1 /* 2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "memory/barrierSet.hpp" 31 #include "memory/cardTableModRefBS.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/graphKit.hpp" 34 #include "opto/idealKit.hpp" 35 #include "opto/locknode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/parse.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "runtime/deoptimization.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 43 //----------------------------GraphKit----------------------------------------- 44 // Main utility constructor. 45 GraphKit::GraphKit(JVMState* jvms) 46 : Phase(Phase::Parser), 47 _env(C->env()), 48 _gvn(*C->initial_gvn()) 49 { 50 _exceptions = jvms->map()->next_exception(); 51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 52 set_jvms(jvms); 53 } 54 55 // Private constructor for parser. 56 GraphKit::GraphKit() 57 : Phase(Phase::Parser), 58 _env(C->env()), 59 _gvn(*C->initial_gvn()) 60 { 61 _exceptions = NULL; 62 set_map(NULL); 63 debug_only(_sp = -99); 64 debug_only(set_bci(-99)); 65 } 66 67 68 69 //---------------------------clean_stack--------------------------------------- 70 // Clear away rubbish from the stack area of the JVM state. 71 // This destroys any arguments that may be waiting on the stack. 72 void GraphKit::clean_stack(int from_sp) { 73 SafePointNode* map = this->map(); 74 JVMState* jvms = this->jvms(); 75 int stk_size = jvms->stk_size(); 76 int stkoff = jvms->stkoff(); 77 Node* top = this->top(); 78 for (int i = from_sp; i < stk_size; i++) { 79 if (map->in(stkoff + i) != top) { 80 map->set_req(stkoff + i, top); 81 } 82 } 83 } 84 85 86 //--------------------------------sync_jvms----------------------------------- 87 // Make sure our current jvms agrees with our parse state. 88 JVMState* GraphKit::sync_jvms() const { 89 JVMState* jvms = this->jvms(); 90 jvms->set_bci(bci()); // Record the new bci in the JVMState 91 jvms->set_sp(sp()); // Record the new sp in the JVMState 92 assert(jvms_in_sync(), "jvms is now in sync"); 93 return jvms; 94 } 95 96 //--------------------------------sync_jvms_for_reexecute--------------------- 97 // Make sure our current jvms agrees with our parse state. This version 98 // uses the reexecute_sp for reexecuting bytecodes. 99 JVMState* GraphKit::sync_jvms_for_reexecute() { 100 JVMState* jvms = this->jvms(); 101 jvms->set_bci(bci()); // Record the new bci in the JVMState 102 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 103 return jvms; 104 } 105 106 #ifdef ASSERT 107 bool GraphKit::jvms_in_sync() const { 108 Parse* parse = is_Parse(); 109 if (parse == NULL) { 110 if (bci() != jvms()->bci()) return false; 111 if (sp() != (int)jvms()->sp()) return false; 112 return true; 113 } 114 if (jvms()->method() != parse->method()) return false; 115 if (jvms()->bci() != parse->bci()) return false; 116 int jvms_sp = jvms()->sp(); 117 if (jvms_sp != parse->sp()) return false; 118 int jvms_depth = jvms()->depth(); 119 if (jvms_depth != parse->depth()) return false; 120 return true; 121 } 122 123 // Local helper checks for special internal merge points 124 // used to accumulate and merge exception states. 125 // They are marked by the region's in(0) edge being the map itself. 126 // Such merge points must never "escape" into the parser at large, 127 // until they have been handed to gvn.transform. 128 static bool is_hidden_merge(Node* reg) { 129 if (reg == NULL) return false; 130 if (reg->is_Phi()) { 131 reg = reg->in(0); 132 if (reg == NULL) return false; 133 } 134 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 135 } 136 137 void GraphKit::verify_map() const { 138 if (map() == NULL) return; // null map is OK 139 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 140 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 141 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 142 } 143 144 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 145 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 146 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 147 } 148 #endif 149 150 //---------------------------stop_and_kill_map--------------------------------- 151 // Set _map to NULL, signalling a stop to further bytecode execution. 152 // First smash the current map's control to a constant, to mark it dead. 153 void GraphKit::stop_and_kill_map() { 154 SafePointNode* dead_map = stop(); 155 if (dead_map != NULL) { 156 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 157 assert(dead_map->is_killed(), "must be so marked"); 158 } 159 } 160 161 162 //--------------------------------stopped-------------------------------------- 163 // Tell if _map is NULL, or control is top. 164 bool GraphKit::stopped() { 165 if (map() == NULL) return true; 166 else if (control() == top()) return true; 167 else return false; 168 } 169 170 171 //-----------------------------has_ex_handler---------------------------------- 172 // Tell if this method or any caller method has exception handlers. 173 bool GraphKit::has_ex_handler() { 174 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 175 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 176 return true; 177 } 178 } 179 return false; 180 } 181 182 //------------------------------save_ex_oop------------------------------------ 183 // Save an exception without blowing stack contents or other JVM state. 184 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 185 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 186 ex_map->add_req(ex_oop); 187 debug_only(verify_exception_state(ex_map)); 188 } 189 190 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 191 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 192 Node* ex_oop = ex_map->in(ex_map->req()-1); 193 if (clear_it) ex_map->del_req(ex_map->req()-1); 194 return ex_oop; 195 } 196 197 //-----------------------------saved_ex_oop------------------------------------ 198 // Recover a saved exception from its map. 199 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 200 return common_saved_ex_oop(ex_map, false); 201 } 202 203 //--------------------------clear_saved_ex_oop--------------------------------- 204 // Erase a previously saved exception from its map. 205 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 206 return common_saved_ex_oop(ex_map, true); 207 } 208 209 #ifdef ASSERT 210 //---------------------------has_saved_ex_oop---------------------------------- 211 // Erase a previously saved exception from its map. 212 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 213 return ex_map->req() == ex_map->jvms()->endoff()+1; 214 } 215 #endif 216 217 //-------------------------make_exception_state-------------------------------- 218 // Turn the current JVM state into an exception state, appending the ex_oop. 219 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 220 sync_jvms(); 221 SafePointNode* ex_map = stop(); // do not manipulate this map any more 222 set_saved_ex_oop(ex_map, ex_oop); 223 return ex_map; 224 } 225 226 227 //--------------------------add_exception_state-------------------------------- 228 // Add an exception to my list of exceptions. 229 void GraphKit::add_exception_state(SafePointNode* ex_map) { 230 if (ex_map == NULL || ex_map->control() == top()) { 231 return; 232 } 233 #ifdef ASSERT 234 verify_exception_state(ex_map); 235 if (has_exceptions()) { 236 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 237 } 238 #endif 239 240 // If there is already an exception of exactly this type, merge with it. 241 // In particular, null-checks and other low-level exceptions common up here. 242 Node* ex_oop = saved_ex_oop(ex_map); 243 const Type* ex_type = _gvn.type(ex_oop); 244 if (ex_oop == top()) { 245 // No action needed. 246 return; 247 } 248 assert(ex_type->isa_instptr(), "exception must be an instance"); 249 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 250 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 251 // We check sp also because call bytecodes can generate exceptions 252 // both before and after arguments are popped! 253 if (ex_type2 == ex_type 254 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 255 combine_exception_states(ex_map, e2); 256 return; 257 } 258 } 259 260 // No pre-existing exception of the same type. Chain it on the list. 261 push_exception_state(ex_map); 262 } 263 264 //-----------------------add_exception_states_from----------------------------- 265 void GraphKit::add_exception_states_from(JVMState* jvms) { 266 SafePointNode* ex_map = jvms->map()->next_exception(); 267 if (ex_map != NULL) { 268 jvms->map()->set_next_exception(NULL); 269 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 270 next_map = ex_map->next_exception(); 271 ex_map->set_next_exception(NULL); 272 add_exception_state(ex_map); 273 } 274 } 275 } 276 277 //-----------------------transfer_exceptions_into_jvms------------------------- 278 JVMState* GraphKit::transfer_exceptions_into_jvms() { 279 if (map() == NULL) { 280 // We need a JVMS to carry the exceptions, but the map has gone away. 281 // Create a scratch JVMS, cloned from any of the exception states... 282 if (has_exceptions()) { 283 _map = _exceptions; 284 _map = clone_map(); 285 _map->set_next_exception(NULL); 286 clear_saved_ex_oop(_map); 287 debug_only(verify_map()); 288 } else { 289 // ...or created from scratch 290 JVMState* jvms = new (C) JVMState(_method, NULL); 291 jvms->set_bci(_bci); 292 jvms->set_sp(_sp); 293 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms)); 294 set_jvms(jvms); 295 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 296 set_all_memory(top()); 297 while (map()->req() < jvms->endoff()) map()->add_req(top()); 298 } 299 // (This is a kludge, in case you didn't notice.) 300 set_control(top()); 301 } 302 JVMState* jvms = sync_jvms(); 303 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 304 jvms->map()->set_next_exception(_exceptions); 305 _exceptions = NULL; // done with this set of exceptions 306 return jvms; 307 } 308 309 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 310 assert(is_hidden_merge(dstphi), "must be a special merge node"); 311 assert(is_hidden_merge(srcphi), "must be a special merge node"); 312 uint limit = srcphi->req(); 313 for (uint i = PhiNode::Input; i < limit; i++) { 314 dstphi->add_req(srcphi->in(i)); 315 } 316 } 317 static inline void add_one_req(Node* dstphi, Node* src) { 318 assert(is_hidden_merge(dstphi), "must be a special merge node"); 319 assert(!is_hidden_merge(src), "must not be a special merge node"); 320 dstphi->add_req(src); 321 } 322 323 //-----------------------combine_exception_states------------------------------ 324 // This helper function combines exception states by building phis on a 325 // specially marked state-merging region. These regions and phis are 326 // untransformed, and can build up gradually. The region is marked by 327 // having a control input of its exception map, rather than NULL. Such 328 // regions do not appear except in this function, and in use_exception_state. 329 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 330 if (failing()) return; // dying anyway... 331 JVMState* ex_jvms = ex_map->_jvms; 332 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 333 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 334 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 335 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 336 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 337 assert(ex_map->req() == phi_map->req(), "matching maps"); 338 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 339 Node* hidden_merge_mark = root(); 340 Node* region = phi_map->control(); 341 MergeMemNode* phi_mem = phi_map->merged_memory(); 342 MergeMemNode* ex_mem = ex_map->merged_memory(); 343 if (region->in(0) != hidden_merge_mark) { 344 // The control input is not (yet) a specially-marked region in phi_map. 345 // Make it so, and build some phis. 346 region = new (C) RegionNode(2); 347 _gvn.set_type(region, Type::CONTROL); 348 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 349 region->init_req(1, phi_map->control()); 350 phi_map->set_control(region); 351 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 352 record_for_igvn(io_phi); 353 _gvn.set_type(io_phi, Type::ABIO); 354 phi_map->set_i_o(io_phi); 355 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 356 Node* m = mms.memory(); 357 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 358 record_for_igvn(m_phi); 359 _gvn.set_type(m_phi, Type::MEMORY); 360 mms.set_memory(m_phi); 361 } 362 } 363 364 // Either or both of phi_map and ex_map might already be converted into phis. 365 Node* ex_control = ex_map->control(); 366 // if there is special marking on ex_map also, we add multiple edges from src 367 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 368 // how wide was the destination phi_map, originally? 369 uint orig_width = region->req(); 370 371 if (add_multiple) { 372 add_n_reqs(region, ex_control); 373 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 374 } else { 375 // ex_map has no merges, so we just add single edges everywhere 376 add_one_req(region, ex_control); 377 add_one_req(phi_map->i_o(), ex_map->i_o()); 378 } 379 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 380 if (mms.is_empty()) { 381 // get a copy of the base memory, and patch some inputs into it 382 const TypePtr* adr_type = mms.adr_type(C); 383 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 384 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 385 mms.set_memory(phi); 386 // Prepare to append interesting stuff onto the newly sliced phi: 387 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 388 } 389 // Append stuff from ex_map: 390 if (add_multiple) { 391 add_n_reqs(mms.memory(), mms.memory2()); 392 } else { 393 add_one_req(mms.memory(), mms.memory2()); 394 } 395 } 396 uint limit = ex_map->req(); 397 for (uint i = TypeFunc::Parms; i < limit; i++) { 398 // Skip everything in the JVMS after tos. (The ex_oop follows.) 399 if (i == tos) i = ex_jvms->monoff(); 400 Node* src = ex_map->in(i); 401 Node* dst = phi_map->in(i); 402 if (src != dst) { 403 PhiNode* phi; 404 if (dst->in(0) != region) { 405 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 406 record_for_igvn(phi); 407 _gvn.set_type(phi, phi->type()); 408 phi_map->set_req(i, dst); 409 // Prepare to append interesting stuff onto the new phi: 410 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 411 } else { 412 assert(dst->is_Phi(), "nobody else uses a hidden region"); 413 phi = dst->as_Phi(); 414 } 415 if (add_multiple && src->in(0) == ex_control) { 416 // Both are phis. 417 add_n_reqs(dst, src); 418 } else { 419 while (dst->req() < region->req()) add_one_req(dst, src); 420 } 421 const Type* srctype = _gvn.type(src); 422 if (phi->type() != srctype) { 423 const Type* dsttype = phi->type()->meet_speculative(srctype); 424 if (phi->type() != dsttype) { 425 phi->set_type(dsttype); 426 _gvn.set_type(phi, dsttype); 427 } 428 } 429 } 430 } 431 phi_map->merge_replaced_nodes_with(ex_map); 432 } 433 434 //--------------------------use_exception_state-------------------------------- 435 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 436 if (failing()) { stop(); return top(); } 437 Node* region = phi_map->control(); 438 Node* hidden_merge_mark = root(); 439 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 440 Node* ex_oop = clear_saved_ex_oop(phi_map); 441 if (region->in(0) == hidden_merge_mark) { 442 // Special marking for internal ex-states. Process the phis now. 443 region->set_req(0, region); // now it's an ordinary region 444 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 445 // Note: Setting the jvms also sets the bci and sp. 446 set_control(_gvn.transform(region)); 447 uint tos = jvms()->stkoff() + sp(); 448 for (uint i = 1; i < tos; i++) { 449 Node* x = phi_map->in(i); 450 if (x->in(0) == region) { 451 assert(x->is_Phi(), "expected a special phi"); 452 phi_map->set_req(i, _gvn.transform(x)); 453 } 454 } 455 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 456 Node* x = mms.memory(); 457 if (x->in(0) == region) { 458 assert(x->is_Phi(), "nobody else uses a hidden region"); 459 mms.set_memory(_gvn.transform(x)); 460 } 461 } 462 if (ex_oop->in(0) == region) { 463 assert(ex_oop->is_Phi(), "expected a special phi"); 464 ex_oop = _gvn.transform(ex_oop); 465 } 466 } else { 467 set_jvms(phi_map->jvms()); 468 } 469 470 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 471 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 472 return ex_oop; 473 } 474 475 //---------------------------------java_bc------------------------------------- 476 Bytecodes::Code GraphKit::java_bc() const { 477 ciMethod* method = this->method(); 478 int bci = this->bci(); 479 if (method != NULL && bci != InvocationEntryBci) 480 return method->java_code_at_bci(bci); 481 else 482 return Bytecodes::_illegal; 483 } 484 485 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 486 bool must_throw) { 487 // if the exception capability is set, then we will generate code 488 // to check the JavaThread.should_post_on_exceptions flag to see 489 // if we actually need to report exception events (for this 490 // thread). If we don't need to report exception events, we will 491 // take the normal fast path provided by add_exception_events. If 492 // exception event reporting is enabled for this thread, we will 493 // take the uncommon_trap in the BuildCutout below. 494 495 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 496 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 497 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 498 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 499 500 // Test the should_post_on_exceptions_flag vs. 0 501 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 502 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 503 504 // Branch to slow_path if should_post_on_exceptions_flag was true 505 { BuildCutout unless(this, tst, PROB_MAX); 506 // Do not try anything fancy if we're notifying the VM on every throw. 507 // Cf. case Bytecodes::_athrow in parse2.cpp. 508 uncommon_trap(reason, Deoptimization::Action_none, 509 (ciKlass*)NULL, (char*)NULL, must_throw); 510 } 511 512 } 513 514 //------------------------------builtin_throw---------------------------------- 515 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 516 bool must_throw = true; 517 518 if (env()->jvmti_can_post_on_exceptions()) { 519 // check if we must post exception events, take uncommon trap if so 520 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 521 // here if should_post_on_exceptions is false 522 // continue on with the normal codegen 523 } 524 525 // If this particular condition has not yet happened at this 526 // bytecode, then use the uncommon trap mechanism, and allow for 527 // a future recompilation if several traps occur here. 528 // If the throw is hot, try to use a more complicated inline mechanism 529 // which keeps execution inside the compiled code. 530 bool treat_throw_as_hot = false; 531 ciMethodData* md = method()->method_data(); 532 533 if (ProfileTraps) { 534 if (too_many_traps(reason)) { 535 treat_throw_as_hot = true; 536 } 537 // (If there is no MDO at all, assume it is early in 538 // execution, and that any deopts are part of the 539 // startup transient, and don't need to be remembered.) 540 541 // Also, if there is a local exception handler, treat all throws 542 // as hot if there has been at least one in this method. 543 if (C->trap_count(reason) != 0 544 && method()->method_data()->trap_count(reason) != 0 545 && has_ex_handler()) { 546 treat_throw_as_hot = true; 547 } 548 } 549 550 // If this throw happens frequently, an uncommon trap might cause 551 // a performance pothole. If there is a local exception handler, 552 // and if this particular bytecode appears to be deoptimizing often, 553 // let us handle the throw inline, with a preconstructed instance. 554 // Note: If the deopt count has blown up, the uncommon trap 555 // runtime is going to flush this nmethod, not matter what. 556 if (treat_throw_as_hot 557 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 558 // If the throw is local, we use a pre-existing instance and 559 // punt on the backtrace. This would lead to a missing backtrace 560 // (a repeat of 4292742) if the backtrace object is ever asked 561 // for its backtrace. 562 // Fixing this remaining case of 4292742 requires some flavor of 563 // escape analysis. Leave that for the future. 564 ciInstance* ex_obj = NULL; 565 switch (reason) { 566 case Deoptimization::Reason_null_check: 567 ex_obj = env()->NullPointerException_instance(); 568 break; 569 case Deoptimization::Reason_div0_check: 570 ex_obj = env()->ArithmeticException_instance(); 571 break; 572 case Deoptimization::Reason_range_check: 573 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 574 break; 575 case Deoptimization::Reason_class_check: 576 if (java_bc() == Bytecodes::_aastore) { 577 ex_obj = env()->ArrayStoreException_instance(); 578 } else { 579 ex_obj = env()->ClassCastException_instance(); 580 } 581 break; 582 } 583 if (failing()) { stop(); return; } // exception allocation might fail 584 if (ex_obj != NULL) { 585 // Cheat with a preallocated exception object. 586 if (C->log() != NULL) 587 C->log()->elem("hot_throw preallocated='1' reason='%s'", 588 Deoptimization::trap_reason_name(reason)); 589 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 590 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 591 592 // Clear the detail message of the preallocated exception object. 593 // Weblogic sometimes mutates the detail message of exceptions 594 // using reflection. 595 int offset = java_lang_Throwable::get_detailMessage_offset(); 596 const TypePtr* adr_typ = ex_con->add_offset(offset); 597 598 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 599 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 600 // Conservatively release stores of object references. 601 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 602 603 add_exception_state(make_exception_state(ex_node)); 604 return; 605 } 606 } 607 608 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 609 // It won't be much cheaper than bailing to the interp., since we'll 610 // have to pass up all the debug-info, and the runtime will have to 611 // create the stack trace. 612 613 // Usual case: Bail to interpreter. 614 // Reserve the right to recompile if we haven't seen anything yet. 615 616 assert(!Deoptimization::reason_is_speculate(reason), "unsupported"); 617 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 618 if (treat_throw_as_hot 619 && (method()->method_data()->trap_recompiled_at(bci(), NULL) 620 || C->too_many_traps(reason))) { 621 // We cannot afford to take more traps here. Suffer in the interpreter. 622 if (C->log() != NULL) 623 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 624 Deoptimization::trap_reason_name(reason), 625 C->trap_count(reason)); 626 action = Deoptimization::Action_none; 627 } 628 629 // "must_throw" prunes the JVM state to include only the stack, if there 630 // are no local exception handlers. This should cut down on register 631 // allocation time and code size, by drastically reducing the number 632 // of in-edges on the call to the uncommon trap. 633 634 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 635 } 636 637 638 //----------------------------PreserveJVMState--------------------------------- 639 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 640 debug_only(kit->verify_map()); 641 _kit = kit; 642 _map = kit->map(); // preserve the map 643 _sp = kit->sp(); 644 kit->set_map(clone_map ? kit->clone_map() : NULL); 645 #ifdef ASSERT 646 _bci = kit->bci(); 647 Parse* parser = kit->is_Parse(); 648 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 649 _block = block; 650 #endif 651 } 652 PreserveJVMState::~PreserveJVMState() { 653 GraphKit* kit = _kit; 654 #ifdef ASSERT 655 assert(kit->bci() == _bci, "bci must not shift"); 656 Parse* parser = kit->is_Parse(); 657 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 658 assert(block == _block, "block must not shift"); 659 #endif 660 kit->set_map(_map); 661 kit->set_sp(_sp); 662 } 663 664 665 //-----------------------------BuildCutout------------------------------------- 666 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 667 : PreserveJVMState(kit) 668 { 669 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 670 SafePointNode* outer_map = _map; // preserved map is caller's 671 SafePointNode* inner_map = kit->map(); 672 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 673 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 674 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 675 } 676 BuildCutout::~BuildCutout() { 677 GraphKit* kit = _kit; 678 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 679 } 680 681 //---------------------------PreserveReexecuteState---------------------------- 682 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 683 assert(!kit->stopped(), "must call stopped() before"); 684 _kit = kit; 685 _sp = kit->sp(); 686 _reexecute = kit->jvms()->_reexecute; 687 } 688 PreserveReexecuteState::~PreserveReexecuteState() { 689 if (_kit->stopped()) return; 690 _kit->jvms()->_reexecute = _reexecute; 691 _kit->set_sp(_sp); 692 } 693 694 //------------------------------clone_map-------------------------------------- 695 // Implementation of PreserveJVMState 696 // 697 // Only clone_map(...) here. If this function is only used in the 698 // PreserveJVMState class we may want to get rid of this extra 699 // function eventually and do it all there. 700 701 SafePointNode* GraphKit::clone_map() { 702 if (map() == NULL) return NULL; 703 704 // Clone the memory edge first 705 Node* mem = MergeMemNode::make(C, map()->memory()); 706 gvn().set_type_bottom(mem); 707 708 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 709 JVMState* jvms = this->jvms(); 710 JVMState* clonejvms = jvms->clone_shallow(C); 711 clonemap->set_memory(mem); 712 clonemap->set_jvms(clonejvms); 713 clonejvms->set_map(clonemap); 714 record_for_igvn(clonemap); 715 gvn().set_type_bottom(clonemap); 716 return clonemap; 717 } 718 719 720 //-----------------------------set_map_clone----------------------------------- 721 void GraphKit::set_map_clone(SafePointNode* m) { 722 _map = m; 723 _map = clone_map(); 724 _map->set_next_exception(NULL); 725 debug_only(verify_map()); 726 } 727 728 729 //----------------------------kill_dead_locals--------------------------------- 730 // Detect any locals which are known to be dead, and force them to top. 731 void GraphKit::kill_dead_locals() { 732 // Consult the liveness information for the locals. If any 733 // of them are unused, then they can be replaced by top(). This 734 // should help register allocation time and cut down on the size 735 // of the deoptimization information. 736 737 // This call is made from many of the bytecode handling 738 // subroutines called from the Big Switch in do_one_bytecode. 739 // Every bytecode which might include a slow path is responsible 740 // for killing its dead locals. The more consistent we 741 // are about killing deads, the fewer useless phis will be 742 // constructed for them at various merge points. 743 744 // bci can be -1 (InvocationEntryBci). We return the entry 745 // liveness for the method. 746 747 if (method() == NULL || method()->code_size() == 0) { 748 // We are building a graph for a call to a native method. 749 // All locals are live. 750 return; 751 } 752 753 ResourceMark rm; 754 755 // Consult the liveness information for the locals. If any 756 // of them are unused, then they can be replaced by top(). This 757 // should help register allocation time and cut down on the size 758 // of the deoptimization information. 759 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 760 761 int len = (int)live_locals.size(); 762 assert(len <= jvms()->loc_size(), "too many live locals"); 763 for (int local = 0; local < len; local++) { 764 if (!live_locals.at(local)) { 765 set_local(local, top()); 766 } 767 } 768 } 769 770 #ifdef ASSERT 771 //-------------------------dead_locals_are_killed------------------------------ 772 // Return true if all dead locals are set to top in the map. 773 // Used to assert "clean" debug info at various points. 774 bool GraphKit::dead_locals_are_killed() { 775 if (method() == NULL || method()->code_size() == 0) { 776 // No locals need to be dead, so all is as it should be. 777 return true; 778 } 779 780 // Make sure somebody called kill_dead_locals upstream. 781 ResourceMark rm; 782 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 783 if (jvms->loc_size() == 0) continue; // no locals to consult 784 SafePointNode* map = jvms->map(); 785 ciMethod* method = jvms->method(); 786 int bci = jvms->bci(); 787 if (jvms == this->jvms()) { 788 bci = this->bci(); // it might not yet be synched 789 } 790 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 791 int len = (int)live_locals.size(); 792 if (!live_locals.is_valid() || len == 0) 793 // This method is trivial, or is poisoned by a breakpoint. 794 return true; 795 assert(len == jvms->loc_size(), "live map consistent with locals map"); 796 for (int local = 0; local < len; local++) { 797 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 798 if (PrintMiscellaneous && (Verbose || WizardMode)) { 799 tty->print_cr("Zombie local %d: ", local); 800 jvms->dump(); 801 } 802 return false; 803 } 804 } 805 } 806 return true; 807 } 808 809 #endif //ASSERT 810 811 // Helper function for enforcing certain bytecodes to reexecute if 812 // deoptimization happens 813 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 814 ciMethod* cur_method = jvms->method(); 815 int cur_bci = jvms->bci(); 816 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 817 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 818 return Interpreter::bytecode_should_reexecute(code) || 819 is_anewarray && code == Bytecodes::_multianewarray; 820 // Reexecute _multianewarray bytecode which was replaced with 821 // sequence of [a]newarray. See Parse::do_multianewarray(). 822 // 823 // Note: interpreter should not have it set since this optimization 824 // is limited by dimensions and guarded by flag so in some cases 825 // multianewarray() runtime calls will be generated and 826 // the bytecode should not be reexecutes (stack will not be reset). 827 } else 828 return false; 829 } 830 831 // Helper function for adding JVMState and debug information to node 832 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 833 // Add the safepoint edges to the call (or other safepoint). 834 835 // Make sure dead locals are set to top. This 836 // should help register allocation time and cut down on the size 837 // of the deoptimization information. 838 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 839 840 // Walk the inline list to fill in the correct set of JVMState's 841 // Also fill in the associated edges for each JVMState. 842 843 // If the bytecode needs to be reexecuted we need to put 844 // the arguments back on the stack. 845 const bool should_reexecute = jvms()->should_reexecute(); 846 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 847 848 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 849 // undefined if the bci is different. This is normal for Parse but it 850 // should not happen for LibraryCallKit because only one bci is processed. 851 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 852 "in LibraryCallKit the reexecute bit should not change"); 853 854 // If we are guaranteed to throw, we can prune everything but the 855 // input to the current bytecode. 856 bool can_prune_locals = false; 857 uint stack_slots_not_pruned = 0; 858 int inputs = 0, depth = 0; 859 if (must_throw) { 860 assert(method() == youngest_jvms->method(), "sanity"); 861 if (compute_stack_effects(inputs, depth)) { 862 can_prune_locals = true; 863 stack_slots_not_pruned = inputs; 864 } 865 } 866 867 if (env()->jvmti_can_access_local_variables()) { 868 // At any safepoint, this method can get breakpointed, which would 869 // then require an immediate deoptimization. 870 can_prune_locals = false; // do not prune locals 871 stack_slots_not_pruned = 0; 872 } 873 874 // do not scribble on the input jvms 875 JVMState* out_jvms = youngest_jvms->clone_deep(C); 876 call->set_jvms(out_jvms); // Start jvms list for call node 877 878 // For a known set of bytecodes, the interpreter should reexecute them if 879 // deoptimization happens. We set the reexecute state for them here 880 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 881 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 882 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 883 } 884 885 // Presize the call: 886 DEBUG_ONLY(uint non_debug_edges = call->req()); 887 call->add_req_batch(top(), youngest_jvms->debug_depth()); 888 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 889 890 // Set up edges so that the call looks like this: 891 // Call [state:] ctl io mem fptr retadr 892 // [parms:] parm0 ... parmN 893 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 894 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 895 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 896 // Note that caller debug info precedes callee debug info. 897 898 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 899 uint debug_ptr = call->req(); 900 901 // Loop over the map input edges associated with jvms, add them 902 // to the call node, & reset all offsets to match call node array. 903 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 904 uint debug_end = debug_ptr; 905 uint debug_start = debug_ptr - in_jvms->debug_size(); 906 debug_ptr = debug_start; // back up the ptr 907 908 uint p = debug_start; // walks forward in [debug_start, debug_end) 909 uint j, k, l; 910 SafePointNode* in_map = in_jvms->map(); 911 out_jvms->set_map(call); 912 913 if (can_prune_locals) { 914 assert(in_jvms->method() == out_jvms->method(), "sanity"); 915 // If the current throw can reach an exception handler in this JVMS, 916 // then we must keep everything live that can reach that handler. 917 // As a quick and dirty approximation, we look for any handlers at all. 918 if (in_jvms->method()->has_exception_handlers()) { 919 can_prune_locals = false; 920 } 921 } 922 923 // Add the Locals 924 k = in_jvms->locoff(); 925 l = in_jvms->loc_size(); 926 out_jvms->set_locoff(p); 927 if (!can_prune_locals) { 928 for (j = 0; j < l; j++) 929 call->set_req(p++, in_map->in(k+j)); 930 } else { 931 p += l; // already set to top above by add_req_batch 932 } 933 934 // Add the Expression Stack 935 k = in_jvms->stkoff(); 936 l = in_jvms->sp(); 937 out_jvms->set_stkoff(p); 938 if (!can_prune_locals) { 939 for (j = 0; j < l; j++) 940 call->set_req(p++, in_map->in(k+j)); 941 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 942 // Divide stack into {S0,...,S1}, where S0 is set to top. 943 uint s1 = stack_slots_not_pruned; 944 stack_slots_not_pruned = 0; // for next iteration 945 if (s1 > l) s1 = l; 946 uint s0 = l - s1; 947 p += s0; // skip the tops preinstalled by add_req_batch 948 for (j = s0; j < l; j++) 949 call->set_req(p++, in_map->in(k+j)); 950 } else { 951 p += l; // already set to top above by add_req_batch 952 } 953 954 // Add the Monitors 955 k = in_jvms->monoff(); 956 l = in_jvms->mon_size(); 957 out_jvms->set_monoff(p); 958 for (j = 0; j < l; j++) 959 call->set_req(p++, in_map->in(k+j)); 960 961 // Copy any scalar object fields. 962 k = in_jvms->scloff(); 963 l = in_jvms->scl_size(); 964 out_jvms->set_scloff(p); 965 for (j = 0; j < l; j++) 966 call->set_req(p++, in_map->in(k+j)); 967 968 // Finish the new jvms. 969 out_jvms->set_endoff(p); 970 971 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 972 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 973 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 974 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 975 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 976 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 977 978 // Update the two tail pointers in parallel. 979 out_jvms = out_jvms->caller(); 980 in_jvms = in_jvms->caller(); 981 } 982 983 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 984 985 // Test the correctness of JVMState::debug_xxx accessors: 986 assert(call->jvms()->debug_start() == non_debug_edges, ""); 987 assert(call->jvms()->debug_end() == call->req(), ""); 988 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 989 } 990 991 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 992 Bytecodes::Code code = java_bc(); 993 if (code == Bytecodes::_wide) { 994 code = method()->java_code_at_bci(bci() + 1); 995 } 996 997 BasicType rtype = T_ILLEGAL; 998 int rsize = 0; 999 1000 if (code != Bytecodes::_illegal) { 1001 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1002 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1003 if (rtype < T_CONFLICT) 1004 rsize = type2size[rtype]; 1005 } 1006 1007 switch (code) { 1008 case Bytecodes::_illegal: 1009 return false; 1010 1011 case Bytecodes::_ldc: 1012 case Bytecodes::_ldc_w: 1013 case Bytecodes::_ldc2_w: 1014 inputs = 0; 1015 break; 1016 1017 case Bytecodes::_dup: inputs = 1; break; 1018 case Bytecodes::_dup_x1: inputs = 2; break; 1019 case Bytecodes::_dup_x2: inputs = 3; break; 1020 case Bytecodes::_dup2: inputs = 2; break; 1021 case Bytecodes::_dup2_x1: inputs = 3; break; 1022 case Bytecodes::_dup2_x2: inputs = 4; break; 1023 case Bytecodes::_swap: inputs = 2; break; 1024 case Bytecodes::_arraylength: inputs = 1; break; 1025 1026 case Bytecodes::_getstatic: 1027 case Bytecodes::_putstatic: 1028 case Bytecodes::_getfield: 1029 case Bytecodes::_putfield: 1030 { 1031 bool ignored_will_link; 1032 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1033 int size = field->type()->size(); 1034 bool is_get = (depth >= 0), is_static = (depth & 1); 1035 inputs = (is_static ? 0 : 1); 1036 if (is_get) { 1037 depth = size - inputs; 1038 } else { 1039 inputs += size; // putxxx pops the value from the stack 1040 depth = - inputs; 1041 } 1042 } 1043 break; 1044 1045 case Bytecodes::_invokevirtual: 1046 case Bytecodes::_invokespecial: 1047 case Bytecodes::_invokestatic: 1048 case Bytecodes::_invokedynamic: 1049 case Bytecodes::_invokeinterface: 1050 { 1051 bool ignored_will_link; 1052 ciSignature* declared_signature = NULL; 1053 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1054 assert(declared_signature != NULL, "cannot be null"); 1055 inputs = declared_signature->arg_size_for_bc(code); 1056 int size = declared_signature->return_type()->size(); 1057 depth = size - inputs; 1058 } 1059 break; 1060 1061 case Bytecodes::_multianewarray: 1062 { 1063 ciBytecodeStream iter(method()); 1064 iter.reset_to_bci(bci()); 1065 iter.next(); 1066 inputs = iter.get_dimensions(); 1067 assert(rsize == 1, ""); 1068 depth = rsize - inputs; 1069 } 1070 break; 1071 1072 case Bytecodes::_ireturn: 1073 case Bytecodes::_lreturn: 1074 case Bytecodes::_freturn: 1075 case Bytecodes::_dreturn: 1076 case Bytecodes::_areturn: 1077 assert(rsize = -depth, ""); 1078 inputs = rsize; 1079 break; 1080 1081 case Bytecodes::_jsr: 1082 case Bytecodes::_jsr_w: 1083 inputs = 0; 1084 depth = 1; // S.B. depth=1, not zero 1085 break; 1086 1087 default: 1088 // bytecode produces a typed result 1089 inputs = rsize - depth; 1090 assert(inputs >= 0, ""); 1091 break; 1092 } 1093 1094 #ifdef ASSERT 1095 // spot check 1096 int outputs = depth + inputs; 1097 assert(outputs >= 0, "sanity"); 1098 switch (code) { 1099 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1100 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1101 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1102 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1103 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1104 } 1105 #endif //ASSERT 1106 1107 return true; 1108 } 1109 1110 1111 1112 //------------------------------basic_plus_adr--------------------------------- 1113 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1114 // short-circuit a common case 1115 if (offset == intcon(0)) return ptr; 1116 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1117 } 1118 1119 Node* GraphKit::ConvI2L(Node* offset) { 1120 // short-circuit a common case 1121 jint offset_con = find_int_con(offset, Type::OffsetBot); 1122 if (offset_con != Type::OffsetBot) { 1123 return longcon((jlong) offset_con); 1124 } 1125 return _gvn.transform( new (C) ConvI2LNode(offset)); 1126 } 1127 1128 Node* GraphKit::ConvI2UL(Node* offset) { 1129 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1130 if (offset_con != (juint) Type::OffsetBot) { 1131 return longcon((julong) offset_con); 1132 } 1133 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset)); 1134 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) ); 1135 return _gvn.transform( new (C) AndLNode(conv, mask) ); 1136 } 1137 1138 Node* GraphKit::ConvL2I(Node* offset) { 1139 // short-circuit a common case 1140 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1141 if (offset_con != (jlong)Type::OffsetBot) { 1142 return intcon((int) offset_con); 1143 } 1144 return _gvn.transform( new (C) ConvL2INode(offset)); 1145 } 1146 1147 //-------------------------load_object_klass----------------------------------- 1148 Node* GraphKit::load_object_klass(Node* obj) { 1149 // Special-case a fresh allocation to avoid building nodes: 1150 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1151 if (akls != NULL) return akls; 1152 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1153 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1154 } 1155 1156 //-------------------------load_array_length----------------------------------- 1157 Node* GraphKit::load_array_length(Node* array) { 1158 // Special-case a fresh allocation to avoid building nodes: 1159 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1160 Node *alen; 1161 if (alloc == NULL) { 1162 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1163 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1164 } else { 1165 alen = alloc->Ideal_length(); 1166 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1167 if (ccast != alen) { 1168 alen = _gvn.transform(ccast); 1169 } 1170 } 1171 return alen; 1172 } 1173 1174 //------------------------------do_null_check---------------------------------- 1175 // Helper function to do a NULL pointer check. Returned value is 1176 // the incoming address with NULL casted away. You are allowed to use the 1177 // not-null value only if you are control dependent on the test. 1178 extern int explicit_null_checks_inserted, 1179 explicit_null_checks_elided; 1180 Node* GraphKit::null_check_common(Node* value, BasicType type, 1181 // optional arguments for variations: 1182 bool assert_null, 1183 Node* *null_control) { 1184 assert(!assert_null || null_control == NULL, "not both at once"); 1185 if (stopped()) return top(); 1186 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1187 // For some performance testing, we may wish to suppress null checking. 1188 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1189 return value; 1190 } 1191 explicit_null_checks_inserted++; 1192 1193 // Construct NULL check 1194 Node *chk = NULL; 1195 switch(type) { 1196 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1197 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1198 case T_ARRAY : // fall through 1199 type = T_OBJECT; // simplify further tests 1200 case T_OBJECT : { 1201 const Type *t = _gvn.type( value ); 1202 1203 const TypeOopPtr* tp = t->isa_oopptr(); 1204 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1205 // Only for do_null_check, not any of its siblings: 1206 && !assert_null && null_control == NULL) { 1207 // Usually, any field access or invocation on an unloaded oop type 1208 // will simply fail to link, since the statically linked class is 1209 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1210 // the static class is loaded but the sharper oop type is not. 1211 // Rather than checking for this obscure case in lots of places, 1212 // we simply observe that a null check on an unloaded class 1213 // will always be followed by a nonsense operation, so we 1214 // can just issue the uncommon trap here. 1215 // Our access to the unloaded class will only be correct 1216 // after it has been loaded and initialized, which requires 1217 // a trip through the interpreter. 1218 #ifndef PRODUCT 1219 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1220 #endif 1221 uncommon_trap(Deoptimization::Reason_unloaded, 1222 Deoptimization::Action_reinterpret, 1223 tp->klass(), "!loaded"); 1224 return top(); 1225 } 1226 1227 if (assert_null) { 1228 // See if the type is contained in NULL_PTR. 1229 // If so, then the value is already null. 1230 if (t->higher_equal(TypePtr::NULL_PTR)) { 1231 explicit_null_checks_elided++; 1232 return value; // Elided null assert quickly! 1233 } 1234 } else { 1235 // See if mixing in the NULL pointer changes type. 1236 // If so, then the NULL pointer was not allowed in the original 1237 // type. In other words, "value" was not-null. 1238 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1239 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1240 explicit_null_checks_elided++; 1241 return value; // Elided null check quickly! 1242 } 1243 } 1244 chk = new (C) CmpPNode( value, null() ); 1245 break; 1246 } 1247 1248 default: 1249 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1250 } 1251 assert(chk != NULL, "sanity check"); 1252 chk = _gvn.transform(chk); 1253 1254 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1255 BoolNode *btst = new (C) BoolNode( chk, btest); 1256 Node *tst = _gvn.transform( btst ); 1257 1258 //----------- 1259 // if peephole optimizations occurred, a prior test existed. 1260 // If a prior test existed, maybe it dominates as we can avoid this test. 1261 if (tst != btst && type == T_OBJECT) { 1262 // At this point we want to scan up the CFG to see if we can 1263 // find an identical test (and so avoid this test altogether). 1264 Node *cfg = control(); 1265 int depth = 0; 1266 while( depth < 16 ) { // Limit search depth for speed 1267 if( cfg->Opcode() == Op_IfTrue && 1268 cfg->in(0)->in(1) == tst ) { 1269 // Found prior test. Use "cast_not_null" to construct an identical 1270 // CastPP (and hence hash to) as already exists for the prior test. 1271 // Return that casted value. 1272 if (assert_null) { 1273 replace_in_map(value, null()); 1274 return null(); // do not issue the redundant test 1275 } 1276 Node *oldcontrol = control(); 1277 set_control(cfg); 1278 Node *res = cast_not_null(value); 1279 set_control(oldcontrol); 1280 explicit_null_checks_elided++; 1281 return res; 1282 } 1283 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1284 if (cfg == NULL) break; // Quit at region nodes 1285 depth++; 1286 } 1287 } 1288 1289 //----------- 1290 // Branch to failure if null 1291 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1292 Deoptimization::DeoptReason reason; 1293 if (assert_null) 1294 reason = Deoptimization::Reason_null_assert; 1295 else if (type == T_OBJECT) 1296 reason = Deoptimization::Reason_null_check; 1297 else 1298 reason = Deoptimization::Reason_div0_check; 1299 1300 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1301 // ciMethodData::has_trap_at will return a conservative -1 if any 1302 // must-be-null assertion has failed. This could cause performance 1303 // problems for a method after its first do_null_assert failure. 1304 // Consider using 'Reason_class_check' instead? 1305 1306 // To cause an implicit null check, we set the not-null probability 1307 // to the maximum (PROB_MAX). For an explicit check the probability 1308 // is set to a smaller value. 1309 if (null_control != NULL || too_many_traps(reason)) { 1310 // probability is less likely 1311 ok_prob = PROB_LIKELY_MAG(3); 1312 } else if (!assert_null && 1313 (ImplicitNullCheckThreshold > 0) && 1314 method() != NULL && 1315 (method()->method_data()->trap_count(reason) 1316 >= (uint)ImplicitNullCheckThreshold)) { 1317 ok_prob = PROB_LIKELY_MAG(3); 1318 } 1319 1320 if (null_control != NULL) { 1321 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1322 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1323 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1324 if (null_true == top()) 1325 explicit_null_checks_elided++; 1326 (*null_control) = null_true; 1327 } else { 1328 BuildCutout unless(this, tst, ok_prob); 1329 // Check for optimizer eliding test at parse time 1330 if (stopped()) { 1331 // Failure not possible; do not bother making uncommon trap. 1332 explicit_null_checks_elided++; 1333 } else if (assert_null) { 1334 uncommon_trap(reason, 1335 Deoptimization::Action_make_not_entrant, 1336 NULL, "assert_null"); 1337 } else { 1338 replace_in_map(value, zerocon(type)); 1339 builtin_throw(reason); 1340 } 1341 } 1342 1343 // Must throw exception, fall-thru not possible? 1344 if (stopped()) { 1345 return top(); // No result 1346 } 1347 1348 if (assert_null) { 1349 // Cast obj to null on this path. 1350 replace_in_map(value, zerocon(type)); 1351 return zerocon(type); 1352 } 1353 1354 // Cast obj to not-null on this path, if there is no null_control. 1355 // (If there is a null_control, a non-null value may come back to haunt us.) 1356 if (type == T_OBJECT) { 1357 Node* cast = cast_not_null(value, false); 1358 if (null_control == NULL || (*null_control) == top()) 1359 replace_in_map(value, cast); 1360 value = cast; 1361 } 1362 1363 return value; 1364 } 1365 1366 1367 //------------------------------cast_not_null---------------------------------- 1368 // Cast obj to not-null on this path 1369 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1370 const Type *t = _gvn.type(obj); 1371 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1372 // Object is already not-null? 1373 if( t == t_not_null ) return obj; 1374 1375 Node *cast = new (C) CastPPNode(obj,t_not_null); 1376 cast->init_req(0, control()); 1377 cast = _gvn.transform( cast ); 1378 1379 // Scan for instances of 'obj' in the current JVM mapping. 1380 // These instances are known to be not-null after the test. 1381 if (do_replace_in_map) 1382 replace_in_map(obj, cast); 1383 1384 return cast; // Return casted value 1385 } 1386 1387 1388 //--------------------------replace_in_map------------------------------------- 1389 void GraphKit::replace_in_map(Node* old, Node* neww) { 1390 if (old == neww) { 1391 return; 1392 } 1393 1394 map()->replace_edge(old, neww); 1395 1396 // Note: This operation potentially replaces any edge 1397 // on the map. This includes locals, stack, and monitors 1398 // of the current (innermost) JVM state. 1399 1400 // don't let inconsistent types from profiling escape this 1401 // method 1402 1403 const Type* told = _gvn.type(old); 1404 const Type* tnew = _gvn.type(neww); 1405 1406 if (!tnew->higher_equal(told)) { 1407 return; 1408 } 1409 1410 map()->record_replaced_node(old, neww); 1411 } 1412 1413 1414 //============================================================================= 1415 //--------------------------------memory--------------------------------------- 1416 Node* GraphKit::memory(uint alias_idx) { 1417 MergeMemNode* mem = merged_memory(); 1418 Node* p = mem->memory_at(alias_idx); 1419 _gvn.set_type(p, Type::MEMORY); // must be mapped 1420 return p; 1421 } 1422 1423 //-----------------------------reset_memory------------------------------------ 1424 Node* GraphKit::reset_memory() { 1425 Node* mem = map()->memory(); 1426 // do not use this node for any more parsing! 1427 debug_only( map()->set_memory((Node*)NULL) ); 1428 return _gvn.transform( mem ); 1429 } 1430 1431 //------------------------------set_all_memory--------------------------------- 1432 void GraphKit::set_all_memory(Node* newmem) { 1433 Node* mergemem = MergeMemNode::make(C, newmem); 1434 gvn().set_type_bottom(mergemem); 1435 map()->set_memory(mergemem); 1436 } 1437 1438 //------------------------------set_all_memory_call---------------------------- 1439 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1440 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1441 set_all_memory(newmem); 1442 } 1443 1444 //============================================================================= 1445 // 1446 // parser factory methods for MemNodes 1447 // 1448 // These are layered on top of the factory methods in LoadNode and StoreNode, 1449 // and integrate with the parser's memory state and _gvn engine. 1450 // 1451 1452 // factory methods in "int adr_idx" 1453 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1454 int adr_idx, 1455 MemNode::MemOrd mo, LoadNode::ControlDependency control_dependency, bool require_atomic_access) { 1456 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1457 const TypePtr* adr_type = NULL; // debug-mode-only argument 1458 debug_only(adr_type = C->get_adr_type(adr_idx)); 1459 Node* mem = memory(adr_idx); 1460 Node* ld; 1461 if (require_atomic_access && bt == T_LONG) { 1462 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency); 1463 } else if (require_atomic_access && bt == T_DOUBLE) { 1464 ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency); 1465 } else { 1466 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency); 1467 } 1468 ld = _gvn.transform(ld); 1469 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1470 // Improve graph before escape analysis and boxing elimination. 1471 record_for_igvn(ld); 1472 } 1473 return ld; 1474 } 1475 1476 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1477 int adr_idx, 1478 MemNode::MemOrd mo, 1479 bool require_atomic_access) { 1480 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1481 const TypePtr* adr_type = NULL; 1482 debug_only(adr_type = C->get_adr_type(adr_idx)); 1483 Node *mem = memory(adr_idx); 1484 Node* st; 1485 if (require_atomic_access && bt == T_LONG) { 1486 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1487 } else if (require_atomic_access && bt == T_DOUBLE) { 1488 st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1489 } else { 1490 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1491 } 1492 st = _gvn.transform(st); 1493 set_memory(st, adr_idx); 1494 // Back-to-back stores can only remove intermediate store with DU info 1495 // so push on worklist for optimizer. 1496 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1497 record_for_igvn(st); 1498 1499 return st; 1500 } 1501 1502 1503 void GraphKit::pre_barrier(bool do_load, 1504 Node* ctl, 1505 Node* obj, 1506 Node* adr, 1507 uint adr_idx, 1508 Node* val, 1509 const TypeOopPtr* val_type, 1510 Node* pre_val, 1511 BasicType bt) { 1512 1513 BarrierSet* bs = Universe::heap()->barrier_set(); 1514 set_control(ctl); 1515 switch (bs->kind()) { 1516 case BarrierSet::G1SATBCT: 1517 case BarrierSet::G1SATBCTLogging: 1518 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1519 break; 1520 1521 case BarrierSet::CardTableModRef: 1522 case BarrierSet::CardTableExtension: 1523 case BarrierSet::ModRef: 1524 break; 1525 1526 case BarrierSet::Other: 1527 default : 1528 ShouldNotReachHere(); 1529 1530 } 1531 } 1532 1533 bool GraphKit::can_move_pre_barrier() const { 1534 BarrierSet* bs = Universe::heap()->barrier_set(); 1535 switch (bs->kind()) { 1536 case BarrierSet::G1SATBCT: 1537 case BarrierSet::G1SATBCTLogging: 1538 return true; // Can move it if no safepoint 1539 1540 case BarrierSet::CardTableModRef: 1541 case BarrierSet::CardTableExtension: 1542 case BarrierSet::ModRef: 1543 return true; // There is no pre-barrier 1544 1545 case BarrierSet::Other: 1546 default : 1547 ShouldNotReachHere(); 1548 } 1549 return false; 1550 } 1551 1552 void GraphKit::post_barrier(Node* ctl, 1553 Node* store, 1554 Node* obj, 1555 Node* adr, 1556 uint adr_idx, 1557 Node* val, 1558 BasicType bt, 1559 bool use_precise) { 1560 BarrierSet* bs = Universe::heap()->barrier_set(); 1561 set_control(ctl); 1562 switch (bs->kind()) { 1563 case BarrierSet::G1SATBCT: 1564 case BarrierSet::G1SATBCTLogging: 1565 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1566 break; 1567 1568 case BarrierSet::CardTableModRef: 1569 case BarrierSet::CardTableExtension: 1570 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1571 break; 1572 1573 case BarrierSet::ModRef: 1574 break; 1575 1576 case BarrierSet::Other: 1577 default : 1578 ShouldNotReachHere(); 1579 1580 } 1581 } 1582 1583 Node* GraphKit::store_oop(Node* ctl, 1584 Node* obj, 1585 Node* adr, 1586 const TypePtr* adr_type, 1587 Node* val, 1588 const TypeOopPtr* val_type, 1589 BasicType bt, 1590 bool use_precise, 1591 MemNode::MemOrd mo) { 1592 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1593 // could be delayed during Parse (for example, in adjust_map_after_if()). 1594 // Execute transformation here to avoid barrier generation in such case. 1595 if (_gvn.type(val) == TypePtr::NULL_PTR) 1596 val = _gvn.makecon(TypePtr::NULL_PTR); 1597 1598 set_control(ctl); 1599 if (stopped()) return top(); // Dead path ? 1600 1601 assert(bt == T_OBJECT, "sanity"); 1602 assert(val != NULL, "not dead path"); 1603 uint adr_idx = C->get_alias_index(adr_type); 1604 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1605 1606 pre_barrier(true /* do_load */, 1607 control(), obj, adr, adr_idx, val, val_type, 1608 NULL /* pre_val */, 1609 bt); 1610 1611 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1612 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1613 return store; 1614 } 1615 1616 // Could be an array or object we don't know at compile time (unsafe ref.) 1617 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1618 Node* obj, // containing obj 1619 Node* adr, // actual adress to store val at 1620 const TypePtr* adr_type, 1621 Node* val, 1622 BasicType bt, 1623 MemNode::MemOrd mo) { 1624 Compile::AliasType* at = C->alias_type(adr_type); 1625 const TypeOopPtr* val_type = NULL; 1626 if (adr_type->isa_instptr()) { 1627 if (at->field() != NULL) { 1628 // known field. This code is a copy of the do_put_xxx logic. 1629 ciField* field = at->field(); 1630 if (!field->type()->is_loaded()) { 1631 val_type = TypeInstPtr::BOTTOM; 1632 } else { 1633 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1634 } 1635 } 1636 } else if (adr_type->isa_aryptr()) { 1637 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1638 } 1639 if (val_type == NULL) { 1640 val_type = TypeInstPtr::BOTTOM; 1641 } 1642 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1643 } 1644 1645 1646 //-------------------------array_element_address------------------------- 1647 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1648 const TypeInt* sizetype, Node* ctrl) { 1649 uint shift = exact_log2(type2aelembytes(elembt)); 1650 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1651 1652 // short-circuit a common case (saves lots of confusing waste motion) 1653 jint idx_con = find_int_con(idx, -1); 1654 if (idx_con >= 0) { 1655 intptr_t offset = header + ((intptr_t)idx_con << shift); 1656 return basic_plus_adr(ary, offset); 1657 } 1658 1659 // must be correct type for alignment purposes 1660 Node* base = basic_plus_adr(ary, header); 1661 #ifdef _LP64 1662 // The scaled index operand to AddP must be a clean 64-bit value. 1663 // Java allows a 32-bit int to be incremented to a negative 1664 // value, which appears in a 64-bit register as a large 1665 // positive number. Using that large positive number as an 1666 // operand in pointer arithmetic has bad consequences. 1667 // On the other hand, 32-bit overflow is rare, and the possibility 1668 // can often be excluded, if we annotate the ConvI2L node with 1669 // a type assertion that its value is known to be a small positive 1670 // number. (The prior range check has ensured this.) 1671 // This assertion is used by ConvI2LNode::Ideal. 1672 int index_max = max_jint - 1; // array size is max_jint, index is one less 1673 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1674 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 1675 idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl); 1676 #endif 1677 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1678 return basic_plus_adr(ary, base, scale); 1679 } 1680 1681 //-------------------------load_array_element------------------------- 1682 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1683 const Type* elemtype = arytype->elem(); 1684 BasicType elembt = elemtype->array_element_basic_type(); 1685 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1686 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1687 return ld; 1688 } 1689 1690 //-------------------------set_arguments_for_java_call------------------------- 1691 // Arguments (pre-popped from the stack) are taken from the JVMS. 1692 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1693 // Add the call arguments: 1694 uint nargs = call->method()->arg_size(); 1695 for (uint i = 0; i < nargs; i++) { 1696 Node* arg = argument(i); 1697 call->init_req(i + TypeFunc::Parms, arg); 1698 } 1699 } 1700 1701 //---------------------------set_edges_for_java_call--------------------------- 1702 // Connect a newly created call into the current JVMS. 1703 // A return value node (if any) is returned from set_edges_for_java_call. 1704 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1705 1706 // Add the predefined inputs: 1707 call->init_req( TypeFunc::Control, control() ); 1708 call->init_req( TypeFunc::I_O , i_o() ); 1709 call->init_req( TypeFunc::Memory , reset_memory() ); 1710 call->init_req( TypeFunc::FramePtr, frameptr() ); 1711 call->init_req( TypeFunc::ReturnAdr, top() ); 1712 1713 add_safepoint_edges(call, must_throw); 1714 1715 Node* xcall = _gvn.transform(call); 1716 1717 if (xcall == top()) { 1718 set_control(top()); 1719 return; 1720 } 1721 assert(xcall == call, "call identity is stable"); 1722 1723 // Re-use the current map to produce the result. 1724 1725 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1726 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1727 set_all_memory_call(xcall, separate_io_proj); 1728 1729 //return xcall; // no need, caller already has it 1730 } 1731 1732 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1733 if (stopped()) return top(); // maybe the call folded up? 1734 1735 // Capture the return value, if any. 1736 Node* ret; 1737 if (call->method() == NULL || 1738 call->method()->return_type()->basic_type() == T_VOID) 1739 ret = top(); 1740 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1741 1742 // Note: Since any out-of-line call can produce an exception, 1743 // we always insert an I_O projection from the call into the result. 1744 1745 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1746 1747 if (separate_io_proj) { 1748 // The caller requested separate projections be used by the fall 1749 // through and exceptional paths, so replace the projections for 1750 // the fall through path. 1751 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1752 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1753 } 1754 return ret; 1755 } 1756 1757 //--------------------set_predefined_input_for_runtime_call-------------------- 1758 // Reading and setting the memory state is way conservative here. 1759 // The real problem is that I am not doing real Type analysis on memory, 1760 // so I cannot distinguish card mark stores from other stores. Across a GC 1761 // point the Store Barrier and the card mark memory has to agree. I cannot 1762 // have a card mark store and its barrier split across the GC point from 1763 // either above or below. Here I get that to happen by reading ALL of memory. 1764 // A better answer would be to separate out card marks from other memory. 1765 // For now, return the input memory state, so that it can be reused 1766 // after the call, if this call has restricted memory effects. 1767 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1768 // Set fixed predefined input arguments 1769 Node* memory = reset_memory(); 1770 call->init_req( TypeFunc::Control, control() ); 1771 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1772 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1773 call->init_req( TypeFunc::FramePtr, frameptr() ); 1774 call->init_req( TypeFunc::ReturnAdr, top() ); 1775 return memory; 1776 } 1777 1778 //-------------------set_predefined_output_for_runtime_call-------------------- 1779 // Set control and memory (not i_o) from the call. 1780 // If keep_mem is not NULL, use it for the output state, 1781 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1782 // If hook_mem is NULL, this call produces no memory effects at all. 1783 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1784 // then only that memory slice is taken from the call. 1785 // In the last case, we must put an appropriate memory barrier before 1786 // the call, so as to create the correct anti-dependencies on loads 1787 // preceding the call. 1788 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1789 Node* keep_mem, 1790 const TypePtr* hook_mem) { 1791 // no i/o 1792 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1793 if (keep_mem) { 1794 // First clone the existing memory state 1795 set_all_memory(keep_mem); 1796 if (hook_mem != NULL) { 1797 // Make memory for the call 1798 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1799 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1800 // We also use hook_mem to extract specific effects from arraycopy stubs. 1801 set_memory(mem, hook_mem); 1802 } 1803 // ...else the call has NO memory effects. 1804 1805 // Make sure the call advertises its memory effects precisely. 1806 // This lets us build accurate anti-dependences in gcm.cpp. 1807 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1808 "call node must be constructed correctly"); 1809 } else { 1810 assert(hook_mem == NULL, ""); 1811 // This is not a "slow path" call; all memory comes from the call. 1812 set_all_memory_call(call); 1813 } 1814 } 1815 1816 1817 // Replace the call with the current state of the kit. 1818 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1819 JVMState* ejvms = NULL; 1820 if (has_exceptions()) { 1821 ejvms = transfer_exceptions_into_jvms(); 1822 } 1823 1824 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1825 ReplacedNodes replaced_nodes_exception; 1826 Node* ex_ctl = top(); 1827 1828 SafePointNode* final_state = stop(); 1829 1830 // Find all the needed outputs of this call 1831 CallProjections callprojs; 1832 call->extract_projections(&callprojs, true); 1833 1834 Node* init_mem = call->in(TypeFunc::Memory); 1835 Node* final_mem = final_state->in(TypeFunc::Memory); 1836 Node* final_ctl = final_state->in(TypeFunc::Control); 1837 Node* final_io = final_state->in(TypeFunc::I_O); 1838 1839 // Replace all the old call edges with the edges from the inlining result 1840 if (callprojs.fallthrough_catchproj != NULL) { 1841 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1842 } 1843 if (callprojs.fallthrough_memproj != NULL) { 1844 if (final_mem->is_MergeMem()) { 1845 // Parser's exits MergeMem was not transformed but may be optimized 1846 final_mem = _gvn.transform(final_mem); 1847 } 1848 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1849 } 1850 if (callprojs.fallthrough_ioproj != NULL) { 1851 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1852 } 1853 1854 // Replace the result with the new result if it exists and is used 1855 if (callprojs.resproj != NULL && result != NULL) { 1856 C->gvn_replace_by(callprojs.resproj, result); 1857 } 1858 1859 if (ejvms == NULL) { 1860 // No exception edges to simply kill off those paths 1861 if (callprojs.catchall_catchproj != NULL) { 1862 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1863 } 1864 if (callprojs.catchall_memproj != NULL) { 1865 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1866 } 1867 if (callprojs.catchall_ioproj != NULL) { 1868 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1869 } 1870 // Replace the old exception object with top 1871 if (callprojs.exobj != NULL) { 1872 C->gvn_replace_by(callprojs.exobj, C->top()); 1873 } 1874 } else { 1875 GraphKit ekit(ejvms); 1876 1877 // Load my combined exception state into the kit, with all phis transformed: 1878 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1879 replaced_nodes_exception = ex_map->replaced_nodes(); 1880 1881 Node* ex_oop = ekit.use_exception_state(ex_map); 1882 1883 if (callprojs.catchall_catchproj != NULL) { 1884 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1885 ex_ctl = ekit.control(); 1886 } 1887 if (callprojs.catchall_memproj != NULL) { 1888 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1889 } 1890 if (callprojs.catchall_ioproj != NULL) { 1891 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1892 } 1893 1894 // Replace the old exception object with the newly created one 1895 if (callprojs.exobj != NULL) { 1896 C->gvn_replace_by(callprojs.exobj, ex_oop); 1897 } 1898 } 1899 1900 // Disconnect the call from the graph 1901 call->disconnect_inputs(NULL, C); 1902 C->gvn_replace_by(call, C->top()); 1903 1904 // Clean up any MergeMems that feed other MergeMems since the 1905 // optimizer doesn't like that. 1906 if (final_mem->is_MergeMem()) { 1907 Node_List wl; 1908 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1909 Node* m = i.get(); 1910 if (m->is_MergeMem() && !wl.contains(m)) { 1911 wl.push(m); 1912 } 1913 } 1914 while (wl.size() > 0) { 1915 _gvn.transform(wl.pop()); 1916 } 1917 } 1918 1919 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1920 replaced_nodes.apply(C, final_ctl); 1921 } 1922 if (!ex_ctl->is_top() && do_replaced_nodes) { 1923 replaced_nodes_exception.apply(C, ex_ctl); 1924 } 1925 } 1926 1927 1928 //------------------------------increment_counter------------------------------ 1929 // for statistics: increment a VM counter by 1 1930 1931 void GraphKit::increment_counter(address counter_addr) { 1932 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1933 increment_counter(adr1); 1934 } 1935 1936 void GraphKit::increment_counter(Node* counter_addr) { 1937 int adr_type = Compile::AliasIdxRaw; 1938 Node* ctrl = control(); 1939 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1940 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1941 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1942 } 1943 1944 1945 //------------------------------uncommon_trap---------------------------------- 1946 // Bail out to the interpreter in mid-method. Implemented by calling the 1947 // uncommon_trap blob. This helper function inserts a runtime call with the 1948 // right debug info. 1949 void GraphKit::uncommon_trap(int trap_request, 1950 ciKlass* klass, const char* comment, 1951 bool must_throw, 1952 bool keep_exact_action) { 1953 if (failing()) stop(); 1954 if (stopped()) return; // trap reachable? 1955 1956 // Note: If ProfileTraps is true, and if a deopt. actually 1957 // occurs here, the runtime will make sure an MDO exists. There is 1958 // no need to call method()->ensure_method_data() at this point. 1959 1960 // Set the stack pointer to the right value for reexecution: 1961 set_sp(reexecute_sp()); 1962 1963 #ifdef ASSERT 1964 if (!must_throw) { 1965 // Make sure the stack has at least enough depth to execute 1966 // the current bytecode. 1967 int inputs, ignored_depth; 1968 if (compute_stack_effects(inputs, ignored_depth)) { 1969 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1970 Bytecodes::name(java_bc()), sp(), inputs)); 1971 } 1972 } 1973 #endif 1974 1975 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1976 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1977 1978 switch (action) { 1979 case Deoptimization::Action_maybe_recompile: 1980 case Deoptimization::Action_reinterpret: 1981 // Temporary fix for 6529811 to allow virtual calls to be sure they 1982 // get the chance to go from mono->bi->mega 1983 if (!keep_exact_action && 1984 Deoptimization::trap_request_index(trap_request) < 0 && 1985 too_many_recompiles(reason)) { 1986 // This BCI is causing too many recompilations. 1987 if (C->log() != NULL) { 1988 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1989 Deoptimization::trap_reason_name(reason), 1990 Deoptimization::trap_action_name(action)); 1991 } 1992 action = Deoptimization::Action_none; 1993 trap_request = Deoptimization::make_trap_request(reason, action); 1994 } else { 1995 C->set_trap_can_recompile(true); 1996 } 1997 break; 1998 case Deoptimization::Action_make_not_entrant: 1999 C->set_trap_can_recompile(true); 2000 break; 2001 #ifdef ASSERT 2002 case Deoptimization::Action_none: 2003 case Deoptimization::Action_make_not_compilable: 2004 break; 2005 default: 2006 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2007 break; 2008 #endif 2009 } 2010 2011 if (TraceOptoParse) { 2012 char buf[100]; 2013 tty->print_cr("Uncommon trap %s at bci:%d", 2014 Deoptimization::format_trap_request(buf, sizeof(buf), 2015 trap_request), bci()); 2016 } 2017 2018 CompileLog* log = C->log(); 2019 if (log != NULL) { 2020 int kid = (klass == NULL)? -1: log->identify(klass); 2021 log->begin_elem("uncommon_trap bci='%d'", bci()); 2022 char buf[100]; 2023 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2024 trap_request)); 2025 if (kid >= 0) log->print(" klass='%d'", kid); 2026 if (comment != NULL) log->print(" comment='%s'", comment); 2027 log->end_elem(); 2028 } 2029 2030 // Make sure any guarding test views this path as very unlikely 2031 Node *i0 = control()->in(0); 2032 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2033 IfNode *iff = i0->as_If(); 2034 float f = iff->_prob; // Get prob 2035 if (control()->Opcode() == Op_IfTrue) { 2036 if (f > PROB_UNLIKELY_MAG(4)) 2037 iff->_prob = PROB_MIN; 2038 } else { 2039 if (f < PROB_LIKELY_MAG(4)) 2040 iff->_prob = PROB_MAX; 2041 } 2042 } 2043 2044 // Clear out dead values from the debug info. 2045 kill_dead_locals(); 2046 2047 // Now insert the uncommon trap subroutine call 2048 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2049 const TypePtr* no_memory_effects = NULL; 2050 // Pass the index of the class to be loaded 2051 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2052 (must_throw ? RC_MUST_THROW : 0), 2053 OptoRuntime::uncommon_trap_Type(), 2054 call_addr, "uncommon_trap", no_memory_effects, 2055 intcon(trap_request)); 2056 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2057 "must extract request correctly from the graph"); 2058 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2059 2060 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2061 // The debug info is the only real input to this call. 2062 2063 // Halt-and-catch fire here. The above call should never return! 2064 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2065 _gvn.set_type_bottom(halt); 2066 root()->add_req(halt); 2067 2068 stop_and_kill_map(); 2069 } 2070 2071 2072 //--------------------------just_allocated_object------------------------------ 2073 // Report the object that was just allocated. 2074 // It must be the case that there are no intervening safepoints. 2075 // We use this to determine if an object is so "fresh" that 2076 // it does not require card marks. 2077 Node* GraphKit::just_allocated_object(Node* current_control) { 2078 if (C->recent_alloc_ctl() == current_control) 2079 return C->recent_alloc_obj(); 2080 return NULL; 2081 } 2082 2083 2084 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2085 // (Note: TypeFunc::make has a cache that makes this fast.) 2086 const TypeFunc* tf = TypeFunc::make(dest_method); 2087 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2088 for (int j = 0; j < nargs; j++) { 2089 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2090 if( targ->basic_type() == T_DOUBLE ) { 2091 // If any parameters are doubles, they must be rounded before 2092 // the call, dstore_rounding does gvn.transform 2093 Node *arg = argument(j); 2094 arg = dstore_rounding(arg); 2095 set_argument(j, arg); 2096 } 2097 } 2098 } 2099 2100 /** 2101 * Record profiling data exact_kls for Node n with the type system so 2102 * that it can propagate it (speculation) 2103 * 2104 * @param n node that the type applies to 2105 * @param exact_kls type from profiling 2106 * 2107 * @return node with improved type 2108 */ 2109 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) { 2110 const Type* current_type = _gvn.type(n); 2111 assert(UseTypeSpeculation, "type speculation must be on"); 2112 2113 const TypeOopPtr* speculative = current_type->speculative(); 2114 2115 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2116 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2117 const TypeOopPtr* xtype = tklass->as_instance_type(); 2118 assert(xtype->klass_is_exact(), "Should be exact"); 2119 // record the new speculative type's depth 2120 speculative = xtype->with_inline_depth(jvms()->depth()); 2121 } 2122 2123 if (speculative != current_type->speculative()) { 2124 // Build a type with a speculative type (what we think we know 2125 // about the type but will need a guard when we use it) 2126 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2127 // We're changing the type, we need a new CheckCast node to carry 2128 // the new type. The new type depends on the control: what 2129 // profiling tells us is only valid from here as far as we can 2130 // tell. 2131 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2132 cast = _gvn.transform(cast); 2133 replace_in_map(n, cast); 2134 n = cast; 2135 } 2136 2137 return n; 2138 } 2139 2140 /** 2141 * Record profiling data from receiver profiling at an invoke with the 2142 * type system so that it can propagate it (speculation) 2143 * 2144 * @param n receiver node 2145 * 2146 * @return node with improved type 2147 */ 2148 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2149 if (!UseTypeSpeculation) { 2150 return n; 2151 } 2152 ciKlass* exact_kls = profile_has_unique_klass(); 2153 return record_profile_for_speculation(n, exact_kls); 2154 } 2155 2156 /** 2157 * Record profiling data from argument profiling at an invoke with the 2158 * type system so that it can propagate it (speculation) 2159 * 2160 * @param dest_method target method for the call 2161 * @param bc what invoke bytecode is this? 2162 */ 2163 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2164 if (!UseTypeSpeculation) { 2165 return; 2166 } 2167 const TypeFunc* tf = TypeFunc::make(dest_method); 2168 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2169 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2170 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2171 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2172 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2173 ciKlass* better_type = method()->argument_profiled_type(bci(), i); 2174 if (better_type != NULL) { 2175 record_profile_for_speculation(argument(j), better_type); 2176 } 2177 i++; 2178 } 2179 } 2180 } 2181 2182 /** 2183 * Record profiling data from parameter profiling at an invoke with 2184 * the type system so that it can propagate it (speculation) 2185 */ 2186 void GraphKit::record_profiled_parameters_for_speculation() { 2187 if (!UseTypeSpeculation) { 2188 return; 2189 } 2190 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2191 if (_gvn.type(local(i))->isa_oopptr()) { 2192 ciKlass* better_type = method()->parameter_profiled_type(j); 2193 if (better_type != NULL) { 2194 record_profile_for_speculation(local(i), better_type); 2195 } 2196 j++; 2197 } 2198 } 2199 } 2200 2201 void GraphKit::round_double_result(ciMethod* dest_method) { 2202 // A non-strict method may return a double value which has an extended 2203 // exponent, but this must not be visible in a caller which is 'strict' 2204 // If a strict caller invokes a non-strict callee, round a double result 2205 2206 BasicType result_type = dest_method->return_type()->basic_type(); 2207 assert( method() != NULL, "must have caller context"); 2208 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2209 // Destination method's return value is on top of stack 2210 // dstore_rounding() does gvn.transform 2211 Node *result = pop_pair(); 2212 result = dstore_rounding(result); 2213 push_pair(result); 2214 } 2215 } 2216 2217 // rounding for strict float precision conformance 2218 Node* GraphKit::precision_rounding(Node* n) { 2219 return UseStrictFP && _method->flags().is_strict() 2220 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2221 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2222 : n; 2223 } 2224 2225 // rounding for strict double precision conformance 2226 Node* GraphKit::dprecision_rounding(Node *n) { 2227 return UseStrictFP && _method->flags().is_strict() 2228 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2229 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2230 : n; 2231 } 2232 2233 // rounding for non-strict double stores 2234 Node* GraphKit::dstore_rounding(Node* n) { 2235 return Matcher::strict_fp_requires_explicit_rounding 2236 && UseSSE <= 1 2237 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2238 : n; 2239 } 2240 2241 //============================================================================= 2242 // Generate a fast path/slow path idiom. Graph looks like: 2243 // [foo] indicates that 'foo' is a parameter 2244 // 2245 // [in] NULL 2246 // \ / 2247 // CmpP 2248 // Bool ne 2249 // If 2250 // / \ 2251 // True False-<2> 2252 // / | 2253 // / cast_not_null 2254 // Load | | ^ 2255 // [fast_test] | | 2256 // gvn to opt_test | | 2257 // / \ | <1> 2258 // True False | 2259 // | \\ | 2260 // [slow_call] \[fast_result] 2261 // Ctl Val \ \ 2262 // | \ \ 2263 // Catch <1> \ \ 2264 // / \ ^ \ \ 2265 // Ex No_Ex | \ \ 2266 // | \ \ | \ <2> \ 2267 // ... \ [slow_res] | | \ [null_result] 2268 // \ \--+--+--- | | 2269 // \ | / \ | / 2270 // --------Region Phi 2271 // 2272 //============================================================================= 2273 // Code is structured as a series of driver functions all called 'do_XXX' that 2274 // call a set of helper functions. Helper functions first, then drivers. 2275 2276 //------------------------------null_check_oop--------------------------------- 2277 // Null check oop. Set null-path control into Region in slot 3. 2278 // Make a cast-not-nullness use the other not-null control. Return cast. 2279 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2280 bool never_see_null, bool safe_for_replace) { 2281 // Initial NULL check taken path 2282 (*null_control) = top(); 2283 Node* cast = null_check_common(value, T_OBJECT, false, null_control); 2284 2285 // Generate uncommon_trap: 2286 if (never_see_null && (*null_control) != top()) { 2287 // If we see an unexpected null at a check-cast we record it and force a 2288 // recompile; the offending check-cast will be compiled to handle NULLs. 2289 // If we see more than one offending BCI, then all checkcasts in the 2290 // method will be compiled to handle NULLs. 2291 PreserveJVMState pjvms(this); 2292 set_control(*null_control); 2293 replace_in_map(value, null()); 2294 uncommon_trap(Deoptimization::Reason_null_check, 2295 Deoptimization::Action_make_not_entrant); 2296 (*null_control) = top(); // NULL path is dead 2297 } 2298 if ((*null_control) == top() && safe_for_replace) { 2299 replace_in_map(value, cast); 2300 } 2301 2302 // Cast away null-ness on the result 2303 return cast; 2304 } 2305 2306 //------------------------------opt_iff---------------------------------------- 2307 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2308 // Return slow-path control. 2309 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2310 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2311 2312 // Fast path taken; set region slot 2 2313 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2314 region->init_req(2,fast_taken); // Capture fast-control 2315 2316 // Fast path not-taken, i.e. slow path 2317 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2318 return slow_taken; 2319 } 2320 2321 //-----------------------------make_runtime_call------------------------------- 2322 Node* GraphKit::make_runtime_call(int flags, 2323 const TypeFunc* call_type, address call_addr, 2324 const char* call_name, 2325 const TypePtr* adr_type, 2326 // The following parms are all optional. 2327 // The first NULL ends the list. 2328 Node* parm0, Node* parm1, 2329 Node* parm2, Node* parm3, 2330 Node* parm4, Node* parm5, 2331 Node* parm6, Node* parm7) { 2332 // Slow-path call 2333 bool is_leaf = !(flags & RC_NO_LEAF); 2334 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2335 if (call_name == NULL) { 2336 assert(!is_leaf, "must supply name for leaf"); 2337 call_name = OptoRuntime::stub_name(call_addr); 2338 } 2339 CallNode* call; 2340 if (!is_leaf) { 2341 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2342 bci(), adr_type); 2343 } else if (flags & RC_NO_FP) { 2344 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2345 } else { 2346 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2347 } 2348 2349 // The following is similar to set_edges_for_java_call, 2350 // except that the memory effects of the call are restricted to AliasIdxRaw. 2351 2352 // Slow path call has no side-effects, uses few values 2353 bool wide_in = !(flags & RC_NARROW_MEM); 2354 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2355 2356 Node* prev_mem = NULL; 2357 if (wide_in) { 2358 prev_mem = set_predefined_input_for_runtime_call(call); 2359 } else { 2360 assert(!wide_out, "narrow in => narrow out"); 2361 Node* narrow_mem = memory(adr_type); 2362 prev_mem = reset_memory(); 2363 map()->set_memory(narrow_mem); 2364 set_predefined_input_for_runtime_call(call); 2365 } 2366 2367 // Hook each parm in order. Stop looking at the first NULL. 2368 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2369 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2370 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2371 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2372 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2373 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2374 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2375 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2376 /* close each nested if ===> */ } } } } } } } } 2377 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2378 2379 if (!is_leaf) { 2380 // Non-leaves can block and take safepoints: 2381 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2382 } 2383 // Non-leaves can throw exceptions: 2384 if (has_io) { 2385 call->set_req(TypeFunc::I_O, i_o()); 2386 } 2387 2388 if (flags & RC_UNCOMMON) { 2389 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2390 // (An "if" probability corresponds roughly to an unconditional count. 2391 // Sort of.) 2392 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2393 } 2394 2395 Node* c = _gvn.transform(call); 2396 assert(c == call, "cannot disappear"); 2397 2398 if (wide_out) { 2399 // Slow path call has full side-effects. 2400 set_predefined_output_for_runtime_call(call); 2401 } else { 2402 // Slow path call has few side-effects, and/or sets few values. 2403 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2404 } 2405 2406 if (has_io) { 2407 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2408 } 2409 return call; 2410 2411 } 2412 2413 //------------------------------merge_memory----------------------------------- 2414 // Merge memory from one path into the current memory state. 2415 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2416 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2417 Node* old_slice = mms.force_memory(); 2418 Node* new_slice = mms.memory2(); 2419 if (old_slice != new_slice) { 2420 PhiNode* phi; 2421 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2422 if (mms.is_empty()) { 2423 // clone base memory Phi's inputs for this memory slice 2424 assert(old_slice == mms.base_memory(), "sanity"); 2425 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2426 _gvn.set_type(phi, Type::MEMORY); 2427 for (uint i = 1; i < phi->req(); i++) { 2428 phi->init_req(i, old_slice->in(i)); 2429 } 2430 } else { 2431 phi = old_slice->as_Phi(); // Phi was generated already 2432 } 2433 } else { 2434 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2435 _gvn.set_type(phi, Type::MEMORY); 2436 } 2437 phi->set_req(new_path, new_slice); 2438 mms.set_memory(phi); 2439 } 2440 } 2441 } 2442 2443 //------------------------------make_slow_call_ex------------------------------ 2444 // Make the exception handler hookups for the slow call 2445 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2446 if (stopped()) return; 2447 2448 // Make a catch node with just two handlers: fall-through and catch-all 2449 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2450 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2451 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2452 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2453 2454 { PreserveJVMState pjvms(this); 2455 set_control(excp); 2456 set_i_o(i_o); 2457 2458 if (excp != top()) { 2459 if (deoptimize) { 2460 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2461 uncommon_trap(Deoptimization::Reason_unhandled, 2462 Deoptimization::Action_none); 2463 } else { 2464 // Create an exception state also. 2465 // Use an exact type if the caller has specified a specific exception. 2466 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2467 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2468 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2469 } 2470 } 2471 } 2472 2473 // Get the no-exception control from the CatchNode. 2474 set_control(norm); 2475 } 2476 2477 2478 //-------------------------------gen_subtype_check----------------------------- 2479 // Generate a subtyping check. Takes as input the subtype and supertype. 2480 // Returns 2 values: sets the default control() to the true path and returns 2481 // the false path. Only reads invariant memory; sets no (visible) memory. 2482 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2483 // but that's not exposed to the optimizer. This call also doesn't take in an 2484 // Object; if you wish to check an Object you need to load the Object's class 2485 // prior to coming here. 2486 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2487 // Fast check for identical types, perhaps identical constants. 2488 // The types can even be identical non-constants, in cases 2489 // involving Array.newInstance, Object.clone, etc. 2490 if (subklass == superklass) 2491 return top(); // false path is dead; no test needed. 2492 2493 if (_gvn.type(superklass)->singleton()) { 2494 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2495 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2496 2497 // In the common case of an exact superklass, try to fold up the 2498 // test before generating code. You may ask, why not just generate 2499 // the code and then let it fold up? The answer is that the generated 2500 // code will necessarily include null checks, which do not always 2501 // completely fold away. If they are also needless, then they turn 2502 // into a performance loss. Example: 2503 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2504 // Here, the type of 'fa' is often exact, so the store check 2505 // of fa[1]=x will fold up, without testing the nullness of x. 2506 switch (static_subtype_check(superk, subk)) { 2507 case SSC_always_false: 2508 { 2509 Node* always_fail = control(); 2510 set_control(top()); 2511 return always_fail; 2512 } 2513 case SSC_always_true: 2514 return top(); 2515 case SSC_easy_test: 2516 { 2517 // Just do a direct pointer compare and be done. 2518 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2519 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2520 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2521 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2522 return _gvn.transform( new(C) IfFalseNode(iff) ); 2523 } 2524 case SSC_full_test: 2525 break; 2526 default: 2527 ShouldNotReachHere(); 2528 } 2529 } 2530 2531 // %%% Possible further optimization: Even if the superklass is not exact, 2532 // if the subklass is the unique subtype of the superklass, the check 2533 // will always succeed. We could leave a dependency behind to ensure this. 2534 2535 // First load the super-klass's check-offset 2536 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2537 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), 2538 TypeInt::INT, MemNode::unordered)); 2539 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2540 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2541 2542 // Load from the sub-klass's super-class display list, or a 1-word cache of 2543 // the secondary superclass list, or a failing value with a sentinel offset 2544 // if the super-klass is an interface or exceptionally deep in the Java 2545 // hierarchy and we have to scan the secondary superclass list the hard way. 2546 // Worst-case type is a little odd: NULL is allowed as a result (usually 2547 // klass loads can never produce a NULL). 2548 Node *chk_off_X = ConvI2X(chk_off); 2549 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2550 // For some types like interfaces the following loadKlass is from a 1-word 2551 // cache which is mutable so can't use immutable memory. Other 2552 // types load from the super-class display table which is immutable. 2553 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2554 Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2555 2556 // Compile speed common case: ARE a subtype and we canNOT fail 2557 if( superklass == nkls ) 2558 return top(); // false path is dead; no test needed. 2559 2560 // See if we get an immediate positive hit. Happens roughly 83% of the 2561 // time. Test to see if the value loaded just previously from the subklass 2562 // is exactly the superklass. 2563 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2564 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2565 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2566 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2567 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2568 2569 // Compile speed common case: Check for being deterministic right now. If 2570 // chk_off is a constant and not equal to cacheoff then we are NOT a 2571 // subklass. In this case we need exactly the 1 test above and we can 2572 // return those results immediately. 2573 if (!might_be_cache) { 2574 Node* not_subtype_ctrl = control(); 2575 set_control(iftrue1); // We need exactly the 1 test above 2576 return not_subtype_ctrl; 2577 } 2578 2579 // Gather the various success & failures here 2580 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2581 record_for_igvn(r_ok_subtype); 2582 RegionNode *r_not_subtype = new (C) RegionNode(3); 2583 record_for_igvn(r_not_subtype); 2584 2585 r_ok_subtype->init_req(1, iftrue1); 2586 2587 // Check for immediate negative hit. Happens roughly 11% of the time (which 2588 // is roughly 63% of the remaining cases). Test to see if the loaded 2589 // check-offset points into the subklass display list or the 1-element 2590 // cache. If it points to the display (and NOT the cache) and the display 2591 // missed then it's not a subtype. 2592 Node *cacheoff = _gvn.intcon(cacheoff_con); 2593 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2594 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2595 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2596 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2597 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2598 2599 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2600 // No performance impact (too rare) but allows sharing of secondary arrays 2601 // which has some footprint reduction. 2602 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2603 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2604 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2605 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2606 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2607 2608 // -- Roads not taken here: -- 2609 // We could also have chosen to perform the self-check at the beginning 2610 // of this code sequence, as the assembler does. This would not pay off 2611 // the same way, since the optimizer, unlike the assembler, can perform 2612 // static type analysis to fold away many successful self-checks. 2613 // Non-foldable self checks work better here in second position, because 2614 // the initial primary superclass check subsumes a self-check for most 2615 // types. An exception would be a secondary type like array-of-interface, 2616 // which does not appear in its own primary supertype display. 2617 // Finally, we could have chosen to move the self-check into the 2618 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2619 // dependent manner. But it is worthwhile to have the check here, 2620 // where it can be perhaps be optimized. The cost in code space is 2621 // small (register compare, branch). 2622 2623 // Now do a linear scan of the secondary super-klass array. Again, no real 2624 // performance impact (too rare) but it's gotta be done. 2625 // Since the code is rarely used, there is no penalty for moving it 2626 // out of line, and it can only improve I-cache density. 2627 // The decision to inline or out-of-line this final check is platform 2628 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2629 Node* psc = _gvn.transform( 2630 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2631 2632 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2633 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2634 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2635 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2636 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2637 2638 // Return false path; set default control to true path. 2639 set_control( _gvn.transform(r_ok_subtype) ); 2640 return _gvn.transform(r_not_subtype); 2641 } 2642 2643 //----------------------------static_subtype_check----------------------------- 2644 // Shortcut important common cases when superklass is exact: 2645 // (0) superklass is java.lang.Object (can occur in reflective code) 2646 // (1) subklass is already limited to a subtype of superklass => always ok 2647 // (2) subklass does not overlap with superklass => always fail 2648 // (3) superklass has NO subtypes and we can check with a simple compare. 2649 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2650 if (StressReflectiveCode) { 2651 return SSC_full_test; // Let caller generate the general case. 2652 } 2653 2654 if (superk == env()->Object_klass()) { 2655 return SSC_always_true; // (0) this test cannot fail 2656 } 2657 2658 ciType* superelem = superk; 2659 if (superelem->is_array_klass()) 2660 superelem = superelem->as_array_klass()->base_element_type(); 2661 2662 if (!subk->is_interface()) { // cannot trust static interface types yet 2663 if (subk->is_subtype_of(superk)) { 2664 return SSC_always_true; // (1) false path dead; no dynamic test needed 2665 } 2666 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2667 !superk->is_subtype_of(subk)) { 2668 return SSC_always_false; 2669 } 2670 } 2671 2672 // If casting to an instance klass, it must have no subtypes 2673 if (superk->is_interface()) { 2674 // Cannot trust interfaces yet. 2675 // %%% S.B. superk->nof_implementors() == 1 2676 } else if (superelem->is_instance_klass()) { 2677 ciInstanceKlass* ik = superelem->as_instance_klass(); 2678 if (!ik->has_subklass() && !ik->is_interface()) { 2679 if (!ik->is_final()) { 2680 // Add a dependency if there is a chance of a later subclass. 2681 C->dependencies()->assert_leaf_type(ik); 2682 } 2683 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2684 } 2685 } else { 2686 // A primitive array type has no subtypes. 2687 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2688 } 2689 2690 return SSC_full_test; 2691 } 2692 2693 // Profile-driven exact type check: 2694 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2695 float prob, 2696 Node* *casted_receiver) { 2697 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2698 Node* recv_klass = load_object_klass(receiver); 2699 Node* want_klass = makecon(tklass); 2700 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2701 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2702 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2703 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2704 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2705 2706 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2707 assert(recv_xtype->klass_is_exact(), ""); 2708 2709 // Subsume downstream occurrences of receiver with a cast to 2710 // recv_xtype, since now we know what the type will be. 2711 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2712 (*casted_receiver) = _gvn.transform(cast); 2713 // (User must make the replace_in_map call.) 2714 2715 return fail; 2716 } 2717 2718 2719 //------------------------------seems_never_null------------------------------- 2720 // Use null_seen information if it is available from the profile. 2721 // If we see an unexpected null at a type check we record it and force a 2722 // recompile; the offending check will be recompiled to handle NULLs. 2723 // If we see several offending BCIs, then all checks in the 2724 // method will be recompiled. 2725 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) { 2726 if (UncommonNullCast // Cutout for this technique 2727 && obj != null() // And not the -Xcomp stupid case? 2728 && !too_many_traps(Deoptimization::Reason_null_check) 2729 ) { 2730 if (data == NULL) 2731 // Edge case: no mature data. Be optimistic here. 2732 return true; 2733 // If the profile has not seen a null, assume it won't happen. 2734 assert(java_bc() == Bytecodes::_checkcast || 2735 java_bc() == Bytecodes::_instanceof || 2736 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2737 return !data->as_BitData()->null_seen(); 2738 } 2739 return false; 2740 } 2741 2742 //------------------------maybe_cast_profiled_receiver------------------------- 2743 // If the profile has seen exactly one type, narrow to exactly that type. 2744 // Subsequent type checks will always fold up. 2745 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2746 ciKlass* require_klass, 2747 ciKlass* spec_klass, 2748 bool safe_for_replace) { 2749 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2750 2751 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check; 2752 2753 // Make sure we haven't already deoptimized from this tactic. 2754 if (too_many_traps(reason) || too_many_recompiles(reason)) 2755 return NULL; 2756 2757 // (No, this isn't a call, but it's enough like a virtual call 2758 // to use the same ciMethod accessor to get the profile info...) 2759 // If we have a speculative type use it instead of profiling (which 2760 // may not help us) 2761 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2762 if (exact_kls != NULL) {// no cast failures here 2763 if (require_klass == NULL || 2764 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2765 // If we narrow the type to match what the type profile sees or 2766 // the speculative type, we can then remove the rest of the 2767 // cast. 2768 // This is a win, even if the exact_kls is very specific, 2769 // because downstream operations, such as method calls, 2770 // will often benefit from the sharper type. 2771 Node* exact_obj = not_null_obj; // will get updated in place... 2772 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2773 &exact_obj); 2774 { PreserveJVMState pjvms(this); 2775 set_control(slow_ctl); 2776 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2777 } 2778 if (safe_for_replace) { 2779 replace_in_map(not_null_obj, exact_obj); 2780 } 2781 return exact_obj; 2782 } 2783 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2784 } 2785 2786 return NULL; 2787 } 2788 2789 /** 2790 * Cast obj to type and emit guard unless we had too many traps here 2791 * already 2792 * 2793 * @param obj node being casted 2794 * @param type type to cast the node to 2795 * @param not_null true if we know node cannot be null 2796 */ 2797 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2798 ciKlass* type, 2799 bool not_null) { 2800 // type == NULL if profiling tells us this object is always null 2801 if (type != NULL) { 2802 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2803 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check; 2804 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2805 !too_many_traps(class_reason) && !too_many_recompiles(class_reason)) { 2806 Node* not_null_obj = NULL; 2807 // not_null is true if we know the object is not null and 2808 // there's no need for a null check 2809 if (!not_null) { 2810 Node* null_ctl = top(); 2811 not_null_obj = null_check_oop(obj, &null_ctl, true, true); 2812 assert(null_ctl->is_top(), "no null control here"); 2813 } else { 2814 not_null_obj = obj; 2815 } 2816 2817 Node* exact_obj = not_null_obj; 2818 ciKlass* exact_kls = type; 2819 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2820 &exact_obj); 2821 { 2822 PreserveJVMState pjvms(this); 2823 set_control(slow_ctl); 2824 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2825 } 2826 replace_in_map(not_null_obj, exact_obj); 2827 obj = exact_obj; 2828 } 2829 } else { 2830 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2831 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2832 Node* exact_obj = null_assert(obj); 2833 replace_in_map(obj, exact_obj); 2834 obj = exact_obj; 2835 } 2836 } 2837 return obj; 2838 } 2839 2840 //-------------------------------gen_instanceof-------------------------------- 2841 // Generate an instance-of idiom. Used by both the instance-of bytecode 2842 // and the reflective instance-of call. 2843 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2844 kill_dead_locals(); // Benefit all the uncommon traps 2845 assert( !stopped(), "dead parse path should be checked in callers" ); 2846 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2847 "must check for not-null not-dead klass in callers"); 2848 2849 // Make the merge point 2850 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2851 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2852 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2853 C->set_has_split_ifs(true); // Has chance for split-if optimization 2854 2855 ciProfileData* data = NULL; 2856 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2857 data = method()->method_data()->bci_to_data(bci()); 2858 } 2859 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2860 && seems_never_null(obj, data)); 2861 2862 // Null check; get casted pointer; set region slot 3 2863 Node* null_ctl = top(); 2864 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2865 2866 // If not_null_obj is dead, only null-path is taken 2867 if (stopped()) { // Doing instance-of on a NULL? 2868 set_control(null_ctl); 2869 return intcon(0); 2870 } 2871 region->init_req(_null_path, null_ctl); 2872 phi ->init_req(_null_path, intcon(0)); // Set null path value 2873 if (null_ctl == top()) { 2874 // Do this eagerly, so that pattern matches like is_diamond_phi 2875 // will work even during parsing. 2876 assert(_null_path == PATH_LIMIT-1, "delete last"); 2877 region->del_req(_null_path); 2878 phi ->del_req(_null_path); 2879 } 2880 2881 // Do we know the type check always succeed? 2882 bool known_statically = false; 2883 if (_gvn.type(superklass)->singleton()) { 2884 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2885 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2886 if (subk != NULL && subk->is_loaded()) { 2887 int static_res = static_subtype_check(superk, subk); 2888 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2889 } 2890 } 2891 2892 if (known_statically && UseTypeSpeculation) { 2893 // If we know the type check always succeeds then we don't use the 2894 // profiling data at this bytecode. Don't lose it, feed it to the 2895 // type system as a speculative type. 2896 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2897 } else { 2898 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2899 // We may not have profiling here or it may not help us. If we 2900 // have a speculative type use it to perform an exact cast. 2901 ciKlass* spec_obj_type = obj_type->speculative_type(); 2902 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2903 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2904 if (stopped()) { // Profile disagrees with this path. 2905 set_control(null_ctl); // Null is the only remaining possibility. 2906 return intcon(0); 2907 } 2908 if (cast_obj != NULL) { 2909 not_null_obj = cast_obj; 2910 } 2911 } 2912 } 2913 2914 // Load the object's klass 2915 Node* obj_klass = load_object_klass(not_null_obj); 2916 2917 // Generate the subtype check 2918 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2919 2920 // Plug in the success path to the general merge in slot 1. 2921 region->init_req(_obj_path, control()); 2922 phi ->init_req(_obj_path, intcon(1)); 2923 2924 // Plug in the failing path to the general merge in slot 2. 2925 region->init_req(_fail_path, not_subtype_ctrl); 2926 phi ->init_req(_fail_path, intcon(0)); 2927 2928 // Return final merged results 2929 set_control( _gvn.transform(region) ); 2930 record_for_igvn(region); 2931 return _gvn.transform(phi); 2932 } 2933 2934 //-------------------------------gen_checkcast--------------------------------- 2935 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2936 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2937 // uncommon-trap paths work. Adjust stack after this call. 2938 // If failure_control is supplied and not null, it is filled in with 2939 // the control edge for the cast failure. Otherwise, an appropriate 2940 // uncommon trap or exception is thrown. 2941 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2942 Node* *failure_control) { 2943 kill_dead_locals(); // Benefit all the uncommon traps 2944 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2945 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2946 2947 // Fast cutout: Check the case that the cast is vacuously true. 2948 // This detects the common cases where the test will short-circuit 2949 // away completely. We do this before we perform the null check, 2950 // because if the test is going to turn into zero code, we don't 2951 // want a residual null check left around. (Causes a slowdown, 2952 // for example, in some objArray manipulations, such as a[i]=a[j].) 2953 if (tk->singleton()) { 2954 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2955 if (objtp != NULL && objtp->klass() != NULL) { 2956 switch (static_subtype_check(tk->klass(), objtp->klass())) { 2957 case SSC_always_true: 2958 // If we know the type check always succeed then we don't use 2959 // the profiling data at this bytecode. Don't lose it, feed it 2960 // to the type system as a speculative type. 2961 return record_profiled_receiver_for_speculation(obj); 2962 case SSC_always_false: 2963 // It needs a null check because a null will *pass* the cast check. 2964 // A non-null value will always produce an exception. 2965 return null_assert(obj); 2966 } 2967 } 2968 } 2969 2970 ciProfileData* data = NULL; 2971 bool safe_for_replace = false; 2972 if (failure_control == NULL) { // use MDO in regular case only 2973 assert(java_bc() == Bytecodes::_aastore || 2974 java_bc() == Bytecodes::_checkcast, 2975 "interpreter profiles type checks only for these BCs"); 2976 data = method()->method_data()->bci_to_data(bci()); 2977 safe_for_replace = true; 2978 } 2979 2980 // Make the merge point 2981 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2982 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 2983 Node* phi = new (C) PhiNode(region, toop); 2984 C->set_has_split_ifs(true); // Has chance for split-if optimization 2985 2986 // Use null-cast information if it is available 2987 bool never_see_null = ((failure_control == NULL) // regular case only 2988 && seems_never_null(obj, data)); 2989 2990 // Null check; get casted pointer; set region slot 3 2991 Node* null_ctl = top(); 2992 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace); 2993 2994 // If not_null_obj is dead, only null-path is taken 2995 if (stopped()) { // Doing instance-of on a NULL? 2996 set_control(null_ctl); 2997 return null(); 2998 } 2999 region->init_req(_null_path, null_ctl); 3000 phi ->init_req(_null_path, null()); // Set null path value 3001 if (null_ctl == top()) { 3002 // Do this eagerly, so that pattern matches like is_diamond_phi 3003 // will work even during parsing. 3004 assert(_null_path == PATH_LIMIT-1, "delete last"); 3005 region->del_req(_null_path); 3006 phi ->del_req(_null_path); 3007 } 3008 3009 Node* cast_obj = NULL; 3010 if (tk->klass_is_exact()) { 3011 // The following optimization tries to statically cast the speculative type of the object 3012 // (for example obtained during profiling) to the type of the superklass and then do a 3013 // dynamic check that the type of the object is what we expect. To work correctly 3014 // for checkcast and aastore the type of superklass should be exact. 3015 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3016 // We may not have profiling here or it may not help us. If we have 3017 // a speculative type use it to perform an exact cast. 3018 ciKlass* spec_obj_type = obj_type->speculative_type(); 3019 if (spec_obj_type != NULL || 3020 (data != NULL && 3021 // Counter has never been decremented (due to cast failure). 3022 // ...This is a reasonable thing to expect. It is true of 3023 // all casts inserted by javac to implement generic types. 3024 data->as_CounterData()->count() >= 0)) { 3025 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3026 if (cast_obj != NULL) { 3027 if (failure_control != NULL) // failure is now impossible 3028 (*failure_control) = top(); 3029 // adjust the type of the phi to the exact klass: 3030 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3031 } 3032 } 3033 } 3034 3035 if (cast_obj == NULL) { 3036 // Load the object's klass 3037 Node* obj_klass = load_object_klass(not_null_obj); 3038 3039 // Generate the subtype check 3040 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3041 3042 // Plug in success path into the merge 3043 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 3044 not_null_obj, toop)); 3045 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3046 if (failure_control == NULL) { 3047 if (not_subtype_ctrl != top()) { // If failure is possible 3048 PreserveJVMState pjvms(this); 3049 set_control(not_subtype_ctrl); 3050 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3051 } 3052 } else { 3053 (*failure_control) = not_subtype_ctrl; 3054 } 3055 } 3056 3057 region->init_req(_obj_path, control()); 3058 phi ->init_req(_obj_path, cast_obj); 3059 3060 // A merge of NULL or Casted-NotNull obj 3061 Node* res = _gvn.transform(phi); 3062 3063 // Note I do NOT always 'replace_in_map(obj,result)' here. 3064 // if( tk->klass()->can_be_primary_super() ) 3065 // This means that if I successfully store an Object into an array-of-String 3066 // I 'forget' that the Object is really now known to be a String. I have to 3067 // do this because we don't have true union types for interfaces - if I store 3068 // a Baz into an array-of-Interface and then tell the optimizer it's an 3069 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3070 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3071 // replace_in_map( obj, res ); 3072 3073 // Return final merged results 3074 set_control( _gvn.transform(region) ); 3075 record_for_igvn(region); 3076 return res; 3077 } 3078 3079 //------------------------------next_monitor----------------------------------- 3080 // What number should be given to the next monitor? 3081 int GraphKit::next_monitor() { 3082 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3083 int next = current + C->sync_stack_slots(); 3084 // Keep the toplevel high water mark current: 3085 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3086 return current; 3087 } 3088 3089 //------------------------------insert_mem_bar--------------------------------- 3090 // Memory barrier to avoid floating things around 3091 // The membar serves as a pinch point between both control and all memory slices. 3092 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3093 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3094 mb->init_req(TypeFunc::Control, control()); 3095 mb->init_req(TypeFunc::Memory, reset_memory()); 3096 Node* membar = _gvn.transform(mb); 3097 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3098 set_all_memory_call(membar); 3099 return membar; 3100 } 3101 3102 //-------------------------insert_mem_bar_volatile---------------------------- 3103 // Memory barrier to avoid floating things around 3104 // The membar serves as a pinch point between both control and memory(alias_idx). 3105 // If you want to make a pinch point on all memory slices, do not use this 3106 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3107 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3108 // When Parse::do_put_xxx updates a volatile field, it appends a series 3109 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3110 // The first membar is on the same memory slice as the field store opcode. 3111 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3112 // All the other membars (for other volatile slices, including AliasIdxBot, 3113 // which stands for all unknown volatile slices) are control-dependent 3114 // on the first membar. This prevents later volatile loads or stores 3115 // from sliding up past the just-emitted store. 3116 3117 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3118 mb->set_req(TypeFunc::Control,control()); 3119 if (alias_idx == Compile::AliasIdxBot) { 3120 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3121 } else { 3122 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3123 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3124 } 3125 Node* membar = _gvn.transform(mb); 3126 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3127 if (alias_idx == Compile::AliasIdxBot) { 3128 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 3129 } else { 3130 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 3131 } 3132 return membar; 3133 } 3134 3135 //------------------------------shared_lock------------------------------------ 3136 // Emit locking code. 3137 FastLockNode* GraphKit::shared_lock(Node* obj) { 3138 // bci is either a monitorenter bc or InvocationEntryBci 3139 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3140 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3141 3142 if( !GenerateSynchronizationCode ) 3143 return NULL; // Not locking things? 3144 if (stopped()) // Dead monitor? 3145 return NULL; 3146 3147 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3148 3149 // Box the stack location 3150 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 3151 Node* mem = reset_memory(); 3152 3153 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 3154 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3155 // Create the counters for this fast lock. 3156 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3157 } 3158 3159 // Create the rtm counters for this fast lock if needed. 3160 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3161 3162 // Add monitor to debug info for the slow path. If we block inside the 3163 // slow path and de-opt, we need the monitor hanging around 3164 map()->push_monitor( flock ); 3165 3166 const TypeFunc *tf = LockNode::lock_type(); 3167 LockNode *lock = new (C) LockNode(C, tf); 3168 3169 lock->init_req( TypeFunc::Control, control() ); 3170 lock->init_req( TypeFunc::Memory , mem ); 3171 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3172 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3173 lock->init_req( TypeFunc::ReturnAdr, top() ); 3174 3175 lock->init_req(TypeFunc::Parms + 0, obj); 3176 lock->init_req(TypeFunc::Parms + 1, box); 3177 lock->init_req(TypeFunc::Parms + 2, flock); 3178 add_safepoint_edges(lock); 3179 3180 lock = _gvn.transform( lock )->as_Lock(); 3181 3182 // lock has no side-effects, sets few values 3183 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3184 3185 insert_mem_bar(Op_MemBarAcquireLock); 3186 3187 // Add this to the worklist so that the lock can be eliminated 3188 record_for_igvn(lock); 3189 3190 #ifndef PRODUCT 3191 if (PrintLockStatistics) { 3192 // Update the counter for this lock. Don't bother using an atomic 3193 // operation since we don't require absolute accuracy. 3194 lock->create_lock_counter(map()->jvms()); 3195 increment_counter(lock->counter()->addr()); 3196 } 3197 #endif 3198 3199 return flock; 3200 } 3201 3202 3203 //------------------------------shared_unlock---------------------------------- 3204 // Emit unlocking code. 3205 void GraphKit::shared_unlock(Node* box, Node* obj) { 3206 // bci is either a monitorenter bc or InvocationEntryBci 3207 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3208 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3209 3210 if( !GenerateSynchronizationCode ) 3211 return; 3212 if (stopped()) { // Dead monitor? 3213 map()->pop_monitor(); // Kill monitor from debug info 3214 return; 3215 } 3216 3217 // Memory barrier to avoid floating things down past the locked region 3218 insert_mem_bar(Op_MemBarReleaseLock); 3219 3220 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3221 UnlockNode *unlock = new (C) UnlockNode(C, tf); 3222 #ifdef ASSERT 3223 unlock->set_dbg_jvms(sync_jvms()); 3224 #endif 3225 uint raw_idx = Compile::AliasIdxRaw; 3226 unlock->init_req( TypeFunc::Control, control() ); 3227 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3228 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3229 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3230 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3231 3232 unlock->init_req(TypeFunc::Parms + 0, obj); 3233 unlock->init_req(TypeFunc::Parms + 1, box); 3234 unlock = _gvn.transform(unlock)->as_Unlock(); 3235 3236 Node* mem = reset_memory(); 3237 3238 // unlock has no side-effects, sets few values 3239 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3240 3241 // Kill monitor from debug info 3242 map()->pop_monitor( ); 3243 } 3244 3245 //-------------------------------get_layout_helper----------------------------- 3246 // If the given klass is a constant or known to be an array, 3247 // fetch the constant layout helper value into constant_value 3248 // and return (Node*)NULL. Otherwise, load the non-constant 3249 // layout helper value, and return the node which represents it. 3250 // This two-faced routine is useful because allocation sites 3251 // almost always feature constant types. 3252 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3253 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3254 if (!StressReflectiveCode && inst_klass != NULL) { 3255 ciKlass* klass = inst_klass->klass(); 3256 bool xklass = inst_klass->klass_is_exact(); 3257 if (xklass || klass->is_array_klass()) { 3258 jint lhelper = klass->layout_helper(); 3259 if (lhelper != Klass::_lh_neutral_value) { 3260 constant_value = lhelper; 3261 return (Node*) NULL; 3262 } 3263 } 3264 } 3265 constant_value = Klass::_lh_neutral_value; // put in a known value 3266 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3267 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3268 } 3269 3270 // We just put in an allocate/initialize with a big raw-memory effect. 3271 // Hook selected additional alias categories on the initialization. 3272 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3273 MergeMemNode* init_in_merge, 3274 Node* init_out_raw) { 3275 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3276 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3277 3278 Node* prevmem = kit.memory(alias_idx); 3279 init_in_merge->set_memory_at(alias_idx, prevmem); 3280 kit.set_memory(init_out_raw, alias_idx); 3281 } 3282 3283 //---------------------------set_output_for_allocation------------------------- 3284 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3285 const TypeOopPtr* oop_type, 3286 bool deoptimize_on_exception) { 3287 int rawidx = Compile::AliasIdxRaw; 3288 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3289 add_safepoint_edges(alloc); 3290 Node* allocx = _gvn.transform(alloc); 3291 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3292 // create memory projection for i_o 3293 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3294 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3295 3296 // create a memory projection as for the normal control path 3297 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3298 set_memory(malloc, rawidx); 3299 3300 // a normal slow-call doesn't change i_o, but an allocation does 3301 // we create a separate i_o projection for the normal control path 3302 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3303 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3304 3305 // put in an initialization barrier 3306 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3307 rawoop)->as_Initialize(); 3308 assert(alloc->initialization() == init, "2-way macro link must work"); 3309 assert(init ->allocation() == alloc, "2-way macro link must work"); 3310 { 3311 // Extract memory strands which may participate in the new object's 3312 // initialization, and source them from the new InitializeNode. 3313 // This will allow us to observe initializations when they occur, 3314 // and link them properly (as a group) to the InitializeNode. 3315 assert(init->in(InitializeNode::Memory) == malloc, ""); 3316 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3317 init->set_req(InitializeNode::Memory, minit_in); 3318 record_for_igvn(minit_in); // fold it up later, if possible 3319 Node* minit_out = memory(rawidx); 3320 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3321 if (oop_type->isa_aryptr()) { 3322 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3323 int elemidx = C->get_alias_index(telemref); 3324 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3325 } else if (oop_type->isa_instptr()) { 3326 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3327 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3328 ciField* field = ik->nonstatic_field_at(i); 3329 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3330 continue; // do not bother to track really large numbers of fields 3331 // Find (or create) the alias category for this field: 3332 int fieldidx = C->alias_type(field)->index(); 3333 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3334 } 3335 } 3336 } 3337 3338 // Cast raw oop to the real thing... 3339 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3340 javaoop = _gvn.transform(javaoop); 3341 C->set_recent_alloc(control(), javaoop); 3342 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3343 3344 #ifdef ASSERT 3345 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3346 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3347 "Ideal_allocation works"); 3348 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3349 "Ideal_allocation works"); 3350 if (alloc->is_AllocateArray()) { 3351 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3352 "Ideal_allocation works"); 3353 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3354 "Ideal_allocation works"); 3355 } else { 3356 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3357 } 3358 } 3359 #endif //ASSERT 3360 3361 return javaoop; 3362 } 3363 3364 //---------------------------new_instance-------------------------------------- 3365 // This routine takes a klass_node which may be constant (for a static type) 3366 // or may be non-constant (for reflective code). It will work equally well 3367 // for either, and the graph will fold nicely if the optimizer later reduces 3368 // the type to a constant. 3369 // The optional arguments are for specialized use by intrinsics: 3370 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3371 // - If 'return_size_val', report the the total object size to the caller. 3372 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3373 Node* GraphKit::new_instance(Node* klass_node, 3374 Node* extra_slow_test, 3375 Node* *return_size_val, 3376 bool deoptimize_on_exception) { 3377 // Compute size in doublewords 3378 // The size is always an integral number of doublewords, represented 3379 // as a positive bytewise size stored in the klass's layout_helper. 3380 // The layout_helper also encodes (in a low bit) the need for a slow path. 3381 jint layout_con = Klass::_lh_neutral_value; 3382 Node* layout_val = get_layout_helper(klass_node, layout_con); 3383 int layout_is_con = (layout_val == NULL); 3384 3385 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3386 // Generate the initial go-slow test. It's either ALWAYS (return a 3387 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3388 // case) a computed value derived from the layout_helper. 3389 Node* initial_slow_test = NULL; 3390 if (layout_is_con) { 3391 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3392 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3393 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3394 3395 } else { // reflective case 3396 // This reflective path is used by Unsafe.allocateInstance. 3397 // (It may be stress-tested by specifying StressReflectiveCode.) 3398 // Basically, we want to get into the VM is there's an illegal argument. 3399 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3400 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3401 if (extra_slow_test != intcon(0)) { 3402 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3403 } 3404 // (Macro-expander will further convert this to a Bool, if necessary.) 3405 } 3406 3407 // Find the size in bytes. This is easy; it's the layout_helper. 3408 // The size value must be valid even if the slow path is taken. 3409 Node* size = NULL; 3410 if (layout_is_con) { 3411 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3412 } else { // reflective case 3413 // This reflective path is used by clone and Unsafe.allocateInstance. 3414 size = ConvI2X(layout_val); 3415 3416 // Clear the low bits to extract layout_helper_size_in_bytes: 3417 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3418 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3419 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3420 } 3421 if (return_size_val != NULL) { 3422 (*return_size_val) = size; 3423 } 3424 3425 // This is a precise notnull oop of the klass. 3426 // (Actually, it need not be precise if this is a reflective allocation.) 3427 // It's what we cast the result to. 3428 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3429 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3430 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3431 3432 // Now generate allocation code 3433 3434 // The entire memory state is needed for slow path of the allocation 3435 // since GC and deoptimization can happened. 3436 Node *mem = reset_memory(); 3437 set_all_memory(mem); // Create new memory state 3438 3439 AllocateNode* alloc 3440 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3441 control(), mem, i_o(), 3442 size, klass_node, 3443 initial_slow_test); 3444 3445 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3446 } 3447 3448 //-------------------------------new_array------------------------------------- 3449 // helper for both newarray and anewarray 3450 // The 'length' parameter is (obviously) the length of the array. 3451 // See comments on new_instance for the meaning of the other arguments. 3452 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3453 Node* length, // number of array elements 3454 int nargs, // number of arguments to push back for uncommon trap 3455 Node* *return_size_val, 3456 bool deoptimize_on_exception) { 3457 jint layout_con = Klass::_lh_neutral_value; 3458 Node* layout_val = get_layout_helper(klass_node, layout_con); 3459 int layout_is_con = (layout_val == NULL); 3460 3461 if (!layout_is_con && !StressReflectiveCode && 3462 !too_many_traps(Deoptimization::Reason_class_check)) { 3463 // This is a reflective array creation site. 3464 // Optimistically assume that it is a subtype of Object[], 3465 // so that we can fold up all the address arithmetic. 3466 layout_con = Klass::array_layout_helper(T_OBJECT); 3467 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3468 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3469 { BuildCutout unless(this, bol_lh, PROB_MAX); 3470 inc_sp(nargs); 3471 uncommon_trap(Deoptimization::Reason_class_check, 3472 Deoptimization::Action_maybe_recompile); 3473 } 3474 layout_val = NULL; 3475 layout_is_con = true; 3476 } 3477 3478 // Generate the initial go-slow test. Make sure we do not overflow 3479 // if length is huge (near 2Gig) or negative! We do not need 3480 // exact double-words here, just a close approximation of needed 3481 // double-words. We can't add any offset or rounding bits, lest we 3482 // take a size -1 of bytes and make it positive. Use an unsigned 3483 // compare, so negative sizes look hugely positive. 3484 int fast_size_limit = FastAllocateSizeLimit; 3485 if (layout_is_con) { 3486 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3487 // Increase the size limit if we have exact knowledge of array type. 3488 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3489 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3490 } 3491 3492 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3493 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3494 3495 // --- Size Computation --- 3496 // array_size = round_to_heap(array_header + (length << elem_shift)); 3497 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3498 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3499 // The rounding mask is strength-reduced, if possible. 3500 int round_mask = MinObjAlignmentInBytes - 1; 3501 Node* header_size = NULL; 3502 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3503 // (T_BYTE has the weakest alignment and size restrictions...) 3504 if (layout_is_con) { 3505 int hsize = Klass::layout_helper_header_size(layout_con); 3506 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3507 BasicType etype = Klass::layout_helper_element_type(layout_con); 3508 if ((round_mask & ~right_n_bits(eshift)) == 0) 3509 round_mask = 0; // strength-reduce it if it goes away completely 3510 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3511 assert(header_size_min <= hsize, "generic minimum is smallest"); 3512 header_size_min = hsize; 3513 header_size = intcon(hsize + round_mask); 3514 } else { 3515 Node* hss = intcon(Klass::_lh_header_size_shift); 3516 Node* hsm = intcon(Klass::_lh_header_size_mask); 3517 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3518 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3519 Node* mask = intcon(round_mask); 3520 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3521 } 3522 3523 Node* elem_shift = NULL; 3524 if (layout_is_con) { 3525 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3526 if (eshift != 0) 3527 elem_shift = intcon(eshift); 3528 } else { 3529 // There is no need to mask or shift this value. 3530 // The semantics of LShiftINode include an implicit mask to 0x1F. 3531 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3532 elem_shift = layout_val; 3533 } 3534 3535 // Transition to native address size for all offset calculations: 3536 Node* lengthx = ConvI2X(length); 3537 Node* headerx = ConvI2X(header_size); 3538 #ifdef _LP64 3539 { const TypeInt* tilen = _gvn.find_int_type(length); 3540 if (tilen != NULL && tilen->_lo < 0) { 3541 // Add a manual constraint to a positive range. Cf. array_element_address. 3542 jlong size_max = fast_size_limit; 3543 if (size_max > tilen->_hi) size_max = tilen->_hi; 3544 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3545 3546 // Only do a narrow I2L conversion if the range check passed. 3547 IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3548 _gvn.transform(iff); 3549 RegionNode* region = new (C) RegionNode(3); 3550 _gvn.set_type(region, Type::CONTROL); 3551 lengthx = new (C) PhiNode(region, TypeLong::LONG); 3552 _gvn.set_type(lengthx, TypeLong::LONG); 3553 3554 // Range check passed. Use ConvI2L node with narrow type. 3555 Node* passed = IfFalse(iff); 3556 region->init_req(1, passed); 3557 // Make I2L conversion control dependent to prevent it from 3558 // floating above the range check during loop optimizations. 3559 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3560 3561 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3562 region->init_req(2, IfTrue(iff)); 3563 lengthx->init_req(2, ConvI2X(length)); 3564 3565 set_control(region); 3566 record_for_igvn(region); 3567 record_for_igvn(lengthx); 3568 } 3569 } 3570 #endif 3571 3572 // Combine header size (plus rounding) and body size. Then round down. 3573 // This computation cannot overflow, because it is used only in two 3574 // places, one where the length is sharply limited, and the other 3575 // after a successful allocation. 3576 Node* abody = lengthx; 3577 if (elem_shift != NULL) 3578 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3579 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3580 if (round_mask != 0) { 3581 Node* mask = MakeConX(~round_mask); 3582 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3583 } 3584 // else if round_mask == 0, the size computation is self-rounding 3585 3586 if (return_size_val != NULL) { 3587 // This is the size 3588 (*return_size_val) = size; 3589 } 3590 3591 // Now generate allocation code 3592 3593 // The entire memory state is needed for slow path of the allocation 3594 // since GC and deoptimization can happened. 3595 Node *mem = reset_memory(); 3596 set_all_memory(mem); // Create new memory state 3597 3598 if (initial_slow_test->is_Bool()) { 3599 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3600 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3601 } 3602 3603 // Create the AllocateArrayNode and its result projections 3604 AllocateArrayNode* alloc 3605 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3606 control(), mem, i_o(), 3607 size, klass_node, 3608 initial_slow_test, 3609 length); 3610 3611 // Cast to correct type. Note that the klass_node may be constant or not, 3612 // and in the latter case the actual array type will be inexact also. 3613 // (This happens via a non-constant argument to inline_native_newArray.) 3614 // In any case, the value of klass_node provides the desired array type. 3615 const TypeInt* length_type = _gvn.find_int_type(length); 3616 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3617 if (ary_type->isa_aryptr() && length_type != NULL) { 3618 // Try to get a better type than POS for the size 3619 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3620 } 3621 3622 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3623 3624 // Cast length on remaining path to be as narrow as possible 3625 if (map()->find_edge(length) >= 0) { 3626 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3627 if (ccast != length) { 3628 _gvn.set_type_bottom(ccast); 3629 record_for_igvn(ccast); 3630 replace_in_map(length, ccast); 3631 } 3632 } 3633 3634 return javaoop; 3635 } 3636 3637 // The following "Ideal_foo" functions are placed here because they recognize 3638 // the graph shapes created by the functions immediately above. 3639 3640 //---------------------------Ideal_allocation---------------------------------- 3641 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3642 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3643 if (ptr == NULL) { // reduce dumb test in callers 3644 return NULL; 3645 } 3646 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3647 ptr = ptr->in(1); 3648 if (ptr == NULL) return NULL; 3649 } 3650 // Return NULL for allocations with several casts: 3651 // j.l.reflect.Array.newInstance(jobject, jint) 3652 // Object.clone() 3653 // to keep more precise type from last cast. 3654 if (ptr->is_Proj()) { 3655 Node* allo = ptr->in(0); 3656 if (allo != NULL && allo->is_Allocate()) { 3657 return allo->as_Allocate(); 3658 } 3659 } 3660 // Report failure to match. 3661 return NULL; 3662 } 3663 3664 // Fancy version which also strips off an offset (and reports it to caller). 3665 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3666 intptr_t& offset) { 3667 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3668 if (base == NULL) return NULL; 3669 return Ideal_allocation(base, phase); 3670 } 3671 3672 // Trace Initialize <- Proj[Parm] <- Allocate 3673 AllocateNode* InitializeNode::allocation() { 3674 Node* rawoop = in(InitializeNode::RawAddress); 3675 if (rawoop->is_Proj()) { 3676 Node* alloc = rawoop->in(0); 3677 if (alloc->is_Allocate()) { 3678 return alloc->as_Allocate(); 3679 } 3680 } 3681 return NULL; 3682 } 3683 3684 // Trace Allocate -> Proj[Parm] -> Initialize 3685 InitializeNode* AllocateNode::initialization() { 3686 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3687 if (rawoop == NULL) return NULL; 3688 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3689 Node* init = rawoop->fast_out(i); 3690 if (init->is_Initialize()) { 3691 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3692 return init->as_Initialize(); 3693 } 3694 } 3695 return NULL; 3696 } 3697 3698 //----------------------------- loop predicates --------------------------- 3699 3700 //------------------------------add_predicate_impl---------------------------- 3701 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3702 // Too many traps seen? 3703 if (too_many_traps(reason)) { 3704 #ifdef ASSERT 3705 if (TraceLoopPredicate) { 3706 int tc = C->trap_count(reason); 3707 tty->print("too many traps=%s tcount=%d in ", 3708 Deoptimization::trap_reason_name(reason), tc); 3709 method()->print(); // which method has too many predicate traps 3710 tty->cr(); 3711 } 3712 #endif 3713 // We cannot afford to take more traps here, 3714 // do not generate predicate. 3715 return; 3716 } 3717 3718 Node *cont = _gvn.intcon(1); 3719 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3720 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3721 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3722 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3723 C->add_predicate_opaq(opq); 3724 { 3725 PreserveJVMState pjvms(this); 3726 set_control(iffalse); 3727 inc_sp(nargs); 3728 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3729 } 3730 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3731 set_control(iftrue); 3732 } 3733 3734 //------------------------------add_predicate--------------------------------- 3735 void GraphKit::add_predicate(int nargs) { 3736 if (UseLoopPredicate) { 3737 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3738 } 3739 // loop's limit check predicate should be near the loop. 3740 if (LoopLimitCheck) { 3741 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3742 } 3743 } 3744 3745 //----------------------------- store barriers ---------------------------- 3746 #define __ ideal. 3747 3748 void GraphKit::sync_kit(IdealKit& ideal) { 3749 set_all_memory(__ merged_memory()); 3750 set_i_o(__ i_o()); 3751 set_control(__ ctrl()); 3752 } 3753 3754 void GraphKit::final_sync(IdealKit& ideal) { 3755 // Final sync IdealKit and graphKit. 3756 sync_kit(ideal); 3757 } 3758 3759 // vanilla/CMS post barrier 3760 // Insert a write-barrier store. This is to let generational GC work; we have 3761 // to flag all oop-stores before the next GC point. 3762 void GraphKit::write_barrier_post(Node* oop_store, 3763 Node* obj, 3764 Node* adr, 3765 uint adr_idx, 3766 Node* val, 3767 bool use_precise) { 3768 // No store check needed if we're storing a NULL or an old object 3769 // (latter case is probably a string constant). The concurrent 3770 // mark sweep garbage collector, however, needs to have all nonNull 3771 // oop updates flagged via card-marks. 3772 if (val != NULL && val->is_Con()) { 3773 // must be either an oop or NULL 3774 const Type* t = val->bottom_type(); 3775 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3776 // stores of null never (?) need barriers 3777 return; 3778 } 3779 3780 if (use_ReduceInitialCardMarks() 3781 && obj == just_allocated_object(control())) { 3782 // We can skip marks on a freshly-allocated object in Eden. 3783 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3784 // That routine informs GC to take appropriate compensating steps, 3785 // upon a slow-path allocation, so as to make this card-mark 3786 // elision safe. 3787 return; 3788 } 3789 3790 if (!use_precise) { 3791 // All card marks for a (non-array) instance are in one place: 3792 adr = obj; 3793 } 3794 // (Else it's an array (or unknown), and we want more precise card marks.) 3795 assert(adr != NULL, ""); 3796 3797 IdealKit ideal(this, true); 3798 3799 // Convert the pointer to an int prior to doing math on it 3800 Node* cast = __ CastPX(__ ctrl(), adr); 3801 3802 // Divide by card size 3803 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3804 "Only one we handle so far."); 3805 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3806 3807 // Combine card table base and card offset 3808 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3809 3810 // Get the alias_index for raw card-mark memory 3811 int adr_type = Compile::AliasIdxRaw; 3812 Node* zero = __ ConI(0); // Dirty card value 3813 BasicType bt = T_BYTE; 3814 3815 if (UseCondCardMark) { 3816 // The classic GC reference write barrier is typically implemented 3817 // as a store into the global card mark table. Unfortunately 3818 // unconditional stores can result in false sharing and excessive 3819 // coherence traffic as well as false transactional aborts. 3820 // UseCondCardMark enables MP "polite" conditional card mark 3821 // stores. In theory we could relax the load from ctrl() to 3822 // no_ctrl, but that doesn't buy much latitude. 3823 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3824 __ if_then(card_val, BoolTest::ne, zero); 3825 } 3826 3827 // Smash zero into card 3828 if( !UseConcMarkSweepGC ) { 3829 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); 3830 } else { 3831 // Specialized path for CM store barrier 3832 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3833 } 3834 3835 if (UseCondCardMark) { 3836 __ end_if(); 3837 } 3838 3839 // Final sync IdealKit and GraphKit. 3840 final_sync(ideal); 3841 } 3842 3843 // G1 pre/post barriers 3844 void GraphKit::g1_write_barrier_pre(bool do_load, 3845 Node* obj, 3846 Node* adr, 3847 uint alias_idx, 3848 Node* val, 3849 const TypeOopPtr* val_type, 3850 Node* pre_val, 3851 BasicType bt) { 3852 3853 // Some sanity checks 3854 // Note: val is unused in this routine. 3855 3856 if (do_load) { 3857 // We need to generate the load of the previous value 3858 assert(obj != NULL, "must have a base"); 3859 assert(adr != NULL, "where are loading from?"); 3860 assert(pre_val == NULL, "loaded already?"); 3861 assert(val_type != NULL, "need a type"); 3862 } else { 3863 // In this case both val_type and alias_idx are unused. 3864 assert(pre_val != NULL, "must be loaded already"); 3865 // Nothing to be done if pre_val is null. 3866 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3867 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3868 } 3869 assert(bt == T_OBJECT, "or we shouldn't be here"); 3870 3871 IdealKit ideal(this, true); 3872 3873 Node* tls = __ thread(); // ThreadLocalStorage 3874 3875 Node* no_ctrl = NULL; 3876 Node* no_base = __ top(); 3877 Node* zero = __ ConI(0); 3878 Node* zeroX = __ ConX(0); 3879 3880 float likely = PROB_LIKELY(0.999); 3881 float unlikely = PROB_UNLIKELY(0.999); 3882 3883 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3884 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3885 3886 // Offsets into the thread 3887 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3888 PtrQueue::byte_offset_of_active()); 3889 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3890 PtrQueue::byte_offset_of_index()); 3891 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3892 PtrQueue::byte_offset_of_buf()); 3893 3894 // Now the actual pointers into the thread 3895 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3896 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3897 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3898 3899 // Now some of the values 3900 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3901 3902 // if (!marking) 3903 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3904 BasicType index_bt = TypeX_X->basic_type(); 3905 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3906 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3907 3908 if (do_load) { 3909 // load original value 3910 // alias_idx correct?? 3911 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3912 } 3913 3914 // if (pre_val != NULL) 3915 __ if_then(pre_val, BoolTest::ne, null()); { 3916 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3917 3918 // is the queue for this thread full? 3919 __ if_then(index, BoolTest::ne, zeroX, likely); { 3920 3921 // decrement the index 3922 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3923 3924 // Now get the buffer location we will log the previous value into and store it 3925 Node *log_addr = __ AddP(no_base, buffer, next_index); 3926 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 3927 // update the index 3928 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 3929 3930 } __ else_(); { 3931 3932 // logging buffer is full, call the runtime 3933 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3934 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3935 } __ end_if(); // (!index) 3936 } __ end_if(); // (pre_val != NULL) 3937 } __ end_if(); // (!marking) 3938 3939 // Final sync IdealKit and GraphKit. 3940 final_sync(ideal); 3941 } 3942 3943 // 3944 // Update the card table and add card address to the queue 3945 // 3946 void GraphKit::g1_mark_card(IdealKit& ideal, 3947 Node* card_adr, 3948 Node* oop_store, 3949 uint oop_alias_idx, 3950 Node* index, 3951 Node* index_adr, 3952 Node* buffer, 3953 const TypeFunc* tf) { 3954 3955 Node* zero = __ ConI(0); 3956 Node* zeroX = __ ConX(0); 3957 Node* no_base = __ top(); 3958 BasicType card_bt = T_BYTE; 3959 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3960 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3961 3962 // Now do the queue work 3963 __ if_then(index, BoolTest::ne, zeroX); { 3964 3965 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3966 Node* log_addr = __ AddP(no_base, buffer, next_index); 3967 3968 // Order, see storeCM. 3969 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 3970 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 3971 3972 } __ else_(); { 3973 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 3974 } __ end_if(); 3975 3976 } 3977 3978 void GraphKit::g1_write_barrier_post(Node* oop_store, 3979 Node* obj, 3980 Node* adr, 3981 uint alias_idx, 3982 Node* val, 3983 BasicType bt, 3984 bool use_precise) { 3985 // If we are writing a NULL then we need no post barrier 3986 3987 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 3988 // Must be NULL 3989 const Type* t = val->bottom_type(); 3990 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 3991 // No post barrier if writing NULLx 3992 return; 3993 } 3994 3995 if (!use_precise) { 3996 // All card marks for a (non-array) instance are in one place: 3997 adr = obj; 3998 } 3999 // (Else it's an array (or unknown), and we want more precise card marks.) 4000 assert(adr != NULL, ""); 4001 4002 IdealKit ideal(this, true); 4003 4004 Node* tls = __ thread(); // ThreadLocalStorage 4005 4006 Node* no_base = __ top(); 4007 float likely = PROB_LIKELY(0.999); 4008 float unlikely = PROB_UNLIKELY(0.999); 4009 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4010 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4011 Node* zeroX = __ ConX(0); 4012 4013 // Get the alias_index for raw card-mark memory 4014 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4015 4016 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4017 4018 // Offsets into the thread 4019 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4020 PtrQueue::byte_offset_of_index()); 4021 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4022 PtrQueue::byte_offset_of_buf()); 4023 4024 // Pointers into the thread 4025 4026 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4027 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4028 4029 // Now some values 4030 // Use ctrl to avoid hoisting these values past a safepoint, which could 4031 // potentially reset these fields in the JavaThread. 4032 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4033 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4034 4035 // Convert the store obj pointer to an int prior to doing math on it 4036 // Must use ctrl to prevent "integerized oop" existing across safepoint 4037 Node* cast = __ CastPX(__ ctrl(), adr); 4038 4039 // Divide pointer by card size 4040 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4041 4042 // Combine card table base and card offset 4043 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4044 4045 // If we know the value being stored does it cross regions? 4046 4047 if (val != NULL) { 4048 // Does the store cause us to cross regions? 4049 4050 // Should be able to do an unsigned compare of region_size instead of 4051 // and extra shift. Do we have an unsigned compare?? 4052 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4053 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4054 4055 // if (xor_res == 0) same region so skip 4056 __ if_then(xor_res, BoolTest::ne, zeroX); { 4057 4058 // No barrier if we are storing a NULL 4059 __ if_then(val, BoolTest::ne, null(), unlikely); { 4060 4061 // Ok must mark the card if not already dirty 4062 4063 // load the original value of the card 4064 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4065 4066 __ if_then(card_val, BoolTest::ne, young_card); { 4067 sync_kit(ideal); 4068 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4069 insert_mem_bar(Op_MemBarVolatile, oop_store); 4070 __ sync_kit(this); 4071 4072 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4073 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4074 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4075 } __ end_if(); 4076 } __ end_if(); 4077 } __ end_if(); 4078 } __ end_if(); 4079 } else { 4080 // Object.clone() instrinsic uses this path. 4081 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4082 } 4083 4084 // Final sync IdealKit and GraphKit. 4085 final_sync(ideal); 4086 } 4087 #undef __ 4088 4089 4090 4091 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4092 if (java_lang_String::has_offset_field()) { 4093 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4094 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4095 false, NULL, 0); 4096 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4097 int offset_field_idx = C->get_alias_index(offset_field_type); 4098 return make_load(ctrl, 4099 basic_plus_adr(str, str, offset_offset), 4100 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4101 } else { 4102 return intcon(0); 4103 } 4104 } 4105 4106 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4107 if (java_lang_String::has_count_field()) { 4108 int count_offset = java_lang_String::count_offset_in_bytes(); 4109 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4110 false, NULL, 0); 4111 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4112 int count_field_idx = C->get_alias_index(count_field_type); 4113 return make_load(ctrl, 4114 basic_plus_adr(str, str, count_offset), 4115 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4116 } else { 4117 return load_array_length(load_String_value(ctrl, str)); 4118 } 4119 } 4120 4121 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4122 int value_offset = java_lang_String::value_offset_in_bytes(); 4123 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4124 false, NULL, 0); 4125 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4126 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4127 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4128 ciTypeArrayKlass::make(T_CHAR), true, 0); 4129 int value_field_idx = C->get_alias_index(value_field_type); 4130 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4131 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4132 // String.value field is known to be @Stable. 4133 if (UseImplicitStableValues) { 4134 load = cast_array_to_stable(load, value_type); 4135 } 4136 return load; 4137 } 4138 4139 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4140 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4141 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4142 false, NULL, 0); 4143 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4144 int offset_field_idx = C->get_alias_index(offset_field_type); 4145 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4146 value, T_INT, offset_field_idx, MemNode::unordered); 4147 } 4148 4149 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4150 int value_offset = java_lang_String::value_offset_in_bytes(); 4151 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4152 false, NULL, 0); 4153 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4154 4155 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4156 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4157 } 4158 4159 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4160 int count_offset = java_lang_String::count_offset_in_bytes(); 4161 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4162 false, NULL, 0); 4163 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4164 int count_field_idx = C->get_alias_index(count_field_type); 4165 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4166 value, T_INT, count_field_idx, MemNode::unordered); 4167 } 4168 4169 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4170 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4171 // assumption of CCP analysis. 4172 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); 4173 }