1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp" 28 #include "gc_implementation/g1/heapRegion.hpp" 29 #include "gc_interface/collectedHeap.hpp" 30 #include "memory/barrierSet.hpp" 31 #include "memory/cardTableModRefBS.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/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 } 432 433 //--------------------------use_exception_state-------------------------------- 434 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 435 if (failing()) { stop(); return top(); } 436 Node* region = phi_map->control(); 437 Node* hidden_merge_mark = root(); 438 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 439 Node* ex_oop = clear_saved_ex_oop(phi_map); 440 if (region->in(0) == hidden_merge_mark) { 441 // Special marking for internal ex-states. Process the phis now. 442 region->set_req(0, region); // now it's an ordinary region 443 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 444 // Note: Setting the jvms also sets the bci and sp. 445 set_control(_gvn.transform(region)); 446 uint tos = jvms()->stkoff() + sp(); 447 for (uint i = 1; i < tos; i++) { 448 Node* x = phi_map->in(i); 449 if (x->in(0) == region) { 450 assert(x->is_Phi(), "expected a special phi"); 451 phi_map->set_req(i, _gvn.transform(x)); 452 } 453 } 454 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 455 Node* x = mms.memory(); 456 if (x->in(0) == region) { 457 assert(x->is_Phi(), "nobody else uses a hidden region"); 458 mms.set_memory(_gvn.transform(x)); 459 } 460 } 461 if (ex_oop->in(0) == region) { 462 assert(ex_oop->is_Phi(), "expected a special phi"); 463 ex_oop = _gvn.transform(ex_oop); 464 } 465 } else { 466 set_jvms(phi_map->jvms()); 467 } 468 469 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 470 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 471 return ex_oop; 472 } 473 474 //---------------------------------java_bc------------------------------------- 475 Bytecodes::Code GraphKit::java_bc() const { 476 ciMethod* method = this->method(); 477 int bci = this->bci(); 478 if (method != NULL && bci != InvocationEntryBci) 479 return method->java_code_at_bci(bci); 480 else 481 return Bytecodes::_illegal; 482 } 483 484 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 485 bool must_throw) { 486 // if the exception capability is set, then we will generate code 487 // to check the JavaThread.should_post_on_exceptions flag to see 488 // if we actually need to report exception events (for this 489 // thread). If we don't need to report exception events, we will 490 // take the normal fast path provided by add_exception_events. If 491 // exception event reporting is enabled for this thread, we will 492 // take the uncommon_trap in the BuildCutout below. 493 494 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 495 Node* jthread = _gvn.transform(new (C) ThreadLocalNode()); 496 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 497 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 498 499 // Test the should_post_on_exceptions_flag vs. 0 500 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) ); 501 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 502 503 // Branch to slow_path if should_post_on_exceptions_flag was true 504 { BuildCutout unless(this, tst, PROB_MAX); 505 // Do not try anything fancy if we're notifying the VM on every throw. 506 // Cf. case Bytecodes::_athrow in parse2.cpp. 507 uncommon_trap(reason, Deoptimization::Action_none, 508 (ciKlass*)NULL, (char*)NULL, must_throw); 509 } 510 511 } 512 513 //------------------------------builtin_throw---------------------------------- 514 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 515 bool must_throw = true; 516 517 if (env()->jvmti_can_post_on_exceptions()) { 518 // check if we must post exception events, take uncommon trap if so 519 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 520 // here if should_post_on_exceptions is false 521 // continue on with the normal codegen 522 } 523 524 // If this particular condition has not yet happened at this 525 // bytecode, then use the uncommon trap mechanism, and allow for 526 // a future recompilation if several traps occur here. 527 // If the throw is hot, try to use a more complicated inline mechanism 528 // which keeps execution inside the compiled code. 529 bool treat_throw_as_hot = false; 530 ciMethodData* md = method()->method_data(); 531 532 if (ProfileTraps) { 533 if (too_many_traps(reason)) { 534 treat_throw_as_hot = true; 535 } 536 // (If there is no MDO at all, assume it is early in 537 // execution, and that any deopts are part of the 538 // startup transient, and don't need to be remembered.) 539 540 // Also, if there is a local exception handler, treat all throws 541 // as hot if there has been at least one in this method. 542 if (C->trap_count(reason) != 0 543 && method()->method_data()->trap_count(reason) != 0 544 && has_ex_handler()) { 545 treat_throw_as_hot = true; 546 } 547 } 548 549 // If this throw happens frequently, an uncommon trap might cause 550 // a performance pothole. If there is a local exception handler, 551 // and if this particular bytecode appears to be deoptimizing often, 552 // let us handle the throw inline, with a preconstructed instance. 553 // Note: If the deopt count has blown up, the uncommon trap 554 // runtime is going to flush this nmethod, not matter what. 555 if (treat_throw_as_hot 556 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 557 // If the throw is local, we use a pre-existing instance and 558 // punt on the backtrace. This would lead to a missing backtrace 559 // (a repeat of 4292742) if the backtrace object is ever asked 560 // for its backtrace. 561 // Fixing this remaining case of 4292742 requires some flavor of 562 // escape analysis. Leave that for the future. 563 ciInstance* ex_obj = NULL; 564 switch (reason) { 565 case Deoptimization::Reason_null_check: 566 ex_obj = env()->NullPointerException_instance(); 567 break; 568 case Deoptimization::Reason_div0_check: 569 ex_obj = env()->ArithmeticException_instance(); 570 break; 571 case Deoptimization::Reason_range_check: 572 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 573 break; 574 case Deoptimization::Reason_class_check: 575 if (java_bc() == Bytecodes::_aastore) { 576 ex_obj = env()->ArrayStoreException_instance(); 577 } else { 578 ex_obj = env()->ClassCastException_instance(); 579 } 580 break; 581 } 582 if (failing()) { stop(); return; } // exception allocation might fail 583 if (ex_obj != NULL) { 584 // Cheat with a preallocated exception object. 585 if (C->log() != NULL) 586 C->log()->elem("hot_throw preallocated='1' reason='%s'", 587 Deoptimization::trap_reason_name(reason)); 588 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 589 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) ); 590 591 // Clear the detail message of the preallocated exception object. 592 // Weblogic sometimes mutates the detail message of exceptions 593 // using reflection. 594 int offset = java_lang_Throwable::get_detailMessage_offset(); 595 const TypePtr* adr_typ = ex_con->add_offset(offset); 596 597 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 598 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 599 // Conservatively release stores of object references. 600 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release); 601 602 add_exception_state(make_exception_state(ex_node)); 603 return; 604 } 605 } 606 607 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 608 // It won't be much cheaper than bailing to the interp., since we'll 609 // have to pass up all the debug-info, and the runtime will have to 610 // create the stack trace. 611 612 // Usual case: Bail to interpreter. 613 // Reserve the right to recompile if we haven't seen anything yet. 614 615 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 616 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 617 if (treat_throw_as_hot 618 && (method()->method_data()->trap_recompiled_at(bci(), m) 619 || C->too_many_traps(reason))) { 620 // We cannot afford to take more traps here. Suffer in the interpreter. 621 if (C->log() != NULL) 622 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 623 Deoptimization::trap_reason_name(reason), 624 C->trap_count(reason)); 625 action = Deoptimization::Action_none; 626 } 627 628 // "must_throw" prunes the JVM state to include only the stack, if there 629 // are no local exception handlers. This should cut down on register 630 // allocation time and code size, by drastically reducing the number 631 // of in-edges on the call to the uncommon trap. 632 633 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 634 } 635 636 637 //----------------------------PreserveJVMState--------------------------------- 638 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 639 debug_only(kit->verify_map()); 640 _kit = kit; 641 _map = kit->map(); // preserve the map 642 _sp = kit->sp(); 643 kit->set_map(clone_map ? kit->clone_map() : NULL); 644 Compile::current()->inc_preserve_jvm_state(); 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 Compile::current()->dec_preserve_jvm_state(); 663 } 664 665 666 //-----------------------------BuildCutout------------------------------------- 667 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 668 : PreserveJVMState(kit) 669 { 670 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 671 SafePointNode* outer_map = _map; // preserved map is caller's 672 SafePointNode* inner_map = kit->map(); 673 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 674 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) )); 675 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) )); 676 } 677 BuildCutout::~BuildCutout() { 678 GraphKit* kit = _kit; 679 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 680 } 681 682 //---------------------------PreserveReexecuteState---------------------------- 683 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 684 assert(!kit->stopped(), "must call stopped() before"); 685 _kit = kit; 686 _sp = kit->sp(); 687 _reexecute = kit->jvms()->_reexecute; 688 } 689 PreserveReexecuteState::~PreserveReexecuteState() { 690 if (_kit->stopped()) return; 691 _kit->jvms()->_reexecute = _reexecute; 692 _kit->set_sp(_sp); 693 } 694 695 //------------------------------clone_map-------------------------------------- 696 // Implementation of PreserveJVMState 697 // 698 // Only clone_map(...) here. If this function is only used in the 699 // PreserveJVMState class we may want to get rid of this extra 700 // function eventually and do it all there. 701 702 SafePointNode* GraphKit::clone_map() { 703 if (map() == NULL) return NULL; 704 705 // Clone the memory edge first 706 Node* mem = MergeMemNode::make(C, map()->memory()); 707 gvn().set_type_bottom(mem); 708 709 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 710 JVMState* jvms = this->jvms(); 711 JVMState* clonejvms = jvms->clone_shallow(C); 712 clonemap->set_memory(mem); 713 clonemap->set_jvms(clonejvms); 714 clonejvms->set_map(clonemap); 715 record_for_igvn(clonemap); 716 gvn().set_type_bottom(clonemap); 717 return clonemap; 718 } 719 720 721 //-----------------------------set_map_clone----------------------------------- 722 void GraphKit::set_map_clone(SafePointNode* m) { 723 _map = m; 724 _map = clone_map(); 725 _map->set_next_exception(NULL); 726 debug_only(verify_map()); 727 } 728 729 730 //----------------------------kill_dead_locals--------------------------------- 731 // Detect any locals which are known to be dead, and force them to top. 732 void GraphKit::kill_dead_locals() { 733 // Consult the liveness information for the locals. If any 734 // of them are unused, then they can be replaced by top(). This 735 // should help register allocation time and cut down on the size 736 // of the deoptimization information. 737 738 // This call is made from many of the bytecode handling 739 // subroutines called from the Big Switch in do_one_bytecode. 740 // Every bytecode which might include a slow path is responsible 741 // for killing its dead locals. The more consistent we 742 // are about killing deads, the fewer useless phis will be 743 // constructed for them at various merge points. 744 745 // bci can be -1 (InvocationEntryBci). We return the entry 746 // liveness for the method. 747 748 if (method() == NULL || method()->code_size() == 0) { 749 // We are building a graph for a call to a native method. 750 // All locals are live. 751 return; 752 } 753 754 ResourceMark rm; 755 756 // Consult the liveness information for the locals. If any 757 // of them are unused, then they can be replaced by top(). This 758 // should help register allocation time and cut down on the size 759 // of the deoptimization information. 760 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 761 762 int len = (int)live_locals.size(); 763 assert(len <= jvms()->loc_size(), "too many live locals"); 764 for (int local = 0; local < len; local++) { 765 if (!live_locals.at(local)) { 766 set_local(local, top()); 767 } 768 } 769 } 770 771 #ifdef ASSERT 772 //-------------------------dead_locals_are_killed------------------------------ 773 // Return true if all dead locals are set to top in the map. 774 // Used to assert "clean" debug info at various points. 775 bool GraphKit::dead_locals_are_killed() { 776 if (method() == NULL || method()->code_size() == 0) { 777 // No locals need to be dead, so all is as it should be. 778 return true; 779 } 780 781 // Make sure somebody called kill_dead_locals upstream. 782 ResourceMark rm; 783 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 784 if (jvms->loc_size() == 0) continue; // no locals to consult 785 SafePointNode* map = jvms->map(); 786 ciMethod* method = jvms->method(); 787 int bci = jvms->bci(); 788 if (jvms == this->jvms()) { 789 bci = this->bci(); // it might not yet be synched 790 } 791 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 792 int len = (int)live_locals.size(); 793 if (!live_locals.is_valid() || len == 0) 794 // This method is trivial, or is poisoned by a breakpoint. 795 return true; 796 assert(len == jvms->loc_size(), "live map consistent with locals map"); 797 for (int local = 0; local < len; local++) { 798 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 799 if (PrintMiscellaneous && (Verbose || WizardMode)) { 800 tty->print_cr("Zombie local %d: ", local); 801 jvms->dump(); 802 } 803 return false; 804 } 805 } 806 } 807 return true; 808 } 809 810 #endif //ASSERT 811 812 // Helper function for enforcing certain bytecodes to reexecute if 813 // deoptimization happens 814 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 815 ciMethod* cur_method = jvms->method(); 816 int cur_bci = jvms->bci(); 817 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 818 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 819 return Interpreter::bytecode_should_reexecute(code) || 820 is_anewarray && code == Bytecodes::_multianewarray; 821 // Reexecute _multianewarray bytecode which was replaced with 822 // sequence of [a]newarray. See Parse::do_multianewarray(). 823 // 824 // Note: interpreter should not have it set since this optimization 825 // is limited by dimensions and guarded by flag so in some cases 826 // multianewarray() runtime calls will be generated and 827 // the bytecode should not be reexecutes (stack will not be reset). 828 } else 829 return false; 830 } 831 832 // Helper function for adding JVMState and debug information to node 833 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 834 // Add the safepoint edges to the call (or other safepoint). 835 836 // Make sure dead locals are set to top. This 837 // should help register allocation time and cut down on the size 838 // of the deoptimization information. 839 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 840 841 // Walk the inline list to fill in the correct set of JVMState's 842 // Also fill in the associated edges for each JVMState. 843 844 // If the bytecode needs to be reexecuted we need to put 845 // the arguments back on the stack. 846 const bool should_reexecute = jvms()->should_reexecute(); 847 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 848 849 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 850 // undefined if the bci is different. This is normal for Parse but it 851 // should not happen for LibraryCallKit because only one bci is processed. 852 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 853 "in LibraryCallKit the reexecute bit should not change"); 854 855 // If we are guaranteed to throw, we can prune everything but the 856 // input to the current bytecode. 857 bool can_prune_locals = false; 858 uint stack_slots_not_pruned = 0; 859 int inputs = 0, depth = 0; 860 if (must_throw) { 861 assert(method() == youngest_jvms->method(), "sanity"); 862 if (compute_stack_effects(inputs, depth)) { 863 can_prune_locals = true; 864 stack_slots_not_pruned = inputs; 865 } 866 } 867 868 if (env()->should_retain_local_variables()) { 869 // At any safepoint, this method can get breakpointed, which would 870 // then require an immediate deoptimization. 871 can_prune_locals = false; // do not prune locals 872 stack_slots_not_pruned = 0; 873 } 874 875 // do not scribble on the input jvms 876 JVMState* out_jvms = youngest_jvms->clone_deep(C); 877 call->set_jvms(out_jvms); // Start jvms list for call node 878 879 // For a known set of bytecodes, the interpreter should reexecute them if 880 // deoptimization happens. We set the reexecute state for them here 881 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 882 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 883 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 884 } 885 886 // Presize the call: 887 DEBUG_ONLY(uint non_debug_edges = call->req()); 888 call->add_req_batch(top(), youngest_jvms->debug_depth()); 889 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 890 891 // Set up edges so that the call looks like this: 892 // Call [state:] ctl io mem fptr retadr 893 // [parms:] parm0 ... parmN 894 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 895 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 896 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 897 // Note that caller debug info precedes callee debug info. 898 899 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 900 uint debug_ptr = call->req(); 901 902 // Loop over the map input edges associated with jvms, add them 903 // to the call node, & reset all offsets to match call node array. 904 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 905 uint debug_end = debug_ptr; 906 uint debug_start = debug_ptr - in_jvms->debug_size(); 907 debug_ptr = debug_start; // back up the ptr 908 909 uint p = debug_start; // walks forward in [debug_start, debug_end) 910 uint j, k, l; 911 SafePointNode* in_map = in_jvms->map(); 912 out_jvms->set_map(call); 913 914 if (can_prune_locals) { 915 assert(in_jvms->method() == out_jvms->method(), "sanity"); 916 // If the current throw can reach an exception handler in this JVMS, 917 // then we must keep everything live that can reach that handler. 918 // As a quick and dirty approximation, we look for any handlers at all. 919 if (in_jvms->method()->has_exception_handlers()) { 920 can_prune_locals = false; 921 } 922 } 923 924 // Add the Locals 925 k = in_jvms->locoff(); 926 l = in_jvms->loc_size(); 927 out_jvms->set_locoff(p); 928 if (!can_prune_locals) { 929 for (j = 0; j < l; j++) 930 call->set_req(p++, in_map->in(k+j)); 931 } else { 932 p += l; // already set to top above by add_req_batch 933 } 934 935 // Add the Expression Stack 936 k = in_jvms->stkoff(); 937 l = in_jvms->sp(); 938 out_jvms->set_stkoff(p); 939 if (!can_prune_locals) { 940 for (j = 0; j < l; j++) 941 call->set_req(p++, in_map->in(k+j)); 942 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 943 // Divide stack into {S0,...,S1}, where S0 is set to top. 944 uint s1 = stack_slots_not_pruned; 945 stack_slots_not_pruned = 0; // for next iteration 946 if (s1 > l) s1 = l; 947 uint s0 = l - s1; 948 p += s0; // skip the tops preinstalled by add_req_batch 949 for (j = s0; j < l; j++) 950 call->set_req(p++, in_map->in(k+j)); 951 } else { 952 p += l; // already set to top above by add_req_batch 953 } 954 955 // Add the Monitors 956 k = in_jvms->monoff(); 957 l = in_jvms->mon_size(); 958 out_jvms->set_monoff(p); 959 for (j = 0; j < l; j++) 960 call->set_req(p++, in_map->in(k+j)); 961 962 // Copy any scalar object fields. 963 k = in_jvms->scloff(); 964 l = in_jvms->scl_size(); 965 out_jvms->set_scloff(p); 966 for (j = 0; j < l; j++) 967 call->set_req(p++, in_map->in(k+j)); 968 969 // Finish the new jvms. 970 out_jvms->set_endoff(p); 971 972 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 973 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 974 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 975 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 976 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 977 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 978 979 // Update the two tail pointers in parallel. 980 out_jvms = out_jvms->caller(); 981 in_jvms = in_jvms->caller(); 982 } 983 984 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 985 986 // Test the correctness of JVMState::debug_xxx accessors: 987 assert(call->jvms()->debug_start() == non_debug_edges, ""); 988 assert(call->jvms()->debug_end() == call->req(), ""); 989 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 990 } 991 992 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 993 Bytecodes::Code code = java_bc(); 994 if (code == Bytecodes::_wide) { 995 code = method()->java_code_at_bci(bci() + 1); 996 } 997 998 BasicType rtype = T_ILLEGAL; 999 int rsize = 0; 1000 1001 if (code != Bytecodes::_illegal) { 1002 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1003 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1004 if (rtype < T_CONFLICT) 1005 rsize = type2size[rtype]; 1006 } 1007 1008 switch (code) { 1009 case Bytecodes::_illegal: 1010 return false; 1011 1012 case Bytecodes::_ldc: 1013 case Bytecodes::_ldc_w: 1014 case Bytecodes::_ldc2_w: 1015 inputs = 0; 1016 break; 1017 1018 case Bytecodes::_dup: inputs = 1; break; 1019 case Bytecodes::_dup_x1: inputs = 2; break; 1020 case Bytecodes::_dup_x2: inputs = 3; break; 1021 case Bytecodes::_dup2: inputs = 2; break; 1022 case Bytecodes::_dup2_x1: inputs = 3; break; 1023 case Bytecodes::_dup2_x2: inputs = 4; break; 1024 case Bytecodes::_swap: inputs = 2; break; 1025 case Bytecodes::_arraylength: inputs = 1; break; 1026 1027 case Bytecodes::_getstatic: 1028 case Bytecodes::_putstatic: 1029 case Bytecodes::_getfield: 1030 case Bytecodes::_putfield: 1031 { 1032 bool ignored_will_link; 1033 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1034 int size = field->type()->size(); 1035 bool is_get = (depth >= 0), is_static = (depth & 1); 1036 inputs = (is_static ? 0 : 1); 1037 if (is_get) { 1038 depth = size - inputs; 1039 } else { 1040 inputs += size; // putxxx pops the value from the stack 1041 depth = - inputs; 1042 } 1043 } 1044 break; 1045 1046 case Bytecodes::_invokevirtual: 1047 case Bytecodes::_invokespecial: 1048 case Bytecodes::_invokestatic: 1049 case Bytecodes::_invokedynamic: 1050 case Bytecodes::_invokeinterface: 1051 { 1052 bool ignored_will_link; 1053 ciSignature* declared_signature = NULL; 1054 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1055 assert(declared_signature != NULL, "cannot be null"); 1056 inputs = declared_signature->arg_size_for_bc(code); 1057 int size = declared_signature->return_type()->size(); 1058 depth = size - inputs; 1059 } 1060 break; 1061 1062 case Bytecodes::_multianewarray: 1063 { 1064 ciBytecodeStream iter(method()); 1065 iter.reset_to_bci(bci()); 1066 iter.next(); 1067 inputs = iter.get_dimensions(); 1068 assert(rsize == 1, ""); 1069 depth = rsize - inputs; 1070 } 1071 break; 1072 1073 case Bytecodes::_ireturn: 1074 case Bytecodes::_lreturn: 1075 case Bytecodes::_freturn: 1076 case Bytecodes::_dreturn: 1077 case Bytecodes::_areturn: 1078 assert(rsize = -depth, ""); 1079 inputs = rsize; 1080 break; 1081 1082 case Bytecodes::_jsr: 1083 case Bytecodes::_jsr_w: 1084 inputs = 0; 1085 depth = 1; // S.B. depth=1, not zero 1086 break; 1087 1088 default: 1089 // bytecode produces a typed result 1090 inputs = rsize - depth; 1091 assert(inputs >= 0, ""); 1092 break; 1093 } 1094 1095 #ifdef ASSERT 1096 // spot check 1097 int outputs = depth + inputs; 1098 assert(outputs >= 0, "sanity"); 1099 switch (code) { 1100 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1101 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1102 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1103 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1104 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1105 } 1106 #endif //ASSERT 1107 1108 return true; 1109 } 1110 1111 1112 1113 //------------------------------basic_plus_adr--------------------------------- 1114 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1115 // short-circuit a common case 1116 if (offset == intcon(0)) return ptr; 1117 return _gvn.transform( new (C) AddPNode(base, ptr, offset) ); 1118 } 1119 1120 Node* GraphKit::ConvI2L(Node* offset) { 1121 // short-circuit a common case 1122 jint offset_con = find_int_con(offset, Type::OffsetBot); 1123 if (offset_con != Type::OffsetBot) { 1124 return longcon((jlong) offset_con); 1125 } 1126 return _gvn.transform( new (C) ConvI2LNode(offset)); 1127 } 1128 1129 Node* GraphKit::ConvI2UL(Node* offset) { 1130 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1131 if (offset_con != (juint) Type::OffsetBot) { 1132 return longcon((julong) offset_con); 1133 } 1134 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset)); 1135 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) ); 1136 return _gvn.transform( new (C) AndLNode(conv, mask) ); 1137 } 1138 1139 Node* GraphKit::ConvL2I(Node* offset) { 1140 // short-circuit a common case 1141 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1142 if (offset_con != (jlong)Type::OffsetBot) { 1143 return intcon((int) offset_con); 1144 } 1145 return _gvn.transform( new (C) ConvL2INode(offset)); 1146 } 1147 1148 //-------------------------load_object_klass----------------------------------- 1149 Node* GraphKit::load_object_klass(Node* obj) { 1150 // Special-case a fresh allocation to avoid building nodes: 1151 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1152 if (akls != NULL) return akls; 1153 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1154 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) ); 1155 } 1156 1157 //-------------------------load_array_length----------------------------------- 1158 Node* GraphKit::load_array_length(Node* array) { 1159 // Special-case a fresh allocation to avoid building nodes: 1160 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1161 Node *alen; 1162 if (alloc == NULL) { 1163 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1164 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1165 } else { 1166 alen = alloc->Ideal_length(); 1167 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1168 if (ccast != alen) { 1169 alen = _gvn.transform(ccast); 1170 } 1171 } 1172 return alen; 1173 } 1174 1175 //------------------------------do_null_check---------------------------------- 1176 // Helper function to do a NULL pointer check. Returned value is 1177 // the incoming address with NULL casted away. You are allowed to use the 1178 // not-null value only if you are control dependent on the test. 1179 extern int explicit_null_checks_inserted, 1180 explicit_null_checks_elided; 1181 Node* GraphKit::null_check_common(Node* value, BasicType type, 1182 // optional arguments for variations: 1183 bool assert_null, 1184 Node* *null_control, 1185 bool speculative) { 1186 assert(!assert_null || null_control == NULL, "not both at once"); 1187 if (stopped()) return top(); 1188 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) { 1189 // For some performance testing, we may wish to suppress null checking. 1190 value = cast_not_null(value); // Make it appear to be non-null (4962416). 1191 return value; 1192 } 1193 explicit_null_checks_inserted++; 1194 1195 // Construct NULL check 1196 Node *chk = NULL; 1197 switch(type) { 1198 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1199 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break; 1200 case T_ARRAY : // fall through 1201 type = T_OBJECT; // simplify further tests 1202 case T_OBJECT : { 1203 const Type *t = _gvn.type( value ); 1204 1205 const TypeOopPtr* tp = t->isa_oopptr(); 1206 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1207 // Only for do_null_check, not any of its siblings: 1208 && !assert_null && null_control == NULL) { 1209 // Usually, any field access or invocation on an unloaded oop type 1210 // will simply fail to link, since the statically linked class is 1211 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1212 // the static class is loaded but the sharper oop type is not. 1213 // Rather than checking for this obscure case in lots of places, 1214 // we simply observe that a null check on an unloaded class 1215 // will always be followed by a nonsense operation, so we 1216 // can just issue the uncommon trap here. 1217 // Our access to the unloaded class will only be correct 1218 // after it has been loaded and initialized, which requires 1219 // a trip through the interpreter. 1220 #ifndef PRODUCT 1221 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1222 #endif 1223 uncommon_trap(Deoptimization::Reason_unloaded, 1224 Deoptimization::Action_reinterpret, 1225 tp->klass(), "!loaded"); 1226 return top(); 1227 } 1228 1229 if (assert_null) { 1230 // See if the type is contained in NULL_PTR. 1231 // If so, then the value is already null. 1232 if (t->higher_equal(TypePtr::NULL_PTR)) { 1233 explicit_null_checks_elided++; 1234 return value; // Elided null assert quickly! 1235 } 1236 } else { 1237 // See if mixing in the NULL pointer changes type. 1238 // If so, then the NULL pointer was not allowed in the original 1239 // type. In other words, "value" was not-null. 1240 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1241 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1242 explicit_null_checks_elided++; 1243 return value; // Elided null check quickly! 1244 } 1245 } 1246 chk = new (C) CmpPNode( value, null() ); 1247 break; 1248 } 1249 1250 default: 1251 fatal(err_msg_res("unexpected type: %s", type2name(type))); 1252 } 1253 assert(chk != NULL, "sanity check"); 1254 chk = _gvn.transform(chk); 1255 1256 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1257 BoolNode *btst = new (C) BoolNode( chk, btest); 1258 Node *tst = _gvn.transform( btst ); 1259 1260 //----------- 1261 // if peephole optimizations occurred, a prior test existed. 1262 // If a prior test existed, maybe it dominates as we can avoid this test. 1263 if (tst != btst && type == T_OBJECT) { 1264 // At this point we want to scan up the CFG to see if we can 1265 // find an identical test (and so avoid this test altogether). 1266 Node *cfg = control(); 1267 int depth = 0; 1268 while( depth < 16 ) { // Limit search depth for speed 1269 if( cfg->Opcode() == Op_IfTrue && 1270 cfg->in(0)->in(1) == tst ) { 1271 // Found prior test. Use "cast_not_null" to construct an identical 1272 // CastPP (and hence hash to) as already exists for the prior test. 1273 // Return that casted value. 1274 if (assert_null) { 1275 replace_in_map(value, null()); 1276 return null(); // do not issue the redundant test 1277 } 1278 Node *oldcontrol = control(); 1279 set_control(cfg); 1280 Node *res = cast_not_null(value); 1281 set_control(oldcontrol); 1282 explicit_null_checks_elided++; 1283 return res; 1284 } 1285 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1286 if (cfg == NULL) break; // Quit at region nodes 1287 depth++; 1288 } 1289 } 1290 1291 //----------- 1292 // Branch to failure if null 1293 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1294 Deoptimization::DeoptReason reason; 1295 if (assert_null) { 1296 reason = Deoptimization::Reason_null_assert; 1297 } else if (type == T_OBJECT) { 1298 reason = Deoptimization::reason_null_check(speculative); 1299 } else { 1300 reason = Deoptimization::Reason_div0_check; 1301 } 1302 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1303 // ciMethodData::has_trap_at will return a conservative -1 if any 1304 // must-be-null assertion has failed. This could cause performance 1305 // problems for a method after its first do_null_assert failure. 1306 // Consider using 'Reason_class_check' instead? 1307 1308 // To cause an implicit null check, we set the not-null probability 1309 // to the maximum (PROB_MAX). For an explicit check the probability 1310 // is set to a smaller value. 1311 if (null_control != NULL || too_many_traps(reason)) { 1312 // probability is less likely 1313 ok_prob = PROB_LIKELY_MAG(3); 1314 } else if (!assert_null && 1315 (ImplicitNullCheckThreshold > 0) && 1316 method() != NULL && 1317 (method()->method_data()->trap_count(reason) 1318 >= (uint)ImplicitNullCheckThreshold)) { 1319 ok_prob = PROB_LIKELY_MAG(3); 1320 } 1321 1322 if (null_control != NULL) { 1323 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1324 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff)); 1325 set_control( _gvn.transform( new (C) IfTrueNode(iff))); 1326 if (null_true == top()) 1327 explicit_null_checks_elided++; 1328 (*null_control) = null_true; 1329 } else { 1330 BuildCutout unless(this, tst, ok_prob); 1331 // Check for optimizer eliding test at parse time 1332 if (stopped()) { 1333 // Failure not possible; do not bother making uncommon trap. 1334 explicit_null_checks_elided++; 1335 } else if (assert_null) { 1336 uncommon_trap(reason, 1337 Deoptimization::Action_make_not_entrant, 1338 NULL, "assert_null"); 1339 } else { 1340 replace_in_map(value, zerocon(type)); 1341 builtin_throw(reason); 1342 } 1343 } 1344 1345 // Must throw exception, fall-thru not possible? 1346 if (stopped()) { 1347 return top(); // No result 1348 } 1349 1350 if (assert_null) { 1351 // Cast obj to null on this path. 1352 replace_in_map(value, zerocon(type)); 1353 return zerocon(type); 1354 } 1355 1356 // Cast obj to not-null on this path, if there is no null_control. 1357 // (If there is a null_control, a non-null value may come back to haunt us.) 1358 if (type == T_OBJECT) { 1359 Node* cast = cast_not_null(value, false); 1360 if (null_control == NULL || (*null_control) == top()) 1361 replace_in_map(value, cast); 1362 value = cast; 1363 } 1364 1365 return value; 1366 } 1367 1368 1369 //------------------------------cast_not_null---------------------------------- 1370 // Cast obj to not-null on this path 1371 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1372 const Type *t = _gvn.type(obj); 1373 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1374 // Object is already not-null? 1375 if( t == t_not_null ) return obj; 1376 1377 Node *cast = new (C) CastPPNode(obj,t_not_null); 1378 cast->init_req(0, control()); 1379 cast = _gvn.transform( cast ); 1380 1381 // Scan for instances of 'obj' in the current JVM mapping. 1382 // These instances are known to be not-null after the test. 1383 if (do_replace_in_map) 1384 replace_in_map(obj, cast); 1385 1386 return cast; // Return casted value 1387 } 1388 1389 1390 //--------------------------replace_in_map------------------------------------- 1391 void GraphKit::replace_in_map(Node* old, Node* neww) { 1392 if (old == neww) { 1393 return; 1394 } 1395 1396 map()->replace_edge(old, neww); 1397 1398 // Note: This operation potentially replaces any edge 1399 // on the map. This includes locals, stack, and monitors 1400 // of the current (innermost) JVM state. 1401 1402 if (!ReplaceInParentMaps) { 1403 return; 1404 } 1405 1406 // PreserveJVMState doesn't do a deep copy so we can't modify 1407 // parents 1408 if (Compile::current()->has_preserve_jvm_state()) { 1409 return; 1410 } 1411 1412 Parse* parser = is_Parse(); 1413 bool progress = true; 1414 Node* ctrl = map()->in(0); 1415 // Follow the chain of parsers and see whether the update can be 1416 // done in the map of callers. We can do the replace for a caller if 1417 // the current control post dominates the control of a caller. 1418 while (parser != NULL && parser->caller() != NULL && progress) { 1419 progress = false; 1420 Node* parent_map = parser->caller()->map(); 1421 assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch"); 1422 1423 Node* parent_ctrl = parent_map->in(0); 1424 1425 while (parent_ctrl->is_Region()) { 1426 Node* n = parent_ctrl->as_Region()->is_copy(); 1427 if (n == NULL) { 1428 break; 1429 } 1430 parent_ctrl = n; 1431 } 1432 1433 for (;;) { 1434 if (ctrl == parent_ctrl) { 1435 // update the map of the exits which is the one that will be 1436 // used when compilation resume after inlining 1437 parser->exits().map()->replace_edge(old, neww); 1438 progress = true; 1439 break; 1440 } 1441 if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 1442 ctrl = ctrl->in(0)->in(0); 1443 } else if (ctrl->is_Region()) { 1444 Node* n = ctrl->as_Region()->is_copy(); 1445 if (n == NULL) { 1446 break; 1447 } 1448 ctrl = n; 1449 } else { 1450 break; 1451 } 1452 } 1453 1454 parser = parser->parent_parser(); 1455 } 1456 } 1457 1458 1459 //============================================================================= 1460 //--------------------------------memory--------------------------------------- 1461 Node* GraphKit::memory(uint alias_idx) { 1462 MergeMemNode* mem = merged_memory(); 1463 Node* p = mem->memory_at(alias_idx); 1464 _gvn.set_type(p, Type::MEMORY); // must be mapped 1465 return p; 1466 } 1467 1468 //-----------------------------reset_memory------------------------------------ 1469 Node* GraphKit::reset_memory() { 1470 Node* mem = map()->memory(); 1471 // do not use this node for any more parsing! 1472 debug_only( map()->set_memory((Node*)NULL) ); 1473 return _gvn.transform( mem ); 1474 } 1475 1476 //------------------------------set_all_memory--------------------------------- 1477 void GraphKit::set_all_memory(Node* newmem) { 1478 Node* mergemem = MergeMemNode::make(C, newmem); 1479 gvn().set_type_bottom(mergemem); 1480 map()->set_memory(mergemem); 1481 } 1482 1483 //------------------------------set_all_memory_call---------------------------- 1484 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1485 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1486 set_all_memory(newmem); 1487 } 1488 1489 //============================================================================= 1490 // 1491 // parser factory methods for MemNodes 1492 // 1493 // These are layered on top of the factory methods in LoadNode and StoreNode, 1494 // and integrate with the parser's memory state and _gvn engine. 1495 // 1496 1497 // factory methods in "int adr_idx" 1498 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1499 int adr_idx, 1500 MemNode::MemOrd mo, bool require_atomic_access) { 1501 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1502 const TypePtr* adr_type = NULL; // debug-mode-only argument 1503 debug_only(adr_type = C->get_adr_type(adr_idx)); 1504 Node* mem = memory(adr_idx); 1505 Node* ld; 1506 if (require_atomic_access && bt == T_LONG) { 1507 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo); 1508 } else { 1509 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo); 1510 } 1511 ld = _gvn.transform(ld); 1512 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1513 // Improve graph before escape analysis and boxing elimination. 1514 record_for_igvn(ld); 1515 } 1516 return ld; 1517 } 1518 1519 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1520 int adr_idx, 1521 MemNode::MemOrd mo, 1522 bool require_atomic_access) { 1523 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1524 const TypePtr* adr_type = NULL; 1525 debug_only(adr_type = C->get_adr_type(adr_idx)); 1526 Node *mem = memory(adr_idx); 1527 Node* st; 1528 if (require_atomic_access && bt == T_LONG) { 1529 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo); 1530 } else { 1531 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1532 } 1533 st = _gvn.transform(st); 1534 set_memory(st, adr_idx); 1535 // Back-to-back stores can only remove intermediate store with DU info 1536 // so push on worklist for optimizer. 1537 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1538 record_for_igvn(st); 1539 1540 return st; 1541 } 1542 1543 1544 void GraphKit::pre_barrier(bool do_load, 1545 Node* ctl, 1546 Node* obj, 1547 Node* adr, 1548 uint adr_idx, 1549 Node* val, 1550 const TypeOopPtr* val_type, 1551 Node* pre_val, 1552 BasicType bt) { 1553 1554 BarrierSet* bs = Universe::heap()->barrier_set(); 1555 set_control(ctl); 1556 switch (bs->kind()) { 1557 case BarrierSet::G1SATBCT: 1558 case BarrierSet::G1SATBCTLogging: 1559 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt); 1560 break; 1561 1562 case BarrierSet::CardTableModRef: 1563 case BarrierSet::CardTableExtension: 1564 case BarrierSet::ModRef: 1565 break; 1566 1567 case BarrierSet::Other: 1568 default : 1569 ShouldNotReachHere(); 1570 1571 } 1572 } 1573 1574 bool GraphKit::can_move_pre_barrier() const { 1575 BarrierSet* bs = Universe::heap()->barrier_set(); 1576 switch (bs->kind()) { 1577 case BarrierSet::G1SATBCT: 1578 case BarrierSet::G1SATBCTLogging: 1579 return true; // Can move it if no safepoint 1580 1581 case BarrierSet::CardTableModRef: 1582 case BarrierSet::CardTableExtension: 1583 case BarrierSet::ModRef: 1584 return true; // There is no pre-barrier 1585 1586 case BarrierSet::Other: 1587 default : 1588 ShouldNotReachHere(); 1589 } 1590 return false; 1591 } 1592 1593 void GraphKit::post_barrier(Node* ctl, 1594 Node* store, 1595 Node* obj, 1596 Node* adr, 1597 uint adr_idx, 1598 Node* val, 1599 BasicType bt, 1600 bool use_precise) { 1601 BarrierSet* bs = Universe::heap()->barrier_set(); 1602 set_control(ctl); 1603 switch (bs->kind()) { 1604 case BarrierSet::G1SATBCT: 1605 case BarrierSet::G1SATBCTLogging: 1606 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise); 1607 break; 1608 1609 case BarrierSet::CardTableModRef: 1610 case BarrierSet::CardTableExtension: 1611 write_barrier_post(store, obj, adr, adr_idx, val, use_precise); 1612 break; 1613 1614 case BarrierSet::ModRef: 1615 break; 1616 1617 case BarrierSet::Other: 1618 default : 1619 ShouldNotReachHere(); 1620 1621 } 1622 } 1623 1624 Node* GraphKit::store_oop(Node* ctl, 1625 Node* obj, 1626 Node* adr, 1627 const TypePtr* adr_type, 1628 Node* val, 1629 const TypeOopPtr* val_type, 1630 BasicType bt, 1631 bool use_precise, 1632 MemNode::MemOrd mo) { 1633 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1634 // could be delayed during Parse (for example, in adjust_map_after_if()). 1635 // Execute transformation here to avoid barrier generation in such case. 1636 if (_gvn.type(val) == TypePtr::NULL_PTR) 1637 val = _gvn.makecon(TypePtr::NULL_PTR); 1638 1639 set_control(ctl); 1640 if (stopped()) return top(); // Dead path ? 1641 1642 assert(bt == T_OBJECT, "sanity"); 1643 assert(val != NULL, "not dead path"); 1644 uint adr_idx = C->get_alias_index(adr_type); 1645 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1646 1647 pre_barrier(true /* do_load */, 1648 control(), obj, adr, adr_idx, val, val_type, 1649 NULL /* pre_val */, 1650 bt); 1651 1652 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo); 1653 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise); 1654 return store; 1655 } 1656 1657 // Could be an array or object we don't know at compile time (unsafe ref.) 1658 Node* GraphKit::store_oop_to_unknown(Node* ctl, 1659 Node* obj, // containing obj 1660 Node* adr, // actual adress to store val at 1661 const TypePtr* adr_type, 1662 Node* val, 1663 BasicType bt, 1664 MemNode::MemOrd mo) { 1665 Compile::AliasType* at = C->alias_type(adr_type); 1666 const TypeOopPtr* val_type = NULL; 1667 if (adr_type->isa_instptr()) { 1668 if (at->field() != NULL) { 1669 // known field. This code is a copy of the do_put_xxx logic. 1670 ciField* field = at->field(); 1671 if (!field->type()->is_loaded()) { 1672 val_type = TypeInstPtr::BOTTOM; 1673 } else { 1674 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass()); 1675 } 1676 } 1677 } else if (adr_type->isa_aryptr()) { 1678 val_type = adr_type->is_aryptr()->elem()->make_oopptr(); 1679 } 1680 if (val_type == NULL) { 1681 val_type = TypeInstPtr::BOTTOM; 1682 } 1683 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo); 1684 } 1685 1686 1687 //-------------------------array_element_address------------------------- 1688 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1689 const TypeInt* sizetype) { 1690 uint shift = exact_log2(type2aelembytes(elembt)); 1691 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1692 1693 // short-circuit a common case (saves lots of confusing waste motion) 1694 jint idx_con = find_int_con(idx, -1); 1695 if (idx_con >= 0) { 1696 intptr_t offset = header + ((intptr_t)idx_con << shift); 1697 return basic_plus_adr(ary, offset); 1698 } 1699 1700 // must be correct type for alignment purposes 1701 Node* base = basic_plus_adr(ary, header); 1702 #ifdef _LP64 1703 // The scaled index operand to AddP must be a clean 64-bit value. 1704 // Java allows a 32-bit int to be incremented to a negative 1705 // value, which appears in a 64-bit register as a large 1706 // positive number. Using that large positive number as an 1707 // operand in pointer arithmetic has bad consequences. 1708 // On the other hand, 32-bit overflow is rare, and the possibility 1709 // can often be excluded, if we annotate the ConvI2L node with 1710 // a type assertion that its value is known to be a small positive 1711 // number. (The prior range check has ensured this.) 1712 // This assertion is used by ConvI2LNode::Ideal. 1713 int index_max = max_jint - 1; // array size is max_jint, index is one less 1714 if (sizetype != NULL) index_max = sizetype->_hi - 1; 1715 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 1716 idx = _gvn.transform( new (C) ConvI2LNode(idx, lidxtype) ); 1717 #endif 1718 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) ); 1719 return basic_plus_adr(ary, base, scale); 1720 } 1721 1722 //-------------------------load_array_element------------------------- 1723 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1724 const Type* elemtype = arytype->elem(); 1725 BasicType elembt = elemtype->array_element_basic_type(); 1726 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1727 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1728 return ld; 1729 } 1730 1731 //-------------------------set_arguments_for_java_call------------------------- 1732 // Arguments (pre-popped from the stack) are taken from the JVMS. 1733 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1734 // Add the call arguments: 1735 uint nargs = call->method()->arg_size(); 1736 for (uint i = 0; i < nargs; i++) { 1737 Node* arg = argument(i); 1738 call->init_req(i + TypeFunc::Parms, arg); 1739 } 1740 } 1741 1742 //---------------------------set_edges_for_java_call--------------------------- 1743 // Connect a newly created call into the current JVMS. 1744 // A return value node (if any) is returned from set_edges_for_java_call. 1745 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1746 1747 // Add the predefined inputs: 1748 call->init_req( TypeFunc::Control, control() ); 1749 call->init_req( TypeFunc::I_O , i_o() ); 1750 call->init_req( TypeFunc::Memory , reset_memory() ); 1751 call->init_req( TypeFunc::FramePtr, frameptr() ); 1752 call->init_req( TypeFunc::ReturnAdr, top() ); 1753 1754 add_safepoint_edges(call, must_throw); 1755 1756 Node* xcall = _gvn.transform(call); 1757 1758 if (xcall == top()) { 1759 set_control(top()); 1760 return; 1761 } 1762 assert(xcall == call, "call identity is stable"); 1763 1764 // Re-use the current map to produce the result. 1765 1766 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control))); 1767 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1768 set_all_memory_call(xcall, separate_io_proj); 1769 1770 //return xcall; // no need, caller already has it 1771 } 1772 1773 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1774 if (stopped()) return top(); // maybe the call folded up? 1775 1776 // Capture the return value, if any. 1777 Node* ret; 1778 if (call->method() == NULL || 1779 call->method()->return_type()->basic_type() == T_VOID) 1780 ret = top(); 1781 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1782 1783 // Note: Since any out-of-line call can produce an exception, 1784 // we always insert an I_O projection from the call into the result. 1785 1786 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1787 1788 if (separate_io_proj) { 1789 // The caller requested separate projections be used by the fall 1790 // through and exceptional paths, so replace the projections for 1791 // the fall through path. 1792 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) )); 1793 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) )); 1794 } 1795 return ret; 1796 } 1797 1798 //--------------------set_predefined_input_for_runtime_call-------------------- 1799 // Reading and setting the memory state is way conservative here. 1800 // The real problem is that I am not doing real Type analysis on memory, 1801 // so I cannot distinguish card mark stores from other stores. Across a GC 1802 // point the Store Barrier and the card mark memory has to agree. I cannot 1803 // have a card mark store and its barrier split across the GC point from 1804 // either above or below. Here I get that to happen by reading ALL of memory. 1805 // A better answer would be to separate out card marks from other memory. 1806 // For now, return the input memory state, so that it can be reused 1807 // after the call, if this call has restricted memory effects. 1808 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1809 // Set fixed predefined input arguments 1810 Node* memory = reset_memory(); 1811 call->init_req( TypeFunc::Control, control() ); 1812 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1813 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1814 call->init_req( TypeFunc::FramePtr, frameptr() ); 1815 call->init_req( TypeFunc::ReturnAdr, top() ); 1816 return memory; 1817 } 1818 1819 //-------------------set_predefined_output_for_runtime_call-------------------- 1820 // Set control and memory (not i_o) from the call. 1821 // If keep_mem is not NULL, use it for the output state, 1822 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1823 // If hook_mem is NULL, this call produces no memory effects at all. 1824 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1825 // then only that memory slice is taken from the call. 1826 // In the last case, we must put an appropriate memory barrier before 1827 // the call, so as to create the correct anti-dependencies on loads 1828 // preceding the call. 1829 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1830 Node* keep_mem, 1831 const TypePtr* hook_mem) { 1832 // no i/o 1833 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) )); 1834 if (keep_mem) { 1835 // First clone the existing memory state 1836 set_all_memory(keep_mem); 1837 if (hook_mem != NULL) { 1838 // Make memory for the call 1839 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ); 1840 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1841 // We also use hook_mem to extract specific effects from arraycopy stubs. 1842 set_memory(mem, hook_mem); 1843 } 1844 // ...else the call has NO memory effects. 1845 1846 // Make sure the call advertises its memory effects precisely. 1847 // This lets us build accurate anti-dependences in gcm.cpp. 1848 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1849 "call node must be constructed correctly"); 1850 } else { 1851 assert(hook_mem == NULL, ""); 1852 // This is not a "slow path" call; all memory comes from the call. 1853 set_all_memory_call(call); 1854 } 1855 } 1856 1857 1858 // Replace the call with the current state of the kit. 1859 void GraphKit::replace_call(CallNode* call, Node* result) { 1860 JVMState* ejvms = NULL; 1861 if (has_exceptions()) { 1862 ejvms = transfer_exceptions_into_jvms(); 1863 } 1864 1865 SafePointNode* final_state = stop(); 1866 1867 // Find all the needed outputs of this call 1868 CallProjections callprojs; 1869 call->extract_projections(&callprojs, true); 1870 1871 Node* init_mem = call->in(TypeFunc::Memory); 1872 Node* final_mem = final_state->in(TypeFunc::Memory); 1873 Node* final_ctl = final_state->in(TypeFunc::Control); 1874 Node* final_io = final_state->in(TypeFunc::I_O); 1875 1876 // Replace all the old call edges with the edges from the inlining result 1877 if (callprojs.fallthrough_catchproj != NULL) { 1878 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1879 } 1880 if (callprojs.fallthrough_memproj != NULL) { 1881 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1882 } 1883 if (callprojs.fallthrough_ioproj != NULL) { 1884 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1885 } 1886 1887 // Replace the result with the new result if it exists and is used 1888 if (callprojs.resproj != NULL && result != NULL) { 1889 C->gvn_replace_by(callprojs.resproj, result); 1890 } 1891 1892 if (ejvms == NULL) { 1893 // No exception edges to simply kill off those paths 1894 if (callprojs.catchall_catchproj != NULL) { 1895 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1896 } 1897 if (callprojs.catchall_memproj != NULL) { 1898 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1899 } 1900 if (callprojs.catchall_ioproj != NULL) { 1901 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1902 } 1903 // Replace the old exception object with top 1904 if (callprojs.exobj != NULL) { 1905 C->gvn_replace_by(callprojs.exobj, C->top()); 1906 } 1907 } else { 1908 GraphKit ekit(ejvms); 1909 1910 // Load my combined exception state into the kit, with all phis transformed: 1911 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1912 1913 Node* ex_oop = ekit.use_exception_state(ex_map); 1914 if (callprojs.catchall_catchproj != NULL) { 1915 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1916 } 1917 if (callprojs.catchall_memproj != NULL) { 1918 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1919 } 1920 if (callprojs.catchall_ioproj != NULL) { 1921 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1922 } 1923 1924 // Replace the old exception object with the newly created one 1925 if (callprojs.exobj != NULL) { 1926 C->gvn_replace_by(callprojs.exobj, ex_oop); 1927 } 1928 } 1929 1930 // Disconnect the call from the graph 1931 call->disconnect_inputs(NULL, C); 1932 C->gvn_replace_by(call, C->top()); 1933 1934 // Clean up any MergeMems that feed other MergeMems since the 1935 // optimizer doesn't like that. 1936 if (final_mem->is_MergeMem()) { 1937 Node_List wl; 1938 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1939 Node* m = i.get(); 1940 if (m->is_MergeMem() && !wl.contains(m)) { 1941 wl.push(m); 1942 } 1943 } 1944 while (wl.size() > 0) { 1945 _gvn.transform(wl.pop()); 1946 } 1947 } 1948 } 1949 1950 1951 //------------------------------increment_counter------------------------------ 1952 // for statistics: increment a VM counter by 1 1953 1954 void GraphKit::increment_counter(address counter_addr) { 1955 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1956 increment_counter(adr1); 1957 } 1958 1959 void GraphKit::increment_counter(Node* counter_addr) { 1960 int adr_type = Compile::AliasIdxRaw; 1961 Node* ctrl = control(); 1962 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1963 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1))); 1964 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1965 } 1966 1967 1968 //------------------------------uncommon_trap---------------------------------- 1969 // Bail out to the interpreter in mid-method. Implemented by calling the 1970 // uncommon_trap blob. This helper function inserts a runtime call with the 1971 // right debug info. 1972 void GraphKit::uncommon_trap(int trap_request, 1973 ciKlass* klass, const char* comment, 1974 bool must_throw, 1975 bool keep_exact_action) { 1976 if (failing()) stop(); 1977 if (stopped()) return; // trap reachable? 1978 1979 // Note: If ProfileTraps is true, and if a deopt. actually 1980 // occurs here, the runtime will make sure an MDO exists. There is 1981 // no need to call method()->ensure_method_data() at this point. 1982 1983 // Set the stack pointer to the right value for reexecution: 1984 set_sp(reexecute_sp()); 1985 1986 #ifdef ASSERT 1987 if (!must_throw) { 1988 // Make sure the stack has at least enough depth to execute 1989 // the current bytecode. 1990 int inputs, ignored_depth; 1991 if (compute_stack_effects(inputs, ignored_depth)) { 1992 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1993 Bytecodes::name(java_bc()), sp(), inputs)); 1994 } 1995 } 1996 #endif 1997 1998 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1999 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2000 2001 switch (action) { 2002 case Deoptimization::Action_maybe_recompile: 2003 case Deoptimization::Action_reinterpret: 2004 // Temporary fix for 6529811 to allow virtual calls to be sure they 2005 // get the chance to go from mono->bi->mega 2006 if (!keep_exact_action && 2007 Deoptimization::trap_request_index(trap_request) < 0 && 2008 too_many_recompiles(reason)) { 2009 // This BCI is causing too many recompilations. 2010 action = Deoptimization::Action_none; 2011 trap_request = Deoptimization::make_trap_request(reason, action); 2012 } else { 2013 C->set_trap_can_recompile(true); 2014 } 2015 break; 2016 case Deoptimization::Action_make_not_entrant: 2017 C->set_trap_can_recompile(true); 2018 break; 2019 #ifdef ASSERT 2020 case Deoptimization::Action_none: 2021 case Deoptimization::Action_make_not_compilable: 2022 break; 2023 default: 2024 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action))); 2025 break; 2026 #endif 2027 } 2028 2029 if (TraceOptoParse) { 2030 char buf[100]; 2031 tty->print_cr("Uncommon trap %s at bci:%d", 2032 Deoptimization::format_trap_request(buf, sizeof(buf), 2033 trap_request), bci()); 2034 } 2035 2036 CompileLog* log = C->log(); 2037 if (log != NULL) { 2038 int kid = (klass == NULL)? -1: log->identify(klass); 2039 log->begin_elem("uncommon_trap bci='%d'", bci()); 2040 char buf[100]; 2041 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2042 trap_request)); 2043 if (kid >= 0) log->print(" klass='%d'", kid); 2044 if (comment != NULL) log->print(" comment='%s'", comment); 2045 log->end_elem(); 2046 } 2047 2048 // Make sure any guarding test views this path as very unlikely 2049 Node *i0 = control()->in(0); 2050 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2051 IfNode *iff = i0->as_If(); 2052 float f = iff->_prob; // Get prob 2053 if (control()->Opcode() == Op_IfTrue) { 2054 if (f > PROB_UNLIKELY_MAG(4)) 2055 iff->_prob = PROB_MIN; 2056 } else { 2057 if (f < PROB_LIKELY_MAG(4)) 2058 iff->_prob = PROB_MAX; 2059 } 2060 } 2061 2062 // Clear out dead values from the debug info. 2063 kill_dead_locals(); 2064 2065 // Now insert the uncommon trap subroutine call 2066 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2067 const TypePtr* no_memory_effects = NULL; 2068 // Pass the index of the class to be loaded 2069 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2070 (must_throw ? RC_MUST_THROW : 0), 2071 OptoRuntime::uncommon_trap_Type(), 2072 call_addr, "uncommon_trap", no_memory_effects, 2073 intcon(trap_request)); 2074 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2075 "must extract request correctly from the graph"); 2076 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2077 2078 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2079 // The debug info is the only real input to this call. 2080 2081 // Halt-and-catch fire here. The above call should never return! 2082 HaltNode* halt = new(C) HaltNode(control(), frameptr()); 2083 _gvn.set_type_bottom(halt); 2084 root()->add_req(halt); 2085 2086 stop_and_kill_map(); 2087 } 2088 2089 2090 //--------------------------just_allocated_object------------------------------ 2091 // Report the object that was just allocated. 2092 // It must be the case that there are no intervening safepoints. 2093 // We use this to determine if an object is so "fresh" that 2094 // it does not require card marks. 2095 Node* GraphKit::just_allocated_object(Node* current_control) { 2096 if (C->recent_alloc_ctl() == current_control) 2097 return C->recent_alloc_obj(); 2098 return NULL; 2099 } 2100 2101 2102 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2103 // (Note: TypeFunc::make has a cache that makes this fast.) 2104 const TypeFunc* tf = TypeFunc::make(dest_method); 2105 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2106 for (int j = 0; j < nargs; j++) { 2107 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2108 if( targ->basic_type() == T_DOUBLE ) { 2109 // If any parameters are doubles, they must be rounded before 2110 // the call, dstore_rounding does gvn.transform 2111 Node *arg = argument(j); 2112 arg = dstore_rounding(arg); 2113 set_argument(j, arg); 2114 } 2115 } 2116 } 2117 2118 /** 2119 * Record profiling data exact_kls for Node n with the type system so 2120 * that it can propagate it (speculation) 2121 * 2122 * @param n node that the type applies to 2123 * @param exact_kls type from profiling 2124 * @param maybe_null did profiling see null? 2125 * 2126 * @return node with improved type 2127 */ 2128 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2129 const Type* current_type = _gvn.type(n); 2130 assert(UseTypeSpeculation, "type speculation must be on"); 2131 2132 const TypePtr* speculative = current_type->speculative(); 2133 2134 // Should the klass from the profile be recorded in the speculative type? 2135 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2136 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2137 const TypeOopPtr* xtype = tklass->as_instance_type(); 2138 assert(xtype->klass_is_exact(), "Should be exact"); 2139 // Any reason to believe n is not null (from this profiling or a previous one)? 2140 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2141 // record the new speculative type's depth 2142 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2143 speculative = speculative->with_inline_depth(jvms()->depth()); 2144 } else if (current_type->would_improve_ptr(maybe_null)) { 2145 // Profiling report that null was never seen so we can change the 2146 // speculative type to non null ptr. 2147 assert(!maybe_null, "nothing to improve"); 2148 if (speculative == NULL) { 2149 speculative = TypePtr::NOTNULL; 2150 } else { 2151 const TypePtr* ptr = TypePtr::NOTNULL; 2152 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2153 } 2154 } 2155 2156 if (speculative != current_type->speculative()) { 2157 // Build a type with a speculative type (what we think we know 2158 // about the type but will need a guard when we use it) 2159 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2160 // We're changing the type, we need a new CheckCast node to carry 2161 // the new type. The new type depends on the control: what 2162 // profiling tells us is only valid from here as far as we can 2163 // tell. 2164 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2165 cast = _gvn.transform(cast); 2166 replace_in_map(n, cast); 2167 n = cast; 2168 } 2169 2170 return n; 2171 } 2172 2173 /** 2174 * Record profiling data from receiver profiling at an invoke with the 2175 * type system so that it can propagate it (speculation) 2176 * 2177 * @param n receiver node 2178 * 2179 * @return node with improved type 2180 */ 2181 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2182 if (!UseTypeSpeculation) { 2183 return n; 2184 } 2185 ciKlass* exact_kls = profile_has_unique_klass(); 2186 bool maybe_null = true; 2187 if (java_bc() == Bytecodes::_checkcast || 2188 java_bc() == Bytecodes::_instanceof || 2189 java_bc() == Bytecodes::_aastore) { 2190 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2191 bool maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2192 } 2193 return record_profile_for_speculation(n, exact_kls, maybe_null); 2194 return n; 2195 } 2196 2197 /** 2198 * Record profiling data from argument profiling at an invoke with the 2199 * type system so that it can propagate it (speculation) 2200 * 2201 * @param dest_method target method for the call 2202 * @param bc what invoke bytecode is this? 2203 */ 2204 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2205 if (!UseTypeSpeculation) { 2206 return; 2207 } 2208 const TypeFunc* tf = TypeFunc::make(dest_method); 2209 int nargs = tf->_domain->_cnt - TypeFunc::Parms; 2210 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2211 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2212 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms); 2213 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2214 bool maybe_null = true; 2215 ciKlass* better_type = NULL; 2216 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2217 record_profile_for_speculation(argument(j), better_type, maybe_null); 2218 } 2219 i++; 2220 } 2221 } 2222 } 2223 2224 /** 2225 * Record profiling data from parameter profiling at an invoke with 2226 * the type system so that it can propagate it (speculation) 2227 */ 2228 void GraphKit::record_profiled_parameters_for_speculation() { 2229 if (!UseTypeSpeculation) { 2230 return; 2231 } 2232 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2233 if (_gvn.type(local(i))->isa_oopptr()) { 2234 bool maybe_null = true; 2235 ciKlass* better_type = NULL; 2236 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2237 record_profile_for_speculation(local(i), better_type, maybe_null); 2238 } 2239 j++; 2240 } 2241 } 2242 } 2243 2244 /** 2245 * Record profiling data from return value profiling at an invoke with 2246 * the type system so that it can propagate it (speculation) 2247 */ 2248 void GraphKit::record_profiled_return_for_speculation() { 2249 if (!UseTypeSpeculation) { 2250 return; 2251 } 2252 bool maybe_null = true; 2253 ciKlass* better_type = NULL; 2254 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2255 // If profiling reports a single type for the return value, 2256 // feed it to the type system so it can propagate it as a 2257 // speculative type 2258 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2259 } 2260 } 2261 2262 void GraphKit::round_double_result(ciMethod* dest_method) { 2263 // A non-strict method may return a double value which has an extended 2264 // exponent, but this must not be visible in a caller which is 'strict' 2265 // If a strict caller invokes a non-strict callee, round a double result 2266 2267 BasicType result_type = dest_method->return_type()->basic_type(); 2268 assert( method() != NULL, "must have caller context"); 2269 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2270 // Destination method's return value is on top of stack 2271 // dstore_rounding() does gvn.transform 2272 Node *result = pop_pair(); 2273 result = dstore_rounding(result); 2274 push_pair(result); 2275 } 2276 } 2277 2278 // rounding for strict float precision conformance 2279 Node* GraphKit::precision_rounding(Node* n) { 2280 return UseStrictFP && _method->flags().is_strict() 2281 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2282 ? _gvn.transform( new (C) RoundFloatNode(0, n) ) 2283 : n; 2284 } 2285 2286 // rounding for strict double precision conformance 2287 Node* GraphKit::dprecision_rounding(Node *n) { 2288 return UseStrictFP && _method->flags().is_strict() 2289 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2290 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2291 : n; 2292 } 2293 2294 // rounding for non-strict double stores 2295 Node* GraphKit::dstore_rounding(Node* n) { 2296 return Matcher::strict_fp_requires_explicit_rounding 2297 && UseSSE <= 1 2298 ? _gvn.transform( new (C) RoundDoubleNode(0, n) ) 2299 : n; 2300 } 2301 2302 //============================================================================= 2303 // Generate a fast path/slow path idiom. Graph looks like: 2304 // [foo] indicates that 'foo' is a parameter 2305 // 2306 // [in] NULL 2307 // \ / 2308 // CmpP 2309 // Bool ne 2310 // If 2311 // / \ 2312 // True False-<2> 2313 // / | 2314 // / cast_not_null 2315 // Load | | ^ 2316 // [fast_test] | | 2317 // gvn to opt_test | | 2318 // / \ | <1> 2319 // True False | 2320 // | \\ | 2321 // [slow_call] \[fast_result] 2322 // Ctl Val \ \ 2323 // | \ \ 2324 // Catch <1> \ \ 2325 // / \ ^ \ \ 2326 // Ex No_Ex | \ \ 2327 // | \ \ | \ <2> \ 2328 // ... \ [slow_res] | | \ [null_result] 2329 // \ \--+--+--- | | 2330 // \ | / \ | / 2331 // --------Region Phi 2332 // 2333 //============================================================================= 2334 // Code is structured as a series of driver functions all called 'do_XXX' that 2335 // call a set of helper functions. Helper functions first, then drivers. 2336 2337 //------------------------------null_check_oop--------------------------------- 2338 // Null check oop. Set null-path control into Region in slot 3. 2339 // Make a cast-not-nullness use the other not-null control. Return cast. 2340 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2341 bool never_see_null, 2342 bool safe_for_replace, 2343 bool speculative) { 2344 // Initial NULL check taken path 2345 (*null_control) = top(); 2346 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2347 2348 // Generate uncommon_trap: 2349 if (never_see_null && (*null_control) != top()) { 2350 // If we see an unexpected null at a check-cast we record it and force a 2351 // recompile; the offending check-cast will be compiled to handle NULLs. 2352 // If we see more than one offending BCI, then all checkcasts in the 2353 // method will be compiled to handle NULLs. 2354 PreserveJVMState pjvms(this); 2355 set_control(*null_control); 2356 replace_in_map(value, null()); 2357 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2358 uncommon_trap(reason, 2359 Deoptimization::Action_make_not_entrant); 2360 (*null_control) = top(); // NULL path is dead 2361 } 2362 if ((*null_control) == top() && safe_for_replace) { 2363 replace_in_map(value, cast); 2364 } 2365 2366 // Cast away null-ness on the result 2367 return cast; 2368 } 2369 2370 //------------------------------opt_iff---------------------------------------- 2371 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2372 // Return slow-path control. 2373 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2374 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2375 2376 // Fast path taken; set region slot 2 2377 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) ); 2378 region->init_req(2,fast_taken); // Capture fast-control 2379 2380 // Fast path not-taken, i.e. slow path 2381 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) ); 2382 return slow_taken; 2383 } 2384 2385 //-----------------------------make_runtime_call------------------------------- 2386 Node* GraphKit::make_runtime_call(int flags, 2387 const TypeFunc* call_type, address call_addr, 2388 const char* call_name, 2389 const TypePtr* adr_type, 2390 // The following parms are all optional. 2391 // The first NULL ends the list. 2392 Node* parm0, Node* parm1, 2393 Node* parm2, Node* parm3, 2394 Node* parm4, Node* parm5, 2395 Node* parm6, Node* parm7) { 2396 // Slow-path call 2397 bool is_leaf = !(flags & RC_NO_LEAF); 2398 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2399 if (call_name == NULL) { 2400 assert(!is_leaf, "must supply name for leaf"); 2401 call_name = OptoRuntime::stub_name(call_addr); 2402 } 2403 CallNode* call; 2404 if (!is_leaf) { 2405 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name, 2406 bci(), adr_type); 2407 } else if (flags & RC_NO_FP) { 2408 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2409 } else { 2410 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type); 2411 } 2412 2413 // The following is similar to set_edges_for_java_call, 2414 // except that the memory effects of the call are restricted to AliasIdxRaw. 2415 2416 // Slow path call has no side-effects, uses few values 2417 bool wide_in = !(flags & RC_NARROW_MEM); 2418 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2419 2420 Node* prev_mem = NULL; 2421 if (wide_in) { 2422 prev_mem = set_predefined_input_for_runtime_call(call); 2423 } else { 2424 assert(!wide_out, "narrow in => narrow out"); 2425 Node* narrow_mem = memory(adr_type); 2426 prev_mem = reset_memory(); 2427 map()->set_memory(narrow_mem); 2428 set_predefined_input_for_runtime_call(call); 2429 } 2430 2431 // Hook each parm in order. Stop looking at the first NULL. 2432 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2433 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2434 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2435 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2436 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2437 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2438 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2439 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2440 /* close each nested if ===> */ } } } } } } } } 2441 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2442 2443 if (!is_leaf) { 2444 // Non-leaves can block and take safepoints: 2445 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2446 } 2447 // Non-leaves can throw exceptions: 2448 if (has_io) { 2449 call->set_req(TypeFunc::I_O, i_o()); 2450 } 2451 2452 if (flags & RC_UNCOMMON) { 2453 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2454 // (An "if" probability corresponds roughly to an unconditional count. 2455 // Sort of.) 2456 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2457 } 2458 2459 Node* c = _gvn.transform(call); 2460 assert(c == call, "cannot disappear"); 2461 2462 if (wide_out) { 2463 // Slow path call has full side-effects. 2464 set_predefined_output_for_runtime_call(call); 2465 } else { 2466 // Slow path call has few side-effects, and/or sets few values. 2467 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2468 } 2469 2470 if (has_io) { 2471 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O))); 2472 } 2473 return call; 2474 2475 } 2476 2477 //------------------------------merge_memory----------------------------------- 2478 // Merge memory from one path into the current memory state. 2479 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2480 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2481 Node* old_slice = mms.force_memory(); 2482 Node* new_slice = mms.memory2(); 2483 if (old_slice != new_slice) { 2484 PhiNode* phi; 2485 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) { 2486 phi = new_slice->as_Phi(); 2487 #ifdef ASSERT 2488 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) 2489 old_slice = old_slice->in(new_path); 2490 // Caller is responsible for ensuring that any pre-existing 2491 // phis are already aware of old memory. 2492 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path 2493 assert(phi->in(old_path) == old_slice, "pre-existing phis OK"); 2494 #endif 2495 mms.set_memory(phi); 2496 } else { 2497 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2498 _gvn.set_type(phi, Type::MEMORY); 2499 phi->set_req(new_path, new_slice); 2500 mms.set_memory(_gvn.transform(phi)); // assume it is complete 2501 } 2502 } 2503 } 2504 } 2505 2506 //------------------------------make_slow_call_ex------------------------------ 2507 // Make the exception handler hookups for the slow call 2508 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) { 2509 if (stopped()) return; 2510 2511 // Make a catch node with just two handlers: fall-through and catch-all 2512 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2513 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) ); 2514 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2515 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2516 2517 { PreserveJVMState pjvms(this); 2518 set_control(excp); 2519 set_i_o(i_o); 2520 2521 if (excp != top()) { 2522 // Create an exception state also. 2523 // Use an exact type if the caller has specified a specific exception. 2524 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2525 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o); 2526 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2527 } 2528 } 2529 2530 // Get the no-exception control from the CatchNode. 2531 set_control(norm); 2532 } 2533 2534 2535 //-------------------------------gen_subtype_check----------------------------- 2536 // Generate a subtyping check. Takes as input the subtype and supertype. 2537 // Returns 2 values: sets the default control() to the true path and returns 2538 // the false path. Only reads invariant memory; sets no (visible) memory. 2539 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2540 // but that's not exposed to the optimizer. This call also doesn't take in an 2541 // Object; if you wish to check an Object you need to load the Object's class 2542 // prior to coming here. 2543 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) { 2544 // Fast check for identical types, perhaps identical constants. 2545 // The types can even be identical non-constants, in cases 2546 // involving Array.newInstance, Object.clone, etc. 2547 if (subklass == superklass) 2548 return top(); // false path is dead; no test needed. 2549 2550 if (_gvn.type(superklass)->singleton()) { 2551 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2552 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass(); 2553 2554 // In the common case of an exact superklass, try to fold up the 2555 // test before generating code. You may ask, why not just generate 2556 // the code and then let it fold up? The answer is that the generated 2557 // code will necessarily include null checks, which do not always 2558 // completely fold away. If they are also needless, then they turn 2559 // into a performance loss. Example: 2560 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2561 // Here, the type of 'fa' is often exact, so the store check 2562 // of fa[1]=x will fold up, without testing the nullness of x. 2563 switch (static_subtype_check(superk, subk)) { 2564 case SSC_always_false: 2565 { 2566 Node* always_fail = control(); 2567 set_control(top()); 2568 return always_fail; 2569 } 2570 case SSC_always_true: 2571 return top(); 2572 case SSC_easy_test: 2573 { 2574 // Just do a direct pointer compare and be done. 2575 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) ); 2576 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2577 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 2578 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) ); 2579 return _gvn.transform( new(C) IfFalseNode(iff) ); 2580 } 2581 case SSC_full_test: 2582 break; 2583 default: 2584 ShouldNotReachHere(); 2585 } 2586 } 2587 2588 // %%% Possible further optimization: Even if the superklass is not exact, 2589 // if the subklass is the unique subtype of the superklass, the check 2590 // will always succeed. We could leave a dependency behind to ensure this. 2591 2592 // First load the super-klass's check-offset 2593 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) ); 2594 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(), 2595 TypeInt::INT, MemNode::unordered)); 2596 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2597 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2598 2599 // Load from the sub-klass's super-class display list, or a 1-word cache of 2600 // the secondary superclass list, or a failing value with a sentinel offset 2601 // if the super-klass is an interface or exceptionally deep in the Java 2602 // hierarchy and we have to scan the secondary superclass list the hard way. 2603 // Worst-case type is a little odd: NULL is allowed as a result (usually 2604 // klass loads can never produce a NULL). 2605 Node *chk_off_X = ConvI2X(chk_off); 2606 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) ); 2607 // For some types like interfaces the following loadKlass is from a 1-word 2608 // cache which is mutable so can't use immutable memory. Other 2609 // types load from the super-class display table which is immutable. 2610 Node *kmem = might_be_cache ? memory(p2) : immutable_memory(); 2611 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) ); 2612 2613 // Compile speed common case: ARE a subtype and we canNOT fail 2614 if( superklass == nkls ) 2615 return top(); // false path is dead; no test needed. 2616 2617 // See if we get an immediate positive hit. Happens roughly 83% of the 2618 // time. Test to see if the value loaded just previously from the subklass 2619 // is exactly the superklass. 2620 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) ); 2621 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) ); 2622 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN ); 2623 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) ); 2624 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) ); 2625 2626 // Compile speed common case: Check for being deterministic right now. If 2627 // chk_off is a constant and not equal to cacheoff then we are NOT a 2628 // subklass. In this case we need exactly the 1 test above and we can 2629 // return those results immediately. 2630 if (!might_be_cache) { 2631 Node* not_subtype_ctrl = control(); 2632 set_control(iftrue1); // We need exactly the 1 test above 2633 return not_subtype_ctrl; 2634 } 2635 2636 // Gather the various success & failures here 2637 RegionNode *r_ok_subtype = new (C) RegionNode(4); 2638 record_for_igvn(r_ok_subtype); 2639 RegionNode *r_not_subtype = new (C) RegionNode(3); 2640 record_for_igvn(r_not_subtype); 2641 2642 r_ok_subtype->init_req(1, iftrue1); 2643 2644 // Check for immediate negative hit. Happens roughly 11% of the time (which 2645 // is roughly 63% of the remaining cases). Test to see if the loaded 2646 // check-offset points into the subklass display list or the 1-element 2647 // cache. If it points to the display (and NOT the cache) and the display 2648 // missed then it's not a subtype. 2649 Node *cacheoff = _gvn.intcon(cacheoff_con); 2650 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) ); 2651 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) ); 2652 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN ); 2653 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) ); 2654 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) ); 2655 2656 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2657 // No performance impact (too rare) but allows sharing of secondary arrays 2658 // which has some footprint reduction. 2659 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) ); 2660 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) ); 2661 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN ); 2662 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) ); 2663 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) ); 2664 2665 // -- Roads not taken here: -- 2666 // We could also have chosen to perform the self-check at the beginning 2667 // of this code sequence, as the assembler does. This would not pay off 2668 // the same way, since the optimizer, unlike the assembler, can perform 2669 // static type analysis to fold away many successful self-checks. 2670 // Non-foldable self checks work better here in second position, because 2671 // the initial primary superclass check subsumes a self-check for most 2672 // types. An exception would be a secondary type like array-of-interface, 2673 // which does not appear in its own primary supertype display. 2674 // Finally, we could have chosen to move the self-check into the 2675 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2676 // dependent manner. But it is worthwhile to have the check here, 2677 // where it can be perhaps be optimized. The cost in code space is 2678 // small (register compare, branch). 2679 2680 // Now do a linear scan of the secondary super-klass array. Again, no real 2681 // performance impact (too rare) but it's gotta be done. 2682 // Since the code is rarely used, there is no penalty for moving it 2683 // out of line, and it can only improve I-cache density. 2684 // The decision to inline or out-of-line this final check is platform 2685 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2686 Node* psc = _gvn.transform( 2687 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) ); 2688 2689 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) ); 2690 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) ); 2691 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN ); 2692 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) ); 2693 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) ); 2694 2695 // Return false path; set default control to true path. 2696 set_control( _gvn.transform(r_ok_subtype) ); 2697 return _gvn.transform(r_not_subtype); 2698 } 2699 2700 //----------------------------static_subtype_check----------------------------- 2701 // Shortcut important common cases when superklass is exact: 2702 // (0) superklass is java.lang.Object (can occur in reflective code) 2703 // (1) subklass is already limited to a subtype of superklass => always ok 2704 // (2) subklass does not overlap with superklass => always fail 2705 // (3) superklass has NO subtypes and we can check with a simple compare. 2706 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) { 2707 if (StressReflectiveCode) { 2708 return SSC_full_test; // Let caller generate the general case. 2709 } 2710 2711 if (superk == env()->Object_klass()) { 2712 return SSC_always_true; // (0) this test cannot fail 2713 } 2714 2715 ciType* superelem = superk; 2716 if (superelem->is_array_klass()) 2717 superelem = superelem->as_array_klass()->base_element_type(); 2718 2719 if (!subk->is_interface()) { // cannot trust static interface types yet 2720 if (subk->is_subtype_of(superk)) { 2721 return SSC_always_true; // (1) false path dead; no dynamic test needed 2722 } 2723 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 2724 !superk->is_subtype_of(subk)) { 2725 return SSC_always_false; 2726 } 2727 } 2728 2729 // If casting to an instance klass, it must have no subtypes 2730 if (superk->is_interface()) { 2731 // Cannot trust interfaces yet. 2732 // %%% S.B. superk->nof_implementors() == 1 2733 } else if (superelem->is_instance_klass()) { 2734 ciInstanceKlass* ik = superelem->as_instance_klass(); 2735 if (!ik->has_subklass() && !ik->is_interface()) { 2736 if (!ik->is_final()) { 2737 // Add a dependency if there is a chance of a later subclass. 2738 C->dependencies()->assert_leaf_type(ik); 2739 } 2740 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2741 } 2742 } else { 2743 // A primitive array type has no subtypes. 2744 return SSC_easy_test; // (3) caller can do a simple ptr comparison 2745 } 2746 2747 return SSC_full_test; 2748 } 2749 2750 // Profile-driven exact type check: 2751 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2752 float prob, 2753 Node* *casted_receiver) { 2754 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2755 Node* recv_klass = load_object_klass(receiver); 2756 Node* want_klass = makecon(tklass); 2757 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) ); 2758 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) ); 2759 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2760 set_control( _gvn.transform( new(C) IfTrueNode (iff) )); 2761 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) ); 2762 2763 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2764 assert(recv_xtype->klass_is_exact(), ""); 2765 2766 // Subsume downstream occurrences of receiver with a cast to 2767 // recv_xtype, since now we know what the type will be. 2768 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype); 2769 (*casted_receiver) = _gvn.transform(cast); 2770 // (User must make the replace_in_map call.) 2771 2772 return fail; 2773 } 2774 2775 2776 //------------------------------seems_never_null------------------------------- 2777 // Use null_seen information if it is available from the profile. 2778 // If we see an unexpected null at a type check we record it and force a 2779 // recompile; the offending check will be recompiled to handle NULLs. 2780 // If we see several offending BCIs, then all checks in the 2781 // method will be recompiled. 2782 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2783 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2784 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2785 if (UncommonNullCast // Cutout for this technique 2786 && obj != null() // And not the -Xcomp stupid case? 2787 && !too_many_traps(reason) 2788 ) { 2789 if (speculating) { 2790 return true; 2791 } 2792 if (data == NULL) 2793 // Edge case: no mature data. Be optimistic here. 2794 return true; 2795 // If the profile has not seen a null, assume it won't happen. 2796 assert(java_bc() == Bytecodes::_checkcast || 2797 java_bc() == Bytecodes::_instanceof || 2798 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2799 return !data->as_BitData()->null_seen(); 2800 } 2801 speculating = false; 2802 return false; 2803 } 2804 2805 //------------------------maybe_cast_profiled_receiver------------------------- 2806 // If the profile has seen exactly one type, narrow to exactly that type. 2807 // Subsequent type checks will always fold up. 2808 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2809 ciKlass* require_klass, 2810 ciKlass* spec_klass, 2811 bool safe_for_replace) { 2812 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2813 2814 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2815 2816 // Make sure we haven't already deoptimized from this tactic. 2817 if (too_many_traps(reason)) 2818 return NULL; 2819 2820 // (No, this isn't a call, but it's enough like a virtual call 2821 // to use the same ciMethod accessor to get the profile info...) 2822 // If we have a speculative type use it instead of profiling (which 2823 // may not help us) 2824 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2825 if (exact_kls != NULL) {// no cast failures here 2826 if (require_klass == NULL || 2827 static_subtype_check(require_klass, exact_kls) == SSC_always_true) { 2828 // If we narrow the type to match what the type profile sees or 2829 // the speculative type, we can then remove the rest of the 2830 // cast. 2831 // This is a win, even if the exact_kls is very specific, 2832 // because downstream operations, such as method calls, 2833 // will often benefit from the sharper type. 2834 Node* exact_obj = not_null_obj; // will get updated in place... 2835 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2836 &exact_obj); 2837 { PreserveJVMState pjvms(this); 2838 set_control(slow_ctl); 2839 uncommon_trap(reason, 2840 Deoptimization::Action_maybe_recompile); 2841 } 2842 if (safe_for_replace) { 2843 replace_in_map(not_null_obj, exact_obj); 2844 } 2845 return exact_obj; 2846 } 2847 // assert(ssc == SSC_always_true)... except maybe the profile lied to us. 2848 } 2849 2850 return NULL; 2851 } 2852 2853 /** 2854 * Cast obj to type and emit guard unless we had too many traps here 2855 * already 2856 * 2857 * @param obj node being casted 2858 * @param type type to cast the node to 2859 * @param not_null true if we know node cannot be null 2860 */ 2861 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2862 ciKlass* type, 2863 bool not_null) { 2864 // type == NULL if profiling tells us this object is always null 2865 if (type != NULL) { 2866 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2867 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2868 if (!too_many_traps(null_reason) && 2869 !too_many_traps(class_reason)) { 2870 Node* not_null_obj = NULL; 2871 // not_null is true if we know the object is not null and 2872 // there's no need for a null check 2873 if (!not_null) { 2874 Node* null_ctl = top(); 2875 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2876 assert(null_ctl->is_top(), "no null control here"); 2877 } else { 2878 not_null_obj = obj; 2879 } 2880 2881 Node* exact_obj = not_null_obj; 2882 ciKlass* exact_kls = type; 2883 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2884 &exact_obj); 2885 { 2886 PreserveJVMState pjvms(this); 2887 set_control(slow_ctl); 2888 uncommon_trap(class_reason, 2889 Deoptimization::Action_maybe_recompile); 2890 } 2891 replace_in_map(not_null_obj, exact_obj); 2892 obj = exact_obj; 2893 } 2894 } else { 2895 if (!too_many_traps(Deoptimization::Reason_null_assert)) { 2896 Node* exact_obj = null_assert(obj); 2897 replace_in_map(obj, exact_obj); 2898 obj = exact_obj; 2899 } 2900 } 2901 return obj; 2902 } 2903 2904 //-------------------------------gen_instanceof-------------------------------- 2905 // Generate an instance-of idiom. Used by both the instance-of bytecode 2906 // and the reflective instance-of call. 2907 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2908 kill_dead_locals(); // Benefit all the uncommon traps 2909 assert( !stopped(), "dead parse path should be checked in callers" ); 2910 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2911 "must check for not-null not-dead klass in callers"); 2912 2913 // Make the merge point 2914 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2915 RegionNode* region = new(C) RegionNode(PATH_LIMIT); 2916 Node* phi = new(C) PhiNode(region, TypeInt::BOOL); 2917 C->set_has_split_ifs(true); // Has chance for split-if optimization 2918 2919 ciProfileData* data = NULL; 2920 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2921 data = method()->method_data()->bci_to_data(bci()); 2922 } 2923 bool speculative_not_null = false; 2924 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2925 && seems_never_null(obj, data, speculative_not_null)); 2926 2927 // Null check; get casted pointer; set region slot 3 2928 Node* null_ctl = top(); 2929 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2930 2931 // If not_null_obj is dead, only null-path is taken 2932 if (stopped()) { // Doing instance-of on a NULL? 2933 set_control(null_ctl); 2934 return intcon(0); 2935 } 2936 region->init_req(_null_path, null_ctl); 2937 phi ->init_req(_null_path, intcon(0)); // Set null path value 2938 if (null_ctl == top()) { 2939 // Do this eagerly, so that pattern matches like is_diamond_phi 2940 // will work even during parsing. 2941 assert(_null_path == PATH_LIMIT-1, "delete last"); 2942 region->del_req(_null_path); 2943 phi ->del_req(_null_path); 2944 } 2945 2946 // Do we know the type check always succeed? 2947 bool known_statically = false; 2948 if (_gvn.type(superklass)->singleton()) { 2949 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2950 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2951 if (subk != NULL && subk->is_loaded()) { 2952 int static_res = static_subtype_check(superk, subk); 2953 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false); 2954 } 2955 } 2956 2957 if (known_statically && UseTypeSpeculation) { 2958 // If we know the type check always succeeds then we don't use the 2959 // profiling data at this bytecode. Don't lose it, feed it to the 2960 // type system as a speculative type. 2961 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2962 } else { 2963 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2964 // We may not have profiling here or it may not help us. If we 2965 // have a speculative type use it to perform an exact cast. 2966 ciKlass* spec_obj_type = obj_type->speculative_type(); 2967 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2968 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2969 if (stopped()) { // Profile disagrees with this path. 2970 set_control(null_ctl); // Null is the only remaining possibility. 2971 return intcon(0); 2972 } 2973 if (cast_obj != NULL) { 2974 not_null_obj = cast_obj; 2975 } 2976 } 2977 } 2978 2979 // Load the object's klass 2980 Node* obj_klass = load_object_klass(not_null_obj); 2981 2982 // Generate the subtype check 2983 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2984 2985 // Plug in the success path to the general merge in slot 1. 2986 region->init_req(_obj_path, control()); 2987 phi ->init_req(_obj_path, intcon(1)); 2988 2989 // Plug in the failing path to the general merge in slot 2. 2990 region->init_req(_fail_path, not_subtype_ctrl); 2991 phi ->init_req(_fail_path, intcon(0)); 2992 2993 // Return final merged results 2994 set_control( _gvn.transform(region) ); 2995 record_for_igvn(region); 2996 return _gvn.transform(phi); 2997 } 2998 2999 //-------------------------------gen_checkcast--------------------------------- 3000 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3001 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3002 // uncommon-trap paths work. Adjust stack after this call. 3003 // If failure_control is supplied and not null, it is filled in with 3004 // the control edge for the cast failure. Otherwise, an appropriate 3005 // uncommon trap or exception is thrown. 3006 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3007 Node* *failure_control) { 3008 kill_dead_locals(); // Benefit all the uncommon traps 3009 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3010 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3011 3012 // Fast cutout: Check the case that the cast is vacuously true. 3013 // This detects the common cases where the test will short-circuit 3014 // away completely. We do this before we perform the null check, 3015 // because if the test is going to turn into zero code, we don't 3016 // want a residual null check left around. (Causes a slowdown, 3017 // for example, in some objArray manipulations, such as a[i]=a[j].) 3018 if (tk->singleton()) { 3019 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3020 if (objtp != NULL && objtp->klass() != NULL) { 3021 switch (static_subtype_check(tk->klass(), objtp->klass())) { 3022 case SSC_always_true: 3023 // If we know the type check always succeed then we don't use 3024 // the profiling data at this bytecode. Don't lose it, feed it 3025 // to the type system as a speculative type. 3026 return record_profiled_receiver_for_speculation(obj); 3027 case SSC_always_false: 3028 // It needs a null check because a null will *pass* the cast check. 3029 // A non-null value will always produce an exception. 3030 return null_assert(obj); 3031 } 3032 } 3033 } 3034 3035 ciProfileData* data = NULL; 3036 bool safe_for_replace = false; 3037 if (failure_control == NULL) { // use MDO in regular case only 3038 assert(java_bc() == Bytecodes::_aastore || 3039 java_bc() == Bytecodes::_checkcast, 3040 "interpreter profiles type checks only for these BCs"); 3041 data = method()->method_data()->bci_to_data(bci()); 3042 safe_for_replace = true; 3043 } 3044 3045 // Make the merge point 3046 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3047 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3048 Node* phi = new (C) PhiNode(region, toop); 3049 C->set_has_split_ifs(true); // Has chance for split-if optimization 3050 3051 // Use null-cast information if it is available 3052 bool speculative_not_null = false; 3053 bool never_see_null = ((failure_control == NULL) // regular case only 3054 && seems_never_null(obj, data, speculative_not_null)); 3055 3056 // Null check; get casted pointer; set region slot 3 3057 Node* null_ctl = top(); 3058 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3059 3060 // If not_null_obj is dead, only null-path is taken 3061 if (stopped()) { // Doing instance-of on a NULL? 3062 set_control(null_ctl); 3063 return null(); 3064 } 3065 region->init_req(_null_path, null_ctl); 3066 phi ->init_req(_null_path, null()); // Set null path value 3067 if (null_ctl == top()) { 3068 // Do this eagerly, so that pattern matches like is_diamond_phi 3069 // will work even during parsing. 3070 assert(_null_path == PATH_LIMIT-1, "delete last"); 3071 region->del_req(_null_path); 3072 phi ->del_req(_null_path); 3073 } 3074 3075 Node* cast_obj = NULL; 3076 if (tk->klass_is_exact()) { 3077 // The following optimization tries to statically cast the speculative type of the object 3078 // (for example obtained during profiling) to the type of the superklass and then do a 3079 // dynamic check that the type of the object is what we expect. To work correctly 3080 // for checkcast and aastore the type of superklass should be exact. 3081 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3082 // We may not have profiling here or it may not help us. If we have 3083 // a speculative type use it to perform an exact cast. 3084 ciKlass* spec_obj_type = obj_type->speculative_type(); 3085 if (spec_obj_type != NULL || 3086 (data != NULL && 3087 // Counter has never been decremented (due to cast failure). 3088 // ...This is a reasonable thing to expect. It is true of 3089 // all casts inserted by javac to implement generic types. 3090 data->as_CounterData()->count() >= 0)) { 3091 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3092 if (cast_obj != NULL) { 3093 if (failure_control != NULL) // failure is now impossible 3094 (*failure_control) = top(); 3095 // adjust the type of the phi to the exact klass: 3096 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3097 } 3098 } 3099 } 3100 3101 if (cast_obj == NULL) { 3102 // Load the object's klass 3103 Node* obj_klass = load_object_klass(not_null_obj); 3104 3105 // Generate the subtype check 3106 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3107 3108 // Plug in success path into the merge 3109 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), 3110 not_null_obj, toop)); 3111 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3112 if (failure_control == NULL) { 3113 if (not_subtype_ctrl != top()) { // If failure is possible 3114 PreserveJVMState pjvms(this); 3115 set_control(not_subtype_ctrl); 3116 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3117 } 3118 } else { 3119 (*failure_control) = not_subtype_ctrl; 3120 } 3121 } 3122 3123 region->init_req(_obj_path, control()); 3124 phi ->init_req(_obj_path, cast_obj); 3125 3126 // A merge of NULL or Casted-NotNull obj 3127 Node* res = _gvn.transform(phi); 3128 3129 // Note I do NOT always 'replace_in_map(obj,result)' here. 3130 // if( tk->klass()->can_be_primary_super() ) 3131 // This means that if I successfully store an Object into an array-of-String 3132 // I 'forget' that the Object is really now known to be a String. I have to 3133 // do this because we don't have true union types for interfaces - if I store 3134 // a Baz into an array-of-Interface and then tell the optimizer it's an 3135 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3136 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3137 // replace_in_map( obj, res ); 3138 3139 // Return final merged results 3140 set_control( _gvn.transform(region) ); 3141 record_for_igvn(region); 3142 return res; 3143 } 3144 3145 //------------------------------next_monitor----------------------------------- 3146 // What number should be given to the next monitor? 3147 int GraphKit::next_monitor() { 3148 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3149 int next = current + C->sync_stack_slots(); 3150 // Keep the toplevel high water mark current: 3151 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3152 return current; 3153 } 3154 3155 //------------------------------insert_mem_bar--------------------------------- 3156 // Memory barrier to avoid floating things around 3157 // The membar serves as a pinch point between both control and all memory slices. 3158 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3159 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3160 mb->init_req(TypeFunc::Control, control()); 3161 mb->init_req(TypeFunc::Memory, reset_memory()); 3162 Node* membar = _gvn.transform(mb); 3163 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3164 set_all_memory_call(membar); 3165 return membar; 3166 } 3167 3168 //-------------------------insert_mem_bar_volatile---------------------------- 3169 // Memory barrier to avoid floating things around 3170 // The membar serves as a pinch point between both control and memory(alias_idx). 3171 // If you want to make a pinch point on all memory slices, do not use this 3172 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3173 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3174 // When Parse::do_put_xxx updates a volatile field, it appends a series 3175 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3176 // The first membar is on the same memory slice as the field store opcode. 3177 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3178 // All the other membars (for other volatile slices, including AliasIdxBot, 3179 // which stands for all unknown volatile slices) are control-dependent 3180 // on the first membar. This prevents later volatile loads or stores 3181 // from sliding up past the just-emitted store. 3182 3183 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3184 mb->set_req(TypeFunc::Control,control()); 3185 if (alias_idx == Compile::AliasIdxBot) { 3186 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3187 } else { 3188 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3189 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3190 } 3191 Node* membar = _gvn.transform(mb); 3192 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control))); 3193 if (alias_idx == Compile::AliasIdxBot) { 3194 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory))); 3195 } else { 3196 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx); 3197 } 3198 return membar; 3199 } 3200 3201 //------------------------------shared_lock------------------------------------ 3202 // Emit locking code. 3203 FastLockNode* GraphKit::shared_lock(Node* obj) { 3204 // bci is either a monitorenter bc or InvocationEntryBci 3205 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3206 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3207 3208 if( !GenerateSynchronizationCode ) 3209 return NULL; // Not locking things? 3210 if (stopped()) // Dead monitor? 3211 return NULL; 3212 3213 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3214 3215 // Box the stack location 3216 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 3217 Node* mem = reset_memory(); 3218 3219 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock(); 3220 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3221 // Create the counters for this fast lock. 3222 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3223 } 3224 3225 // Create the rtm counters for this fast lock if needed. 3226 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3227 3228 // Add monitor to debug info for the slow path. If we block inside the 3229 // slow path and de-opt, we need the monitor hanging around 3230 map()->push_monitor( flock ); 3231 3232 const TypeFunc *tf = LockNode::lock_type(); 3233 LockNode *lock = new (C) LockNode(C, tf); 3234 3235 lock->init_req( TypeFunc::Control, control() ); 3236 lock->init_req( TypeFunc::Memory , mem ); 3237 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3238 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3239 lock->init_req( TypeFunc::ReturnAdr, top() ); 3240 3241 lock->init_req(TypeFunc::Parms + 0, obj); 3242 lock->init_req(TypeFunc::Parms + 1, box); 3243 lock->init_req(TypeFunc::Parms + 2, flock); 3244 add_safepoint_edges(lock); 3245 3246 lock = _gvn.transform( lock )->as_Lock(); 3247 3248 // lock has no side-effects, sets few values 3249 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3250 3251 insert_mem_bar(Op_MemBarAcquireLock); 3252 3253 // Add this to the worklist so that the lock can be eliminated 3254 record_for_igvn(lock); 3255 3256 #ifndef PRODUCT 3257 if (PrintLockStatistics) { 3258 // Update the counter for this lock. Don't bother using an atomic 3259 // operation since we don't require absolute accuracy. 3260 lock->create_lock_counter(map()->jvms()); 3261 increment_counter(lock->counter()->addr()); 3262 } 3263 #endif 3264 3265 return flock; 3266 } 3267 3268 3269 //------------------------------shared_unlock---------------------------------- 3270 // Emit unlocking code. 3271 void GraphKit::shared_unlock(Node* box, Node* obj) { 3272 // bci is either a monitorenter bc or InvocationEntryBci 3273 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3274 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3275 3276 if( !GenerateSynchronizationCode ) 3277 return; 3278 if (stopped()) { // Dead monitor? 3279 map()->pop_monitor(); // Kill monitor from debug info 3280 return; 3281 } 3282 3283 // Memory barrier to avoid floating things down past the locked region 3284 insert_mem_bar(Op_MemBarReleaseLock); 3285 3286 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3287 UnlockNode *unlock = new (C) UnlockNode(C, tf); 3288 uint raw_idx = Compile::AliasIdxRaw; 3289 unlock->init_req( TypeFunc::Control, control() ); 3290 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3291 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3292 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3293 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3294 3295 unlock->init_req(TypeFunc::Parms + 0, obj); 3296 unlock->init_req(TypeFunc::Parms + 1, box); 3297 unlock = _gvn.transform(unlock)->as_Unlock(); 3298 3299 Node* mem = reset_memory(); 3300 3301 // unlock has no side-effects, sets few values 3302 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3303 3304 // Kill monitor from debug info 3305 map()->pop_monitor( ); 3306 } 3307 3308 //-------------------------------get_layout_helper----------------------------- 3309 // If the given klass is a constant or known to be an array, 3310 // fetch the constant layout helper value into constant_value 3311 // and return (Node*)NULL. Otherwise, load the non-constant 3312 // layout helper value, and return the node which represents it. 3313 // This two-faced routine is useful because allocation sites 3314 // almost always feature constant types. 3315 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3316 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3317 if (!StressReflectiveCode && inst_klass != NULL) { 3318 ciKlass* klass = inst_klass->klass(); 3319 bool xklass = inst_klass->klass_is_exact(); 3320 if (xklass || klass->is_array_klass()) { 3321 jint lhelper = klass->layout_helper(); 3322 if (lhelper != Klass::_lh_neutral_value) { 3323 constant_value = lhelper; 3324 return (Node*) NULL; 3325 } 3326 } 3327 } 3328 constant_value = Klass::_lh_neutral_value; // put in a known value 3329 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3330 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3331 } 3332 3333 // We just put in an allocate/initialize with a big raw-memory effect. 3334 // Hook selected additional alias categories on the initialization. 3335 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3336 MergeMemNode* init_in_merge, 3337 Node* init_out_raw) { 3338 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3339 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3340 3341 Node* prevmem = kit.memory(alias_idx); 3342 init_in_merge->set_memory_at(alias_idx, prevmem); 3343 kit.set_memory(init_out_raw, alias_idx); 3344 } 3345 3346 //---------------------------set_output_for_allocation------------------------- 3347 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3348 const TypeOopPtr* oop_type) { 3349 int rawidx = Compile::AliasIdxRaw; 3350 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3351 add_safepoint_edges(alloc); 3352 Node* allocx = _gvn.transform(alloc); 3353 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) ); 3354 // create memory projection for i_o 3355 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3356 make_slow_call_ex(allocx, env()->Throwable_klass(), true); 3357 3358 // create a memory projection as for the normal control path 3359 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory)); 3360 set_memory(malloc, rawidx); 3361 3362 // a normal slow-call doesn't change i_o, but an allocation does 3363 // we create a separate i_o projection for the normal control path 3364 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) ); 3365 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) ); 3366 3367 // put in an initialization barrier 3368 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3369 rawoop)->as_Initialize(); 3370 assert(alloc->initialization() == init, "2-way macro link must work"); 3371 assert(init ->allocation() == alloc, "2-way macro link must work"); 3372 { 3373 // Extract memory strands which may participate in the new object's 3374 // initialization, and source them from the new InitializeNode. 3375 // This will allow us to observe initializations when they occur, 3376 // and link them properly (as a group) to the InitializeNode. 3377 assert(init->in(InitializeNode::Memory) == malloc, ""); 3378 MergeMemNode* minit_in = MergeMemNode::make(C, malloc); 3379 init->set_req(InitializeNode::Memory, minit_in); 3380 record_for_igvn(minit_in); // fold it up later, if possible 3381 Node* minit_out = memory(rawidx); 3382 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3383 if (oop_type->isa_aryptr()) { 3384 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3385 int elemidx = C->get_alias_index(telemref); 3386 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3387 } else if (oop_type->isa_instptr()) { 3388 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3389 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3390 ciField* field = ik->nonstatic_field_at(i); 3391 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3392 continue; // do not bother to track really large numbers of fields 3393 // Find (or create) the alias category for this field: 3394 int fieldidx = C->alias_type(field)->index(); 3395 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3396 } 3397 } 3398 } 3399 3400 // Cast raw oop to the real thing... 3401 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type); 3402 javaoop = _gvn.transform(javaoop); 3403 C->set_recent_alloc(control(), javaoop); 3404 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3405 3406 #ifdef ASSERT 3407 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3408 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3409 "Ideal_allocation works"); 3410 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3411 "Ideal_allocation works"); 3412 if (alloc->is_AllocateArray()) { 3413 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3414 "Ideal_allocation works"); 3415 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3416 "Ideal_allocation works"); 3417 } else { 3418 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3419 } 3420 } 3421 #endif //ASSERT 3422 3423 return javaoop; 3424 } 3425 3426 //---------------------------new_instance-------------------------------------- 3427 // This routine takes a klass_node which may be constant (for a static type) 3428 // or may be non-constant (for reflective code). It will work equally well 3429 // for either, and the graph will fold nicely if the optimizer later reduces 3430 // the type to a constant. 3431 // The optional arguments are for specialized use by intrinsics: 3432 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3433 // - If 'return_size_val', report the the total object size to the caller. 3434 Node* GraphKit::new_instance(Node* klass_node, 3435 Node* extra_slow_test, 3436 Node* *return_size_val) { 3437 // Compute size in doublewords 3438 // The size is always an integral number of doublewords, represented 3439 // as a positive bytewise size stored in the klass's layout_helper. 3440 // The layout_helper also encodes (in a low bit) the need for a slow path. 3441 jint layout_con = Klass::_lh_neutral_value; 3442 Node* layout_val = get_layout_helper(klass_node, layout_con); 3443 int layout_is_con = (layout_val == NULL); 3444 3445 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3446 // Generate the initial go-slow test. It's either ALWAYS (return a 3447 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3448 // case) a computed value derived from the layout_helper. 3449 Node* initial_slow_test = NULL; 3450 if (layout_is_con) { 3451 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3452 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3453 initial_slow_test = must_go_slow? intcon(1): extra_slow_test; 3454 3455 } else { // reflective case 3456 // This reflective path is used by Unsafe.allocateInstance. 3457 // (It may be stress-tested by specifying StressReflectiveCode.) 3458 // Basically, we want to get into the VM is there's an illegal argument. 3459 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3460 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) ); 3461 if (extra_slow_test != intcon(0)) { 3462 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) ); 3463 } 3464 // (Macro-expander will further convert this to a Bool, if necessary.) 3465 } 3466 3467 // Find the size in bytes. This is easy; it's the layout_helper. 3468 // The size value must be valid even if the slow path is taken. 3469 Node* size = NULL; 3470 if (layout_is_con) { 3471 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3472 } else { // reflective case 3473 // This reflective path is used by clone and Unsafe.allocateInstance. 3474 size = ConvI2X(layout_val); 3475 3476 // Clear the low bits to extract layout_helper_size_in_bytes: 3477 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3478 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3479 size = _gvn.transform( new (C) AndXNode(size, mask) ); 3480 } 3481 if (return_size_val != NULL) { 3482 (*return_size_val) = size; 3483 } 3484 3485 // This is a precise notnull oop of the klass. 3486 // (Actually, it need not be precise if this is a reflective allocation.) 3487 // It's what we cast the result to. 3488 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3489 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3490 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3491 3492 // Now generate allocation code 3493 3494 // The entire memory state is needed for slow path of the allocation 3495 // since GC and deoptimization can happened. 3496 Node *mem = reset_memory(); 3497 set_all_memory(mem); // Create new memory state 3498 3499 AllocateNode* alloc 3500 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3501 control(), mem, i_o(), 3502 size, klass_node, 3503 initial_slow_test); 3504 3505 return set_output_for_allocation(alloc, oop_type); 3506 } 3507 3508 //-------------------------------new_array------------------------------------- 3509 // helper for both newarray and anewarray 3510 // The 'length' parameter is (obviously) the length of the array. 3511 // See comments on new_instance for the meaning of the other arguments. 3512 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3513 Node* length, // number of array elements 3514 int nargs, // number of arguments to push back for uncommon trap 3515 Node* *return_size_val) { 3516 jint layout_con = Klass::_lh_neutral_value; 3517 Node* layout_val = get_layout_helper(klass_node, layout_con); 3518 int layout_is_con = (layout_val == NULL); 3519 3520 if (!layout_is_con && !StressReflectiveCode && 3521 !too_many_traps(Deoptimization::Reason_class_check)) { 3522 // This is a reflective array creation site. 3523 // Optimistically assume that it is a subtype of Object[], 3524 // so that we can fold up all the address arithmetic. 3525 layout_con = Klass::array_layout_helper(T_OBJECT); 3526 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) ); 3527 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) ); 3528 { BuildCutout unless(this, bol_lh, PROB_MAX); 3529 inc_sp(nargs); 3530 uncommon_trap(Deoptimization::Reason_class_check, 3531 Deoptimization::Action_maybe_recompile); 3532 } 3533 layout_val = NULL; 3534 layout_is_con = true; 3535 } 3536 3537 // Generate the initial go-slow test. Make sure we do not overflow 3538 // if length is huge (near 2Gig) or negative! We do not need 3539 // exact double-words here, just a close approximation of needed 3540 // double-words. We can't add any offset or rounding bits, lest we 3541 // take a size -1 of bytes and make it positive. Use an unsigned 3542 // compare, so negative sizes look hugely positive. 3543 int fast_size_limit = FastAllocateSizeLimit; 3544 if (layout_is_con) { 3545 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3546 // Increase the size limit if we have exact knowledge of array type. 3547 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3548 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3549 } 3550 3551 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) ); 3552 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3553 if (initial_slow_test->is_Bool()) { 3554 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3555 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3556 } 3557 3558 // --- Size Computation --- 3559 // array_size = round_to_heap(array_header + (length << elem_shift)); 3560 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3561 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3562 // The rounding mask is strength-reduced, if possible. 3563 int round_mask = MinObjAlignmentInBytes - 1; 3564 Node* header_size = NULL; 3565 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3566 // (T_BYTE has the weakest alignment and size restrictions...) 3567 if (layout_is_con) { 3568 int hsize = Klass::layout_helper_header_size(layout_con); 3569 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3570 BasicType etype = Klass::layout_helper_element_type(layout_con); 3571 if ((round_mask & ~right_n_bits(eshift)) == 0) 3572 round_mask = 0; // strength-reduce it if it goes away completely 3573 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3574 assert(header_size_min <= hsize, "generic minimum is smallest"); 3575 header_size_min = hsize; 3576 header_size = intcon(hsize + round_mask); 3577 } else { 3578 Node* hss = intcon(Klass::_lh_header_size_shift); 3579 Node* hsm = intcon(Klass::_lh_header_size_mask); 3580 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) ); 3581 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) ); 3582 Node* mask = intcon(round_mask); 3583 header_size = _gvn.transform( new(C) AddINode(hsize, mask) ); 3584 } 3585 3586 Node* elem_shift = NULL; 3587 if (layout_is_con) { 3588 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3589 if (eshift != 0) 3590 elem_shift = intcon(eshift); 3591 } else { 3592 // There is no need to mask or shift this value. 3593 // The semantics of LShiftINode include an implicit mask to 0x1F. 3594 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3595 elem_shift = layout_val; 3596 } 3597 3598 // Transition to native address size for all offset calculations: 3599 Node* lengthx = ConvI2X(length); 3600 Node* headerx = ConvI2X(header_size); 3601 #ifdef _LP64 3602 { const TypeLong* tllen = _gvn.find_long_type(lengthx); 3603 if (tllen != NULL && tllen->_lo < 0) { 3604 // Add a manual constraint to a positive range. Cf. array_element_address. 3605 jlong size_max = arrayOopDesc::max_array_length(T_BYTE); 3606 if (size_max > tllen->_hi) size_max = tllen->_hi; 3607 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin); 3608 lengthx = _gvn.transform( new (C) ConvI2LNode(length, tlcon)); 3609 } 3610 } 3611 #endif 3612 3613 // Combine header size (plus rounding) and body size. Then round down. 3614 // This computation cannot overflow, because it is used only in two 3615 // places, one where the length is sharply limited, and the other 3616 // after a successful allocation. 3617 Node* abody = lengthx; 3618 if (elem_shift != NULL) 3619 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) ); 3620 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) ); 3621 if (round_mask != 0) { 3622 Node* mask = MakeConX(~round_mask); 3623 size = _gvn.transform( new(C) AndXNode(size, mask) ); 3624 } 3625 // else if round_mask == 0, the size computation is self-rounding 3626 3627 if (return_size_val != NULL) { 3628 // This is the size 3629 (*return_size_val) = size; 3630 } 3631 3632 // Now generate allocation code 3633 3634 // The entire memory state is needed for slow path of the allocation 3635 // since GC and deoptimization can happened. 3636 Node *mem = reset_memory(); 3637 set_all_memory(mem); // Create new memory state 3638 3639 // Create the AllocateArrayNode and its result projections 3640 AllocateArrayNode* alloc 3641 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3642 control(), mem, i_o(), 3643 size, klass_node, 3644 initial_slow_test, 3645 length); 3646 3647 // Cast to correct type. Note that the klass_node may be constant or not, 3648 // and in the latter case the actual array type will be inexact also. 3649 // (This happens via a non-constant argument to inline_native_newArray.) 3650 // In any case, the value of klass_node provides the desired array type. 3651 const TypeInt* length_type = _gvn.find_int_type(length); 3652 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3653 if (ary_type->isa_aryptr() && length_type != NULL) { 3654 // Try to get a better type than POS for the size 3655 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3656 } 3657 3658 Node* javaoop = set_output_for_allocation(alloc, ary_type); 3659 3660 // Cast length on remaining path to be as narrow as possible 3661 if (map()->find_edge(length) >= 0) { 3662 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3663 if (ccast != length) { 3664 _gvn.set_type_bottom(ccast); 3665 record_for_igvn(ccast); 3666 replace_in_map(length, ccast); 3667 } 3668 } 3669 3670 return javaoop; 3671 } 3672 3673 // The following "Ideal_foo" functions are placed here because they recognize 3674 // the graph shapes created by the functions immediately above. 3675 3676 //---------------------------Ideal_allocation---------------------------------- 3677 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3678 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3679 if (ptr == NULL) { // reduce dumb test in callers 3680 return NULL; 3681 } 3682 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3683 ptr = ptr->in(1); 3684 if (ptr == NULL) return NULL; 3685 } 3686 // Return NULL for allocations with several casts: 3687 // j.l.reflect.Array.newInstance(jobject, jint) 3688 // Object.clone() 3689 // to keep more precise type from last cast. 3690 if (ptr->is_Proj()) { 3691 Node* allo = ptr->in(0); 3692 if (allo != NULL && allo->is_Allocate()) { 3693 return allo->as_Allocate(); 3694 } 3695 } 3696 // Report failure to match. 3697 return NULL; 3698 } 3699 3700 // Fancy version which also strips off an offset (and reports it to caller). 3701 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3702 intptr_t& offset) { 3703 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3704 if (base == NULL) return NULL; 3705 return Ideal_allocation(base, phase); 3706 } 3707 3708 // Trace Initialize <- Proj[Parm] <- Allocate 3709 AllocateNode* InitializeNode::allocation() { 3710 Node* rawoop = in(InitializeNode::RawAddress); 3711 if (rawoop->is_Proj()) { 3712 Node* alloc = rawoop->in(0); 3713 if (alloc->is_Allocate()) { 3714 return alloc->as_Allocate(); 3715 } 3716 } 3717 return NULL; 3718 } 3719 3720 // Trace Allocate -> Proj[Parm] -> Initialize 3721 InitializeNode* AllocateNode::initialization() { 3722 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3723 if (rawoop == NULL) return NULL; 3724 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3725 Node* init = rawoop->fast_out(i); 3726 if (init->is_Initialize()) { 3727 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3728 return init->as_Initialize(); 3729 } 3730 } 3731 return NULL; 3732 } 3733 3734 //----------------------------- loop predicates --------------------------- 3735 3736 //------------------------------add_predicate_impl---------------------------- 3737 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3738 // Too many traps seen? 3739 if (too_many_traps(reason)) { 3740 #ifdef ASSERT 3741 if (TraceLoopPredicate) { 3742 int tc = C->trap_count(reason); 3743 tty->print("too many traps=%s tcount=%d in ", 3744 Deoptimization::trap_reason_name(reason), tc); 3745 method()->print(); // which method has too many predicate traps 3746 tty->cr(); 3747 } 3748 #endif 3749 // We cannot afford to take more traps here, 3750 // do not generate predicate. 3751 return; 3752 } 3753 3754 Node *cont = _gvn.intcon(1); 3755 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont)); 3756 Node *bol = _gvn.transform(new (C) Conv2BNode(opq)); 3757 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3758 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff)); 3759 C->add_predicate_opaq(opq); 3760 { 3761 PreserveJVMState pjvms(this); 3762 set_control(iffalse); 3763 inc_sp(nargs); 3764 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3765 } 3766 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff)); 3767 set_control(iftrue); 3768 } 3769 3770 //------------------------------add_predicate--------------------------------- 3771 void GraphKit::add_predicate(int nargs) { 3772 if (UseLoopPredicate) { 3773 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3774 } 3775 // loop's limit check predicate should be near the loop. 3776 if (LoopLimitCheck) { 3777 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3778 } 3779 } 3780 3781 //----------------------------- store barriers ---------------------------- 3782 #define __ ideal. 3783 3784 void GraphKit::sync_kit(IdealKit& ideal) { 3785 set_all_memory(__ merged_memory()); 3786 set_i_o(__ i_o()); 3787 set_control(__ ctrl()); 3788 } 3789 3790 void GraphKit::final_sync(IdealKit& ideal) { 3791 // Final sync IdealKit and graphKit. 3792 sync_kit(ideal); 3793 } 3794 3795 // vanilla/CMS post barrier 3796 // Insert a write-barrier store. This is to let generational GC work; we have 3797 // to flag all oop-stores before the next GC point. 3798 void GraphKit::write_barrier_post(Node* oop_store, 3799 Node* obj, 3800 Node* adr, 3801 uint adr_idx, 3802 Node* val, 3803 bool use_precise) { 3804 // No store check needed if we're storing a NULL or an old object 3805 // (latter case is probably a string constant). The concurrent 3806 // mark sweep garbage collector, however, needs to have all nonNull 3807 // oop updates flagged via card-marks. 3808 if (val != NULL && val->is_Con()) { 3809 // must be either an oop or NULL 3810 const Type* t = val->bottom_type(); 3811 if (t == TypePtr::NULL_PTR || t == Type::TOP) 3812 // stores of null never (?) need barriers 3813 return; 3814 } 3815 3816 if (use_ReduceInitialCardMarks() 3817 && obj == just_allocated_object(control())) { 3818 // We can skip marks on a freshly-allocated object in Eden. 3819 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp. 3820 // That routine informs GC to take appropriate compensating steps, 3821 // upon a slow-path allocation, so as to make this card-mark 3822 // elision safe. 3823 return; 3824 } 3825 3826 if (!use_precise) { 3827 // All card marks for a (non-array) instance are in one place: 3828 adr = obj; 3829 } 3830 // (Else it's an array (or unknown), and we want more precise card marks.) 3831 assert(adr != NULL, ""); 3832 3833 IdealKit ideal(this, true); 3834 3835 // Convert the pointer to an int prior to doing math on it 3836 Node* cast = __ CastPX(__ ctrl(), adr); 3837 3838 // Divide by card size 3839 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef, 3840 "Only one we handle so far."); 3841 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 3842 3843 // Combine card table base and card offset 3844 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset ); 3845 3846 // Get the alias_index for raw card-mark memory 3847 int adr_type = Compile::AliasIdxRaw; 3848 Node* zero = __ ConI(0); // Dirty card value 3849 BasicType bt = T_BYTE; 3850 3851 if (UseCondCardMark) { 3852 // The classic GC reference write barrier is typically implemented 3853 // as a store into the global card mark table. Unfortunately 3854 // unconditional stores can result in false sharing and excessive 3855 // coherence traffic as well as false transactional aborts. 3856 // UseCondCardMark enables MP "polite" conditional card mark 3857 // stores. In theory we could relax the load from ctrl() to 3858 // no_ctrl, but that doesn't buy much latitude. 3859 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type); 3860 __ if_then(card_val, BoolTest::ne, zero); 3861 } 3862 3863 // Smash zero into card 3864 if( !UseConcMarkSweepGC ) { 3865 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::release); 3866 } else { 3867 // Specialized path for CM store barrier 3868 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type); 3869 } 3870 3871 if (UseCondCardMark) { 3872 __ end_if(); 3873 } 3874 3875 // Final sync IdealKit and GraphKit. 3876 final_sync(ideal); 3877 } 3878 3879 // G1 pre/post barriers 3880 void GraphKit::g1_write_barrier_pre(bool do_load, 3881 Node* obj, 3882 Node* adr, 3883 uint alias_idx, 3884 Node* val, 3885 const TypeOopPtr* val_type, 3886 Node* pre_val, 3887 BasicType bt) { 3888 3889 // Some sanity checks 3890 // Note: val is unused in this routine. 3891 3892 if (do_load) { 3893 // We need to generate the load of the previous value 3894 assert(obj != NULL, "must have a base"); 3895 assert(adr != NULL, "where are loading from?"); 3896 assert(pre_val == NULL, "loaded already?"); 3897 assert(val_type != NULL, "need a type"); 3898 } else { 3899 // In this case both val_type and alias_idx are unused. 3900 assert(pre_val != NULL, "must be loaded already"); 3901 // Nothing to be done if pre_val is null. 3902 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return; 3903 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here"); 3904 } 3905 assert(bt == T_OBJECT, "or we shouldn't be here"); 3906 3907 IdealKit ideal(this, true); 3908 3909 Node* tls = __ thread(); // ThreadLocalStorage 3910 3911 Node* no_ctrl = NULL; 3912 Node* no_base = __ top(); 3913 Node* zero = __ ConI(0); 3914 Node* zeroX = __ ConX(0); 3915 3916 float likely = PROB_LIKELY(0.999); 3917 float unlikely = PROB_UNLIKELY(0.999); 3918 3919 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE; 3920 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width"); 3921 3922 // Offsets into the thread 3923 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648 3924 PtrQueue::byte_offset_of_active()); 3925 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656 3926 PtrQueue::byte_offset_of_index()); 3927 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652 3928 PtrQueue::byte_offset_of_buf()); 3929 3930 // Now the actual pointers into the thread 3931 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset)); 3932 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 3933 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 3934 3935 // Now some of the values 3936 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw); 3937 3938 // if (!marking) 3939 __ if_then(marking, BoolTest::ne, zero, unlikely); { 3940 BasicType index_bt = TypeX_X->basic_type(); 3941 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size."); 3942 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw); 3943 3944 if (do_load) { 3945 // load original value 3946 // alias_idx correct?? 3947 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx); 3948 } 3949 3950 // if (pre_val != NULL) 3951 __ if_then(pre_val, BoolTest::ne, null()); { 3952 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 3953 3954 // is the queue for this thread full? 3955 __ if_then(index, BoolTest::ne, zeroX, likely); { 3956 3957 // decrement the index 3958 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 3959 3960 // Now get the buffer location we will log the previous value into and store it 3961 Node *log_addr = __ AddP(no_base, buffer, next_index); 3962 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered); 3963 // update the index 3964 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered); 3965 3966 } __ else_(); { 3967 3968 // logging buffer is full, call the runtime 3969 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type(); 3970 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls); 3971 } __ end_if(); // (!index) 3972 } __ end_if(); // (pre_val != NULL) 3973 } __ end_if(); // (!marking) 3974 3975 // Final sync IdealKit and GraphKit. 3976 final_sync(ideal); 3977 } 3978 3979 // 3980 // Update the card table and add card address to the queue 3981 // 3982 void GraphKit::g1_mark_card(IdealKit& ideal, 3983 Node* card_adr, 3984 Node* oop_store, 3985 uint oop_alias_idx, 3986 Node* index, 3987 Node* index_adr, 3988 Node* buffer, 3989 const TypeFunc* tf) { 3990 3991 Node* zero = __ ConI(0); 3992 Node* zeroX = __ ConX(0); 3993 Node* no_base = __ top(); 3994 BasicType card_bt = T_BYTE; 3995 // Smash zero into card. MUST BE ORDERED WRT TO STORE 3996 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw); 3997 3998 // Now do the queue work 3999 __ if_then(index, BoolTest::ne, zeroX); { 4000 4001 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t)))); 4002 Node* log_addr = __ AddP(no_base, buffer, next_index); 4003 4004 // Order, see storeCM. 4005 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 4006 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered); 4007 4008 } __ else_(); { 4009 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread()); 4010 } __ end_if(); 4011 4012 } 4013 4014 void GraphKit::g1_write_barrier_post(Node* oop_store, 4015 Node* obj, 4016 Node* adr, 4017 uint alias_idx, 4018 Node* val, 4019 BasicType bt, 4020 bool use_precise) { 4021 // If we are writing a NULL then we need no post barrier 4022 4023 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) { 4024 // Must be NULL 4025 const Type* t = val->bottom_type(); 4026 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL"); 4027 // No post barrier if writing NULLx 4028 return; 4029 } 4030 4031 if (!use_precise) { 4032 // All card marks for a (non-array) instance are in one place: 4033 adr = obj; 4034 } 4035 // (Else it's an array (or unknown), and we want more precise card marks.) 4036 assert(adr != NULL, ""); 4037 4038 IdealKit ideal(this, true); 4039 4040 Node* tls = __ thread(); // ThreadLocalStorage 4041 4042 Node* no_base = __ top(); 4043 float likely = PROB_LIKELY(0.999); 4044 float unlikely = PROB_UNLIKELY(0.999); 4045 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val()); 4046 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val()); 4047 Node* zeroX = __ ConX(0); 4048 4049 // Get the alias_index for raw card-mark memory 4050 const TypePtr* card_type = TypeRawPtr::BOTTOM; 4051 4052 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type(); 4053 4054 // Offsets into the thread 4055 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4056 PtrQueue::byte_offset_of_index()); 4057 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() + 4058 PtrQueue::byte_offset_of_buf()); 4059 4060 // Pointers into the thread 4061 4062 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset)); 4063 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset)); 4064 4065 // Now some values 4066 // Use ctrl to avoid hoisting these values past a safepoint, which could 4067 // potentially reset these fields in the JavaThread. 4068 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw); 4069 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw); 4070 4071 // Convert the store obj pointer to an int prior to doing math on it 4072 // Must use ctrl to prevent "integerized oop" existing across safepoint 4073 Node* cast = __ CastPX(__ ctrl(), adr); 4074 4075 // Divide pointer by card size 4076 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) ); 4077 4078 // Combine card table base and card offset 4079 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset ); 4080 4081 // If we know the value being stored does it cross regions? 4082 4083 if (val != NULL) { 4084 // Does the store cause us to cross regions? 4085 4086 // Should be able to do an unsigned compare of region_size instead of 4087 // and extra shift. Do we have an unsigned compare?? 4088 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes); 4089 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes)); 4090 4091 // if (xor_res == 0) same region so skip 4092 __ if_then(xor_res, BoolTest::ne, zeroX); { 4093 4094 // No barrier if we are storing a NULL 4095 __ if_then(val, BoolTest::ne, null(), unlikely); { 4096 4097 // Ok must mark the card if not already dirty 4098 4099 // load the original value of the card 4100 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4101 4102 __ if_then(card_val, BoolTest::ne, young_card); { 4103 sync_kit(ideal); 4104 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier. 4105 insert_mem_bar(Op_MemBarVolatile, oop_store); 4106 __ sync_kit(this); 4107 4108 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw); 4109 __ if_then(card_val_reload, BoolTest::ne, dirty_card); { 4110 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4111 } __ end_if(); 4112 } __ end_if(); 4113 } __ end_if(); 4114 } __ end_if(); 4115 } else { 4116 // Object.clone() instrinsic uses this path. 4117 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf); 4118 } 4119 4120 // Final sync IdealKit and GraphKit. 4121 final_sync(ideal); 4122 } 4123 #undef __ 4124 4125 4126 4127 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) { 4128 if (java_lang_String::has_offset_field()) { 4129 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4130 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4131 false, NULL, 0); 4132 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4133 int offset_field_idx = C->get_alias_index(offset_field_type); 4134 return make_load(ctrl, 4135 basic_plus_adr(str, str, offset_offset), 4136 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered); 4137 } else { 4138 return intcon(0); 4139 } 4140 } 4141 4142 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 4143 if (java_lang_String::has_count_field()) { 4144 int count_offset = java_lang_String::count_offset_in_bytes(); 4145 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4146 false, NULL, 0); 4147 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4148 int count_field_idx = C->get_alias_index(count_field_type); 4149 return make_load(ctrl, 4150 basic_plus_adr(str, str, count_offset), 4151 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered); 4152 } else { 4153 return load_array_length(load_String_value(ctrl, str)); 4154 } 4155 } 4156 4157 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 4158 int value_offset = java_lang_String::value_offset_in_bytes(); 4159 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4160 false, NULL, 0); 4161 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4162 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 4163 TypeAry::make(TypeInt::CHAR,TypeInt::POS), 4164 ciTypeArrayKlass::make(T_CHAR), true, 0); 4165 int value_field_idx = C->get_alias_index(value_field_type); 4166 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset), 4167 value_type, T_OBJECT, value_field_idx, MemNode::unordered); 4168 // String.value field is known to be @Stable. 4169 if (UseImplicitStableValues) { 4170 load = cast_array_to_stable(load, value_type); 4171 } 4172 return load; 4173 } 4174 4175 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) { 4176 int offset_offset = java_lang_String::offset_offset_in_bytes(); 4177 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4178 false, NULL, 0); 4179 const TypePtr* offset_field_type = string_type->add_offset(offset_offset); 4180 int offset_field_idx = C->get_alias_index(offset_field_type); 4181 store_to_memory(ctrl, basic_plus_adr(str, offset_offset), 4182 value, T_INT, offset_field_idx, MemNode::unordered); 4183 } 4184 4185 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 4186 int value_offset = java_lang_String::value_offset_in_bytes(); 4187 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4188 false, NULL, 0); 4189 const TypePtr* value_field_type = string_type->add_offset(value_offset); 4190 4191 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 4192 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered); 4193 } 4194 4195 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) { 4196 int count_offset = java_lang_String::count_offset_in_bytes(); 4197 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 4198 false, NULL, 0); 4199 const TypePtr* count_field_type = string_type->add_offset(count_offset); 4200 int count_field_idx = C->get_alias_index(count_field_type); 4201 store_to_memory(ctrl, basic_plus_adr(str, count_offset), 4202 value, T_INT, count_field_idx, MemNode::unordered); 4203 } 4204 4205 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 4206 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 4207 // assumption of CCP analysis. 4208 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true))); 4209 }