_req( TypeFunc::Memory , mem ); 3171 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3172 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3173 lock->init_req( TypeFunc::ReturnAdr, top() ); 3174 3175 lock->init_req(TypeFunc::Parms + 0, obj); 3176 lock->init_req(TypeFunc::Parms + 1, box); 3177 lock->init_req(TypeFunc::Parms + 2, flock); 3178 add_safepoint_edges(lock); 3179 3180 lock = _gvn.transform( lock )->as_Lock(); 3181 3182 // lock has no side-effects, sets few values 3183 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3184 3185 insert_mem_bar(Op_MemBarAcquireLock); 3186 3187 // Add this to the worklist so that the lock can be eliminated 3188 record_for_igvn(lock); 3189 3190 #ifndef PRODUCT 3191 if (PrintLockStatistics) { 3192 // Update the counter for this lock. Don't bother using an atomic 3193 // operation since we don't require absolute accuracy. 3194 lock->create_lock_counter(map()->jvms()); 3195 increment_counter(lock->counter()->addr()); 3196 } 3197 #endif 3198 3199 return flock; 3200 } 3201 3202 3203 //------------------------------shared_unlock---------------------------------- 3204 // Emit unlocking code. 3205 void GraphKit::shared_unlock(Node* box, Node* obj) { 3206 // bci is either a monitorenter bc or InvocationEntryBci 3207 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3208 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3209 3210 if( !GenerateSynchronizationCode ) 3211 return; 3212 if (stopped()) { // Dead monitor? 3213 map()->pop_monitor(); // Kill monitor from debug info 3214 return; 3215 } 3216 3217 // Memory barrier to avoid floating things down past the locked region 3218 insert_mem_bar(Op_MemBarReleaseLock); 3219 3220 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3221 UnlockNode *unlock = new (C) UnlockNode(C, tf); 3222 uint raw_idx = Compile::AliasIdxRaw; 3223 unlock->init_req( TypeFunc::Control, control() ); 3224 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3225 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3226 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3227 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3228 3229 unlock->init_req(TypeFunc::Parms + 0, obj); 3230 unlock->init_req(TypeFunc::Parms + 1, box); 3231 unlock = _gvn.transform(unlock)->as_Unlock(); 3232 3233 Node* mem = reset_memory(); 3234 3235 // unlock has no side-effects, sets few values 3236 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3237 3238 // Kill monitor from debug info 3239 map()->pop_monitor( ); 3240 } 3241 3242 //-------------------------------get_layout_helper----------------------------- 3243 // If the given klass is a constant or known to be an array, 3244 // fetch the constant layout helper value into constant_value 3245 // and return (Node*)NULL. Otherwise, load the non-constant 3246 // layout helper value, and return the node which represents it. 3247 // This two-faced routine is useful because allocation sites 3248 // almost always feature constant types. 3249 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3250 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3251 if (!StressReflectiveCode && inst_klass != NULL) { 3252 ciKlass* klass = inst_klass->klass(); 3253 bool xklass = inst_klass->klass_is_exact(); 3254 if (xklass || klass->is_array_klass()) { 3255 jint lhelper = klass->layout_helper(); 3256 if (lhelper != Klass::_lh_neutral_value) { 3257 constant_value = lhelper; 3258 return (Node*) NULL; 3259 } 3260 } 3261 } 3262 constant_value = Klass::_lh_neutral_value; // put in a known value 3263 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3264 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3265 } 3266 3267 // We just put in an allocate/initialize with a big raw-memory effect. 3268 // Hook selected additional alias categories on the initialization. 3269 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3270 MergeMemNode* init_in_merge, 3271 Node* init_out_raw) { 3272 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3273 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3274 3275 Node* prevmem = kit.memory(alias_idx); 3276 init_in_merge->set_memory_at(alias_idx, prevmem); 3277 kit.set_memory(init_out_raw, alias_idx); 3278 } 3279 3280 //---------------------------set_output_for_allocation------------------------- 3281 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3282 const TypeOopPtr* oop_type) { 3283 int rawidx = Compile::AliasIdxRaw; 3284 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3285 add_safepoint_edges(alloc); 3286 Node* allocx = _gvn.transform(alloc); 3287 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3288 // create memory projection for i_o 3289 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3290 make_slow_call_ex(allocx, env()->Throwable_klass(), true); 3291 3292 // create a memory projection as for the normal control path 3293 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3294 set_memory(malloc, rawidx); 3295 3296 // a normal slow-call doesn't change i_o, but an allocation does 3297 // we create a separate i_o projection for the normal control path 3298 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3299 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3300 3301 // put in an initialization barrier 3302 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3303 rawoop)->as_Initialize(); 3304 assert(alloc->initialization() == init, "2-way macro link must work"); 3305 assert(init ->allocation() == alloc, "2-way macro link must work"); 3306 { 3307 // Extract memory strands which may participate in the new object's 3308 // initialization, and source them from the new InitializeNode. 3309 // This will allow us to observe initializations when they occur, 3310 // and link them properly (as a group) to the InitializeNode. 3311 assert(init->in(InitializeNode::Memory) == malloc, ""); 3312 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3313 init->set_req(InitializeNode::Memory, minit_in); 3314 record_for_igvn(minit_in); // fold it up later, if possible 3315 Node* minit_out = memory(rawidx); 3316 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3317 if (oop_type->isa_aryptr()) { 3318 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3319 int elemidx = C->get_alias_index(telemref); 3320 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3321 } else if (oop_type->isa_instptr()) { 3322 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3323 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3324 ciField* field = ik->nonstatic_field_at(i); 3325 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3326 continue; // do not bother to track really large numbers of fields 3327 // Find (or create) the alias category for this field: 3328 int fieldidx = C->alias_type(field)->index(); 3329 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3330 } 3331 } 3332 } 3333 3334 // Cast raw oop to the real thing... 3335 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3336 javaoop = _gvn.transform(javaoop); 3337 C->set_recent_alloc(control(), javaoop); 3338 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3339 3340 #ifdef ASSERT 3341 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3342 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3343 "Ideal_allocation works"); 3344 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3345 "Ideal_allocation works"); 3346 if (alloc->is_AllocateArray()) { 3347 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3348 "Ideal_allocation works"); 3349 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3350 "Ideal_allocation works"); 3351 } else { 3352 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3353 } 3354 } 3355 #endif //ASSERT 3356 3357 return javaoop; 3358 } 3359 3360 //---------------------------new_instance-------------------------------------- 3361 // This routine takes a klass_node which may be constant (for a static type) 3362 // or may be non-constant (for reflective code). It will work equally well 3363 // for either, and the graph will fold nicely if the optimizer later reduces 3364 // the type to a constant. 3365 // The optional arguments are for specialized use by intrinsics: 3366 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3367 // - If 'return_size_val', report the the total object size to the caller. 3368 Node* GraphKit::new_instance(Node* klass_node, 3369 Node* extra_slow_test, 3370 Node* *return_size_val) { 3371 // Compute size in doublewords 3372 // The size is always an integral number of doublewords, represented 3373 // as a positive bytewise size stored in the klass's layout_helper. 3374 // The layout_helper also encodes (in a low bit) the need for a slow path. 3375 jint layout_con = Klass::_lh_neutral_value; 3376 Node* layout_val = get_layout_helper(klass_node, layout_con); 3377 int layout_is_con = (layout_val == NULL); 3378 3379 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3380 // Generate the initial go-slow test. It's either ALWAYS (return a 3381 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3382 // case) a computed value derived from the layout_helper. 3383 Node* initial_slow_test = NULL; 3384 if (layout_is_con) { 3385 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3386 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3387 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3388 3389 } else { // reflective case 3390 // This reflective path is used by Unsafe.allocateInstance. 3391 // (It may be stress-tested by specifying StressReflectiveCode.) 3392 // Basically, we want to get into the VM is there's an illegal argument. 3393 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3394 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3395 if (extra_slow_test != intcon(0)) { 3396 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3397 } 3398 // (Macro-expander will further convert this to a Bool, if necessary.) 3399 } 3400 3401 // Find the size in bytes. This is easy; it's the layout_helper. 3402 // The size value must be valid even if the slow path is taken. 3403 Node* size = NULL; 3404 if (layout_is_con) { 3405 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3406 } else { // reflective case 3407 // This reflective path is used by clone and Unsafe.allocateInstance. 3408 size = ConvI2X(layout_val); 3409 3410 // Clear the low bits to extract layout_helper_size_in_bytes: 3411 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3412 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3413 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3414 } 3415 if (return_size_val != NULL) { 3416 (*return_size_val) = size; 3417 } 3418 3419 // This is a precise notnull oop of the klass. 3420 // (Actually, it need not be precise if this is a reflective allocation.) 3421 // It's what we cast the result to. 3422 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3423 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3424 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3425 3426 // Now generate allocation code 3427 3428 // The entire memory state is needed for slow path of the allocation 3429 // since GC and deoptimization can happened. 3430 Node *mem = reset_memory(); 3431 set_all_memory(mem); // Create new memory state 3432 3433 AllocateNode* alloc 3434 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3435 control(), mem, i_o(), 3436 size, klass_node, 3437 initial_slow_test); 3438 3439 return set_output_for_allocation(alloc, oop_type); 3440 } 3441 3442 //-------------------------------new_array------------------------------------- 3443 // helper for both newarray and anewarray 3444 // The 'length' parameter is (obviously) the length of the array. 3445 // See comments on new_instance for the meaning of the other arguments. 3446 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3447 Node* length, // number of array elements 3448 int nargs, // number of arguments to push back for uncommon trap 3449 Node* *return_size_val) { 3450 jint layout_con = Klass::_lh_neutral_value; 3451 Node* layout_val = get_layout_helper(klass_node, layout_con); 3452 int layout_is_con = (layout_val == NULL); 3453 3454 if (!layout_is_con && !StressReflectiveCode && 3455 !too_many_traps(Deoptimization::Reason_class_check)) { 3456 // This is a reflective array creation site. 3457 // Optimistically assume that it is a subtype of Object[], 3458 // so that we can fold up all the address arithmetic. 3459 layout_con = Klass::array_layout_helper(T_OBJECT); 3460 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3461 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3462 { BuildCutout unless(this, bol_lh, PROB_MAX); 3463 inc_sp(nargs); 3464 uncommon_trap(Deoptimization::Reason_class_check, 3465 Deoptimization::Action_maybe_recompile); 3466 } 3467 layout_val = NULL; 3468 layout_is_con = true; 3469 } 3470 3471 // Generate the initial go-slow test. Make sure we do not overflow 3472 // if length is huge (near 2Gig) or negative! We do not need 3473 // exact double-words here, just a close approximation of needed 3474 // double-words. We can't add any offset or rounding bits, lest we 3475 // take a size -1 of bytes and make it positive. Use an unsigned 3476 // compare, so negative sizes look hugely positive. 3477 int fast_size_limit = FastAllocateSizeLimit; 3478 if (layout_is_con) { 3479 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3480 // Increase the size limit if we have exact knowledge of array type. 3481 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3482 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3483 } 3484 3485 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3486 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3487 if (initial_slow_test->is_Bool()) { 3488 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3489 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3490 } 3491 3492 // --- Size Computation --- 3493 // array_size = round_to_heap(array_header + (length << elem_shift)); 3494 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3495 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3496 // The rounding mask is strength-reduced, if possible. 3497 int round_mask = MinObjAlignmentInBytes - 1; 3498 Node* header_size = NULL; 3499 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3500 // (T_BYTE has the weakest alignment and size restrictions...) 3501 if (layout_is_con) { 3502 int hsize = Klass::layout_helper_header_size(layout_con); 3503 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3504 BasicType etype = Klass::layout_helper_element_type(layout_con); 3505 if ((round_mask & ~right_n_bits(eshift)) == 0) 3506 round_mask = 0; // strength-reduce it if it goes away completely 3507 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3508 assert(header_size_min <= hsize, "generic minimum is smallest"); 3509 header_size_min = hsize; 3510 header_size = intcon(hsize + round_mask); 3511 } else { 3512 Node* hss = intcon(Klass::_lh_header_size_shift); 3513 Node* hsm = intcon(Klass::_lh_header_size_mask); 3514 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3515 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3516 Node* mask = intcon(round_mask); 3517 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3518 } 3519 3520 Node* elem_shift = NULL; 3521 if (layout_is_con) { 3522 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3523 if (eshift != 0) 3524 elem_shift = intcon(eshift); 3525 } else { 3526 // There is no need to mask or shift this value. 3527 // The semantics of LShiftINode include an implicit mask to 0x1F. 3528 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3529 elem_shift = layout_val; 3530 } 3531 3532 // Transition to native address size for all offset calculations: 3533 Node* lengthx = ConvI2X(length); 3534 Node* headerx = ConvI2X(header_size); 3535 #ifdef _LP64 3536 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3537 if (tllen != NULL && tllen->_lo < 0) { 3538 // Add a manual constraint to a positive range. Cf. array_element_address. 3539 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3540 if (size_max > tllen->_hi) size_max = tllen->_hi; 3541 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3542 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon)); 3543 } 3544 } 3545 #endif 3546 3547 // Combine header size (plus rounding) and body size. Then round down. 3548 // This computation cannot overflow, because it is used only in two 3549 // places, one where the length is sharply limited, and the other 3550 // after a successful allocation. 3551 Node* abody = lengthx; 3552 if (elem_shift != NULL) 3553 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3554 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3555 if (round_mask != 0) { 3556 Node* mask = MakeConX(~round_mask); 3557 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3558 } 3559 // else if round_mask == 0, the size computation is self-rounding 3560 3561 if (return_size_val != NULL) { 3562 // This is the size 3563 (*return_size_val) = size; 3564 } 3565 3566 // Now generate allocation code 3567 3568 // The entire memory state is needed for slow path of the allocation 3569 // since GC and deoptimization can happened. 3570 Node *mem = reset_memory(); 3571 set_all_memory(mem); // Create new memory state 3572 3573 // Create the AllocateArrayNode and its result projections 3574 AllocateArrayNode* alloc 3575 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3576 control(), mem, i_o(), 3577 size, klass_node, 3578 initial_slow_test, 3579 length); 3580 3581 // Cast to correct type. Note that the klass_node may be constant or not, 3582 // and in the latter case the actual array type will be inexact also. 3583 // (This happens via a non-constant argument to inline_native_newArray.) 3584 // In any case, the value of klass_node provides the desired array type. 3585 const TypeInt* length_type = _gvn.find_int_type(length); 3586 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3587 if (ary_type->isa_aryptr() && length_type != NULL) { 3588 // Try to get a better type than POS for the size 3589 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3590 } 3591 3592 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3593 3594 // Cast length on remaining path to be as narrow as possible 3595 if (map()->find_edge(length) >= 0) { 3596 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3597 if (ccast != length) { 3598 _gvn.set_type_bottom(ccast); 3599 record_for_igvn(ccast); 3600 replace_in_map(length, ccast); 3601 } 3602 } 3603 3604 return javaoop; 3605 } 3606 3607 // The following "Ideal_foo" functions are placed here because they recognize 3608 // the graph shapes created by the functions immediately above. 3609 3610 //---------------------------Ideal_allocation---------------------------------- 3611 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3612 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3613 if (ptr == NULL) { // reduce dumb test in callers 3614 return NULL; 3615 } 3616 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3617 ptr = ptr->in(1); 3618 if (ptr == NULL) return NULL; 3619 } 3620 // Return NULL for allocations with several casts: 3621 // j.l.reflect.Array.newInstance(jobject, jint) 3622 // Object.clone() 3623 // to keep more precise type from last cast. 3624 if (ptr->is_Proj()) { 3625 Node* allo = ptr->in(0); 3626 if (allo != NULL && allo->is_Allocate()) { 3627 return allo->as_Allocate(); 3628 } 3629 } 3630 // Report failure to match. 3631 return NULL; 3632 } 3633 3634 // Fancy version which also strips off an offset (and reports it to caller). 3635 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3636 intptr_t& offset) { 3637 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3638 if (base == NULL) return NULL; 3639 return Ideal_allocation(base, phase); 3640 } 3641 3642 // Trace Initialize <- Proj[Parm] <- Allocate 3643 AllocateNode* InitializeNode::allocation() { 3644 Node* rawoop = in(InitializeNode::RawAddress); 3645 if (rawoop->is_Proj()) { 3646 Node* alloc = rawoop->in(0); 3647 if (alloc->is_Allocate()) { 3648 return alloc->as_Allocate(); 3649 } 3650 } 3651 return NULL; 3652 } 3653 3654 // Trace Allocate -> Proj[Parm] -> Initialize 3655 InitializeNode* AllocateNode::initialization() { 3656 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3657 if (rawoop == NULL) return NULL; 3658 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3659 Node* init = rawoop->fast_out(i); 3660 if (init->is_Initialize()) { 3661 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3662 return init->as_Initialize(); 3663 } 3664 } 3665 return NULL; 3666 } 3667 3668 //----------------------------- loop predicates --------------------------- 3669 3670 //------------------------------add_predicate_impl---------------------------- 3671 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3672 // Too many traps seen? 3673 if (too_many_traps(reason)) { 3674 #ifdef ASSERT 3675 if (TraceLoopPredicate) { 3676 int tc = C->trap_count(reason); 3677 tty->print("too many traps=%s tcount=%d in ", 3678 Deoptimization::trap_reason_name(reason), tc); 3679 method()->print(); // which method has too many predicate traps 3680 tty->cr(); 3681 } 3682 #endif 3683 // We cannot afford to take more traps here, 3684 // do not generate predicate. 3685 return; 3686 } 3687 3688 Node *cont = _gvn.intcon(1); 3689 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3690 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3691 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3692 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3693 C->add_predicate_opaq(opq); 3694 { 3695 PreserveJVMState pjvms(this); 3696 set_control(iffalse); 3697 inc_sp(nargs); 3698 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3699 } 3700 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3701 set_control(iftrue); 3702 } 3703 3704 //------------------------------add_predicate--------------------------------- 3705 void GraphKit::add_predicate(int nargs) { 3706 if (UseLoopPredicate) { 3707 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3708 } 3709 // loop's limit check predicate should be near the loop. 3710 if (LoopLimitCheck) { 3711 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3712 } 3713 } 3714 3715 //----------------------------- store barriers ---------------------------- 3716 #define __ ideal. 3717 3718 void GraphKit::sync_kit(IdealKit& ideal) { 3719 set_all_memory(__ merged_memory()); 3720 set_i_o(__ i_o()); 3721 set_control(__ ctrl()); 3722 } 3723 3724 void GraphKit::final_sync(IdealKit& ideal) { 3725 // Final sync IdealKit and graphKit. 3726 sync_kit(ideal); 3727 } 3728 3729 // vanilla/CMS post barrier 3730 // Insert a write-barrier store. This is to let generational GC work; we have 3731 // to flag all oop-stores before the next GC point. 3732 void GraphKit::write_barrier_post(Node* oop_store, 3733 Node* obj, 3734 Node* adr, 3735 uint adr_idx, 3736 Node* val, 3737 bool use_precise) { 3738 // No store check needed if we're storing a NULL or an old object 3739 // (latter case is probably a string constant). The concurrent 3740 // mark sweep garbage collector, however, needs to have all nonNull 3741 // oop updates flagged via card-marks. 3742 if (val != NULL && val->is_Con()) { 3743 // must be either an oop or NULL 3744 const Type* t = val->bottom_type(); 3745 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3746 // stores of null never (?) need barriers 3747 return; 3748 } 3749 3750 if (use_ReduceInitialCardMarks() 3751 && obj == just_allocated_object(control())) { 3752 // We can skip marks on a freshly-allocated object in Eden. 3753 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3754 // That routine informs GC to take appropriate compensating steps, 3755 // upon a slow-path allocation, so as to make this card-mark 3756 // elision safe. 3757 return; 3758 } 3759 3760 if (!use_precise) { 3761 // All card marks for a (non-array) instance are in one place: 3762 adr = obj; 3763 } 3764 // (Else it's an array (or unknown), and we want more precise card marks.) 3765 assert(adr != NULL, ""); 3766 3767 IdealKit ideal(this, true); 3768 3769 // Convert the pointer to an int prior to doing math on it 3770 Node* cast = __ CastPX(__ ctrl(), adr); 3771 3772 // Divide by card size 3773 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3774 "Only one we handle so far."); 3775 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3776 3777 // Combine card table base and card offset 3778 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3779 3780 // Get the alias_index for raw card-mark memory 3781 int adr_type = Compile::AliasIdxRaw; 3782 Node* zero = __ ConI(0); // Dirty card value 3783 BasicType bt = T_BYTE; 3784 3785 if (UseCondCardMark) { 3786 // The classic GC reference write barrier is typically implemented 3787 // as a store into the global card mark table. Unfortunately 3788 // unconditional stores can result in false sharing and excessive 3789 // coherence traffic as well as false transactional aborts. 3790 // UseCondCardMark enables MP "polite" conditional card mark 3791 // stores. In theory we could relax the load from ctrl() to 3792 // no_ctrl, but that doesn't buy much latitude. 3793 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3794 __ if_then(card_val, BoolTest::ne, zero); 3795 } 3796 3797 // Smash zero into card 3798 if( !UseConcMarkSweepGC ) { 3799 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); 3800 } else { 3801 // Specialized path for CM store barrier 3802 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3803 } 3804 3805 if (UseCondCardMark) { 3806 __ end_if(); 3807 } 3808 3809 // Final sync IdealKit and GraphKit. 3810 final_sync(ideal); 3811 } 3812 3813 // G1 pre/post barriers 3814 void GraphKit::g1_write_barrier_pre(bool do_load, 3815 Node* obj, 3816 Node* adr, 3817 uint alias_idx, 3818 Node* val, 3819 const TypeOopPtr* val_type, 3820 Node* pre_val, 3821 BasicType bt) { 3822 3823 // Some sanity checks 3824 // Note: val is unused in this routine. 3825 3826 if (do_load) { 3827 // We need to generate the load of the previous value 3828 assert(obj != NULL, "must have a base"); 3829 assert(adr != NULL, "where are loading from?"); 3830 assert(pre_val == NULL, "loaded already?"); 3831 assert(val_type != NULL, "need a type"); 3832 } else { 3833 // In this case both val_type and alias_idx are unused. 3834 assert(pre_val != NULL, "must be loaded already"); 3835 // Nothing to be done if pre_val is null. 3836 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3837 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3838 } 3839 assert(bt == T_OBJECT, "or we shouldn't be here"); 3840 3841 IdealKit ideal(this, true); 3842 3843 Node* tls = __ thread(); // ThreadLocalStorage 3844 3845 Node* no_ctrl = NULL; 3846 Node* no_base = __ top(); 3847 Node* zero = __ ConI(0); 3848 Node* zeroX = __ ConX(0); 3849 3850 float likely = PROB_LIKELY(0.999); 3851 float unlikely = PROB_UNLIKELY(0.999); 3852 3853 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3854 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3855 3856 // Offsets into the thread 3857 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3858 PtrQueue::byte_offset_of_active()); 3859 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3860 PtrQueue::byte_offset_of_index()); 3861 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3862 PtrQueue::byte_offset_of_buf()); 3863 3864 // Now the actual pointers into the thread 3865 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3866 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3867 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3868 3869 // Now some of the values 3870 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3871 3872 // if (!marking) 3873 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3874 BasicType index_bt = TypeX_X->basic_type(); 3875 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3876 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3877 3878 if (do_load) { 3879 // load original value 3880 // alias_idx correct?? 3881 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3882 } 3883 3884 // if (pre_val != NULL) 3885 __ if_then(pre_val, BoolTest::ne, null()); { 3886 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3887 3888 // is the queue for this thread full? 3889 __ if_then(index, BoolTest::ne, zeroX, likely); { 3890 3891 // decrement the index 3892 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3893 3894 // Now get the buffer location we will log the previous value into and store it 3895 Node *log_addr = __ AddP(no_base, buffer, next_index); 3896 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 3897 // update the index 3898 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 3899 3900 } __ else_(); { 3901 3902 // logging buffer is full, call the runtime 3903 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3904 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3905 } __ end_if(); // (!index) 3906 } __ end_if(); // (pre_val != NULL) 3907 } __ end_if(); // (!marking) 3908 3909 // Final sync IdealKit and GraphKit. 3910 final_sync(ideal); 3911 } 3912 3913 // 3914 // Update the card table and add card address to the queue 3915 // 3916 void GraphKit::g1_mark_card(IdealKit& ideal, 3917 Node* card_adr, 3918 Node* oop_store, 3919 uint oop_alias_idx, 3920 Node* index, 3921 Node* index_adr, 3922 Node* buffer, 3923 const TypeFunc* tf) { 3924 3925 Node* zero = __ ConI(0); 3926 Node* zeroX = __ ConX(0); 3927 Node* no_base = __ top(); 3928 BasicType card_bt = T_BYTE; 3929 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3930 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3931 3932 // Now do the queue work 3933 __ if_then(index, BoolTest::ne, zeroX); { 3934 3935 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3936 Node* log_addr = __ AddP(no_base, buffer, next_index); 3937 3938 // Order, see storeCM. 3939 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 3940 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 3941 3942 } __ else_(); { 3943 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3944 } __ end_if(); 3945 3946 } 3947 3948 void GraphKit::g1_write_barrier_post(Node* oop_store, 3949 Node* obj, 3950 Node* adr, 3951 uint alias_idx, 3952 Node* val, 3953 BasicType bt, 3954 bool use_precise) { 3955 // If we are writing a NULL then we need no post barrier 3956 3957 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3958 // Must be NULL 3959 const Type* t = val->bottom_type(); 3960 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3961 // No post barrier if writing NULLx 3962 return; 3963 } 3964 3965 if (!use_precise) { 3966 // All card marks for a (non-array) instance are in one place: 3967 adr = obj; 3968 } 3969 // (Else it's an array (or unknown), and we want more precise card marks.) 3970 assert(adr != NULL, ""); 3971 3972 IdealKit ideal(this, true); 3973 3974 Node* tls = __ thread(); // ThreadLocalStorage 3975 3976 Node* no_base = __ top(); 3977 float likely = PROB_LIKELY(0.999); 3978 float unlikely = PROB_UNLIKELY(0.999); 3979 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 3980 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 3981 Node* zeroX = __ ConX(0); 3982 3983 // Get the alias_index for raw card-mark memory 3984 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3985 3986 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3987 3988 // Offsets into the thread 3989 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3990 PtrQueue::byte_offset_of_index()); 3991 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3992 PtrQueue::byte_offset_of_buf()); 3993 3994 // Pointers into the thread 3995 3996 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3997 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3998 3999 // Now some values 4000 // Use ctrl to avoid hoisting these values past a safepoint, which could 4001 // potentially reset these fields in the JavaThread. 4002 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4003 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4004 4005 // Convert the store obj pointer to an int prior to doing math on it 4006 // Must use ctrl to prevent "integerized oop" existing across safepoint 4007 Node* cast = __ CastPX(__ ctrl(), adr); 4008 4009 // Divide pointer by card size 4010 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4011 4012 // Combine card table base and card offset 4013 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4014 4015 // If we know the value being stored does it cross regions? 4016 4017 if (val != NULL) { 4018 // Does the store cause us to cross regions? 4019 4020 // Should be able to do an unsigned compare of region_size instead of 4021 // and extra shift. Do we have an unsigned compare?? 4022 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4023 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4024 4025 // if (xor_res == 0) same region so skip 4026 __ if_then(xor_res, BoolTest::ne, zeroX); { 4027 4028 // No barrier if we are storing a NULL 4029 __ if_then(val, BoolTest::ne, null(), unlikely); { 4030 4031 // Ok must mark the card if not already dirty 4032 4033 // load the original value of the card 4034 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4035 4036 __ if_then(card_val, BoolTest::ne, young_card); { 4037 sync_kit(ideal); 4038 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4039 insert_mem_bar(Op_MemBarVolatile, oop_store); 4040 __ sync_kit(this); 4041 4042 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4043 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4044 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4045 } __ end_if(); 4046 } __ end_if(); 4047 } __ end_if(); 4048 } __ end_if(); 4049 } else { 4050 // Object.clone() instrinsic uses this path. 4051 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4052 } 4053 4054 // Final sync IdealKit and GraphKit. 4055 final_sync(ideal); 4056 } 4057 #undef __ 4058 4059 4060 4061 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4062 if (java_lang_String::has_offset_field()) { 4063 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4064 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4065 false, NULL, 0); 4066 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4067 int offset_field_idx = C->get_alias_index(offset_field_type); 4068 return make_load(ctrl, 4069 basic_plus_adr(str, str, offset_offset), 4070 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4071 } else { 4072 return intcon(0); 4073 } 4074 } 4075 4076 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4077 if (java_lang_String::has_count_field()) { 4078 int count_offset = java_lang_String::count_offset_in_bytes(); 4079 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4080 false, NULL, 0); 4081 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4082 int count_field_idx = C->get_alias_index(count_field_type); 4083 return make_load(ctrl, 4084 basic_plus_adr(str, str, count_offset), 4085 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4086 } else { 4087 return load_array_length(load_String_value(ctrl, str)); 4088 } 4089 } 4090 4091 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4092 int value_offset = java_lang_String::value_offset_in_bytes(); 4093 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4094 false, NULL, 0); 4095 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4096 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4097 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4098 ciTypeArrayKlass::make(T_CHAR), true, 0); 4099 int value_field_idx = C->get_alias_index(value_field_type); 4100 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4101 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4102 // String.value field is known to be @Stable. 4103 if (UseImplicitStableValues) { 4104 load = cast_array_to_stable(load, value_type); 4105 } 4106 return load; 4107 } 4108 4109 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4110 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4111 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4112 false, NULL, 0); 4113 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4114 int offset_field_idx = C->get_alias_index(offset_field_type); 4115 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4116 value, T_INT, offset_field_idx, MemNode::unordered); 4117 } 4118 4119 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4120 int value_offset = java_lang_String::value_offset_in_bytes(); 4121 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4122 false, NULL, 0); 4123 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4124 4125 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4126 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4127 } 4128 4129 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4130 int count_offset = java_lang_String::count_offset_in_bytes(); 4131 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4132 false, NULL, 0); 4133 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4134 int count_field_idx = C->get_alias_index(count_field_type); 4135 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4136 value, T_INT, count_field_idx, MemNode::unordered); 4137 } 4138 4139 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4140 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4141 // assumption of CCP analysis. 4142 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); 4143 }