1 /* 2 * Copyright 2001-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_graphKit.cpp.incl" 27 28 //----------------------------GraphKit----------------------------------------- 29 // Main utility constructor. 30 GraphKit::GraphKit(JVMState* jvms) 31 : Phase(Phase::Parser), 32 _env(C->env()), 33 _gvn(*C->initial_gvn()) 34 { 35 _exceptions = jvms->map()->next_exception(); 36 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 37 set_jvms(jvms); 38 } 39 40 // Private constructor for parser. 41 GraphKit::GraphKit() 42 : Phase(Phase::Parser), 43 _env(C->env()), 44 _gvn(*C->initial_gvn()) 45 { 46 _exceptions = NULL; 47 set_map(NULL); 48 debug_only(_sp = -99); 49 debug_only(set_bci(-99)); 50 } 51 52 53 54 //---------------------------clean_stack--------------------------------------- 55 // Clear away rubbish from the stack area of the JVM state. 56 // This destroys any arguments that may be waiting on the stack. 57 void GraphKit::clean_stack(int from_sp) { 58 SafePointNode* map = this->map(); 59 JVMState* jvms = this->jvms(); 60 int stk_size = jvms->stk_size(); 61 int stkoff = jvms->stkoff(); 62 Node* top = this->top(); 63 for (int i = from_sp; i < stk_size; i++) { 64 if (map->in(stkoff + i) != top) { 65 map->set_req(stkoff + i, top); 66 } 67 } 68 } 69 70 71 //--------------------------------sync_jvms----------------------------------- 72 // Make sure our current jvms agrees with our parse state. 73 JVMState* GraphKit::sync_jvms() const { 74 JVMState* jvms = this->jvms(); 75 jvms->set_bci(bci()); // Record the new bci in the JVMState 76 jvms->set_sp(sp()); // Record the new sp in the JVMState 77 assert(jvms_in_sync(), "jvms is now in sync"); 78 return jvms; 79 } 80 81 #ifdef ASSERT 82 bool GraphKit::jvms_in_sync() const { 83 Parse* parse = is_Parse(); 84 if (parse == NULL) { 85 if (bci() != jvms()->bci()) return false; 86 if (sp() != (int)jvms()->sp()) return false; 87 return true; 88 } 89 if (jvms()->method() != parse->method()) return false; 90 if (jvms()->bci() != parse->bci()) return false; 91 int jvms_sp = jvms()->sp(); 92 if (jvms_sp != parse->sp()) return false; 93 int jvms_depth = jvms()->depth(); 94 if (jvms_depth != parse->depth()) return false; 95 return true; 96 } 97 98 // Local helper checks for special internal merge points 99 // used to accumulate and merge exception states. 100 // They are marked by the region's in(0) edge being the map itself. 101 // Such merge points must never "escape" into the parser at large, 102 // until they have been handed to gvn.transform. 103 static bool is_hidden_merge(Node* reg) { 104 if (reg == NULL) return false; 105 if (reg->is_Phi()) { 106 reg = reg->in(0); 107 if (reg == NULL) return false; 108 } 109 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 110 } 111 112 void GraphKit::verify_map() const { 113 if (map() == NULL) return; // null map is OK 114 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 115 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 116 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 117 } 118 119 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 120 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 121 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 122 } 123 #endif 124 125 //---------------------------stop_and_kill_map--------------------------------- 126 // Set _map to NULL, signalling a stop to further bytecode execution. 127 // First smash the current map's control to a constant, to mark it dead. 128 void GraphKit::stop_and_kill_map() { 129 SafePointNode* dead_map = stop(); 130 if (dead_map != NULL) { 131 dead_map->disconnect_inputs(NULL); // Mark the map as killed. 132 assert(dead_map->is_killed(), "must be so marked"); 133 } 134 } 135 136 137 //--------------------------------stopped-------------------------------------- 138 // Tell if _map is NULL, or control is top. 139 bool GraphKit::stopped() { 140 if (map() == NULL) return true; 141 else if (control() == top()) return true; 142 else return false; 143 } 144 145 146 //-----------------------------has_ex_handler---------------------------------- 147 // Tell if this method or any caller method has exception handlers. 148 bool GraphKit::has_ex_handler() { 149 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 150 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 151 return true; 152 } 153 } 154 return false; 155 } 156 157 //------------------------------save_ex_oop------------------------------------ 158 // Save an exception without blowing stack contents or other JVM state. 159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 160 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 161 ex_map->add_req(ex_oop); 162 debug_only(verify_exception_state(ex_map)); 163 } 164 165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 166 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 167 Node* ex_oop = ex_map->in(ex_map->req()-1); 168 if (clear_it) ex_map->del_req(ex_map->req()-1); 169 return ex_oop; 170 } 171 172 //-----------------------------saved_ex_oop------------------------------------ 173 // Recover a saved exception from its map. 174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 175 return common_saved_ex_oop(ex_map, false); 176 } 177 178 //--------------------------clear_saved_ex_oop--------------------------------- 179 // Erase a previously saved exception from its map. 180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 181 return common_saved_ex_oop(ex_map, true); 182 } 183 184 #ifdef ASSERT 185 //---------------------------has_saved_ex_oop---------------------------------- 186 // Erase a previously saved exception from its map. 187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 188 return ex_map->req() == ex_map->jvms()->endoff()+1; 189 } 190 #endif 191 192 //-------------------------make_exception_state-------------------------------- 193 // Turn the current JVM state into an exception state, appending the ex_oop. 194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 195 sync_jvms(); 196 SafePointNode* ex_map = stop(); // do not manipulate this map any more 197 set_saved_ex_oop(ex_map, ex_oop); 198 return ex_map; 199 } 200 201 202 //--------------------------add_exception_state-------------------------------- 203 // Add an exception to my list of exceptions. 204 void GraphKit::add_exception_state(SafePointNode* ex_map) { 205 if (ex_map == NULL || ex_map->control() == top()) { 206 return; 207 } 208 #ifdef ASSERT 209 verify_exception_state(ex_map); 210 if (has_exceptions()) { 211 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 212 } 213 #endif 214 215 // If there is already an exception of exactly this type, merge with it. 216 // In particular, null-checks and other low-level exceptions common up here. 217 Node* ex_oop = saved_ex_oop(ex_map); 218 const Type* ex_type = _gvn.type(ex_oop); 219 if (ex_oop == top()) { 220 // No action needed. 221 return; 222 } 223 assert(ex_type->isa_instptr(), "exception must be an instance"); 224 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 225 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 226 // We check sp also because call bytecodes can generate exceptions 227 // both before and after arguments are popped! 228 if (ex_type2 == ex_type 229 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 230 combine_exception_states(ex_map, e2); 231 return; 232 } 233 } 234 235 // No pre-existing exception of the same type. Chain it on the list. 236 push_exception_state(ex_map); 237 } 238 239 //-----------------------add_exception_states_from----------------------------- 240 void GraphKit::add_exception_states_from(JVMState* jvms) { 241 SafePointNode* ex_map = jvms->map()->next_exception(); 242 if (ex_map != NULL) { 243 jvms->map()->set_next_exception(NULL); 244 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 245 next_map = ex_map->next_exception(); 246 ex_map->set_next_exception(NULL); 247 add_exception_state(ex_map); 248 } 249 } 250 } 251 252 //-----------------------transfer_exceptions_into_jvms------------------------- 253 JVMState* GraphKit::transfer_exceptions_into_jvms() { 254 if (map() == NULL) { 255 // We need a JVMS to carry the exceptions, but the map has gone away. 256 // Create a scratch JVMS, cloned from any of the exception states... 257 if (has_exceptions()) { 258 _map = _exceptions; 259 _map = clone_map(); 260 _map->set_next_exception(NULL); 261 clear_saved_ex_oop(_map); 262 debug_only(verify_map()); 263 } else { 264 // ...or created from scratch 265 JVMState* jvms = new (C) JVMState(_method, NULL); 266 jvms->set_bci(_bci); 267 jvms->set_sp(_sp); 268 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms)); 269 set_jvms(jvms); 270 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 271 set_all_memory(top()); 272 while (map()->req() < jvms->endoff()) map()->add_req(top()); 273 } 274 // (This is a kludge, in case you didn't notice.) 275 set_control(top()); 276 } 277 JVMState* jvms = sync_jvms(); 278 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 279 jvms->map()->set_next_exception(_exceptions); 280 _exceptions = NULL; // done with this set of exceptions 281 return jvms; 282 } 283 284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 285 assert(is_hidden_merge(dstphi), "must be a special merge node"); 286 assert(is_hidden_merge(srcphi), "must be a special merge node"); 287 uint limit = srcphi->req(); 288 for (uint i = PhiNode::Input; i < limit; i++) { 289 dstphi->add_req(srcphi->in(i)); 290 } 291 } 292 static inline void add_one_req(Node* dstphi, Node* src) { 293 assert(is_hidden_merge(dstphi), "must be a special merge node"); 294 assert(!is_hidden_merge(src), "must not be a special merge node"); 295 dstphi->add_req(src); 296 } 297 298 //-----------------------combine_exception_states------------------------------ 299 // This helper function combines exception states by building phis on a 300 // specially marked state-merging region. These regions and phis are 301 // untransformed, and can build up gradually. The region is marked by 302 // having a control input of its exception map, rather than NULL. Such 303 // regions do not appear except in this function, and in use_exception_state. 304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 305 if (failing()) return; // dying anyway... 306 JVMState* ex_jvms = ex_map->_jvms; 307 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 308 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 309 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 310 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 311 assert(ex_map->req() == phi_map->req(), "matching maps"); 312 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 313 Node* hidden_merge_mark = root(); 314 Node* region = phi_map->control(); 315 MergeMemNode* phi_mem = phi_map->merged_memory(); 316 MergeMemNode* ex_mem = ex_map->merged_memory(); 317 if (region->in(0) != hidden_merge_mark) { 318 // The control input is not (yet) a specially-marked region in phi_map. 319 // Make it so, and build some phis. 320 region = new (C, 2) RegionNode(2); 321 _gvn.set_type(region, Type::CONTROL); 322 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 323 region->init_req(1, phi_map->control()); 324 phi_map->set_control(region); 325 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 326 record_for_igvn(io_phi); 327 _gvn.set_type(io_phi, Type::ABIO); 328 phi_map->set_i_o(io_phi); 329 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 330 Node* m = mms.memory(); 331 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 332 record_for_igvn(m_phi); 333 _gvn.set_type(m_phi, Type::MEMORY); 334 mms.set_memory(m_phi); 335 } 336 } 337 338 // Either or both of phi_map and ex_map might already be converted into phis. 339 Node* ex_control = ex_map->control(); 340 // if there is special marking on ex_map also, we add multiple edges from src 341 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 342 // how wide was the destination phi_map, originally? 343 uint orig_width = region->req(); 344 345 if (add_multiple) { 346 add_n_reqs(region, ex_control); 347 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 348 } else { 349 // ex_map has no merges, so we just add single edges everywhere 350 add_one_req(region, ex_control); 351 add_one_req(phi_map->i_o(), ex_map->i_o()); 352 } 353 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 354 if (mms.is_empty()) { 355 // get a copy of the base memory, and patch some inputs into it 356 const TypePtr* adr_type = mms.adr_type(C); 357 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 358 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 359 mms.set_memory(phi); 360 // Prepare to append interesting stuff onto the newly sliced phi: 361 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 362 } 363 // Append stuff from ex_map: 364 if (add_multiple) { 365 add_n_reqs(mms.memory(), mms.memory2()); 366 } else { 367 add_one_req(mms.memory(), mms.memory2()); 368 } 369 } 370 uint limit = ex_map->req(); 371 for (uint i = TypeFunc::Parms; i < limit; i++) { 372 // Skip everything in the JVMS after tos. (The ex_oop follows.) 373 if (i == tos) i = ex_jvms->monoff(); 374 Node* src = ex_map->in(i); 375 Node* dst = phi_map->in(i); 376 if (src != dst) { 377 PhiNode* phi; 378 if (dst->in(0) != region) { 379 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 380 record_for_igvn(phi); 381 _gvn.set_type(phi, phi->type()); 382 phi_map->set_req(i, dst); 383 // Prepare to append interesting stuff onto the new phi: 384 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 385 } else { 386 assert(dst->is_Phi(), "nobody else uses a hidden region"); 387 phi = (PhiNode*)dst; 388 } 389 if (add_multiple && src->in(0) == ex_control) { 390 // Both are phis. 391 add_n_reqs(dst, src); 392 } else { 393 while (dst->req() < region->req()) add_one_req(dst, src); 394 } 395 const Type* srctype = _gvn.type(src); 396 if (phi->type() != srctype) { 397 const Type* dsttype = phi->type()->meet(srctype); 398 if (phi->type() != dsttype) { 399 phi->set_type(dsttype); 400 _gvn.set_type(phi, dsttype); 401 } 402 } 403 } 404 } 405 } 406 407 //--------------------------use_exception_state-------------------------------- 408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 409 if (failing()) { stop(); return top(); } 410 Node* region = phi_map->control(); 411 Node* hidden_merge_mark = root(); 412 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 413 Node* ex_oop = clear_saved_ex_oop(phi_map); 414 if (region->in(0) == hidden_merge_mark) { 415 // Special marking for internal ex-states. Process the phis now. 416 region->set_req(0, region); // now it's an ordinary region 417 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 418 // Note: Setting the jvms also sets the bci and sp. 419 set_control(_gvn.transform(region)); 420 uint tos = jvms()->stkoff() + sp(); 421 for (uint i = 1; i < tos; i++) { 422 Node* x = phi_map->in(i); 423 if (x->in(0) == region) { 424 assert(x->is_Phi(), "expected a special phi"); 425 phi_map->set_req(i, _gvn.transform(x)); 426 } 427 } 428 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 429 Node* x = mms.memory(); 430 if (x->in(0) == region) { 431 assert(x->is_Phi(), "nobody else uses a hidden region"); 432 mms.set_memory(_gvn.transform(x)); 433 } 434 } 435 if (ex_oop->in(0) == region) { 436 assert(ex_oop->is_Phi(), "expected a special phi"); 437 ex_oop = _gvn.transform(ex_oop); 438 } 439 } else { 440 set_jvms(phi_map->jvms()); 441 } 442 443 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 444 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 445 return ex_oop; 446 } 447 448 //---------------------------------java_bc------------------------------------- 449 Bytecodes::Code GraphKit::java_bc() const { 450 ciMethod* method = this->method(); 451 int bci = this->bci(); 452 if (method != NULL && bci != InvocationEntryBci) 453 return method->java_code_at_bci(bci); 454 else 455 return Bytecodes::_illegal; 456 } 457 458 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 459 bool must_throw) { 460 // if the exception capability is set, then we will generate code 461 // to check the JavaThread.should_post_on_exceptions flag to see 462 // if we actually need to report exception events (for this 463 // thread). If we don't need to report exception events, we will 464 // take the normal fast path provided by add_exception_events. If 465 // exception event reporting is enabled for this thread, we will 466 // take the uncommon_trap in the BuildCutout below. 467 468 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 469 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode()); 470 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 471 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false); 472 473 // Test the should_post_on_exceptions_flag vs. 0 474 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) ); 475 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) ); 476 477 // Branch to slow_path if should_post_on_exceptions_flag was true 478 { BuildCutout unless(this, tst, PROB_MAX); 479 // Do not try anything fancy if we're notifying the VM on every throw. 480 // Cf. case Bytecodes::_athrow in parse2.cpp. 481 uncommon_trap(reason, Deoptimization::Action_none, 482 (ciKlass*)NULL, (char*)NULL, must_throw); 483 } 484 485 } 486 487 //------------------------------builtin_throw---------------------------------- 488 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 489 bool must_throw = true; 490 491 if (env()->jvmti_can_post_on_exceptions()) { 492 // check if we must post exception events, take uncommon trap if so 493 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 494 // here if should_post_on_exceptions is false 495 // continue on with the normal codegen 496 } 497 498 // If this particular condition has not yet happened at this 499 // bytecode, then use the uncommon trap mechanism, and allow for 500 // a future recompilation if several traps occur here. 501 // If the throw is hot, try to use a more complicated inline mechanism 502 // which keeps execution inside the compiled code. 503 bool treat_throw_as_hot = false; 504 ciMethodData* md = method()->method_data(); 505 506 if (ProfileTraps) { 507 if (too_many_traps(reason)) { 508 treat_throw_as_hot = true; 509 } 510 // (If there is no MDO at all, assume it is early in 511 // execution, and that any deopts are part of the 512 // startup transient, and don't need to be remembered.) 513 514 // Also, if there is a local exception handler, treat all throws 515 // as hot if there has been at least one in this method. 516 if (C->trap_count(reason) != 0 517 && method()->method_data()->trap_count(reason) != 0 518 && has_ex_handler()) { 519 treat_throw_as_hot = true; 520 } 521 } 522 523 // If this throw happens frequently, an uncommon trap might cause 524 // a performance pothole. If there is a local exception handler, 525 // and if this particular bytecode appears to be deoptimizing often, 526 // let us handle the throw inline, with a preconstructed instance. 527 // Note: If the deopt count has blown up, the uncommon trap 528 // runtime is going to flush this nmethod, not matter what. 529 if (treat_throw_as_hot 530 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 531 // If the throw is local, we use a pre-existing instance and 532 // punt on the backtrace. This would lead to a missing backtrace 533 // (a repeat of 4292742) if the backtrace object is ever asked 534 // for its backtrace. 535 // Fixing this remaining case of 4292742 requires some flavor of 536 // escape analysis. Leave that for the future. 537 ciInstance* ex_obj = NULL; 538 switch (reason) { 539 case Deoptimization::Reason_null_check: 540 ex_obj = env()->NullPointerException_instance(); 541 break; 542 case Deoptimization::Reason_div0_check: 543 ex_obj = env()->ArithmeticException_instance(); 544 break; 545 case Deoptimization::Reason_range_check: 546 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 547 break; 548 case Deoptimization::Reason_class_check: 549 if (java_bc() == Bytecodes::_aastore) { 550 ex_obj = env()->ArrayStoreException_instance(); 551 } else { 552 ex_obj = env()->ClassCastException_instance(); 553 } 554 break; 555 } 556 if (failing()) { stop(); return; } // exception allocation might fail 557 if (ex_obj != NULL) { 558 // Cheat with a preallocated exception object. 559 if (C->log() != NULL) 560 C->log()->elem("hot_throw preallocated='1' reason='%s'", 561 Deoptimization::trap_reason_name(reason)); 562 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 563 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 564 565 // Clear the detail message of the preallocated exception object. 566 // Weblogic sometimes mutates the detail message of exceptions 567 // using reflection. 568 int offset = java_lang_Throwable::get_detailMessage_offset(); 569 const TypePtr* adr_typ = ex_con->add_offset(offset); 570 571 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 572 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT); 573 574 add_exception_state(make_exception_state(ex_node)); 575 return; 576 } 577 } 578 579 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 580 // It won't be much cheaper than bailing to the interp., since we'll 581 // have to pass up all the debug-info, and the runtime will have to 582 // create the stack trace. 583 584 // Usual case: Bail to interpreter. 585 // Reserve the right to recompile if we haven't seen anything yet. 586 587 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 588 if (treat_throw_as_hot 589 && (method()->method_data()->trap_recompiled_at(bci()) 590 || C->too_many_traps(reason))) { 591 // We cannot afford to take more traps here. Suffer in the interpreter. 592 if (C->log() != NULL) 593 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 594 Deoptimization::trap_reason_name(reason), 595 C->trap_count(reason)); 596 action = Deoptimization::Action_none; 597 } 598 599 // "must_throw" prunes the JVM state to include only the stack, if there 600 // are no local exception handlers. This should cut down on register 601 // allocation time and code size, by drastically reducing the number 602 // of in-edges on the call to the uncommon trap. 603 604 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 605 } 606 607 608 //----------------------------PreserveJVMState--------------------------------- 609 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 610 debug_only(kit->verify_map()); 611 _kit = kit; 612 _map = kit->map(); // preserve the map 613 _sp = kit->sp(); 614 kit->set_map(clone_map ? kit->clone_map() : NULL); 615 #ifdef ASSERT 616 _bci = kit->bci(); 617 Parse* parser = kit->is_Parse(); 618 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 619 _block = block; 620 #endif 621 } 622 PreserveJVMState::~PreserveJVMState() { 623 GraphKit* kit = _kit; 624 #ifdef ASSERT 625 assert(kit->bci() == _bci, "bci must not shift"); 626 Parse* parser = kit->is_Parse(); 627 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 628 assert(block == _block, "block must not shift"); 629 #endif 630 kit->set_map(_map); 631 kit->set_sp(_sp); 632 } 633 634 635 //-----------------------------BuildCutout------------------------------------- 636 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 637 : PreserveJVMState(kit) 638 { 639 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 640 SafePointNode* outer_map = _map; // preserved map is caller's 641 SafePointNode* inner_map = kit->map(); 642 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 643 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) )); 644 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) )); 645 } 646 BuildCutout::~BuildCutout() { 647 GraphKit* kit = _kit; 648 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 649 } 650 651 //---------------------------PreserveReexecuteState---------------------------- 652 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 653 assert(!kit->stopped(), "must call stopped() before"); 654 _kit = kit; 655 _sp = kit->sp(); 656 _reexecute = kit->jvms()->_reexecute; 657 } 658 PreserveReexecuteState::~PreserveReexecuteState() { 659 if (_kit->stopped()) return; 660 _kit->jvms()->_reexecute = _reexecute; 661 _kit->set_sp(_sp); 662 } 663 664 //------------------------------clone_map-------------------------------------- 665 // Implementation of PreserveJVMState 666 // 667 // Only clone_map(...) here. If this function is only used in the 668 // PreserveJVMState class we may want to get rid of this extra 669 // function eventually and do it all there. 670 671 SafePointNode* GraphKit::clone_map() { 672 if (map() == NULL) return NULL; 673 674 // Clone the memory edge first 675 Node* mem = MergeMemNode::make(C, map()->memory()); 676 gvn().set_type_bottom(mem); 677 678 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 679 JVMState* jvms = this->jvms(); 680 JVMState* clonejvms = jvms->clone_shallow(C); 681 clonemap->set_memory(mem); 682 clonemap->set_jvms(clonejvms); 683 clonejvms->set_map(clonemap); 684 record_for_igvn(clonemap); 685 gvn().set_type_bottom(clonemap); 686 return clonemap; 687 } 688 689 690 //-----------------------------set_map_clone----------------------------------- 691 void GraphKit::set_map_clone(SafePointNode* m) { 692 _map = m; 693 _map = clone_map(); 694 _map->set_next_exception(NULL); 695 debug_only(verify_map()); 696 } 697 698 699 //----------------------------kill_dead_locals--------------------------------- 700 // Detect any locals which are known to be dead, and force them to top. 701 void GraphKit::kill_dead_locals() { 702 // Consult the liveness information for the locals. If any 703 // of them are unused, then they can be replaced by top(). This 704 // should help register allocation time and cut down on the size 705 // of the deoptimization information. 706 707 // This call is made from many of the bytecode handling 708 // subroutines called from the Big Switch in do_one_bytecode. 709 // Every bytecode which might include a slow path is responsible 710 // for killing its dead locals. The more consistent we 711 // are about killing deads, the fewer useless phis will be 712 // constructed for them at various merge points. 713 714 // bci can be -1 (InvocationEntryBci). We return the entry 715 // liveness for the method. 716 717 if (method() == NULL || method()->code_size() == 0) { 718 // We are building a graph for a call to a native method. 719 // All locals are live. 720 return; 721 } 722 723 ResourceMark rm; 724 725 // Consult the liveness information for the locals. If any 726 // of them are unused, then they can be replaced by top(). This 727 // should help register allocation time and cut down on the size 728 // of the deoptimization information. 729 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 730 731 int len = (int)live_locals.size(); 732 assert(len <= jvms()->loc_size(), "too many live locals"); 733 for (int local = 0; local < len; local++) { 734 if (!live_locals.at(local)) { 735 set_local(local, top()); 736 } 737 } 738 } 739 740 #ifdef ASSERT 741 //-------------------------dead_locals_are_killed------------------------------ 742 // Return true if all dead locals are set to top in the map. 743 // Used to assert "clean" debug info at various points. 744 bool GraphKit::dead_locals_are_killed() { 745 if (method() == NULL || method()->code_size() == 0) { 746 // No locals need to be dead, so all is as it should be. 747 return true; 748 } 749 750 // Make sure somebody called kill_dead_locals upstream. 751 ResourceMark rm; 752 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 753 if (jvms->loc_size() == 0) continue; // no locals to consult 754 SafePointNode* map = jvms->map(); 755 ciMethod* method = jvms->method(); 756 int bci = jvms->bci(); 757 if (jvms == this->jvms()) { 758 bci = this->bci(); // it might not yet be synched 759 } 760 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 761 int len = (int)live_locals.size(); 762 if (!live_locals.is_valid() || len == 0) 763 // This method is trivial, or is poisoned by a breakpoint. 764 return true; 765 assert(len == jvms->loc_size(), "live map consistent with locals map"); 766 for (int local = 0; local < len; local++) { 767 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 768 if (PrintMiscellaneous && (Verbose || WizardMode)) { 769 tty->print_cr("Zombie local %d: ", local); 770 jvms->dump(); 771 } 772 return false; 773 } 774 } 775 } 776 return true; 777 } 778 779 #endif //ASSERT 780 781 // Helper function for enforcing certain bytecodes to reexecute if 782 // deoptimization happens 783 static bool should_reexecute_implied_by_bytecode(JVMState *jvms) { 784 ciMethod* cur_method = jvms->method(); 785 int cur_bci = jvms->bci(); 786 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 787 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 788 return Interpreter::bytecode_should_reexecute(code); 789 } else 790 return false; 791 } 792 793 // Helper function for adding JVMState and debug information to node 794 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 795 // Add the safepoint edges to the call (or other safepoint). 796 797 // Make sure dead locals are set to top. This 798 // should help register allocation time and cut down on the size 799 // of the deoptimization information. 800 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 801 802 // Walk the inline list to fill in the correct set of JVMState's 803 // Also fill in the associated edges for each JVMState. 804 805 JVMState* youngest_jvms = sync_jvms(); 806 807 // Do we need debug info here? If it is a SafePoint and this method 808 // cannot de-opt, then we do NOT need any debug info. 809 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint); 810 811 // If we are guaranteed to throw, we can prune everything but the 812 // input to the current bytecode. 813 bool can_prune_locals = false; 814 uint stack_slots_not_pruned = 0; 815 int inputs = 0, depth = 0; 816 if (must_throw) { 817 assert(method() == youngest_jvms->method(), "sanity"); 818 if (compute_stack_effects(inputs, depth)) { 819 can_prune_locals = true; 820 stack_slots_not_pruned = inputs; 821 } 822 } 823 824 if (env()->jvmti_can_examine_or_deopt_anywhere()) { 825 // At any safepoint, this method can get breakpointed, which would 826 // then require an immediate deoptimization. 827 full_info = true; 828 can_prune_locals = false; // do not prune locals 829 stack_slots_not_pruned = 0; 830 } 831 832 // do not scribble on the input jvms 833 JVMState* out_jvms = youngest_jvms->clone_deep(C); 834 call->set_jvms(out_jvms); // Start jvms list for call node 835 836 // For a known set of bytecodes, the interpreter should reexecute them if 837 // deoptimization happens. We set the reexecute state for them here 838 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 839 should_reexecute_implied_by_bytecode(out_jvms)) { 840 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 841 } 842 843 // Presize the call: 844 debug_only(uint non_debug_edges = call->req()); 845 call->add_req_batch(top(), youngest_jvms->debug_depth()); 846 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 847 848 // Set up edges so that the call looks like this: 849 // Call [state:] ctl io mem fptr retadr 850 // [parms:] parm0 ... parmN 851 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 852 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 853 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 854 // Note that caller debug info precedes callee debug info. 855 856 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 857 uint debug_ptr = call->req(); 858 859 // Loop over the map input edges associated with jvms, add them 860 // to the call node, & reset all offsets to match call node array. 861 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 862 uint debug_end = debug_ptr; 863 uint debug_start = debug_ptr - in_jvms->debug_size(); 864 debug_ptr = debug_start; // back up the ptr 865 866 uint p = debug_start; // walks forward in [debug_start, debug_end) 867 uint j, k, l; 868 SafePointNode* in_map = in_jvms->map(); 869 out_jvms->set_map(call); 870 871 if (can_prune_locals) { 872 assert(in_jvms->method() == out_jvms->method(), "sanity"); 873 // If the current throw can reach an exception handler in this JVMS, 874 // then we must keep everything live that can reach that handler. 875 // As a quick and dirty approximation, we look for any handlers at all. 876 if (in_jvms->method()->has_exception_handlers()) { 877 can_prune_locals = false; 878 } 879 } 880 881 // Add the Locals 882 k = in_jvms->locoff(); 883 l = in_jvms->loc_size(); 884 out_jvms->set_locoff(p); 885 if (full_info && !can_prune_locals) { 886 for (j = 0; j < l; j++) 887 call->set_req(p++, in_map->in(k+j)); 888 } else { 889 p += l; // already set to top above by add_req_batch 890 } 891 892 // Add the Expression Stack 893 k = in_jvms->stkoff(); 894 l = in_jvms->sp(); 895 out_jvms->set_stkoff(p); 896 if (full_info && !can_prune_locals) { 897 for (j = 0; j < l; j++) 898 call->set_req(p++, in_map->in(k+j)); 899 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 900 // Divide stack into {S0,...,S1}, where S0 is set to top. 901 uint s1 = stack_slots_not_pruned; 902 stack_slots_not_pruned = 0; // for next iteration 903 if (s1 > l) s1 = l; 904 uint s0 = l - s1; 905 p += s0; // skip the tops preinstalled by add_req_batch 906 for (j = s0; j < l; j++) 907 call->set_req(p++, in_map->in(k+j)); 908 } else { 909 p += l; // already set to top above by add_req_batch 910 } 911 912 // Add the Monitors 913 k = in_jvms->monoff(); 914 l = in_jvms->mon_size(); 915 out_jvms->set_monoff(p); 916 for (j = 0; j < l; j++) 917 call->set_req(p++, in_map->in(k+j)); 918 919 // Copy any scalar object fields. 920 k = in_jvms->scloff(); 921 l = in_jvms->scl_size(); 922 out_jvms->set_scloff(p); 923 for (j = 0; j < l; j++) 924 call->set_req(p++, in_map->in(k+j)); 925 926 // Finish the new jvms. 927 out_jvms->set_endoff(p); 928 929 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 930 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 931 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 932 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 933 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 934 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 935 936 // Update the two tail pointers in parallel. 937 out_jvms = out_jvms->caller(); 938 in_jvms = in_jvms->caller(); 939 } 940 941 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 942 943 // Test the correctness of JVMState::debug_xxx accessors: 944 assert(call->jvms()->debug_start() == non_debug_edges, ""); 945 assert(call->jvms()->debug_end() == call->req(), ""); 946 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 947 } 948 949 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 950 Bytecodes::Code code = java_bc(); 951 if (code == Bytecodes::_wide) { 952 code = method()->java_code_at_bci(bci() + 1); 953 } 954 955 BasicType rtype = T_ILLEGAL; 956 int rsize = 0; 957 958 if (code != Bytecodes::_illegal) { 959 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 960 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 961 if (rtype < T_CONFLICT) 962 rsize = type2size[rtype]; 963 } 964 965 switch (code) { 966 case Bytecodes::_illegal: 967 return false; 968 969 case Bytecodes::_ldc: 970 case Bytecodes::_ldc_w: 971 case Bytecodes::_ldc2_w: 972 inputs = 0; 973 break; 974 975 case Bytecodes::_dup: inputs = 1; break; 976 case Bytecodes::_dup_x1: inputs = 2; break; 977 case Bytecodes::_dup_x2: inputs = 3; break; 978 case Bytecodes::_dup2: inputs = 2; break; 979 case Bytecodes::_dup2_x1: inputs = 3; break; 980 case Bytecodes::_dup2_x2: inputs = 4; break; 981 case Bytecodes::_swap: inputs = 2; break; 982 case Bytecodes::_arraylength: inputs = 1; break; 983 984 case Bytecodes::_getstatic: 985 case Bytecodes::_putstatic: 986 case Bytecodes::_getfield: 987 case Bytecodes::_putfield: 988 { 989 bool is_get = (depth >= 0), is_static = (depth & 1); 990 bool ignore; 991 ciBytecodeStream iter(method()); 992 iter.reset_to_bci(bci()); 993 iter.next(); 994 ciField* field = iter.get_field(ignore); 995 int size = field->type()->size(); 996 inputs = (is_static ? 0 : 1); 997 if (is_get) { 998 depth = size - inputs; 999 } else { 1000 inputs += size; // putxxx pops the value from the stack 1001 depth = - inputs; 1002 } 1003 } 1004 break; 1005 1006 case Bytecodes::_invokevirtual: 1007 case Bytecodes::_invokespecial: 1008 case Bytecodes::_invokestatic: 1009 case Bytecodes::_invokedynamic: 1010 case Bytecodes::_invokeinterface: 1011 { 1012 bool ignore; 1013 ciBytecodeStream iter(method()); 1014 iter.reset_to_bci(bci()); 1015 iter.next(); 1016 ciMethod* method = iter.get_method(ignore); 1017 inputs = method->arg_size_no_receiver(); 1018 // Add a receiver argument, maybe: 1019 if (code != Bytecodes::_invokestatic && 1020 code != Bytecodes::_invokedynamic) 1021 inputs += 1; 1022 // (Do not use ciMethod::arg_size(), because 1023 // it might be an unloaded method, which doesn't 1024 // know whether it is static or not.) 1025 int size = method->return_type()->size(); 1026 depth = size - inputs; 1027 } 1028 break; 1029 1030 case Bytecodes::_multianewarray: 1031 { 1032 ciBytecodeStream iter(method()); 1033 iter.reset_to_bci(bci()); 1034 iter.next(); 1035 inputs = iter.get_dimensions(); 1036 assert(rsize == 1, ""); 1037 depth = rsize - inputs; 1038 } 1039 break; 1040 1041 case Bytecodes::_ireturn: 1042 case Bytecodes::_lreturn: 1043 case Bytecodes::_freturn: 1044 case Bytecodes::_dreturn: 1045 case Bytecodes::_areturn: 1046 assert(rsize = -depth, ""); 1047 inputs = rsize; 1048 break; 1049 1050 case Bytecodes::_jsr: 1051 case Bytecodes::_jsr_w: 1052 inputs = 0; 1053 depth = 1; // S.B. depth=1, not zero 1054 break; 1055 1056 default: 1057 // bytecode produces a typed result 1058 inputs = rsize - depth; 1059 assert(inputs >= 0, ""); 1060 break; 1061 } 1062 1063 #ifdef ASSERT 1064 // spot check 1065 int outputs = depth + inputs; 1066 assert(outputs >= 0, "sanity"); 1067 switch (code) { 1068 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1069 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1070 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1071 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1072 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1073 } 1074 #endif //ASSERT 1075 1076 return true; 1077 } 1078 1079 1080 1081 //------------------------------basic_plus_adr--------------------------------- 1082 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1083 // short-circuit a common case 1084 if (offset == intcon(0)) return ptr; 1085 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) ); 1086 } 1087 1088 Node* GraphKit::ConvI2L(Node* offset) { 1089 // short-circuit a common case 1090 jint offset_con = find_int_con(offset, Type::OffsetBot); 1091 if (offset_con != Type::OffsetBot) { 1092 return longcon((long) offset_con); 1093 } 1094 return _gvn.transform( new (C, 2) ConvI2LNode(offset)); 1095 } 1096 Node* GraphKit::ConvL2I(Node* offset) { 1097 // short-circuit a common case 1098 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1099 if (offset_con != (jlong)Type::OffsetBot) { 1100 return intcon((int) offset_con); 1101 } 1102 return _gvn.transform( new (C, 2) ConvL2INode(offset)); 1103 } 1104 1105 //-------------------------load_object_klass----------------------------------- 1106 Node* GraphKit::load_object_klass(Node* obj) { 1107 // Special-case a fresh allocation to avoid building nodes: 1108 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1109 if (akls != NULL) return akls; 1110 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1111 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1112 } 1113 1114 //-------------------------load_array_length----------------------------------- 1115 Node* GraphKit::load_array_length(Node* array) { 1116 // Special-case a fresh allocation to avoid building nodes: 1117 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1118 Node *alen; 1119 if (alloc == NULL) { 1120 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1121 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1122 } else { 1123 alen = alloc->Ideal_length(); 1124 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1125 if (ccast != alen) { 1126 alen = _gvn.transform(ccast); 1127 } 1128 } 1129 return alen; 1130 } 1131 1132 //------------------------------do_null_check---------------------------------- 1133 // Helper function to do a NULL pointer check. Returned value is 1134 // the incoming address with NULL casted away. You are allowed to use the 1135 // not-null value only if you are control dependent on the test. 1136 extern int explicit_null_checks_inserted, 1137 explicit_null_checks_elided; 1138 Node* GraphKit::null_check_common(Node* value, BasicType type, 1139 // optional arguments for variations: 1140 bool assert_null, 1141 Node* *null_control) { 1142 assert(!assert_null || null_control == NULL, "not both at once"); 1143 if (stopped()) return top(); 1144 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1145 // For some performance testing, we may wish to suppress null checking. 1146 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1147 return value; 1148 } 1149 explicit_null_checks_inserted++; 1150 1151 // Construct NULL check 1152 Node *chk = NULL; 1153 switch(type) { 1154 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1155 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break; 1156 case T_ARRAY : // fall through 1157 type = T_OBJECT; // simplify further tests 1158 case T_OBJECT : { 1159 const Type *t = _gvn.type( value ); 1160 1161 const TypeOopPtr* tp = t->isa_oopptr(); 1162 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1163 // Only for do_null_check, not any of its siblings: 1164 && !assert_null && null_control == NULL) { 1165 // Usually, any field access or invocation on an unloaded oop type 1166 // will simply fail to link, since the statically linked class is 1167 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1168 // the static class is loaded but the sharper oop type is not. 1169 // Rather than checking for this obscure case in lots of places, 1170 // we simply observe that a null check on an unloaded class 1171 // will always be followed by a nonsense operation, so we 1172 // can just issue the uncommon trap here. 1173 // Our access to the unloaded class will only be correct 1174 // after it has been loaded and initialized, which requires 1175 // a trip through the interpreter. 1176 #ifndef PRODUCT 1177 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1178 #endif 1179 uncommon_trap(Deoptimization::Reason_unloaded, 1180 Deoptimization::Action_reinterpret, 1181 tp->klass(), "!loaded"); 1182 return top(); 1183 } 1184 1185 if (assert_null) { 1186 // See if the type is contained in NULL_PTR. 1187 // If so, then the value is already null. 1188 if (t->higher_equal(TypePtr::NULL_PTR)) { 1189 explicit_null_checks_elided++; 1190 return value; // Elided null assert quickly! 1191 } 1192 } else { 1193 // See if mixing in the NULL pointer changes type. 1194 // If so, then the NULL pointer was not allowed in the original 1195 // type. In other words, "value" was not-null. 1196 if (t->meet(TypePtr::NULL_PTR) != t) { 1197 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1198 explicit_null_checks_elided++; 1199 return value; // Elided null check quickly! 1200 } 1201 } 1202 chk = new (C, 3) CmpPNode( value, null() ); 1203 break; 1204 } 1205 1206 default : ShouldNotReachHere(); 1207 } 1208 assert(chk != NULL, "sanity check"); 1209 chk = _gvn.transform(chk); 1210 1211 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1212 BoolNode *btst = new (C, 2) BoolNode( chk, btest); 1213 Node *tst = _gvn.transform( btst ); 1214 1215 //----------- 1216 // if peephole optimizations occurred, a prior test existed. 1217 // If a prior test existed, maybe it dominates as we can avoid this test. 1218 if (tst != btst && type == T_OBJECT) { 1219 // At this point we want to scan up the CFG to see if we can 1220 // find an identical test (and so avoid this test altogether). 1221 Node *cfg = control(); 1222 int depth = 0; 1223 while( depth < 16 ) { // Limit search depth for speed 1224 if( cfg->Opcode() == Op_IfTrue && 1225 cfg->in(0)->in(1) == tst ) { 1226 // Found prior test. Use "cast_not_null" to construct an identical 1227 // CastPP (and hence hash to) as already exists for the prior test. 1228 // Return that casted value. 1229 if (assert_null) { 1230 replace_in_map(value, null()); 1231 return null(); // do not issue the redundant test 1232 } 1233 Node *oldcontrol = control(); 1234 set_control(cfg); 1235 Node *res = cast_not_null(value); 1236 set_control(oldcontrol); 1237 explicit_null_checks_elided++; 1238 return res; 1239 } 1240 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1241 if (cfg == NULL) break; // Quit at region nodes 1242 depth++; 1243 } 1244 } 1245 1246 //----------- 1247 // Branch to failure if null 1248 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1249 Deoptimization::DeoptReason reason; 1250 if (assert_null) 1251 reason = Deoptimization::Reason_null_assert; 1252 else if (type == T_OBJECT) 1253 reason = Deoptimization::Reason_null_check; 1254 else 1255 reason = Deoptimization::Reason_div0_check; 1256 1257 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1258 // ciMethodData::has_trap_at will return a conservative -1 if any 1259 // must-be-null assertion has failed. This could cause performance 1260 // problems for a method after its first do_null_assert failure. 1261 // Consider using 'Reason_class_check' instead? 1262 1263 // To cause an implicit null check, we set the not-null probability 1264 // to the maximum (PROB_MAX). For an explicit check the probability 1265 // is set to a smaller value. 1266 if (null_control != NULL || too_many_traps(reason)) { 1267 // probability is less likely 1268 ok_prob = PROB_LIKELY_MAG(3); 1269 } else if (!assert_null && 1270 (ImplicitNullCheckThreshold > 0) && 1271 method() != NULL && 1272 (method()->method_data()->trap_count(reason) 1273 >= (uint)ImplicitNullCheckThreshold)) { 1274 ok_prob = PROB_LIKELY_MAG(3); 1275 } 1276 1277 if (null_control != NULL) { 1278 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1279 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff)); 1280 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff))); 1281 if (null_true == top()) 1282 explicit_null_checks_elided++; 1283 (*null_control) = null_true; 1284 } else { 1285 BuildCutout unless(this, tst, ok_prob); 1286 // Check for optimizer eliding test at parse time 1287 if (stopped()) { 1288 // Failure not possible; do not bother making uncommon trap. 1289 explicit_null_checks_elided++; 1290 } else if (assert_null) { 1291 uncommon_trap(reason, 1292 Deoptimization::Action_make_not_entrant, 1293 NULL, "assert_null"); 1294 } else { 1295 replace_in_map(value, zerocon(type)); 1296 builtin_throw(reason); 1297 } 1298 } 1299 1300 // Must throw exception, fall-thru not possible? 1301 if (stopped()) { 1302 return top(); // No result 1303 } 1304 1305 if (assert_null) { 1306 // Cast obj to null on this path. 1307 replace_in_map(value, zerocon(type)); 1308 return zerocon(type); 1309 } 1310 1311 // Cast obj to not-null on this path, if there is no null_control. 1312 // (If there is a null_control, a non-null value may come back to haunt us.) 1313 if (type == T_OBJECT) { 1314 Node* cast = cast_not_null(value, false); 1315 if (null_control == NULL || (*null_control) == top()) 1316 replace_in_map(value, cast); 1317 value = cast; 1318 } 1319 1320 return value; 1321 } 1322 1323 1324 //------------------------------cast_not_null---------------------------------- 1325 // Cast obj to not-null on this path 1326 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1327 const Type *t = _gvn.type(obj); 1328 const Type *t_not_null = t->join(TypePtr::NOTNULL); 1329 // Object is already not-null? 1330 if( t == t_not_null ) return obj; 1331 1332 Node *cast = new (C, 2) CastPPNode(obj,t_not_null); 1333 cast->init_req(0, control()); 1334 cast = _gvn.transform( cast ); 1335 1336 // Scan for instances of 'obj' in the current JVM mapping. 1337 // These instances are known to be not-null after the test. 1338 if (do_replace_in_map) 1339 replace_in_map(obj, cast); 1340 1341 return cast; // Return casted value 1342 } 1343 1344 1345 //--------------------------replace_in_map------------------------------------- 1346 void GraphKit::replace_in_map(Node* old, Node* neww) { 1347 this->map()->replace_edge(old, neww); 1348 1349 // Note: This operation potentially replaces any edge 1350 // on the map. This includes locals, stack, and monitors 1351 // of the current (innermost) JVM state. 1352 1353 // We can consider replacing in caller maps. 1354 // The idea would be that an inlined function's null checks 1355 // can be shared with the entire inlining tree. 1356 // The expense of doing this is that the PreserveJVMState class 1357 // would have to preserve caller states too, with a deep copy. 1358 } 1359 1360 1361 1362 //============================================================================= 1363 //--------------------------------memory--------------------------------------- 1364 Node* GraphKit::memory(uint alias_idx) { 1365 MergeMemNode* mem = merged_memory(); 1366 Node* p = mem->memory_at(alias_idx); 1367 _gvn.set_type(p, Type::MEMORY); // must be mapped 1368 return p; 1369 } 1370 1371 //-----------------------------reset_memory------------------------------------ 1372 Node* GraphKit::reset_memory() { 1373 Node* mem = map()->memory(); 1374 // do not use this node for any more parsing! 1375 debug_only( map()->set_memory((Node*)NULL) ); 1376 return _gvn.transform( mem ); 1377 } 1378 1379 //------------------------------set_all_memory--------------------------------- 1380 void GraphKit::set_all_memory(Node* newmem) { 1381 Node* mergemem = MergeMemNode::make(C, newmem); 1382 gvn().set_type_bottom(mergemem); 1383 map()->set_memory(mergemem); 1384 } 1385 1386 //------------------------------set_all_memory_call---------------------------- 1387 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1388 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1389 set_all_memory(newmem); 1390 } 1391 1392 //============================================================================= 1393 // 1394 // parser factory methods for MemNodes 1395 // 1396 // These are layered on top of the factory methods in LoadNode and StoreNode, 1397 // and integrate with the parser's memory state and _gvn engine. 1398 // 1399 1400 // factory methods in "int adr_idx" 1401 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1402 int adr_idx, 1403 bool require_atomic_access) { 1404 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1405 const TypePtr* adr_type = NULL; // debug-mode-only argument 1406 debug_only(adr_type = C->get_adr_type(adr_idx)); 1407 Node* mem = memory(adr_idx); 1408 Node* ld; 1409 if (require_atomic_access && bt == T_LONG) { 1410 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t); 1411 } else { 1412 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt); 1413 } 1414 return _gvn.transform(ld); 1415 } 1416 1417 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1418 int adr_idx, 1419 bool require_atomic_access) { 1420 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1421 const TypePtr* adr_type = NULL; 1422 debug_only(adr_type = C->get_adr_type(adr_idx)); 1423 Node *mem = memory(adr_idx); 1424 Node* st; 1425 if (require_atomic_access && bt == T_LONG) { 1426 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val); 1427 } else { 1428 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt); 1429 } 1430 st = _gvn.transform(st); 1431 set_memory(st, adr_idx); 1432 // Back-to-back stores can only remove intermediate store with DU info 1433 // so push on worklist for optimizer. 1434 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1435 record_for_igvn(st); 1436 1437 return st; 1438 } 1439 1440 1441 void GraphKit::pre_barrier(Node* ctl, 1442 Node* obj, 1443 Node* adr, 1444 uint adr_idx, 1445 Node* val, 1446 const TypeOopPtr* val_type, 1447 BasicType bt) { 1448 BarrierSet* bs = Universe::heap()->barrier_set(); 1449 set_control(ctl); 1450 switch (bs->kind()) { 1451 case BarrierSet::G1SATBCT: 1452 case BarrierSet::G1SATBCTLogging: 1453 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt); 1454 break; 1455 1456 case BarrierSet::CardTableModRef: 1457 case BarrierSet::CardTableExtension: 1458 case BarrierSet::ModRef: 1459 break; 1460 1461 case BarrierSet::Other: 1462 default : 1463 ShouldNotReachHere(); 1464 1465 } 1466 } 1467 1468 void GraphKit::post_barrier(Node* ctl, 1469 Node* store, 1470 Node* obj, 1471 Node* adr, 1472 uint adr_idx, 1473 Node* val, 1474 BasicType bt, 1475 bool use_precise) { 1476 BarrierSet* bs = Universe::heap()->barrier_set(); 1477 set_control(ctl); 1478 switch (bs->kind()) { 1479 case BarrierSet::G1SATBCT: 1480 case BarrierSet::G1SATBCTLogging: 1481 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1482 break; 1483 1484 case BarrierSet::CardTableModRef: 1485 case BarrierSet::CardTableExtension: 1486 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1487 break; 1488 1489 case BarrierSet::ModRef: 1490 break; 1491 1492 case BarrierSet::Other: 1493 default : 1494 ShouldNotReachHere(); 1495 1496 } 1497 } 1498 1499 Node* GraphKit::store_oop(Node* ctl, 1500 Node* obj, 1501 Node* adr, 1502 const TypePtr* adr_type, 1503 Node* val, 1504 const TypeOopPtr* val_type, 1505 BasicType bt, 1506 bool use_precise) { 1507 1508 set_control(ctl); 1509 if (stopped()) return top(); // Dead path ? 1510 1511 assert(bt == T_OBJECT, "sanity"); 1512 assert(val != NULL, "not dead path"); 1513 uint adr_idx = C->get_alias_index(adr_type); 1514 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1515 1516 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt); 1517 Node* store = store_to_memory(control(), adr, val, bt, adr_idx); 1518 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1519 return store; 1520 } 1521 1522 // Could be an array or object we don't know at compile time (unsafe ref.) 1523 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1524 Node* obj, // containing obj 1525 Node* adr, // actual adress to store val at 1526 const TypePtr* adr_type, 1527 Node* val, 1528 BasicType bt) { 1529 Compile::AliasType* at = C->alias_type(adr_type); 1530 const TypeOopPtr* val_type = NULL; 1531 if (adr_type->isa_instptr()) { 1532 if (at->field() != NULL) { 1533 // known field. This code is a copy of the do_put_xxx logic. 1534 ciField* field = at->field(); 1535 if (!field->type()->is_loaded()) { 1536 val_type = TypeInstPtr::BOTTOM; 1537 } else { 1538 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1539 } 1540 } 1541 } else if (adr_type->isa_aryptr()) { 1542 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1543 } 1544 if (val_type == NULL) { 1545 val_type = TypeInstPtr::BOTTOM; 1546 } 1547 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true); 1548 } 1549 1550 1551 //-------------------------array_element_address------------------------- 1552 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1553 const TypeInt* sizetype) { 1554 uint shift = exact_log2(type2aelembytes(elembt)); 1555 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1556 1557 // short-circuit a common case (saves lots of confusing waste motion) 1558 jint idx_con = find_int_con(idx, -1); 1559 if (idx_con >= 0) { 1560 intptr_t offset = header + ((intptr_t)idx_con << shift); 1561 return basic_plus_adr(ary, offset); 1562 } 1563 1564 // must be correct type for alignment purposes 1565 Node* base = basic_plus_adr(ary, header); 1566 #ifdef _LP64 1567 // The scaled index operand to AddP must be a clean 64-bit value. 1568 // Java allows a 32-bit int to be incremented to a negative 1569 // value, which appears in a 64-bit register as a large 1570 // positive number. Using that large positive number as an 1571 // operand in pointer arithmetic has bad consequences. 1572 // On the other hand, 32-bit overflow is rare, and the possibility 1573 // can often be excluded, if we annotate the ConvI2L node with 1574 // a type assertion that its value is known to be a small positive 1575 // number. (The prior range check has ensured this.) 1576 // This assertion is used by ConvI2LNode::Ideal. 1577 int index_max = max_jint - 1; // array size is max_jint, index is one less 1578 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1579 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1580 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) ); 1581 #endif 1582 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) ); 1583 return basic_plus_adr(ary, base, scale); 1584 } 1585 1586 //-------------------------load_array_element------------------------- 1587 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1588 const Type* elemtype = arytype->elem(); 1589 BasicType elembt = elemtype->array_element_basic_type(); 1590 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1591 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype); 1592 return ld; 1593 } 1594 1595 //-------------------------set_arguments_for_java_call------------------------- 1596 // Arguments (pre-popped from the stack) are taken from the JVMS. 1597 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1598 // Add the call arguments: 1599 uint nargs = call->method()->arg_size(); 1600 for (uint i = 0; i < nargs; i++) { 1601 Node* arg = argument(i); 1602 call->init_req(i + TypeFunc::Parms, arg); 1603 } 1604 } 1605 1606 //---------------------------set_edges_for_java_call--------------------------- 1607 // Connect a newly created call into the current JVMS. 1608 // A return value node (if any) is returned from set_edges_for_java_call. 1609 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1610 1611 // Add the predefined inputs: 1612 call->init_req( TypeFunc::Control, control() ); 1613 call->init_req( TypeFunc::I_O , i_o() ); 1614 call->init_req( TypeFunc::Memory , reset_memory() ); 1615 call->init_req( TypeFunc::FramePtr, frameptr() ); 1616 call->init_req( TypeFunc::ReturnAdr, top() ); 1617 1618 add_safepoint_edges(call, must_throw); 1619 1620 Node* xcall = _gvn.transform(call); 1621 1622 if (xcall == top()) { 1623 set_control(top()); 1624 return; 1625 } 1626 assert(xcall == call, "call identity is stable"); 1627 1628 // Re-use the current map to produce the result. 1629 1630 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control))); 1631 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1632 set_all_memory_call(xcall, separate_io_proj); 1633 1634 //return xcall; // no need, caller already has it 1635 } 1636 1637 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1638 if (stopped()) return top(); // maybe the call folded up? 1639 1640 // Capture the return value, if any. 1641 Node* ret; 1642 if (call->method() == NULL || 1643 call->method()->return_type()->basic_type() == T_VOID) 1644 ret = top(); 1645 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms)); 1646 1647 // Note: Since any out-of-line call can produce an exception, 1648 // we always insert an I_O projection from the call into the result. 1649 1650 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1651 1652 if (separate_io_proj) { 1653 // The caller requested separate projections be used by the fall 1654 // through and exceptional paths, so replace the projections for 1655 // the fall through path. 1656 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) )); 1657 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) )); 1658 } 1659 return ret; 1660 } 1661 1662 //--------------------set_predefined_input_for_runtime_call-------------------- 1663 // Reading and setting the memory state is way conservative here. 1664 // The real problem is that I am not doing real Type analysis on memory, 1665 // so I cannot distinguish card mark stores from other stores. Across a GC 1666 // point the Store Barrier and the card mark memory has to agree. I cannot 1667 // have a card mark store and its barrier split across the GC point from 1668 // either above or below. Here I get that to happen by reading ALL of memory. 1669 // A better answer would be to separate out card marks from other memory. 1670 // For now, return the input memory state, so that it can be reused 1671 // after the call, if this call has restricted memory effects. 1672 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1673 // Set fixed predefined input arguments 1674 Node* memory = reset_memory(); 1675 call->init_req( TypeFunc::Control, control() ); 1676 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1677 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1678 call->init_req( TypeFunc::FramePtr, frameptr() ); 1679 call->init_req( TypeFunc::ReturnAdr, top() ); 1680 return memory; 1681 } 1682 1683 //-------------------set_predefined_output_for_runtime_call-------------------- 1684 // Set control and memory (not i_o) from the call. 1685 // If keep_mem is not NULL, use it for the output state, 1686 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1687 // If hook_mem is NULL, this call produces no memory effects at all. 1688 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1689 // then only that memory slice is taken from the call. 1690 // In the last case, we must put an appropriate memory barrier before 1691 // the call, so as to create the correct anti-dependencies on loads 1692 // preceding the call. 1693 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1694 Node* keep_mem, 1695 const TypePtr* hook_mem) { 1696 // no i/o 1697 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) )); 1698 if (keep_mem) { 1699 // First clone the existing memory state 1700 set_all_memory(keep_mem); 1701 if (hook_mem != NULL) { 1702 // Make memory for the call 1703 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ); 1704 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1705 // We also use hook_mem to extract specific effects from arraycopy stubs. 1706 set_memory(mem, hook_mem); 1707 } 1708 // ...else the call has NO memory effects. 1709 1710 // Make sure the call advertises its memory effects precisely. 1711 // This lets us build accurate anti-dependences in gcm.cpp. 1712 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1713 "call node must be constructed correctly"); 1714 } else { 1715 assert(hook_mem == NULL, ""); 1716 // This is not a "slow path" call; all memory comes from the call. 1717 set_all_memory_call(call); 1718 } 1719 } 1720 1721 1722 // Replace the call with the current state of the kit. 1723 void GraphKit::replace_call(CallNode* call, Node* result) { 1724 JVMState* ejvms = NULL; 1725 if (has_exceptions()) { 1726 ejvms = transfer_exceptions_into_jvms(); 1727 } 1728 1729 SafePointNode* final_state = stop(); 1730 1731 // Find all the needed outputs of this call 1732 CallProjections callprojs; 1733 call->extract_projections(&callprojs, true); 1734 1735 // Replace all the old call edges with the edges from the inlining result 1736 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control)); 1737 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory)); 1738 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O)); 1739 1740 // Replace the result with the new result if it exists and is used 1741 if (callprojs.resproj != NULL && result != NULL) { 1742 C->gvn_replace_by(callprojs.resproj, result); 1743 } 1744 1745 if (ejvms == NULL) { 1746 // No exception edges to simply kill off those paths 1747 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1748 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1749 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1750 1751 // Replace the old exception object with top 1752 if (callprojs.exobj != NULL) { 1753 C->gvn_replace_by(callprojs.exobj, C->top()); 1754 } 1755 } else { 1756 GraphKit ekit(ejvms); 1757 1758 // Load my combined exception state into the kit, with all phis transformed: 1759 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1760 1761 Node* ex_oop = ekit.use_exception_state(ex_map); 1762 1763 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1764 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1765 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1766 1767 // Replace the old exception object with the newly created one 1768 if (callprojs.exobj != NULL) { 1769 C->gvn_replace_by(callprojs.exobj, ex_oop); 1770 } 1771 } 1772 1773 // Disconnect the call from the graph 1774 call->disconnect_inputs(NULL); 1775 C->gvn_replace_by(call, C->top()); 1776 } 1777 1778 1779 //------------------------------increment_counter------------------------------ 1780 // for statistics: increment a VM counter by 1 1781 1782 void GraphKit::increment_counter(address counter_addr) { 1783 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1784 increment_counter(adr1); 1785 } 1786 1787 void GraphKit::increment_counter(Node* counter_addr) { 1788 int adr_type = Compile::AliasIdxRaw; 1789 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 1790 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 1791 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type ); 1792 } 1793 1794 1795 //------------------------------uncommon_trap---------------------------------- 1796 // Bail out to the interpreter in mid-method. Implemented by calling the 1797 // uncommon_trap blob. This helper function inserts a runtime call with the 1798 // right debug info. 1799 void GraphKit::uncommon_trap(int trap_request, 1800 ciKlass* klass, const char* comment, 1801 bool must_throw, 1802 bool keep_exact_action) { 1803 if (failing()) stop(); 1804 if (stopped()) return; // trap reachable? 1805 1806 // Note: If ProfileTraps is true, and if a deopt. actually 1807 // occurs here, the runtime will make sure an MDO exists. There is 1808 // no need to call method()->build_method_data() at this point. 1809 1810 #ifdef ASSERT 1811 if (!must_throw) { 1812 // Make sure the stack has at least enough depth to execute 1813 // the current bytecode. 1814 int inputs, ignore; 1815 if (compute_stack_effects(inputs, ignore)) { 1816 assert(sp() >= inputs, "must have enough JVMS stack to execute"); 1817 // It is a frequent error in library_call.cpp to issue an 1818 // uncommon trap with the _sp value already popped. 1819 } 1820 } 1821 #endif 1822 1823 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1824 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1825 1826 switch (action) { 1827 case Deoptimization::Action_maybe_recompile: 1828 case Deoptimization::Action_reinterpret: 1829 // Temporary fix for 6529811 to allow virtual calls to be sure they 1830 // get the chance to go from mono->bi->mega 1831 if (!keep_exact_action && 1832 Deoptimization::trap_request_index(trap_request) < 0 && 1833 too_many_recompiles(reason)) { 1834 // This BCI is causing too many recompilations. 1835 action = Deoptimization::Action_none; 1836 trap_request = Deoptimization::make_trap_request(reason, action); 1837 } else { 1838 C->set_trap_can_recompile(true); 1839 } 1840 break; 1841 case Deoptimization::Action_make_not_entrant: 1842 C->set_trap_can_recompile(true); 1843 break; 1844 #ifdef ASSERT 1845 case Deoptimization::Action_none: 1846 case Deoptimization::Action_make_not_compilable: 1847 break; 1848 default: 1849 assert(false, "bad action"); 1850 #endif 1851 } 1852 1853 if (TraceOptoParse) { 1854 char buf[100]; 1855 tty->print_cr("Uncommon trap %s at bci:%d", 1856 Deoptimization::format_trap_request(buf, sizeof(buf), 1857 trap_request), bci()); 1858 } 1859 1860 CompileLog* log = C->log(); 1861 if (log != NULL) { 1862 int kid = (klass == NULL)? -1: log->identify(klass); 1863 log->begin_elem("uncommon_trap bci='%d'", bci()); 1864 char buf[100]; 1865 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1866 trap_request)); 1867 if (kid >= 0) log->print(" klass='%d'", kid); 1868 if (comment != NULL) log->print(" comment='%s'", comment); 1869 log->end_elem(); 1870 } 1871 1872 // Make sure any guarding test views this path as very unlikely 1873 Node *i0 = control()->in(0); 1874 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1875 IfNode *iff = i0->as_If(); 1876 float f = iff->_prob; // Get prob 1877 if (control()->Opcode() == Op_IfTrue) { 1878 if (f > PROB_UNLIKELY_MAG(4)) 1879 iff->_prob = PROB_MIN; 1880 } else { 1881 if (f < PROB_LIKELY_MAG(4)) 1882 iff->_prob = PROB_MAX; 1883 } 1884 } 1885 1886 // Clear out dead values from the debug info. 1887 kill_dead_locals(); 1888 1889 // Now insert the uncommon trap subroutine call 1890 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin(); 1891 const TypePtr* no_memory_effects = NULL; 1892 // Pass the index of the class to be loaded 1893 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 1894 (must_throw ? RC_MUST_THROW : 0), 1895 OptoRuntime::uncommon_trap_Type(), 1896 call_addr, "uncommon_trap", no_memory_effects, 1897 intcon(trap_request)); 1898 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 1899 "must extract request correctly from the graph"); 1900 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 1901 1902 call->set_req(TypeFunc::ReturnAdr, returnadr()); 1903 // The debug info is the only real input to this call. 1904 1905 // Halt-and-catch fire here. The above call should never return! 1906 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr()); 1907 _gvn.set_type_bottom(halt); 1908 root()->add_req(halt); 1909 1910 stop_and_kill_map(); 1911 } 1912 1913 1914 //--------------------------just_allocated_object------------------------------ 1915 // Report the object that was just allocated. 1916 // It must be the case that there are no intervening safepoints. 1917 // We use this to determine if an object is so "fresh" that 1918 // it does not require card marks. 1919 Node* GraphKit::just_allocated_object(Node* current_control) { 1920 if (C->recent_alloc_ctl() == current_control) 1921 return C->recent_alloc_obj(); 1922 return NULL; 1923 } 1924 1925 1926 void GraphKit::round_double_arguments(ciMethod* dest_method) { 1927 // (Note: TypeFunc::make has a cache that makes this fast.) 1928 const TypeFunc* tf = TypeFunc::make(dest_method); 1929 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 1930 for (int j = 0; j < nargs; j++) { 1931 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 1932 if( targ->basic_type() == T_DOUBLE ) { 1933 // If any parameters are doubles, they must be rounded before 1934 // the call, dstore_rounding does gvn.transform 1935 Node *arg = argument(j); 1936 arg = dstore_rounding(arg); 1937 set_argument(j, arg); 1938 } 1939 } 1940 } 1941 1942 void GraphKit::round_double_result(ciMethod* dest_method) { 1943 // A non-strict method may return a double value which has an extended 1944 // exponent, but this must not be visible in a caller which is 'strict' 1945 // If a strict caller invokes a non-strict callee, round a double result 1946 1947 BasicType result_type = dest_method->return_type()->basic_type(); 1948 assert( method() != NULL, "must have caller context"); 1949 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 1950 // Destination method's return value is on top of stack 1951 // dstore_rounding() does gvn.transform 1952 Node *result = pop_pair(); 1953 result = dstore_rounding(result); 1954 push_pair(result); 1955 } 1956 } 1957 1958 // rounding for strict float precision conformance 1959 Node* GraphKit::precision_rounding(Node* n) { 1960 return UseStrictFP && _method->flags().is_strict() 1961 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 1962 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) ) 1963 : n; 1964 } 1965 1966 // rounding for strict double precision conformance 1967 Node* GraphKit::dprecision_rounding(Node *n) { 1968 return UseStrictFP && _method->flags().is_strict() 1969 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 1970 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1971 : n; 1972 } 1973 1974 // rounding for non-strict double stores 1975 Node* GraphKit::dstore_rounding(Node* n) { 1976 return Matcher::strict_fp_requires_explicit_rounding 1977 && UseSSE <= 1 1978 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) ) 1979 : n; 1980 } 1981 1982 //============================================================================= 1983 // Generate a fast path/slow path idiom. Graph looks like: 1984 // [foo] indicates that 'foo' is a parameter 1985 // 1986 // [in] NULL 1987 // \ / 1988 // CmpP 1989 // Bool ne 1990 // If 1991 // / \ 1992 // True False-<2> 1993 // / | 1994 // / cast_not_null 1995 // Load | | ^ 1996 // [fast_test] | | 1997 // gvn to opt_test | | 1998 // / \ | <1> 1999 // True False | 2000 // | \\ | 2001 // [slow_call] \[fast_result] 2002 // Ctl Val \ \ 2003 // | \ \ 2004 // Catch <1> \ \ 2005 // / \ ^ \ \ 2006 // Ex No_Ex | \ \ 2007 // | \ \ | \ <2> \ 2008 // ... \ [slow_res] | | \ [null_result] 2009 // \ \--+--+--- | | 2010 // \ | / \ | / 2011 // --------Region Phi 2012 // 2013 //============================================================================= 2014 // Code is structured as a series of driver functions all called 'do_XXX' that 2015 // call a set of helper functions. Helper functions first, then drivers. 2016 2017 //------------------------------null_check_oop--------------------------------- 2018 // Null check oop. Set null-path control into Region in slot 3. 2019 // Make a cast-not-nullness use the other not-null control. Return cast. 2020 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2021 bool never_see_null) { 2022 // Initial NULL check taken path 2023 (*null_control) = top(); 2024 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2025 2026 // Generate uncommon_trap: 2027 if (never_see_null && (*null_control) != top()) { 2028 // If we see an unexpected null at a check-cast we record it and force a 2029 // recompile; the offending check-cast will be compiled to handle NULLs. 2030 // If we see more than one offending BCI, then all checkcasts in the 2031 // method will be compiled to handle NULLs. 2032 PreserveJVMState pjvms(this); 2033 set_control(*null_control); 2034 replace_in_map(value, null()); 2035 uncommon_trap(Deoptimization::Reason_null_check, 2036 Deoptimization::Action_make_not_entrant); 2037 (*null_control) = top(); // NULL path is dead 2038 } 2039 2040 // Cast away null-ness on the result 2041 return cast; 2042 } 2043 2044 //------------------------------opt_iff---------------------------------------- 2045 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2046 // Return slow-path control. 2047 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2048 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2049 2050 // Fast path taken; set region slot 2 2051 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) ); 2052 region->init_req(2,fast_taken); // Capture fast-control 2053 2054 // Fast path not-taken, i.e. slow path 2055 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) ); 2056 return slow_taken; 2057 } 2058 2059 //-----------------------------make_runtime_call------------------------------- 2060 Node* GraphKit::make_runtime_call(int flags, 2061 const TypeFunc* call_type, address call_addr, 2062 const char* call_name, 2063 const TypePtr* adr_type, 2064 // The following parms are all optional. 2065 // The first NULL ends the list. 2066 Node* parm0, Node* parm1, 2067 Node* parm2, Node* parm3, 2068 Node* parm4, Node* parm5, 2069 Node* parm6, Node* parm7) { 2070 // Slow-path call 2071 int size = call_type->domain()->cnt(); 2072 bool is_leaf = !(flags & RC_NO_LEAF); 2073 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2074 if (call_name == NULL) { 2075 assert(!is_leaf, "must supply name for leaf"); 2076 call_name = OptoRuntime::stub_name(call_addr); 2077 } 2078 CallNode* call; 2079 if (!is_leaf) { 2080 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name, 2081 bci(), adr_type); 2082 } else if (flags & RC_NO_FP) { 2083 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2084 } else { 2085 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type); 2086 } 2087 2088 // The following is similar to set_edges_for_java_call, 2089 // except that the memory effects of the call are restricted to AliasIdxRaw. 2090 2091 // Slow path call has no side-effects, uses few values 2092 bool wide_in = !(flags & RC_NARROW_MEM); 2093 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2094 2095 Node* prev_mem = NULL; 2096 if (wide_in) { 2097 prev_mem = set_predefined_input_for_runtime_call(call); 2098 } else { 2099 assert(!wide_out, "narrow in => narrow out"); 2100 Node* narrow_mem = memory(adr_type); 2101 prev_mem = reset_memory(); 2102 map()->set_memory(narrow_mem); 2103 set_predefined_input_for_runtime_call(call); 2104 } 2105 2106 // Hook each parm in order. Stop looking at the first NULL. 2107 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2108 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2109 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2110 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2111 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2112 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2113 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2114 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2115 /* close each nested if ===> */ } } } } } } } } 2116 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2117 2118 if (!is_leaf) { 2119 // Non-leaves can block and take safepoints: 2120 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2121 } 2122 // Non-leaves can throw exceptions: 2123 if (has_io) { 2124 call->set_req(TypeFunc::I_O, i_o()); 2125 } 2126 2127 if (flags & RC_UNCOMMON) { 2128 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2129 // (An "if" probability corresponds roughly to an unconditional count. 2130 // Sort of.) 2131 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2132 } 2133 2134 Node* c = _gvn.transform(call); 2135 assert(c == call, "cannot disappear"); 2136 2137 if (wide_out) { 2138 // Slow path call has full side-effects. 2139 set_predefined_output_for_runtime_call(call); 2140 } else { 2141 // Slow path call has few side-effects, and/or sets few values. 2142 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2143 } 2144 2145 if (has_io) { 2146 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O))); 2147 } 2148 return call; 2149 2150 } 2151 2152 //------------------------------merge_memory----------------------------------- 2153 // Merge memory from one path into the current memory state. 2154 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2155 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2156 Node* old_slice = mms.force_memory(); 2157 Node* new_slice = mms.memory2(); 2158 if (old_slice != new_slice) { 2159 PhiNode* phi; 2160 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2161 phi = new_slice->as_Phi(); 2162 #ifdef ASSERT 2163 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2164 old_slice = old_slice->in(new_path); 2165 // Caller is responsible for ensuring that any pre-existing 2166 // phis are already aware of old memory. 2167 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2168 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2169 #endif 2170 mms.set_memory(phi); 2171 } else { 2172 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2173 _gvn.set_type(phi, Type::MEMORY); 2174 phi->set_req(new_path, new_slice); 2175 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2176 } 2177 } 2178 } 2179 } 2180 2181 //------------------------------make_slow_call_ex------------------------------ 2182 // Make the exception handler hookups for the slow call 2183 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2184 if (stopped()) return; 2185 2186 // Make a catch node with just two handlers: fall-through and catch-all 2187 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2188 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) ); 2189 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2190 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2191 2192 { PreserveJVMState pjvms(this); 2193 set_control(excp); 2194 set_i_o(i_o); 2195 2196 if (excp != top()) { 2197 // Create an exception state also. 2198 // Use an exact type if the caller has specified a specific exception. 2199 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2200 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o); 2201 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2202 } 2203 } 2204 2205 // Get the no-exception control from the CatchNode. 2206 set_control(norm); 2207 } 2208 2209 2210 //-------------------------------gen_subtype_check----------------------------- 2211 // Generate a subtyping check. Takes as input the subtype and supertype. 2212 // Returns 2 values: sets the default control() to the true path and returns 2213 // the false path. Only reads invariant memory; sets no (visible) memory. 2214 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2215 // but that's not exposed to the optimizer. This call also doesn't take in an 2216 // Object; if you wish to check an Object you need to load the Object's class 2217 // prior to coming here. 2218 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2219 // Fast check for identical types, perhaps identical constants. 2220 // The types can even be identical non-constants, in cases 2221 // involving Array.newInstance, Object.clone, etc. 2222 if (subklass == superklass) 2223 return top(); // false path is dead; no test needed. 2224 2225 if (_gvn.type(superklass)->singleton()) { 2226 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2227 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2228 2229 // In the common case of an exact superklass, try to fold up the 2230 // test before generating code. You may ask, why not just generate 2231 // the code and then let it fold up? The answer is that the generated 2232 // code will necessarily include null checks, which do not always 2233 // completely fold away. If they are also needless, then they turn 2234 // into a performance loss. Example: 2235 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2236 // Here, the type of 'fa' is often exact, so the store check 2237 // of fa[1]=x will fold up, without testing the nullness of x. 2238 switch (static_subtype_check(superk, subk)) { 2239 case SSC_always_false: 2240 { 2241 Node* always_fail = control(); 2242 set_control(top()); 2243 return always_fail; 2244 } 2245 case SSC_always_true: 2246 return top(); 2247 case SSC_easy_test: 2248 { 2249 // Just do a direct pointer compare and be done. 2250 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) ); 2251 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2252 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2253 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) ); 2254 return _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2255 } 2256 case SSC_full_test: 2257 break; 2258 default: 2259 ShouldNotReachHere(); 2260 } 2261 } 2262 2263 // %%% Possible further optimization: Even if the superklass is not exact, 2264 // if the subklass is the unique subtype of the superklass, the check 2265 // will always succeed. We could leave a dependency behind to ensure this. 2266 2267 // First load the super-klass's check-offset 2268 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() ); 2269 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) ); 2270 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes(); 2271 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2272 2273 // Load from the sub-klass's super-class display list, or a 1-word cache of 2274 // the secondary superclass list, or a failing value with a sentinel offset 2275 // if the super-klass is an interface or exceptionally deep in the Java 2276 // hierarchy and we have to scan the secondary superclass list the hard way. 2277 // Worst-case type is a little odd: NULL is allowed as a result (usually 2278 // klass loads can never produce a NULL). 2279 Node *chk_off_X = ConvI2X(chk_off); 2280 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) ); 2281 // For some types like interfaces the following loadKlass is from a 1-word 2282 // cache which is mutable so can't use immutable memory. Other 2283 // types load from the super-class display table which is immutable. 2284 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2285 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2286 2287 // Compile speed common case: ARE a subtype and we canNOT fail 2288 if( superklass == nkls ) 2289 return top(); // false path is dead; no test needed. 2290 2291 // See if we get an immediate positive hit. Happens roughly 83% of the 2292 // time. Test to see if the value loaded just previously from the subklass 2293 // is exactly the superklass. 2294 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) ); 2295 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) ); 2296 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2297 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) ); 2298 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) ); 2299 2300 // Compile speed common case: Check for being deterministic right now. If 2301 // chk_off is a constant and not equal to cacheoff then we are NOT a 2302 // subklass. In this case we need exactly the 1 test above and we can 2303 // return those results immediately. 2304 if (!might_be_cache) { 2305 Node* not_subtype_ctrl = control(); 2306 set_control(iftrue1); // We need exactly the 1 test above 2307 return not_subtype_ctrl; 2308 } 2309 2310 // Gather the various success & failures here 2311 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4); 2312 record_for_igvn(r_ok_subtype); 2313 RegionNode *r_not_subtype = new (C, 3) RegionNode(3); 2314 record_for_igvn(r_not_subtype); 2315 2316 r_ok_subtype->init_req(1, iftrue1); 2317 2318 // Check for immediate negative hit. Happens roughly 11% of the time (which 2319 // is roughly 63% of the remaining cases). Test to see if the loaded 2320 // check-offset points into the subklass display list or the 1-element 2321 // cache. If it points to the display (and NOT the cache) and the display 2322 // missed then it's not a subtype. 2323 Node *cacheoff = _gvn.intcon(cacheoff_con); 2324 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) ); 2325 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) ); 2326 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2327 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) ); 2328 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) ); 2329 2330 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2331 // No performance impact (too rare) but allows sharing of secondary arrays 2332 // which has some footprint reduction. 2333 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) ); 2334 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) ); 2335 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2336 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) ); 2337 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) ); 2338 2339 // -- Roads not taken here: -- 2340 // We could also have chosen to perform the self-check at the beginning 2341 // of this code sequence, as the assembler does. This would not pay off 2342 // the same way, since the optimizer, unlike the assembler, can perform 2343 // static type analysis to fold away many successful self-checks. 2344 // Non-foldable self checks work better here in second position, because 2345 // the initial primary superclass check subsumes a self-check for most 2346 // types. An exception would be a secondary type like array-of-interface, 2347 // which does not appear in its own primary supertype display. 2348 // Finally, we could have chosen to move the self-check into the 2349 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2350 // dependent manner. But it is worthwhile to have the check here, 2351 // where it can be perhaps be optimized. The cost in code space is 2352 // small (register compare, branch). 2353 2354 // Now do a linear scan of the secondary super-klass array. Again, no real 2355 // performance impact (too rare) but it's gotta be done. 2356 // Since the code is rarely used, there is no penalty for moving it 2357 // out of line, and it can only improve I-cache density. 2358 // The decision to inline or out-of-line this final check is platform 2359 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2360 Node* psc = _gvn.transform( 2361 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2362 2363 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) ); 2364 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) ); 2365 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2366 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) ); 2367 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) ); 2368 2369 // Return false path; set default control to true path. 2370 set_control( _gvn.transform(r_ok_subtype) ); 2371 return _gvn.transform(r_not_subtype); 2372 } 2373 2374 //----------------------------static_subtype_check----------------------------- 2375 // Shortcut important common cases when superklass is exact: 2376 // (0) superklass is java.lang.Object (can occur in reflective code) 2377 // (1) subklass is already limited to a subtype of superklass => always ok 2378 // (2) subklass does not overlap with superklass => always fail 2379 // (3) superklass has NO subtypes and we can check with a simple compare. 2380 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2381 if (StressReflectiveCode) { 2382 return SSC_full_test; // Let caller generate the general case. 2383 } 2384 2385 if (superk == env()->Object_klass()) { 2386 return SSC_always_true; // (0) this test cannot fail 2387 } 2388 2389 ciType* superelem = superk; 2390 if (superelem->is_array_klass()) 2391 superelem = superelem->as_array_klass()->base_element_type(); 2392 2393 if (!subk->is_interface()) { // cannot trust static interface types yet 2394 if (subk->is_subtype_of(superk)) { 2395 return SSC_always_true; // (1) false path dead; no dynamic test needed 2396 } 2397 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2398 !superk->is_subtype_of(subk)) { 2399 return SSC_always_false; 2400 } 2401 } 2402 2403 // If casting to an instance klass, it must have no subtypes 2404 if (superk->is_interface()) { 2405 // Cannot trust interfaces yet. 2406 // %%% S.B. superk->nof_implementors() == 1 2407 } else if (superelem->is_instance_klass()) { 2408 ciInstanceKlass* ik = superelem->as_instance_klass(); 2409 if (!ik->has_subklass() && !ik->is_interface()) { 2410 if (!ik->is_final()) { 2411 // Add a dependency if there is a chance of a later subclass. 2412 C->dependencies()->assert_leaf_type(ik); 2413 } 2414 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2415 } 2416 } else { 2417 // A primitive array type has no subtypes. 2418 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2419 } 2420 2421 return SSC_full_test; 2422 } 2423 2424 // Profile-driven exact type check: 2425 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2426 float prob, 2427 Node* *casted_receiver) { 2428 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2429 Node* recv_klass = load_object_klass(receiver); 2430 Node* want_klass = makecon(tklass); 2431 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) ); 2432 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) ); 2433 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2434 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) )); 2435 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) ); 2436 2437 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2438 assert(recv_xtype->klass_is_exact(), ""); 2439 2440 // Subsume downstream occurrences of receiver with a cast to 2441 // recv_xtype, since now we know what the type will be. 2442 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype); 2443 (*casted_receiver) = _gvn.transform(cast); 2444 // (User must make the replace_in_map call.) 2445 2446 return fail; 2447 } 2448 2449 2450 //-------------------------------gen_instanceof-------------------------------- 2451 // Generate an instance-of idiom. Used by both the instance-of bytecode 2452 // and the reflective instance-of call. 2453 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) { 2454 C->set_has_split_ifs(true); // Has chance for split-if optimization 2455 assert( !stopped(), "dead parse path should be checked in callers" ); 2456 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2457 "must check for not-null not-dead klass in callers"); 2458 2459 // Make the merge point 2460 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2461 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2462 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL); 2463 C->set_has_split_ifs(true); // Has chance for split-if optimization 2464 2465 // Null check; get casted pointer; set region slot 3 2466 Node* null_ctl = top(); 2467 Node* not_null_obj = null_check_oop(subobj, &null_ctl); 2468 2469 // If not_null_obj is dead, only null-path is taken 2470 if (stopped()) { // Doing instance-of on a NULL? 2471 set_control(null_ctl); 2472 return intcon(0); 2473 } 2474 region->init_req(_null_path, null_ctl); 2475 phi ->init_req(_null_path, intcon(0)); // Set null path value 2476 2477 // Load the object's klass 2478 Node* obj_klass = load_object_klass(not_null_obj); 2479 2480 // Generate the subtype check 2481 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2482 2483 // Plug in the success path to the general merge in slot 1. 2484 region->init_req(_obj_path, control()); 2485 phi ->init_req(_obj_path, intcon(1)); 2486 2487 // Plug in the failing path to the general merge in slot 2. 2488 region->init_req(_fail_path, not_subtype_ctrl); 2489 phi ->init_req(_fail_path, intcon(0)); 2490 2491 // Return final merged results 2492 set_control( _gvn.transform(region) ); 2493 record_for_igvn(region); 2494 return _gvn.transform(phi); 2495 } 2496 2497 //-------------------------------gen_checkcast--------------------------------- 2498 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2499 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2500 // uncommon-trap paths work. Adjust stack after this call. 2501 // If failure_control is supplied and not null, it is filled in with 2502 // the control edge for the cast failure. Otherwise, an appropriate 2503 // uncommon trap or exception is thrown. 2504 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2505 Node* *failure_control) { 2506 kill_dead_locals(); // Benefit all the uncommon traps 2507 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2508 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2509 2510 // Fast cutout: Check the case that the cast is vacuously true. 2511 // This detects the common cases where the test will short-circuit 2512 // away completely. We do this before we perform the null check, 2513 // because if the test is going to turn into zero code, we don't 2514 // want a residual null check left around. (Causes a slowdown, 2515 // for example, in some objArray manipulations, such as a[i]=a[j].) 2516 if (tk->singleton()) { 2517 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2518 if (objtp != NULL && objtp->klass() != NULL) { 2519 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2520 case SSC_always_true: 2521 return obj; 2522 case SSC_always_false: 2523 // It needs a null check because a null will *pass* the cast check. 2524 // A non-null value will always produce an exception. 2525 return do_null_assert(obj, T_OBJECT); 2526 } 2527 } 2528 } 2529 2530 ciProfileData* data = NULL; 2531 if (failure_control == NULL) { // use MDO in regular case only 2532 assert(java_bc() == Bytecodes::_aastore || 2533 java_bc() == Bytecodes::_checkcast, 2534 "interpreter profiles type checks only for these BCs"); 2535 data = method()->method_data()->bci_to_data(bci()); 2536 } 2537 2538 // Make the merge point 2539 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2540 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT); 2541 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop); 2542 C->set_has_split_ifs(true); // Has chance for split-if optimization 2543 2544 // Use null-cast information if it is available 2545 bool never_see_null = false; 2546 // If we see an unexpected null at a check-cast we record it and force a 2547 // recompile; the offending check-cast will be compiled to handle NULLs. 2548 // If we see several offending BCIs, then all checkcasts in the 2549 // method will be compiled to handle NULLs. 2550 if (UncommonNullCast // Cutout for this technique 2551 && failure_control == NULL // regular case 2552 && obj != null() // And not the -Xcomp stupid case? 2553 && !too_many_traps(Deoptimization::Reason_null_check)) { 2554 // Finally, check the "null_seen" bit from the interpreter. 2555 if (data == NULL || !data->as_BitData()->null_seen()) { 2556 never_see_null = true; 2557 } 2558 } 2559 2560 // Null check; get casted pointer; set region slot 3 2561 Node* null_ctl = top(); 2562 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null); 2563 2564 // If not_null_obj is dead, only null-path is taken 2565 if (stopped()) { // Doing instance-of on a NULL? 2566 set_control(null_ctl); 2567 return null(); 2568 } 2569 region->init_req(_null_path, null_ctl); 2570 phi ->init_req(_null_path, null()); // Set null path value 2571 2572 Node* cast_obj = NULL; // the casted version of the object 2573 2574 // If the profile has seen exactly one type, narrow to that type. 2575 // (The subsequent subtype check will always fold up.) 2576 if (UseTypeProfile && TypeProfileCasts && data != NULL && 2577 // Counter has never been decremented (due to cast failure). 2578 // ...This is a reasonable thing to expect. It is true of 2579 // all casts inserted by javac to implement generic types. 2580 data->as_CounterData()->count() >= 0 && 2581 !too_many_traps(Deoptimization::Reason_class_check)) { 2582 // (No, this isn't a call, but it's enough like a virtual call 2583 // to use the same ciMethod accessor to get the profile info...) 2584 ciCallProfile profile = method()->call_profile_at_bci(bci()); 2585 if (profile.count() >= 0 && // no cast failures here 2586 profile.has_receiver(0) && 2587 profile.morphism() == 1) { 2588 ciKlass* exact_kls = profile.receiver(0); 2589 int ssc = static_subtype_check(tk->klass(), exact_kls); 2590 if (ssc == SSC_always_true) { 2591 // If we narrow the type to match what the type profile sees, 2592 // we can then remove the rest of the cast. 2593 // This is a win, even if the exact_kls is very specific, 2594 // because downstream operations, such as method calls, 2595 // will often benefit from the sharper type. 2596 Node* exact_obj = not_null_obj; // will get updated in place... 2597 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2598 &exact_obj); 2599 { PreserveJVMState pjvms(this); 2600 set_control(slow_ctl); 2601 uncommon_trap(Deoptimization::Reason_class_check, 2602 Deoptimization::Action_maybe_recompile); 2603 } 2604 if (failure_control != NULL) // failure is now impossible 2605 (*failure_control) = top(); 2606 replace_in_map(not_null_obj, exact_obj); 2607 // adjust the type of the phi to the exact klass: 2608 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR)); 2609 cast_obj = exact_obj; 2610 } 2611 // assert(cast_obj != NULL)... except maybe the profile lied to us. 2612 } 2613 } 2614 2615 if (cast_obj == NULL) { 2616 // Load the object's klass 2617 Node* obj_klass = load_object_klass(not_null_obj); 2618 2619 // Generate the subtype check 2620 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 2621 2622 // Plug in success path into the merge 2623 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(), 2624 not_null_obj, toop)); 2625 // Failure path ends in uncommon trap (or may be dead - failure impossible) 2626 if (failure_control == NULL) { 2627 if (not_subtype_ctrl != top()) { // If failure is possible 2628 PreserveJVMState pjvms(this); 2629 set_control(not_subtype_ctrl); 2630 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 2631 } 2632 } else { 2633 (*failure_control) = not_subtype_ctrl; 2634 } 2635 } 2636 2637 region->init_req(_obj_path, control()); 2638 phi ->init_req(_obj_path, cast_obj); 2639 2640 // A merge of NULL or Casted-NotNull obj 2641 Node* res = _gvn.transform(phi); 2642 2643 // Note I do NOT always 'replace_in_map(obj,result)' here. 2644 // if( tk->klass()->can_be_primary_super() ) 2645 // This means that if I successfully store an Object into an array-of-String 2646 // I 'forget' that the Object is really now known to be a String. I have to 2647 // do this because we don't have true union types for interfaces - if I store 2648 // a Baz into an array-of-Interface and then tell the optimizer it's an 2649 // Interface, I forget that it's also a Baz and cannot do Baz-like field 2650 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 2651 // replace_in_map( obj, res ); 2652 2653 // Return final merged results 2654 set_control( _gvn.transform(region) ); 2655 record_for_igvn(region); 2656 return res; 2657 } 2658 2659 //------------------------------next_monitor----------------------------------- 2660 // What number should be given to the next monitor? 2661 int GraphKit::next_monitor() { 2662 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 2663 int next = current + C->sync_stack_slots(); 2664 // Keep the toplevel high water mark current: 2665 if (C->fixed_slots() < next) C->set_fixed_slots(next); 2666 return current; 2667 } 2668 2669 //------------------------------insert_mem_bar--------------------------------- 2670 // Memory barrier to avoid floating things around 2671 // The membar serves as a pinch point between both control and all memory slices. 2672 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 2673 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 2674 mb->init_req(TypeFunc::Control, control()); 2675 mb->init_req(TypeFunc::Memory, reset_memory()); 2676 Node* membar = _gvn.transform(mb); 2677 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) )); 2678 set_all_memory_call(membar); 2679 return membar; 2680 } 2681 2682 //-------------------------insert_mem_bar_volatile---------------------------- 2683 // Memory barrier to avoid floating things around 2684 // The membar serves as a pinch point between both control and memory(alias_idx). 2685 // If you want to make a pinch point on all memory slices, do not use this 2686 // function (even with AliasIdxBot); use insert_mem_bar() instead. 2687 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 2688 // When Parse::do_put_xxx updates a volatile field, it appends a series 2689 // of MemBarVolatile nodes, one for *each* volatile field alias category. 2690 // The first membar is on the same memory slice as the field store opcode. 2691 // This forces the membar to follow the store. (Bug 6500685 broke this.) 2692 // All the other membars (for other volatile slices, including AliasIdxBot, 2693 // which stands for all unknown volatile slices) are control-dependent 2694 // on the first membar. This prevents later volatile loads or stores 2695 // from sliding up past the just-emitted store. 2696 2697 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 2698 mb->set_req(TypeFunc::Control,control()); 2699 if (alias_idx == Compile::AliasIdxBot) { 2700 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 2701 } else { 2702 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 2703 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 2704 } 2705 Node* membar = _gvn.transform(mb); 2706 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control))); 2707 if (alias_idx == Compile::AliasIdxBot) { 2708 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory))); 2709 } else { 2710 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx); 2711 } 2712 return membar; 2713 } 2714 2715 //------------------------------shared_lock------------------------------------ 2716 // Emit locking code. 2717 FastLockNode* GraphKit::shared_lock(Node* obj) { 2718 // bci is either a monitorenter bc or InvocationEntryBci 2719 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2720 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2721 2722 if( !GenerateSynchronizationCode ) 2723 return NULL; // Not locking things? 2724 if (stopped()) // Dead monitor? 2725 return NULL; 2726 2727 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 2728 2729 // Box the stack location 2730 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor())); 2731 Node* mem = reset_memory(); 2732 2733 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock(); 2734 if (PrintPreciseBiasedLockingStatistics) { 2735 // Create the counters for this fast lock. 2736 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 2737 } 2738 // Add monitor to debug info for the slow path. If we block inside the 2739 // slow path and de-opt, we need the monitor hanging around 2740 map()->push_monitor( flock ); 2741 2742 const TypeFunc *tf = LockNode::lock_type(); 2743 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf); 2744 2745 lock->init_req( TypeFunc::Control, control() ); 2746 lock->init_req( TypeFunc::Memory , mem ); 2747 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2748 lock->init_req( TypeFunc::FramePtr, frameptr() ); 2749 lock->init_req( TypeFunc::ReturnAdr, top() ); 2750 2751 lock->init_req(TypeFunc::Parms + 0, obj); 2752 lock->init_req(TypeFunc::Parms + 1, box); 2753 lock->init_req(TypeFunc::Parms + 2, flock); 2754 add_safepoint_edges(lock); 2755 2756 lock = _gvn.transform( lock )->as_Lock(); 2757 2758 // lock has no side-effects, sets few values 2759 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 2760 2761 insert_mem_bar(Op_MemBarAcquire); 2762 2763 // Add this to the worklist so that the lock can be eliminated 2764 record_for_igvn(lock); 2765 2766 #ifndef PRODUCT 2767 if (PrintLockStatistics) { 2768 // Update the counter for this lock. Don't bother using an atomic 2769 // operation since we don't require absolute accuracy. 2770 lock->create_lock_counter(map()->jvms()); 2771 int adr_type = Compile::AliasIdxRaw; 2772 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr())); 2773 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type); 2774 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1))); 2775 store_to_memory(control(), counter_addr, incr, T_INT, adr_type); 2776 } 2777 #endif 2778 2779 return flock; 2780 } 2781 2782 2783 //------------------------------shared_unlock---------------------------------- 2784 // Emit unlocking code. 2785 void GraphKit::shared_unlock(Node* box, Node* obj) { 2786 // bci is either a monitorenter bc or InvocationEntryBci 2787 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 2788 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 2789 2790 if( !GenerateSynchronizationCode ) 2791 return; 2792 if (stopped()) { // Dead monitor? 2793 map()->pop_monitor(); // Kill monitor from debug info 2794 return; 2795 } 2796 2797 // Memory barrier to avoid floating things down past the locked region 2798 insert_mem_bar(Op_MemBarRelease); 2799 2800 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 2801 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf); 2802 uint raw_idx = Compile::AliasIdxRaw; 2803 unlock->init_req( TypeFunc::Control, control() ); 2804 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 2805 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 2806 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 2807 unlock->init_req( TypeFunc::ReturnAdr, top() ); 2808 2809 unlock->init_req(TypeFunc::Parms + 0, obj); 2810 unlock->init_req(TypeFunc::Parms + 1, box); 2811 unlock = _gvn.transform(unlock)->as_Unlock(); 2812 2813 Node* mem = reset_memory(); 2814 2815 // unlock has no side-effects, sets few values 2816 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 2817 2818 // Kill monitor from debug info 2819 map()->pop_monitor( ); 2820 } 2821 2822 //-------------------------------get_layout_helper----------------------------- 2823 // If the given klass is a constant or known to be an array, 2824 // fetch the constant layout helper value into constant_value 2825 // and return (Node*)NULL. Otherwise, load the non-constant 2826 // layout helper value, and return the node which represents it. 2827 // This two-faced routine is useful because allocation sites 2828 // almost always feature constant types. 2829 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 2830 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 2831 if (!StressReflectiveCode && inst_klass != NULL) { 2832 ciKlass* klass = inst_klass->klass(); 2833 bool xklass = inst_klass->klass_is_exact(); 2834 if (xklass || klass->is_array_klass()) { 2835 jint lhelper = klass->layout_helper(); 2836 if (lhelper != Klass::_lh_neutral_value) { 2837 constant_value = lhelper; 2838 return (Node*) NULL; 2839 } 2840 } 2841 } 2842 constant_value = Klass::_lh_neutral_value; // put in a known value 2843 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc)); 2844 return make_load(NULL, lhp, TypeInt::INT, T_INT); 2845 } 2846 2847 // We just put in an allocate/initialize with a big raw-memory effect. 2848 // Hook selected additional alias categories on the initialization. 2849 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 2850 MergeMemNode* init_in_merge, 2851 Node* init_out_raw) { 2852 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 2853 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 2854 2855 Node* prevmem = kit.memory(alias_idx); 2856 init_in_merge->set_memory_at(alias_idx, prevmem); 2857 kit.set_memory(init_out_raw, alias_idx); 2858 } 2859 2860 //---------------------------set_output_for_allocation------------------------- 2861 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 2862 const TypeOopPtr* oop_type, 2863 bool raw_mem_only) { 2864 int rawidx = Compile::AliasIdxRaw; 2865 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 2866 add_safepoint_edges(alloc); 2867 Node* allocx = _gvn.transform(alloc); 2868 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) ); 2869 // create memory projection for i_o 2870 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 2871 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true); 2872 2873 // create a memory projection as for the normal control path 2874 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory)); 2875 set_memory(malloc, rawidx); 2876 2877 // a normal slow-call doesn't change i_o, but an allocation does 2878 // we create a separate i_o projection for the normal control path 2879 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) ); 2880 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) ); 2881 2882 // put in an initialization barrier 2883 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 2884 rawoop)->as_Initialize(); 2885 assert(alloc->initialization() == init, "2-way macro link must work"); 2886 assert(init ->allocation() == alloc, "2-way macro link must work"); 2887 if (ReduceFieldZeroing && !raw_mem_only) { 2888 // Extract memory strands which may participate in the new object's 2889 // initialization, and source them from the new InitializeNode. 2890 // This will allow us to observe initializations when they occur, 2891 // and link them properly (as a group) to the InitializeNode. 2892 assert(init->in(InitializeNode::Memory) == malloc, ""); 2893 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 2894 init->set_req(InitializeNode::Memory, minit_in); 2895 record_for_igvn(minit_in); // fold it up later, if possible 2896 Node* minit_out = memory(rawidx); 2897 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 2898 if (oop_type->isa_aryptr()) { 2899 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 2900 int elemidx = C->get_alias_index(telemref); 2901 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 2902 } else if (oop_type->isa_instptr()) { 2903 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 2904 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 2905 ciField* field = ik->nonstatic_field_at(i); 2906 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 2907 continue; // do not bother to track really large numbers of fields 2908 // Find (or create) the alias category for this field: 2909 int fieldidx = C->alias_type(field)->index(); 2910 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 2911 } 2912 } 2913 } 2914 2915 // Cast raw oop to the real thing... 2916 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type); 2917 javaoop = _gvn.transform(javaoop); 2918 C->set_recent_alloc(control(), javaoop); 2919 assert(just_allocated_object(control()) == javaoop, "just allocated"); 2920 2921 #ifdef ASSERT 2922 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 2923 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 2924 "Ideal_allocation works"); 2925 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 2926 "Ideal_allocation works"); 2927 if (alloc->is_AllocateArray()) { 2928 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 2929 "Ideal_allocation works"); 2930 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 2931 "Ideal_allocation works"); 2932 } else { 2933 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 2934 } 2935 } 2936 #endif //ASSERT 2937 2938 return javaoop; 2939 } 2940 2941 //---------------------------new_instance-------------------------------------- 2942 // This routine takes a klass_node which may be constant (for a static type) 2943 // or may be non-constant (for reflective code). It will work equally well 2944 // for either, and the graph will fold nicely if the optimizer later reduces 2945 // the type to a constant. 2946 // The optional arguments are for specialized use by intrinsics: 2947 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 2948 // - If 'raw_mem_only', do not cast the result to an oop. 2949 // - If 'return_size_val', report the the total object size to the caller. 2950 Node* GraphKit::new_instance(Node* klass_node, 2951 Node* extra_slow_test, 2952 bool raw_mem_only, // affect only raw memory 2953 Node* *return_size_val) { 2954 // Compute size in doublewords 2955 // The size is always an integral number of doublewords, represented 2956 // as a positive bytewise size stored in the klass's layout_helper. 2957 // The layout_helper also encodes (in a low bit) the need for a slow path. 2958 jint layout_con = Klass::_lh_neutral_value; 2959 Node* layout_val = get_layout_helper(klass_node, layout_con); 2960 int layout_is_con = (layout_val == NULL); 2961 2962 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 2963 // Generate the initial go-slow test. It's either ALWAYS (return a 2964 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 2965 // case) a computed value derived from the layout_helper. 2966 Node* initial_slow_test = NULL; 2967 if (layout_is_con) { 2968 assert(!StressReflectiveCode, "stress mode does not use these paths"); 2969 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 2970 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 2971 2972 } else { // reflective case 2973 // This reflective path is used by Unsafe.allocateInstance. 2974 // (It may be stress-tested by specifying StressReflectiveCode.) 2975 // Basically, we want to get into the VM is there's an illegal argument. 2976 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 2977 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) ); 2978 if (extra_slow_test != intcon(0)) { 2979 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) ); 2980 } 2981 // (Macro-expander will further convert this to a Bool, if necessary.) 2982 } 2983 2984 // Find the size in bytes. This is easy; it's the layout_helper. 2985 // The size value must be valid even if the slow path is taken. 2986 Node* size = NULL; 2987 if (layout_is_con) { 2988 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 2989 } else { // reflective case 2990 // This reflective path is used by clone and Unsafe.allocateInstance. 2991 size = ConvI2X(layout_val); 2992 2993 // Clear the low bits to extract layout_helper_size_in_bytes: 2994 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 2995 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 2996 size = _gvn.transform( new (C, 3) AndXNode(size, mask) ); 2997 } 2998 if (return_size_val != NULL) { 2999 (*return_size_val) = size; 3000 } 3001 3002 // This is a precise notnull oop of the klass. 3003 // (Actually, it need not be precise if this is a reflective allocation.) 3004 // It's what we cast the result to. 3005 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3006 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3007 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3008 3009 // Now generate allocation code 3010 3011 // The entire memory state is needed for slow path of the allocation 3012 // since GC and deoptimization can happened. 3013 Node *mem = reset_memory(); 3014 set_all_memory(mem); // Create new memory state 3015 3016 AllocateNode* alloc 3017 = new (C, AllocateNode::ParmLimit) 3018 AllocateNode(C, AllocateNode::alloc_type(), 3019 control(), mem, i_o(), 3020 size, klass_node, 3021 initial_slow_test); 3022 3023 return set_output_for_allocation(alloc, oop_type, raw_mem_only); 3024 } 3025 3026 //-------------------------------new_array------------------------------------- 3027 // helper for both newarray and anewarray 3028 // The 'length' parameter is (obviously) the length of the array. 3029 // See comments on new_instance for the meaning of the other arguments. 3030 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3031 Node* length, // number of array elements 3032 int nargs, // number of arguments to push back for uncommon trap 3033 bool raw_mem_only, // affect only raw memory 3034 Node* *return_size_val) { 3035 jint layout_con = Klass::_lh_neutral_value; 3036 Node* layout_val = get_layout_helper(klass_node, layout_con); 3037 int layout_is_con = (layout_val == NULL); 3038 3039 if (!layout_is_con && !StressReflectiveCode && 3040 !too_many_traps(Deoptimization::Reason_class_check)) { 3041 // This is a reflective array creation site. 3042 // Optimistically assume that it is a subtype of Object[], 3043 // so that we can fold up all the address arithmetic. 3044 layout_con = Klass::array_layout_helper(T_OBJECT); 3045 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) ); 3046 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) ); 3047 { BuildCutout unless(this, bol_lh, PROB_MAX); 3048 _sp += nargs; 3049 uncommon_trap(Deoptimization::Reason_class_check, 3050 Deoptimization::Action_maybe_recompile); 3051 } 3052 layout_val = NULL; 3053 layout_is_con = true; 3054 } 3055 3056 // Generate the initial go-slow test. Make sure we do not overflow 3057 // if length is huge (near 2Gig) or negative! We do not need 3058 // exact double-words here, just a close approximation of needed 3059 // double-words. We can't add any offset or rounding bits, lest we 3060 // take a size -1 of bytes and make it positive. Use an unsigned 3061 // compare, so negative sizes look hugely positive. 3062 int fast_size_limit = FastAllocateSizeLimit; 3063 if (layout_is_con) { 3064 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3065 // Increase the size limit if we have exact knowledge of array type. 3066 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3067 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3068 } 3069 3070 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) ); 3071 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3072 if (initial_slow_test->is_Bool()) { 3073 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3074 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3075 } 3076 3077 // --- Size Computation --- 3078 // array_size = round_to_heap(array_header + (length << elem_shift)); 3079 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3080 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3081 // The rounding mask is strength-reduced, if possible. 3082 int round_mask = MinObjAlignmentInBytes - 1; 3083 Node* header_size = NULL; 3084 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3085 // (T_BYTE has the weakest alignment and size restrictions...) 3086 if (layout_is_con) { 3087 int hsize = Klass::layout_helper_header_size(layout_con); 3088 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3089 BasicType etype = Klass::layout_helper_element_type(layout_con); 3090 if ((round_mask & ~right_n_bits(eshift)) == 0) 3091 round_mask = 0; // strength-reduce it if it goes away completely 3092 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3093 assert(header_size_min <= hsize, "generic minimum is smallest"); 3094 header_size_min = hsize; 3095 header_size = intcon(hsize + round_mask); 3096 } else { 3097 Node* hss = intcon(Klass::_lh_header_size_shift); 3098 Node* hsm = intcon(Klass::_lh_header_size_mask); 3099 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) ); 3100 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) ); 3101 Node* mask = intcon(round_mask); 3102 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) ); 3103 } 3104 3105 Node* elem_shift = NULL; 3106 if (layout_is_con) { 3107 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3108 if (eshift != 0) 3109 elem_shift = intcon(eshift); 3110 } else { 3111 // There is no need to mask or shift this value. 3112 // The semantics of LShiftINode include an implicit mask to 0x1F. 3113 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3114 elem_shift = layout_val; 3115 } 3116 3117 // Transition to native address size for all offset calculations: 3118 Node* lengthx = ConvI2X(length); 3119 Node* headerx = ConvI2X(header_size); 3120 #ifdef _LP64 3121 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3122 if (tllen != NULL && tllen->_lo < 0) { 3123 // Add a manual constraint to a positive range. Cf. array_element_address. 3124 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3125 if (size_max > tllen->_hi) size_max = tllen->_hi; 3126 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3127 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon)); 3128 } 3129 } 3130 #endif 3131 3132 // Combine header size (plus rounding) and body size. Then round down. 3133 // This computation cannot overflow, because it is used only in two 3134 // places, one where the length is sharply limited, and the other 3135 // after a successful allocation. 3136 Node* abody = lengthx; 3137 if (elem_shift != NULL) 3138 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) ); 3139 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) ); 3140 if (round_mask != 0) { 3141 Node* mask = MakeConX(~round_mask); 3142 size = _gvn.transform( new(C, 3) AndXNode(size, mask) ); 3143 } 3144 // else if round_mask == 0, the size computation is self-rounding 3145 3146 if (return_size_val != NULL) { 3147 // This is the size 3148 (*return_size_val) = size; 3149 } 3150 3151 // Now generate allocation code 3152 3153 // The entire memory state is needed for slow path of the allocation 3154 // since GC and deoptimization can happened. 3155 Node *mem = reset_memory(); 3156 set_all_memory(mem); // Create new memory state 3157 3158 // Create the AllocateArrayNode and its result projections 3159 AllocateArrayNode* alloc 3160 = new (C, AllocateArrayNode::ParmLimit) 3161 AllocateArrayNode(C, AllocateArrayNode::alloc_type(), 3162 control(), mem, i_o(), 3163 size, klass_node, 3164 initial_slow_test, 3165 length); 3166 3167 // Cast to correct type. Note that the klass_node may be constant or not, 3168 // and in the latter case the actual array type will be inexact also. 3169 // (This happens via a non-constant argument to inline_native_newArray.) 3170 // In any case, the value of klass_node provides the desired array type. 3171 const TypeInt* length_type = _gvn.find_int_type(length); 3172 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3173 if (ary_type->isa_aryptr() && length_type != NULL) { 3174 // Try to get a better type than POS for the size 3175 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3176 } 3177 3178 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only); 3179 3180 // Cast length on remaining path to be as narrow as possible 3181 if (map()->find_edge(length) >= 0) { 3182 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3183 if (ccast != length) { 3184 _gvn.set_type_bottom(ccast); 3185 record_for_igvn(ccast); 3186 replace_in_map(length, ccast); 3187 } 3188 } 3189 3190 return javaoop; 3191 } 3192 3193 // The following "Ideal_foo" functions are placed here because they recognize 3194 // the graph shapes created by the functions immediately above. 3195 3196 //---------------------------Ideal_allocation---------------------------------- 3197 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3198 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3199 if (ptr == NULL) { // reduce dumb test in callers 3200 return NULL; 3201 } 3202 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast 3203 ptr = ptr->in(1); 3204 if (ptr == NULL) return NULL; 3205 } 3206 if (ptr->is_Proj()) { 3207 Node* allo = ptr->in(0); 3208 if (allo != NULL && allo->is_Allocate()) { 3209 return allo->as_Allocate(); 3210 } 3211 } 3212 // Report failure to match. 3213 return NULL; 3214 } 3215 3216 // Fancy version which also strips off an offset (and reports it to caller). 3217 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3218 intptr_t& offset) { 3219 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3220 if (base == NULL) return NULL; 3221 return Ideal_allocation(base, phase); 3222 } 3223 3224 // Trace Initialize <- Proj[Parm] <- Allocate 3225 AllocateNode* InitializeNode::allocation() { 3226 Node* rawoop = in(InitializeNode::RawAddress); 3227 if (rawoop->is_Proj()) { 3228 Node* alloc = rawoop->in(0); 3229 if (alloc->is_Allocate()) { 3230 return alloc->as_Allocate(); 3231 } 3232 } 3233 return NULL; 3234 } 3235 3236 // Trace Allocate -> Proj[Parm] -> Initialize 3237 InitializeNode* AllocateNode::initialization() { 3238 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3239 if (rawoop == NULL) return NULL; 3240 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3241 Node* init = rawoop->fast_out(i); 3242 if (init->is_Initialize()) { 3243 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3244 return init->as_Initialize(); 3245 } 3246 } 3247 return NULL; 3248 } 3249 3250 //----------------------------- store barriers ---------------------------- 3251 #define __ ideal. 3252 3253 void GraphKit::sync_kit(IdealKit& ideal) { 3254 // Final sync IdealKit and graphKit. 3255 __ drain_delay_transform(); 3256 set_all_memory(__ merged_memory()); 3257 set_control(__ ctrl()); 3258 } 3259 3260 // vanilla/CMS post barrier 3261 // Insert a write-barrier store. This is to let generational GC work; we have 3262 // to flag all oop-stores before the next GC point. 3263 void GraphKit::write_barrier_post(Node* oop_store, 3264 Node* obj, 3265 Node* adr, 3266 uint adr_idx, 3267 Node* val, 3268 bool use_precise) { 3269 // No store check needed if we're storing a NULL or an old object 3270 // (latter case is probably a string constant). The concurrent 3271 // mark sweep garbage collector, however, needs to have all nonNull 3272 // oop updates flagged via card-marks. 3273 if (val != NULL && val->is_Con()) { 3274 // must be either an oop or NULL 3275 const Type* t = val->bottom_type(); 3276 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3277 // stores of null never (?) need barriers 3278 return; 3279 ciObject* con = t->is_oopptr()->const_oop(); 3280 if (con != NULL 3281 && con->is_perm() 3282 && Universe::heap()->can_elide_permanent_oop_store_barriers()) 3283 // no store barrier needed, because no old-to-new ref created 3284 return; 3285 } 3286 3287 if (use_ReduceInitialCardMarks() 3288 && obj == just_allocated_object(control())) { 3289 // We can skip marks on a freshly-allocated object in Eden. 3290 // Keep this code in sync with maybe_defer_card_mark() in runtime.cpp. 3291 // That routine informs GC to take appropriate compensating steps 3292 // so as to make this card-mark elision safe. 3293 return; 3294 } 3295 3296 if (!use_precise) { 3297 // All card marks for a (non-array) instance are in one place: 3298 adr = obj; 3299 } 3300 // (Else it's an array (or unknown), and we want more precise card marks.) 3301 assert(adr != NULL, ""); 3302 3303 IdealKit ideal(gvn(), control(), merged_memory(), true); 3304 3305 // Convert the pointer to an int prior to doing math on it 3306 Node* cast = __ CastPX(__ ctrl(), adr); 3307 3308 // Divide by card size 3309 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3310 "Only one we handle so far."); 3311 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3312 3313 // Combine card table base and card offset 3314 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3315 3316 // Get the alias_index for raw card-mark memory 3317 int adr_type = Compile::AliasIdxRaw; 3318 // Smash zero into card 3319 Node* zero = __ ConI(0); 3320 BasicType bt = T_BYTE; 3321 if( !UseConcMarkSweepGC ) { 3322 __ store(__ ctrl(), card_adr, zero, bt, adr_type); 3323 } else { 3324 // Specialized path for CM store barrier 3325 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3326 } 3327 3328 // Final sync IdealKit and GraphKit. 3329 sync_kit(ideal); 3330 } 3331 3332 // G1 pre/post barriers 3333 void GraphKit::g1_write_barrier_pre(Node* obj, 3334 Node* adr, 3335 uint alias_idx, 3336 Node* val, 3337 const TypeOopPtr* val_type, 3338 BasicType bt) { 3339 IdealKit ideal(gvn(), control(), merged_memory(), true); 3340 3341 Node* tls = __ thread(); // ThreadLocalStorage 3342 3343 Node* no_ctrl = NULL; 3344 Node* no_base = __ top(); 3345 Node* zero = __ ConI(0); 3346 3347 float likely = PROB_LIKELY(0.999); 3348 float unlikely = PROB_UNLIKELY(0.999); 3349 3350 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3351 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3352 3353 // Offsets into the thread 3354 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3355 PtrQueue::byte_offset_of_active()); 3356 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3357 PtrQueue::byte_offset_of_index()); 3358 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3359 PtrQueue::byte_offset_of_buf()); 3360 // Now the actual pointers into the thread 3361 3362 // set_control( ctl); 3363 3364 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3365 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3366 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3367 3368 // Now some of the values 3369 3370 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3371 3372 // if (!marking) 3373 __ if_then(marking, BoolTest::ne, zero); { 3374 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3375 3376 const Type* t1 = adr->bottom_type(); 3377 const Type* t2 = val->bottom_type(); 3378 3379 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx); 3380 // if (orig != NULL) 3381 __ if_then(orig, BoolTest::ne, null()); { 3382 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3383 3384 // load original value 3385 // alias_idx correct?? 3386 3387 // is the queue for this thread full? 3388 __ if_then(index, BoolTest::ne, zero, likely); { 3389 3390 // decrement the index 3391 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3392 Node* next_indexX = next_index; 3393 #ifdef _LP64 3394 // We could refine the type for what it's worth 3395 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3396 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3397 #endif 3398 3399 // Now get the buffer location we will log the original value into and store it 3400 Node *log_addr = __ AddP(no_base, buffer, next_indexX); 3401 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw); 3402 3403 // update the index 3404 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3405 3406 } __ else_(); { 3407 3408 // logging buffer is full, call the runtime 3409 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3410 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls); 3411 } __ end_if(); // (!index) 3412 } __ end_if(); // (orig != NULL) 3413 } __ end_if(); // (!marking) 3414 3415 // Final sync IdealKit and GraphKit. 3416 sync_kit(ideal); 3417 } 3418 3419 // 3420 // Update the card table and add card address to the queue 3421 // 3422 void GraphKit::g1_mark_card(IdealKit& ideal, 3423 Node* card_adr, 3424 Node* oop_store, 3425 uint oop_alias_idx, 3426 Node* index, 3427 Node* index_adr, 3428 Node* buffer, 3429 const TypeFunc* tf) { 3430 3431 Node* zero = __ ConI(0); 3432 Node* no_base = __ top(); 3433 BasicType card_bt = T_BYTE; 3434 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3435 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3436 3437 // Now do the queue work 3438 __ if_then(index, BoolTest::ne, zero); { 3439 3440 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t))); 3441 Node* next_indexX = next_index; 3442 #ifdef _LP64 3443 // We could refine the type for what it's worth 3444 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue); 3445 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) ); 3446 #endif // _LP64 3447 Node* log_addr = __ AddP(no_base, buffer, next_indexX); 3448 3449 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw); 3450 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw); 3451 3452 } __ else_(); { 3453 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3454 } __ end_if(); 3455 3456 } 3457 3458 void GraphKit::g1_write_barrier_post(Node* oop_store, 3459 Node* obj, 3460 Node* adr, 3461 uint alias_idx, 3462 Node* val, 3463 BasicType bt, 3464 bool use_precise) { 3465 // If we are writing a NULL then we need no post barrier 3466 3467 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3468 // Must be NULL 3469 const Type* t = val->bottom_type(); 3470 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3471 // No post barrier if writing NULLx 3472 return; 3473 } 3474 3475 if (!use_precise) { 3476 // All card marks for a (non-array) instance are in one place: 3477 adr = obj; 3478 } 3479 // (Else it's an array (or unknown), and we want more precise card marks.) 3480 assert(adr != NULL, ""); 3481 3482 IdealKit ideal(gvn(), control(), merged_memory(), true); 3483 3484 Node* tls = __ thread(); // ThreadLocalStorage 3485 3486 Node* no_ctrl = NULL; 3487 Node* no_base = __ top(); 3488 float likely = PROB_LIKELY(0.999); 3489 float unlikely = PROB_UNLIKELY(0.999); 3490 Node* zero = __ ConI(0); 3491 Node* zeroX = __ ConX(0); 3492 3493 // Get the alias_index for raw card-mark memory 3494 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3495 3496 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3497 3498 // Offsets into the thread 3499 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3500 PtrQueue::byte_offset_of_index()); 3501 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3502 PtrQueue::byte_offset_of_buf()); 3503 3504 // Pointers into the thread 3505 3506 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3507 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3508 3509 // Now some values 3510 3511 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw); 3512 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3513 3514 3515 // Convert the store obj pointer to an int prior to doing math on it 3516 // Must use ctrl to prevent "integerized oop" existing across safepoint 3517 Node* cast = __ CastPX(__ ctrl(), adr); 3518 3519 // Divide pointer by card size 3520 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3521 3522 // Combine card table base and card offset 3523 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 3524 3525 // If we know the value being stored does it cross regions? 3526 3527 if (val != NULL) { 3528 // Does the store cause us to cross regions? 3529 3530 // Should be able to do an unsigned compare of region_size instead of 3531 // and extra shift. Do we have an unsigned compare?? 3532 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 3533 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 3534 3535 // if (xor_res == 0) same region so skip 3536 __ if_then(xor_res, BoolTest::ne, zeroX); { 3537 3538 // No barrier if we are storing a NULL 3539 __ if_then(val, BoolTest::ne, null(), unlikely); { 3540 3541 // Ok must mark the card if not already dirty 3542 3543 // load the original value of the card 3544 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 3545 3546 __ if_then(card_val, BoolTest::ne, zero); { 3547 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3548 } __ end_if(); 3549 } __ end_if(); 3550 } __ end_if(); 3551 } else { 3552 // Object.clone() instrinsic uses this path. 3553 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 3554 } 3555 3556 // Final sync IdealKit and GraphKit. 3557 sync_kit(ideal); 3558 } 3559 #undef __