1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "memory/barrierSet.hpp" 31 #include "memory/cardTableModRefBS.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/intrinsicnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/threadnode.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 //----------------------------GraphKit----------------------------------------- 49 // Main utility constructor. 50 GraphKit::GraphKit(JVMState* jvms) 51 : Phase(Phase::Parser), 52 _env(C->env()), 53 _gvn(*C->initial_gvn()) 54 { 55 _exceptions = jvms->map()->next_exception(); 56 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 57 set_jvms(jvms); 58 } 59 60 // Private constructor for parser. 61 GraphKit::GraphKit() 62 : Phase(Phase::Parser), 63 _env(C->env()), 64 _gvn(*C->initial_gvn()) 65 { 66 _exceptions = NULL; 67 set_map(NULL); 68 debug_only(_sp = -99); 69 debug_only(set_bci(-99)); 70 } 71 72 73 74 //---------------------------clean_stack--------------------------------------- 75 // Clear away rubbish from the stack area of the JVM state. 76 // This destroys any arguments that may be waiting on the stack. 77 void GraphKit::clean_stack(int from_sp) { 78 SafePointNode* map = this->map(); 79 JVMState* jvms = this->jvms(); 80 int stk_size = jvms->stk_size(); 81 int stkoff = jvms->stkoff(); 82 Node* top = this->top(); 83 for (int i = from_sp; i < stk_size; i++) { 84 if (map->in(stkoff + i) != top) { 85 map->set_req(stkoff + i, top); 86 } 87 } 88 } 89 90 91 //--------------------------------sync_jvms----------------------------------- 92 // Make sure our current jvms agrees with our parse state. 93 JVMState* GraphKit::sync_jvms() const { 94 JVMState* jvms = this->jvms(); 95 jvms->set_bci(bci()); // Record the new bci in the JVMState 96 jvms->set_sp(sp()); // Record the new sp in the JVMState 97 assert(jvms_in_sync(), "jvms is now in sync"); 98 return jvms; 99 } 100 101 //--------------------------------sync_jvms_for_reexecute--------------------- 102 // Make sure our current jvms agrees with our parse state. This version 103 // uses the reexecute_sp for reexecuting bytecodes. 104 JVMState* GraphKit::sync_jvms_for_reexecute() { 105 JVMState* jvms = this->jvms(); 106 jvms->set_bci(bci()); // Record the new bci in the JVMState 107 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 108 return jvms; 109 } 110 111 #ifdef ASSERT 112 bool GraphKit::jvms_in_sync() const { 113 Parse* parse = is_Parse(); 114 if (parse == NULL) { 115 if (bci() != jvms()->bci()) return false; 116 if (sp() != (int)jvms()->sp()) return false; 117 return true; 118 } 119 if (jvms()->method() != parse->method()) return false; 120 if (jvms()->bci() != parse->bci()) return false; 121 int jvms_sp = jvms()->sp(); 122 if (jvms_sp != parse->sp()) return false; 123 int jvms_depth = jvms()->depth(); 124 if (jvms_depth != parse->depth()) return false; 125 return true; 126 } 127 128 // Local helper checks for special internal merge points 129 // used to accumulate and merge exception states. 130 // They are marked by the region's in(0) edge being the map itself. 131 // Such merge points must never "escape" into the parser at large, 132 // until they have been handed to gvn.transform. 133 static bool is_hidden_merge(Node* reg) { 134 if (reg == NULL) return false; 135 if (reg->is_Phi()) { 136 reg = reg->in(0); 137 if (reg == NULL) return false; 138 } 139 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 140 } 141 142 void GraphKit::verify_map() const { 143 if (map() == NULL) return; // null map is OK 144 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 145 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 146 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 147 } 148 149 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 150 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 151 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 152 } 153 #endif 154 155 //---------------------------stop_and_kill_map--------------------------------- 156 // Set _map to NULL, signalling a stop to further bytecode execution. 157 // First smash the current map's control to a constant, to mark it dead. 158 void GraphKit::stop_and_kill_map() { 159 SafePointNode* dead_map = stop(); 160 if (dead_map != NULL) { 161 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 162 assert(dead_map->is_killed(), "must be so marked"); 163 } 164 } 165 166 167 //--------------------------------stopped-------------------------------------- 168 // Tell if _map is NULL, or control is top. 169 bool GraphKit::stopped() { 170 if (map() == NULL) return true; 171 else if (control() == top()) return true; 172 else return false; 173 } 174 175 176 //-----------------------------has_ex_handler---------------------------------- 177 // Tell if this method or any caller method has exception handlers. 178 bool GraphKit::has_ex_handler() { 179 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 180 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 181 return true; 182 } 183 } 184 return false; 185 } 186 187 //------------------------------save_ex_oop------------------------------------ 188 // Save an exception without blowing stack contents or other JVM state. 189 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 190 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 191 ex_map->add_req(ex_oop); 192 debug_only(verify_exception_state(ex_map)); 193 } 194 195 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 196 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 197 Node* ex_oop = ex_map->in(ex_map->req()-1); 198 if (clear_it) ex_map->del_req(ex_map->req()-1); 199 return ex_oop; 200 } 201 202 //-----------------------------saved_ex_oop------------------------------------ 203 // Recover a saved exception from its map. 204 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 205 return common_saved_ex_oop(ex_map, false); 206 } 207 208 //--------------------------clear_saved_ex_oop--------------------------------- 209 // Erase a previously saved exception from its map. 210 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 211 return common_saved_ex_oop(ex_map, true); 212 } 213 214 #ifdef ASSERT 215 //---------------------------has_saved_ex_oop---------------------------------- 216 // Erase a previously saved exception from its map. 217 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 218 return ex_map->req() == ex_map->jvms()->endoff()+1; 219 } 220 #endif 221 222 //-------------------------make_exception_state-------------------------------- 223 // Turn the current JVM state into an exception state, appending the ex_oop. 224 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 225 sync_jvms(); 226 SafePointNode* ex_map = stop(); // do not manipulate this map any more 227 set_saved_ex_oop(ex_map, ex_oop); 228 return ex_map; 229 } 230 231 232 //--------------------------add_exception_state-------------------------------- 233 // Add an exception to my list of exceptions. 234 void GraphKit::add_exception_state(SafePointNode* ex_map) { 235 if (ex_map == NULL || ex_map->control() == top()) { 236 return; 237 } 238 #ifdef ASSERT 239 verify_exception_state(ex_map); 240 if (has_exceptions()) { 241 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 242 } 243 #endif 244 245 // If there is already an exception of exactly this type, merge with it. 246 // In particular, null-checks and other low-level exceptions common up here. 247 Node* ex_oop = saved_ex_oop(ex_map); 248 const Type* ex_type = _gvn.type(ex_oop); 249 if (ex_oop == top()) { 250 // No action needed. 251 return; 252 } 253 assert(ex_type->isa_instptr(), "exception must be an instance"); 254 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 255 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 256 // We check sp also because call bytecodes can generate exceptions 257 // both before and after arguments are popped! 258 if (ex_type2 == ex_type 259 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 260 combine_exception_states(ex_map, e2); 261 return; 262 } 263 } 264 265 // No pre-existing exception of the same type. Chain it on the list. 266 push_exception_state(ex_map); 267 } 268 269 //-----------------------add_exception_states_from----------------------------- 270 void GraphKit::add_exception_states_from(JVMState* jvms) { 271 SafePointNode* ex_map = jvms->map()->next_exception(); 272 if (ex_map != NULL) { 273 jvms->map()->set_next_exception(NULL); 274 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 275 next_map = ex_map->next_exception(); 276 ex_map->set_next_exception(NULL); 277 add_exception_state(ex_map); 278 } 279 } 280 } 281 282 //-----------------------transfer_exceptions_into_jvms------------------------- 283 JVMState* GraphKit::transfer_exceptions_into_jvms() { 284 if (map() == NULL) { 285 // We need a JVMS to carry the exceptions, but the map has gone away. 286 // Create a scratch JVMS, cloned from any of the exception states... 287 if (has_exceptions()) { 288 _map = _exceptions; 289 _map = clone_map(); 290 _map->set_next_exception(NULL); 291 clear_saved_ex_oop(_map); 292 debug_only(verify_map()); 293 } else { 294 // ...or created from scratch 295 JVMState* jvms = new (C) JVMState(_method, NULL); 296 jvms->set_bci(_bci); 297 jvms->set_sp(_sp); 298 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); 299 set_jvms(jvms); 300 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 301 set_all_memory(top()); 302 while (map()->req() < jvms->endoff()) map()->add_req(top()); 303 } 304 // (This is a kludge, in case you didn't notice.) 305 set_control(top()); 306 } 307 JVMState* jvms = sync_jvms(); 308 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 309 jvms->map()->set_next_exception(_exceptions); 310 _exceptions = NULL; // done with this set of exceptions 311 return jvms; 312 } 313 314 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 315 assert(is_hidden_merge(dstphi), "must be a special merge node"); 316 assert(is_hidden_merge(srcphi), "must be a special merge node"); 317 uint limit = srcphi->req(); 318 for (uint i = PhiNode::Input; i < limit; i++) { 319 dstphi->add_req(srcphi->in(i)); 320 } 321 } 322 static inline void add_one_req(Node* dstphi, Node* src) { 323 assert(is_hidden_merge(dstphi), "must be a special merge node"); 324 assert(!is_hidden_merge(src), "must not be a special merge node"); 325 dstphi->add_req(src); 326 } 327 328 //-----------------------combine_exception_states------------------------------ 329 // This helper function combines exception states by building phis on a 330 // specially marked state-merging region. These regions and phis are 331 // untransformed, and can build up gradually. The region is marked by 332 // having a control input of its exception map, rather than NULL. Such 333 // regions do not appear except in this function, and in use_exception_state. 334 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 335 if (failing()) return; // dying anyway... 336 JVMState* ex_jvms = ex_map->_jvms; 337 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 338 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 339 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 340 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 341 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 342 assert(ex_map->req() == phi_map->req(), "matching maps"); 343 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 344 Node* hidden_merge_mark = root(); 345 Node* region = phi_map->control(); 346 MergeMemNode* phi_mem = phi_map->merged_memory(); 347 MergeMemNode* ex_mem = ex_map->merged_memory(); 348 if (region->in(0) != hidden_merge_mark) { 349 // The control input is not (yet) a specially-marked region in phi_map. 350 // Make it so, and build some phis. 351 region = new (C) RegionNode(2); 352 _gvn.set_type(region, Type::CONTROL); 353 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 354 region->init_req(1, phi_map->control()); 355 phi_map->set_control(region); 356 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 357 record_for_igvn(io_phi); 358 _gvn.set_type(io_phi, Type::ABIO); 359 phi_map->set_i_o(io_phi); 360 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 361 Node* m = mms.memory(); 362 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 363 record_for_igvn(m_phi); 364 _gvn.set_type(m_phi, Type::MEMORY); 365 mms.set_memory(m_phi); 366 } 367 } 368 369 // Either or both of phi_map and ex_map might already be converted into phis. 370 Node* ex_control = ex_map->control(); 371 // if there is special marking on ex_map also, we add multiple edges from src 372 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 373 // how wide was the destination phi_map, originally? 374 uint orig_width = region->req(); 375 376 if (add_multiple) { 377 add_n_reqs(region, ex_control); 378 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 379 } else { 380 // ex_map has no merges, so we just add single edges everywhere 381 add_one_req(region, ex_control); 382 add_one_req(phi_map->i_o(), ex_map->i_o()); 383 } 384 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 385 if (mms.is_empty()) { 386 // get a copy of the base memory, and patch some inputs into it 387 const TypePtr* adr_type = mms.adr_type(C); 388 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 389 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 390 mms.set_memory(phi); 391 // Prepare to append interesting stuff onto the newly sliced phi: 392 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 393 } 394 // Append stuff from ex_map: 395 if (add_multiple) { 396 add_n_reqs(mms.memory(), mms.memory2()); 397 } else { 398 add_one_req(mms.memory(), mms.memory2()); 399 } 400 } 401 uint limit = ex_map->req(); 402 for (uint i = TypeFunc::Parms; i < limit; i++) { 403 // Skip everything in the JVMS after tos. (The ex_oop follows.) 404 if (i == tos) i = ex_jvms->monoff(); 405 Node* src = ex_map->in(i); 406 Node* dst = phi_map->in(i); 407 if (src != dst) { 408 PhiNode* phi; 409 if (dst->in(0) != region) { 410 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 411 record_for_igvn(phi); 412 _gvn.set_type(phi, phi->type()); 413 phi_map->set_req(i, dst); 414 // Prepare to append interesting stuff onto the new phi: 415 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 416 } else { 417 assert(dst->is_Phi(), "nobody else uses a hidden region"); 418 phi = dst->as_Phi(); 419 } 420 if (add_multiple && src->in(0) == ex_control) { 421 // Both are phis. 422 add_n_reqs(dst, src); 423 } else { 424 while (dst->req() < region->req()) add_one_req(dst, src); 425 } 426 const Type* srctype = _gvn.type(src); 427 if (phi->type() != srctype) { 428 const Type* dsttype = phi->type()->meet_speculative(srctype); 429 if (phi->type() != dsttype) { 430 phi->set_type(dsttype); 431 _gvn.set_type(phi, dsttype); 432 } 433 } 434 } 435 } 436 } 437 438 //--------------------------use_exception_state-------------------------------- 439 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 440 if (failing()) { stop(); return top(); } 441 Node* region = phi_map->control(); 442 Node* hidden_merge_mark = root(); 443 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 444 Node* ex_oop = clear_saved_ex_oop(phi_map); 445 if (region->in(0) == hidden_merge_mark) { 446 // Special marking for internal ex-states. Process the phis now. 447 region->set_req(0, region); // now it's an ordinary region 448 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 449 // Note: Setting the jvms also sets the bci and sp. 450 set_control(_gvn.transform(region)); 451 uint tos = jvms()->stkoff() + sp(); 452 for (uint i = 1; i < tos; i++) { 453 Node* x = phi_map->in(i); 454 if (x->in(0) == region) { 455 assert(x->is_Phi(), "expected a special phi"); 456 phi_map->set_req(i, _gvn.transform(x)); 457 } 458 } 459 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 460 Node* x = mms.memory(); 461 if (x->in(0) == region) { 462 assert(x->is_Phi(), "nobody else uses a hidden region"); 463 mms.set_memory(_gvn.transform(x)); 464 } 465 } 466 if (ex_oop->in(0) == region) { 467 assert(ex_oop->is_Phi(), "expected a special phi"); 468 ex_oop = _gvn.transform(ex_oop); 469 } 470 } else { 471 set_jvms(phi_map->jvms()); 472 } 473 474 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 475 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 476 return ex_oop; 477 } 478 479 //---------------------------------java_bc------------------------------------- 480 Bytecodes::Code GraphKit::java_bc() const { 481 ciMethod* method = this->method(); 482 int bci = this->bci(); 483 if (method != NULL && bci != InvocationEntryBci) 484 return method->java_code_at_bci(bci); 485 else 486 return Bytecodes::_illegal; 487 } 488 489 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 490 bool must_throw) { 491 // if the exception capability is set, then we will generate code 492 // to check the JavaThread.should_post_on_exceptions flag to see 493 // if we actually need to report exception events (for this 494 // thread). If we don't need to report exception events, we will 495 // take the normal fast path provided by add_exception_events. If 496 // exception event reporting is enabled for this thread, we will 497 // take the uncommon_trap in the BuildCutout below. 498 499 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 500 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 501 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 502 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 503 504 // Test the should_post_on_exceptions_flag vs. 0 505 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 506 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 507 508 // Branch to slow_path if should_post_on_exceptions_flag was true 509 { BuildCutout unless(this, tst, PROB_MAX); 510 // Do not try anything fancy if we're notifying the VM on every throw. 511 // Cf. case Bytecodes::_athrow in parse2.cpp. 512 uncommon_trap(reason, Deoptimization::Action_none, 513 (ciKlass*)NULL, (char*)NULL, must_throw); 514 } 515 516 } 517 518 //------------------------------builtin_throw---------------------------------- 519 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 520 bool must_throw = true; 521 522 if (env()->jvmti_can_post_on_exceptions()) { 523 // check if we must post exception events, take uncommon trap if so 524 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 525 // here if should_post_on_exceptions is false 526 // continue on with the normal codegen 527 } 528 529 // If this particular condition has not yet happened at this 530 // bytecode, then use the uncommon trap mechanism, and allow for 531 // a future recompilation if several traps occur here. 532 // If the throw is hot, try to use a more complicated inline mechanism 533 // which keeps execution inside the compiled code. 534 bool treat_throw_as_hot = false; 535 ciMethodData* md = method()->method_data(); 536 537 if (ProfileTraps) { 538 if (too_many_traps(reason)) { 539 treat_throw_as_hot = true; 540 } 541 // (If there is no MDO at all, assume it is early in 542 // execution, and that any deopts are part of the 543 // startup transient, and don't need to be remembered.) 544 545 // Also, if there is a local exception handler, treat all throws 546 // as hot if there has been at least one in this method. 547 if (C->trap_count(reason) != 0 548 && method()->method_data()->trap_count(reason) != 0 549 && has_ex_handler()) { 550 treat_throw_as_hot = true; 551 } 552 } 553 554 // If this throw happens frequently, an uncommon trap might cause 555 // a performance pothole. If there is a local exception handler, 556 // and if this particular bytecode appears to be deoptimizing often, 557 // let us handle the throw inline, with a preconstructed instance. 558 // Note: If the deopt count has blown up, the uncommon trap 559 // runtime is going to flush this nmethod, not matter what. 560 if (treat_throw_as_hot 561 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 562 // If the throw is local, we use a pre-existing instance and 563 // punt on the backtrace. This would lead to a missing backtrace 564 // (a repeat of 4292742) if the backtrace object is ever asked 565 // for its backtrace. 566 // Fixing this remaining case of 4292742 requires some flavor of 567 // escape analysis. Leave that for the future. 568 ciInstance* ex_obj = NULL; 569 switch (reason) { 570 case Deoptimization::Reason_null_check: 571 ex_obj = env()->NullPointerException_instance(); 572 break; 573 case Deoptimization::Reason_div0_check: 574 ex_obj = env()->ArithmeticException_instance(); 575 break; 576 case Deoptimization::Reason_range_check: 577 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 578 break; 579 case Deoptimization::Reason_class_check: 580 if (java_bc() == Bytecodes::_aastore) { 581 ex_obj = env()->ArrayStoreException_instance(); 582 } else { 583 ex_obj = env()->ClassCastException_instance(); 584 } 585 break; 586 } 587 if (failing()) { stop(); return; } // exception allocation might fail 588 if (ex_obj != NULL) { 589 // Cheat with a preallocated exception object. 590 if (C->log() != NULL) 591 C->log()->elem("hot_throw preallocated='1' reason='%s'", 592 Deoptimization::trap_reason_name(reason)); 593 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 594 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 595 596 // Clear the detail message of the preallocated exception object. 597 // Weblogic sometimes mutates the detail message of exceptions 598 // using reflection. 599 int offset = java_lang_Throwable::get_detailMessage_offset(); 600 const TypePtr* adr_typ = ex_con->add_offset(offset); 601 602 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 603 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 604 // Conservatively release stores of object references. 605 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 606 607 add_exception_state(make_exception_state(ex_node)); 608 return; 609 } 610 } 611 612 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 613 // It won't be much cheaper than bailing to the interp., since we'll 614 // have to pass up all the debug-info, and the runtime will have to 615 // create the stack trace. 616 617 // Usual case: Bail to interpreter. 618 // Reserve the right to recompile if we haven't seen anything yet. 619 620 assert(!Deoptimization::reason_is_speculate(reason), "unsupported"); 621 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 622 if (treat_throw_as_hot 623 && (method()->method_data()->trap_recompiled_at(bci(), NULL) 624 || C->too_many_traps(reason))) { 625 // We cannot afford to take more traps here. Suffer in the interpreter. 626 if (C->log() != NULL) 627 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 628 Deoptimization::trap_reason_name(reason), 629 C->trap_count(reason)); 630 action = Deoptimization::Action_none; 631 } 632 633 // "must_throw" prunes the JVM state to include only the stack, if there 634 // are no local exception handlers. This should cut down on register 635 // allocation time and code size, by drastically reducing the number 636 // of in-edges on the call to the uncommon trap. 637 638 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 639 } 640 641 642 //----------------------------PreserveJVMState--------------------------------- 643 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 644 debug_only(kit->verify_map()); 645 _kit = kit; 646 _map = kit->map(); // preserve the map 647 _sp = kit->sp(); 648 kit->set_map(clone_map ? kit->clone_map() : NULL); 649 Compile::current()->inc_preserve_jvm_state(); 650 #ifdef ASSERT 651 _bci = kit->bci(); 652 Parse* parser = kit->is_Parse(); 653 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 654 _block = block; 655 #endif 656 } 657 PreserveJVMState::~PreserveJVMState() { 658 GraphKit* kit = _kit; 659 #ifdef ASSERT 660 assert(kit->bci() == _bci, "bci must not shift"); 661 Parse* parser = kit->is_Parse(); 662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 663 assert(block == _block, "block must not shift"); 664 #endif 665 kit->set_map(_map); 666 kit->set_sp(_sp); 667 Compile::current()->dec_preserve_jvm_state(); 668 } 669 670 671 //-----------------------------BuildCutout------------------------------------- 672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 673 : PreserveJVMState(kit) 674 { 675 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 676 SafePointNode* outer_map = _map; // preserved map is caller's 677 SafePointNode* inner_map = kit->map(); 678 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 679 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 680 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 681 } 682 BuildCutout::~BuildCutout() { 683 GraphKit* kit = _kit; 684 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 685 } 686 687 //---------------------------PreserveReexecuteState---------------------------- 688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 689 assert(!kit->stopped(), "must call stopped() before"); 690 _kit = kit; 691 _sp = kit->sp(); 692 _reexecute = kit->jvms()->_reexecute; 693 } 694 PreserveReexecuteState::~PreserveReexecuteState() { 695 if (_kit->stopped()) return; 696 _kit->jvms()->_reexecute = _reexecute; 697 _kit->set_sp(_sp); 698 } 699 700 //------------------------------clone_map-------------------------------------- 701 // Implementation of PreserveJVMState 702 // 703 // Only clone_map(...) here. If this function is only used in the 704 // PreserveJVMState class we may want to get rid of this extra 705 // function eventually and do it all there. 706 707 SafePointNode* GraphKit::clone_map() { 708 if (map() == NULL) return NULL; 709 710 // Clone the memory edge first 711 Node* mem = MergeMemNode::make(C, map()->memory()); 712 gvn().set_type_bottom(mem); 713 714 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 715 JVMState* jvms = this->jvms(); 716 JVMState* clonejvms = jvms->clone_shallow(C); 717 clonemap->set_memory(mem); 718 clonemap->set_jvms(clonejvms); 719 clonejvms->set_map(clonemap); 720 record_for_igvn(clonemap); 721 gvn().set_type_bottom(clonemap); 722 return clonemap; 723 } 724 725 726 //-----------------------------set_map_clone----------------------------------- 727 void GraphKit::set_map_clone(SafePointNode* m) { 728 _map = m; 729 _map = clone_map(); 730 _map->set_next_exception(NULL); 731 debug_only(verify_map()); 732 } 733 734 735 //----------------------------kill_dead_locals--------------------------------- 736 // Detect any locals which are known to be dead, and force them to top. 737 void GraphKit::kill_dead_locals() { 738 // Consult the liveness information for the locals. If any 739 // of them are unused, then they can be replaced by top(). This 740 // should help register allocation time and cut down on the size 741 // of the deoptimization information. 742 743 // This call is made from many of the bytecode handling 744 // subroutines called from the Big Switch in do_one_bytecode. 745 // Every bytecode which might include a slow path is responsible 746 // for killing its dead locals. The more consistent we 747 // are about killing deads, the fewer useless phis will be 748 // constructed for them at various merge points. 749 750 // bci can be -1 (InvocationEntryBci). We return the entry 751 // liveness for the method. 752 753 if (method() == NULL || method()->code_size() == 0) { 754 // We are building a graph for a call to a native method. 755 // All locals are live. 756 return; 757 } 758 759 ResourceMark rm; 760 761 // Consult the liveness information for the locals. If any 762 // of them are unused, then they can be replaced by top(). This 763 // should help register allocation time and cut down on the size 764 // of the deoptimization information. 765 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 766 767 int len = (int)live_locals.size(); 768 assert(len <= jvms()->loc_size(), "too many live locals"); 769 for (int local = 0; local < len; local++) { 770 if (!live_locals.at(local)) { 771 set_local(local, top()); 772 } 773 } 774 } 775 776 #ifdef ASSERT 777 //-------------------------dead_locals_are_killed------------------------------ 778 // Return true if all dead locals are set to top in the map. 779 // Used to assert "clean" debug info at various points. 780 bool GraphKit::dead_locals_are_killed() { 781 if (method() == NULL || method()->code_size() == 0) { 782 // No locals need to be dead, so all is as it should be. 783 return true; 784 } 785 786 // Make sure somebody called kill_dead_locals upstream. 787 ResourceMark rm; 788 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 789 if (jvms->loc_size() == 0) continue; // no locals to consult 790 SafePointNode* map = jvms->map(); 791 ciMethod* method = jvms->method(); 792 int bci = jvms->bci(); 793 if (jvms == this->jvms()) { 794 bci = this->bci(); // it might not yet be synched 795 } 796 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 797 int len = (int)live_locals.size(); 798 if (!live_locals.is_valid() || len == 0) 799 // This method is trivial, or is poisoned by a breakpoint. 800 return true; 801 assert(len == jvms->loc_size(), "live map consistent with locals map"); 802 for (int local = 0; local < len; local++) { 803 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 804 if (PrintMiscellaneous && (Verbose || WizardMode)) { 805 tty->print_cr("Zombie local %d: ", local); 806 jvms->dump(); 807 } 808 return false; 809 } 810 } 811 } 812 return true; 813 } 814 815 #endif //ASSERT 816 817 // Helper function for enforcing certain bytecodes to reexecute if 818 // deoptimization happens 819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 820 ciMethod* cur_method = jvms->method(); 821 int cur_bci = jvms->bci(); 822 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 823 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 824 return Interpreter::bytecode_should_reexecute(code) || 825 is_anewarray && code == Bytecodes::_multianewarray; 826 // Reexecute _multianewarray bytecode which was replaced with 827 // sequence of [a]newarray. See Parse::do_multianewarray(). 828 // 829 // Note: interpreter should not have it set since this optimization 830 // is limited by dimensions and guarded by flag so in some cases 831 // multianewarray() runtime calls will be generated and 832 // the bytecode should not be reexecutes (stack will not be reset). 833 } else 834 return false; 835 } 836 837 // Helper function for adding JVMState and debug information to node 838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 839 // Add the safepoint edges to the call (or other safepoint). 840 841 // Make sure dead locals are set to top. This 842 // should help register allocation time and cut down on the size 843 // of the deoptimization information. 844 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 845 846 // Walk the inline list to fill in the correct set of JVMState's 847 // Also fill in the associated edges for each JVMState. 848 849 // If the bytecode needs to be reexecuted we need to put 850 // the arguments back on the stack. 851 const bool should_reexecute = jvms()->should_reexecute(); 852 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 853 854 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 855 // undefined if the bci is different. This is normal for Parse but it 856 // should not happen for LibraryCallKit because only one bci is processed. 857 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 858 "in LibraryCallKit the reexecute bit should not change"); 859 860 // If we are guaranteed to throw, we can prune everything but the 861 // input to the current bytecode. 862 bool can_prune_locals = false; 863 uint stack_slots_not_pruned = 0; 864 int inputs = 0, depth = 0; 865 if (must_throw) { 866 assert(method() == youngest_jvms->method(), "sanity"); 867 if (compute_stack_effects(inputs, depth)) { 868 can_prune_locals = true; 869 stack_slots_not_pruned = inputs; 870 } 871 } 872 873 if (env()->should_retain_local_variables()) { 874 // At any safepoint, this method can get breakpointed, which would 875 // then require an immediate deoptimization. 876 can_prune_locals = false; // do not prune locals 877 stack_slots_not_pruned = 0; 878 } 879 880 // do not scribble on the input jvms 881 JVMState* out_jvms = youngest_jvms->clone_deep(C); 882 call->set_jvms(out_jvms); // Start jvms list for call node 883 884 // For a known set of bytecodes, the interpreter should reexecute them if 885 // deoptimization happens. We set the reexecute state for them here 886 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 887 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 888 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 889 } 890 891 // Presize the call: 892 DEBUG_ONLY(uint non_debug_edges = call->req()); 893 call->add_req_batch(top(), youngest_jvms->debug_depth()); 894 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 895 896 // Set up edges so that the call looks like this: 897 // Call [state:] ctl io mem fptr retadr 898 // [parms:] parm0 ... parmN 899 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 900 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 901 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 902 // Note that caller debug info precedes callee debug info. 903 904 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 905 uint debug_ptr = call->req(); 906 907 // Loop over the map input edges associated with jvms, add them 908 // to the call node, & reset all offsets to match call node array. 909 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 910 uint debug_end = debug_ptr; 911 uint debug_start = debug_ptr - in_jvms->debug_size(); 912 debug_ptr = debug_start; // back up the ptr 913 914 uint p = debug_start; // walks forward in [debug_start, debug_end) 915 uint j, k, l; 916 SafePointNode* in_map = in_jvms->map(); 917 out_jvms->set_map(call); 918 919 if (can_prune_locals) { 920 assert(in_jvms->method() == out_jvms->method(), "sanity"); 921 // If the current throw can reach an exception handler in this JVMS, 922 // then we must keep everything live that can reach that handler. 923 // As a quick and dirty approximation, we look for any handlers at all. 924 if (in_jvms->method()->has_exception_handlers()) { 925 can_prune_locals = false; 926 } 927 } 928 929 // Add the Locals 930 k = in_jvms->locoff(); 931 l = in_jvms->loc_size(); 932 out_jvms->set_locoff(p); 933 if (!can_prune_locals) { 934 for (j = 0; j < l; j++) 935 call->set_req(p++, in_map->in(k+j)); 936 } else { 937 p += l; // already set to top above by add_req_batch 938 } 939 940 // Add the Expression Stack 941 k = in_jvms->stkoff(); 942 l = in_jvms->sp(); 943 out_jvms->set_stkoff(p); 944 if (!can_prune_locals) { 945 for (j = 0; j < l; j++) 946 call->set_req(p++, in_map->in(k+j)); 947 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 948 // Divide stack into {S0,...,S1}, where S0 is set to top. 949 uint s1 = stack_slots_not_pruned; 950 stack_slots_not_pruned = 0; // for next iteration 951 if (s1 > l) s1 = l; 952 uint s0 = l - s1; 953 p += s0; // skip the tops preinstalled by add_req_batch 954 for (j = s0; j < l; j++) 955 call->set_req(p++, in_map->in(k+j)); 956 } else { 957 p += l; // already set to top above by add_req_batch 958 } 959 960 // Add the Monitors 961 k = in_jvms->monoff(); 962 l = in_jvms->mon_size(); 963 out_jvms->set_monoff(p); 964 for (j = 0; j < l; j++) 965 call->set_req(p++, in_map->in(k+j)); 966 967 // Copy any scalar object fields. 968 k = in_jvms->scloff(); 969 l = in_jvms->scl_size(); 970 out_jvms->set_scloff(p); 971 for (j = 0; j < l; j++) 972 call->set_req(p++, in_map->in(k+j)); 973 974 // Finish the new jvms. 975 out_jvms->set_endoff(p); 976 977 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 978 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 979 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 980 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 981 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 982 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 983 984 // Update the two tail pointers in parallel. 985 out_jvms = out_jvms->caller(); 986 in_jvms = in_jvms->caller(); 987 } 988 989 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 990 991 // Test the correctness of JVMState::debug_xxx accessors: 992 assert(call->jvms()->debug_start() == non_debug_edges, ""); 993 assert(call->jvms()->debug_end() == call->req(), ""); 994 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 995 } 996 997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 998 Bytecodes::Code code = java_bc(); 999 if (code == Bytecodes::_wide) { 1000 code = method()->java_code_at_bci(bci() + 1); 1001 } 1002 1003 BasicType rtype = T_ILLEGAL; 1004 int rsize = 0; 1005 1006 if (code != Bytecodes::_illegal) { 1007 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1008 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1009 if (rtype < T_CONFLICT) 1010 rsize = type2size[rtype]; 1011 } 1012 1013 switch (code) { 1014 case Bytecodes::_illegal: 1015 return false; 1016 1017 case Bytecodes::_ldc: 1018 case Bytecodes::_ldc_w: 1019 case Bytecodes::_ldc2_w: 1020 inputs = 0; 1021 break; 1022 1023 case Bytecodes::_dup: inputs = 1; break; 1024 case Bytecodes::_dup_x1: inputs = 2; break; 1025 case Bytecodes::_dup_x2: inputs = 3; break; 1026 case Bytecodes::_dup2: inputs = 2; break; 1027 case Bytecodes::_dup2_x1: inputs = 3; break; 1028 case Bytecodes::_dup2_x2: inputs = 4; break; 1029 case Bytecodes::_swap: inputs = 2; break; 1030 case Bytecodes::_arraylength: inputs = 1; break; 1031 1032 case Bytecodes::_getstatic: 1033 case Bytecodes::_putstatic: 1034 case Bytecodes::_getfield: 1035 case Bytecodes::_putfield: 1036 { 1037 bool ignored_will_link; 1038 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1039 int size = field->type()->size(); 1040 bool is_get = (depth >= 0), is_static = (depth & 1); 1041 inputs = (is_static ? 0 : 1); 1042 if (is_get) { 1043 depth = size - inputs; 1044 } else { 1045 inputs += size; // putxxx pops the value from the stack 1046 depth = - inputs; 1047 } 1048 } 1049 break; 1050 1051 case Bytecodes::_invokevirtual: 1052 case Bytecodes::_invokespecial: 1053 case Bytecodes::_invokestatic: 1054 case Bytecodes::_invokedynamic: 1055 case Bytecodes::_invokeinterface: 1056 { 1057 bool ignored_will_link; 1058 ciSignature* declared_signature = NULL; 1059 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1060 assert(declared_signature != NULL, "cannot be null"); 1061 inputs = declared_signature->arg_size_for_bc(code); 1062 int size = declared_signature->return_type()->size(); 1063 depth = size - inputs; 1064 } 1065 break; 1066 1067 case Bytecodes::_multianewarray: 1068 { 1069 ciBytecodeStream iter(method()); 1070 iter.reset_to_bci(bci()); 1071 iter.next(); 1072 inputs = iter.get_dimensions(); 1073 assert(rsize == 1, ""); 1074 depth = rsize - inputs; 1075 } 1076 break; 1077 1078 case Bytecodes::_ireturn: 1079 case Bytecodes::_lreturn: 1080 case Bytecodes::_freturn: 1081 case Bytecodes::_dreturn: 1082 case Bytecodes::_areturn: 1083 assert(rsize = -depth, ""); 1084 inputs = rsize; 1085 break; 1086 1087 case Bytecodes::_jsr: 1088 case Bytecodes::_jsr_w: 1089 inputs = 0; 1090 depth = 1; // S.B. depth=1, not zero 1091 break; 1092 1093 default: 1094 // bytecode produces a typed result 1095 inputs = rsize - depth; 1096 assert(inputs >= 0, ""); 1097 break; 1098 } 1099 1100 #ifdef ASSERT 1101 // spot check 1102 int outputs = depth + inputs; 1103 assert(outputs >= 0, "sanity"); 1104 switch (code) { 1105 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1106 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1107 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1108 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1109 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1110 } 1111 #endif //ASSERT 1112 1113 return true; 1114 } 1115 1116 1117 1118 //------------------------------basic_plus_adr--------------------------------- 1119 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1120 // short-circuit a common case 1121 if (offset == intcon(0)) return ptr; 1122 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1123 } 1124 1125 Node* GraphKit::ConvI2L(Node* offset) { 1126 // short-circuit a common case 1127 jint offset_con = find_int_con(offset, Type::OffsetBot); 1128 if (offset_con != Type::OffsetBot) { 1129 return longcon((jlong) offset_con); 1130 } 1131 return _gvn.transform( new (C) ConvI2LNode(offset)); 1132 } 1133 Node* GraphKit::ConvL2I(Node* offset) { 1134 // short-circuit a common case 1135 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1136 if (offset_con != (jlong)Type::OffsetBot) { 1137 return intcon((int) offset_con); 1138 } 1139 return _gvn.transform( new (C) ConvL2INode(offset)); 1140 } 1141 1142 //-------------------------load_object_klass----------------------------------- 1143 Node* GraphKit::load_object_klass(Node* obj) { 1144 // Special-case a fresh allocation to avoid building nodes: 1145 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1146 if (akls != NULL) return akls; 1147 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1148 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1149 } 1150 1151 //-------------------------load_array_length----------------------------------- 1152 Node* GraphKit::load_array_length(Node* array) { 1153 // Special-case a fresh allocation to avoid building nodes: 1154 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1155 Node *alen; 1156 if (alloc == NULL) { 1157 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1158 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1159 } else { 1160 alen = alloc->Ideal_length(); 1161 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1162 if (ccast != alen) { 1163 alen = _gvn.transform(ccast); 1164 } 1165 } 1166 return alen; 1167 } 1168 1169 //------------------------------do_null_check---------------------------------- 1170 // Helper function to do a NULL pointer check. Returned value is 1171 // the incoming address with NULL casted away. You are allowed to use the 1172 // not-null value only if you are control dependent on the test. 1173 extern int explicit_null_checks_inserted, 1174 explicit_null_checks_elided; 1175 Node* GraphKit::null_check_common(Node* value, BasicType type, 1176 // optional arguments for variations: 1177 bool assert_null, 1178 Node* *null_control) { 1179 assert(!assert_null || null_control == NULL, "not both at once"); 1180 if (stopped()) return top(); 1181 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1182 // For some performance testing, we may wish to suppress null checking. 1183 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1184 return value; 1185 } 1186 explicit_null_checks_inserted++; 1187 1188 // Construct NULL check 1189 Node *chk = NULL; 1190 switch(type) { 1191 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1192 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1193 case T_ARRAY : // fall through 1194 type = T_OBJECT; // simplify further tests 1195 case T_OBJECT : { 1196 const Type *t = _gvn.type( value ); 1197 1198 const TypeOopPtr* tp = t->isa_oopptr(); 1199 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1200 // Only for do_null_check, not any of its siblings: 1201 && !assert_null && null_control == NULL) { 1202 // Usually, any field access or invocation on an unloaded oop type 1203 // will simply fail to link, since the statically linked class is 1204 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1205 // the static class is loaded but the sharper oop type is not. 1206 // Rather than checking for this obscure case in lots of places, 1207 // we simply observe that a null check on an unloaded class 1208 // will always be followed by a nonsense operation, so we 1209 // can just issue the uncommon trap here. 1210 // Our access to the unloaded class will only be correct 1211 // after it has been loaded and initialized, which requires 1212 // a trip through the interpreter. 1213 #ifndef PRODUCT 1214 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1215 #endif 1216 uncommon_trap(Deoptimization::Reason_unloaded, 1217 Deoptimization::Action_reinterpret, 1218 tp->klass(), "!loaded"); 1219 return top(); 1220 } 1221 1222 if (assert_null) { 1223 // See if the type is contained in NULL_PTR. 1224 // If so, then the value is already null. 1225 if (t->higher_equal(TypePtr::NULL_PTR)) { 1226 explicit_null_checks_elided++; 1227 return value; // Elided null assert quickly! 1228 } 1229 } else { 1230 // See if mixing in the NULL pointer changes type. 1231 // If so, then the NULL pointer was not allowed in the original 1232 // type. In other words, "value" was not-null. 1233 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1234 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1235 explicit_null_checks_elided++; 1236 return value; // Elided null check quickly! 1237 } 1238 } 1239 chk = new (C) CmpPNode( value, null() ); 1240 break; 1241 } 1242 1243 default: 1244 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1245 } 1246 assert(chk != NULL, "sanity check"); 1247 chk = _gvn.transform(chk); 1248 1249 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1250 BoolNode *btst = new (C) BoolNode( chk, btest); 1251 Node *tst = _gvn.transform( btst ); 1252 1253 //----------- 1254 // if peephole optimizations occurred, a prior test existed. 1255 // If a prior test existed, maybe it dominates as we can avoid this test. 1256 if (tst != btst && type == T_OBJECT) { 1257 // At this point we want to scan up the CFG to see if we can 1258 // find an identical test (and so avoid this test altogether). 1259 Node *cfg = control(); 1260 int depth = 0; 1261 while( depth < 16 ) { // Limit search depth for speed 1262 if( cfg->Opcode() == Op_IfTrue && 1263 cfg->in(0)->in(1) == tst ) { 1264 // Found prior test. Use "cast_not_null" to construct an identical 1265 // CastPP (and hence hash to) as already exists for the prior test. 1266 // Return that casted value. 1267 if (assert_null) { 1268 replace_in_map(value, null()); 1269 return null(); // do not issue the redundant test 1270 } 1271 Node *oldcontrol = control(); 1272 set_control(cfg); 1273 Node *res = cast_not_null(value); 1274 set_control(oldcontrol); 1275 explicit_null_checks_elided++; 1276 return res; 1277 } 1278 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1279 if (cfg == NULL) break; // Quit at region nodes 1280 depth++; 1281 } 1282 } 1283 1284 //----------- 1285 // Branch to failure if null 1286 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1287 Deoptimization::DeoptReason reason; 1288 if (assert_null) 1289 reason = Deoptimization::Reason_null_assert; 1290 else if (type == T_OBJECT) 1291 reason = Deoptimization::Reason_null_check; 1292 else 1293 reason = Deoptimization::Reason_div0_check; 1294 1295 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1296 // ciMethodData::has_trap_at will return a conservative -1 if any 1297 // must-be-null assertion has failed. This could cause performance 1298 // problems for a method after its first do_null_assert failure. 1299 // Consider using 'Reason_class_check' instead? 1300 1301 // To cause an implicit null check, we set the not-null probability 1302 // to the maximum (PROB_MAX). For an explicit check the probability 1303 // is set to a smaller value. 1304 if (null_control != NULL || too_many_traps(reason)) { 1305 // probability is less likely 1306 ok_prob = PROB_LIKELY_MAG(3); 1307 } else if (!assert_null && 1308 (ImplicitNullCheckThreshold > 0) && 1309 method() != NULL && 1310 (method()->method_data()->trap_count(reason) 1311 >= (uint)ImplicitNullCheckThreshold)) { 1312 ok_prob = PROB_LIKELY_MAG(3); 1313 } 1314 1315 if (null_control != NULL) { 1316 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1317 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1318 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1319 if (null_true == top()) 1320 explicit_null_checks_elided++; 1321 (*null_control) = null_true; 1322 } else { 1323 BuildCutout unless(this, tst, ok_prob); 1324 // Check for optimizer eliding test at parse time 1325 if (stopped()) { 1326 // Failure not possible; do not bother making uncommon trap. 1327 explicit_null_checks_elided++; 1328 } else if (assert_null) { 1329 uncommon_trap(reason, 1330 Deoptimization::Action_make_not_entrant, 1331 NULL, "assert_null"); 1332 } else { 1333 replace_in_map(value, zerocon(type)); 1334 builtin_throw(reason); 1335 } 1336 } 1337 1338 // Must throw exception, fall-thru not possible? 1339 if (stopped()) { 1340 return top(); // No result 1341 } 1342 1343 if (assert_null) { 1344 // Cast obj to null on this path. 1345 replace_in_map(value, zerocon(type)); 1346 return zerocon(type); 1347 } 1348 1349 // Cast obj to not-null on this path, if there is no null_control. 1350 // (If there is a null_control, a non-null value may come back to haunt us.) 1351 if (type == T_OBJECT) { 1352 Node* cast = cast_not_null(value, false); 1353 if (null_control == NULL || (*null_control) == top()) 1354 replace_in_map(value, cast); 1355 value = cast; 1356 } 1357 1358 return value; 1359 } 1360 1361 1362 //------------------------------cast_not_null---------------------------------- 1363 // Cast obj to not-null on this path 1364 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1365 const Type *t = _gvn.type(obj); 1366 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1367 // Object is already not-null? 1368 if( t == t_not_null ) return obj; 1369 1370 Node *cast = new (C) CastPPNode(obj,t_not_null); 1371 cast->init_req(0, control()); 1372 cast = _gvn.transform( cast ); 1373 1374 // Scan for instances of 'obj' in the current JVM mapping. 1375 // These instances are known to be not-null after the test. 1376 if (do_replace_in_map) 1377 replace_in_map(obj, cast); 1378 1379 return cast; // Return casted value 1380 } 1381 1382 1383 //--------------------------replace_in_map------------------------------------- 1384 void GraphKit::replace_in_map(Node* old, Node* neww) { 1385 if (old == neww) { 1386 return; 1387 } 1388 1389 map()->replace_edge(old, neww); 1390 1391 // Note: This operation potentially replaces any edge 1392 // on the map. This includes locals, stack, and monitors 1393 // of the current (innermost) JVM state. 1394 1395 if (!ReplaceInParentMaps) { 1396 return; 1397 } 1398 1399 // PreserveJVMState doesn't do a deep copy so we can't modify 1400 // parents 1401 if (Compile::current()->has_preserve_jvm_state()) { 1402 return; 1403 } 1404 1405 Parse* parser = is_Parse(); 1406 bool progress = true; 1407 Node* ctrl = map()->in(0); 1408 // Follow the chain of parsers and see whether the update can be 1409 // done in the map of callers. We can do the replace for a caller if 1410 // the current control post dominates the control of a caller. 1411 while (parser != NULL && parser->caller() != NULL && progress) { 1412 progress = false; 1413 Node* parent_map = parser->caller()->map(); 1414 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch"); 1415 1416 Node* parent_ctrl = parent_map->in(0); 1417 1418 while (parent_ctrl->is_Region()) { 1419 Node* n = parent_ctrl->as_Region()->is_copy(); 1420 if (n == NULL) { 1421 break; 1422 } 1423 parent_ctrl = n; 1424 } 1425 1426 for (;;) { 1427 if (ctrl == parent_ctrl) { 1428 // update the map of the exits which is the one that will be 1429 // used when compilation resume after inlining 1430 parser->exits().map()->replace_edge(old, neww); 1431 progress = true; 1432 break; 1433 } 1434 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 1435 ctrl = ctrl->in(0)->in(0); 1436 } else if (ctrl->is_Region()) { 1437 Node* n = ctrl->as_Region()->is_copy(); 1438 if (n == NULL) { 1439 break; 1440 } 1441 ctrl = n; 1442 } else { 1443 break; 1444 } 1445 } 1446 1447 parser = parser->parent_parser(); 1448 } 1449 } 1450 1451 1452 //============================================================================= 1453 //--------------------------------memory--------------------------------------- 1454 Node* GraphKit::memory(uint alias_idx) { 1455 MergeMemNode* mem = merged_memory(); 1456 Node* p = mem->memory_at(alias_idx); 1457 _gvn.set_type(p, Type::MEMORY); // must be mapped 1458 return p; 1459 } 1460 1461 //-----------------------------reset_memory------------------------------------ 1462 Node* GraphKit::reset_memory() { 1463 Node* mem = map()->memory(); 1464 // do not use this node for any more parsing! 1465 debug_only( map()->set_memory((Node*)NULL) ); 1466 return _gvn.transform( mem ); 1467 } 1468 1469 //------------------------------set_all_memory--------------------------------- 1470 void GraphKit::set_all_memory(Node* newmem) { 1471 Node* mergemem = MergeMemNode::make(C, newmem); 1472 gvn().set_type_bottom(mergemem); 1473 map()->set_memory(mergemem); 1474 } 1475 1476 //------------------------------set_all_memory_call---------------------------- 1477 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1478 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1479 set_all_memory(newmem); 1480 } 1481 1482 //============================================================================= 1483 // 1484 // parser factory methods for MemNodes 1485 // 1486 // These are layered on top of the factory methods in LoadNode and StoreNode, 1487 // and integrate with the parser's memory state and _gvn engine. 1488 // 1489 1490 // factory methods in "int adr_idx" 1491 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1492 int adr_idx, 1493 MemNode::MemOrd mo, bool require_atomic_access) { 1494 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1495 const TypePtr* adr_type = NULL; // debug-mode-only argument 1496 debug_only(adr_type = C->get_adr_type(adr_idx)); 1497 Node* mem = memory(adr_idx); 1498 Node* ld; 1499 if (require_atomic_access && bt == T_LONG) { 1500 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo); 1501 } else { 1502 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); 1503 } 1504 ld = _gvn.transform(ld); 1505 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1506 // Improve graph before escape analysis and boxing elimination. 1507 record_for_igvn(ld); 1508 } 1509 return ld; 1510 } 1511 1512 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1513 int adr_idx, 1514 MemNode::MemOrd mo, 1515 bool require_atomic_access) { 1516 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1517 const TypePtr* adr_type = NULL; 1518 debug_only(adr_type = C->get_adr_type(adr_idx)); 1519 Node *mem = memory(adr_idx); 1520 Node* st; 1521 if (require_atomic_access && bt == T_LONG) { 1522 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1523 } else { 1524 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1525 } 1526 st = _gvn.transform(st); 1527 set_memory(st, adr_idx); 1528 // Back-to-back stores can only remove intermediate store with DU info 1529 // so push on worklist for optimizer. 1530 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1531 record_for_igvn(st); 1532 1533 return st; 1534 } 1535 1536 1537 void GraphKit::pre_barrier(bool do_load, 1538 Node* ctl, 1539 Node* obj, 1540 Node* adr, 1541 uint adr_idx, 1542 Node* val, 1543 const TypeOopPtr* val_type, 1544 Node* pre_val, 1545 BasicType bt) { 1546 1547 BarrierSet* bs = Universe::heap()->barrier_set(); 1548 set_control(ctl); 1549 switch (bs->kind()) { 1550 case BarrierSet::G1SATBCT: 1551 case BarrierSet::G1SATBCTLogging: 1552 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1553 break; 1554 1555 case BarrierSet::CardTableModRef: 1556 case BarrierSet::CardTableExtension: 1557 case BarrierSet::ModRef: 1558 break; 1559 1560 case BarrierSet::Other: 1561 default : 1562 ShouldNotReachHere(); 1563 1564 } 1565 } 1566 1567 bool GraphKit::can_move_pre_barrier() const { 1568 BarrierSet* bs = Universe::heap()->barrier_set(); 1569 switch (bs->kind()) { 1570 case BarrierSet::G1SATBCT: 1571 case BarrierSet::G1SATBCTLogging: 1572 return true; // Can move it if no safepoint 1573 1574 case BarrierSet::CardTableModRef: 1575 case BarrierSet::CardTableExtension: 1576 case BarrierSet::ModRef: 1577 return true; // There is no pre-barrier 1578 1579 case BarrierSet::Other: 1580 default : 1581 ShouldNotReachHere(); 1582 } 1583 return false; 1584 } 1585 1586 void GraphKit::post_barrier(Node* ctl, 1587 Node* store, 1588 Node* obj, 1589 Node* adr, 1590 uint adr_idx, 1591 Node* val, 1592 BasicType bt, 1593 bool use_precise) { 1594 BarrierSet* bs = Universe::heap()->barrier_set(); 1595 set_control(ctl); 1596 switch (bs->kind()) { 1597 case BarrierSet::G1SATBCT: 1598 case BarrierSet::G1SATBCTLogging: 1599 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1600 break; 1601 1602 case BarrierSet::CardTableModRef: 1603 case BarrierSet::CardTableExtension: 1604 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1605 break; 1606 1607 case BarrierSet::ModRef: 1608 break; 1609 1610 case BarrierSet::Other: 1611 default : 1612 ShouldNotReachHere(); 1613 1614 } 1615 } 1616 1617 Node* GraphKit::store_oop(Node* ctl, 1618 Node* obj, 1619 Node* adr, 1620 const TypePtr* adr_type, 1621 Node* val, 1622 const TypeOopPtr* val_type, 1623 BasicType bt, 1624 bool use_precise, 1625 MemNode::MemOrd mo) { 1626 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1627 // could be delayed during Parse (for example, in adjust_map_after_if()). 1628 // Execute transformation here to avoid barrier generation in such case. 1629 if (_gvn.type(val) == TypePtr::NULL_PTR) 1630 val = _gvn.makecon(TypePtr::NULL_PTR); 1631 1632 set_control(ctl); 1633 if (stopped()) return top(); // Dead path ? 1634 1635 assert(bt == T_OBJECT, "sanity"); 1636 assert(val != NULL, "not dead path"); 1637 uint adr_idx = C->get_alias_index(adr_type); 1638 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1639 1640 pre_barrier(true /* do_load */, 1641 control(), obj, adr, adr_idx, val, val_type, 1642 NULL /* pre_val */, 1643 bt); 1644 1645 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1646 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1647 return store; 1648 } 1649 1650 // Could be an array or object we don't know at compile time (unsafe ref.) 1651 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1652 Node* obj, // containing obj 1653 Node* adr, // actual adress to store val at 1654 const TypePtr* adr_type, 1655 Node* val, 1656 BasicType bt, 1657 MemNode::MemOrd mo) { 1658 Compile::AliasType* at = C->alias_type(adr_type); 1659 const TypeOopPtr* val_type = NULL; 1660 if (adr_type->isa_instptr()) { 1661 if (at->field() != NULL) { 1662 // known field. This code is a copy of the do_put_xxx logic. 1663 ciField* field = at->field(); 1664 if (!field->type()->is_loaded()) { 1665 val_type = TypeInstPtr::BOTTOM; 1666 } else { 1667 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1668 } 1669 } 1670 } else if (adr_type->isa_aryptr()) { 1671 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1672 } 1673 if (val_type == NULL) { 1674 val_type = TypeInstPtr::BOTTOM; 1675 } 1676 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1677 } 1678 1679 1680 //-------------------------array_element_address------------------------- 1681 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1682 const TypeInt* sizetype) { 1683 uint shift = exact_log2(type2aelembytes(elembt)); 1684 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1685 1686 // short-circuit a common case (saves lots of confusing waste motion) 1687 jint idx_con = find_int_con(idx, -1); 1688 if (idx_con >= 0) { 1689 intptr_t offset = header + ((intptr_t)idx_con << shift); 1690 return basic_plus_adr(ary, offset); 1691 } 1692 1693 // must be correct type for alignment purposes 1694 Node* base = basic_plus_adr(ary, header); 1695 #ifdef _LP64 1696 // The scaled index operand to AddP must be a clean 64-bit value. 1697 // Java allows a 32-bit int to be incremented to a negative 1698 // value, which appears in a 64-bit register as a large 1699 // positive number. Using that large positive number as an 1700 // operand in pointer arithmetic has bad consequences. 1701 // On the other hand, 32-bit overflow is rare, and the possibility 1702 // can often be excluded, if we annotate the ConvI2L node with 1703 // a type assertion that its value is known to be a small positive 1704 // number. (The prior range check has ensured this.) 1705 // This assertion is used by ConvI2LNode::Ideal. 1706 int index_max = max_jint - 1; // array size is max_jint, index is one less 1707 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1708 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1709 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) ); 1710 #endif 1711 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1712 return basic_plus_adr(ary, base, scale); 1713 } 1714 1715 //-------------------------load_array_element------------------------- 1716 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1717 const Type* elemtype = arytype->elem(); 1718 BasicType elembt = elemtype->array_element_basic_type(); 1719 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1720 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1721 return ld; 1722 } 1723 1724 //-------------------------set_arguments_for_java_call------------------------- 1725 // Arguments (pre-popped from the stack) are taken from the JVMS. 1726 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1727 // Add the call arguments: 1728 uint nargs = call->method()->arg_size(); 1729 for (uint i = 0; i < nargs; i++) { 1730 Node* arg = argument(i); 1731 call->init_req(i + TypeFunc::Parms, arg); 1732 } 1733 } 1734 1735 //---------------------------set_edges_for_java_call--------------------------- 1736 // Connect a newly created call into the current JVMS. 1737 // A return value node (if any) is returned from set_edges_for_java_call. 1738 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1739 1740 // Add the predefined inputs: 1741 call->init_req( TypeFunc::Control, control() ); 1742 call->init_req( TypeFunc::I_O , i_o() ); 1743 call->init_req( TypeFunc::Memory , reset_memory() ); 1744 call->init_req( TypeFunc::FramePtr, frameptr() ); 1745 call->init_req( TypeFunc::ReturnAdr, top() ); 1746 1747 add_safepoint_edges(call, must_throw); 1748 1749 Node* xcall = _gvn.transform(call); 1750 1751 if (xcall == top()) { 1752 set_control(top()); 1753 return; 1754 } 1755 assert(xcall == call, "call identity is stable"); 1756 1757 // Re-use the current map to produce the result. 1758 1759 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1760 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1761 set_all_memory_call(xcall, separate_io_proj); 1762 1763 //return xcall; // no need, caller already has it 1764 } 1765 1766 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1767 if (stopped()) return top(); // maybe the call folded up? 1768 1769 // Capture the return value, if any. 1770 Node* ret; 1771 if (call->method() == NULL || 1772 call->method()->return_type()->basic_type() == T_VOID) 1773 ret = top(); 1774 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1775 1776 // Note: Since any out-of-line call can produce an exception, 1777 // we always insert an I_O projection from the call into the result. 1778 1779 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1780 1781 if (separate_io_proj) { 1782 // The caller requested separate projections be used by the fall 1783 // through and exceptional paths, so replace the projections for 1784 // the fall through path. 1785 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1786 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1787 } 1788 return ret; 1789 } 1790 1791 //--------------------set_predefined_input_for_runtime_call-------------------- 1792 // Reading and setting the memory state is way conservative here. 1793 // The real problem is that I am not doing real Type analysis on memory, 1794 // so I cannot distinguish card mark stores from other stores. Across a GC 1795 // point the Store Barrier and the card mark memory has to agree. I cannot 1796 // have a card mark store and its barrier split across the GC point from 1797 // either above or below. Here I get that to happen by reading ALL of memory. 1798 // A better answer would be to separate out card marks from other memory. 1799 // For now, return the input memory state, so that it can be reused 1800 // after the call, if this call has restricted memory effects. 1801 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1802 // Set fixed predefined input arguments 1803 Node* memory = reset_memory(); 1804 call->init_req( TypeFunc::Control, control() ); 1805 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1806 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1807 call->init_req( TypeFunc::FramePtr, frameptr() ); 1808 call->init_req( TypeFunc::ReturnAdr, top() ); 1809 return memory; 1810 } 1811 1812 //-------------------set_predefined_output_for_runtime_call-------------------- 1813 // Set control and memory (not i_o) from the call. 1814 // If keep_mem is not NULL, use it for the output state, 1815 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1816 // If hook_mem is NULL, this call produces no memory effects at all. 1817 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1818 // then only that memory slice is taken from the call. 1819 // In the last case, we must put an appropriate memory barrier before 1820 // the call, so as to create the correct anti-dependencies on loads 1821 // preceding the call. 1822 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1823 Node* keep_mem, 1824 const TypePtr* hook_mem) { 1825 // no i/o 1826 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1827 if (keep_mem) { 1828 // First clone the existing memory state 1829 set_all_memory(keep_mem); 1830 if (hook_mem != NULL) { 1831 // Make memory for the call 1832 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1833 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1834 // We also use hook_mem to extract specific effects from arraycopy stubs. 1835 set_memory(mem, hook_mem); 1836 } 1837 // ...else the call has NO memory effects. 1838 1839 // Make sure the call advertises its memory effects precisely. 1840 // This lets us build accurate anti-dependences in gcm.cpp. 1841 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1842 "call node must be constructed correctly"); 1843 } else { 1844 assert(hook_mem == NULL, ""); 1845 // This is not a "slow path" call; all memory comes from the call. 1846 set_all_memory_call(call); 1847 } 1848 } 1849 1850 1851 // Replace the call with the current state of the kit. 1852 void GraphKit::replace_call(CallNode* call, Node* result) { 1853 JVMState* ejvms = NULL; 1854 if (has_exceptions()) { 1855 ejvms = transfer_exceptions_into_jvms(); 1856 } 1857 1858 SafePointNode* final_state = stop(); 1859 1860 // Find all the needed outputs of this call 1861 CallProjections callprojs; 1862 call->extract_projections(&callprojs, true); 1863 1864 Node* init_mem = call->in(TypeFunc::Memory); 1865 Node* final_mem = final_state->in(TypeFunc::Memory); 1866 Node* final_ctl = final_state->in(TypeFunc::Control); 1867 Node* final_io = final_state->in(TypeFunc::I_O); 1868 1869 // Replace all the old call edges with the edges from the inlining result 1870 if (callprojs.fallthrough_catchproj != NULL) { 1871 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1872 } 1873 if (callprojs.fallthrough_memproj != NULL) { 1874 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1875 } 1876 if (callprojs.fallthrough_ioproj != NULL) { 1877 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1878 } 1879 1880 // Replace the result with the new result if it exists and is used 1881 if (callprojs.resproj != NULL && result != NULL) { 1882 C->gvn_replace_by(callprojs.resproj, result); 1883 } 1884 1885 if (ejvms == NULL) { 1886 // No exception edges to simply kill off those paths 1887 if (callprojs.catchall_catchproj != NULL) { 1888 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1889 } 1890 if (callprojs.catchall_memproj != NULL) { 1891 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1892 } 1893 if (callprojs.catchall_ioproj != NULL) { 1894 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1895 } 1896 // Replace the old exception object with top 1897 if (callprojs.exobj != NULL) { 1898 C->gvn_replace_by(callprojs.exobj, C->top()); 1899 } 1900 } else { 1901 GraphKit ekit(ejvms); 1902 1903 // Load my combined exception state into the kit, with all phis transformed: 1904 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1905 1906 Node* ex_oop = ekit.use_exception_state(ex_map); 1907 if (callprojs.catchall_catchproj != NULL) { 1908 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1909 } 1910 if (callprojs.catchall_memproj != NULL) { 1911 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1912 } 1913 if (callprojs.catchall_ioproj != NULL) { 1914 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1915 } 1916 1917 // Replace the old exception object with the newly created one 1918 if (callprojs.exobj != NULL) { 1919 C->gvn_replace_by(callprojs.exobj, ex_oop); 1920 } 1921 } 1922 1923 // Disconnect the call from the graph 1924 call->disconnect_inputs(NULL, C); 1925 C->gvn_replace_by(call, C->top()); 1926 1927 // Clean up any MergeMems that feed other MergeMems since the 1928 // optimizer doesn't like that. 1929 if (final_mem->is_MergeMem()) { 1930 Node_List wl; 1931 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1932 Node* m = i.get(); 1933 if (m->is_MergeMem() && !wl.contains(m)) { 1934 wl.push(m); 1935 } 1936 } 1937 while (wl.size() > 0) { 1938 _gvn.transform(wl.pop()); 1939 } 1940 } 1941 } 1942 1943 1944 //------------------------------increment_counter------------------------------ 1945 // for statistics: increment a VM counter by 1 1946 1947 void GraphKit::increment_counter(address counter_addr) { 1948 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1949 increment_counter(adr1); 1950 } 1951 1952 void GraphKit::increment_counter(Node* counter_addr) { 1953 int adr_type = Compile::AliasIdxRaw; 1954 Node* ctrl = control(); 1955 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1956 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1957 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1958 } 1959 1960 1961 //------------------------------uncommon_trap---------------------------------- 1962 // Bail out to the interpreter in mid-method. Implemented by calling the 1963 // uncommon_trap blob. This helper function inserts a runtime call with the 1964 // right debug info. 1965 void GraphKit::uncommon_trap(int trap_request, 1966 ciKlass* klass, const char* comment, 1967 bool must_throw, 1968 bool keep_exact_action) { 1969 if (failing()) stop(); 1970 if (stopped()) return; // trap reachable? 1971 1972 // Note: If ProfileTraps is true, and if a deopt. actually 1973 // occurs here, the runtime will make sure an MDO exists. There is 1974 // no need to call method()->ensure_method_data() at this point. 1975 1976 // Set the stack pointer to the right value for reexecution: 1977 set_sp(reexecute_sp()); 1978 1979 #ifdef ASSERT 1980 if (!must_throw) { 1981 // Make sure the stack has at least enough depth to execute 1982 // the current bytecode. 1983 int inputs, ignored_depth; 1984 if (compute_stack_effects(inputs, ignored_depth)) { 1985 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1986 Bytecodes::name(java_bc()), sp(), inputs)); 1987 } 1988 } 1989 #endif 1990 1991 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1992 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1993 1994 switch (action) { 1995 case Deoptimization::Action_maybe_recompile: 1996 case Deoptimization::Action_reinterpret: 1997 // Temporary fix for 6529811 to allow virtual calls to be sure they 1998 // get the chance to go from mono->bi->mega 1999 if (!keep_exact_action && 2000 Deoptimization::trap_request_index(trap_request) < 0 && 2001 too_many_recompiles(reason)) { 2002 // This BCI is causing too many recompilations. 2003 action = Deoptimization::Action_none; 2004 trap_request = Deoptimization::make_trap_request(reason, action); 2005 } else { 2006 C->set_trap_can_recompile(true); 2007 } 2008 break; 2009 case Deoptimization::Action_make_not_entrant: 2010 C->set_trap_can_recompile(true); 2011 break; 2012 #ifdef ASSERT 2013 case Deoptimization::Action_none: 2014 case Deoptimization::Action_make_not_compilable: 2015 break; 2016 default: 2017 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2018 break; 2019 #endif 2020 } 2021 2022 if (TraceOptoParse) { 2023 char buf[100]; 2024 tty->print_cr("Uncommon trap %s at bci:%d", 2025 Deoptimization::format_trap_request(buf, sizeof(buf), 2026 trap_request), bci()); 2027 } 2028 2029 CompileLog* log = C->log(); 2030 if (log != NULL) { 2031 int kid = (klass == NULL)? -1: log->identify(klass); 2032 log->begin_elem("uncommon_trap bci='%d'", bci()); 2033 char buf[100]; 2034 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2035 trap_request)); 2036 if (kid >= 0) log->print(" klass='%d'", kid); 2037 if (comment != NULL) log->print(" comment='%s'", comment); 2038 log->end_elem(); 2039 } 2040 2041 // Make sure any guarding test views this path as very unlikely 2042 Node *i0 = control()->in(0); 2043 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2044 IfNode *iff = i0->as_If(); 2045 float f = iff->_prob; // Get prob 2046 if (control()->Opcode() == Op_IfTrue) { 2047 if (f > PROB_UNLIKELY_MAG(4)) 2048 iff->_prob = PROB_MIN; 2049 } else { 2050 if (f < PROB_LIKELY_MAG(4)) 2051 iff->_prob = PROB_MAX; 2052 } 2053 } 2054 2055 // Clear out dead values from the debug info. 2056 kill_dead_locals(); 2057 2058 // Now insert the uncommon trap subroutine call 2059 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2060 const TypePtr* no_memory_effects = NULL; 2061 // Pass the index of the class to be loaded 2062 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2063 (must_throw ? RC_MUST_THROW : 0), 2064 OptoRuntime::uncommon_trap_Type(), 2065 call_addr, "uncommon_trap", no_memory_effects, 2066 intcon(trap_request)); 2067 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2068 "must extract request correctly from the graph"); 2069 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2070 2071 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2072 // The debug info is the only real input to this call. 2073 2074 // Halt-and-catch fire here. The above call should never return! 2075 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2076 _gvn.set_type_bottom(halt); 2077 root()->add_req(halt); 2078 2079 stop_and_kill_map(); 2080 } 2081 2082 2083 //--------------------------just_allocated_object------------------------------ 2084 // Report the object that was just allocated. 2085 // It must be the case that there are no intervening safepoints. 2086 // We use this to determine if an object is so "fresh" that 2087 // it does not require card marks. 2088 Node* GraphKit::just_allocated_object(Node* current_control) { 2089 if (C->recent_alloc_ctl() == current_control) 2090 return C->recent_alloc_obj(); 2091 return NULL; 2092 } 2093 2094 2095 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2096 // (Note: TypeFunc::make has a cache that makes this fast.) 2097 const TypeFunc* tf = TypeFunc::make(dest_method); 2098 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2099 for (int j = 0; j < nargs; j++) { 2100 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2101 if( targ->basic_type() == T_DOUBLE ) { 2102 // If any parameters are doubles, they must be rounded before 2103 // the call, dstore_rounding does gvn.transform 2104 Node *arg = argument(j); 2105 arg = dstore_rounding(arg); 2106 set_argument(j, arg); 2107 } 2108 } 2109 } 2110 2111 /** 2112 * Record profiling data exact_kls for Node n with the type system so 2113 * that it can propagate it (speculation) 2114 * 2115 * @param n node that the type applies to 2116 * @param exact_kls type from profiling 2117 * 2118 * @return node with improved type 2119 */ 2120 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) { 2121 const Type* current_type = _gvn.type(n); 2122 assert(UseTypeSpeculation, "type speculation must be on"); 2123 2124 const TypeOopPtr* speculative = current_type->speculative(); 2125 2126 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2127 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2128 const TypeOopPtr* xtype = tklass->as_instance_type(); 2129 assert(xtype->klass_is_exact(), "Should be exact"); 2130 // record the new speculative type's depth 2131 speculative = xtype->with_inline_depth(jvms()->depth()); 2132 } 2133 2134 if (speculative != current_type->speculative()) { 2135 // Build a type with a speculative type (what we think we know 2136 // about the type but will need a guard when we use it) 2137 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2138 // We're changing the type, we need a new CheckCast node to carry 2139 // the new type. The new type depends on the control: what 2140 // profiling tells us is only valid from here as far as we can 2141 // tell. 2142 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2143 cast = _gvn.transform(cast); 2144 replace_in_map(n, cast); 2145 n = cast; 2146 } 2147 2148 return n; 2149 } 2150 2151 /** 2152 * Record profiling data from receiver profiling at an invoke with the 2153 * type system so that it can propagate it (speculation) 2154 * 2155 * @param n receiver node 2156 * 2157 * @return node with improved type 2158 */ 2159 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2160 if (!UseTypeSpeculation) { 2161 return n; 2162 } 2163 ciKlass* exact_kls = profile_has_unique_klass(); 2164 return record_profile_for_speculation(n, exact_kls); 2165 } 2166 2167 /** 2168 * Record profiling data from argument profiling at an invoke with the 2169 * type system so that it can propagate it (speculation) 2170 * 2171 * @param dest_method target method for the call 2172 * @param bc what invoke bytecode is this? 2173 */ 2174 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2175 if (!UseTypeSpeculation) { 2176 return; 2177 } 2178 const TypeFunc* tf = TypeFunc::make(dest_method); 2179 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2180 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2181 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2182 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2183 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2184 ciKlass* better_type = method()->argument_profiled_type(bci(), i); 2185 if (better_type != NULL) { 2186 record_profile_for_speculation(argument(j), better_type); 2187 } 2188 i++; 2189 } 2190 } 2191 } 2192 2193 /** 2194 * Record profiling data from parameter profiling at an invoke with 2195 * the type system so that it can propagate it (speculation) 2196 */ 2197 void GraphKit::record_profiled_parameters_for_speculation() { 2198 if (!UseTypeSpeculation) { 2199 return; 2200 } 2201 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2202 if (_gvn.type(local(i))->isa_oopptr()) { 2203 ciKlass* better_type = method()->parameter_profiled_type(j); 2204 if (better_type != NULL) { 2205 record_profile_for_speculation(local(i), better_type); 2206 } 2207 j++; 2208 } 2209 } 2210 } 2211 2212 void GraphKit::round_double_result(ciMethod* dest_method) { 2213 // A non-strict method may return a double value which has an extended 2214 // exponent, but this must not be visible in a caller which is 'strict' 2215 // If a strict caller invokes a non-strict callee, round a double result 2216 2217 BasicType result_type = dest_method->return_type()->basic_type(); 2218 assert( method() != NULL, "must have caller context"); 2219 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2220 // Destination method's return value is on top of stack 2221 // dstore_rounding() does gvn.transform 2222 Node *result = pop_pair(); 2223 result = dstore_rounding(result); 2224 push_pair(result); 2225 } 2226 } 2227 2228 // rounding for strict float precision conformance 2229 Node* GraphKit::precision_rounding(Node* n) { 2230 return UseStrictFP && _method->flags().is_strict() 2231 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2232 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2233 : n; 2234 } 2235 2236 // rounding for strict double precision conformance 2237 Node* GraphKit::dprecision_rounding(Node *n) { 2238 return UseStrictFP && _method->flags().is_strict() 2239 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2240 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2241 : n; 2242 } 2243 2244 // rounding for non-strict double stores 2245 Node* GraphKit::dstore_rounding(Node* n) { 2246 return Matcher::strict_fp_requires_explicit_rounding 2247 && UseSSE <= 1 2248 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2249 : n; 2250 } 2251 2252 //============================================================================= 2253 // Generate a fast path/slow path idiom. Graph looks like: 2254 // [foo] indicates that 'foo' is a parameter 2255 // 2256 // [in] NULL 2257 // \ / 2258 // CmpP 2259 // Bool ne 2260 // If 2261 // / \ 2262 // True False-<2> 2263 // / | 2264 // / cast_not_null 2265 // Load | | ^ 2266 // [fast_test] | | 2267 // gvn to opt_test | | 2268 // / \ | <1> 2269 // True False | 2270 // | \\ | 2271 // [slow_call] \[fast_result] 2272 // Ctl Val \ \ 2273 // | \ \ 2274 // Catch <1> \ \ 2275 // / \ ^ \ \ 2276 // Ex No_Ex | \ \ 2277 // | \ \ | \ <2> \ 2278 // ... \ [slow_res] | | \ [null_result] 2279 // \ \--+--+--- | | 2280 // \ | / \ | / 2281 // --------Region Phi 2282 // 2283 //============================================================================= 2284 // Code is structured as a series of driver functions all called 'do_XXX' that 2285 // call a set of helper functions. Helper functions first, then drivers. 2286 2287 //------------------------------null_check_oop--------------------------------- 2288 // Null check oop. Set null-path control into Region in slot 3. 2289 // Make a cast-not-nullness use the other not-null control. Return cast. 2290 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2291 bool never_see_null, bool safe_for_replace) { 2292 // Initial NULL check taken path 2293 (*null_control) = top(); 2294 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2295 2296 // Generate uncommon_trap: 2297 if (never_see_null && (*null_control) != top()) { 2298 // If we see an unexpected null at a check-cast we record it and force a 2299 // recompile; the offending check-cast will be compiled to handle NULLs. 2300 // If we see more than one offending BCI, then all checkcasts in the 2301 // method will be compiled to handle NULLs. 2302 PreserveJVMState pjvms(this); 2303 set_control(*null_control); 2304 replace_in_map(value, null()); 2305 uncommon_trap(Deoptimization::Reason_null_check, 2306 Deoptimization::Action_make_not_entrant); 2307 (*null_control) = top(); // NULL path is dead 2308 } 2309 if ((*null_control) == top() && safe_for_replace) { 2310 replace_in_map(value, cast); 2311 } 2312 2313 // Cast away null-ness on the result 2314 return cast; 2315 } 2316 2317 //------------------------------opt_iff---------------------------------------- 2318 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2319 // Return slow-path control. 2320 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2321 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2322 2323 // Fast path taken; set region slot 2 2324 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2325 region->init_req(2,fast_taken); // Capture fast-control 2326 2327 // Fast path not-taken, i.e. slow path 2328 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2329 return slow_taken; 2330 } 2331 2332 //-----------------------------make_runtime_call------------------------------- 2333 Node* GraphKit::make_runtime_call(int flags, 2334 const TypeFunc* call_type, address call_addr, 2335 const char* call_name, 2336 const TypePtr* adr_type, 2337 // The following parms are all optional. 2338 // The first NULL ends the list. 2339 Node* parm0, Node* parm1, 2340 Node* parm2, Node* parm3, 2341 Node* parm4, Node* parm5, 2342 Node* parm6, Node* parm7) { 2343 // Slow-path call 2344 bool is_leaf = !(flags & RC_NO_LEAF); 2345 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2346 if (call_name == NULL) { 2347 assert(!is_leaf, "must supply name for leaf"); 2348 call_name = OptoRuntime::stub_name(call_addr); 2349 } 2350 CallNode* call; 2351 if (!is_leaf) { 2352 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2353 bci(), adr_type); 2354 } else if (flags & RC_NO_FP) { 2355 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2356 } else { 2357 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2358 } 2359 2360 // The following is similar to set_edges_for_java_call, 2361 // except that the memory effects of the call are restricted to AliasIdxRaw. 2362 2363 // Slow path call has no side-effects, uses few values 2364 bool wide_in = !(flags & RC_NARROW_MEM); 2365 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2366 2367 Node* prev_mem = NULL; 2368 if (wide_in) { 2369 prev_mem = set_predefined_input_for_runtime_call(call); 2370 } else { 2371 assert(!wide_out, "narrow in => narrow out"); 2372 Node* narrow_mem = memory(adr_type); 2373 prev_mem = reset_memory(); 2374 map()->set_memory(narrow_mem); 2375 set_predefined_input_for_runtime_call(call); 2376 } 2377 2378 // Hook each parm in order. Stop looking at the first NULL. 2379 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2380 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2381 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2382 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2383 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2384 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2385 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2386 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2387 /* close each nested if ===> */ } } } } } } } } 2388 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2389 2390 if (!is_leaf) { 2391 // Non-leaves can block and take safepoints: 2392 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2393 } 2394 // Non-leaves can throw exceptions: 2395 if (has_io) { 2396 call->set_req(TypeFunc::I_O, i_o()); 2397 } 2398 2399 if (flags & RC_UNCOMMON) { 2400 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2401 // (An "if" probability corresponds roughly to an unconditional count. 2402 // Sort of.) 2403 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2404 } 2405 2406 Node* c = _gvn.transform(call); 2407 assert(c == call, "cannot disappear"); 2408 2409 if (wide_out) { 2410 // Slow path call has full side-effects. 2411 set_predefined_output_for_runtime_call(call); 2412 } else { 2413 // Slow path call has few side-effects, and/or sets few values. 2414 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2415 } 2416 2417 if (has_io) { 2418 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2419 } 2420 return call; 2421 2422 } 2423 2424 //------------------------------merge_memory----------------------------------- 2425 // Merge memory from one path into the current memory state. 2426 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2427 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2428 Node* old_slice = mms.force_memory(); 2429 Node* new_slice = mms.memory2(); 2430 if (old_slice != new_slice) { 2431 PhiNode* phi; 2432 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2433 phi = new_slice->as_Phi(); 2434 #ifdef ASSERT 2435 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2436 old_slice = old_slice->in(new_path); 2437 // Caller is responsible for ensuring that any pre-existing 2438 // phis are already aware of old memory. 2439 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2440 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2441 #endif 2442 mms.set_memory(phi); 2443 } else { 2444 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2445 _gvn.set_type(phi, Type::MEMORY); 2446 phi->set_req(new_path, new_slice); 2447 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2448 } 2449 } 2450 } 2451 } 2452 2453 //------------------------------make_slow_call_ex------------------------------ 2454 // Make the exception handler hookups for the slow call 2455 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2456 if (stopped()) return; 2457 2458 // Make a catch node with just two handlers: fall-through and catch-all 2459 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2460 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2461 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2462 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2463 2464 { PreserveJVMState pjvms(this); 2465 set_control(excp); 2466 set_i_o(i_o); 2467 2468 if (excp != top()) { 2469 // Create an exception state also. 2470 // Use an exact type if the caller has specified a specific exception. 2471 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2472 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2473 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2474 } 2475 } 2476 2477 // Get the no-exception control from the CatchNode. 2478 set_control(norm); 2479 } 2480 2481 2482 //-------------------------------gen_subtype_check----------------------------- 2483 // Generate a subtyping check. Takes as input the subtype and supertype. 2484 // Returns 2 values: sets the default control() to the true path and returns 2485 // the false path. Only reads invariant memory; sets no (visible) memory. 2486 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2487 // but that's not exposed to the optimizer. This call also doesn't take in an 2488 // Object; if you wish to check an Object you need to load the Object's class 2489 // prior to coming here. 2490 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2491 // Fast check for identical types, perhaps identical constants. 2492 // The types can even be identical non-constants, in cases 2493 // involving Array.newInstance, Object.clone, etc. 2494 if (subklass == superklass) 2495 return top(); // false path is dead; no test needed. 2496 2497 if (_gvn.type(superklass)->singleton()) { 2498 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2499 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2500 2501 // In the common case of an exact superklass, try to fold up the 2502 // test before generating code. You may ask, why not just generate 2503 // the code and then let it fold up? The answer is that the generated 2504 // code will necessarily include null checks, which do not always 2505 // completely fold away. If they are also needless, then they turn 2506 // into a performance loss. Example: 2507 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2508 // Here, the type of 'fa' is often exact, so the store check 2509 // of fa[1]=x will fold up, without testing the nullness of x. 2510 switch (static_subtype_check(superk, subk)) { 2511 case SSC_always_false: 2512 { 2513 Node* always_fail = control(); 2514 set_control(top()); 2515 return always_fail; 2516 } 2517 case SSC_always_true: 2518 return top(); 2519 case SSC_easy_test: 2520 { 2521 // Just do a direct pointer compare and be done. 2522 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2523 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2524 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2525 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2526 return _gvn.transform( new(C) IfFalseNode(iff) ); 2527 } 2528 case SSC_full_test: 2529 break; 2530 default: 2531 ShouldNotReachHere(); 2532 } 2533 } 2534 2535 // %%% Possible further optimization: Even if the superklass is not exact, 2536 // if the subklass is the unique subtype of the superklass, the check 2537 // will always succeed. We could leave a dependency behind to ensure this. 2538 2539 // First load the super-klass's check-offset 2540 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2541 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), 2542 TypeInt::INT, MemNode::unordered)); 2543 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2544 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2545 2546 // Load from the sub-klass's super-class display list, or a 1-word cache of 2547 // the secondary superclass list, or a failing value with a sentinel offset 2548 // if the super-klass is an interface or exceptionally deep in the Java 2549 // hierarchy and we have to scan the secondary superclass list the hard way. 2550 // Worst-case type is a little odd: NULL is allowed as a result (usually 2551 // klass loads can never produce a NULL). 2552 Node *chk_off_X = ConvI2X(chk_off); 2553 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2554 // For some types like interfaces the following loadKlass is from a 1-word 2555 // cache which is mutable so can't use immutable memory. Other 2556 // types load from the super-class display table which is immutable. 2557 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2558 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2559 2560 // Compile speed common case: ARE a subtype and we canNOT fail 2561 if( superklass == nkls ) 2562 return top(); // false path is dead; no test needed. 2563 2564 // See if we get an immediate positive hit. Happens roughly 83% of the 2565 // time. Test to see if the value loaded just previously from the subklass 2566 // is exactly the superklass. 2567 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2568 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2569 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2570 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2571 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2572 2573 // Compile speed common case: Check for being deterministic right now. If 2574 // chk_off is a constant and not equal to cacheoff then we are NOT a 2575 // subklass. In this case we need exactly the 1 test above and we can 2576 // return those results immediately. 2577 if (!might_be_cache) { 2578 Node* not_subtype_ctrl = control(); 2579 set_control(iftrue1); // We need exactly the 1 test above 2580 return not_subtype_ctrl; 2581 } 2582 2583 // Gather the various success & failures here 2584 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2585 record_for_igvn(r_ok_subtype); 2586 RegionNode *r_not_subtype = new (C) RegionNode(3); 2587 record_for_igvn(r_not_subtype); 2588 2589 r_ok_subtype->init_req(1, iftrue1); 2590 2591 // Check for immediate negative hit. Happens roughly 11% of the time (which 2592 // is roughly 63% of the remaining cases). Test to see if the loaded 2593 // check-offset points into the subklass display list or the 1-element 2594 // cache. If it points to the display (and NOT the cache) and the display 2595 // missed then it's not a subtype. 2596 Node *cacheoff = _gvn.intcon(cacheoff_con); 2597 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2598 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2599 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2600 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2601 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2602 2603 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2604 // No performance impact (too rare) but allows sharing of secondary arrays 2605 // which has some footprint reduction. 2606 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2607 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2608 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2609 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2610 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2611 2612 // -- Roads not taken here: -- 2613 // We could also have chosen to perform the self-check at the beginning 2614 // of this code sequence, as the assembler does. This would not pay off 2615 // the same way, since the optimizer, unlike the assembler, can perform 2616 // static type analysis to fold away many successful self-checks. 2617 // Non-foldable self checks work better here in second position, because 2618 // the initial primary superclass check subsumes a self-check for most 2619 // types. An exception would be a secondary type like array-of-interface, 2620 // which does not appear in its own primary supertype display. 2621 // Finally, we could have chosen to move the self-check into the 2622 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2623 // dependent manner. But it is worthwhile to have the check here, 2624 // where it can be perhaps be optimized. The cost in code space is 2625 // small (register compare, branch). 2626 2627 // Now do a linear scan of the secondary super-klass array. Again, no real 2628 // performance impact (too rare) but it's gotta be done. 2629 // Since the code is rarely used, there is no penalty for moving it 2630 // out of line, and it can only improve I-cache density. 2631 // The decision to inline or out-of-line this final check is platform 2632 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2633 Node* psc = _gvn.transform( 2634 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2635 2636 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2637 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2638 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2639 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2640 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2641 2642 // Return false path; set default control to true path. 2643 set_control( _gvn.transform(r_ok_subtype) ); 2644 return _gvn.transform(r_not_subtype); 2645 } 2646 2647 //----------------------------static_subtype_check----------------------------- 2648 // Shortcut important common cases when superklass is exact: 2649 // (0) superklass is java.lang.Object (can occur in reflective code) 2650 // (1) subklass is already limited to a subtype of superklass => always ok 2651 // (2) subklass does not overlap with superklass => always fail 2652 // (3) superklass has NO subtypes and we can check with a simple compare. 2653 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2654 if (StressReflectiveCode) { 2655 return SSC_full_test; // Let caller generate the general case. 2656 } 2657 2658 if (superk == env()->Object_klass()) { 2659 return SSC_always_true; // (0) this test cannot fail 2660 } 2661 2662 ciType* superelem = superk; 2663 if (superelem->is_array_klass()) 2664 superelem = superelem->as_array_klass()->base_element_type(); 2665 2666 if (!subk->is_interface()) { // cannot trust static interface types yet 2667 if (subk->is_subtype_of(superk)) { 2668 return SSC_always_true; // (1) false path dead; no dynamic test needed 2669 } 2670 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2671 !superk->is_subtype_of(subk)) { 2672 return SSC_always_false; 2673 } 2674 } 2675 2676 // If casting to an instance klass, it must have no subtypes 2677 if (superk->is_interface()) { 2678 // Cannot trust interfaces yet. 2679 // %%% S.B. superk->nof_implementors() == 1 2680 } else if (superelem->is_instance_klass()) { 2681 ciInstanceKlass* ik = superelem->as_instance_klass(); 2682 if (!ik->has_subklass() && !ik->is_interface()) { 2683 if (!ik->is_final()) { 2684 // Add a dependency if there is a chance of a later subclass. 2685 C->dependencies()->assert_leaf_type(ik); 2686 } 2687 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2688 } 2689 } else { 2690 // A primitive array type has no subtypes. 2691 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2692 } 2693 2694 return SSC_full_test; 2695 } 2696 2697 // Profile-driven exact type check: 2698 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2699 float prob, 2700 Node* *casted_receiver) { 2701 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2702 Node* recv_klass = load_object_klass(receiver); 2703 Node* want_klass = makecon(tklass); 2704 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2705 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2706 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2707 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2708 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2709 2710 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2711 assert(recv_xtype->klass_is_exact(), ""); 2712 2713 // Subsume downstream occurrences of receiver with a cast to 2714 // recv_xtype, since now we know what the type will be. 2715 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2716 (*casted_receiver) = _gvn.transform(cast); 2717 // (User must make the replace_in_map call.) 2718 2719 return fail; 2720 } 2721 2722 2723 //------------------------------seems_never_null------------------------------- 2724 // Use null_seen information if it is available from the profile. 2725 // If we see an unexpected null at a type check we record it and force a 2726 // recompile; the offending check will be recompiled to handle NULLs. 2727 // If we see several offending BCIs, then all checks in the 2728 // method will be recompiled. 2729 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2730 if (UncommonNullCast // Cutout for this technique 2731 && obj != null() // And not the -Xcomp stupid case? 2732 && !too_many_traps(Deoptimization::Reason_null_check) 2733 ) { 2734 if (data == NULL) 2735 // Edge case: no mature data. Be optimistic here. 2736 return true; 2737 // If the profile has not seen a null, assume it won't happen. 2738 assert(java_bc() == Bytecodes::_checkcast || 2739 java_bc() == Bytecodes::_instanceof || 2740 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2741 return !data->as_BitData()->null_seen(); 2742 } 2743 return false; 2744 } 2745 2746 //------------------------maybe_cast_profiled_receiver------------------------- 2747 // If the profile has seen exactly one type, narrow to exactly that type. 2748 // Subsequent type checks will always fold up. 2749 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2750 ciKlass* require_klass, 2751 ciKlass* spec_klass, 2752 bool safe_for_replace) { 2753 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2754 2755 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check; 2756 2757 // Make sure we haven't already deoptimized from this tactic. 2758 if (too_many_traps(reason)) 2759 return NULL; 2760 2761 // (No, this isn't a call, but it's enough like a virtual call 2762 // to use the same ciMethod accessor to get the profile info...) 2763 // If we have a speculative type use it instead of profiling (which 2764 // may not help us) 2765 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2766 if (exact_kls != NULL) {// no cast failures here 2767 if (require_klass == NULL || 2768 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2769 // If we narrow the type to match what the type profile sees or 2770 // the speculative type, we can then remove the rest of the 2771 // cast. 2772 // This is a win, even if the exact_kls is very specific, 2773 // because downstream operations, such as method calls, 2774 // will often benefit from the sharper type. 2775 Node* exact_obj = not_null_obj; // will get updated in place... 2776 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2777 &exact_obj); 2778 { PreserveJVMState pjvms(this); 2779 set_control(slow_ctl); 2780 uncommon_trap(reason, 2781 Deoptimization::Action_maybe_recompile); 2782 } 2783 if (safe_for_replace) { 2784 replace_in_map(not_null_obj, exact_obj); 2785 } 2786 return exact_obj; 2787 } 2788 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2789 } 2790 2791 return NULL; 2792 } 2793 2794 /** 2795 * Cast obj to type and emit guard unless we had too many traps here 2796 * already 2797 * 2798 * @param obj node being casted 2799 * @param type type to cast the node to 2800 * @param not_null true if we know node cannot be null 2801 */ 2802 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2803 ciKlass* type, 2804 bool not_null) { 2805 // type == NULL if profiling tells us this object is always null 2806 if (type != NULL) { 2807 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2808 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check; 2809 if (!too_many_traps(null_reason) && 2810 !too_many_traps(class_reason)) { 2811 Node* not_null_obj = NULL; 2812 // not_null is true if we know the object is not null and 2813 // there's no need for a null check 2814 if (!not_null) { 2815 Node* null_ctl = top(); 2816 not_null_obj = null_check_oop(obj, &null_ctl, true, true); 2817 assert(null_ctl->is_top(), "no null control here"); 2818 } else { 2819 not_null_obj = obj; 2820 } 2821 2822 Node* exact_obj = not_null_obj; 2823 ciKlass* exact_kls = type; 2824 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2825 &exact_obj); 2826 { 2827 PreserveJVMState pjvms(this); 2828 set_control(slow_ctl); 2829 uncommon_trap(class_reason, 2830 Deoptimization::Action_maybe_recompile); 2831 } 2832 replace_in_map(not_null_obj, exact_obj); 2833 obj = exact_obj; 2834 } 2835 } else { 2836 if (!too_many_traps(Deoptimization::Reason_null_assert)) { 2837 Node* exact_obj = null_assert(obj); 2838 replace_in_map(obj, exact_obj); 2839 obj = exact_obj; 2840 } 2841 } 2842 return obj; 2843 } 2844 2845 //-------------------------------gen_instanceof-------------------------------- 2846 // Generate an instance-of idiom. Used by both the instance-of bytecode 2847 // and the reflective instance-of call. 2848 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2849 kill_dead_locals(); // Benefit all the uncommon traps 2850 assert( !stopped(), "dead parse path should be checked in callers" ); 2851 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2852 "must check for not-null not-dead klass in callers"); 2853 2854 // Make the merge point 2855 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2856 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2857 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2858 C->set_has_split_ifs(true); // Has chance for split-if optimization 2859 2860 ciProfileData* data = NULL; 2861 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2862 data = method()->method_data()->bci_to_data(bci()); 2863 } 2864 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2865 && seems_never_null(obj, data)); 2866 2867 // Null check; get casted pointer; set region slot 3 2868 Node* null_ctl = top(); 2869 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2870 2871 // If not_null_obj is dead, only null-path is taken 2872 if (stopped()) { // Doing instance-of on a NULL? 2873 set_control(null_ctl); 2874 return intcon(0); 2875 } 2876 region->init_req(_null_path, null_ctl); 2877 phi ->init_req(_null_path, intcon(0)); // Set null path value 2878 if (null_ctl == top()) { 2879 // Do this eagerly, so that pattern matches like is_diamond_phi 2880 // will work even during parsing. 2881 assert(_null_path == PATH_LIMIT-1, "delete last"); 2882 region->del_req(_null_path); 2883 phi ->del_req(_null_path); 2884 } 2885 2886 // Do we know the type check always succeed? 2887 bool known_statically = false; 2888 if (_gvn.type(superklass)->singleton()) { 2889 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2890 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2891 if (subk != NULL && subk->is_loaded()) { 2892 int static_res = static_subtype_check(superk, subk); 2893 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2894 } 2895 } 2896 2897 if (known_statically && UseTypeSpeculation) { 2898 // If we know the type check always succeeds then we don't use the 2899 // profiling data at this bytecode. Don't lose it, feed it to the 2900 // type system as a speculative type. 2901 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2902 } else { 2903 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2904 // We may not have profiling here or it may not help us. If we 2905 // have a speculative type use it to perform an exact cast. 2906 ciKlass* spec_obj_type = obj_type->speculative_type(); 2907 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2908 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2909 if (stopped()) { // Profile disagrees with this path. 2910 set_control(null_ctl); // Null is the only remaining possibility. 2911 return intcon(0); 2912 } 2913 if (cast_obj != NULL) { 2914 not_null_obj = cast_obj; 2915 } 2916 } 2917 } 2918 2919 // Load the object's klass 2920 Node* obj_klass = load_object_klass(not_null_obj); 2921 2922 // Generate the subtype check 2923 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2924 2925 // Plug in the success path to the general merge in slot 1. 2926 region->init_req(_obj_path, control()); 2927 phi ->init_req(_obj_path, intcon(1)); 2928 2929 // Plug in the failing path to the general merge in slot 2. 2930 region->init_req(_fail_path, not_subtype_ctrl); 2931 phi ->init_req(_fail_path, intcon(0)); 2932 2933 // Return final merged results 2934 set_control( _gvn.transform(region) ); 2935 record_for_igvn(region); 2936 return _gvn.transform(phi); 2937 } 2938 2939 //-------------------------------gen_checkcast--------------------------------- 2940 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2941 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2942 // uncommon-trap paths work. Adjust stack after this call. 2943 // If failure_control is supplied and not null, it is filled in with 2944 // the control edge for the cast failure. Otherwise, an appropriate 2945 // uncommon trap or exception is thrown. 2946 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2947 Node* *failure_control) { 2948 kill_dead_locals(); // Benefit all the uncommon traps 2949 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2950 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2951 2952 // Fast cutout: Check the case that the cast is vacuously true. 2953 // This detects the common cases where the test will short-circuit 2954 // away completely. We do this before we perform the null check, 2955 // because if the test is going to turn into zero code, we don't 2956 // want a residual null check left around. (Causes a slowdown, 2957 // for example, in some objArray manipulations, such as a[i]=a[j].) 2958 if (tk->singleton()) { 2959 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2960 if (objtp != NULL && objtp->klass() != NULL) { 2961 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2962 case SSC_always_true: 2963 // If we know the type check always succeed then we don't use 2964 // the profiling data at this bytecode. Don't lose it, feed it 2965 // to the type system as a speculative type. 2966 return record_profiled_receiver_for_speculation(obj); 2967 case SSC_always_false: 2968 // It needs a null check because a null will *pass* the cast check. 2969 // A non-null value will always produce an exception. 2970 return null_assert(obj); 2971 } 2972 } 2973 } 2974 2975 ciProfileData* data = NULL; 2976 bool safe_for_replace = false; 2977 if (failure_control == NULL) { // use MDO in regular case only 2978 assert(java_bc() == Bytecodes::_aastore || 2979 java_bc() == Bytecodes::_checkcast, 2980 "interpreter profiles type checks only for these BCs"); 2981 data = method()->method_data()->bci_to_data(bci()); 2982 safe_for_replace = true; 2983 } 2984 2985 // Make the merge point 2986 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2987 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 2988 Node* phi = new (C) PhiNode(region, toop); 2989 C->set_has_split_ifs(true); // Has chance for split-if optimization 2990 2991 // Use null-cast information if it is available 2992 bool never_see_null = ((failure_control == NULL) // regular case only 2993 && seems_never_null(obj, data)); 2994 2995 // Null check; get casted pointer; set region slot 3 2996 Node* null_ctl = top(); 2997 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2998 2999 // If not_null_obj is dead, only null-path is taken 3000 if (stopped()) { // Doing instance-of on a NULL? 3001 set_control(null_ctl); 3002 return null(); 3003 } 3004 region->init_req(_null_path, null_ctl); 3005 phi ->init_req(_null_path, null()); // Set null path value 3006 if (null_ctl == top()) { 3007 // Do this eagerly, so that pattern matches like is_diamond_phi 3008 // will work even during parsing. 3009 assert(_null_path == PATH_LIMIT-1, "delete last"); 3010 region->del_req(_null_path); 3011 phi ->del_req(_null_path); 3012 } 3013 3014 Node* cast_obj = NULL; 3015 if (tk->klass_is_exact()) { 3016 // The following optimization tries to statically cast the speculative type of the object 3017 // (for example obtained during profiling) to the type of the superklass and then do a 3018 // dynamic check that the type of the object is what we expect. To work correctly 3019 // for checkcast and aastore the type of superklass should be exact. 3020 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3021 // We may not have profiling here or it may not help us. If we have 3022 // a speculative type use it to perform an exact cast. 3023 ciKlass* spec_obj_type = obj_type->speculative_type(); 3024 if (spec_obj_type != NULL || 3025 (data != NULL && 3026 // Counter has never been decremented (due to cast failure). 3027 // ...This is a reasonable thing to expect. It is true of 3028 // all casts inserted by javac to implement generic types. 3029 data->as_CounterData()->count() >= 0)) { 3030 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3031 if (cast_obj != NULL) { 3032 if (failure_control != NULL) // failure is now impossible 3033 (*failure_control) = top(); 3034 // adjust the type of the phi to the exact klass: 3035 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3036 } 3037 } 3038 } 3039 3040 if (cast_obj == NULL) { 3041 // Load the object's klass 3042 Node* obj_klass = load_object_klass(not_null_obj); 3043 3044 // Generate the subtype check 3045 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3046 3047 // Plug in success path into the merge 3048 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 3049 not_null_obj, toop)); 3050 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3051 if (failure_control == NULL) { 3052 if (not_subtype_ctrl != top()) { // If failure is possible 3053 PreserveJVMState pjvms(this); 3054 set_control(not_subtype_ctrl); 3055 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3056 } 3057 } else { 3058 (*failure_control) = not_subtype_ctrl; 3059 } 3060 } 3061 3062 region->init_req(_obj_path, control()); 3063 phi ->init_req(_obj_path, cast_obj); 3064 3065 // A merge of NULL or Casted-NotNull obj 3066 Node* res = _gvn.transform(phi); 3067 3068 // Note I do NOT always 'replace_in_map(obj,result)' here. 3069 // if( tk->klass()->can_be_primary_super() ) 3070 // This means that if I successfully store an Object into an array-of-String 3071 // I 'forget' that the Object is really now known to be a String. I have to 3072 // do this because we don't have true union types for interfaces - if I store 3073 // a Baz into an array-of-Interface and then tell the optimizer it's an 3074 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3075 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3076 // replace_in_map( obj, res ); 3077 3078 // Return final merged results 3079 set_control( _gvn.transform(region) ); 3080 record_for_igvn(region); 3081 return res; 3082 } 3083 3084 //------------------------------next_monitor----------------------------------- 3085 // What number should be given to the next monitor? 3086 int GraphKit::next_monitor() { 3087 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3088 int next = current + C->sync_stack_slots(); 3089 // Keep the toplevel high water mark current: 3090 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3091 return current; 3092 } 3093 3094 //------------------------------insert_mem_bar--------------------------------- 3095 // Memory barrier to avoid floating things around 3096 // The membar serves as a pinch point between both control and all memory slices. 3097 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3098 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3099 mb->init_req(TypeFunc::Control, control()); 3100 mb->init_req(TypeFunc::Memory, reset_memory()); 3101 Node* membar = _gvn.transform(mb); 3102 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3103 set_all_memory_call(membar); 3104 return membar; 3105 } 3106 3107 //-------------------------insert_mem_bar_volatile---------------------------- 3108 // Memory barrier to avoid floating things around 3109 // The membar serves as a pinch point between both control and memory(alias_idx). 3110 // If you want to make a pinch point on all memory slices, do not use this 3111 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3112 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3113 // When Parse::do_put_xxx updates a volatile field, it appends a series 3114 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3115 // The first membar is on the same memory slice as the field store opcode. 3116 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3117 // All the other membars (for other volatile slices, including AliasIdxBot, 3118 // which stands for all unknown volatile slices) are control-dependent 3119 // on the first membar. This prevents later volatile loads or stores 3120 // from sliding up past the just-emitted store. 3121 3122 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3123 mb->set_req(TypeFunc::Control,control()); 3124 if (alias_idx == Compile::AliasIdxBot) { 3125 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3126 } else { 3127 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3128 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3129 } 3130 Node* membar = _gvn.transform(mb); 3131 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3132 if (alias_idx == Compile::AliasIdxBot) { 3133 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 3134 } else { 3135 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 3136 } 3137 return membar; 3138 } 3139 3140 //------------------------------shared_lock------------------------------------ 3141 // Emit locking code. 3142 FastLockNode* GraphKit::shared_lock(Node* obj) { 3143 // bci is either a monitorenter bc or InvocationEntryBci 3144 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3145 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3146 3147 if( !GenerateSynchronizationCode ) 3148 return NULL; // Not locking things? 3149 if (stopped()) // Dead monitor? 3150 return NULL; 3151 3152 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3153 3154 // Box the stack location 3155 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 3156 Node* mem = reset_memory(); 3157 3158 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 3159 if (PrintPreciseBiasedLockingStatistics) { 3160 // Create the counters for this fast lock. 3161 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3162 } 3163 // Add monitor to debug info for the slow path. If we block inside the 3164 // slow path and de-opt, we need the monitor hanging around 3165 map()->push_monitor( flock ); 3166 3167 const TypeFunc *tf = LockNode::lock_type(); 3168 LockNode *lock = new (C) LockNode(C, tf); 3169 3170 lock->init_req( TypeFunc::Control, control() ); 3171 lock->init_req( TypeFunc::Memory , mem ); 3172 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3173 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3174 lock->init_req( TypeFunc::ReturnAdr, top() ); 3175 3176 lock->init_req(TypeFunc::Parms + 0, obj); 3177 lock->init_req(TypeFunc::Parms + 1, box); 3178 lock->init_req(TypeFunc::Parms + 2, flock); 3179 add_safepoint_edges(lock); 3180 3181 lock = _gvn.transform( lock )->as_Lock(); 3182 3183 // lock has no side-effects, sets few values 3184 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3185 3186 insert_mem_bar(Op_MemBarAcquireLock); 3187 3188 // Add this to the worklist so that the lock can be eliminated 3189 record_for_igvn(lock); 3190 3191 #ifndef PRODUCT 3192 if (PrintLockStatistics) { 3193 // Update the counter for this lock. Don't bother using an atomic 3194 // operation since we don't require absolute accuracy. 3195 lock->create_lock_counter(map()->jvms()); 3196 increment_counter(lock->counter()->addr()); 3197 } 3198 #endif 3199 3200 return flock; 3201 } 3202 3203 3204 //------------------------------shared_unlock---------------------------------- 3205 // Emit unlocking code. 3206 void GraphKit::shared_unlock(Node* box, Node* obj) { 3207 // bci is either a monitorenter bc or InvocationEntryBci 3208 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3209 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3210 3211 if( !GenerateSynchronizationCode ) 3212 return; 3213 if (stopped()) { // Dead monitor? 3214 map()->pop_monitor(); // Kill monitor from debug info 3215 return; 3216 } 3217 3218 // Memory barrier to avoid floating things down past the locked region 3219 insert_mem_bar(Op_MemBarReleaseLock); 3220 3221 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3222 UnlockNode *unlock = new (C) UnlockNode(C, tf); 3223 uint raw_idx = Compile::AliasIdxRaw; 3224 unlock->init_req( TypeFunc::Control, control() ); 3225 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3226 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3227 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3228 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3229 3230 unlock->init_req(TypeFunc::Parms + 0, obj); 3231 unlock->init_req(TypeFunc::Parms + 1, box); 3232 unlock = _gvn.transform(unlock)->as_Unlock(); 3233 3234 Node* mem = reset_memory(); 3235 3236 // unlock has no side-effects, sets few values 3237 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3238 3239 // Kill monitor from debug info 3240 map()->pop_monitor( ); 3241 } 3242 3243 //-------------------------------get_layout_helper----------------------------- 3244 // If the given klass is a constant or known to be an array, 3245 // fetch the constant layout helper value into constant_value 3246 // and return (Node*)NULL. Otherwise, load the non-constant 3247 // layout helper value, and return the node which represents it. 3248 // This two-faced routine is useful because allocation sites 3249 // almost always feature constant types. 3250 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3251 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3252 if (!StressReflectiveCode && inst_klass != NULL) { 3253 ciKlass* klass = inst_klass->klass(); 3254 bool xklass = inst_klass->klass_is_exact(); 3255 if (xklass || klass->is_array_klass()) { 3256 jint lhelper = klass->layout_helper(); 3257 if (lhelper != Klass::_lh_neutral_value) { 3258 constant_value = lhelper; 3259 return (Node*) NULL; 3260 } 3261 } 3262 } 3263 constant_value = Klass::_lh_neutral_value; // put in a known value 3264 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3265 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3266 } 3267 3268 // We just put in an allocate/initialize with a big raw-memory effect. 3269 // Hook selected additional alias categories on the initialization. 3270 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3271 MergeMemNode* init_in_merge, 3272 Node* init_out_raw) { 3273 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3274 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3275 3276 Node* prevmem = kit.memory(alias_idx); 3277 init_in_merge->set_memory_at(alias_idx, prevmem); 3278 kit.set_memory(init_out_raw, alias_idx); 3279 } 3280 3281 //---------------------------set_output_for_allocation------------------------- 3282 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3283 const TypeOopPtr* oop_type) { 3284 int rawidx = Compile::AliasIdxRaw; 3285 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3286 add_safepoint_edges(alloc); 3287 Node* allocx = _gvn.transform(alloc); 3288 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3289 // create memory projection for i_o 3290 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3291 make_slow_call_ex(allocx, env()->Throwable_klass(), true); 3292 3293 // create a memory projection as for the normal control path 3294 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3295 set_memory(malloc, rawidx); 3296 3297 // a normal slow-call doesn't change i_o, but an allocation does 3298 // we create a separate i_o projection for the normal control path 3299 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3300 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3301 3302 // put in an initialization barrier 3303 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3304 rawoop)->as_Initialize(); 3305 assert(alloc->initialization() == init, "2-way macro link must work"); 3306 assert(init ->allocation() == alloc, "2-way macro link must work"); 3307 { 3308 // Extract memory strands which may participate in the new object's 3309 // initialization, and source them from the new InitializeNode. 3310 // This will allow us to observe initializations when they occur, 3311 // and link them properly (as a group) to the InitializeNode. 3312 assert(init->in(InitializeNode::Memory) == malloc, ""); 3313 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3314 init->set_req(InitializeNode::Memory, minit_in); 3315 record_for_igvn(minit_in); // fold it up later, if possible 3316 Node* minit_out = memory(rawidx); 3317 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3318 if (oop_type->isa_aryptr()) { 3319 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3320 int elemidx = C->get_alias_index(telemref); 3321 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3322 } else if (oop_type->isa_instptr()) { 3323 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3324 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3325 ciField* field = ik->nonstatic_field_at(i); 3326 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3327 continue; // do not bother to track really large numbers of fields 3328 // Find (or create) the alias category for this field: 3329 int fieldidx = C->alias_type(field)->index(); 3330 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3331 } 3332 } 3333 } 3334 3335 // Cast raw oop to the real thing... 3336 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3337 javaoop = _gvn.transform(javaoop); 3338 C->set_recent_alloc(control(), javaoop); 3339 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3340 3341 #ifdef ASSERT 3342 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3343 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3344 "Ideal_allocation works"); 3345 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3346 "Ideal_allocation works"); 3347 if (alloc->is_AllocateArray()) { 3348 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3349 "Ideal_allocation works"); 3350 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3351 "Ideal_allocation works"); 3352 } else { 3353 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3354 } 3355 } 3356 #endif //ASSERT 3357 3358 return javaoop; 3359 } 3360 3361 //---------------------------new_instance-------------------------------------- 3362 // This routine takes a klass_node which may be constant (for a static type) 3363 // or may be non-constant (for reflective code). It will work equally well 3364 // for either, and the graph will fold nicely if the optimizer later reduces 3365 // the type to a constant. 3366 // The optional arguments are for specialized use by intrinsics: 3367 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3368 // - If 'return_size_val', report the the total object size to the caller. 3369 Node* GraphKit::new_instance(Node* klass_node, 3370 Node* extra_slow_test, 3371 Node* *return_size_val) { 3372 // Compute size in doublewords 3373 // The size is always an integral number of doublewords, represented 3374 // as a positive bytewise size stored in the klass's layout_helper. 3375 // The layout_helper also encodes (in a low bit) the need for a slow path. 3376 jint layout_con = Klass::_lh_neutral_value; 3377 Node* layout_val = get_layout_helper(klass_node, layout_con); 3378 int layout_is_con = (layout_val == NULL); 3379 3380 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3381 // Generate the initial go-slow test. It's either ALWAYS (return a 3382 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3383 // case) a computed value derived from the layout_helper. 3384 Node* initial_slow_test = NULL; 3385 if (layout_is_con) { 3386 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3387 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3388 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3389 3390 } else { // reflective case 3391 // This reflective path is used by Unsafe.allocateInstance. 3392 // (It may be stress-tested by specifying StressReflectiveCode.) 3393 // Basically, we want to get into the VM is there's an illegal argument. 3394 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3395 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3396 if (extra_slow_test != intcon(0)) { 3397 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3398 } 3399 // (Macro-expander will further convert this to a Bool, if necessary.) 3400 } 3401 3402 // Find the size in bytes. This is easy; it's the layout_helper. 3403 // The size value must be valid even if the slow path is taken. 3404 Node* size = NULL; 3405 if (layout_is_con) { 3406 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3407 } else { // reflective case 3408 // This reflective path is used by clone and Unsafe.allocateInstance. 3409 size = ConvI2X(layout_val); 3410 3411 // Clear the low bits to extract layout_helper_size_in_bytes: 3412 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3413 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3414 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3415 } 3416 if (return_size_val != NULL) { 3417 (*return_size_val) = size; 3418 } 3419 3420 // This is a precise notnull oop of the klass. 3421 // (Actually, it need not be precise if this is a reflective allocation.) 3422 // It's what we cast the result to. 3423 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3424 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3425 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3426 3427 // Now generate allocation code 3428 3429 // The entire memory state is needed for slow path of the allocation 3430 // since GC and deoptimization can happened. 3431 Node *mem = reset_memory(); 3432 set_all_memory(mem); // Create new memory state 3433 3434 AllocateNode* alloc 3435 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3436 control(), mem, i_o(), 3437 size, klass_node, 3438 initial_slow_test); 3439 3440 return set_output_for_allocation(alloc, oop_type); 3441 } 3442 3443 //-------------------------------new_array------------------------------------- 3444 // helper for both newarray and anewarray 3445 // The 'length' parameter is (obviously) the length of the array. 3446 // See comments on new_instance for the meaning of the other arguments. 3447 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3448 Node* length, // number of array elements 3449 int nargs, // number of arguments to push back for uncommon trap 3450 Node* *return_size_val) { 3451 jint layout_con = Klass::_lh_neutral_value; 3452 Node* layout_val = get_layout_helper(klass_node, layout_con); 3453 int layout_is_con = (layout_val == NULL); 3454 3455 if (!layout_is_con && !StressReflectiveCode && 3456 !too_many_traps(Deoptimization::Reason_class_check)) { 3457 // This is a reflective array creation site. 3458 // Optimistically assume that it is a subtype of Object[], 3459 // so that we can fold up all the address arithmetic. 3460 layout_con = Klass::array_layout_helper(T_OBJECT); 3461 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3462 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3463 { BuildCutout unless(this, bol_lh, PROB_MAX); 3464 inc_sp(nargs); 3465 uncommon_trap(Deoptimization::Reason_class_check, 3466 Deoptimization::Action_maybe_recompile); 3467 } 3468 layout_val = NULL; 3469 layout_is_con = true; 3470 } 3471 3472 // Generate the initial go-slow test. Make sure we do not overflow 3473 // if length is huge (near 2Gig) or negative! We do not need 3474 // exact double-words here, just a close approximation of needed 3475 // double-words. We can't add any offset or rounding bits, lest we 3476 // take a size -1 of bytes and make it positive. Use an unsigned 3477 // compare, so negative sizes look hugely positive. 3478 int fast_size_limit = FastAllocateSizeLimit; 3479 if (layout_is_con) { 3480 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3481 // Increase the size limit if we have exact knowledge of array type. 3482 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3483 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3484 } 3485 3486 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3487 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3488 if (initial_slow_test->is_Bool()) { 3489 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3490 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3491 } 3492 3493 // --- Size Computation --- 3494 // array_size = round_to_heap(array_header + (length << elem_shift)); 3495 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3496 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3497 // The rounding mask is strength-reduced, if possible. 3498 int round_mask = MinObjAlignmentInBytes - 1; 3499 Node* header_size = NULL; 3500 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3501 // (T_BYTE has the weakest alignment and size restrictions...) 3502 if (layout_is_con) { 3503 int hsize = Klass::layout_helper_header_size(layout_con); 3504 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3505 BasicType etype = Klass::layout_helper_element_type(layout_con); 3506 if ((round_mask & ~right_n_bits(eshift)) == 0) 3507 round_mask = 0; // strength-reduce it if it goes away completely 3508 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3509 assert(header_size_min <= hsize, "generic minimum is smallest"); 3510 header_size_min = hsize; 3511 header_size = intcon(hsize + round_mask); 3512 } else { 3513 Node* hss = intcon(Klass::_lh_header_size_shift); 3514 Node* hsm = intcon(Klass::_lh_header_size_mask); 3515 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3516 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3517 Node* mask = intcon(round_mask); 3518 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3519 } 3520 3521 Node* elem_shift = NULL; 3522 if (layout_is_con) { 3523 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3524 if (eshift != 0) 3525 elem_shift = intcon(eshift); 3526 } else { 3527 // There is no need to mask or shift this value. 3528 // The semantics of LShiftINode include an implicit mask to 0x1F. 3529 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3530 elem_shift = layout_val; 3531 } 3532 3533 // Transition to native address size for all offset calculations: 3534 Node* lengthx = ConvI2X(length); 3535 Node* headerx = ConvI2X(header_size); 3536 #ifdef _LP64 3537 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3538 if (tllen != NULL && tllen->_lo < 0) { 3539 // Add a manual constraint to a positive range. Cf. array_element_address. 3540 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3541 if (size_max > tllen->_hi) size_max = tllen->_hi; 3542 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3543 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon)); 3544 } 3545 } 3546 #endif 3547 3548 // Combine header size (plus rounding) and body size. Then round down. 3549 // This computation cannot overflow, because it is used only in two 3550 // places, one where the length is sharply limited, and the other 3551 // after a successful allocation. 3552 Node* abody = lengthx; 3553 if (elem_shift != NULL) 3554 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3555 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3556 if (round_mask != 0) { 3557 Node* mask = MakeConX(~round_mask); 3558 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3559 } 3560 // else if round_mask == 0, the size computation is self-rounding 3561 3562 if (return_size_val != NULL) { 3563 // This is the size 3564 (*return_size_val) = size; 3565 } 3566 3567 // Now generate allocation code 3568 3569 // The entire memory state is needed for slow path of the allocation 3570 // since GC and deoptimization can happened. 3571 Node *mem = reset_memory(); 3572 set_all_memory(mem); // Create new memory state 3573 3574 // Create the AllocateArrayNode and its result projections 3575 AllocateArrayNode* alloc 3576 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3577 control(), mem, i_o(), 3578 size, klass_node, 3579 initial_slow_test, 3580 length); 3581 3582 // Cast to correct type. Note that the klass_node may be constant or not, 3583 // and in the latter case the actual array type will be inexact also. 3584 // (This happens via a non-constant argument to inline_native_newArray.) 3585 // In any case, the value of klass_node provides the desired array type. 3586 const TypeInt* length_type = _gvn.find_int_type(length); 3587 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3588 if (ary_type->isa_aryptr() && length_type != NULL) { 3589 // Try to get a better type than POS for the size 3590 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3591 } 3592 3593 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3594 3595 // Cast length on remaining path to be as narrow as possible 3596 if (map()->find_edge(length) >= 0) { 3597 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3598 if (ccast != length) { 3599 _gvn.set_type_bottom(ccast); 3600 record_for_igvn(ccast); 3601 replace_in_map(length, ccast); 3602 } 3603 } 3604 3605 return javaoop; 3606 } 3607 3608 // The following "Ideal_foo" functions are placed here because they recognize 3609 // the graph shapes created by the functions immediately above. 3610 3611 //---------------------------Ideal_allocation---------------------------------- 3612 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3613 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3614 if (ptr == NULL) { // reduce dumb test in callers 3615 return NULL; 3616 } 3617 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3618 ptr = ptr->in(1); 3619 if (ptr == NULL) return NULL; 3620 } 3621 // Return NULL for allocations with several casts: 3622 // j.l.reflect.Array.newInstance(jobject, jint) 3623 // Object.clone() 3624 // to keep more precise type from last cast. 3625 if (ptr->is_Proj()) { 3626 Node* allo = ptr->in(0); 3627 if (allo != NULL && allo->is_Allocate()) { 3628 return allo->as_Allocate(); 3629 } 3630 } 3631 // Report failure to match. 3632 return NULL; 3633 } 3634 3635 // Fancy version which also strips off an offset (and reports it to caller). 3636 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3637 intptr_t& offset) { 3638 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3639 if (base == NULL) return NULL; 3640 return Ideal_allocation(base, phase); 3641 } 3642 3643 // Trace Initialize <- Proj[Parm] <- Allocate 3644 AllocateNode* InitializeNode::allocation() { 3645 Node* rawoop = in(InitializeNode::RawAddress); 3646 if (rawoop->is_Proj()) { 3647 Node* alloc = rawoop->in(0); 3648 if (alloc->is_Allocate()) { 3649 return alloc->as_Allocate(); 3650 } 3651 } 3652 return NULL; 3653 } 3654 3655 // Trace Allocate -> Proj[Parm] -> Initialize 3656 InitializeNode* AllocateNode::initialization() { 3657 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3658 if (rawoop == NULL) return NULL; 3659 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3660 Node* init = rawoop->fast_out(i); 3661 if (init->is_Initialize()) { 3662 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3663 return init->as_Initialize(); 3664 } 3665 } 3666 return NULL; 3667 } 3668 3669 //----------------------------- loop predicates --------------------------- 3670 3671 //------------------------------add_predicate_impl---------------------------- 3672 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3673 // Too many traps seen? 3674 if (too_many_traps(reason)) { 3675 #ifdef ASSERT 3676 if (TraceLoopPredicate) { 3677 int tc = C->trap_count(reason); 3678 tty->print("too many traps=%s tcount=%d in ", 3679 Deoptimization::trap_reason_name(reason), tc); 3680 method()->print(); // which method has too many predicate traps 3681 tty->cr(); 3682 } 3683 #endif 3684 // We cannot afford to take more traps here, 3685 // do not generate predicate. 3686 return; 3687 } 3688 3689 Node *cont = _gvn.intcon(1); 3690 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3691 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3692 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3693 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3694 C->add_predicate_opaq(opq); 3695 { 3696 PreserveJVMState pjvms(this); 3697 set_control(iffalse); 3698 inc_sp(nargs); 3699 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3700 } 3701 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3702 set_control(iftrue); 3703 } 3704 3705 //------------------------------add_predicate--------------------------------- 3706 void GraphKit::add_predicate(int nargs) { 3707 if (UseLoopPredicate) { 3708 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3709 } 3710 // loop's limit check predicate should be near the loop. 3711 if (LoopLimitCheck) { 3712 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3713 } 3714 } 3715 3716 //----------------------------- store barriers ---------------------------- 3717 #define __ ideal. 3718 3719 void GraphKit::sync_kit(IdealKit& ideal) { 3720 set_all_memory(__ merged_memory()); 3721 set_i_o(__ i_o()); 3722 set_control(__ ctrl()); 3723 } 3724 3725 void GraphKit::final_sync(IdealKit& ideal) { 3726 // Final sync IdealKit and graphKit. 3727 sync_kit(ideal); 3728 } 3729 3730 // vanilla/CMS post barrier 3731 // Insert a write-barrier store. This is to let generational GC work; we have 3732 // to flag all oop-stores before the next GC point. 3733 void GraphKit::write_barrier_post(Node* oop_store, 3734 Node* obj, 3735 Node* adr, 3736 uint adr_idx, 3737 Node* val, 3738 bool use_precise) { 3739 // No store check needed if we're storing a NULL or an old object 3740 // (latter case is probably a string constant). The concurrent 3741 // mark sweep garbage collector, however, needs to have all nonNull 3742 // oop updates flagged via card-marks. 3743 if (val != NULL && val->is_Con()) { 3744 // must be either an oop or NULL 3745 const Type* t = val->bottom_type(); 3746 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3747 // stores of null never (?) need barriers 3748 return; 3749 } 3750 3751 if (use_ReduceInitialCardMarks() 3752 && obj == just_allocated_object(control())) { 3753 // We can skip marks on a freshly-allocated object in Eden. 3754 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3755 // That routine informs GC to take appropriate compensating steps, 3756 // upon a slow-path allocation, so as to make this card-mark 3757 // elision safe. 3758 return; 3759 } 3760 3761 if (!use_precise) { 3762 // All card marks for a (non-array) instance are in one place: 3763 adr = obj; 3764 } 3765 // (Else it's an array (or unknown), and we want more precise card marks.) 3766 assert(adr != NULL, ""); 3767 3768 IdealKit ideal(this, true); 3769 3770 // Convert the pointer to an int prior to doing math on it 3771 Node* cast = __ CastPX(__ ctrl(), adr); 3772 3773 // Divide by card size 3774 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3775 "Only one we handle so far."); 3776 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3777 3778 // Combine card table base and card offset 3779 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3780 3781 // Get the alias_index for raw card-mark memory 3782 int adr_type = Compile::AliasIdxRaw; 3783 Node* zero = __ ConI(0); // Dirty card value 3784 BasicType bt = T_BYTE; 3785 3786 if (UseCondCardMark) { 3787 // The classic GC reference write barrier is typically implemented 3788 // as a store into the global card mark table. Unfortunately 3789 // unconditional stores can result in false sharing and excessive 3790 // coherence traffic as well as false transactional aborts. 3791 // UseCondCardMark enables MP "polite" conditional card mark 3792 // stores. In theory we could relax the load from ctrl() to 3793 // no_ctrl, but that doesn't buy much latitude. 3794 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3795 __ if_then(card_val, BoolTest::ne, zero); 3796 } 3797 3798 // Smash zero into card 3799 if( !UseConcMarkSweepGC ) { 3800 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); 3801 } else { 3802 // Specialized path for CM store barrier 3803 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3804 } 3805 3806 if (UseCondCardMark) { 3807 __ end_if(); 3808 } 3809 3810 // Final sync IdealKit and GraphKit. 3811 final_sync(ideal); 3812 } 3813 3814 // G1 pre/post barriers 3815 void GraphKit::g1_write_barrier_pre(bool do_load, 3816 Node* obj, 3817 Node* adr, 3818 uint alias_idx, 3819 Node* val, 3820 const TypeOopPtr* val_type, 3821 Node* pre_val, 3822 BasicType bt) { 3823 3824 // Some sanity checks 3825 // Note: val is unused in this routine. 3826 3827 if (do_load) { 3828 // We need to generate the load of the previous value 3829 assert(obj != NULL, "must have a base"); 3830 assert(adr != NULL, "where are loading from?"); 3831 assert(pre_val == NULL, "loaded already?"); 3832 assert(val_type != NULL, "need a type"); 3833 } else { 3834 // In this case both val_type and alias_idx are unused. 3835 assert(pre_val != NULL, "must be loaded already"); 3836 // Nothing to be done if pre_val is null. 3837 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3838 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3839 } 3840 assert(bt == T_OBJECT, "or we shouldn't be here"); 3841 3842 IdealKit ideal(this, true); 3843 3844 Node* tls = __ thread(); // ThreadLocalStorage 3845 3846 Node* no_ctrl = NULL; 3847 Node* no_base = __ top(); 3848 Node* zero = __ ConI(0); 3849 Node* zeroX = __ ConX(0); 3850 3851 float likely = PROB_LIKELY(0.999); 3852 float unlikely = PROB_UNLIKELY(0.999); 3853 3854 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3855 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3856 3857 // Offsets into the thread 3858 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3859 PtrQueue::byte_offset_of_active()); 3860 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3861 PtrQueue::byte_offset_of_index()); 3862 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3863 PtrQueue::byte_offset_of_buf()); 3864 3865 // Now the actual pointers into the thread 3866 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3867 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3868 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3869 3870 // Now some of the values 3871 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3872 3873 // if (!marking) 3874 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3875 BasicType index_bt = TypeX_X->basic_type(); 3876 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3877 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3878 3879 if (do_load) { 3880 // load original value 3881 // alias_idx correct?? 3882 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3883 } 3884 3885 // if (pre_val != NULL) 3886 __ if_then(pre_val, BoolTest::ne, null()); { 3887 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3888 3889 // is the queue for this thread full? 3890 __ if_then(index, BoolTest::ne, zeroX, likely); { 3891 3892 // decrement the index 3893 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3894 3895 // Now get the buffer location we will log the previous value into and store it 3896 Node *log_addr = __ AddP(no_base, buffer, next_index); 3897 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 3898 // update the index 3899 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 3900 3901 } __ else_(); { 3902 3903 // logging buffer is full, call the runtime 3904 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3905 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3906 } __ end_if(); // (!index) 3907 } __ end_if(); // (pre_val != NULL) 3908 } __ end_if(); // (!marking) 3909 3910 // Final sync IdealKit and GraphKit. 3911 final_sync(ideal); 3912 } 3913 3914 // 3915 // Update the card table and add card address to the queue 3916 // 3917 void GraphKit::g1_mark_card(IdealKit& ideal, 3918 Node* card_adr, 3919 Node* oop_store, 3920 uint oop_alias_idx, 3921 Node* index, 3922 Node* index_adr, 3923 Node* buffer, 3924 const TypeFunc* tf) { 3925 3926 Node* zero = __ ConI(0); 3927 Node* zeroX = __ ConX(0); 3928 Node* no_base = __ top(); 3929 BasicType card_bt = T_BYTE; 3930 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3931 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3932 3933 // Now do the queue work 3934 __ if_then(index, BoolTest::ne, zeroX); { 3935 3936 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3937 Node* log_addr = __ AddP(no_base, buffer, next_index); 3938 3939 // Order, see storeCM. 3940 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 3941 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 3942 3943 } __ else_(); { 3944 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3945 } __ end_if(); 3946 3947 } 3948 3949 void GraphKit::g1_write_barrier_post(Node* oop_store, 3950 Node* obj, 3951 Node* adr, 3952 uint alias_idx, 3953 Node* val, 3954 BasicType bt, 3955 bool use_precise) { 3956 // If we are writing a NULL then we need no post barrier 3957 3958 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3959 // Must be NULL 3960 const Type* t = val->bottom_type(); 3961 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3962 // No post barrier if writing NULLx 3963 return; 3964 } 3965 3966 if (!use_precise) { 3967 // All card marks for a (non-array) instance are in one place: 3968 adr = obj; 3969 } 3970 // (Else it's an array (or unknown), and we want more precise card marks.) 3971 assert(adr != NULL, ""); 3972 3973 IdealKit ideal(this, true); 3974 3975 Node* tls = __ thread(); // ThreadLocalStorage 3976 3977 Node* no_base = __ top(); 3978 float likely = PROB_LIKELY(0.999); 3979 float unlikely = PROB_UNLIKELY(0.999); 3980 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 3981 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 3982 Node* zeroX = __ ConX(0); 3983 3984 // Get the alias_index for raw card-mark memory 3985 const TypePtr* card_type = TypeRawPtr::BOTTOM; 3986 3987 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 3988 3989 // Offsets into the thread 3990 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3991 PtrQueue::byte_offset_of_index()); 3992 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 3993 PtrQueue::byte_offset_of_buf()); 3994 3995 // Pointers into the thread 3996 3997 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3998 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3999 4000 // Now some values 4001 // Use ctrl to avoid hoisting these values past a safepoint, which could 4002 // potentially reset these fields in the JavaThread. 4003 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4004 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4005 4006 // Convert the store obj pointer to an int prior to doing math on it 4007 // Must use ctrl to prevent "integerized oop" existing across safepoint 4008 Node* cast = __ CastPX(__ ctrl(), adr); 4009 4010 // Divide pointer by card size 4011 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4012 4013 // Combine card table base and card offset 4014 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4015 4016 // If we know the value being stored does it cross regions? 4017 4018 if (val != NULL) { 4019 // Does the store cause us to cross regions? 4020 4021 // Should be able to do an unsigned compare of region_size instead of 4022 // and extra shift. Do we have an unsigned compare?? 4023 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4024 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4025 4026 // if (xor_res == 0) same region so skip 4027 __ if_then(xor_res, BoolTest::ne, zeroX); { 4028 4029 // No barrier if we are storing a NULL 4030 __ if_then(val, BoolTest::ne, null(), unlikely); { 4031 4032 // Ok must mark the card if not already dirty 4033 4034 // load the original value of the card 4035 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4036 4037 __ if_then(card_val, BoolTest::ne, young_card); { 4038 sync_kit(ideal); 4039 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4040 insert_mem_bar(Op_MemBarVolatile, oop_store); 4041 __ sync_kit(this); 4042 4043 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4044 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4045 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4046 } __ end_if(); 4047 } __ end_if(); 4048 } __ end_if(); 4049 } __ end_if(); 4050 } else { 4051 // Object.clone() instrinsic uses this path. 4052 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4053 } 4054 4055 // Final sync IdealKit and GraphKit. 4056 final_sync(ideal); 4057 } 4058 #undef __ 4059 4060 4061 4062 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4063 if (java_lang_String::has_offset_field()) { 4064 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4065 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4066 false, NULL, 0); 4067 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4068 int offset_field_idx = C->get_alias_index(offset_field_type); 4069 return make_load(ctrl, 4070 basic_plus_adr(str, str, offset_offset), 4071 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4072 } else { 4073 return intcon(0); 4074 } 4075 } 4076 4077 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4078 if (java_lang_String::has_count_field()) { 4079 int count_offset = java_lang_String::count_offset_in_bytes(); 4080 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4081 false, NULL, 0); 4082 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4083 int count_field_idx = C->get_alias_index(count_field_type); 4084 return make_load(ctrl, 4085 basic_plus_adr(str, str, count_offset), 4086 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4087 } else { 4088 return load_array_length(load_String_value(ctrl, str)); 4089 } 4090 } 4091 4092 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4093 int value_offset = java_lang_String::value_offset_in_bytes(); 4094 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4095 false, NULL, 0); 4096 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4097 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4098 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4099 ciTypeArrayKlass::make(T_CHAR), true, 0); 4100 int value_field_idx = C->get_alias_index(value_field_type); 4101 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4102 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4103 // String.value field is known to be @Stable. 4104 if (UseImplicitStableValues) { 4105 load = cast_array_to_stable(load, value_type); 4106 } 4107 return load; 4108 } 4109 4110 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4111 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4112 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4113 false, NULL, 0); 4114 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4115 int offset_field_idx = C->get_alias_index(offset_field_type); 4116 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4117 value, T_INT, offset_field_idx, MemNode::unordered); 4118 } 4119 4120 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4121 int value_offset = java_lang_String::value_offset_in_bytes(); 4122 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4123 false, NULL, 0); 4124 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4125 4126 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4127 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4128 } 4129 4130 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4131 int count_offset = java_lang_String::count_offset_in_bytes(); 4132 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4133 false, NULL, 0); 4134 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4135 int count_field_idx = C->get_alias_index(count_field_type); 4136 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4137 value, T_INT, count_field_idx, MemNode::unordered); 4138 } 4139 4140 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4141 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4142 // assumption of CCP analysis. 4143 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); 4144 }