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 "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 47 //----------------------------GraphKit----------------------------------------- 48 // Main utility constructor. 49 GraphKit::GraphKit(JVMState* jvms) 50 : Phase(Phase::Parser), 51 _env(C->env()), 52 _gvn(*C->initial_gvn()) 53 { 54 _exceptions = jvms->map()->next_exception(); 55 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 56 set_jvms(jvms); 57 } 58 59 // Private constructor for parser. 60 GraphKit::GraphKit() 61 : Phase(Phase::Parser), 62 _env(C->env()), 63 _gvn(*C->initial_gvn()) 64 { 65 _exceptions = NULL; 66 set_map(NULL); 67 debug_only(_sp = -99); 68 debug_only(set_bci(-99)); 69 } 70 71 72 73 //---------------------------clean_stack--------------------------------------- 74 // Clear away rubbish from the stack area of the JVM state. 75 // This destroys any arguments that may be waiting on the stack. 76 void GraphKit::clean_stack(int from_sp) { 77 SafePointNode* map = this->map(); 78 JVMState* jvms = this->jvms(); 79 int stk_size = jvms->stk_size(); 80 int stkoff = jvms->stkoff(); 81 Node* top = this->top(); 82 for (int i = from_sp; i < stk_size; i++) { 83 if (map->in(stkoff + i) != top) { 84 map->set_req(stkoff + i, top); 85 } 86 } 87 } 88 89 90 //--------------------------------sync_jvms----------------------------------- 91 // Make sure our current jvms agrees with our parse state. 92 JVMState* GraphKit::sync_jvms() const { 93 JVMState* jvms = this->jvms(); 94 jvms->set_bci(bci()); // Record the new bci in the JVMState 95 jvms->set_sp(sp()); // Record the new sp in the JVMState 96 assert(jvms_in_sync(), "jvms is now in sync"); 97 return jvms; 98 } 99 100 //--------------------------------sync_jvms_for_reexecute--------------------- 101 // Make sure our current jvms agrees with our parse state. This version 102 // uses the reexecute_sp for reexecuting bytecodes. 103 JVMState* GraphKit::sync_jvms_for_reexecute() { 104 JVMState* jvms = this->jvms(); 105 jvms->set_bci(bci()); // Record the new bci in the JVMState 106 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 107 return jvms; 108 } 109 110 #ifdef ASSERT 111 bool GraphKit::jvms_in_sync() const { 112 Parse* parse = is_Parse(); 113 if (parse == NULL) { 114 if (bci() != jvms()->bci()) return false; 115 if (sp() != (int)jvms()->sp()) return false; 116 return true; 117 } 118 if (jvms()->method() != parse->method()) return false; 119 if (jvms()->bci() != parse->bci()) return false; 120 int jvms_sp = jvms()->sp(); 121 if (jvms_sp != parse->sp()) return false; 122 int jvms_depth = jvms()->depth(); 123 if (jvms_depth != parse->depth()) return false; 124 return true; 125 } 126 127 // Local helper checks for special internal merge points 128 // used to accumulate and merge exception states. 129 // They are marked by the region's in(0) edge being the map itself. 130 // Such merge points must never "escape" into the parser at large, 131 // until they have been handed to gvn.transform. 132 static bool is_hidden_merge(Node* reg) { 133 if (reg == NULL) return false; 134 if (reg->is_Phi()) { 135 reg = reg->in(0); 136 if (reg == NULL) return false; 137 } 138 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 139 } 140 141 void GraphKit::verify_map() const { 142 if (map() == NULL) return; // null map is OK 143 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 144 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 145 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 146 } 147 148 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 149 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 150 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 151 } 152 #endif 153 154 //---------------------------stop_and_kill_map--------------------------------- 155 // Set _map to NULL, signalling a stop to further bytecode execution. 156 // First smash the current map's control to a constant, to mark it dead. 157 void GraphKit::stop_and_kill_map() { 158 SafePointNode* dead_map = stop(); 159 if (dead_map != NULL) { 160 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 161 assert(dead_map->is_killed(), "must be so marked"); 162 } 163 } 164 165 166 //--------------------------------stopped-------------------------------------- 167 // Tell if _map is NULL, or control is top. 168 bool GraphKit::stopped() { 169 if (map() == NULL) return true; 170 else if (control() == top()) return true; 171 else return false; 172 } 173 174 175 //-----------------------------has_ex_handler---------------------------------- 176 // Tell if this method or any caller method has exception handlers. 177 bool GraphKit::has_ex_handler() { 178 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 179 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 180 return true; 181 } 182 } 183 return false; 184 } 185 186 //------------------------------save_ex_oop------------------------------------ 187 // Save an exception without blowing stack contents or other JVM state. 188 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 189 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 190 ex_map->add_req(ex_oop); 191 debug_only(verify_exception_state(ex_map)); 192 } 193 194 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 195 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 196 Node* ex_oop = ex_map->in(ex_map->req()-1); 197 if (clear_it) ex_map->del_req(ex_map->req()-1); 198 return ex_oop; 199 } 200 201 //-----------------------------saved_ex_oop------------------------------------ 202 // Recover a saved exception from its map. 203 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 204 return common_saved_ex_oop(ex_map, false); 205 } 206 207 //--------------------------clear_saved_ex_oop--------------------------------- 208 // Erase a previously saved exception from its map. 209 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, true); 211 } 212 213 #ifdef ASSERT 214 //---------------------------has_saved_ex_oop---------------------------------- 215 // Erase a previously saved exception from its map. 216 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 217 return ex_map->req() == ex_map->jvms()->endoff()+1; 218 } 219 #endif 220 221 //-------------------------make_exception_state-------------------------------- 222 // Turn the current JVM state into an exception state, appending the ex_oop. 223 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 224 sync_jvms(); 225 SafePointNode* ex_map = stop(); // do not manipulate this map any more 226 set_saved_ex_oop(ex_map, ex_oop); 227 return ex_map; 228 } 229 230 231 //--------------------------add_exception_state-------------------------------- 232 // Add an exception to my list of exceptions. 233 void GraphKit::add_exception_state(SafePointNode* ex_map) { 234 if (ex_map == NULL || ex_map->control() == top()) { 235 return; 236 } 237 #ifdef ASSERT 238 verify_exception_state(ex_map); 239 if (has_exceptions()) { 240 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 241 } 242 #endif 243 244 // If there is already an exception of exactly this type, merge with it. 245 // In particular, null-checks and other low-level exceptions common up here. 246 Node* ex_oop = saved_ex_oop(ex_map); 247 const Type* ex_type = _gvn.type(ex_oop); 248 if (ex_oop == top()) { 249 // No action needed. 250 return; 251 } 252 assert(ex_type->isa_instptr(), "exception must be an instance"); 253 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 254 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 255 // We check sp also because call bytecodes can generate exceptions 256 // both before and after arguments are popped! 257 if (ex_type2 == ex_type 258 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 259 combine_exception_states(ex_map, e2); 260 return; 261 } 262 } 263 264 // No pre-existing exception of the same type. Chain it on the list. 265 push_exception_state(ex_map); 266 } 267 268 //-----------------------add_exception_states_from----------------------------- 269 void GraphKit::add_exception_states_from(JVMState* jvms) { 270 SafePointNode* ex_map = jvms->map()->next_exception(); 271 if (ex_map != NULL) { 272 jvms->map()->set_next_exception(NULL); 273 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 274 next_map = ex_map->next_exception(); 275 ex_map->set_next_exception(NULL); 276 add_exception_state(ex_map); 277 } 278 } 279 } 280 281 //-----------------------transfer_exceptions_into_jvms------------------------- 282 JVMState* GraphKit::transfer_exceptions_into_jvms() { 283 if (map() == NULL) { 284 // We need a JVMS to carry the exceptions, but the map has gone away. 285 // Create a scratch JVMS, cloned from any of the exception states... 286 if (has_exceptions()) { 287 _map = _exceptions; 288 _map = clone_map(); 289 _map->set_next_exception(NULL); 290 clear_saved_ex_oop(_map); 291 debug_only(verify_map()); 292 } else { 293 // ...or created from scratch 294 JVMState* jvms = new (C) JVMState(_method, NULL); 295 jvms->set_bci(_bci); 296 jvms->set_sp(_sp); 297 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 298 set_jvms(jvms); 299 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 300 set_all_memory(top()); 301 while (map()->req() < jvms->endoff()) map()->add_req(top()); 302 } 303 // (This is a kludge, in case you didn't notice.) 304 set_control(top()); 305 } 306 JVMState* jvms = sync_jvms(); 307 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 308 jvms->map()->set_next_exception(_exceptions); 309 _exceptions = NULL; // done with this set of exceptions 310 return jvms; 311 } 312 313 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 314 assert(is_hidden_merge(dstphi), "must be a special merge node"); 315 assert(is_hidden_merge(srcphi), "must be a special merge node"); 316 uint limit = srcphi->req(); 317 for (uint i = PhiNode::Input; i < limit; i++) { 318 dstphi->add_req(srcphi->in(i)); 319 } 320 } 321 static inline void add_one_req(Node* dstphi, Node* src) { 322 assert(is_hidden_merge(dstphi), "must be a special merge node"); 323 assert(!is_hidden_merge(src), "must not be a special merge node"); 324 dstphi->add_req(src); 325 } 326 327 //-----------------------combine_exception_states------------------------------ 328 // This helper function combines exception states by building phis on a 329 // specially marked state-merging region. These regions and phis are 330 // untransformed, and can build up gradually. The region is marked by 331 // having a control input of its exception map, rather than NULL. Such 332 // regions do not appear except in this function, and in use_exception_state. 333 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 334 if (failing()) return; // dying anyway... 335 JVMState* ex_jvms = ex_map->_jvms; 336 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 337 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 338 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 339 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 340 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 341 assert(ex_map->req() == phi_map->req(), "matching maps"); 342 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 343 Node* hidden_merge_mark = root(); 344 Node* region = phi_map->control(); 345 MergeMemNode* phi_mem = phi_map->merged_memory(); 346 MergeMemNode* ex_mem = ex_map->merged_memory(); 347 if (region->in(0) != hidden_merge_mark) { 348 // The control input is not (yet) a specially-marked region in phi_map. 349 // Make it so, and build some phis. 350 region = new RegionNode(2); 351 _gvn.set_type(region, Type::CONTROL); 352 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 353 region->init_req(1, phi_map->control()); 354 phi_map->set_control(region); 355 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 356 record_for_igvn(io_phi); 357 _gvn.set_type(io_phi, Type::ABIO); 358 phi_map->set_i_o(io_phi); 359 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 360 Node* m = mms.memory(); 361 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 362 record_for_igvn(m_phi); 363 _gvn.set_type(m_phi, Type::MEMORY); 364 mms.set_memory(m_phi); 365 } 366 } 367 368 // Either or both of phi_map and ex_map might already be converted into phis. 369 Node* ex_control = ex_map->control(); 370 // if there is special marking on ex_map also, we add multiple edges from src 371 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 372 // how wide was the destination phi_map, originally? 373 uint orig_width = region->req(); 374 375 if (add_multiple) { 376 add_n_reqs(region, ex_control); 377 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 378 } else { 379 // ex_map has no merges, so we just add single edges everywhere 380 add_one_req(region, ex_control); 381 add_one_req(phi_map->i_o(), ex_map->i_o()); 382 } 383 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 384 if (mms.is_empty()) { 385 // get a copy of the base memory, and patch some inputs into it 386 const TypePtr* adr_type = mms.adr_type(C); 387 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 388 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 389 mms.set_memory(phi); 390 // Prepare to append interesting stuff onto the newly sliced phi: 391 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 392 } 393 // Append stuff from ex_map: 394 if (add_multiple) { 395 add_n_reqs(mms.memory(), mms.memory2()); 396 } else { 397 add_one_req(mms.memory(), mms.memory2()); 398 } 399 } 400 uint limit = ex_map->req(); 401 for (uint i = TypeFunc::Parms; i < limit; i++) { 402 // Skip everything in the JVMS after tos. (The ex_oop follows.) 403 if (i == tos) i = ex_jvms->monoff(); 404 Node* src = ex_map->in(i); 405 Node* dst = phi_map->in(i); 406 if (src != dst) { 407 PhiNode* phi; 408 if (dst->in(0) != region) { 409 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 410 record_for_igvn(phi); 411 _gvn.set_type(phi, phi->type()); 412 phi_map->set_req(i, dst); 413 // Prepare to append interesting stuff onto the new phi: 414 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 415 } else { 416 assert(dst->is_Phi(), "nobody else uses a hidden region"); 417 phi = dst->as_Phi(); 418 } 419 if (add_multiple && src->in(0) == ex_control) { 420 // Both are phis. 421 add_n_reqs(dst, src); 422 } else { 423 while (dst->req() < region->req()) add_one_req(dst, src); 424 } 425 const Type* srctype = _gvn.type(src); 426 if (phi->type() != srctype) { 427 const Type* dsttype = phi->type()->meet_speculative(srctype); 428 if (phi->type() != dsttype) { 429 phi->set_type(dsttype); 430 _gvn.set_type(phi, dsttype); 431 } 432 } 433 } 434 } 435 } 436 437 //--------------------------use_exception_state-------------------------------- 438 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 439 if (failing()) { stop(); return top(); } 440 Node* region = phi_map->control(); 441 Node* hidden_merge_mark = root(); 442 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 443 Node* ex_oop = clear_saved_ex_oop(phi_map); 444 if (region->in(0) == hidden_merge_mark) { 445 // Special marking for internal ex-states. Process the phis now. 446 region->set_req(0, region); // now it's an ordinary region 447 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 448 // Note: Setting the jvms also sets the bci and sp. 449 set_control(_gvn.transform(region)); 450 uint tos = jvms()->stkoff() + sp(); 451 for (uint i = 1; i < tos; i++) { 452 Node* x = phi_map->in(i); 453 if (x->in(0) == region) { 454 assert(x->is_Phi(), "expected a special phi"); 455 phi_map->set_req(i, _gvn.transform(x)); 456 } 457 } 458 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 459 Node* x = mms.memory(); 460 if (x->in(0) == region) { 461 assert(x->is_Phi(), "nobody else uses a hidden region"); 462 mms.set_memory(_gvn.transform(x)); 463 } 464 } 465 if (ex_oop->in(0) == region) { 466 assert(ex_oop->is_Phi(), "expected a special phi"); 467 ex_oop = _gvn.transform(ex_oop); 468 } 469 } else { 470 set_jvms(phi_map->jvms()); 471 } 472 473 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 474 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 475 return ex_oop; 476 } 477 478 //---------------------------------java_bc------------------------------------- 479 Bytecodes::Code GraphKit::java_bc() const { 480 ciMethod* method = this->method(); 481 int bci = this->bci(); 482 if (method != NULL && bci != InvocationEntryBci) 483 return method->java_code_at_bci(bci); 484 else 485 return Bytecodes::_illegal; 486 } 487 488 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 489 bool must_throw) { 490 // if the exception capability is set, then we will generate code 491 // to check the JavaThread.should_post_on_exceptions flag to see 492 // if we actually need to report exception events (for this 493 // thread). If we don't need to report exception events, we will 494 // take the normal fast path provided by add_exception_events. If 495 // exception event reporting is enabled for this thread, we will 496 // take the uncommon_trap in the BuildCutout below. 497 498 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 499 Node* jthread = _gvn.transform(new ThreadLocalNode()); 500 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 501 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 502 503 // Test the should_post_on_exceptions_flag vs. 0 504 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 505 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 506 507 // Branch to slow_path if should_post_on_exceptions_flag was true 508 { BuildCutout unless(this, tst, PROB_MAX); 509 // Do not try anything fancy if we're notifying the VM on every throw. 510 // Cf. case Bytecodes::_athrow in parse2.cpp. 511 uncommon_trap(reason, Deoptimization::Action_none, 512 (ciKlass*)NULL, (char*)NULL, must_throw); 513 } 514 515 } 516 517 //------------------------------builtin_throw---------------------------------- 518 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 519 bool must_throw = true; 520 521 if (env()->jvmti_can_post_on_exceptions()) { 522 // check if we must post exception events, take uncommon trap if so 523 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 524 // here if should_post_on_exceptions is false 525 // continue on with the normal codegen 526 } 527 528 // If this particular condition has not yet happened at this 529 // bytecode, then use the uncommon trap mechanism, and allow for 530 // a future recompilation if several traps occur here. 531 // If the throw is hot, try to use a more complicated inline mechanism 532 // which keeps execution inside the compiled code. 533 bool treat_throw_as_hot = false; 534 ciMethodData* md = method()->method_data(); 535 536 if (ProfileTraps) { 537 if (too_many_traps(reason)) { 538 treat_throw_as_hot = true; 539 } 540 // (If there is no MDO at all, assume it is early in 541 // execution, and that any deopts are part of the 542 // startup transient, and don't need to be remembered.) 543 544 // Also, if there is a local exception handler, treat all throws 545 // as hot if there has been at least one in this method. 546 if (C->trap_count(reason) != 0 547 && method()->method_data()->trap_count(reason) != 0 548 && has_ex_handler()) { 549 treat_throw_as_hot = true; 550 } 551 } 552 553 // If this throw happens frequently, an uncommon trap might cause 554 // a performance pothole. If there is a local exception handler, 555 // and if this particular bytecode appears to be deoptimizing often, 556 // let us handle the throw inline, with a preconstructed instance. 557 // Note: If the deopt count has blown up, the uncommon trap 558 // runtime is going to flush this nmethod, not matter what. 559 if (treat_throw_as_hot 560 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 561 // If the throw is local, we use a pre-existing instance and 562 // punt on the backtrace. This would lead to a missing backtrace 563 // (a repeat of 4292742) if the backtrace object is ever asked 564 // for its backtrace. 565 // Fixing this remaining case of 4292742 requires some flavor of 566 // escape analysis. Leave that for the future. 567 ciInstance* ex_obj = NULL; 568 switch (reason) { 569 case Deoptimization::Reason_null_check: 570 ex_obj = env()->NullPointerException_instance(); 571 break; 572 case Deoptimization::Reason_div0_check: 573 ex_obj = env()->ArithmeticException_instance(); 574 break; 575 case Deoptimization::Reason_range_check: 576 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 577 break; 578 case Deoptimization::Reason_class_check: 579 if (java_bc() == Bytecodes::_aastore) { 580 ex_obj = env()->ArrayStoreException_instance(); 581 } else { 582 ex_obj = env()->ClassCastException_instance(); 583 } 584 break; 585 } 586 if (failing()) { stop(); return; } // exception allocation might fail 587 if (ex_obj != NULL) { 588 // Cheat with a preallocated exception object. 589 if (C->log() != NULL) 590 C->log()->elem("hot_throw preallocated='1' reason='%s'", 591 Deoptimization::trap_reason_name(reason)); 592 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 593 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 594 595 // Clear the detail message of the preallocated exception object. 596 // Weblogic sometimes mutates the detail message of exceptions 597 // using reflection. 598 int offset = java_lang_Throwable::get_detailMessage_offset(); 599 const TypePtr* adr_typ = ex_con->add_offset(offset); 600 601 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 602 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 603 // Conservatively release stores of object references. 604 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 605 606 add_exception_state(make_exception_state(ex_node)); 607 return; 608 } 609 } 610 611 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 612 // It won't be much cheaper than bailing to the interp., since we'll 613 // have to pass up all the debug-info, and the runtime will have to 614 // create the stack trace. 615 616 // Usual case: Bail to interpreter. 617 // Reserve the right to recompile if we haven't seen anything yet. 618 619 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 620 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 621 if (treat_throw_as_hot 622 && (method()->method_data()->trap_recompiled_at(bci(), m) 623 || C->too_many_traps(reason))) { 624 // We cannot afford to take more traps here. Suffer in the interpreter. 625 if (C->log() != NULL) 626 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 627 Deoptimization::trap_reason_name(reason), 628 C->trap_count(reason)); 629 action = Deoptimization::Action_none; 630 } 631 632 // "must_throw" prunes the JVM state to include only the stack, if there 633 // are no local exception handlers. This should cut down on register 634 // allocation time and code size, by drastically reducing the number 635 // of in-edges on the call to the uncommon trap. 636 637 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 638 } 639 640 641 //----------------------------PreserveJVMState--------------------------------- 642 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 643 debug_only(kit->verify_map()); 644 _kit = kit; 645 _map = kit->map(); // preserve the map 646 _sp = kit->sp(); 647 kit->set_map(clone_map ? kit->clone_map() : NULL); 648 Compile::current()->inc_preserve_jvm_state(); 649 #ifdef ASSERT 650 _bci = kit->bci(); 651 Parse* parser = kit->is_Parse(); 652 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 653 _block = block; 654 #endif 655 } 656 PreserveJVMState::~PreserveJVMState() { 657 GraphKit* kit = _kit; 658 #ifdef ASSERT 659 assert(kit->bci() == _bci, "bci must not shift"); 660 Parse* parser = kit->is_Parse(); 661 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 662 assert(block == _block, "block must not shift"); 663 #endif 664 kit->set_map(_map); 665 kit->set_sp(_sp); 666 Compile::current()->dec_preserve_jvm_state(); 667 } 668 669 670 //-----------------------------BuildCutout------------------------------------- 671 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 672 : PreserveJVMState(kit) 673 { 674 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 675 SafePointNode* outer_map = _map; // preserved map is caller's 676 SafePointNode* inner_map = kit->map(); 677 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 678 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 679 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 680 } 681 BuildCutout::~BuildCutout() { 682 GraphKit* kit = _kit; 683 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 684 } 685 686 //---------------------------PreserveReexecuteState---------------------------- 687 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 688 assert(!kit->stopped(), "must call stopped() before"); 689 _kit = kit; 690 _sp = kit->sp(); 691 _reexecute = kit->jvms()->_reexecute; 692 } 693 PreserveReexecuteState::~PreserveReexecuteState() { 694 if (_kit->stopped()) return; 695 _kit->jvms()->_reexecute = _reexecute; 696 _kit->set_sp(_sp); 697 } 698 699 //------------------------------clone_map-------------------------------------- 700 // Implementation of PreserveJVMState 701 // 702 // Only clone_map(...) here. If this function is only used in the 703 // PreserveJVMState class we may want to get rid of this extra 704 // function eventually and do it all there. 705 706 SafePointNode* GraphKit::clone_map() { 707 if (map() == NULL) return NULL; 708 709 // Clone the memory edge first 710 Node* mem = MergeMemNode::make(C, map()->memory()); 711 gvn().set_type_bottom(mem); 712 713 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 714 JVMState* jvms = this->jvms(); 715 JVMState* clonejvms = jvms->clone_shallow(C); 716 clonemap->set_memory(mem); 717 clonemap->set_jvms(clonejvms); 718 clonejvms->set_map(clonemap); 719 record_for_igvn(clonemap); 720 gvn().set_type_bottom(clonemap); 721 return clonemap; 722 } 723 724 725 //-----------------------------set_map_clone----------------------------------- 726 void GraphKit::set_map_clone(SafePointNode* m) { 727 _map = m; 728 _map = clone_map(); 729 _map->set_next_exception(NULL); 730 debug_only(verify_map()); 731 } 732 733 734 //----------------------------kill_dead_locals--------------------------------- 735 // Detect any locals which are known to be dead, and force them to top. 736 void GraphKit::kill_dead_locals() { 737 // Consult the liveness information for the locals. If any 738 // of them are unused, then they can be replaced by top(). This 739 // should help register allocation time and cut down on the size 740 // of the deoptimization information. 741 742 // This call is made from many of the bytecode handling 743 // subroutines called from the Big Switch in do_one_bytecode. 744 // Every bytecode which might include a slow path is responsible 745 // for killing its dead locals. The more consistent we 746 // are about killing deads, the fewer useless phis will be 747 // constructed for them at various merge points. 748 749 // bci can be -1 (InvocationEntryBci). We return the entry 750 // liveness for the method. 751 752 if (method() == NULL || method()->code_size() == 0) { 753 // We are building a graph for a call to a native method. 754 // All locals are live. 755 return; 756 } 757 758 ResourceMark rm; 759 760 // Consult the liveness information for the locals. If any 761 // of them are unused, then they can be replaced by top(). This 762 // should help register allocation time and cut down on the size 763 // of the deoptimization information. 764 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 765 766 int len = (int)live_locals.size(); 767 assert(len <= jvms()->loc_size(), "too many live locals"); 768 for (int local = 0; local < len; local++) { 769 if (!live_locals.at(local)) { 770 set_local(local, top()); 771 } 772 } 773 } 774 775 #ifdef ASSERT 776 //-------------------------dead_locals_are_killed------------------------------ 777 // Return true if all dead locals are set to top in the map. 778 // Used to assert "clean" debug info at various points. 779 bool GraphKit::dead_locals_are_killed() { 780 if (method() == NULL || method()->code_size() == 0) { 781 // No locals need to be dead, so all is as it should be. 782 return true; 783 } 784 785 // Make sure somebody called kill_dead_locals upstream. 786 ResourceMark rm; 787 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 788 if (jvms->loc_size() == 0) continue; // no locals to consult 789 SafePointNode* map = jvms->map(); 790 ciMethod* method = jvms->method(); 791 int bci = jvms->bci(); 792 if (jvms == this->jvms()) { 793 bci = this->bci(); // it might not yet be synched 794 } 795 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 796 int len = (int)live_locals.size(); 797 if (!live_locals.is_valid() || len == 0) 798 // This method is trivial, or is poisoned by a breakpoint. 799 return true; 800 assert(len == jvms->loc_size(), "live map consistent with locals map"); 801 for (int local = 0; local < len; local++) { 802 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 803 if (PrintMiscellaneous && (Verbose || WizardMode)) { 804 tty->print_cr("Zombie local %d: ", local); 805 jvms->dump(); 806 } 807 return false; 808 } 809 } 810 } 811 return true; 812 } 813 814 #endif //ASSERT 815 816 // Helper function for enforcing certain bytecodes to reexecute if 817 // deoptimization happens 818 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 819 ciMethod* cur_method = jvms->method(); 820 int cur_bci = jvms->bci(); 821 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 822 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 823 return Interpreter::bytecode_should_reexecute(code) || 824 is_anewarray && code == Bytecodes::_multianewarray; 825 // Reexecute _multianewarray bytecode which was replaced with 826 // sequence of [a]newarray. See Parse::do_multianewarray(). 827 // 828 // Note: interpreter should not have it set since this optimization 829 // is limited by dimensions and guarded by flag so in some cases 830 // multianewarray() runtime calls will be generated and 831 // the bytecode should not be reexecutes (stack will not be reset). 832 } else 833 return false; 834 } 835 836 // Helper function for adding JVMState and debug information to node 837 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 838 // Add the safepoint edges to the call (or other safepoint). 839 840 // Make sure dead locals are set to top. This 841 // should help register allocation time and cut down on the size 842 // of the deoptimization information. 843 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 844 845 // Walk the inline list to fill in the correct set of JVMState's 846 // Also fill in the associated edges for each JVMState. 847 848 // If the bytecode needs to be reexecuted we need to put 849 // the arguments back on the stack. 850 const bool should_reexecute = jvms()->should_reexecute(); 851 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 852 853 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 854 // undefined if the bci is different. This is normal for Parse but it 855 // should not happen for LibraryCallKit because only one bci is processed. 856 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 857 "in LibraryCallKit the reexecute bit should not change"); 858 859 // If we are guaranteed to throw, we can prune everything but the 860 // input to the current bytecode. 861 bool can_prune_locals = false; 862 uint stack_slots_not_pruned = 0; 863 int inputs = 0, depth = 0; 864 if (must_throw) { 865 assert(method() == youngest_jvms->method(), "sanity"); 866 if (compute_stack_effects(inputs, depth)) { 867 can_prune_locals = true; 868 stack_slots_not_pruned = inputs; 869 } 870 } 871 872 if (env()->should_retain_local_variables()) { 873 // At any safepoint, this method can get breakpointed, which would 874 // then require an immediate deoptimization. 875 can_prune_locals = false; // do not prune locals 876 stack_slots_not_pruned = 0; 877 } 878 879 // do not scribble on the input jvms 880 JVMState* out_jvms = youngest_jvms->clone_deep(C); 881 call->set_jvms(out_jvms); // Start jvms list for call node 882 883 // For a known set of bytecodes, the interpreter should reexecute them if 884 // deoptimization happens. We set the reexecute state for them here 885 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 886 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 887 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 888 } 889 890 // Presize the call: 891 DEBUG_ONLY(uint non_debug_edges = call->req()); 892 call->add_req_batch(top(), youngest_jvms->debug_depth()); 893 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 894 895 // Set up edges so that the call looks like this: 896 // Call [state:] ctl io mem fptr retadr 897 // [parms:] parm0 ... parmN 898 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 899 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 900 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 901 // Note that caller debug info precedes callee debug info. 902 903 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 904 uint debug_ptr = call->req(); 905 906 // Loop over the map input edges associated with jvms, add them 907 // to the call node, & reset all offsets to match call node array. 908 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 909 uint debug_end = debug_ptr; 910 uint debug_start = debug_ptr - in_jvms->debug_size(); 911 debug_ptr = debug_start; // back up the ptr 912 913 uint p = debug_start; // walks forward in [debug_start, debug_end) 914 uint j, k, l; 915 SafePointNode* in_map = in_jvms->map(); 916 out_jvms->set_map(call); 917 918 if (can_prune_locals) { 919 assert(in_jvms->method() == out_jvms->method(), "sanity"); 920 // If the current throw can reach an exception handler in this JVMS, 921 // then we must keep everything live that can reach that handler. 922 // As a quick and dirty approximation, we look for any handlers at all. 923 if (in_jvms->method()->has_exception_handlers()) { 924 can_prune_locals = false; 925 } 926 } 927 928 // Add the Locals 929 k = in_jvms->locoff(); 930 l = in_jvms->loc_size(); 931 out_jvms->set_locoff(p); 932 if (!can_prune_locals) { 933 for (j = 0; j < l; j++) 934 call->set_req(p++, in_map->in(k+j)); 935 } else { 936 p += l; // already set to top above by add_req_batch 937 } 938 939 // Add the Expression Stack 940 k = in_jvms->stkoff(); 941 l = in_jvms->sp(); 942 out_jvms->set_stkoff(p); 943 if (!can_prune_locals) { 944 for (j = 0; j < l; j++) 945 call->set_req(p++, in_map->in(k+j)); 946 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 947 // Divide stack into {S0,...,S1}, where S0 is set to top. 948 uint s1 = stack_slots_not_pruned; 949 stack_slots_not_pruned = 0; // for next iteration 950 if (s1 > l) s1 = l; 951 uint s0 = l - s1; 952 p += s0; // skip the tops preinstalled by add_req_batch 953 for (j = s0; j < l; j++) 954 call->set_req(p++, in_map->in(k+j)); 955 } else { 956 p += l; // already set to top above by add_req_batch 957 } 958 959 // Add the Monitors 960 k = in_jvms->monoff(); 961 l = in_jvms->mon_size(); 962 out_jvms->set_monoff(p); 963 for (j = 0; j < l; j++) 964 call->set_req(p++, in_map->in(k+j)); 965 966 // Copy any scalar object fields. 967 k = in_jvms->scloff(); 968 l = in_jvms->scl_size(); 969 out_jvms->set_scloff(p); 970 for (j = 0; j < l; j++) 971 call->set_req(p++, in_map->in(k+j)); 972 973 // Finish the new jvms. 974 out_jvms->set_endoff(p); 975 976 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 977 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 978 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 979 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 980 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 981 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 982 983 // Update the two tail pointers in parallel. 984 out_jvms = out_jvms->caller(); 985 in_jvms = in_jvms->caller(); 986 } 987 988 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 989 990 // Test the correctness of JVMState::debug_xxx accessors: 991 assert(call->jvms()->debug_start() == non_debug_edges, ""); 992 assert(call->jvms()->debug_end() == call->req(), ""); 993 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 994 } 995 996 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 997 Bytecodes::Code code = java_bc(); 998 if (code == Bytecodes::_wide) { 999 code = method()->java_code_at_bci(bci() + 1); 1000 } 1001 1002 BasicType rtype = T_ILLEGAL; 1003 int rsize = 0; 1004 1005 if (code != Bytecodes::_illegal) { 1006 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1007 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1008 if (rtype < T_CONFLICT) 1009 rsize = type2size[rtype]; 1010 } 1011 1012 switch (code) { 1013 case Bytecodes::_illegal: 1014 return false; 1015 1016 case Bytecodes::_ldc: 1017 case Bytecodes::_ldc_w: 1018 case Bytecodes::_ldc2_w: 1019 inputs = 0; 1020 break; 1021 1022 case Bytecodes::_dup: inputs = 1; break; 1023 case Bytecodes::_dup_x1: inputs = 2; break; 1024 case Bytecodes::_dup_x2: inputs = 3; break; 1025 case Bytecodes::_dup2: inputs = 2; break; 1026 case Bytecodes::_dup2_x1: inputs = 3; break; 1027 case Bytecodes::_dup2_x2: inputs = 4; break; 1028 case Bytecodes::_swap: inputs = 2; break; 1029 case Bytecodes::_arraylength: inputs = 1; break; 1030 1031 case Bytecodes::_getstatic: 1032 case Bytecodes::_putstatic: 1033 case Bytecodes::_getfield: 1034 case Bytecodes::_putfield: 1035 { 1036 bool ignored_will_link; 1037 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1038 int size = field->type()->size(); 1039 bool is_get = (depth >= 0), is_static = (depth & 1); 1040 inputs = (is_static ? 0 : 1); 1041 if (is_get) { 1042 depth = size - inputs; 1043 } else { 1044 inputs += size; // putxxx pops the value from the stack 1045 depth = - inputs; 1046 } 1047 } 1048 break; 1049 1050 case Bytecodes::_invokevirtual: 1051 case Bytecodes::_invokespecial: 1052 case Bytecodes::_invokestatic: 1053 case Bytecodes::_invokedynamic: 1054 case Bytecodes::_invokeinterface: 1055 { 1056 bool ignored_will_link; 1057 ciSignature* declared_signature = NULL; 1058 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1059 assert(declared_signature != NULL, "cannot be null"); 1060 inputs = declared_signature->arg_size_for_bc(code); 1061 int size = declared_signature->return_type()->size(); 1062 depth = size - inputs; 1063 } 1064 break; 1065 1066 case Bytecodes::_multianewarray: 1067 { 1068 ciBytecodeStream iter(method()); 1069 iter.reset_to_bci(bci()); 1070 iter.next(); 1071 inputs = iter.get_dimensions(); 1072 assert(rsize == 1, ""); 1073 depth = rsize - inputs; 1074 } 1075 break; 1076 1077 case Bytecodes::_ireturn: 1078 case Bytecodes::_lreturn: 1079 case Bytecodes::_freturn: 1080 case Bytecodes::_dreturn: 1081 case Bytecodes::_areturn: 1082 assert(rsize = -depth, ""); 1083 inputs = rsize; 1084 break; 1085 1086 case Bytecodes::_jsr: 1087 case Bytecodes::_jsr_w: 1088 inputs = 0; 1089 depth = 1; // S.B. depth=1, not zero 1090 break; 1091 1092 default: 1093 // bytecode produces a typed result 1094 inputs = rsize - depth; 1095 assert(inputs >= 0, ""); 1096 break; 1097 } 1098 1099 #ifdef ASSERT 1100 // spot check 1101 int outputs = depth + inputs; 1102 assert(outputs >= 0, "sanity"); 1103 switch (code) { 1104 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1105 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1106 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1107 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1108 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1109 } 1110 #endif //ASSERT 1111 1112 return true; 1113 } 1114 1115 1116 1117 //------------------------------basic_plus_adr--------------------------------- 1118 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1119 // short-circuit a common case 1120 if (offset == intcon(0)) return ptr; 1121 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1122 } 1123 1124 Node* GraphKit::ConvI2L(Node* offset) { 1125 // short-circuit a common case 1126 jint offset_con = find_int_con(offset, Type::OffsetBot); 1127 if (offset_con != Type::OffsetBot) { 1128 return longcon((jlong) offset_con); 1129 } 1130 return _gvn.transform( new ConvI2LNode(offset)); 1131 } 1132 1133 Node* GraphKit::ConvI2UL(Node* offset) { 1134 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1135 if (offset_con != (juint) Type::OffsetBot) { 1136 return longcon((julong) offset_con); 1137 } 1138 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1139 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) ); 1140 return _gvn.transform( new AndLNode(conv, mask) ); 1141 } 1142 1143 Node* GraphKit::ConvL2I(Node* offset) { 1144 // short-circuit a common case 1145 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1146 if (offset_con != (jlong)Type::OffsetBot) { 1147 return intcon((int) offset_con); 1148 } 1149 return _gvn.transform( new ConvL2INode(offset)); 1150 } 1151 1152 //-------------------------load_object_klass----------------------------------- 1153 Node* GraphKit::load_object_klass(Node* obj) { 1154 // Special-case a fresh allocation to avoid building nodes: 1155 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1156 if (akls != NULL) return akls; 1157 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1158 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1159 } 1160 1161 //-------------------------load_array_length----------------------------------- 1162 Node* GraphKit::load_array_length(Node* array) { 1163 // Special-case a fresh allocation to avoid building nodes: 1164 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1165 Node *alen; 1166 if (alloc == NULL) { 1167 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1168 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1169 } else { 1170 alen = alloc->Ideal_length(); 1171 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1172 if (ccast != alen) { 1173 alen = _gvn.transform(ccast); 1174 } 1175 } 1176 return alen; 1177 } 1178 1179 //------------------------------do_null_check---------------------------------- 1180 // Helper function to do a NULL pointer check. Returned value is 1181 // the incoming address with NULL casted away. You are allowed to use the 1182 // not-null value only if you are control dependent on the test. 1183 extern int explicit_null_checks_inserted, 1184 explicit_null_checks_elided; 1185 Node* GraphKit::null_check_common(Node* value, BasicType type, 1186 // optional arguments for variations: 1187 bool assert_null, 1188 Node* *null_control, 1189 bool speculative) { 1190 assert(!assert_null || null_control == NULL, "not both at once"); 1191 if (stopped()) return top(); 1192 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1193 // For some performance testing, we may wish to suppress null checking. 1194 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1195 return value; 1196 } 1197 explicit_null_checks_inserted++; 1198 1199 // Construct NULL check 1200 Node *chk = NULL; 1201 switch(type) { 1202 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1203 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1204 case T_ARRAY : // fall through 1205 type = T_OBJECT; // simplify further tests 1206 case T_OBJECT : { 1207 const Type *t = _gvn.type( value ); 1208 1209 const TypeOopPtr* tp = t->isa_oopptr(); 1210 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1211 // Only for do_null_check, not any of its siblings: 1212 && !assert_null && null_control == NULL) { 1213 // Usually, any field access or invocation on an unloaded oop type 1214 // will simply fail to link, since the statically linked class is 1215 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1216 // the static class is loaded but the sharper oop type is not. 1217 // Rather than checking for this obscure case in lots of places, 1218 // we simply observe that a null check on an unloaded class 1219 // will always be followed by a nonsense operation, so we 1220 // can just issue the uncommon trap here. 1221 // Our access to the unloaded class will only be correct 1222 // after it has been loaded and initialized, which requires 1223 // a trip through the interpreter. 1224 #ifndef PRODUCT 1225 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1226 #endif 1227 uncommon_trap(Deoptimization::Reason_unloaded, 1228 Deoptimization::Action_reinterpret, 1229 tp->klass(), "!loaded"); 1230 return top(); 1231 } 1232 1233 if (assert_null) { 1234 // See if the type is contained in NULL_PTR. 1235 // If so, then the value is already null. 1236 if (t->higher_equal(TypePtr::NULL_PTR)) { 1237 explicit_null_checks_elided++; 1238 return value; // Elided null assert quickly! 1239 } 1240 } else { 1241 // See if mixing in the NULL pointer changes type. 1242 // If so, then the NULL pointer was not allowed in the original 1243 // type. In other words, "value" was not-null. 1244 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1245 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1246 explicit_null_checks_elided++; 1247 return value; // Elided null check quickly! 1248 } 1249 } 1250 chk = new CmpPNode( value, null() ); 1251 break; 1252 } 1253 1254 default: 1255 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1256 } 1257 assert(chk != NULL, "sanity check"); 1258 chk = _gvn.transform(chk); 1259 1260 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1261 BoolNode *btst = new BoolNode( chk, btest); 1262 Node *tst = _gvn.transform( btst ); 1263 1264 //----------- 1265 // if peephole optimizations occurred, a prior test existed. 1266 // If a prior test existed, maybe it dominates as we can avoid this test. 1267 if (tst != btst && type == T_OBJECT) { 1268 // At this point we want to scan up the CFG to see if we can 1269 // find an identical test (and so avoid this test altogether). 1270 Node *cfg = control(); 1271 int depth = 0; 1272 while( depth < 16 ) { // Limit search depth for speed 1273 if( cfg->Opcode() == Op_IfTrue && 1274 cfg->in(0)->in(1) == tst ) { 1275 // Found prior test. Use "cast_not_null" to construct an identical 1276 // CastPP (and hence hash to) as already exists for the prior test. 1277 // Return that casted value. 1278 if (assert_null) { 1279 replace_in_map(value, null()); 1280 return null(); // do not issue the redundant test 1281 } 1282 Node *oldcontrol = control(); 1283 set_control(cfg); 1284 Node *res = cast_not_null(value); 1285 set_control(oldcontrol); 1286 explicit_null_checks_elided++; 1287 return res; 1288 } 1289 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1290 if (cfg == NULL) break; // Quit at region nodes 1291 depth++; 1292 } 1293 } 1294 1295 //----------- 1296 // Branch to failure if null 1297 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1298 Deoptimization::DeoptReason reason; 1299 if (assert_null) { 1300 reason = Deoptimization::Reason_null_assert; 1301 } else if (type == T_OBJECT) { 1302 reason = Deoptimization::reason_null_check(speculative); 1303 } else { 1304 reason = Deoptimization::Reason_div0_check; 1305 } 1306 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1307 // ciMethodData::has_trap_at will return a conservative -1 if any 1308 // must-be-null assertion has failed. This could cause performance 1309 // problems for a method after its first do_null_assert failure. 1310 // Consider using 'Reason_class_check' instead? 1311 1312 // To cause an implicit null check, we set the not-null probability 1313 // to the maximum (PROB_MAX). For an explicit check the probability 1314 // is set to a smaller value. 1315 if (null_control != NULL || too_many_traps(reason)) { 1316 // probability is less likely 1317 ok_prob = PROB_LIKELY_MAG(3); 1318 } else if (!assert_null && 1319 (ImplicitNullCheckThreshold > 0) && 1320 method() != NULL && 1321 (method()->method_data()->trap_count(reason) 1322 >= (uint)ImplicitNullCheckThreshold)) { 1323 ok_prob = PROB_LIKELY_MAG(3); 1324 } 1325 1326 if (null_control != NULL) { 1327 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1328 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1329 set_control( _gvn.transform( new IfTrueNode(iff))); 1330 if (null_true == top()) 1331 explicit_null_checks_elided++; 1332 (*null_control) = null_true; 1333 } else { 1334 BuildCutout unless(this, tst, ok_prob); 1335 // Check for optimizer eliding test at parse time 1336 if (stopped()) { 1337 // Failure not possible; do not bother making uncommon trap. 1338 explicit_null_checks_elided++; 1339 } else if (assert_null) { 1340 uncommon_trap(reason, 1341 Deoptimization::Action_make_not_entrant, 1342 NULL, "assert_null"); 1343 } else { 1344 replace_in_map(value, zerocon(type)); 1345 builtin_throw(reason); 1346 } 1347 } 1348 1349 // Must throw exception, fall-thru not possible? 1350 if (stopped()) { 1351 return top(); // No result 1352 } 1353 1354 if (assert_null) { 1355 // Cast obj to null on this path. 1356 replace_in_map(value, zerocon(type)); 1357 return zerocon(type); 1358 } 1359 1360 // Cast obj to not-null on this path, if there is no null_control. 1361 // (If there is a null_control, a non-null value may come back to haunt us.) 1362 if (type == T_OBJECT) { 1363 Node* cast = cast_not_null(value, false); 1364 if (null_control == NULL || (*null_control) == top()) 1365 replace_in_map(value, cast); 1366 value = cast; 1367 } 1368 1369 return value; 1370 } 1371 1372 1373 //------------------------------cast_not_null---------------------------------- 1374 // Cast obj to not-null on this path 1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1376 const Type *t = _gvn.type(obj); 1377 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1378 // Object is already not-null? 1379 if( t == t_not_null ) return obj; 1380 1381 Node *cast = new CastPPNode(obj,t_not_null); 1382 cast->init_req(0, control()); 1383 cast = _gvn.transform( cast ); 1384 1385 // Scan for instances of 'obj' in the current JVM mapping. 1386 // These instances are known to be not-null after the test. 1387 if (do_replace_in_map) 1388 replace_in_map(obj, cast); 1389 1390 return cast; // Return casted value 1391 } 1392 1393 1394 //--------------------------replace_in_map------------------------------------- 1395 void GraphKit::replace_in_map(Node* old, Node* neww) { 1396 if (old == neww) { 1397 return; 1398 } 1399 1400 map()->replace_edge(old, neww); 1401 1402 // Note: This operation potentially replaces any edge 1403 // on the map. This includes locals, stack, and monitors 1404 // of the current (innermost) JVM state. 1405 1406 if (!ReplaceInParentMaps) { 1407 return; 1408 } 1409 1410 // PreserveJVMState doesn't do a deep copy so we can't modify 1411 // parents 1412 if (Compile::current()->has_preserve_jvm_state()) { 1413 return; 1414 } 1415 1416 Parse* parser = is_Parse(); 1417 bool progress = true; 1418 Node* ctrl = map()->in(0); 1419 // Follow the chain of parsers and see whether the update can be 1420 // done in the map of callers. We can do the replace for a caller if 1421 // the current control post dominates the control of a caller. 1422 while (parser != NULL && parser->caller() != NULL && progress) { 1423 progress = false; 1424 Node* parent_map = parser->caller()->map(); 1425 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch"); 1426 1427 Node* parent_ctrl = parent_map->in(0); 1428 1429 while (parent_ctrl->is_Region()) { 1430 Node* n = parent_ctrl->as_Region()->is_copy(); 1431 if (n == NULL) { 1432 break; 1433 } 1434 parent_ctrl = n; 1435 } 1436 1437 for (;;) { 1438 if (ctrl == parent_ctrl) { 1439 // update the map of the exits which is the one that will be 1440 // used when compilation resume after inlining 1441 parser->exits().map()->replace_edge(old, neww); 1442 progress = true; 1443 break; 1444 } 1445 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 1446 ctrl = ctrl->in(0)->in(0); 1447 } else if (ctrl->is_Region()) { 1448 Node* n = ctrl->as_Region()->is_copy(); 1449 if (n == NULL) { 1450 break; 1451 } 1452 ctrl = n; 1453 } else { 1454 break; 1455 } 1456 } 1457 1458 parser = parser->parent_parser(); 1459 } 1460 } 1461 1462 1463 //============================================================================= 1464 //--------------------------------memory--------------------------------------- 1465 Node* GraphKit::memory(uint alias_idx) { 1466 MergeMemNode* mem = merged_memory(); 1467 Node* p = mem->memory_at(alias_idx); 1468 _gvn.set_type(p, Type::MEMORY); // must be mapped 1469 return p; 1470 } 1471 1472 //-----------------------------reset_memory------------------------------------ 1473 Node* GraphKit::reset_memory() { 1474 Node* mem = map()->memory(); 1475 // do not use this node for any more parsing! 1476 debug_only( map()->set_memory((Node*)NULL) ); 1477 return _gvn.transform( mem ); 1478 } 1479 1480 //------------------------------set_all_memory--------------------------------- 1481 void GraphKit::set_all_memory(Node* newmem) { 1482 Node* mergemem = MergeMemNode::make(C, newmem); 1483 gvn().set_type_bottom(mergemem); 1484 map()->set_memory(mergemem); 1485 } 1486 1487 //------------------------------set_all_memory_call---------------------------- 1488 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1489 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1490 set_all_memory(newmem); 1491 } 1492 1493 //============================================================================= 1494 // 1495 // parser factory methods for MemNodes 1496 // 1497 // These are layered on top of the factory methods in LoadNode and StoreNode, 1498 // and integrate with the parser's memory state and _gvn engine. 1499 // 1500 1501 // factory methods in "int adr_idx" 1502 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1503 int adr_idx, 1504 MemNode::MemOrd mo, bool require_atomic_access) { 1505 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1506 const TypePtr* adr_type = NULL; // debug-mode-only argument 1507 debug_only(adr_type = C->get_adr_type(adr_idx)); 1508 Node* mem = memory(adr_idx); 1509 Node* ld; 1510 if (require_atomic_access && bt == T_LONG) { 1511 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo); 1512 } else if (require_atomic_access && bt == T_DOUBLE) { 1513 ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo); 1514 } else { 1515 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); 1516 } 1517 ld = _gvn.transform(ld); 1518 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1519 // Improve graph before escape analysis and boxing elimination. 1520 record_for_igvn(ld); 1521 } 1522 return ld; 1523 } 1524 1525 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1526 int adr_idx, 1527 MemNode::MemOrd mo, 1528 bool require_atomic_access) { 1529 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1530 const TypePtr* adr_type = NULL; 1531 debug_only(adr_type = C->get_adr_type(adr_idx)); 1532 Node *mem = memory(adr_idx); 1533 Node* st; 1534 if (require_atomic_access && bt == T_LONG) { 1535 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1536 } else if (require_atomic_access && bt == T_DOUBLE) { 1537 st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1538 } else { 1539 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1540 } 1541 st = _gvn.transform(st); 1542 set_memory(st, adr_idx); 1543 // Back-to-back stores can only remove intermediate store with DU info 1544 // so push on worklist for optimizer. 1545 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1546 record_for_igvn(st); 1547 1548 return st; 1549 } 1550 1551 1552 void GraphKit::pre_barrier(bool do_load, 1553 Node* ctl, 1554 Node* obj, 1555 Node* adr, 1556 uint adr_idx, 1557 Node* val, 1558 const TypeOopPtr* val_type, 1559 Node* pre_val, 1560 BasicType bt) { 1561 1562 BarrierSet* bs = Universe::heap()->barrier_set(); 1563 set_control(ctl); 1564 switch (bs->kind()) { 1565 case BarrierSet::G1SATBCT: 1566 case BarrierSet::G1SATBCTLogging: 1567 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1568 break; 1569 1570 case BarrierSet::CardTableModRef: 1571 case BarrierSet::CardTableExtension: 1572 case BarrierSet::ModRef: 1573 break; 1574 1575 case BarrierSet::Other: 1576 default : 1577 ShouldNotReachHere(); 1578 1579 } 1580 } 1581 1582 bool GraphKit::can_move_pre_barrier() const { 1583 BarrierSet* bs = Universe::heap()->barrier_set(); 1584 switch (bs->kind()) { 1585 case BarrierSet::G1SATBCT: 1586 case BarrierSet::G1SATBCTLogging: 1587 return true; // Can move it if no safepoint 1588 1589 case BarrierSet::CardTableModRef: 1590 case BarrierSet::CardTableExtension: 1591 case BarrierSet::ModRef: 1592 return true; // There is no pre-barrier 1593 1594 case BarrierSet::Other: 1595 default : 1596 ShouldNotReachHere(); 1597 } 1598 return false; 1599 } 1600 1601 void GraphKit::post_barrier(Node* ctl, 1602 Node* store, 1603 Node* obj, 1604 Node* adr, 1605 uint adr_idx, 1606 Node* val, 1607 BasicType bt, 1608 bool use_precise) { 1609 BarrierSet* bs = Universe::heap()->barrier_set(); 1610 set_control(ctl); 1611 switch (bs->kind()) { 1612 case BarrierSet::G1SATBCT: 1613 case BarrierSet::G1SATBCTLogging: 1614 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1615 break; 1616 1617 case BarrierSet::CardTableModRef: 1618 case BarrierSet::CardTableExtension: 1619 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1620 break; 1621 1622 case BarrierSet::ModRef: 1623 break; 1624 1625 case BarrierSet::Other: 1626 default : 1627 ShouldNotReachHere(); 1628 1629 } 1630 } 1631 1632 Node* GraphKit::store_oop(Node* ctl, 1633 Node* obj, 1634 Node* adr, 1635 const TypePtr* adr_type, 1636 Node* val, 1637 const TypeOopPtr* val_type, 1638 BasicType bt, 1639 bool use_precise, 1640 MemNode::MemOrd mo) { 1641 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1642 // could be delayed during Parse (for example, in adjust_map_after_if()). 1643 // Execute transformation here to avoid barrier generation in such case. 1644 if (_gvn.type(val) == TypePtr::NULL_PTR) 1645 val = _gvn.makecon(TypePtr::NULL_PTR); 1646 1647 set_control(ctl); 1648 if (stopped()) return top(); // Dead path ? 1649 1650 assert(bt == T_OBJECT, "sanity"); 1651 assert(val != NULL, "not dead path"); 1652 uint adr_idx = C->get_alias_index(adr_type); 1653 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1654 1655 pre_barrier(true /* do_load */, 1656 control(), obj, adr, adr_idx, val, val_type, 1657 NULL /* pre_val */, 1658 bt); 1659 1660 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1661 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1662 return store; 1663 } 1664 1665 // Could be an array or object we don't know at compile time (unsafe ref.) 1666 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1667 Node* obj, // containing obj 1668 Node* adr, // actual adress to store val at 1669 const TypePtr* adr_type, 1670 Node* val, 1671 BasicType bt, 1672 MemNode::MemOrd mo) { 1673 Compile::AliasType* at = C->alias_type(adr_type); 1674 const TypeOopPtr* val_type = NULL; 1675 if (adr_type->isa_instptr()) { 1676 if (at->field() != NULL) { 1677 // known field. This code is a copy of the do_put_xxx logic. 1678 ciField* field = at->field(); 1679 if (!field->type()->is_loaded()) { 1680 val_type = TypeInstPtr::BOTTOM; 1681 } else { 1682 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1683 } 1684 } 1685 } else if (adr_type->isa_aryptr()) { 1686 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1687 } 1688 if (val_type == NULL) { 1689 val_type = TypeInstPtr::BOTTOM; 1690 } 1691 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1692 } 1693 1694 1695 //-------------------------array_element_address------------------------- 1696 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1697 const TypeInt* sizetype) { 1698 uint shift = exact_log2(type2aelembytes(elembt)); 1699 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1700 1701 // short-circuit a common case (saves lots of confusing waste motion) 1702 jint idx_con = find_int_con(idx, -1); 1703 if (idx_con >= 0) { 1704 intptr_t offset = header + ((intptr_t)idx_con << shift); 1705 return basic_plus_adr(ary, offset); 1706 } 1707 1708 // must be correct type for alignment purposes 1709 Node* base = basic_plus_adr(ary, header); 1710 #ifdef _LP64 1711 // The scaled index operand to AddP must be a clean 64-bit value. 1712 // Java allows a 32-bit int to be incremented to a negative 1713 // value, which appears in a 64-bit register as a large 1714 // positive number. Using that large positive number as an 1715 // operand in pointer arithmetic has bad consequences. 1716 // On the other hand, 32-bit overflow is rare, and the possibility 1717 // can often be excluded, if we annotate the ConvI2L node with 1718 // a type assertion that its value is known to be a small positive 1719 // number. (The prior range check has ensured this.) 1720 // This assertion is used by ConvI2LNode::Ideal. 1721 int index_max = max_jint - 1; // array size is max_jint, index is one less 1722 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1723 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1724 idx = _gvn.transform( new ConvI2LNode(idx, lidxtype) ); 1725 #endif 1726 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1727 return basic_plus_adr(ary, base, scale); 1728 } 1729 1730 //-------------------------load_array_element------------------------- 1731 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1732 const Type* elemtype = arytype->elem(); 1733 BasicType elembt = elemtype->array_element_basic_type(); 1734 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1735 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1736 return ld; 1737 } 1738 1739 //-------------------------set_arguments_for_java_call------------------------- 1740 // Arguments (pre-popped from the stack) are taken from the JVMS. 1741 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1742 // Add the call arguments: 1743 uint nargs = call->method()->arg_size(); 1744 for (uint i = 0; i < nargs; i++) { 1745 Node* arg = argument(i); 1746 call->init_req(i + TypeFunc::Parms, arg); 1747 } 1748 } 1749 1750 //---------------------------set_edges_for_java_call--------------------------- 1751 // Connect a newly created call into the current JVMS. 1752 // A return value node (if any) is returned from set_edges_for_java_call. 1753 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1754 1755 // Add the predefined inputs: 1756 call->init_req( TypeFunc::Control, control() ); 1757 call->init_req( TypeFunc::I_O , i_o() ); 1758 call->init_req( TypeFunc::Memory , reset_memory() ); 1759 call->init_req( TypeFunc::FramePtr, frameptr() ); 1760 call->init_req( TypeFunc::ReturnAdr, top() ); 1761 1762 add_safepoint_edges(call, must_throw); 1763 1764 Node* xcall = _gvn.transform(call); 1765 1766 if (xcall == top()) { 1767 set_control(top()); 1768 return; 1769 } 1770 assert(xcall == call, "call identity is stable"); 1771 1772 // Re-use the current map to produce the result. 1773 1774 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1775 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1776 set_all_memory_call(xcall, separate_io_proj); 1777 1778 //return xcall; // no need, caller already has it 1779 } 1780 1781 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1782 if (stopped()) return top(); // maybe the call folded up? 1783 1784 // Capture the return value, if any. 1785 Node* ret; 1786 if (call->method() == NULL || 1787 call->method()->return_type()->basic_type() == T_VOID) 1788 ret = top(); 1789 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1790 1791 // Note: Since any out-of-line call can produce an exception, 1792 // we always insert an I_O projection from the call into the result. 1793 1794 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1795 1796 if (separate_io_proj) { 1797 // The caller requested separate projections be used by the fall 1798 // through and exceptional paths, so replace the projections for 1799 // the fall through path. 1800 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1801 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1802 } 1803 return ret; 1804 } 1805 1806 //--------------------set_predefined_input_for_runtime_call-------------------- 1807 // Reading and setting the memory state is way conservative here. 1808 // The real problem is that I am not doing real Type analysis on memory, 1809 // so I cannot distinguish card mark stores from other stores. Across a GC 1810 // point the Store Barrier and the card mark memory has to agree. I cannot 1811 // have a card mark store and its barrier split across the GC point from 1812 // either above or below. Here I get that to happen by reading ALL of memory. 1813 // A better answer would be to separate out card marks from other memory. 1814 // For now, return the input memory state, so that it can be reused 1815 // after the call, if this call has restricted memory effects. 1816 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1817 // Set fixed predefined input arguments 1818 Node* memory = reset_memory(); 1819 call->init_req( TypeFunc::Control, control() ); 1820 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1821 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1822 call->init_req( TypeFunc::FramePtr, frameptr() ); 1823 call->init_req( TypeFunc::ReturnAdr, top() ); 1824 return memory; 1825 } 1826 1827 //-------------------set_predefined_output_for_runtime_call-------------------- 1828 // Set control and memory (not i_o) from the call. 1829 // If keep_mem is not NULL, use it for the output state, 1830 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1831 // If hook_mem is NULL, this call produces no memory effects at all. 1832 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1833 // then only that memory slice is taken from the call. 1834 // In the last case, we must put an appropriate memory barrier before 1835 // the call, so as to create the correct anti-dependencies on loads 1836 // preceding the call. 1837 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1838 Node* keep_mem, 1839 const TypePtr* hook_mem) { 1840 // no i/o 1841 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1842 if (keep_mem) { 1843 // First clone the existing memory state 1844 set_all_memory(keep_mem); 1845 if (hook_mem != NULL) { 1846 // Make memory for the call 1847 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1848 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1849 // We also use hook_mem to extract specific effects from arraycopy stubs. 1850 set_memory(mem, hook_mem); 1851 } 1852 // ...else the call has NO memory effects. 1853 1854 // Make sure the call advertises its memory effects precisely. 1855 // This lets us build accurate anti-dependences in gcm.cpp. 1856 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1857 "call node must be constructed correctly"); 1858 } else { 1859 assert(hook_mem == NULL, ""); 1860 // This is not a "slow path" call; all memory comes from the call. 1861 set_all_memory_call(call); 1862 } 1863 } 1864 1865 1866 // Replace the call with the current state of the kit. 1867 void GraphKit::replace_call(CallNode* call, Node* result) { 1868 JVMState* ejvms = NULL; 1869 if (has_exceptions()) { 1870 ejvms = transfer_exceptions_into_jvms(); 1871 } 1872 1873 SafePointNode* final_state = stop(); 1874 1875 // Find all the needed outputs of this call 1876 CallProjections callprojs; 1877 call->extract_projections(&callprojs, true); 1878 1879 Node* init_mem = call->in(TypeFunc::Memory); 1880 Node* final_mem = final_state->in(TypeFunc::Memory); 1881 Node* final_ctl = final_state->in(TypeFunc::Control); 1882 Node* final_io = final_state->in(TypeFunc::I_O); 1883 1884 // Replace all the old call edges with the edges from the inlining result 1885 if (callprojs.fallthrough_catchproj != NULL) { 1886 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1887 } 1888 if (callprojs.fallthrough_memproj != NULL) { 1889 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1890 } 1891 if (callprojs.fallthrough_ioproj != NULL) { 1892 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1893 } 1894 1895 // Replace the result with the new result if it exists and is used 1896 if (callprojs.resproj != NULL && result != NULL) { 1897 C->gvn_replace_by(callprojs.resproj, result); 1898 } 1899 1900 if (ejvms == NULL) { 1901 // No exception edges to simply kill off those paths 1902 if (callprojs.catchall_catchproj != NULL) { 1903 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1904 } 1905 if (callprojs.catchall_memproj != NULL) { 1906 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1907 } 1908 if (callprojs.catchall_ioproj != NULL) { 1909 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1910 } 1911 // Replace the old exception object with top 1912 if (callprojs.exobj != NULL) { 1913 C->gvn_replace_by(callprojs.exobj, C->top()); 1914 } 1915 } else { 1916 GraphKit ekit(ejvms); 1917 1918 // Load my combined exception state into the kit, with all phis transformed: 1919 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1920 1921 Node* ex_oop = ekit.use_exception_state(ex_map); 1922 if (callprojs.catchall_catchproj != NULL) { 1923 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1924 } 1925 if (callprojs.catchall_memproj != NULL) { 1926 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1927 } 1928 if (callprojs.catchall_ioproj != NULL) { 1929 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1930 } 1931 1932 // Replace the old exception object with the newly created one 1933 if (callprojs.exobj != NULL) { 1934 C->gvn_replace_by(callprojs.exobj, ex_oop); 1935 } 1936 } 1937 1938 // Disconnect the call from the graph 1939 call->disconnect_inputs(NULL, C); 1940 C->gvn_replace_by(call, C->top()); 1941 1942 // Clean up any MergeMems that feed other MergeMems since the 1943 // optimizer doesn't like that. 1944 if (final_mem->is_MergeMem()) { 1945 Node_List wl; 1946 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1947 Node* m = i.get(); 1948 if (m->is_MergeMem() && !wl.contains(m)) { 1949 wl.push(m); 1950 } 1951 } 1952 while (wl.size() > 0) { 1953 _gvn.transform(wl.pop()); 1954 } 1955 } 1956 } 1957 1958 1959 //------------------------------increment_counter------------------------------ 1960 // for statistics: increment a VM counter by 1 1961 1962 void GraphKit::increment_counter(address counter_addr) { 1963 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1964 increment_counter(adr1); 1965 } 1966 1967 void GraphKit::increment_counter(Node* counter_addr) { 1968 int adr_type = Compile::AliasIdxRaw; 1969 Node* ctrl = control(); 1970 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1971 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1972 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1973 } 1974 1975 1976 //------------------------------uncommon_trap---------------------------------- 1977 // Bail out to the interpreter in mid-method. Implemented by calling the 1978 // uncommon_trap blob. This helper function inserts a runtime call with the 1979 // right debug info. 1980 void GraphKit::uncommon_trap(int trap_request, 1981 ciKlass* klass, const char* comment, 1982 bool must_throw, 1983 bool keep_exact_action) { 1984 if (failing()) stop(); 1985 if (stopped()) return; // trap reachable? 1986 1987 // Note: If ProfileTraps is true, and if a deopt. actually 1988 // occurs here, the runtime will make sure an MDO exists. There is 1989 // no need to call method()->ensure_method_data() at this point. 1990 1991 // Set the stack pointer to the right value for reexecution: 1992 set_sp(reexecute_sp()); 1993 1994 #ifdef ASSERT 1995 if (!must_throw) { 1996 // Make sure the stack has at least enough depth to execute 1997 // the current bytecode. 1998 int inputs, ignored_depth; 1999 if (compute_stack_effects(inputs, ignored_depth)) { 2000 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 2001 Bytecodes::name(java_bc()), sp(), inputs)); 2002 } 2003 } 2004 #endif 2005 2006 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2007 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2008 2009 switch (action) { 2010 case Deoptimization::Action_maybe_recompile: 2011 case Deoptimization::Action_reinterpret: 2012 // Temporary fix for 6529811 to allow virtual calls to be sure they 2013 // get the chance to go from mono->bi->mega 2014 if (!keep_exact_action && 2015 Deoptimization::trap_request_index(trap_request) < 0 && 2016 too_many_recompiles(reason)) { 2017 // This BCI is causing too many recompilations. 2018 action = Deoptimization::Action_none; 2019 trap_request = Deoptimization::make_trap_request(reason, action); 2020 } else { 2021 C->set_trap_can_recompile(true); 2022 } 2023 break; 2024 case Deoptimization::Action_make_not_entrant: 2025 C->set_trap_can_recompile(true); 2026 break; 2027 #ifdef ASSERT 2028 case Deoptimization::Action_none: 2029 case Deoptimization::Action_make_not_compilable: 2030 break; 2031 default: 2032 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2033 break; 2034 #endif 2035 } 2036 2037 if (TraceOptoParse) { 2038 char buf[100]; 2039 tty->print_cr("Uncommon trap %s at bci:%d", 2040 Deoptimization::format_trap_request(buf, sizeof(buf), 2041 trap_request), bci()); 2042 } 2043 2044 CompileLog* log = C->log(); 2045 if (log != NULL) { 2046 int kid = (klass == NULL)? -1: log->identify(klass); 2047 log->begin_elem("uncommon_trap bci='%d'", bci()); 2048 char buf[100]; 2049 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2050 trap_request)); 2051 if (kid >= 0) log->print(" klass='%d'", kid); 2052 if (comment != NULL) log->print(" comment='%s'", comment); 2053 log->end_elem(); 2054 } 2055 2056 // Make sure any guarding test views this path as very unlikely 2057 Node *i0 = control()->in(0); 2058 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2059 IfNode *iff = i0->as_If(); 2060 float f = iff->_prob; // Get prob 2061 if (control()->Opcode() == Op_IfTrue) { 2062 if (f > PROB_UNLIKELY_MAG(4)) 2063 iff->_prob = PROB_MIN; 2064 } else { 2065 if (f < PROB_LIKELY_MAG(4)) 2066 iff->_prob = PROB_MAX; 2067 } 2068 } 2069 2070 // Clear out dead values from the debug info. 2071 kill_dead_locals(); 2072 2073 // Now insert the uncommon trap subroutine call 2074 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2075 const TypePtr* no_memory_effects = NULL; 2076 // Pass the index of the class to be loaded 2077 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2078 (must_throw ? RC_MUST_THROW : 0), 2079 OptoRuntime::uncommon_trap_Type(), 2080 call_addr, "uncommon_trap", no_memory_effects, 2081 intcon(trap_request)); 2082 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2083 "must extract request correctly from the graph"); 2084 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2085 2086 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2087 // The debug info is the only real input to this call. 2088 2089 // Halt-and-catch fire here. The above call should never return! 2090 HaltNode* halt = new HaltNode(control(), frameptr()); 2091 _gvn.set_type_bottom(halt); 2092 root()->add_req(halt); 2093 2094 stop_and_kill_map(); 2095 } 2096 2097 2098 //--------------------------just_allocated_object------------------------------ 2099 // Report the object that was just allocated. 2100 // It must be the case that there are no intervening safepoints. 2101 // We use this to determine if an object is so "fresh" that 2102 // it does not require card marks. 2103 Node* GraphKit::just_allocated_object(Node* current_control) { 2104 if (C->recent_alloc_ctl() == current_control) 2105 return C->recent_alloc_obj(); 2106 return NULL; 2107 } 2108 2109 2110 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2111 // (Note: TypeFunc::make has a cache that makes this fast.) 2112 const TypeFunc* tf = TypeFunc::make(dest_method); 2113 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2114 for (int j = 0; j < nargs; j++) { 2115 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2116 if( targ->basic_type() == T_DOUBLE ) { 2117 // If any parameters are doubles, they must be rounded before 2118 // the call, dstore_rounding does gvn.transform 2119 Node *arg = argument(j); 2120 arg = dstore_rounding(arg); 2121 set_argument(j, arg); 2122 } 2123 } 2124 } 2125 2126 /** 2127 * Record profiling data exact_kls for Node n with the type system so 2128 * that it can propagate it (speculation) 2129 * 2130 * @param n node that the type applies to 2131 * @param exact_kls type from profiling 2132 * @param maybe_null did profiling see null? 2133 * 2134 * @return node with improved type 2135 */ 2136 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2137 const Type* current_type = _gvn.type(n); 2138 assert(UseTypeSpeculation, "type speculation must be on"); 2139 2140 const TypePtr* speculative = current_type->speculative(); 2141 2142 // Should the klass from the profile be recorded in the speculative type? 2143 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2144 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2145 const TypeOopPtr* xtype = tklass->as_instance_type(); 2146 assert(xtype->klass_is_exact(), "Should be exact"); 2147 // Any reason to believe n is not null (from this profiling or a previous one)? 2148 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2149 // record the new speculative type's depth 2150 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2151 speculative = speculative->with_inline_depth(jvms()->depth()); 2152 } else if (current_type->would_improve_ptr(maybe_null)) { 2153 // Profiling report that null was never seen so we can change the 2154 // speculative type to non null ptr. 2155 assert(!maybe_null, "nothing to improve"); 2156 if (speculative == NULL) { 2157 speculative = TypePtr::NOTNULL; 2158 } else { 2159 const TypePtr* ptr = TypePtr::NOTNULL; 2160 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2161 } 2162 } 2163 2164 if (speculative != current_type->speculative()) { 2165 // Build a type with a speculative type (what we think we know 2166 // about the type but will need a guard when we use it) 2167 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2168 // We're changing the type, we need a new CheckCast node to carry 2169 // the new type. The new type depends on the control: what 2170 // profiling tells us is only valid from here as far as we can 2171 // tell. 2172 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2173 cast = _gvn.transform(cast); 2174 replace_in_map(n, cast); 2175 n = cast; 2176 } 2177 2178 return n; 2179 } 2180 2181 /** 2182 * Record profiling data from receiver profiling at an invoke with the 2183 * type system so that it can propagate it (speculation) 2184 * 2185 * @param n receiver node 2186 * 2187 * @return node with improved type 2188 */ 2189 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2190 if (!UseTypeSpeculation) { 2191 return n; 2192 } 2193 ciKlass* exact_kls = profile_has_unique_klass(); 2194 bool maybe_null = true; 2195 if (java_bc() == Bytecodes::_checkcast || 2196 java_bc() == Bytecodes::_instanceof || 2197 java_bc() == Bytecodes::_aastore) { 2198 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2199 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2200 } 2201 return record_profile_for_speculation(n, exact_kls, maybe_null); 2202 return n; 2203 } 2204 2205 /** 2206 * Record profiling data from argument profiling at an invoke with the 2207 * type system so that it can propagate it (speculation) 2208 * 2209 * @param dest_method target method for the call 2210 * @param bc what invoke bytecode is this? 2211 */ 2212 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2213 if (!UseTypeSpeculation) { 2214 return; 2215 } 2216 const TypeFunc* tf = TypeFunc::make(dest_method); 2217 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2218 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2219 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2220 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2221 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2222 bool maybe_null = true; 2223 ciKlass* better_type = NULL; 2224 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2225 record_profile_for_speculation(argument(j), better_type, maybe_null); 2226 } 2227 i++; 2228 } 2229 } 2230 } 2231 2232 /** 2233 * Record profiling data from parameter profiling at an invoke with 2234 * the type system so that it can propagate it (speculation) 2235 */ 2236 void GraphKit::record_profiled_parameters_for_speculation() { 2237 if (!UseTypeSpeculation) { 2238 return; 2239 } 2240 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2241 if (_gvn.type(local(i))->isa_oopptr()) { 2242 bool maybe_null = true; 2243 ciKlass* better_type = NULL; 2244 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2245 record_profile_for_speculation(local(i), better_type, maybe_null); 2246 } 2247 j++; 2248 } 2249 } 2250 } 2251 2252 /** 2253 * Record profiling data from return value profiling at an invoke with 2254 * the type system so that it can propagate it (speculation) 2255 */ 2256 void GraphKit::record_profiled_return_for_speculation() { 2257 if (!UseTypeSpeculation) { 2258 return; 2259 } 2260 bool maybe_null = true; 2261 ciKlass* better_type = NULL; 2262 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2263 // If profiling reports a single type for the return value, 2264 // feed it to the type system so it can propagate it as a 2265 // speculative type 2266 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2267 } 2268 } 2269 2270 void GraphKit::round_double_result(ciMethod* dest_method) { 2271 // A non-strict method may return a double value which has an extended 2272 // exponent, but this must not be visible in a caller which is 'strict' 2273 // If a strict caller invokes a non-strict callee, round a double result 2274 2275 BasicType result_type = dest_method->return_type()->basic_type(); 2276 assert( method() != NULL, "must have caller context"); 2277 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2278 // Destination method's return value is on top of stack 2279 // dstore_rounding() does gvn.transform 2280 Node *result = pop_pair(); 2281 result = dstore_rounding(result); 2282 push_pair(result); 2283 } 2284 } 2285 2286 // rounding for strict float precision conformance 2287 Node* GraphKit::precision_rounding(Node* n) { 2288 return UseStrictFP && _method->flags().is_strict() 2289 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2290 ? _gvn.transform( new RoundFloatNode(0, n) ) 2291 : n; 2292 } 2293 2294 // rounding for strict double precision conformance 2295 Node* GraphKit::dprecision_rounding(Node *n) { 2296 return UseStrictFP && _method->flags().is_strict() 2297 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2298 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2299 : n; 2300 } 2301 2302 // rounding for non-strict double stores 2303 Node* GraphKit::dstore_rounding(Node* n) { 2304 return Matcher::strict_fp_requires_explicit_rounding 2305 && UseSSE <= 1 2306 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2307 : n; 2308 } 2309 2310 //============================================================================= 2311 // Generate a fast path/slow path idiom. Graph looks like: 2312 // [foo] indicates that 'foo' is a parameter 2313 // 2314 // [in] NULL 2315 // \ / 2316 // CmpP 2317 // Bool ne 2318 // If 2319 // / \ 2320 // True False-<2> 2321 // / | 2322 // / cast_not_null 2323 // Load | | ^ 2324 // [fast_test] | | 2325 // gvn to opt_test | | 2326 // / \ | <1> 2327 // True False | 2328 // | \\ | 2329 // [slow_call] \[fast_result] 2330 // Ctl Val \ \ 2331 // | \ \ 2332 // Catch <1> \ \ 2333 // / \ ^ \ \ 2334 // Ex No_Ex | \ \ 2335 // | \ \ | \ <2> \ 2336 // ... \ [slow_res] | | \ [null_result] 2337 // \ \--+--+--- | | 2338 // \ | / \ | / 2339 // --------Region Phi 2340 // 2341 //============================================================================= 2342 // Code is structured as a series of driver functions all called 'do_XXX' that 2343 // call a set of helper functions. Helper functions first, then drivers. 2344 2345 //------------------------------null_check_oop--------------------------------- 2346 // Null check oop. Set null-path control into Region in slot 3. 2347 // Make a cast-not-nullness use the other not-null control. Return cast. 2348 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2349 bool never_see_null, 2350 bool safe_for_replace, 2351 bool speculative) { 2352 // Initial NULL check taken path 2353 (*null_control) = top(); 2354 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2355 2356 // Generate uncommon_trap: 2357 if (never_see_null && (*null_control) != top()) { 2358 // If we see an unexpected null at a check-cast we record it and force a 2359 // recompile; the offending check-cast will be compiled to handle NULLs. 2360 // If we see more than one offending BCI, then all checkcasts in the 2361 // method will be compiled to handle NULLs. 2362 PreserveJVMState pjvms(this); 2363 set_control(*null_control); 2364 replace_in_map(value, null()); 2365 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2366 uncommon_trap(reason, 2367 Deoptimization::Action_make_not_entrant); 2368 (*null_control) = top(); // NULL path is dead 2369 } 2370 if ((*null_control) == top() && safe_for_replace) { 2371 replace_in_map(value, cast); 2372 } 2373 2374 // Cast away null-ness on the result 2375 return cast; 2376 } 2377 2378 //------------------------------opt_iff---------------------------------------- 2379 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2380 // Return slow-path control. 2381 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2382 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2383 2384 // Fast path taken; set region slot 2 2385 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2386 region->init_req(2,fast_taken); // Capture fast-control 2387 2388 // Fast path not-taken, i.e. slow path 2389 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2390 return slow_taken; 2391 } 2392 2393 //-----------------------------make_runtime_call------------------------------- 2394 Node* GraphKit::make_runtime_call(int flags, 2395 const TypeFunc* call_type, address call_addr, 2396 const char* call_name, 2397 const TypePtr* adr_type, 2398 // The following parms are all optional. 2399 // The first NULL ends the list. 2400 Node* parm0, Node* parm1, 2401 Node* parm2, Node* parm3, 2402 Node* parm4, Node* parm5, 2403 Node* parm6, Node* parm7) { 2404 // Slow-path call 2405 bool is_leaf = !(flags & RC_NO_LEAF); 2406 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2407 if (call_name == NULL) { 2408 assert(!is_leaf, "must supply name for leaf"); 2409 call_name = OptoRuntime::stub_name(call_addr); 2410 } 2411 CallNode* call; 2412 if (!is_leaf) { 2413 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2414 bci(), adr_type); 2415 } else if (flags & RC_NO_FP) { 2416 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2417 } else { 2418 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2419 } 2420 2421 // The following is similar to set_edges_for_java_call, 2422 // except that the memory effects of the call are restricted to AliasIdxRaw. 2423 2424 // Slow path call has no side-effects, uses few values 2425 bool wide_in = !(flags & RC_NARROW_MEM); 2426 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2427 2428 Node* prev_mem = NULL; 2429 if (wide_in) { 2430 prev_mem = set_predefined_input_for_runtime_call(call); 2431 } else { 2432 assert(!wide_out, "narrow in => narrow out"); 2433 Node* narrow_mem = memory(adr_type); 2434 prev_mem = reset_memory(); 2435 map()->set_memory(narrow_mem); 2436 set_predefined_input_for_runtime_call(call); 2437 } 2438 2439 // Hook each parm in order. Stop looking at the first NULL. 2440 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2441 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2442 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2443 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2444 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2445 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2446 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2447 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2448 /* close each nested if ===> */ } } } } } } } } 2449 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2450 2451 if (!is_leaf) { 2452 // Non-leaves can block and take safepoints: 2453 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2454 } 2455 // Non-leaves can throw exceptions: 2456 if (has_io) { 2457 call->set_req(TypeFunc::I_O, i_o()); 2458 } 2459 2460 if (flags & RC_UNCOMMON) { 2461 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2462 // (An "if" probability corresponds roughly to an unconditional count. 2463 // Sort of.) 2464 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2465 } 2466 2467 Node* c = _gvn.transform(call); 2468 assert(c == call, "cannot disappear"); 2469 2470 if (wide_out) { 2471 // Slow path call has full side-effects. 2472 set_predefined_output_for_runtime_call(call); 2473 } else { 2474 // Slow path call has few side-effects, and/or sets few values. 2475 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2476 } 2477 2478 if (has_io) { 2479 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2480 } 2481 return call; 2482 2483 } 2484 2485 //------------------------------merge_memory----------------------------------- 2486 // Merge memory from one path into the current memory state. 2487 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2488 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2489 Node* old_slice = mms.force_memory(); 2490 Node* new_slice = mms.memory2(); 2491 if (old_slice != new_slice) { 2492 PhiNode* phi; 2493 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2494 if (mms.is_empty()) { 2495 // clone base memory Phi's inputs for this memory slice 2496 assert(old_slice == mms.base_memory(), "sanity"); 2497 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2498 _gvn.set_type(phi, Type::MEMORY); 2499 for (uint i = 1; i < phi->req(); i++) { 2500 phi->init_req(i, old_slice->in(i)); 2501 } 2502 } else { 2503 phi = old_slice->as_Phi(); // Phi was generated already 2504 } 2505 } else { 2506 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2507 _gvn.set_type(phi, Type::MEMORY); 2508 } 2509 phi->set_req(new_path, new_slice); 2510 mms.set_memory(phi); 2511 } 2512 } 2513 } 2514 2515 //------------------------------make_slow_call_ex------------------------------ 2516 // Make the exception handler hookups for the slow call 2517 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2518 if (stopped()) return; 2519 2520 // Make a catch node with just two handlers: fall-through and catch-all 2521 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2522 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2523 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2524 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2525 2526 { PreserveJVMState pjvms(this); 2527 set_control(excp); 2528 set_i_o(i_o); 2529 2530 if (excp != top()) { 2531 // Create an exception state also. 2532 // Use an exact type if the caller has specified a specific exception. 2533 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2534 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2535 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2536 } 2537 } 2538 2539 // Get the no-exception control from the CatchNode. 2540 set_control(norm); 2541 } 2542 2543 2544 //-------------------------------gen_subtype_check----------------------------- 2545 // Generate a subtyping check. Takes as input the subtype and supertype. 2546 // Returns 2 values: sets the default control() to the true path and returns 2547 // the false path. Only reads invariant memory; sets no (visible) memory. 2548 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2549 // but that's not exposed to the optimizer. This call also doesn't take in an 2550 // Object; if you wish to check an Object you need to load the Object's class 2551 // prior to coming here. 2552 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2553 // Fast check for identical types, perhaps identical constants. 2554 // The types can even be identical non-constants, in cases 2555 // involving Array.newInstance, Object.clone, etc. 2556 if (subklass == superklass) 2557 return top(); // false path is dead; no test needed. 2558 2559 if (_gvn.type(superklass)->singleton()) { 2560 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2561 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2562 2563 // In the common case of an exact superklass, try to fold up the 2564 // test before generating code. You may ask, why not just generate 2565 // the code and then let it fold up? The answer is that the generated 2566 // code will necessarily include null checks, which do not always 2567 // completely fold away. If they are also needless, then they turn 2568 // into a performance loss. Example: 2569 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2570 // Here, the type of 'fa' is often exact, so the store check 2571 // of fa[1]=x will fold up, without testing the nullness of x. 2572 switch (static_subtype_check(superk, subk)) { 2573 case SSC_always_false: 2574 { 2575 Node* always_fail = control(); 2576 set_control(top()); 2577 return always_fail; 2578 } 2579 case SSC_always_true: 2580 return top(); 2581 case SSC_easy_test: 2582 { 2583 // Just do a direct pointer compare and be done. 2584 Node* cmp = _gvn.transform( new CmpPNode(subklass, superklass) ); 2585 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2586 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2587 set_control( _gvn.transform( new IfTrueNode (iff) ) ); 2588 return _gvn.transform( new IfFalseNode(iff) ); 2589 } 2590 case SSC_full_test: 2591 break; 2592 default: 2593 ShouldNotReachHere(); 2594 } 2595 } 2596 2597 // %%% Possible further optimization: Even if the superklass is not exact, 2598 // if the subklass is the unique subtype of the superklass, the check 2599 // will always succeed. We could leave a dependency behind to ensure this. 2600 2601 // First load the super-klass's check-offset 2602 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2603 Node *chk_off = _gvn.transform(new LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), 2604 TypeInt::INT, MemNode::unordered)); 2605 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2606 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2607 2608 // Load from the sub-klass's super-class display list, or a 1-word cache of 2609 // the secondary superclass list, or a failing value with a sentinel offset 2610 // if the super-klass is an interface or exceptionally deep in the Java 2611 // hierarchy and we have to scan the secondary superclass list the hard way. 2612 // Worst-case type is a little odd: NULL is allowed as a result (usually 2613 // klass loads can never produce a NULL). 2614 Node *chk_off_X = ConvI2X(chk_off); 2615 Node *p2 = _gvn.transform( new AddPNode(subklass,subklass,chk_off_X) ); 2616 // For some types like interfaces the following loadKlass is from a 1-word 2617 // cache which is mutable so can't use immutable memory. Other 2618 // types load from the super-class display table which is immutable. 2619 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2620 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2621 2622 // Compile speed common case: ARE a subtype and we canNOT fail 2623 if( superklass == nkls ) 2624 return top(); // false path is dead; no test needed. 2625 2626 // See if we get an immediate positive hit. Happens roughly 83% of the 2627 // time. Test to see if the value loaded just previously from the subklass 2628 // is exactly the superklass. 2629 Node *cmp1 = _gvn.transform( new CmpPNode( superklass, nkls ) ); 2630 Node *bol1 = _gvn.transform( new BoolNode( cmp1, BoolTest::eq ) ); 2631 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2632 Node *iftrue1 = _gvn.transform( new IfTrueNode ( iff1 ) ); 2633 set_control( _gvn.transform( new IfFalseNode( iff1 ) ) ); 2634 2635 // Compile speed common case: Check for being deterministic right now. If 2636 // chk_off is a constant and not equal to cacheoff then we are NOT a 2637 // subklass. In this case we need exactly the 1 test above and we can 2638 // return those results immediately. 2639 if (!might_be_cache) { 2640 Node* not_subtype_ctrl = control(); 2641 set_control(iftrue1); // We need exactly the 1 test above 2642 return not_subtype_ctrl; 2643 } 2644 2645 // Gather the various success & failures here 2646 RegionNode *r_ok_subtype = new RegionNode(4); 2647 record_for_igvn(r_ok_subtype); 2648 RegionNode *r_not_subtype = new RegionNode(3); 2649 record_for_igvn(r_not_subtype); 2650 2651 r_ok_subtype->init_req(1, iftrue1); 2652 2653 // Check for immediate negative hit. Happens roughly 11% of the time (which 2654 // is roughly 63% of the remaining cases). Test to see if the loaded 2655 // check-offset points into the subklass display list or the 1-element 2656 // cache. If it points to the display (and NOT the cache) and the display 2657 // missed then it's not a subtype. 2658 Node *cacheoff = _gvn.intcon(cacheoff_con); 2659 Node *cmp2 = _gvn.transform( new CmpINode( chk_off, cacheoff ) ); 2660 Node *bol2 = _gvn.transform( new BoolNode( cmp2, BoolTest::ne ) ); 2661 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2662 r_not_subtype->init_req(1, _gvn.transform( new IfTrueNode (iff2) ) ); 2663 set_control( _gvn.transform( new IfFalseNode(iff2) ) ); 2664 2665 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2666 // No performance impact (too rare) but allows sharing of secondary arrays 2667 // which has some footprint reduction. 2668 Node *cmp3 = _gvn.transform( new CmpPNode( subklass, superklass ) ); 2669 Node *bol3 = _gvn.transform( new BoolNode( cmp3, BoolTest::eq ) ); 2670 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2671 r_ok_subtype->init_req(2, _gvn.transform( new IfTrueNode ( iff3 ) ) ); 2672 set_control( _gvn.transform( new IfFalseNode( iff3 ) ) ); 2673 2674 // -- Roads not taken here: -- 2675 // We could also have chosen to perform the self-check at the beginning 2676 // of this code sequence, as the assembler does. This would not pay off 2677 // the same way, since the optimizer, unlike the assembler, can perform 2678 // static type analysis to fold away many successful self-checks. 2679 // Non-foldable self checks work better here in second position, because 2680 // the initial primary superclass check subsumes a self-check for most 2681 // types. An exception would be a secondary type like array-of-interface, 2682 // which does not appear in its own primary supertype display. 2683 // Finally, we could have chosen to move the self-check into the 2684 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2685 // dependent manner. But it is worthwhile to have the check here, 2686 // where it can be perhaps be optimized. The cost in code space is 2687 // small (register compare, branch). 2688 2689 // Now do a linear scan of the secondary super-klass array. Again, no real 2690 // performance impact (too rare) but it's gotta be done. 2691 // Since the code is rarely used, there is no penalty for moving it 2692 // out of line, and it can only improve I-cache density. 2693 // The decision to inline or out-of-line this final check is platform 2694 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2695 Node* psc = _gvn.transform( 2696 new PartialSubtypeCheckNode(control(), subklass, superklass) ); 2697 2698 Node *cmp4 = _gvn.transform( new CmpPNode( psc, null() ) ); 2699 Node *bol4 = _gvn.transform( new BoolNode( cmp4, BoolTest::ne ) ); 2700 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2701 r_not_subtype->init_req(2, _gvn.transform( new IfTrueNode (iff4) ) ); 2702 r_ok_subtype ->init_req(3, _gvn.transform( new IfFalseNode(iff4) ) ); 2703 2704 // Return false path; set default control to true path. 2705 set_control( _gvn.transform(r_ok_subtype) ); 2706 return _gvn.transform(r_not_subtype); 2707 } 2708 2709 //----------------------------static_subtype_check----------------------------- 2710 // Shortcut important common cases when superklass is exact: 2711 // (0) superklass is java.lang.Object (can occur in reflective code) 2712 // (1) subklass is already limited to a subtype of superklass => always ok 2713 // (2) subklass does not overlap with superklass => always fail 2714 // (3) superklass has NO subtypes and we can check with a simple compare. 2715 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2716 if (StressReflectiveCode) { 2717 return SSC_full_test; // Let caller generate the general case. 2718 } 2719 2720 if (superk == env()->Object_klass()) { 2721 return SSC_always_true; // (0) this test cannot fail 2722 } 2723 2724 ciType* superelem = superk; 2725 if (superelem->is_array_klass()) 2726 superelem = superelem->as_array_klass()->base_element_type(); 2727 2728 if (!subk->is_interface()) { // cannot trust static interface types yet 2729 if (subk->is_subtype_of(superk)) { 2730 return SSC_always_true; // (1) false path dead; no dynamic test needed 2731 } 2732 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2733 !superk->is_subtype_of(subk)) { 2734 return SSC_always_false; 2735 } 2736 } 2737 2738 // If casting to an instance klass, it must have no subtypes 2739 if (superk->is_interface()) { 2740 // Cannot trust interfaces yet. 2741 // %%% S.B. superk->nof_implementors() == 1 2742 } else if (superelem->is_instance_klass()) { 2743 ciInstanceKlass* ik = superelem->as_instance_klass(); 2744 if (!ik->has_subklass() && !ik->is_interface()) { 2745 if (!ik->is_final()) { 2746 // Add a dependency if there is a chance of a later subclass. 2747 C->dependencies()->assert_leaf_type(ik); 2748 } 2749 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2750 } 2751 } else { 2752 // A primitive array type has no subtypes. 2753 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2754 } 2755 2756 return SSC_full_test; 2757 } 2758 2759 // Profile-driven exact type check: 2760 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2761 float prob, 2762 Node* *casted_receiver) { 2763 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2764 Node* recv_klass = load_object_klass(receiver); 2765 Node* want_klass = makecon(tklass); 2766 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2767 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2768 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2769 set_control( _gvn.transform( new IfTrueNode (iff) )); 2770 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2771 2772 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2773 assert(recv_xtype->klass_is_exact(), ""); 2774 2775 // Subsume downstream occurrences of receiver with a cast to 2776 // recv_xtype, since now we know what the type will be. 2777 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2778 (*casted_receiver) = _gvn.transform(cast); 2779 // (User must make the replace_in_map call.) 2780 2781 return fail; 2782 } 2783 2784 2785 //------------------------------seems_never_null------------------------------- 2786 // Use null_seen information if it is available from the profile. 2787 // If we see an unexpected null at a type check we record it and force a 2788 // recompile; the offending check will be recompiled to handle NULLs. 2789 // If we see several offending BCIs, then all checks in the 2790 // method will be recompiled. 2791 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2792 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2793 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2794 if (UncommonNullCast // Cutout for this technique 2795 && obj != null() // And not the -Xcomp stupid case? 2796 && !too_many_traps(reason) 2797 ) { 2798 if (speculating) { 2799 return true; 2800 } 2801 if (data == NULL) 2802 // Edge case: no mature data. Be optimistic here. 2803 return true; 2804 // If the profile has not seen a null, assume it won't happen. 2805 assert(java_bc() == Bytecodes::_checkcast || 2806 java_bc() == Bytecodes::_instanceof || 2807 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2808 return !data->as_BitData()->null_seen(); 2809 } 2810 speculating = false; 2811 return false; 2812 } 2813 2814 //------------------------maybe_cast_profiled_receiver------------------------- 2815 // If the profile has seen exactly one type, narrow to exactly that type. 2816 // Subsequent type checks will always fold up. 2817 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2818 ciKlass* require_klass, 2819 ciKlass* spec_klass, 2820 bool safe_for_replace) { 2821 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2822 2823 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2824 2825 // Make sure we haven't already deoptimized from this tactic. 2826 if (too_many_traps(reason)) 2827 return NULL; 2828 2829 // (No, this isn't a call, but it's enough like a virtual call 2830 // to use the same ciMethod accessor to get the profile info...) 2831 // If we have a speculative type use it instead of profiling (which 2832 // may not help us) 2833 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2834 if (exact_kls != NULL) {// no cast failures here 2835 if (require_klass == NULL || 2836 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2837 // If we narrow the type to match what the type profile sees or 2838 // the speculative type, we can then remove the rest of the 2839 // cast. 2840 // This is a win, even if the exact_kls is very specific, 2841 // because downstream operations, such as method calls, 2842 // will often benefit from the sharper type. 2843 Node* exact_obj = not_null_obj; // will get updated in place... 2844 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2845 &exact_obj); 2846 { PreserveJVMState pjvms(this); 2847 set_control(slow_ctl); 2848 uncommon_trap(reason, 2849 Deoptimization::Action_maybe_recompile); 2850 } 2851 if (safe_for_replace) { 2852 replace_in_map(not_null_obj, exact_obj); 2853 } 2854 return exact_obj; 2855 } 2856 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2857 } 2858 2859 return NULL; 2860 } 2861 2862 /** 2863 * Cast obj to type and emit guard unless we had too many traps here 2864 * already 2865 * 2866 * @param obj node being casted 2867 * @param type type to cast the node to 2868 * @param not_null true if we know node cannot be null 2869 */ 2870 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2871 ciKlass* type, 2872 bool not_null) { 2873 // type == NULL if profiling tells us this object is always null 2874 if (type != NULL) { 2875 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2876 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2877 if (!too_many_traps(null_reason) && 2878 !too_many_traps(class_reason)) { 2879 Node* not_null_obj = NULL; 2880 // not_null is true if we know the object is not null and 2881 // there's no need for a null check 2882 if (!not_null) { 2883 Node* null_ctl = top(); 2884 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2885 assert(null_ctl->is_top(), "no null control here"); 2886 } else { 2887 not_null_obj = obj; 2888 } 2889 2890 Node* exact_obj = not_null_obj; 2891 ciKlass* exact_kls = type; 2892 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2893 &exact_obj); 2894 { 2895 PreserveJVMState pjvms(this); 2896 set_control(slow_ctl); 2897 uncommon_trap(class_reason, 2898 Deoptimization::Action_maybe_recompile); 2899 } 2900 replace_in_map(not_null_obj, exact_obj); 2901 obj = exact_obj; 2902 } 2903 } else { 2904 if (!too_many_traps(Deoptimization::Reason_null_assert)) { 2905 Node* exact_obj = null_assert(obj); 2906 replace_in_map(obj, exact_obj); 2907 obj = exact_obj; 2908 } 2909 } 2910 return obj; 2911 } 2912 2913 //-------------------------------gen_instanceof-------------------------------- 2914 // Generate an instance-of idiom. Used by both the instance-of bytecode 2915 // and the reflective instance-of call. 2916 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2917 kill_dead_locals(); // Benefit all the uncommon traps 2918 assert( !stopped(), "dead parse path should be checked in callers" ); 2919 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2920 "must check for not-null not-dead klass in callers"); 2921 2922 // Make the merge point 2923 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2924 RegionNode* region = new RegionNode(PATH_LIMIT); 2925 Node* phi = new PhiNode(region, TypeInt::BOOL); 2926 C->set_has_split_ifs(true); // Has chance for split-if optimization 2927 2928 ciProfileData* data = NULL; 2929 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2930 data = method()->method_data()->bci_to_data(bci()); 2931 } 2932 bool speculative_not_null = false; 2933 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2934 && seems_never_null(obj, data, speculative_not_null)); 2935 2936 // Null check; get casted pointer; set region slot 3 2937 Node* null_ctl = top(); 2938 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2939 2940 // If not_null_obj is dead, only null-path is taken 2941 if (stopped()) { // Doing instance-of on a NULL? 2942 set_control(null_ctl); 2943 return intcon(0); 2944 } 2945 region->init_req(_null_path, null_ctl); 2946 phi ->init_req(_null_path, intcon(0)); // Set null path value 2947 if (null_ctl == top()) { 2948 // Do this eagerly, so that pattern matches like is_diamond_phi 2949 // will work even during parsing. 2950 assert(_null_path == PATH_LIMIT-1, "delete last"); 2951 region->del_req(_null_path); 2952 phi ->del_req(_null_path); 2953 } 2954 2955 // Do we know the type check always succeed? 2956 bool known_statically = false; 2957 if (_gvn.type(superklass)->singleton()) { 2958 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2959 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2960 if (subk != NULL && subk->is_loaded()) { 2961 int static_res = static_subtype_check(superk, subk); 2962 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2963 } 2964 } 2965 2966 if (known_statically && UseTypeSpeculation) { 2967 // If we know the type check always succeeds then we don't use the 2968 // profiling data at this bytecode. Don't lose it, feed it to the 2969 // type system as a speculative type. 2970 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2971 } else { 2972 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2973 // We may not have profiling here or it may not help us. If we 2974 // have a speculative type use it to perform an exact cast. 2975 ciKlass* spec_obj_type = obj_type->speculative_type(); 2976 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2977 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2978 if (stopped()) { // Profile disagrees with this path. 2979 set_control(null_ctl); // Null is the only remaining possibility. 2980 return intcon(0); 2981 } 2982 if (cast_obj != NULL) { 2983 not_null_obj = cast_obj; 2984 } 2985 } 2986 } 2987 2988 // Load the object's klass 2989 Node* obj_klass = load_object_klass(not_null_obj); 2990 2991 // Generate the subtype check 2992 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2993 2994 // Plug in the success path to the general merge in slot 1. 2995 region->init_req(_obj_path, control()); 2996 phi ->init_req(_obj_path, intcon(1)); 2997 2998 // Plug in the failing path to the general merge in slot 2. 2999 region->init_req(_fail_path, not_subtype_ctrl); 3000 phi ->init_req(_fail_path, intcon(0)); 3001 3002 // Return final merged results 3003 set_control( _gvn.transform(region) ); 3004 record_for_igvn(region); 3005 return _gvn.transform(phi); 3006 } 3007 3008 //-------------------------------gen_checkcast--------------------------------- 3009 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3010 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3011 // uncommon-trap paths work. Adjust stack after this call. 3012 // If failure_control is supplied and not null, it is filled in with 3013 // the control edge for the cast failure. Otherwise, an appropriate 3014 // uncommon trap or exception is thrown. 3015 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3016 Node* *failure_control) { 3017 kill_dead_locals(); // Benefit all the uncommon traps 3018 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3019 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3020 3021 // Fast cutout: Check the case that the cast is vacuously true. 3022 // This detects the common cases where the test will short-circuit 3023 // away completely. We do this before we perform the null check, 3024 // because if the test is going to turn into zero code, we don't 3025 // want a residual null check left around. (Causes a slowdown, 3026 // for example, in some objArray manipulations, such as a[i]=a[j].) 3027 if (tk->singleton()) { 3028 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3029 if (objtp != NULL && objtp->klass() != NULL) { 3030 switch (static_subtype_check(tk->klass(), objtp->klass())) { 3031 case SSC_always_true: 3032 // If we know the type check always succeed then we don't use 3033 // the profiling data at this bytecode. Don't lose it, feed it 3034 // to the type system as a speculative type. 3035 return record_profiled_receiver_for_speculation(obj); 3036 case SSC_always_false: 3037 // It needs a null check because a null will *pass* the cast check. 3038 // A non-null value will always produce an exception. 3039 return null_assert(obj); 3040 } 3041 } 3042 } 3043 3044 ciProfileData* data = NULL; 3045 bool safe_for_replace = false; 3046 if (failure_control == NULL) { // use MDO in regular case only 3047 assert(java_bc() == Bytecodes::_aastore || 3048 java_bc() == Bytecodes::_checkcast, 3049 "interpreter profiles type checks only for these BCs"); 3050 data = method()->method_data()->bci_to_data(bci()); 3051 safe_for_replace = true; 3052 } 3053 3054 // Make the merge point 3055 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3056 RegionNode* region = new RegionNode(PATH_LIMIT); 3057 Node* phi = new PhiNode(region, toop); 3058 C->set_has_split_ifs(true); // Has chance for split-if optimization 3059 3060 // Use null-cast information if it is available 3061 bool speculative_not_null = false; 3062 bool never_see_null = ((failure_control == NULL) // regular case only 3063 && seems_never_null(obj, data, speculative_not_null)); 3064 3065 // Null check; get casted pointer; set region slot 3 3066 Node* null_ctl = top(); 3067 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3068 3069 // If not_null_obj is dead, only null-path is taken 3070 if (stopped()) { // Doing instance-of on a NULL? 3071 set_control(null_ctl); 3072 return null(); 3073 } 3074 region->init_req(_null_path, null_ctl); 3075 phi ->init_req(_null_path, null()); // Set null path value 3076 if (null_ctl == top()) { 3077 // Do this eagerly, so that pattern matches like is_diamond_phi 3078 // will work even during parsing. 3079 assert(_null_path == PATH_LIMIT-1, "delete last"); 3080 region->del_req(_null_path); 3081 phi ->del_req(_null_path); 3082 } 3083 3084 Node* cast_obj = NULL; 3085 if (tk->klass_is_exact()) { 3086 // The following optimization tries to statically cast the speculative type of the object 3087 // (for example obtained during profiling) to the type of the superklass and then do a 3088 // dynamic check that the type of the object is what we expect. To work correctly 3089 // for checkcast and aastore the type of superklass should be exact. 3090 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3091 // We may not have profiling here or it may not help us. If we have 3092 // a speculative type use it to perform an exact cast. 3093 ciKlass* spec_obj_type = obj_type->speculative_type(); 3094 if (spec_obj_type != NULL || 3095 (data != NULL && 3096 // Counter has never been decremented (due to cast failure). 3097 // ...This is a reasonable thing to expect. It is true of 3098 // all casts inserted by javac to implement generic types. 3099 data->as_CounterData()->count() >= 0)) { 3100 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3101 if (cast_obj != NULL) { 3102 if (failure_control != NULL) // failure is now impossible 3103 (*failure_control) = top(); 3104 // adjust the type of the phi to the exact klass: 3105 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3106 } 3107 } 3108 } 3109 3110 if (cast_obj == NULL) { 3111 // Load the object's klass 3112 Node* obj_klass = load_object_klass(not_null_obj); 3113 3114 // Generate the subtype check 3115 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3116 3117 // Plug in success path into the merge 3118 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3119 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3120 if (failure_control == NULL) { 3121 if (not_subtype_ctrl != top()) { // If failure is possible 3122 PreserveJVMState pjvms(this); 3123 set_control(not_subtype_ctrl); 3124 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3125 } 3126 } else { 3127 (*failure_control) = not_subtype_ctrl; 3128 } 3129 } 3130 3131 region->init_req(_obj_path, control()); 3132 phi ->init_req(_obj_path, cast_obj); 3133 3134 // A merge of NULL or Casted-NotNull obj 3135 Node* res = _gvn.transform(phi); 3136 3137 // Note I do NOT always 'replace_in_map(obj,result)' here. 3138 // if( tk->klass()->can_be_primary_super() ) 3139 // This means that if I successfully store an Object into an array-of-String 3140 // I 'forget' that the Object is really now known to be a String. I have to 3141 // do this because we don't have true union types for interfaces - if I store 3142 // a Baz into an array-of-Interface and then tell the optimizer it's an 3143 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3144 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3145 // replace_in_map( obj, res ); 3146 3147 // Return final merged results 3148 set_control( _gvn.transform(region) ); 3149 record_for_igvn(region); 3150 return res; 3151 } 3152 3153 //------------------------------next_monitor----------------------------------- 3154 // What number should be given to the next monitor? 3155 int GraphKit::next_monitor() { 3156 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3157 int next = current + C->sync_stack_slots(); 3158 // Keep the toplevel high water mark current: 3159 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3160 return current; 3161 } 3162 3163 //------------------------------insert_mem_bar--------------------------------- 3164 // Memory barrier to avoid floating things around 3165 // The membar serves as a pinch point between both control and all memory slices. 3166 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3167 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3168 mb->init_req(TypeFunc::Control, control()); 3169 mb->init_req(TypeFunc::Memory, reset_memory()); 3170 Node* membar = _gvn.transform(mb); 3171 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3172 set_all_memory_call(membar); 3173 return membar; 3174 } 3175 3176 //-------------------------insert_mem_bar_volatile---------------------------- 3177 // Memory barrier to avoid floating things around 3178 // The membar serves as a pinch point between both control and memory(alias_idx). 3179 // If you want to make a pinch point on all memory slices, do not use this 3180 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3181 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3182 // When Parse::do_put_xxx updates a volatile field, it appends a series 3183 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3184 // The first membar is on the same memory slice as the field store opcode. 3185 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3186 // All the other membars (for other volatile slices, including AliasIdxBot, 3187 // which stands for all unknown volatile slices) are control-dependent 3188 // on the first membar. This prevents later volatile loads or stores 3189 // from sliding up past the just-emitted store. 3190 3191 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3192 mb->set_req(TypeFunc::Control,control()); 3193 if (alias_idx == Compile::AliasIdxBot) { 3194 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3195 } else { 3196 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3197 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3198 } 3199 Node* membar = _gvn.transform(mb); 3200 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3201 if (alias_idx == Compile::AliasIdxBot) { 3202 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3203 } else { 3204 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3205 } 3206 return membar; 3207 } 3208 3209 //------------------------------shared_lock------------------------------------ 3210 // Emit locking code. 3211 FastLockNode* GraphKit::shared_lock(Node* obj) { 3212 // bci is either a monitorenter bc or InvocationEntryBci 3213 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3214 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3215 3216 if( !GenerateSynchronizationCode ) 3217 return NULL; // Not locking things? 3218 if (stopped()) // Dead monitor? 3219 return NULL; 3220 3221 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3222 3223 // Box the stack location 3224 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3225 Node* mem = reset_memory(); 3226 3227 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3228 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3229 // Create the counters for this fast lock. 3230 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3231 } 3232 3233 // Create the rtm counters for this fast lock if needed. 3234 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3235 3236 // Add monitor to debug info for the slow path. If we block inside the 3237 // slow path and de-opt, we need the monitor hanging around 3238 map()->push_monitor( flock ); 3239 3240 const TypeFunc *tf = LockNode::lock_type(); 3241 LockNode *lock = new LockNode(C, tf); 3242 3243 lock->init_req( TypeFunc::Control, control() ); 3244 lock->init_req( TypeFunc::Memory , mem ); 3245 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3246 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3247 lock->init_req( TypeFunc::ReturnAdr, top() ); 3248 3249 lock->init_req(TypeFunc::Parms + 0, obj); 3250 lock->init_req(TypeFunc::Parms + 1, box); 3251 lock->init_req(TypeFunc::Parms + 2, flock); 3252 add_safepoint_edges(lock); 3253 3254 lock = _gvn.transform( lock )->as_Lock(); 3255 3256 // lock has no side-effects, sets few values 3257 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3258 3259 insert_mem_bar(Op_MemBarAcquireLock); 3260 3261 // Add this to the worklist so that the lock can be eliminated 3262 record_for_igvn(lock); 3263 3264 #ifndef PRODUCT 3265 if (PrintLockStatistics) { 3266 // Update the counter for this lock. Don't bother using an atomic 3267 // operation since we don't require absolute accuracy. 3268 lock->create_lock_counter(map()->jvms()); 3269 increment_counter(lock->counter()->addr()); 3270 } 3271 #endif 3272 3273 return flock; 3274 } 3275 3276 3277 //------------------------------shared_unlock---------------------------------- 3278 // Emit unlocking code. 3279 void GraphKit::shared_unlock(Node* box, Node* obj) { 3280 // bci is either a monitorenter bc or InvocationEntryBci 3281 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3282 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3283 3284 if( !GenerateSynchronizationCode ) 3285 return; 3286 if (stopped()) { // Dead monitor? 3287 map()->pop_monitor(); // Kill monitor from debug info 3288 return; 3289 } 3290 3291 // Memory barrier to avoid floating things down past the locked region 3292 insert_mem_bar(Op_MemBarReleaseLock); 3293 3294 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3295 UnlockNode *unlock = new UnlockNode(C, tf); 3296 uint raw_idx = Compile::AliasIdxRaw; 3297 unlock->init_req( TypeFunc::Control, control() ); 3298 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3299 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3300 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3301 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3302 3303 unlock->init_req(TypeFunc::Parms + 0, obj); 3304 unlock->init_req(TypeFunc::Parms + 1, box); 3305 unlock = _gvn.transform(unlock)->as_Unlock(); 3306 3307 Node* mem = reset_memory(); 3308 3309 // unlock has no side-effects, sets few values 3310 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3311 3312 // Kill monitor from debug info 3313 map()->pop_monitor( ); 3314 } 3315 3316 //-------------------------------get_layout_helper----------------------------- 3317 // If the given klass is a constant or known to be an array, 3318 // fetch the constant layout helper value into constant_value 3319 // and return (Node*)NULL. Otherwise, load the non-constant 3320 // layout helper value, and return the node which represents it. 3321 // This two-faced routine is useful because allocation sites 3322 // almost always feature constant types. 3323 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3324 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3325 if (!StressReflectiveCode && inst_klass != NULL) { 3326 ciKlass* klass = inst_klass->klass(); 3327 bool xklass = inst_klass->klass_is_exact(); 3328 if (xklass || klass->is_array_klass()) { 3329 jint lhelper = klass->layout_helper(); 3330 if (lhelper != Klass::_lh_neutral_value) { 3331 constant_value = lhelper; 3332 return (Node*) NULL; 3333 } 3334 } 3335 } 3336 constant_value = Klass::_lh_neutral_value; // put in a known value 3337 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3338 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3339 } 3340 3341 // We just put in an allocate/initialize with a big raw-memory effect. 3342 // Hook selected additional alias categories on the initialization. 3343 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3344 MergeMemNode* init_in_merge, 3345 Node* init_out_raw) { 3346 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3347 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3348 3349 Node* prevmem = kit.memory(alias_idx); 3350 init_in_merge->set_memory_at(alias_idx, prevmem); 3351 kit.set_memory(init_out_raw, alias_idx); 3352 } 3353 3354 //---------------------------set_output_for_allocation------------------------- 3355 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3356 const TypeOopPtr* oop_type) { 3357 int rawidx = Compile::AliasIdxRaw; 3358 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3359 add_safepoint_edges(alloc); 3360 Node* allocx = _gvn.transform(alloc); 3361 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3362 // create memory projection for i_o 3363 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3364 make_slow_call_ex(allocx, env()->Throwable_klass(), true); 3365 3366 // create a memory projection as for the normal control path 3367 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3368 set_memory(malloc, rawidx); 3369 3370 // a normal slow-call doesn't change i_o, but an allocation does 3371 // we create a separate i_o projection for the normal control path 3372 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3373 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3374 3375 // put in an initialization barrier 3376 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3377 rawoop)->as_Initialize(); 3378 assert(alloc->initialization() == init, "2-way macro link must work"); 3379 assert(init ->allocation() == alloc, "2-way macro link must work"); 3380 { 3381 // Extract memory strands which may participate in the new object's 3382 // initialization, and source them from the new InitializeNode. 3383 // This will allow us to observe initializations when they occur, 3384 // and link them properly (as a group) to the InitializeNode. 3385 assert(init->in(InitializeNode::Memory) == malloc, ""); 3386 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3387 init->set_req(InitializeNode::Memory, minit_in); 3388 record_for_igvn(minit_in); // fold it up later, if possible 3389 Node* minit_out = memory(rawidx); 3390 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3391 if (oop_type->isa_aryptr()) { 3392 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3393 int elemidx = C->get_alias_index(telemref); 3394 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3395 } else if (oop_type->isa_instptr()) { 3396 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3397 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3398 ciField* field = ik->nonstatic_field_at(i); 3399 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3400 continue; // do not bother to track really large numbers of fields 3401 // Find (or create) the alias category for this field: 3402 int fieldidx = C->alias_type(field)->index(); 3403 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3404 } 3405 } 3406 } 3407 3408 // Cast raw oop to the real thing... 3409 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3410 javaoop = _gvn.transform(javaoop); 3411 C->set_recent_alloc(control(), javaoop); 3412 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3413 3414 #ifdef ASSERT 3415 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3416 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3417 "Ideal_allocation works"); 3418 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3419 "Ideal_allocation works"); 3420 if (alloc->is_AllocateArray()) { 3421 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3422 "Ideal_allocation works"); 3423 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3424 "Ideal_allocation works"); 3425 } else { 3426 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3427 } 3428 } 3429 #endif //ASSERT 3430 3431 return javaoop; 3432 } 3433 3434 //---------------------------new_instance-------------------------------------- 3435 // This routine takes a klass_node which may be constant (for a static type) 3436 // or may be non-constant (for reflective code). It will work equally well 3437 // for either, and the graph will fold nicely if the optimizer later reduces 3438 // the type to a constant. 3439 // The optional arguments are for specialized use by intrinsics: 3440 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3441 // - If 'return_size_val', report the the total object size to the caller. 3442 Node* GraphKit::new_instance(Node* klass_node, 3443 Node* extra_slow_test, 3444 Node* *return_size_val) { 3445 // Compute size in doublewords 3446 // The size is always an integral number of doublewords, represented 3447 // as a positive bytewise size stored in the klass's layout_helper. 3448 // The layout_helper also encodes (in a low bit) the need for a slow path. 3449 jint layout_con = Klass::_lh_neutral_value; 3450 Node* layout_val = get_layout_helper(klass_node, layout_con); 3451 int layout_is_con = (layout_val == NULL); 3452 3453 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3454 // Generate the initial go-slow test. It's either ALWAYS (return a 3455 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3456 // case) a computed value derived from the layout_helper. 3457 Node* initial_slow_test = NULL; 3458 if (layout_is_con) { 3459 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3460 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3461 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3462 3463 } else { // reflective case 3464 // This reflective path is used by Unsafe.allocateInstance. 3465 // (It may be stress-tested by specifying StressReflectiveCode.) 3466 // Basically, we want to get into the VM is there's an illegal argument. 3467 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3468 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3469 if (extra_slow_test != intcon(0)) { 3470 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3471 } 3472 // (Macro-expander will further convert this to a Bool, if necessary.) 3473 } 3474 3475 // Find the size in bytes. This is easy; it's the layout_helper. 3476 // The size value must be valid even if the slow path is taken. 3477 Node* size = NULL; 3478 if (layout_is_con) { 3479 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3480 } else { // reflective case 3481 // This reflective path is used by clone and Unsafe.allocateInstance. 3482 size = ConvI2X(layout_val); 3483 3484 // Clear the low bits to extract layout_helper_size_in_bytes: 3485 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3486 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3487 size = _gvn.transform( new AndXNode(size, mask) ); 3488 } 3489 if (return_size_val != NULL) { 3490 (*return_size_val) = size; 3491 } 3492 3493 // This is a precise notnull oop of the klass. 3494 // (Actually, it need not be precise if this is a reflective allocation.) 3495 // It's what we cast the result to. 3496 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3497 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3498 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3499 3500 // Now generate allocation code 3501 3502 // The entire memory state is needed for slow path of the allocation 3503 // since GC and deoptimization can happened. 3504 Node *mem = reset_memory(); 3505 set_all_memory(mem); // Create new memory state 3506 3507 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3508 control(), mem, i_o(), 3509 size, klass_node, 3510 initial_slow_test); 3511 3512 return set_output_for_allocation(alloc, oop_type); 3513 } 3514 3515 //-------------------------------new_array------------------------------------- 3516 // helper for both newarray and anewarray 3517 // The 'length' parameter is (obviously) the length of the array. 3518 // See comments on new_instance for the meaning of the other arguments. 3519 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3520 Node* length, // number of array elements 3521 int nargs, // number of arguments to push back for uncommon trap 3522 Node* *return_size_val) { 3523 jint layout_con = Klass::_lh_neutral_value; 3524 Node* layout_val = get_layout_helper(klass_node, layout_con); 3525 int layout_is_con = (layout_val == NULL); 3526 3527 if (!layout_is_con && !StressReflectiveCode && 3528 !too_many_traps(Deoptimization::Reason_class_check)) { 3529 // This is a reflective array creation site. 3530 // Optimistically assume that it is a subtype of Object[], 3531 // so that we can fold up all the address arithmetic. 3532 layout_con = Klass::array_layout_helper(T_OBJECT); 3533 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3534 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3535 { BuildCutout unless(this, bol_lh, PROB_MAX); 3536 inc_sp(nargs); 3537 uncommon_trap(Deoptimization::Reason_class_check, 3538 Deoptimization::Action_maybe_recompile); 3539 } 3540 layout_val = NULL; 3541 layout_is_con = true; 3542 } 3543 3544 // Generate the initial go-slow test. Make sure we do not overflow 3545 // if length is huge (near 2Gig) or negative! We do not need 3546 // exact double-words here, just a close approximation of needed 3547 // double-words. We can't add any offset or rounding bits, lest we 3548 // take a size -1 of bytes and make it positive. Use an unsigned 3549 // compare, so negative sizes look hugely positive. 3550 int fast_size_limit = FastAllocateSizeLimit; 3551 if (layout_is_con) { 3552 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3553 // Increase the size limit if we have exact knowledge of array type. 3554 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3555 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3556 } 3557 3558 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3559 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3560 if (initial_slow_test->is_Bool()) { 3561 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3562 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3563 } 3564 3565 // --- Size Computation --- 3566 // array_size = round_to_heap(array_header + (length << elem_shift)); 3567 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3568 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3569 // The rounding mask is strength-reduced, if possible. 3570 int round_mask = MinObjAlignmentInBytes - 1; 3571 Node* header_size = NULL; 3572 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3573 // (T_BYTE has the weakest alignment and size restrictions...) 3574 if (layout_is_con) { 3575 int hsize = Klass::layout_helper_header_size(layout_con); 3576 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3577 BasicType etype = Klass::layout_helper_element_type(layout_con); 3578 if ((round_mask & ~right_n_bits(eshift)) == 0) 3579 round_mask = 0; // strength-reduce it if it goes away completely 3580 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3581 assert(header_size_min <= hsize, "generic minimum is smallest"); 3582 header_size_min = hsize; 3583 header_size = intcon(hsize + round_mask); 3584 } else { 3585 Node* hss = intcon(Klass::_lh_header_size_shift); 3586 Node* hsm = intcon(Klass::_lh_header_size_mask); 3587 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3588 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3589 Node* mask = intcon(round_mask); 3590 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3591 } 3592 3593 Node* elem_shift = NULL; 3594 if (layout_is_con) { 3595 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3596 if (eshift != 0) 3597 elem_shift = intcon(eshift); 3598 } else { 3599 // There is no need to mask or shift this value. 3600 // The semantics of LShiftINode include an implicit mask to 0x1F. 3601 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3602 elem_shift = layout_val; 3603 } 3604 3605 // Transition to native address size for all offset calculations: 3606 Node* lengthx = ConvI2X(length); 3607 Node* headerx = ConvI2X(header_size); 3608 #ifdef _LP64 3609 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3610 if (tllen != NULL && tllen->_lo < 0) { 3611 // Add a manual constraint to a positive range. Cf. array_element_address. 3612 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3613 if (size_max > tllen->_hi) size_max = tllen->_hi; 3614 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3615 lengthx = _gvn.transform( new ConvI2LNode(length, tlcon)); 3616 } 3617 } 3618 #endif 3619 3620 // Combine header size (plus rounding) and body size. Then round down. 3621 // This computation cannot overflow, because it is used only in two 3622 // places, one where the length is sharply limited, and the other 3623 // after a successful allocation. 3624 Node* abody = lengthx; 3625 if (elem_shift != NULL) 3626 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3627 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3628 if (round_mask != 0) { 3629 Node* mask = MakeConX(~round_mask); 3630 size = _gvn.transform( new AndXNode(size, mask) ); 3631 } 3632 // else if round_mask == 0, the size computation is self-rounding 3633 3634 if (return_size_val != NULL) { 3635 // This is the size 3636 (*return_size_val) = size; 3637 } 3638 3639 // Now generate allocation code 3640 3641 // The entire memory state is needed for slow path of the allocation 3642 // since GC and deoptimization can happened. 3643 Node *mem = reset_memory(); 3644 set_all_memory(mem); // Create new memory state 3645 3646 // Create the AllocateArrayNode and its result projections 3647 AllocateArrayNode* alloc 3648 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3649 control(), mem, i_o(), 3650 size, klass_node, 3651 initial_slow_test, 3652 length); 3653 3654 // Cast to correct type. Note that the klass_node may be constant or not, 3655 // and in the latter case the actual array type will be inexact also. 3656 // (This happens via a non-constant argument to inline_native_newArray.) 3657 // In any case, the value of klass_node provides the desired array type. 3658 const TypeInt* length_type = _gvn.find_int_type(length); 3659 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3660 if (ary_type->isa_aryptr() && length_type != NULL) { 3661 // Try to get a better type than POS for the size 3662 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3663 } 3664 3665 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3666 3667 // Cast length on remaining path to be as narrow as possible 3668 if (map()->find_edge(length) >= 0) { 3669 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3670 if (ccast != length) { 3671 _gvn.set_type_bottom(ccast); 3672 record_for_igvn(ccast); 3673 replace_in_map(length, ccast); 3674 } 3675 } 3676 3677 return javaoop; 3678 } 3679 3680 // The following "Ideal_foo" functions are placed here because they recognize 3681 // the graph shapes created by the functions immediately above. 3682 3683 //---------------------------Ideal_allocation---------------------------------- 3684 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3685 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3686 if (ptr == NULL) { // reduce dumb test in callers 3687 return NULL; 3688 } 3689 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3690 ptr = ptr->in(1); 3691 if (ptr == NULL) return NULL; 3692 } 3693 // Return NULL for allocations with several casts: 3694 // j.l.reflect.Array.newInstance(jobject, jint) 3695 // Object.clone() 3696 // to keep more precise type from last cast. 3697 if (ptr->is_Proj()) { 3698 Node* allo = ptr->in(0); 3699 if (allo != NULL && allo->is_Allocate()) { 3700 return allo->as_Allocate(); 3701 } 3702 } 3703 // Report failure to match. 3704 return NULL; 3705 } 3706 3707 // Fancy version which also strips off an offset (and reports it to caller). 3708 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3709 intptr_t& offset) { 3710 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3711 if (base == NULL) return NULL; 3712 return Ideal_allocation(base, phase); 3713 } 3714 3715 // Trace Initialize <- Proj[Parm] <- Allocate 3716 AllocateNode* InitializeNode::allocation() { 3717 Node* rawoop = in(InitializeNode::RawAddress); 3718 if (rawoop->is_Proj()) { 3719 Node* alloc = rawoop->in(0); 3720 if (alloc->is_Allocate()) { 3721 return alloc->as_Allocate(); 3722 } 3723 } 3724 return NULL; 3725 } 3726 3727 // Trace Allocate -> Proj[Parm] -> Initialize 3728 InitializeNode* AllocateNode::initialization() { 3729 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3730 if (rawoop == NULL) return NULL; 3731 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3732 Node* init = rawoop->fast_out(i); 3733 if (init->is_Initialize()) { 3734 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3735 return init->as_Initialize(); 3736 } 3737 } 3738 return NULL; 3739 } 3740 3741 //----------------------------- loop predicates --------------------------- 3742 3743 //------------------------------add_predicate_impl---------------------------- 3744 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3745 // Too many traps seen? 3746 if (too_many_traps(reason)) { 3747 #ifdef ASSERT 3748 if (TraceLoopPredicate) { 3749 int tc = C->trap_count(reason); 3750 tty->print("too many traps=%s tcount=%d in ", 3751 Deoptimization::trap_reason_name(reason), tc); 3752 method()->print(); // which method has too many predicate traps 3753 tty->cr(); 3754 } 3755 #endif 3756 // We cannot afford to take more traps here, 3757 // do not generate predicate. 3758 return; 3759 } 3760 3761 Node *cont = _gvn.intcon(1); 3762 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3763 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3764 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3765 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3766 C->add_predicate_opaq(opq); 3767 { 3768 PreserveJVMState pjvms(this); 3769 set_control(iffalse); 3770 inc_sp(nargs); 3771 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3772 } 3773 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3774 set_control(iftrue); 3775 } 3776 3777 //------------------------------add_predicate--------------------------------- 3778 void GraphKit::add_predicate(int nargs) { 3779 if (UseLoopPredicate) { 3780 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3781 } 3782 // loop's limit check predicate should be near the loop. 3783 if (LoopLimitCheck) { 3784 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3785 } 3786 } 3787 3788 //----------------------------- store barriers ---------------------------- 3789 #define __ ideal. 3790 3791 void GraphKit::sync_kit(IdealKit& ideal) { 3792 set_all_memory(__ merged_memory()); 3793 set_i_o(__ i_o()); 3794 set_control(__ ctrl()); 3795 } 3796 3797 void GraphKit::final_sync(IdealKit& ideal) { 3798 // Final sync IdealKit and graphKit. 3799 sync_kit(ideal); 3800 } 3801 3802 // vanilla/CMS post barrier 3803 // Insert a write-barrier store. This is to let generational GC work; we have 3804 // to flag all oop-stores before the next GC point. 3805 void GraphKit::write_barrier_post(Node* oop_store, 3806 Node* obj, 3807 Node* adr, 3808 uint adr_idx, 3809 Node* val, 3810 bool use_precise) { 3811 // No store check needed if we're storing a NULL or an old object 3812 // (latter case is probably a string constant). The concurrent 3813 // mark sweep garbage collector, however, needs to have all nonNull 3814 // oop updates flagged via card-marks. 3815 if (val != NULL && val->is_Con()) { 3816 // must be either an oop or NULL 3817 const Type* t = val->bottom_type(); 3818 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3819 // stores of null never (?) need barriers 3820 return; 3821 } 3822 3823 if (use_ReduceInitialCardMarks() 3824 && obj == just_allocated_object(control())) { 3825 // We can skip marks on a freshly-allocated object in Eden. 3826 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3827 // That routine informs GC to take appropriate compensating steps, 3828 // upon a slow-path allocation, so as to make this card-mark 3829 // elision safe. 3830 return; 3831 } 3832 3833 if (!use_precise) { 3834 // All card marks for a (non-array) instance are in one place: 3835 adr = obj; 3836 } 3837 // (Else it's an array (or unknown), and we want more precise card marks.) 3838 assert(adr != NULL, ""); 3839 3840 IdealKit ideal(this, true); 3841 3842 // Convert the pointer to an int prior to doing math on it 3843 Node* cast = __ CastPX(__ ctrl(), adr); 3844 3845 // Divide by card size 3846 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3847 "Only one we handle so far."); 3848 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3849 3850 // Combine card table base and card offset 3851 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3852 3853 // Get the alias_index for raw card-mark memory 3854 int adr_type = Compile::AliasIdxRaw; 3855 Node* zero = __ ConI(0); // Dirty card value 3856 BasicType bt = T_BYTE; 3857 3858 if (UseCondCardMark) { 3859 // The classic GC reference write barrier is typically implemented 3860 // as a store into the global card mark table. Unfortunately 3861 // unconditional stores can result in false sharing and excessive 3862 // coherence traffic as well as false transactional aborts. 3863 // UseCondCardMark enables MP "polite" conditional card mark 3864 // stores. In theory we could relax the load from ctrl() to 3865 // no_ctrl, but that doesn't buy much latitude. 3866 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3867 __ if_then(card_val, BoolTest::ne, zero); 3868 } 3869 3870 // Smash zero into card 3871 if( !UseConcMarkSweepGC ) { 3872 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); 3873 } else { 3874 // Specialized path for CM store barrier 3875 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3876 } 3877 3878 if (UseCondCardMark) { 3879 __ end_if(); 3880 } 3881 3882 // Final sync IdealKit and GraphKit. 3883 final_sync(ideal); 3884 } 3885 3886 // G1 pre/post barriers 3887 void GraphKit::g1_write_barrier_pre(bool do_load, 3888 Node* obj, 3889 Node* adr, 3890 uint alias_idx, 3891 Node* val, 3892 const TypeOopPtr* val_type, 3893 Node* pre_val, 3894 BasicType bt) { 3895 3896 // Some sanity checks 3897 // Note: val is unused in this routine. 3898 3899 if (do_load) { 3900 // We need to generate the load of the previous value 3901 assert(obj != NULL, "must have a base"); 3902 assert(adr != NULL, "where are loading from?"); 3903 assert(pre_val == NULL, "loaded already?"); 3904 assert(val_type != NULL, "need a type"); 3905 } else { 3906 // In this case both val_type and alias_idx are unused. 3907 assert(pre_val != NULL, "must be loaded already"); 3908 // Nothing to be done if pre_val is null. 3909 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3910 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3911 } 3912 assert(bt == T_OBJECT, "or we shouldn't be here"); 3913 3914 IdealKit ideal(this, true); 3915 3916 Node* tls = __ thread(); // ThreadLocalStorage 3917 3918 Node* no_ctrl = NULL; 3919 Node* no_base = __ top(); 3920 Node* zero = __ ConI(0); 3921 Node* zeroX = __ ConX(0); 3922 3923 float likely = PROB_LIKELY(0.999); 3924 float unlikely = PROB_UNLIKELY(0.999); 3925 3926 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3927 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3928 3929 // Offsets into the thread 3930 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3931 PtrQueue::byte_offset_of_active()); 3932 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3933 PtrQueue::byte_offset_of_index()); 3934 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3935 PtrQueue::byte_offset_of_buf()); 3936 3937 // Now the actual pointers into the thread 3938 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3939 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3940 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3941 3942 // Now some of the values 3943 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3944 3945 // if (!marking) 3946 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3947 BasicType index_bt = TypeX_X->basic_type(); 3948 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3949 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3950 3951 if (do_load) { 3952 // load original value 3953 // alias_idx correct?? 3954 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3955 } 3956 3957 // if (pre_val != NULL) 3958 __ if_then(pre_val, BoolTest::ne, null()); { 3959 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3960 3961 // is the queue for this thread full? 3962 __ if_then(index, BoolTest::ne, zeroX, likely); { 3963 3964 // decrement the index 3965 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 3966 3967 // Now get the buffer location we will log the previous value into and store it 3968 Node *log_addr = __ AddP(no_base, buffer, next_index); 3969 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 3970 // update the index 3971 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 3972 3973 } __ else_(); { 3974 3975 // logging buffer is full, call the runtime 3976 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3977 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3978 } __ end_if(); // (!index) 3979 } __ end_if(); // (pre_val != NULL) 3980 } __ end_if(); // (!marking) 3981 3982 // Final sync IdealKit and GraphKit. 3983 final_sync(ideal); 3984 } 3985 3986 // 3987 // Update the card table and add card address to the queue 3988 // 3989 void GraphKit::g1_mark_card(IdealKit& ideal, 3990 Node* card_adr, 3991 Node* oop_store, 3992 uint oop_alias_idx, 3993 Node* index, 3994 Node* index_adr, 3995 Node* buffer, 3996 const TypeFunc* tf) { 3997 3998 Node* zero = __ ConI(0); 3999 Node* zeroX = __ ConX(0); 4000 Node* no_base = __ top(); 4001 BasicType card_bt = T_BYTE; 4002 // Smash zero into card. MUST BE ORDERED WRT TO STORE 4003 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 4004 4005 // Now do the queue work 4006 __ if_then(index, BoolTest::ne, zeroX); { 4007 4008 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t)))); 4009 Node* log_addr = __ AddP(no_base, buffer, next_index); 4010 4011 // Order, see storeCM. 4012 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4013 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4014 4015 } __ else_(); { 4016 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4017 } __ end_if(); 4018 4019 } 4020 4021 void GraphKit::g1_write_barrier_post(Node* oop_store, 4022 Node* obj, 4023 Node* adr, 4024 uint alias_idx, 4025 Node* val, 4026 BasicType bt, 4027 bool use_precise) { 4028 // If we are writing a NULL then we need no post barrier 4029 4030 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4031 // Must be NULL 4032 const Type* t = val->bottom_type(); 4033 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4034 // No post barrier if writing NULLx 4035 return; 4036 } 4037 4038 if (!use_precise) { 4039 // All card marks for a (non-array) instance are in one place: 4040 adr = obj; 4041 } 4042 // (Else it's an array (or unknown), and we want more precise card marks.) 4043 assert(adr != NULL, ""); 4044 4045 IdealKit ideal(this, true); 4046 4047 Node* tls = __ thread(); // ThreadLocalStorage 4048 4049 Node* no_base = __ top(); 4050 float likely = PROB_LIKELY(0.999); 4051 float unlikely = PROB_UNLIKELY(0.999); 4052 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4053 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4054 Node* zeroX = __ ConX(0); 4055 4056 // Get the alias_index for raw card-mark memory 4057 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4058 4059 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4060 4061 // Offsets into the thread 4062 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4063 PtrQueue::byte_offset_of_index()); 4064 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4065 PtrQueue::byte_offset_of_buf()); 4066 4067 // Pointers into the thread 4068 4069 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4070 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4071 4072 // Now some values 4073 // Use ctrl to avoid hoisting these values past a safepoint, which could 4074 // potentially reset these fields in the JavaThread. 4075 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4076 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4077 4078 // Convert the store obj pointer to an int prior to doing math on it 4079 // Must use ctrl to prevent "integerized oop" existing across safepoint 4080 Node* cast = __ CastPX(__ ctrl(), adr); 4081 4082 // Divide pointer by card size 4083 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4084 4085 // Combine card table base and card offset 4086 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4087 4088 // If we know the value being stored does it cross regions? 4089 4090 if (val != NULL) { 4091 // Does the store cause us to cross regions? 4092 4093 // Should be able to do an unsigned compare of region_size instead of 4094 // and extra shift. Do we have an unsigned compare?? 4095 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4096 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4097 4098 // if (xor_res == 0) same region so skip 4099 __ if_then(xor_res, BoolTest::ne, zeroX); { 4100 4101 // No barrier if we are storing a NULL 4102 __ if_then(val, BoolTest::ne, null(), unlikely); { 4103 4104 // Ok must mark the card if not already dirty 4105 4106 // load the original value of the card 4107 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4108 4109 __ if_then(card_val, BoolTest::ne, young_card); { 4110 sync_kit(ideal); 4111 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4112 insert_mem_bar(Op_MemBarVolatile, oop_store); 4113 __ sync_kit(this); 4114 4115 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4116 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4117 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4118 } __ end_if(); 4119 } __ end_if(); 4120 } __ end_if(); 4121 } __ end_if(); 4122 } else { 4123 // Object.clone() instrinsic uses this path. 4124 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4125 } 4126 4127 // Final sync IdealKit and GraphKit. 4128 final_sync(ideal); 4129 } 4130 #undef __ 4131 4132 4133 4134 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4135 if (java_lang_String::has_offset_field()) { 4136 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4137 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4138 false, NULL, 0); 4139 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4140 int offset_field_idx = C->get_alias_index(offset_field_type); 4141 return make_load(ctrl, 4142 basic_plus_adr(str, str, offset_offset), 4143 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4144 } else { 4145 return intcon(0); 4146 } 4147 } 4148 4149 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4150 if (java_lang_String::has_count_field()) { 4151 int count_offset = java_lang_String::count_offset_in_bytes(); 4152 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4153 false, NULL, 0); 4154 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4155 int count_field_idx = C->get_alias_index(count_field_type); 4156 return make_load(ctrl, 4157 basic_plus_adr(str, str, count_offset), 4158 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4159 } else { 4160 return load_array_length(load_String_value(ctrl, str)); 4161 } 4162 } 4163 4164 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4165 int value_offset = java_lang_String::value_offset_in_bytes(); 4166 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4167 false, NULL, 0); 4168 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4169 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4170 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4171 ciTypeArrayKlass::make(T_CHAR), true, 0); 4172 int value_field_idx = C->get_alias_index(value_field_type); 4173 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4174 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4175 // String.value field is known to be @Stable. 4176 if (UseImplicitStableValues) { 4177 load = cast_array_to_stable(load, value_type); 4178 } 4179 return load; 4180 } 4181 4182 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4183 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4184 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4185 false, NULL, 0); 4186 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4187 int offset_field_idx = C->get_alias_index(offset_field_type); 4188 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4189 value, T_INT, offset_field_idx, MemNode::unordered); 4190 } 4191 4192 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4193 int value_offset = java_lang_String::value_offset_in_bytes(); 4194 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4195 false, NULL, 0); 4196 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4197 4198 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4199 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4200 } 4201 4202 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4203 int count_offset = java_lang_String::count_offset_in_bytes(); 4204 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4205 false, NULL, 0); 4206 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4207 int count_field_idx = C->get_alias_index(count_field_type); 4208 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4209 value, T_INT, count_field_idx, MemNode::unordered); 4210 } 4211 4212 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4213 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4214 // assumption of CCP analysis. 4215 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 4216 }