1 /* 2 * Copyright (c) 2001, 2018, 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 "ci/ciUtilities.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/g1/heapRegion.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/c2/barrierSetC2.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/graphKit.hpp" 37 #include "opto/idealKit.hpp" 38 #include "opto/intrinsicnode.hpp" 39 #include "opto/locknode.hpp" 40 #include "opto/machnode.hpp" 41 #include "opto/opaquenode.hpp" 42 #include "opto/parse.hpp" 43 #include "opto/rootnode.hpp" 44 #include "opto/runtime.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 //----------------------------GraphKit----------------------------------------- 49 // Main utility constructor. 50 GraphKit::GraphKit(JVMState* jvms) 51 : Phase(Phase::Parser), 52 _env(C->env()), 53 _gvn(*C->initial_gvn()), 54 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 55 { 56 _exceptions = jvms->map()->next_exception(); 57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 58 set_jvms(jvms); 59 } 60 61 // Private constructor for parser. 62 GraphKit::GraphKit() 63 : Phase(Phase::Parser), 64 _env(C->env()), 65 _gvn(*C->initial_gvn()), 66 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) 67 { 68 _exceptions = NULL; 69 set_map(NULL); 70 debug_only(_sp = -99); 71 debug_only(set_bci(-99)); 72 } 73 74 75 76 //---------------------------clean_stack--------------------------------------- 77 // Clear away rubbish from the stack area of the JVM state. 78 // This destroys any arguments that may be waiting on the stack. 79 void GraphKit::clean_stack(int from_sp) { 80 SafePointNode* map = this->map(); 81 JVMState* jvms = this->jvms(); 82 int stk_size = jvms->stk_size(); 83 int stkoff = jvms->stkoff(); 84 Node* top = this->top(); 85 for (int i = from_sp; i < stk_size; i++) { 86 if (map->in(stkoff + i) != top) { 87 map->set_req(stkoff + i, top); 88 } 89 } 90 } 91 92 93 //--------------------------------sync_jvms----------------------------------- 94 // Make sure our current jvms agrees with our parse state. 95 JVMState* GraphKit::sync_jvms() const { 96 JVMState* jvms = this->jvms(); 97 jvms->set_bci(bci()); // Record the new bci in the JVMState 98 jvms->set_sp(sp()); // Record the new sp in the JVMState 99 assert(jvms_in_sync(), "jvms is now in sync"); 100 return jvms; 101 } 102 103 //--------------------------------sync_jvms_for_reexecute--------------------- 104 // Make sure our current jvms agrees with our parse state. This version 105 // uses the reexecute_sp for reexecuting bytecodes. 106 JVMState* GraphKit::sync_jvms_for_reexecute() { 107 JVMState* jvms = this->jvms(); 108 jvms->set_bci(bci()); // Record the new bci in the JVMState 109 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 110 return jvms; 111 } 112 113 #ifdef ASSERT 114 bool GraphKit::jvms_in_sync() const { 115 Parse* parse = is_Parse(); 116 if (parse == NULL) { 117 if (bci() != jvms()->bci()) return false; 118 if (sp() != (int)jvms()->sp()) return false; 119 return true; 120 } 121 if (jvms()->method() != parse->method()) return false; 122 if (jvms()->bci() != parse->bci()) return false; 123 int jvms_sp = jvms()->sp(); 124 if (jvms_sp != parse->sp()) return false; 125 int jvms_depth = jvms()->depth(); 126 if (jvms_depth != parse->depth()) return false; 127 return true; 128 } 129 130 // Local helper checks for special internal merge points 131 // used to accumulate and merge exception states. 132 // They are marked by the region's in(0) edge being the map itself. 133 // Such merge points must never "escape" into the parser at large, 134 // until they have been handed to gvn.transform. 135 static bool is_hidden_merge(Node* reg) { 136 if (reg == NULL) return false; 137 if (reg->is_Phi()) { 138 reg = reg->in(0); 139 if (reg == NULL) return false; 140 } 141 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 142 } 143 144 void GraphKit::verify_map() const { 145 if (map() == NULL) return; // null map is OK 146 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 147 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 148 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 149 } 150 151 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 152 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 153 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 154 } 155 #endif 156 157 //---------------------------stop_and_kill_map--------------------------------- 158 // Set _map to NULL, signalling a stop to further bytecode execution. 159 // First smash the current map's control to a constant, to mark it dead. 160 void GraphKit::stop_and_kill_map() { 161 SafePointNode* dead_map = stop(); 162 if (dead_map != NULL) { 163 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 164 assert(dead_map->is_killed(), "must be so marked"); 165 } 166 } 167 168 169 //--------------------------------stopped-------------------------------------- 170 // Tell if _map is NULL, or control is top. 171 bool GraphKit::stopped() { 172 if (map() == NULL) return true; 173 else if (control() == top()) return true; 174 else return false; 175 } 176 177 178 //-----------------------------has_ex_handler---------------------------------- 179 // Tell if this method or any caller method has exception handlers. 180 bool GraphKit::has_ex_handler() { 181 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 182 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 183 return true; 184 } 185 } 186 return false; 187 } 188 189 //------------------------------save_ex_oop------------------------------------ 190 // Save an exception without blowing stack contents or other JVM state. 191 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 192 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 193 ex_map->add_req(ex_oop); 194 debug_only(verify_exception_state(ex_map)); 195 } 196 197 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 198 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 199 Node* ex_oop = ex_map->in(ex_map->req()-1); 200 if (clear_it) ex_map->del_req(ex_map->req()-1); 201 return ex_oop; 202 } 203 204 //-----------------------------saved_ex_oop------------------------------------ 205 // Recover a saved exception from its map. 206 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 207 return common_saved_ex_oop(ex_map, false); 208 } 209 210 //--------------------------clear_saved_ex_oop--------------------------------- 211 // Erase a previously saved exception from its map. 212 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 213 return common_saved_ex_oop(ex_map, true); 214 } 215 216 #ifdef ASSERT 217 //---------------------------has_saved_ex_oop---------------------------------- 218 // Erase a previously saved exception from its map. 219 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 220 return ex_map->req() == ex_map->jvms()->endoff()+1; 221 } 222 #endif 223 224 //-------------------------make_exception_state-------------------------------- 225 // Turn the current JVM state into an exception state, appending the ex_oop. 226 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 227 sync_jvms(); 228 SafePointNode* ex_map = stop(); // do not manipulate this map any more 229 set_saved_ex_oop(ex_map, ex_oop); 230 return ex_map; 231 } 232 233 234 //--------------------------add_exception_state-------------------------------- 235 // Add an exception to my list of exceptions. 236 void GraphKit::add_exception_state(SafePointNode* ex_map) { 237 if (ex_map == NULL || ex_map->control() == top()) { 238 return; 239 } 240 #ifdef ASSERT 241 verify_exception_state(ex_map); 242 if (has_exceptions()) { 243 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 244 } 245 #endif 246 247 // If there is already an exception of exactly this type, merge with it. 248 // In particular, null-checks and other low-level exceptions common up here. 249 Node* ex_oop = saved_ex_oop(ex_map); 250 const Type* ex_type = _gvn.type(ex_oop); 251 if (ex_oop == top()) { 252 // No action needed. 253 return; 254 } 255 assert(ex_type->isa_instptr(), "exception must be an instance"); 256 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 257 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 258 // We check sp also because call bytecodes can generate exceptions 259 // both before and after arguments are popped! 260 if (ex_type2 == ex_type 261 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 262 combine_exception_states(ex_map, e2); 263 return; 264 } 265 } 266 267 // No pre-existing exception of the same type. Chain it on the list. 268 push_exception_state(ex_map); 269 } 270 271 //-----------------------add_exception_states_from----------------------------- 272 void GraphKit::add_exception_states_from(JVMState* jvms) { 273 SafePointNode* ex_map = jvms->map()->next_exception(); 274 if (ex_map != NULL) { 275 jvms->map()->set_next_exception(NULL); 276 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 277 next_map = ex_map->next_exception(); 278 ex_map->set_next_exception(NULL); 279 add_exception_state(ex_map); 280 } 281 } 282 } 283 284 //-----------------------transfer_exceptions_into_jvms------------------------- 285 JVMState* GraphKit::transfer_exceptions_into_jvms() { 286 if (map() == NULL) { 287 // We need a JVMS to carry the exceptions, but the map has gone away. 288 // Create a scratch JVMS, cloned from any of the exception states... 289 if (has_exceptions()) { 290 _map = _exceptions; 291 _map = clone_map(); 292 _map->set_next_exception(NULL); 293 clear_saved_ex_oop(_map); 294 debug_only(verify_map()); 295 } else { 296 // ...or created from scratch 297 JVMState* jvms = new (C) JVMState(_method, NULL); 298 jvms->set_bci(_bci); 299 jvms->set_sp(_sp); 300 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 301 set_jvms(jvms); 302 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 303 set_all_memory(top()); 304 while (map()->req() < jvms->endoff()) map()->add_req(top()); 305 } 306 // (This is a kludge, in case you didn't notice.) 307 set_control(top()); 308 } 309 JVMState* jvms = sync_jvms(); 310 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 311 jvms->map()->set_next_exception(_exceptions); 312 _exceptions = NULL; // done with this set of exceptions 313 return jvms; 314 } 315 316 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 317 assert(is_hidden_merge(dstphi), "must be a special merge node"); 318 assert(is_hidden_merge(srcphi), "must be a special merge node"); 319 uint limit = srcphi->req(); 320 for (uint i = PhiNode::Input; i < limit; i++) { 321 dstphi->add_req(srcphi->in(i)); 322 } 323 } 324 static inline void add_one_req(Node* dstphi, Node* src) { 325 assert(is_hidden_merge(dstphi), "must be a special merge node"); 326 assert(!is_hidden_merge(src), "must not be a special merge node"); 327 dstphi->add_req(src); 328 } 329 330 //-----------------------combine_exception_states------------------------------ 331 // This helper function combines exception states by building phis on a 332 // specially marked state-merging region. These regions and phis are 333 // untransformed, and can build up gradually. The region is marked by 334 // having a control input of its exception map, rather than NULL. Such 335 // regions do not appear except in this function, and in use_exception_state. 336 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 337 if (failing()) return; // dying anyway... 338 JVMState* ex_jvms = ex_map->_jvms; 339 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 340 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 341 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 342 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 343 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 344 assert(ex_map->req() == phi_map->req(), "matching maps"); 345 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 346 Node* hidden_merge_mark = root(); 347 Node* region = phi_map->control(); 348 MergeMemNode* phi_mem = phi_map->merged_memory(); 349 MergeMemNode* ex_mem = ex_map->merged_memory(); 350 if (region->in(0) != hidden_merge_mark) { 351 // The control input is not (yet) a specially-marked region in phi_map. 352 // Make it so, and build some phis. 353 region = new RegionNode(2); 354 _gvn.set_type(region, Type::CONTROL); 355 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 356 region->init_req(1, phi_map->control()); 357 phi_map->set_control(region); 358 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 359 record_for_igvn(io_phi); 360 _gvn.set_type(io_phi, Type::ABIO); 361 phi_map->set_i_o(io_phi); 362 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 363 Node* m = mms.memory(); 364 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 365 record_for_igvn(m_phi); 366 _gvn.set_type(m_phi, Type::MEMORY); 367 mms.set_memory(m_phi); 368 } 369 } 370 371 // Either or both of phi_map and ex_map might already be converted into phis. 372 Node* ex_control = ex_map->control(); 373 // if there is special marking on ex_map also, we add multiple edges from src 374 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 375 // how wide was the destination phi_map, originally? 376 uint orig_width = region->req(); 377 378 if (add_multiple) { 379 add_n_reqs(region, ex_control); 380 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 381 } else { 382 // ex_map has no merges, so we just add single edges everywhere 383 add_one_req(region, ex_control); 384 add_one_req(phi_map->i_o(), ex_map->i_o()); 385 } 386 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 387 if (mms.is_empty()) { 388 // get a copy of the base memory, and patch some inputs into it 389 const TypePtr* adr_type = mms.adr_type(C); 390 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 391 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 392 mms.set_memory(phi); 393 // Prepare to append interesting stuff onto the newly sliced phi: 394 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 395 } 396 // Append stuff from ex_map: 397 if (add_multiple) { 398 add_n_reqs(mms.memory(), mms.memory2()); 399 } else { 400 add_one_req(mms.memory(), mms.memory2()); 401 } 402 } 403 uint limit = ex_map->req(); 404 for (uint i = TypeFunc::Parms; i < limit; i++) { 405 // Skip everything in the JVMS after tos. (The ex_oop follows.) 406 if (i == tos) i = ex_jvms->monoff(); 407 Node* src = ex_map->in(i); 408 Node* dst = phi_map->in(i); 409 if (src != dst) { 410 PhiNode* phi; 411 if (dst->in(0) != region) { 412 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 413 record_for_igvn(phi); 414 _gvn.set_type(phi, phi->type()); 415 phi_map->set_req(i, dst); 416 // Prepare to append interesting stuff onto the new phi: 417 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 418 } else { 419 assert(dst->is_Phi(), "nobody else uses a hidden region"); 420 phi = dst->as_Phi(); 421 } 422 if (add_multiple && src->in(0) == ex_control) { 423 // Both are phis. 424 add_n_reqs(dst, src); 425 } else { 426 while (dst->req() < region->req()) add_one_req(dst, src); 427 } 428 const Type* srctype = _gvn.type(src); 429 if (phi->type() != srctype) { 430 const Type* dsttype = phi->type()->meet_speculative(srctype); 431 if (phi->type() != dsttype) { 432 phi->set_type(dsttype); 433 _gvn.set_type(phi, dsttype); 434 } 435 } 436 } 437 } 438 phi_map->merge_replaced_nodes_with(ex_map); 439 } 440 441 //--------------------------use_exception_state-------------------------------- 442 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 443 if (failing()) { stop(); return top(); } 444 Node* region = phi_map->control(); 445 Node* hidden_merge_mark = root(); 446 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 447 Node* ex_oop = clear_saved_ex_oop(phi_map); 448 if (region->in(0) == hidden_merge_mark) { 449 // Special marking for internal ex-states. Process the phis now. 450 region->set_req(0, region); // now it's an ordinary region 451 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 452 // Note: Setting the jvms also sets the bci and sp. 453 set_control(_gvn.transform(region)); 454 uint tos = jvms()->stkoff() + sp(); 455 for (uint i = 1; i < tos; i++) { 456 Node* x = phi_map->in(i); 457 if (x->in(0) == region) { 458 assert(x->is_Phi(), "expected a special phi"); 459 phi_map->set_req(i, _gvn.transform(x)); 460 } 461 } 462 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 463 Node* x = mms.memory(); 464 if (x->in(0) == region) { 465 assert(x->is_Phi(), "nobody else uses a hidden region"); 466 mms.set_memory(_gvn.transform(x)); 467 } 468 } 469 if (ex_oop->in(0) == region) { 470 assert(ex_oop->is_Phi(), "expected a special phi"); 471 ex_oop = _gvn.transform(ex_oop); 472 } 473 } else { 474 set_jvms(phi_map->jvms()); 475 } 476 477 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 478 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 479 return ex_oop; 480 } 481 482 //---------------------------------java_bc------------------------------------- 483 Bytecodes::Code GraphKit::java_bc() const { 484 ciMethod* method = this->method(); 485 int bci = this->bci(); 486 if (method != NULL && bci != InvocationEntryBci) 487 return method->java_code_at_bci(bci); 488 else 489 return Bytecodes::_illegal; 490 } 491 492 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 493 bool must_throw) { 494 // if the exception capability is set, then we will generate code 495 // to check the JavaThread.should_post_on_exceptions flag to see 496 // if we actually need to report exception events (for this 497 // thread). If we don't need to report exception events, we will 498 // take the normal fast path provided by add_exception_events. If 499 // exception event reporting is enabled for this thread, we will 500 // take the uncommon_trap in the BuildCutout below. 501 502 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 503 Node* jthread = _gvn.transform(new ThreadLocalNode()); 504 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 505 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 506 507 // Test the should_post_on_exceptions_flag vs. 0 508 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 509 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 510 511 // Branch to slow_path if should_post_on_exceptions_flag was true 512 { BuildCutout unless(this, tst, PROB_MAX); 513 // Do not try anything fancy if we're notifying the VM on every throw. 514 // Cf. case Bytecodes::_athrow in parse2.cpp. 515 uncommon_trap(reason, Deoptimization::Action_none, 516 (ciKlass*)NULL, (char*)NULL, must_throw); 517 } 518 519 } 520 521 //------------------------------builtin_throw---------------------------------- 522 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 523 bool must_throw = true; 524 525 if (env()->jvmti_can_post_on_exceptions()) { 526 // check if we must post exception events, take uncommon trap if so 527 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 528 // here if should_post_on_exceptions is false 529 // continue on with the normal codegen 530 } 531 532 // If this particular condition has not yet happened at this 533 // bytecode, then use the uncommon trap mechanism, and allow for 534 // a future recompilation if several traps occur here. 535 // If the throw is hot, try to use a more complicated inline mechanism 536 // which keeps execution inside the compiled code. 537 bool treat_throw_as_hot = false; 538 ciMethodData* md = method()->method_data(); 539 540 if (ProfileTraps) { 541 if (too_many_traps(reason)) { 542 treat_throw_as_hot = true; 543 } 544 // (If there is no MDO at all, assume it is early in 545 // execution, and that any deopts are part of the 546 // startup transient, and don't need to be remembered.) 547 548 // Also, if there is a local exception handler, treat all throws 549 // as hot if there has been at least one in this method. 550 if (C->trap_count(reason) != 0 551 && method()->method_data()->trap_count(reason) != 0 552 && has_ex_handler()) { 553 treat_throw_as_hot = true; 554 } 555 } 556 557 // If this throw happens frequently, an uncommon trap might cause 558 // a performance pothole. If there is a local exception handler, 559 // and if this particular bytecode appears to be deoptimizing often, 560 // let us handle the throw inline, with a preconstructed instance. 561 // Note: If the deopt count has blown up, the uncommon trap 562 // runtime is going to flush this nmethod, not matter what. 563 if (treat_throw_as_hot 564 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 565 // If the throw is local, we use a pre-existing instance and 566 // punt on the backtrace. This would lead to a missing backtrace 567 // (a repeat of 4292742) if the backtrace object is ever asked 568 // for its backtrace. 569 // Fixing this remaining case of 4292742 requires some flavor of 570 // escape analysis. Leave that for the future. 571 ciInstance* ex_obj = NULL; 572 switch (reason) { 573 case Deoptimization::Reason_null_check: 574 ex_obj = env()->NullPointerException_instance(); 575 break; 576 case Deoptimization::Reason_div0_check: 577 ex_obj = env()->ArithmeticException_instance(); 578 break; 579 case Deoptimization::Reason_range_check: 580 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 581 break; 582 case Deoptimization::Reason_class_check: 583 if (java_bc() == Bytecodes::_aastore) { 584 ex_obj = env()->ArrayStoreException_instance(); 585 } else { 586 ex_obj = env()->ClassCastException_instance(); 587 } 588 break; 589 default: 590 break; 591 } 592 if (failing()) { stop(); return; } // exception allocation might fail 593 if (ex_obj != NULL) { 594 // Cheat with a preallocated exception object. 595 if (C->log() != NULL) 596 C->log()->elem("hot_throw preallocated='1' reason='%s'", 597 Deoptimization::trap_reason_name(reason)); 598 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 599 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 600 601 // Clear the detail message of the preallocated exception object. 602 // Weblogic sometimes mutates the detail message of exceptions 603 // using reflection. 604 int offset = java_lang_Throwable::get_detailMessage_offset(); 605 const TypePtr* adr_typ = ex_con->add_offset(offset); 606 607 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 608 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 609 Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP); 610 611 add_exception_state(make_exception_state(ex_node)); 612 return; 613 } 614 } 615 616 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 617 // It won't be much cheaper than bailing to the interp., since we'll 618 // have to pass up all the debug-info, and the runtime will have to 619 // create the stack trace. 620 621 // Usual case: Bail to interpreter. 622 // Reserve the right to recompile if we haven't seen anything yet. 623 624 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 625 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 626 if (treat_throw_as_hot 627 && (method()->method_data()->trap_recompiled_at(bci(), m) 628 || C->too_many_traps(reason))) { 629 // We cannot afford to take more traps here. Suffer in the interpreter. 630 if (C->log() != NULL) 631 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 632 Deoptimization::trap_reason_name(reason), 633 C->trap_count(reason)); 634 action = Deoptimization::Action_none; 635 } 636 637 // "must_throw" prunes the JVM state to include only the stack, if there 638 // are no local exception handlers. This should cut down on register 639 // allocation time and code size, by drastically reducing the number 640 // of in-edges on the call to the uncommon trap. 641 642 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 643 } 644 645 646 //----------------------------PreserveJVMState--------------------------------- 647 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 648 debug_only(kit->verify_map()); 649 _kit = kit; 650 _map = kit->map(); // preserve the map 651 _sp = kit->sp(); 652 kit->set_map(clone_map ? kit->clone_map() : NULL); 653 #ifdef ASSERT 654 _bci = kit->bci(); 655 Parse* parser = kit->is_Parse(); 656 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 657 _block = block; 658 #endif 659 } 660 PreserveJVMState::~PreserveJVMState() { 661 GraphKit* kit = _kit; 662 #ifdef ASSERT 663 assert(kit->bci() == _bci, "bci must not shift"); 664 Parse* parser = kit->is_Parse(); 665 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 666 assert(block == _block, "block must not shift"); 667 #endif 668 kit->set_map(_map); 669 kit->set_sp(_sp); 670 } 671 672 673 //-----------------------------BuildCutout------------------------------------- 674 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 675 : PreserveJVMState(kit) 676 { 677 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 678 SafePointNode* outer_map = _map; // preserved map is caller's 679 SafePointNode* inner_map = kit->map(); 680 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 681 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 682 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 683 } 684 BuildCutout::~BuildCutout() { 685 GraphKit* kit = _kit; 686 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 687 } 688 689 //---------------------------PreserveReexecuteState---------------------------- 690 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 691 assert(!kit->stopped(), "must call stopped() before"); 692 _kit = kit; 693 _sp = kit->sp(); 694 _reexecute = kit->jvms()->_reexecute; 695 } 696 PreserveReexecuteState::~PreserveReexecuteState() { 697 if (_kit->stopped()) return; 698 _kit->jvms()->_reexecute = _reexecute; 699 _kit->set_sp(_sp); 700 } 701 702 //------------------------------clone_map-------------------------------------- 703 // Implementation of PreserveJVMState 704 // 705 // Only clone_map(...) here. If this function is only used in the 706 // PreserveJVMState class we may want to get rid of this extra 707 // function eventually and do it all there. 708 709 SafePointNode* GraphKit::clone_map() { 710 if (map() == NULL) return NULL; 711 712 // Clone the memory edge first 713 Node* mem = MergeMemNode::make(map()->memory()); 714 gvn().set_type_bottom(mem); 715 716 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 717 JVMState* jvms = this->jvms(); 718 JVMState* clonejvms = jvms->clone_shallow(C); 719 clonemap->set_memory(mem); 720 clonemap->set_jvms(clonejvms); 721 clonejvms->set_map(clonemap); 722 record_for_igvn(clonemap); 723 gvn().set_type_bottom(clonemap); 724 return clonemap; 725 } 726 727 728 //-----------------------------set_map_clone----------------------------------- 729 void GraphKit::set_map_clone(SafePointNode* m) { 730 _map = m; 731 _map = clone_map(); 732 _map->set_next_exception(NULL); 733 debug_only(verify_map()); 734 } 735 736 737 //----------------------------kill_dead_locals--------------------------------- 738 // Detect any locals which are known to be dead, and force them to top. 739 void GraphKit::kill_dead_locals() { 740 // Consult the liveness information for the locals. If any 741 // of them are unused, then they can be replaced by top(). This 742 // should help register allocation time and cut down on the size 743 // of the deoptimization information. 744 745 // This call is made from many of the bytecode handling 746 // subroutines called from the Big Switch in do_one_bytecode. 747 // Every bytecode which might include a slow path is responsible 748 // for killing its dead locals. The more consistent we 749 // are about killing deads, the fewer useless phis will be 750 // constructed for them at various merge points. 751 752 // bci can be -1 (InvocationEntryBci). We return the entry 753 // liveness for the method. 754 755 if (method() == NULL || method()->code_size() == 0) { 756 // We are building a graph for a call to a native method. 757 // All locals are live. 758 return; 759 } 760 761 ResourceMark rm; 762 763 // Consult the liveness information for the locals. If any 764 // of them are unused, then they can be replaced by top(). This 765 // should help register allocation time and cut down on the size 766 // of the deoptimization information. 767 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 768 769 int len = (int)live_locals.size(); 770 assert(len <= jvms()->loc_size(), "too many live locals"); 771 for (int local = 0; local < len; local++) { 772 if (!live_locals.at(local)) { 773 set_local(local, top()); 774 } 775 } 776 } 777 778 #ifdef ASSERT 779 //-------------------------dead_locals_are_killed------------------------------ 780 // Return true if all dead locals are set to top in the map. 781 // Used to assert "clean" debug info at various points. 782 bool GraphKit::dead_locals_are_killed() { 783 if (method() == NULL || method()->code_size() == 0) { 784 // No locals need to be dead, so all is as it should be. 785 return true; 786 } 787 788 // Make sure somebody called kill_dead_locals upstream. 789 ResourceMark rm; 790 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 791 if (jvms->loc_size() == 0) continue; // no locals to consult 792 SafePointNode* map = jvms->map(); 793 ciMethod* method = jvms->method(); 794 int bci = jvms->bci(); 795 if (jvms == this->jvms()) { 796 bci = this->bci(); // it might not yet be synched 797 } 798 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 799 int len = (int)live_locals.size(); 800 if (!live_locals.is_valid() || len == 0) 801 // This method is trivial, or is poisoned by a breakpoint. 802 return true; 803 assert(len == jvms->loc_size(), "live map consistent with locals map"); 804 for (int local = 0; local < len; local++) { 805 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 806 if (PrintMiscellaneous && (Verbose || WizardMode)) { 807 tty->print_cr("Zombie local %d: ", local); 808 jvms->dump(); 809 } 810 return false; 811 } 812 } 813 } 814 return true; 815 } 816 817 #endif //ASSERT 818 819 // Helper function for enforcing certain bytecodes to reexecute if 820 // deoptimization happens 821 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 822 ciMethod* cur_method = jvms->method(); 823 int cur_bci = jvms->bci(); 824 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 825 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 826 return Interpreter::bytecode_should_reexecute(code) || 827 (is_anewarray && code == Bytecodes::_multianewarray); 828 // Reexecute _multianewarray bytecode which was replaced with 829 // sequence of [a]newarray. See Parse::do_multianewarray(). 830 // 831 // Note: interpreter should not have it set since this optimization 832 // is limited by dimensions and guarded by flag so in some cases 833 // multianewarray() runtime calls will be generated and 834 // the bytecode should not be reexecutes (stack will not be reset). 835 } else 836 return false; 837 } 838 839 // Helper function for adding JVMState and debug information to node 840 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 841 // Add the safepoint edges to the call (or other safepoint). 842 843 // Make sure dead locals are set to top. This 844 // should help register allocation time and cut down on the size 845 // of the deoptimization information. 846 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 847 848 // Walk the inline list to fill in the correct set of JVMState's 849 // Also fill in the associated edges for each JVMState. 850 851 // If the bytecode needs to be reexecuted we need to put 852 // the arguments back on the stack. 853 const bool should_reexecute = jvms()->should_reexecute(); 854 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 855 856 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 857 // undefined if the bci is different. This is normal for Parse but it 858 // should not happen for LibraryCallKit because only one bci is processed. 859 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 860 "in LibraryCallKit the reexecute bit should not change"); 861 862 // If we are guaranteed to throw, we can prune everything but the 863 // input to the current bytecode. 864 bool can_prune_locals = false; 865 uint stack_slots_not_pruned = 0; 866 int inputs = 0, depth = 0; 867 if (must_throw) { 868 assert(method() == youngest_jvms->method(), "sanity"); 869 if (compute_stack_effects(inputs, depth)) { 870 can_prune_locals = true; 871 stack_slots_not_pruned = inputs; 872 } 873 } 874 875 if (env()->should_retain_local_variables()) { 876 // At any safepoint, this method can get breakpointed, which would 877 // then require an immediate deoptimization. 878 can_prune_locals = false; // do not prune locals 879 stack_slots_not_pruned = 0; 880 } 881 882 // do not scribble on the input jvms 883 JVMState* out_jvms = youngest_jvms->clone_deep(C); 884 call->set_jvms(out_jvms); // Start jvms list for call node 885 886 // For a known set of bytecodes, the interpreter should reexecute them if 887 // deoptimization happens. We set the reexecute state for them here 888 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 889 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 890 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 891 } 892 893 // Presize the call: 894 DEBUG_ONLY(uint non_debug_edges = call->req()); 895 call->add_req_batch(top(), youngest_jvms->debug_depth()); 896 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 897 898 // Set up edges so that the call looks like this: 899 // Call [state:] ctl io mem fptr retadr 900 // [parms:] parm0 ... parmN 901 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 902 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 903 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 904 // Note that caller debug info precedes callee debug info. 905 906 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 907 uint debug_ptr = call->req(); 908 909 // Loop over the map input edges associated with jvms, add them 910 // to the call node, & reset all offsets to match call node array. 911 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 912 uint debug_end = debug_ptr; 913 uint debug_start = debug_ptr - in_jvms->debug_size(); 914 debug_ptr = debug_start; // back up the ptr 915 916 uint p = debug_start; // walks forward in [debug_start, debug_end) 917 uint j, k, l; 918 SafePointNode* in_map = in_jvms->map(); 919 out_jvms->set_map(call); 920 921 if (can_prune_locals) { 922 assert(in_jvms->method() == out_jvms->method(), "sanity"); 923 // If the current throw can reach an exception handler in this JVMS, 924 // then we must keep everything live that can reach that handler. 925 // As a quick and dirty approximation, we look for any handlers at all. 926 if (in_jvms->method()->has_exception_handlers()) { 927 can_prune_locals = false; 928 } 929 } 930 931 // Add the Locals 932 k = in_jvms->locoff(); 933 l = in_jvms->loc_size(); 934 out_jvms->set_locoff(p); 935 if (!can_prune_locals) { 936 for (j = 0; j < l; j++) 937 call->set_req(p++, in_map->in(k+j)); 938 } else { 939 p += l; // already set to top above by add_req_batch 940 } 941 942 // Add the Expression Stack 943 k = in_jvms->stkoff(); 944 l = in_jvms->sp(); 945 out_jvms->set_stkoff(p); 946 if (!can_prune_locals) { 947 for (j = 0; j < l; j++) 948 call->set_req(p++, in_map->in(k+j)); 949 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 950 // Divide stack into {S0,...,S1}, where S0 is set to top. 951 uint s1 = stack_slots_not_pruned; 952 stack_slots_not_pruned = 0; // for next iteration 953 if (s1 > l) s1 = l; 954 uint s0 = l - s1; 955 p += s0; // skip the tops preinstalled by add_req_batch 956 for (j = s0; j < l; j++) 957 call->set_req(p++, in_map->in(k+j)); 958 } else { 959 p += l; // already set to top above by add_req_batch 960 } 961 962 // Add the Monitors 963 k = in_jvms->monoff(); 964 l = in_jvms->mon_size(); 965 out_jvms->set_monoff(p); 966 for (j = 0; j < l; j++) 967 call->set_req(p++, in_map->in(k+j)); 968 969 // Copy any scalar object fields. 970 k = in_jvms->scloff(); 971 l = in_jvms->scl_size(); 972 out_jvms->set_scloff(p); 973 for (j = 0; j < l; j++) 974 call->set_req(p++, in_map->in(k+j)); 975 976 // Finish the new jvms. 977 out_jvms->set_endoff(p); 978 979 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 980 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 981 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 982 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 983 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 984 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 985 986 // Update the two tail pointers in parallel. 987 out_jvms = out_jvms->caller(); 988 in_jvms = in_jvms->caller(); 989 } 990 991 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 992 993 // Test the correctness of JVMState::debug_xxx accessors: 994 assert(call->jvms()->debug_start() == non_debug_edges, ""); 995 assert(call->jvms()->debug_end() == call->req(), ""); 996 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 997 } 998 999 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 1000 Bytecodes::Code code = java_bc(); 1001 if (code == Bytecodes::_wide) { 1002 code = method()->java_code_at_bci(bci() + 1); 1003 } 1004 1005 BasicType rtype = T_ILLEGAL; 1006 int rsize = 0; 1007 1008 if (code != Bytecodes::_illegal) { 1009 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1010 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1011 if (rtype < T_CONFLICT) 1012 rsize = type2size[rtype]; 1013 } 1014 1015 switch (code) { 1016 case Bytecodes::_illegal: 1017 return false; 1018 1019 case Bytecodes::_ldc: 1020 case Bytecodes::_ldc_w: 1021 case Bytecodes::_ldc2_w: 1022 inputs = 0; 1023 break; 1024 1025 case Bytecodes::_dup: inputs = 1; break; 1026 case Bytecodes::_dup_x1: inputs = 2; break; 1027 case Bytecodes::_dup_x2: inputs = 3; break; 1028 case Bytecodes::_dup2: inputs = 2; break; 1029 case Bytecodes::_dup2_x1: inputs = 3; break; 1030 case Bytecodes::_dup2_x2: inputs = 4; break; 1031 case Bytecodes::_swap: inputs = 2; break; 1032 case Bytecodes::_arraylength: inputs = 1; break; 1033 1034 case Bytecodes::_getstatic: 1035 case Bytecodes::_putstatic: 1036 case Bytecodes::_getfield: 1037 case Bytecodes::_putfield: 1038 { 1039 bool ignored_will_link; 1040 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1041 int size = field->type()->size(); 1042 bool is_get = (depth >= 0), is_static = (depth & 1); 1043 inputs = (is_static ? 0 : 1); 1044 if (is_get) { 1045 depth = size - inputs; 1046 } else { 1047 inputs += size; // putxxx pops the value from the stack 1048 depth = - inputs; 1049 } 1050 } 1051 break; 1052 1053 case Bytecodes::_invokevirtual: 1054 case Bytecodes::_invokespecial: 1055 case Bytecodes::_invokestatic: 1056 case Bytecodes::_invokedynamic: 1057 case Bytecodes::_invokeinterface: 1058 { 1059 bool ignored_will_link; 1060 ciSignature* declared_signature = NULL; 1061 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1062 assert(declared_signature != NULL, "cannot be null"); 1063 inputs = declared_signature->arg_size_for_bc(code); 1064 int size = declared_signature->return_type()->size(); 1065 depth = size - inputs; 1066 } 1067 break; 1068 1069 case Bytecodes::_multianewarray: 1070 { 1071 ciBytecodeStream iter(method()); 1072 iter.reset_to_bci(bci()); 1073 iter.next(); 1074 inputs = iter.get_dimensions(); 1075 assert(rsize == 1, ""); 1076 depth = rsize - inputs; 1077 } 1078 break; 1079 1080 case Bytecodes::_ireturn: 1081 case Bytecodes::_lreturn: 1082 case Bytecodes::_freturn: 1083 case Bytecodes::_dreturn: 1084 case Bytecodes::_areturn: 1085 assert(rsize == -depth, ""); 1086 inputs = rsize; 1087 break; 1088 1089 case Bytecodes::_jsr: 1090 case Bytecodes::_jsr_w: 1091 inputs = 0; 1092 depth = 1; // S.B. depth=1, not zero 1093 break; 1094 1095 default: 1096 // bytecode produces a typed result 1097 inputs = rsize - depth; 1098 assert(inputs >= 0, ""); 1099 break; 1100 } 1101 1102 #ifdef ASSERT 1103 // spot check 1104 int outputs = depth + inputs; 1105 assert(outputs >= 0, "sanity"); 1106 switch (code) { 1107 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1108 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1109 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1110 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1111 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1112 default: break; 1113 } 1114 #endif //ASSERT 1115 1116 return true; 1117 } 1118 1119 1120 1121 //------------------------------basic_plus_adr--------------------------------- 1122 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1123 // short-circuit a common case 1124 if (offset == intcon(0)) return ptr; 1125 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1126 } 1127 1128 Node* GraphKit::ConvI2L(Node* offset) { 1129 // short-circuit a common case 1130 jint offset_con = find_int_con(offset, Type::OffsetBot); 1131 if (offset_con != Type::OffsetBot) { 1132 return longcon((jlong) offset_con); 1133 } 1134 return _gvn.transform( new ConvI2LNode(offset)); 1135 } 1136 1137 Node* GraphKit::ConvI2UL(Node* offset) { 1138 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1139 if (offset_con != (juint) Type::OffsetBot) { 1140 return longcon((julong) offset_con); 1141 } 1142 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1143 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1144 return _gvn.transform( new AndLNode(conv, mask) ); 1145 } 1146 1147 Node* GraphKit::ConvL2I(Node* offset) { 1148 // short-circuit a common case 1149 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1150 if (offset_con != (jlong)Type::OffsetBot) { 1151 return intcon((int) offset_con); 1152 } 1153 return _gvn.transform( new ConvL2INode(offset)); 1154 } 1155 1156 //-------------------------load_object_klass----------------------------------- 1157 Node* GraphKit::load_object_klass(Node* obj) { 1158 // Special-case a fresh allocation to avoid building nodes: 1159 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1160 if (akls != NULL) return akls; 1161 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1162 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1163 } 1164 1165 //-------------------------load_array_length----------------------------------- 1166 Node* GraphKit::load_array_length(Node* array) { 1167 // Special-case a fresh allocation to avoid building nodes: 1168 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1169 Node *alen; 1170 if (alloc == NULL) { 1171 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1172 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1173 } else { 1174 alen = alloc->Ideal_length(); 1175 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1176 if (ccast != alen) { 1177 alen = _gvn.transform(ccast); 1178 } 1179 } 1180 return alen; 1181 } 1182 1183 //------------------------------do_null_check---------------------------------- 1184 // Helper function to do a NULL pointer check. Returned value is 1185 // the incoming address with NULL casted away. You are allowed to use the 1186 // not-null value only if you are control dependent on the test. 1187 #ifndef PRODUCT 1188 extern int explicit_null_checks_inserted, 1189 explicit_null_checks_elided; 1190 #endif 1191 Node* GraphKit::null_check_common(Node* value, BasicType type, 1192 // optional arguments for variations: 1193 bool assert_null, 1194 Node* *null_control, 1195 bool speculative) { 1196 assert(!assert_null || null_control == NULL, "not both at once"); 1197 if (stopped()) return top(); 1198 NOT_PRODUCT(explicit_null_checks_inserted++); 1199 1200 // Construct NULL check 1201 Node *chk = NULL; 1202 switch(type) { 1203 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1204 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1205 case T_ARRAY : // fall through 1206 type = T_OBJECT; // simplify further tests 1207 case T_OBJECT : { 1208 const Type *t = _gvn.type( value ); 1209 1210 const TypeOopPtr* tp = t->isa_oopptr(); 1211 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1212 // Only for do_null_check, not any of its siblings: 1213 && !assert_null && null_control == NULL) { 1214 // Usually, any field access or invocation on an unloaded oop type 1215 // will simply fail to link, since the statically linked class is 1216 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1217 // the static class is loaded but the sharper oop type is not. 1218 // Rather than checking for this obscure case in lots of places, 1219 // we simply observe that a null check on an unloaded class 1220 // will always be followed by a nonsense operation, so we 1221 // can just issue the uncommon trap here. 1222 // Our access to the unloaded class will only be correct 1223 // after it has been loaded and initialized, which requires 1224 // a trip through the interpreter. 1225 #ifndef PRODUCT 1226 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1227 #endif 1228 uncommon_trap(Deoptimization::Reason_unloaded, 1229 Deoptimization::Action_reinterpret, 1230 tp->klass(), "!loaded"); 1231 return top(); 1232 } 1233 1234 if (assert_null) { 1235 // See if the type is contained in NULL_PTR. 1236 // If so, then the value is already null. 1237 if (t->higher_equal(TypePtr::NULL_PTR)) { 1238 NOT_PRODUCT(explicit_null_checks_elided++); 1239 return value; // Elided null assert quickly! 1240 } 1241 } else { 1242 // See if mixing in the NULL pointer changes type. 1243 // If so, then the NULL pointer was not allowed in the original 1244 // type. In other words, "value" was not-null. 1245 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1246 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1247 NOT_PRODUCT(explicit_null_checks_elided++); 1248 return value; // Elided null check quickly! 1249 } 1250 } 1251 chk = new CmpPNode( value, null() ); 1252 break; 1253 } 1254 1255 default: 1256 fatal("unexpected type: %s", type2name(type)); 1257 } 1258 assert(chk != NULL, "sanity check"); 1259 chk = _gvn.transform(chk); 1260 1261 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1262 BoolNode *btst = new BoolNode( chk, btest); 1263 Node *tst = _gvn.transform( btst ); 1264 1265 //----------- 1266 // if peephole optimizations occurred, a prior test existed. 1267 // If a prior test existed, maybe it dominates as we can avoid this test. 1268 if (tst != btst && type == T_OBJECT) { 1269 // At this point we want to scan up the CFG to see if we can 1270 // find an identical test (and so avoid this test altogether). 1271 Node *cfg = control(); 1272 int depth = 0; 1273 while( depth < 16 ) { // Limit search depth for speed 1274 if( cfg->Opcode() == Op_IfTrue && 1275 cfg->in(0)->in(1) == tst ) { 1276 // Found prior test. Use "cast_not_null" to construct an identical 1277 // CastPP (and hence hash to) as already exists for the prior test. 1278 // Return that casted value. 1279 if (assert_null) { 1280 replace_in_map(value, null()); 1281 return null(); // do not issue the redundant test 1282 } 1283 Node *oldcontrol = control(); 1284 set_control(cfg); 1285 Node *res = cast_not_null(value); 1286 set_control(oldcontrol); 1287 NOT_PRODUCT(explicit_null_checks_elided++); 1288 return res; 1289 } 1290 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1291 if (cfg == NULL) break; // Quit at region nodes 1292 depth++; 1293 } 1294 } 1295 1296 //----------- 1297 // Branch to failure if null 1298 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1299 Deoptimization::DeoptReason reason; 1300 if (assert_null) { 1301 reason = Deoptimization::reason_null_assert(speculative); 1302 } else if (type == T_OBJECT) { 1303 reason = Deoptimization::reason_null_check(speculative); 1304 } else { 1305 reason = Deoptimization::Reason_div0_check; 1306 } 1307 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1308 // ciMethodData::has_trap_at will return a conservative -1 if any 1309 // must-be-null assertion has failed. This could cause performance 1310 // problems for a method after its first do_null_assert failure. 1311 // Consider using 'Reason_class_check' instead? 1312 1313 // To cause an implicit null check, we set the not-null probability 1314 // to the maximum (PROB_MAX). For an explicit check the probability 1315 // is set to a smaller value. 1316 if (null_control != NULL || too_many_traps(reason)) { 1317 // probability is less likely 1318 ok_prob = PROB_LIKELY_MAG(3); 1319 } else if (!assert_null && 1320 (ImplicitNullCheckThreshold > 0) && 1321 method() != NULL && 1322 (method()->method_data()->trap_count(reason) 1323 >= (uint)ImplicitNullCheckThreshold)) { 1324 ok_prob = PROB_LIKELY_MAG(3); 1325 } 1326 1327 if (null_control != NULL) { 1328 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1329 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1330 set_control( _gvn.transform( new IfTrueNode(iff))); 1331 #ifndef PRODUCT 1332 if (null_true == top()) { 1333 explicit_null_checks_elided++; 1334 } 1335 #endif 1336 (*null_control) = null_true; 1337 } else { 1338 BuildCutout unless(this, tst, ok_prob); 1339 // Check for optimizer eliding test at parse time 1340 if (stopped()) { 1341 // Failure not possible; do not bother making uncommon trap. 1342 NOT_PRODUCT(explicit_null_checks_elided++); 1343 } else if (assert_null) { 1344 uncommon_trap(reason, 1345 Deoptimization::Action_make_not_entrant, 1346 NULL, "assert_null"); 1347 } else { 1348 replace_in_map(value, zerocon(type)); 1349 builtin_throw(reason); 1350 } 1351 } 1352 1353 // Must throw exception, fall-thru not possible? 1354 if (stopped()) { 1355 return top(); // No result 1356 } 1357 1358 if (assert_null) { 1359 // Cast obj to null on this path. 1360 replace_in_map(value, zerocon(type)); 1361 return zerocon(type); 1362 } 1363 1364 // Cast obj to not-null on this path, if there is no null_control. 1365 // (If there is a null_control, a non-null value may come back to haunt us.) 1366 if (type == T_OBJECT) { 1367 Node* cast = cast_not_null(value, false); 1368 if (null_control == NULL || (*null_control) == top()) 1369 replace_in_map(value, cast); 1370 value = cast; 1371 } 1372 1373 return value; 1374 } 1375 1376 1377 //------------------------------cast_not_null---------------------------------- 1378 // Cast obj to not-null on this path 1379 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1380 const Type *t = _gvn.type(obj); 1381 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1382 // Object is already not-null? 1383 if( t == t_not_null ) return obj; 1384 1385 Node *cast = new CastPPNode(obj,t_not_null); 1386 cast->init_req(0, control()); 1387 cast = _gvn.transform( cast ); 1388 1389 // Scan for instances of 'obj' in the current JVM mapping. 1390 // These instances are known to be not-null after the test. 1391 if (do_replace_in_map) 1392 replace_in_map(obj, cast); 1393 1394 return cast; // Return casted value 1395 } 1396 1397 // Sometimes in intrinsics, we implicitly know an object is not null 1398 // (there's no actual null check) so we can cast it to not null. In 1399 // the course of optimizations, the input to the cast can become null. 1400 // In that case that data path will die and we need the control path 1401 // to become dead as well to keep the graph consistent. So we have to 1402 // add a check for null for which one branch can't be taken. It uses 1403 // an Opaque4 node that will cause the check to be removed after loop 1404 // opts so the test goes away and the compiled code doesn't execute a 1405 // useless check. 1406 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) { 1407 Node* chk = _gvn.transform(new CmpPNode(value, null())); 1408 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne)); 1409 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1))); 1410 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN); 1411 _gvn.set_type(iff, iff->Value(&_gvn)); 1412 Node *if_f = _gvn.transform(new IfFalseNode(iff)); 1413 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr)); 1414 Node *halt = _gvn.transform(new HaltNode(if_f, frame)); 1415 C->root()->add_req(halt); 1416 Node *if_t = _gvn.transform(new IfTrueNode(iff)); 1417 set_control(if_t); 1418 return cast_not_null(value, do_replace_in_map); 1419 } 1420 1421 1422 //--------------------------replace_in_map------------------------------------- 1423 void GraphKit::replace_in_map(Node* old, Node* neww) { 1424 if (old == neww) { 1425 return; 1426 } 1427 1428 map()->replace_edge(old, neww); 1429 1430 // Note: This operation potentially replaces any edge 1431 // on the map. This includes locals, stack, and monitors 1432 // of the current (innermost) JVM state. 1433 1434 // don't let inconsistent types from profiling escape this 1435 // method 1436 1437 const Type* told = _gvn.type(old); 1438 const Type* tnew = _gvn.type(neww); 1439 1440 if (!tnew->higher_equal(told)) { 1441 return; 1442 } 1443 1444 map()->record_replaced_node(old, neww); 1445 } 1446 1447 1448 //============================================================================= 1449 //--------------------------------memory--------------------------------------- 1450 Node* GraphKit::memory(uint alias_idx) { 1451 MergeMemNode* mem = merged_memory(); 1452 Node* p = mem->memory_at(alias_idx); 1453 _gvn.set_type(p, Type::MEMORY); // must be mapped 1454 return p; 1455 } 1456 1457 //-----------------------------reset_memory------------------------------------ 1458 Node* GraphKit::reset_memory() { 1459 Node* mem = map()->memory(); 1460 // do not use this node for any more parsing! 1461 debug_only( map()->set_memory((Node*)NULL) ); 1462 return _gvn.transform( mem ); 1463 } 1464 1465 //------------------------------set_all_memory--------------------------------- 1466 void GraphKit::set_all_memory(Node* newmem) { 1467 Node* mergemem = MergeMemNode::make(newmem); 1468 gvn().set_type_bottom(mergemem); 1469 map()->set_memory(mergemem); 1470 } 1471 1472 //------------------------------set_all_memory_call---------------------------- 1473 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1474 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1475 set_all_memory(newmem); 1476 } 1477 1478 //============================================================================= 1479 // 1480 // parser factory methods for MemNodes 1481 // 1482 // These are layered on top of the factory methods in LoadNode and StoreNode, 1483 // and integrate with the parser's memory state and _gvn engine. 1484 // 1485 1486 // factory methods in "int adr_idx" 1487 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1488 int adr_idx, 1489 MemNode::MemOrd mo, 1490 LoadNode::ControlDependency control_dependency, 1491 bool require_atomic_access, 1492 bool unaligned, 1493 bool mismatched) { 1494 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1495 const TypePtr* adr_type = NULL; // debug-mode-only argument 1496 debug_only(adr_type = C->get_adr_type(adr_idx)); 1497 Node* mem = memory(adr_idx); 1498 Node* ld; 1499 if (require_atomic_access && bt == T_LONG) { 1500 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1501 } else if (require_atomic_access && bt == T_DOUBLE) { 1502 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1503 } else { 1504 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1505 } 1506 ld = _gvn.transform(ld); 1507 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) { 1508 // Improve graph before escape analysis and boxing elimination. 1509 record_for_igvn(ld); 1510 } 1511 return ld; 1512 } 1513 1514 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1515 int adr_idx, 1516 MemNode::MemOrd mo, 1517 bool require_atomic_access, 1518 bool unaligned, 1519 bool mismatched) { 1520 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1521 const TypePtr* adr_type = NULL; 1522 debug_only(adr_type = C->get_adr_type(adr_idx)); 1523 Node *mem = memory(adr_idx); 1524 Node* st; 1525 if (require_atomic_access && bt == T_LONG) { 1526 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1527 } else if (require_atomic_access && bt == T_DOUBLE) { 1528 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1529 } else { 1530 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1531 } 1532 if (unaligned) { 1533 st->as_Store()->set_unaligned_access(); 1534 } 1535 if (mismatched) { 1536 st->as_Store()->set_mismatched_access(); 1537 } 1538 st = _gvn.transform(st); 1539 set_memory(st, adr_idx); 1540 // Back-to-back stores can only remove intermediate store with DU info 1541 // so push on worklist for optimizer. 1542 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1543 record_for_igvn(st); 1544 1545 return st; 1546 } 1547 1548 Node* GraphKit::access_store_at(Node* ctl, 1549 Node* obj, 1550 Node* adr, 1551 const TypePtr* adr_type, 1552 Node* val, 1553 const Type* val_type, 1554 BasicType bt, 1555 DecoratorSet decorators) { 1556 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1557 // could be delayed during Parse (for example, in adjust_map_after_if()). 1558 // Execute transformation here to avoid barrier generation in such case. 1559 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1560 val = _gvn.makecon(TypePtr::NULL_PTR); 1561 } 1562 1563 set_control(ctl); 1564 if (stopped()) { 1565 return top(); // Dead path ? 1566 } 1567 1568 assert(val != NULL, "not dead path"); 1569 1570 C2AccessValuePtr addr(adr, adr_type); 1571 C2AccessValue value(val, val_type); 1572 C2Access access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr); 1573 if (access.is_raw()) { 1574 return _barrier_set->BarrierSetC2::store_at(access, value); 1575 } else { 1576 return _barrier_set->store_at(access, value); 1577 } 1578 } 1579 1580 Node* GraphKit::access_load_at(Node* obj, // containing obj 1581 Node* adr, // actual adress to store val at 1582 const TypePtr* adr_type, 1583 const Type* val_type, 1584 BasicType bt, 1585 DecoratorSet decorators) { 1586 if (stopped()) { 1587 return top(); // Dead path ? 1588 } 1589 1590 C2AccessValuePtr addr(adr, adr_type); 1591 C2Access access(this, decorators | C2_READ_ACCESS, bt, obj, addr); 1592 if (access.is_raw()) { 1593 return _barrier_set->BarrierSetC2::load_at(access, val_type); 1594 } else { 1595 return _barrier_set->load_at(access, val_type); 1596 } 1597 } 1598 1599 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* ctl, 1600 Node* obj, 1601 Node* adr, 1602 const TypePtr* adr_type, 1603 int alias_idx, 1604 Node* expected_val, 1605 Node* new_val, 1606 const Type* value_type, 1607 BasicType bt, 1608 DecoratorSet decorators) { 1609 set_control(ctl); 1610 C2AccessValuePtr addr(adr, adr_type); 1611 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1612 bt, obj, addr, alias_idx); 1613 if (access.is_raw()) { 1614 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1615 } else { 1616 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type); 1617 } 1618 } 1619 1620 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* ctl, 1621 Node* obj, 1622 Node* adr, 1623 const TypePtr* adr_type, 1624 int alias_idx, 1625 Node* expected_val, 1626 Node* new_val, 1627 const Type* value_type, 1628 BasicType bt, 1629 DecoratorSet decorators) { 1630 set_control(ctl); 1631 C2AccessValuePtr addr(adr, adr_type); 1632 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1633 bt, obj, addr, alias_idx); 1634 if (access.is_raw()) { 1635 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1636 } else { 1637 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type); 1638 } 1639 } 1640 1641 Node* GraphKit::access_atomic_xchg_at(Node* ctl, 1642 Node* obj, 1643 Node* adr, 1644 const TypePtr* adr_type, 1645 int alias_idx, 1646 Node* new_val, 1647 const Type* value_type, 1648 BasicType bt, 1649 DecoratorSet decorators) { 1650 set_control(ctl); 1651 C2AccessValuePtr addr(adr, adr_type); 1652 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, 1653 bt, obj, addr, alias_idx); 1654 if (access.is_raw()) { 1655 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type); 1656 } else { 1657 return _barrier_set->atomic_xchg_at(access, new_val, value_type); 1658 } 1659 } 1660 1661 Node* GraphKit::access_atomic_add_at(Node* ctl, 1662 Node* obj, 1663 Node* adr, 1664 const TypePtr* adr_type, 1665 int alias_idx, 1666 Node* new_val, 1667 const Type* value_type, 1668 BasicType bt, 1669 DecoratorSet decorators) { 1670 set_control(ctl); 1671 C2AccessValuePtr addr(adr, adr_type); 1672 C2AtomicAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx); 1673 if (access.is_raw()) { 1674 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type); 1675 } else { 1676 return _barrier_set->atomic_add_at(access, new_val, value_type); 1677 } 1678 } 1679 1680 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) { 1681 set_control(ctl); 1682 return _barrier_set->clone(this, src, dst, size, is_array); 1683 } 1684 1685 //-------------------------array_element_address------------------------- 1686 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1687 const TypeInt* sizetype, Node* ctrl) { 1688 uint shift = exact_log2(type2aelembytes(elembt)); 1689 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1690 1691 // short-circuit a common case (saves lots of confusing waste motion) 1692 jint idx_con = find_int_con(idx, -1); 1693 if (idx_con >= 0) { 1694 intptr_t offset = header + ((intptr_t)idx_con << shift); 1695 return basic_plus_adr(ary, offset); 1696 } 1697 1698 // must be correct type for alignment purposes 1699 Node* base = basic_plus_adr(ary, header); 1700 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1701 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1702 return basic_plus_adr(ary, base, scale); 1703 } 1704 1705 //-------------------------load_array_element------------------------- 1706 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1707 const Type* elemtype = arytype->elem(); 1708 BasicType elembt = elemtype->array_element_basic_type(); 1709 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1710 if (elembt == T_NARROWOOP) { 1711 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1712 } 1713 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1714 return ld; 1715 } 1716 1717 //-------------------------set_arguments_for_java_call------------------------- 1718 // Arguments (pre-popped from the stack) are taken from the JVMS. 1719 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1720 // Add the call arguments: 1721 uint nargs = call->method()->arg_size(); 1722 for (uint i = 0; i < nargs; i++) { 1723 Node* arg = argument(i); 1724 call->init_req(i + TypeFunc::Parms, arg); 1725 } 1726 } 1727 1728 //---------------------------set_edges_for_java_call--------------------------- 1729 // Connect a newly created call into the current JVMS. 1730 // A return value node (if any) is returned from set_edges_for_java_call. 1731 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1732 1733 // Add the predefined inputs: 1734 call->init_req( TypeFunc::Control, control() ); 1735 call->init_req( TypeFunc::I_O , i_o() ); 1736 call->init_req( TypeFunc::Memory , reset_memory() ); 1737 call->init_req( TypeFunc::FramePtr, frameptr() ); 1738 call->init_req( TypeFunc::ReturnAdr, top() ); 1739 1740 add_safepoint_edges(call, must_throw); 1741 1742 Node* xcall = _gvn.transform(call); 1743 1744 if (xcall == top()) { 1745 set_control(top()); 1746 return; 1747 } 1748 assert(xcall == call, "call identity is stable"); 1749 1750 // Re-use the current map to produce the result. 1751 1752 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1753 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1754 set_all_memory_call(xcall, separate_io_proj); 1755 1756 //return xcall; // no need, caller already has it 1757 } 1758 1759 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1760 if (stopped()) return top(); // maybe the call folded up? 1761 1762 // Capture the return value, if any. 1763 Node* ret; 1764 if (call->method() == NULL || 1765 call->method()->return_type()->basic_type() == T_VOID) 1766 ret = top(); 1767 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1768 1769 // Note: Since any out-of-line call can produce an exception, 1770 // we always insert an I_O projection from the call into the result. 1771 1772 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1773 1774 if (separate_io_proj) { 1775 // The caller requested separate projections be used by the fall 1776 // through and exceptional paths, so replace the projections for 1777 // the fall through path. 1778 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1779 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1780 } 1781 return ret; 1782 } 1783 1784 //--------------------set_predefined_input_for_runtime_call-------------------- 1785 // Reading and setting the memory state is way conservative here. 1786 // The real problem is that I am not doing real Type analysis on memory, 1787 // so I cannot distinguish card mark stores from other stores. Across a GC 1788 // point the Store Barrier and the card mark memory has to agree. I cannot 1789 // have a card mark store and its barrier split across the GC point from 1790 // either above or below. Here I get that to happen by reading ALL of memory. 1791 // A better answer would be to separate out card marks from other memory. 1792 // For now, return the input memory state, so that it can be reused 1793 // after the call, if this call has restricted memory effects. 1794 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1795 // Set fixed predefined input arguments 1796 Node* memory = reset_memory(); 1797 call->init_req( TypeFunc::Control, control() ); 1798 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1799 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1800 call->init_req( TypeFunc::FramePtr, frameptr() ); 1801 call->init_req( TypeFunc::ReturnAdr, top() ); 1802 return memory; 1803 } 1804 1805 //-------------------set_predefined_output_for_runtime_call-------------------- 1806 // Set control and memory (not i_o) from the call. 1807 // If keep_mem is not NULL, use it for the output state, 1808 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1809 // If hook_mem is NULL, this call produces no memory effects at all. 1810 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1811 // then only that memory slice is taken from the call. 1812 // In the last case, we must put an appropriate memory barrier before 1813 // the call, so as to create the correct anti-dependencies on loads 1814 // preceding the call. 1815 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1816 Node* keep_mem, 1817 const TypePtr* hook_mem) { 1818 // no i/o 1819 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1820 if (keep_mem) { 1821 // First clone the existing memory state 1822 set_all_memory(keep_mem); 1823 if (hook_mem != NULL) { 1824 // Make memory for the call 1825 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1826 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1827 // We also use hook_mem to extract specific effects from arraycopy stubs. 1828 set_memory(mem, hook_mem); 1829 } 1830 // ...else the call has NO memory effects. 1831 1832 // Make sure the call advertises its memory effects precisely. 1833 // This lets us build accurate anti-dependences in gcm.cpp. 1834 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1835 "call node must be constructed correctly"); 1836 } else { 1837 assert(hook_mem == NULL, ""); 1838 // This is not a "slow path" call; all memory comes from the call. 1839 set_all_memory_call(call); 1840 } 1841 } 1842 1843 1844 // Replace the call with the current state of the kit. 1845 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1846 JVMState* ejvms = NULL; 1847 if (has_exceptions()) { 1848 ejvms = transfer_exceptions_into_jvms(); 1849 } 1850 1851 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1852 ReplacedNodes replaced_nodes_exception; 1853 Node* ex_ctl = top(); 1854 1855 SafePointNode* final_state = stop(); 1856 1857 // Find all the needed outputs of this call 1858 CallProjections callprojs; 1859 call->extract_projections(&callprojs, true); 1860 1861 Node* init_mem = call->in(TypeFunc::Memory); 1862 Node* final_mem = final_state->in(TypeFunc::Memory); 1863 Node* final_ctl = final_state->in(TypeFunc::Control); 1864 Node* final_io = final_state->in(TypeFunc::I_O); 1865 1866 // Replace all the old call edges with the edges from the inlining result 1867 if (callprojs.fallthrough_catchproj != NULL) { 1868 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1869 } 1870 if (callprojs.fallthrough_memproj != NULL) { 1871 if (final_mem->is_MergeMem()) { 1872 // Parser's exits MergeMem was not transformed but may be optimized 1873 final_mem = _gvn.transform(final_mem); 1874 } 1875 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1876 } 1877 if (callprojs.fallthrough_ioproj != NULL) { 1878 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1879 } 1880 1881 // Replace the result with the new result if it exists and is used 1882 if (callprojs.resproj != NULL && result != NULL) { 1883 C->gvn_replace_by(callprojs.resproj, result); 1884 } 1885 1886 if (ejvms == NULL) { 1887 // No exception edges to simply kill off those paths 1888 if (callprojs.catchall_catchproj != NULL) { 1889 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1890 } 1891 if (callprojs.catchall_memproj != NULL) { 1892 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1893 } 1894 if (callprojs.catchall_ioproj != NULL) { 1895 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1896 } 1897 // Replace the old exception object with top 1898 if (callprojs.exobj != NULL) { 1899 C->gvn_replace_by(callprojs.exobj, C->top()); 1900 } 1901 } else { 1902 GraphKit ekit(ejvms); 1903 1904 // Load my combined exception state into the kit, with all phis transformed: 1905 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1906 replaced_nodes_exception = ex_map->replaced_nodes(); 1907 1908 Node* ex_oop = ekit.use_exception_state(ex_map); 1909 1910 if (callprojs.catchall_catchproj != NULL) { 1911 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1912 ex_ctl = ekit.control(); 1913 } 1914 if (callprojs.catchall_memproj != NULL) { 1915 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1916 } 1917 if (callprojs.catchall_ioproj != NULL) { 1918 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1919 } 1920 1921 // Replace the old exception object with the newly created one 1922 if (callprojs.exobj != NULL) { 1923 C->gvn_replace_by(callprojs.exobj, ex_oop); 1924 } 1925 } 1926 1927 // Disconnect the call from the graph 1928 call->disconnect_inputs(NULL, C); 1929 C->gvn_replace_by(call, C->top()); 1930 1931 // Clean up any MergeMems that feed other MergeMems since the 1932 // optimizer doesn't like that. 1933 if (final_mem->is_MergeMem()) { 1934 Node_List wl; 1935 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1936 Node* m = i.get(); 1937 if (m->is_MergeMem() && !wl.contains(m)) { 1938 wl.push(m); 1939 } 1940 } 1941 while (wl.size() > 0) { 1942 _gvn.transform(wl.pop()); 1943 } 1944 } 1945 1946 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1947 replaced_nodes.apply(C, final_ctl); 1948 } 1949 if (!ex_ctl->is_top() && do_replaced_nodes) { 1950 replaced_nodes_exception.apply(C, ex_ctl); 1951 } 1952 } 1953 1954 1955 //------------------------------increment_counter------------------------------ 1956 // for statistics: increment a VM counter by 1 1957 1958 void GraphKit::increment_counter(address counter_addr) { 1959 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1960 increment_counter(adr1); 1961 } 1962 1963 void GraphKit::increment_counter(Node* counter_addr) { 1964 int adr_type = Compile::AliasIdxRaw; 1965 Node* ctrl = control(); 1966 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1967 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1968 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1969 } 1970 1971 1972 //------------------------------uncommon_trap---------------------------------- 1973 // Bail out to the interpreter in mid-method. Implemented by calling the 1974 // uncommon_trap blob. This helper function inserts a runtime call with the 1975 // right debug info. 1976 void GraphKit::uncommon_trap(int trap_request, 1977 ciKlass* klass, const char* comment, 1978 bool must_throw, 1979 bool keep_exact_action) { 1980 if (failing()) stop(); 1981 if (stopped()) return; // trap reachable? 1982 1983 // Note: If ProfileTraps is true, and if a deopt. actually 1984 // occurs here, the runtime will make sure an MDO exists. There is 1985 // no need to call method()->ensure_method_data() at this point. 1986 1987 // Set the stack pointer to the right value for reexecution: 1988 set_sp(reexecute_sp()); 1989 1990 #ifdef ASSERT 1991 if (!must_throw) { 1992 // Make sure the stack has at least enough depth to execute 1993 // the current bytecode. 1994 int inputs, ignored_depth; 1995 if (compute_stack_effects(inputs, ignored_depth)) { 1996 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1997 Bytecodes::name(java_bc()), sp(), inputs); 1998 } 1999 } 2000 #endif 2001 2002 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 2003 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 2004 2005 switch (action) { 2006 case Deoptimization::Action_maybe_recompile: 2007 case Deoptimization::Action_reinterpret: 2008 // Temporary fix for 6529811 to allow virtual calls to be sure they 2009 // get the chance to go from mono->bi->mega 2010 if (!keep_exact_action && 2011 Deoptimization::trap_request_index(trap_request) < 0 && 2012 too_many_recompiles(reason)) { 2013 // This BCI is causing too many recompilations. 2014 if (C->log() != NULL) { 2015 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 2016 Deoptimization::trap_reason_name(reason), 2017 Deoptimization::trap_action_name(action)); 2018 } 2019 action = Deoptimization::Action_none; 2020 trap_request = Deoptimization::make_trap_request(reason, action); 2021 } else { 2022 C->set_trap_can_recompile(true); 2023 } 2024 break; 2025 case Deoptimization::Action_make_not_entrant: 2026 C->set_trap_can_recompile(true); 2027 break; 2028 case Deoptimization::Action_none: 2029 case Deoptimization::Action_make_not_compilable: 2030 break; 2031 default: 2032 #ifdef ASSERT 2033 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 2034 #endif 2035 break; 2036 } 2037 2038 if (TraceOptoParse) { 2039 char buf[100]; 2040 tty->print_cr("Uncommon trap %s at bci:%d", 2041 Deoptimization::format_trap_request(buf, sizeof(buf), 2042 trap_request), bci()); 2043 } 2044 2045 CompileLog* log = C->log(); 2046 if (log != NULL) { 2047 int kid = (klass == NULL)? -1: log->identify(klass); 2048 log->begin_elem("uncommon_trap bci='%d'", bci()); 2049 char buf[100]; 2050 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 2051 trap_request)); 2052 if (kid >= 0) log->print(" klass='%d'", kid); 2053 if (comment != NULL) log->print(" comment='%s'", comment); 2054 log->end_elem(); 2055 } 2056 2057 // Make sure any guarding test views this path as very unlikely 2058 Node *i0 = control()->in(0); 2059 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 2060 IfNode *iff = i0->as_If(); 2061 float f = iff->_prob; // Get prob 2062 if (control()->Opcode() == Op_IfTrue) { 2063 if (f > PROB_UNLIKELY_MAG(4)) 2064 iff->_prob = PROB_MIN; 2065 } else { 2066 if (f < PROB_LIKELY_MAG(4)) 2067 iff->_prob = PROB_MAX; 2068 } 2069 } 2070 2071 // Clear out dead values from the debug info. 2072 kill_dead_locals(); 2073 2074 // Now insert the uncommon trap subroutine call 2075 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2076 const TypePtr* no_memory_effects = NULL; 2077 // Pass the index of the class to be loaded 2078 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2079 (must_throw ? RC_MUST_THROW : 0), 2080 OptoRuntime::uncommon_trap_Type(), 2081 call_addr, "uncommon_trap", no_memory_effects, 2082 intcon(trap_request)); 2083 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2084 "must extract request correctly from the graph"); 2085 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2086 2087 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2088 // The debug info is the only real input to this call. 2089 2090 // Halt-and-catch fire here. The above call should never return! 2091 HaltNode* halt = new HaltNode(control(), frameptr()); 2092 _gvn.set_type_bottom(halt); 2093 root()->add_req(halt); 2094 2095 stop_and_kill_map(); 2096 } 2097 2098 2099 //--------------------------just_allocated_object------------------------------ 2100 // Report the object that was just allocated. 2101 // It must be the case that there are no intervening safepoints. 2102 // We use this to determine if an object is so "fresh" that 2103 // it does not require card marks. 2104 Node* GraphKit::just_allocated_object(Node* current_control) { 2105 if (C->recent_alloc_ctl() == current_control) 2106 return C->recent_alloc_obj(); 2107 return NULL; 2108 } 2109 2110 2111 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2112 // (Note: TypeFunc::make has a cache that makes this fast.) 2113 const TypeFunc* tf = TypeFunc::make(dest_method); 2114 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2115 for (int j = 0; j < nargs; j++) { 2116 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2117 if( targ->basic_type() == T_DOUBLE ) { 2118 // If any parameters are doubles, they must be rounded before 2119 // the call, dstore_rounding does gvn.transform 2120 Node *arg = argument(j); 2121 arg = dstore_rounding(arg); 2122 set_argument(j, arg); 2123 } 2124 } 2125 } 2126 2127 /** 2128 * Record profiling data exact_kls for Node n with the type system so 2129 * that it can propagate it (speculation) 2130 * 2131 * @param n node that the type applies to 2132 * @param exact_kls type from profiling 2133 * @param maybe_null did profiling see null? 2134 * 2135 * @return node with improved type 2136 */ 2137 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) { 2138 const Type* current_type = _gvn.type(n); 2139 assert(UseTypeSpeculation, "type speculation must be on"); 2140 2141 const TypePtr* speculative = current_type->speculative(); 2142 2143 // Should the klass from the profile be recorded in the speculative type? 2144 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2145 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2146 const TypeOopPtr* xtype = tklass->as_instance_type(); 2147 assert(xtype->klass_is_exact(), "Should be exact"); 2148 // Any reason to believe n is not null (from this profiling or a previous one)? 2149 assert(ptr_kind != ProfileAlwaysNull, "impossible here"); 2150 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2151 // record the new speculative type's depth 2152 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2153 speculative = speculative->with_inline_depth(jvms()->depth()); 2154 } else if (current_type->would_improve_ptr(ptr_kind)) { 2155 // Profiling report that null was never seen so we can change the 2156 // speculative type to non null ptr. 2157 if (ptr_kind == ProfileAlwaysNull) { 2158 speculative = TypePtr::NULL_PTR; 2159 } else { 2160 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement"); 2161 const TypePtr* ptr = TypePtr::NOTNULL; 2162 if (speculative != NULL) { 2163 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2164 } else { 2165 speculative = ptr; 2166 } 2167 } 2168 } 2169 2170 if (speculative != current_type->speculative()) { 2171 // Build a type with a speculative type (what we think we know 2172 // about the type but will need a guard when we use it) 2173 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2174 // We're changing the type, we need a new CheckCast node to carry 2175 // the new type. The new type depends on the control: what 2176 // profiling tells us is only valid from here as far as we can 2177 // tell. 2178 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2179 cast = _gvn.transform(cast); 2180 replace_in_map(n, cast); 2181 n = cast; 2182 } 2183 2184 return n; 2185 } 2186 2187 /** 2188 * Record profiling data from receiver profiling at an invoke with the 2189 * type system so that it can propagate it (speculation) 2190 * 2191 * @param n receiver node 2192 * 2193 * @return node with improved type 2194 */ 2195 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2196 if (!UseTypeSpeculation) { 2197 return n; 2198 } 2199 ciKlass* exact_kls = profile_has_unique_klass(); 2200 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2201 if ((java_bc() == Bytecodes::_checkcast || 2202 java_bc() == Bytecodes::_instanceof || 2203 java_bc() == Bytecodes::_aastore) && 2204 method()->method_data()->is_mature()) { 2205 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2206 if (data != NULL) { 2207 if (!data->as_BitData()->null_seen()) { 2208 ptr_kind = ProfileNeverNull; 2209 } else { 2210 assert(data->is_ReceiverTypeData(), "bad profile data type"); 2211 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData(); 2212 uint i = 0; 2213 for (; i < call->row_limit(); i++) { 2214 ciKlass* receiver = call->receiver(i); 2215 if (receiver != NULL) { 2216 break; 2217 } 2218 } 2219 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull; 2220 } 2221 } 2222 } 2223 return record_profile_for_speculation(n, exact_kls, ptr_kind); 2224 } 2225 2226 /** 2227 * Record profiling data from argument profiling at an invoke with the 2228 * type system so that it can propagate it (speculation) 2229 * 2230 * @param dest_method target method for the call 2231 * @param bc what invoke bytecode is this? 2232 */ 2233 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2234 if (!UseTypeSpeculation) { 2235 return; 2236 } 2237 const TypeFunc* tf = TypeFunc::make(dest_method); 2238 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2239 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2240 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2241 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2242 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2243 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2244 ciKlass* better_type = NULL; 2245 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) { 2246 record_profile_for_speculation(argument(j), better_type, ptr_kind); 2247 } 2248 i++; 2249 } 2250 } 2251 } 2252 2253 /** 2254 * Record profiling data from parameter profiling at an invoke with 2255 * the type system so that it can propagate it (speculation) 2256 */ 2257 void GraphKit::record_profiled_parameters_for_speculation() { 2258 if (!UseTypeSpeculation) { 2259 return; 2260 } 2261 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2262 if (_gvn.type(local(i))->isa_oopptr()) { 2263 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2264 ciKlass* better_type = NULL; 2265 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) { 2266 record_profile_for_speculation(local(i), better_type, ptr_kind); 2267 } 2268 j++; 2269 } 2270 } 2271 } 2272 2273 /** 2274 * Record profiling data from return value profiling at an invoke with 2275 * the type system so that it can propagate it (speculation) 2276 */ 2277 void GraphKit::record_profiled_return_for_speculation() { 2278 if (!UseTypeSpeculation) { 2279 return; 2280 } 2281 ProfilePtrKind ptr_kind = ProfileMaybeNull; 2282 ciKlass* better_type = NULL; 2283 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) { 2284 // If profiling reports a single type for the return value, 2285 // feed it to the type system so it can propagate it as a 2286 // speculative type 2287 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind); 2288 } 2289 } 2290 2291 void GraphKit::round_double_result(ciMethod* dest_method) { 2292 // A non-strict method may return a double value which has an extended 2293 // exponent, but this must not be visible in a caller which is 'strict' 2294 // If a strict caller invokes a non-strict callee, round a double result 2295 2296 BasicType result_type = dest_method->return_type()->basic_type(); 2297 assert( method() != NULL, "must have caller context"); 2298 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2299 // Destination method's return value is on top of stack 2300 // dstore_rounding() does gvn.transform 2301 Node *result = pop_pair(); 2302 result = dstore_rounding(result); 2303 push_pair(result); 2304 } 2305 } 2306 2307 // rounding for strict float precision conformance 2308 Node* GraphKit::precision_rounding(Node* n) { 2309 return UseStrictFP && _method->flags().is_strict() 2310 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2311 ? _gvn.transform( new RoundFloatNode(0, n) ) 2312 : n; 2313 } 2314 2315 // rounding for strict double precision conformance 2316 Node* GraphKit::dprecision_rounding(Node *n) { 2317 return UseStrictFP && _method->flags().is_strict() 2318 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2319 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2320 : n; 2321 } 2322 2323 // rounding for non-strict double stores 2324 Node* GraphKit::dstore_rounding(Node* n) { 2325 return Matcher::strict_fp_requires_explicit_rounding 2326 && UseSSE <= 1 2327 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2328 : n; 2329 } 2330 2331 //============================================================================= 2332 // Generate a fast path/slow path idiom. Graph looks like: 2333 // [foo] indicates that 'foo' is a parameter 2334 // 2335 // [in] NULL 2336 // \ / 2337 // CmpP 2338 // Bool ne 2339 // If 2340 // / \ 2341 // True False-<2> 2342 // / | 2343 // / cast_not_null 2344 // Load | | ^ 2345 // [fast_test] | | 2346 // gvn to opt_test | | 2347 // / \ | <1> 2348 // True False | 2349 // | \\ | 2350 // [slow_call] \[fast_result] 2351 // Ctl Val \ \ 2352 // | \ \ 2353 // Catch <1> \ \ 2354 // / \ ^ \ \ 2355 // Ex No_Ex | \ \ 2356 // | \ \ | \ <2> \ 2357 // ... \ [slow_res] | | \ [null_result] 2358 // \ \--+--+--- | | 2359 // \ | / \ | / 2360 // --------Region Phi 2361 // 2362 //============================================================================= 2363 // Code is structured as a series of driver functions all called 'do_XXX' that 2364 // call a set of helper functions. Helper functions first, then drivers. 2365 2366 //------------------------------null_check_oop--------------------------------- 2367 // Null check oop. Set null-path control into Region in slot 3. 2368 // Make a cast-not-nullness use the other not-null control. Return cast. 2369 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2370 bool never_see_null, 2371 bool safe_for_replace, 2372 bool speculative) { 2373 // Initial NULL check taken path 2374 (*null_control) = top(); 2375 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2376 2377 // Generate uncommon_trap: 2378 if (never_see_null && (*null_control) != top()) { 2379 // If we see an unexpected null at a check-cast we record it and force a 2380 // recompile; the offending check-cast will be compiled to handle NULLs. 2381 // If we see more than one offending BCI, then all checkcasts in the 2382 // method will be compiled to handle NULLs. 2383 PreserveJVMState pjvms(this); 2384 set_control(*null_control); 2385 replace_in_map(value, null()); 2386 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2387 uncommon_trap(reason, 2388 Deoptimization::Action_make_not_entrant); 2389 (*null_control) = top(); // NULL path is dead 2390 } 2391 if ((*null_control) == top() && safe_for_replace) { 2392 replace_in_map(value, cast); 2393 } 2394 2395 // Cast away null-ness on the result 2396 return cast; 2397 } 2398 2399 //------------------------------opt_iff---------------------------------------- 2400 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2401 // Return slow-path control. 2402 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2403 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2404 2405 // Fast path taken; set region slot 2 2406 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2407 region->init_req(2,fast_taken); // Capture fast-control 2408 2409 // Fast path not-taken, i.e. slow path 2410 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2411 return slow_taken; 2412 } 2413 2414 //-----------------------------make_runtime_call------------------------------- 2415 Node* GraphKit::make_runtime_call(int flags, 2416 const TypeFunc* call_type, address call_addr, 2417 const char* call_name, 2418 const TypePtr* adr_type, 2419 // The following parms are all optional. 2420 // The first NULL ends the list. 2421 Node* parm0, Node* parm1, 2422 Node* parm2, Node* parm3, 2423 Node* parm4, Node* parm5, 2424 Node* parm6, Node* parm7) { 2425 // Slow-path call 2426 bool is_leaf = !(flags & RC_NO_LEAF); 2427 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2428 if (call_name == NULL) { 2429 assert(!is_leaf, "must supply name for leaf"); 2430 call_name = OptoRuntime::stub_name(call_addr); 2431 } 2432 CallNode* call; 2433 if (!is_leaf) { 2434 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2435 bci(), adr_type); 2436 } else if (flags & RC_NO_FP) { 2437 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2438 } else { 2439 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2440 } 2441 2442 // The following is similar to set_edges_for_java_call, 2443 // except that the memory effects of the call are restricted to AliasIdxRaw. 2444 2445 // Slow path call has no side-effects, uses few values 2446 bool wide_in = !(flags & RC_NARROW_MEM); 2447 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2448 2449 Node* prev_mem = NULL; 2450 if (wide_in) { 2451 prev_mem = set_predefined_input_for_runtime_call(call); 2452 } else { 2453 assert(!wide_out, "narrow in => narrow out"); 2454 Node* narrow_mem = memory(adr_type); 2455 prev_mem = reset_memory(); 2456 map()->set_memory(narrow_mem); 2457 set_predefined_input_for_runtime_call(call); 2458 } 2459 2460 // Hook each parm in order. Stop looking at the first NULL. 2461 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2462 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2463 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2464 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2465 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2466 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2467 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2468 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2469 /* close each nested if ===> */ } } } } } } } } 2470 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2471 2472 if (!is_leaf) { 2473 // Non-leaves can block and take safepoints: 2474 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2475 } 2476 // Non-leaves can throw exceptions: 2477 if (has_io) { 2478 call->set_req(TypeFunc::I_O, i_o()); 2479 } 2480 2481 if (flags & RC_UNCOMMON) { 2482 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2483 // (An "if" probability corresponds roughly to an unconditional count. 2484 // Sort of.) 2485 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2486 } 2487 2488 Node* c = _gvn.transform(call); 2489 assert(c == call, "cannot disappear"); 2490 2491 if (wide_out) { 2492 // Slow path call has full side-effects. 2493 set_predefined_output_for_runtime_call(call); 2494 } else { 2495 // Slow path call has few side-effects, and/or sets few values. 2496 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2497 } 2498 2499 if (has_io) { 2500 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2501 } 2502 return call; 2503 2504 } 2505 2506 //------------------------------merge_memory----------------------------------- 2507 // Merge memory from one path into the current memory state. 2508 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2509 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2510 Node* old_slice = mms.force_memory(); 2511 Node* new_slice = mms.memory2(); 2512 if (old_slice != new_slice) { 2513 PhiNode* phi; 2514 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2515 if (mms.is_empty()) { 2516 // clone base memory Phi's inputs for this memory slice 2517 assert(old_slice == mms.base_memory(), "sanity"); 2518 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2519 _gvn.set_type(phi, Type::MEMORY); 2520 for (uint i = 1; i < phi->req(); i++) { 2521 phi->init_req(i, old_slice->in(i)); 2522 } 2523 } else { 2524 phi = old_slice->as_Phi(); // Phi was generated already 2525 } 2526 } else { 2527 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2528 _gvn.set_type(phi, Type::MEMORY); 2529 } 2530 phi->set_req(new_path, new_slice); 2531 mms.set_memory(phi); 2532 } 2533 } 2534 } 2535 2536 //------------------------------make_slow_call_ex------------------------------ 2537 // Make the exception handler hookups for the slow call 2538 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2539 if (stopped()) return; 2540 2541 // Make a catch node with just two handlers: fall-through and catch-all 2542 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2543 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2544 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2545 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2546 2547 { PreserveJVMState pjvms(this); 2548 set_control(excp); 2549 set_i_o(i_o); 2550 2551 if (excp != top()) { 2552 if (deoptimize) { 2553 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2554 uncommon_trap(Deoptimization::Reason_unhandled, 2555 Deoptimization::Action_none); 2556 } else { 2557 // Create an exception state also. 2558 // Use an exact type if the caller has specified a specific exception. 2559 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2560 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2561 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2562 } 2563 } 2564 } 2565 2566 // Get the no-exception control from the CatchNode. 2567 set_control(norm); 2568 } 2569 2570 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2571 Node* cmp = NULL; 2572 switch(bt) { 2573 case T_INT: cmp = new CmpINode(in1, in2); break; 2574 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2575 default: fatal("unexpected comparison type %s", type2name(bt)); 2576 } 2577 gvn->transform(cmp); 2578 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2579 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2580 gvn->transform(iff); 2581 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2582 return iff; 2583 } 2584 2585 2586 //-------------------------------gen_subtype_check----------------------------- 2587 // Generate a subtyping check. Takes as input the subtype and supertype. 2588 // Returns 2 values: sets the default control() to the true path and returns 2589 // the false path. Only reads invariant memory; sets no (visible) memory. 2590 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2591 // but that's not exposed to the optimizer. This call also doesn't take in an 2592 // Object; if you wish to check an Object you need to load the Object's class 2593 // prior to coming here. 2594 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2595 Compile* C = gvn->C; 2596 2597 if ((*ctrl)->is_top()) { 2598 return C->top(); 2599 } 2600 2601 // Fast check for identical types, perhaps identical constants. 2602 // The types can even be identical non-constants, in cases 2603 // involving Array.newInstance, Object.clone, etc. 2604 if (subklass == superklass) 2605 return C->top(); // false path is dead; no test needed. 2606 2607 if (gvn->type(superklass)->singleton()) { 2608 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2609 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2610 2611 // In the common case of an exact superklass, try to fold up the 2612 // test before generating code. You may ask, why not just generate 2613 // the code and then let it fold up? The answer is that the generated 2614 // code will necessarily include null checks, which do not always 2615 // completely fold away. If they are also needless, then they turn 2616 // into a performance loss. Example: 2617 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2618 // Here, the type of 'fa' is often exact, so the store check 2619 // of fa[1]=x will fold up, without testing the nullness of x. 2620 switch (C->static_subtype_check(superk, subk)) { 2621 case Compile::SSC_always_false: 2622 { 2623 Node* always_fail = *ctrl; 2624 *ctrl = gvn->C->top(); 2625 return always_fail; 2626 } 2627 case Compile::SSC_always_true: 2628 return C->top(); 2629 case Compile::SSC_easy_test: 2630 { 2631 // Just do a direct pointer compare and be done. 2632 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2633 *ctrl = gvn->transform(new IfTrueNode(iff)); 2634 return gvn->transform(new IfFalseNode(iff)); 2635 } 2636 case Compile::SSC_full_test: 2637 break; 2638 default: 2639 ShouldNotReachHere(); 2640 } 2641 } 2642 2643 // %%% Possible further optimization: Even if the superklass is not exact, 2644 // if the subklass is the unique subtype of the superklass, the check 2645 // will always succeed. We could leave a dependency behind to ensure this. 2646 2647 // First load the super-klass's check-offset 2648 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2649 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2650 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2651 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2652 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2653 2654 // Load from the sub-klass's super-class display list, or a 1-word cache of 2655 // the secondary superclass list, or a failing value with a sentinel offset 2656 // if the super-klass is an interface or exceptionally deep in the Java 2657 // hierarchy and we have to scan the secondary superclass list the hard way. 2658 // Worst-case type is a little odd: NULL is allowed as a result (usually 2659 // klass loads can never produce a NULL). 2660 Node *chk_off_X = chk_off; 2661 #ifdef _LP64 2662 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2663 #endif 2664 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2665 // For some types like interfaces the following loadKlass is from a 1-word 2666 // cache which is mutable so can't use immutable memory. Other 2667 // types load from the super-class display table which is immutable. 2668 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2669 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2670 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2671 2672 // Compile speed common case: ARE a subtype and we canNOT fail 2673 if( superklass == nkls ) 2674 return C->top(); // false path is dead; no test needed. 2675 2676 // See if we get an immediate positive hit. Happens roughly 83% of the 2677 // time. Test to see if the value loaded just previously from the subklass 2678 // is exactly the superklass. 2679 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2680 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2681 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2682 2683 // Compile speed common case: Check for being deterministic right now. If 2684 // chk_off is a constant and not equal to cacheoff then we are NOT a 2685 // subklass. In this case we need exactly the 1 test above and we can 2686 // return those results immediately. 2687 if (!might_be_cache) { 2688 Node* not_subtype_ctrl = *ctrl; 2689 *ctrl = iftrue1; // We need exactly the 1 test above 2690 return not_subtype_ctrl; 2691 } 2692 2693 // Gather the various success & failures here 2694 RegionNode *r_ok_subtype = new RegionNode(4); 2695 gvn->record_for_igvn(r_ok_subtype); 2696 RegionNode *r_not_subtype = new RegionNode(3); 2697 gvn->record_for_igvn(r_not_subtype); 2698 2699 r_ok_subtype->init_req(1, iftrue1); 2700 2701 // Check for immediate negative hit. Happens roughly 11% of the time (which 2702 // is roughly 63% of the remaining cases). Test to see if the loaded 2703 // check-offset points into the subklass display list or the 1-element 2704 // cache. If it points to the display (and NOT the cache) and the display 2705 // missed then it's not a subtype. 2706 Node *cacheoff = gvn->intcon(cacheoff_con); 2707 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2708 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2709 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2710 2711 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2712 // No performance impact (too rare) but allows sharing of secondary arrays 2713 // which has some footprint reduction. 2714 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2715 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2716 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2717 2718 // -- Roads not taken here: -- 2719 // We could also have chosen to perform the self-check at the beginning 2720 // of this code sequence, as the assembler does. This would not pay off 2721 // the same way, since the optimizer, unlike the assembler, can perform 2722 // static type analysis to fold away many successful self-checks. 2723 // Non-foldable self checks work better here in second position, because 2724 // the initial primary superclass check subsumes a self-check for most 2725 // types. An exception would be a secondary type like array-of-interface, 2726 // which does not appear in its own primary supertype display. 2727 // Finally, we could have chosen to move the self-check into the 2728 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2729 // dependent manner. But it is worthwhile to have the check here, 2730 // where it can be perhaps be optimized. The cost in code space is 2731 // small (register compare, branch). 2732 2733 // Now do a linear scan of the secondary super-klass array. Again, no real 2734 // performance impact (too rare) but it's gotta be done. 2735 // Since the code is rarely used, there is no penalty for moving it 2736 // out of line, and it can only improve I-cache density. 2737 // The decision to inline or out-of-line this final check is platform 2738 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2739 Node* psc = gvn->transform( 2740 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2741 2742 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2743 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2744 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2745 2746 // Return false path; set default control to true path. 2747 *ctrl = gvn->transform(r_ok_subtype); 2748 return gvn->transform(r_not_subtype); 2749 } 2750 2751 // Profile-driven exact type check: 2752 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2753 float prob, 2754 Node* *casted_receiver) { 2755 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2756 Node* recv_klass = load_object_klass(receiver); 2757 Node* want_klass = makecon(tklass); 2758 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2759 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2760 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2761 set_control( _gvn.transform( new IfTrueNode (iff) )); 2762 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2763 2764 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2765 assert(recv_xtype->klass_is_exact(), ""); 2766 2767 // Subsume downstream occurrences of receiver with a cast to 2768 // recv_xtype, since now we know what the type will be. 2769 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2770 (*casted_receiver) = _gvn.transform(cast); 2771 // (User must make the replace_in_map call.) 2772 2773 return fail; 2774 } 2775 2776 2777 //------------------------------seems_never_null------------------------------- 2778 // Use null_seen information if it is available from the profile. 2779 // If we see an unexpected null at a type check we record it and force a 2780 // recompile; the offending check will be recompiled to handle NULLs. 2781 // If we see several offending BCIs, then all checks in the 2782 // method will be recompiled. 2783 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2784 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2785 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2786 if (UncommonNullCast // Cutout for this technique 2787 && obj != null() // And not the -Xcomp stupid case? 2788 && !too_many_traps(reason) 2789 ) { 2790 if (speculating) { 2791 return true; 2792 } 2793 if (data == NULL) 2794 // Edge case: no mature data. Be optimistic here. 2795 return true; 2796 // If the profile has not seen a null, assume it won't happen. 2797 assert(java_bc() == Bytecodes::_checkcast || 2798 java_bc() == Bytecodes::_instanceof || 2799 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2800 return !data->as_BitData()->null_seen(); 2801 } 2802 speculating = false; 2803 return false; 2804 } 2805 2806 //------------------------maybe_cast_profiled_receiver------------------------- 2807 // If the profile has seen exactly one type, narrow to exactly that type. 2808 // Subsequent type checks will always fold up. 2809 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2810 ciKlass* require_klass, 2811 ciKlass* spec_klass, 2812 bool safe_for_replace) { 2813 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2814 2815 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2816 2817 // Make sure we haven't already deoptimized from this tactic. 2818 if (too_many_traps(reason) || too_many_recompiles(reason)) 2819 return NULL; 2820 2821 // (No, this isn't a call, but it's enough like a virtual call 2822 // to use the same ciMethod accessor to get the profile info...) 2823 // If we have a speculative type use it instead of profiling (which 2824 // may not help us) 2825 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2826 if (exact_kls != NULL) {// no cast failures here 2827 if (require_klass == NULL || 2828 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2829 // If we narrow the type to match what the type profile sees or 2830 // the speculative type, we can then remove the rest of the 2831 // cast. 2832 // This is a win, even if the exact_kls is very specific, 2833 // because downstream operations, such as method calls, 2834 // will often benefit from the sharper type. 2835 Node* exact_obj = not_null_obj; // will get updated in place... 2836 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2837 &exact_obj); 2838 { PreserveJVMState pjvms(this); 2839 set_control(slow_ctl); 2840 uncommon_trap_exact(reason, 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 == Compile::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 if (stopped()) { 2865 return obj; 2866 } 2867 2868 // type == NULL if profiling tells us this object is always null 2869 if (type != NULL) { 2870 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2871 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2872 2873 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2874 !too_many_traps(class_reason) && 2875 !too_many_recompiles(class_reason)) { 2876 Node* not_null_obj = NULL; 2877 // not_null is true if we know the object is not null and 2878 // there's no need for a null check 2879 if (!not_null) { 2880 Node* null_ctl = top(); 2881 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2882 assert(null_ctl->is_top(), "no null control here"); 2883 } else { 2884 not_null_obj = obj; 2885 } 2886 2887 Node* exact_obj = not_null_obj; 2888 ciKlass* exact_kls = type; 2889 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2890 &exact_obj); 2891 { 2892 PreserveJVMState pjvms(this); 2893 set_control(slow_ctl); 2894 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2895 } 2896 replace_in_map(not_null_obj, exact_obj); 2897 obj = exact_obj; 2898 } 2899 } else { 2900 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2901 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2902 Node* exact_obj = null_assert(obj); 2903 replace_in_map(obj, exact_obj); 2904 obj = exact_obj; 2905 } 2906 } 2907 return obj; 2908 } 2909 2910 //-------------------------------gen_instanceof-------------------------------- 2911 // Generate an instance-of idiom. Used by both the instance-of bytecode 2912 // and the reflective instance-of call. 2913 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2914 kill_dead_locals(); // Benefit all the uncommon traps 2915 assert( !stopped(), "dead parse path should be checked in callers" ); 2916 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2917 "must check for not-null not-dead klass in callers"); 2918 2919 // Make the merge point 2920 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2921 RegionNode* region = new RegionNode(PATH_LIMIT); 2922 Node* phi = new PhiNode(region, TypeInt::BOOL); 2923 C->set_has_split_ifs(true); // Has chance for split-if optimization 2924 2925 ciProfileData* data = NULL; 2926 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2927 data = method()->method_data()->bci_to_data(bci()); 2928 } 2929 bool speculative_not_null = false; 2930 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2931 && seems_never_null(obj, data, speculative_not_null)); 2932 2933 // Null check; get casted pointer; set region slot 3 2934 Node* null_ctl = top(); 2935 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2936 2937 // If not_null_obj is dead, only null-path is taken 2938 if (stopped()) { // Doing instance-of on a NULL? 2939 set_control(null_ctl); 2940 return intcon(0); 2941 } 2942 region->init_req(_null_path, null_ctl); 2943 phi ->init_req(_null_path, intcon(0)); // Set null path value 2944 if (null_ctl == top()) { 2945 // Do this eagerly, so that pattern matches like is_diamond_phi 2946 // will work even during parsing. 2947 assert(_null_path == PATH_LIMIT-1, "delete last"); 2948 region->del_req(_null_path); 2949 phi ->del_req(_null_path); 2950 } 2951 2952 // Do we know the type check always succeed? 2953 bool known_statically = false; 2954 if (_gvn.type(superklass)->singleton()) { 2955 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2956 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2957 if (subk != NULL && subk->is_loaded()) { 2958 int static_res = C->static_subtype_check(superk, subk); 2959 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2960 } 2961 } 2962 2963 if (!known_statically) { 2964 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2965 // We may not have profiling here or it may not help us. If we 2966 // have a speculative type use it to perform an exact cast. 2967 ciKlass* spec_obj_type = obj_type->speculative_type(); 2968 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2969 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2970 if (stopped()) { // Profile disagrees with this path. 2971 set_control(null_ctl); // Null is the only remaining possibility. 2972 return intcon(0); 2973 } 2974 if (cast_obj != NULL) { 2975 not_null_obj = cast_obj; 2976 } 2977 } 2978 } 2979 2980 // Load the object's klass 2981 Node* obj_klass = load_object_klass(not_null_obj); 2982 2983 // Generate the subtype check 2984 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2985 2986 // Plug in the success path to the general merge in slot 1. 2987 region->init_req(_obj_path, control()); 2988 phi ->init_req(_obj_path, intcon(1)); 2989 2990 // Plug in the failing path to the general merge in slot 2. 2991 region->init_req(_fail_path, not_subtype_ctrl); 2992 phi ->init_req(_fail_path, intcon(0)); 2993 2994 // Return final merged results 2995 set_control( _gvn.transform(region) ); 2996 record_for_igvn(region); 2997 2998 // If we know the type check always succeeds then we don't use the 2999 // profiling data at this bytecode. Don't lose it, feed it to the 3000 // type system as a speculative type. 3001 if (safe_for_replace) { 3002 Node* casted_obj = record_profiled_receiver_for_speculation(obj); 3003 replace_in_map(obj, casted_obj); 3004 } 3005 3006 return _gvn.transform(phi); 3007 } 3008 3009 //-------------------------------gen_checkcast--------------------------------- 3010 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 3011 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 3012 // uncommon-trap paths work. Adjust stack after this call. 3013 // If failure_control is supplied and not null, it is filled in with 3014 // the control edge for the cast failure. Otherwise, an appropriate 3015 // uncommon trap or exception is thrown. 3016 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 3017 Node* *failure_control) { 3018 kill_dead_locals(); // Benefit all the uncommon traps 3019 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 3020 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 3021 3022 // Fast cutout: Check the case that the cast is vacuously true. 3023 // This detects the common cases where the test will short-circuit 3024 // away completely. We do this before we perform the null check, 3025 // because if the test is going to turn into zero code, we don't 3026 // want a residual null check left around. (Causes a slowdown, 3027 // for example, in some objArray manipulations, such as a[i]=a[j].) 3028 if (tk->singleton()) { 3029 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 3030 if (objtp != NULL && objtp->klass() != NULL) { 3031 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 3032 case Compile::SSC_always_true: 3033 // If we know the type check always succeed then we don't use 3034 // the profiling data at this bytecode. Don't lose it, feed it 3035 // to the type system as a speculative type. 3036 return record_profiled_receiver_for_speculation(obj); 3037 case Compile::SSC_always_false: 3038 // It needs a null check because a null will *pass* the cast check. 3039 // A non-null value will always produce an exception. 3040 return null_assert(obj); 3041 } 3042 } 3043 } 3044 3045 ciProfileData* data = NULL; 3046 bool safe_for_replace = false; 3047 if (failure_control == NULL) { // use MDO in regular case only 3048 assert(java_bc() == Bytecodes::_aastore || 3049 java_bc() == Bytecodes::_checkcast, 3050 "interpreter profiles type checks only for these BCs"); 3051 data = method()->method_data()->bci_to_data(bci()); 3052 safe_for_replace = true; 3053 } 3054 3055 // Make the merge point 3056 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 3057 RegionNode* region = new RegionNode(PATH_LIMIT); 3058 Node* phi = new PhiNode(region, toop); 3059 C->set_has_split_ifs(true); // Has chance for split-if optimization 3060 3061 // Use null-cast information if it is available 3062 bool speculative_not_null = false; 3063 bool never_see_null = ((failure_control == NULL) // regular case only 3064 && seems_never_null(obj, data, speculative_not_null)); 3065 3066 // Null check; get casted pointer; set region slot 3 3067 Node* null_ctl = top(); 3068 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 3069 3070 // If not_null_obj is dead, only null-path is taken 3071 if (stopped()) { // Doing instance-of on a NULL? 3072 set_control(null_ctl); 3073 return null(); 3074 } 3075 region->init_req(_null_path, null_ctl); 3076 phi ->init_req(_null_path, null()); // Set null path value 3077 if (null_ctl == top()) { 3078 // Do this eagerly, so that pattern matches like is_diamond_phi 3079 // will work even during parsing. 3080 assert(_null_path == PATH_LIMIT-1, "delete last"); 3081 region->del_req(_null_path); 3082 phi ->del_req(_null_path); 3083 } 3084 3085 Node* cast_obj = NULL; 3086 if (tk->klass_is_exact()) { 3087 // The following optimization tries to statically cast the speculative type of the object 3088 // (for example obtained during profiling) to the type of the superklass and then do a 3089 // dynamic check that the type of the object is what we expect. To work correctly 3090 // for checkcast and aastore the type of superklass should be exact. 3091 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3092 // We may not have profiling here or it may not help us. If we have 3093 // a speculative type use it to perform an exact cast. 3094 ciKlass* spec_obj_type = obj_type->speculative_type(); 3095 if (spec_obj_type != NULL || data != NULL) { 3096 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3097 if (cast_obj != NULL) { 3098 if (failure_control != NULL) // failure is now impossible 3099 (*failure_control) = top(); 3100 // adjust the type of the phi to the exact klass: 3101 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3102 } 3103 } 3104 } 3105 3106 if (cast_obj == NULL) { 3107 // Load the object's klass 3108 Node* obj_klass = load_object_klass(not_null_obj); 3109 3110 // Generate the subtype check 3111 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3112 3113 // Plug in success path into the merge 3114 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3115 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3116 if (failure_control == NULL) { 3117 if (not_subtype_ctrl != top()) { // If failure is possible 3118 PreserveJVMState pjvms(this); 3119 set_control(not_subtype_ctrl); 3120 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3121 } 3122 } else { 3123 (*failure_control) = not_subtype_ctrl; 3124 } 3125 } 3126 3127 region->init_req(_obj_path, control()); 3128 phi ->init_req(_obj_path, cast_obj); 3129 3130 // A merge of NULL or Casted-NotNull obj 3131 Node* res = _gvn.transform(phi); 3132 3133 // Note I do NOT always 'replace_in_map(obj,result)' here. 3134 // if( tk->klass()->can_be_primary_super() ) 3135 // This means that if I successfully store an Object into an array-of-String 3136 // I 'forget' that the Object is really now known to be a String. I have to 3137 // do this because we don't have true union types for interfaces - if I store 3138 // a Baz into an array-of-Interface and then tell the optimizer it's an 3139 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3140 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3141 // replace_in_map( obj, res ); 3142 3143 // Return final merged results 3144 set_control( _gvn.transform(region) ); 3145 record_for_igvn(region); 3146 3147 return record_profiled_receiver_for_speculation(res); 3148 } 3149 3150 //------------------------------next_monitor----------------------------------- 3151 // What number should be given to the next monitor? 3152 int GraphKit::next_monitor() { 3153 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3154 int next = current + C->sync_stack_slots(); 3155 // Keep the toplevel high water mark current: 3156 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3157 return current; 3158 } 3159 3160 //------------------------------insert_mem_bar--------------------------------- 3161 // Memory barrier to avoid floating things around 3162 // The membar serves as a pinch point between both control and all memory slices. 3163 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3164 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3165 mb->init_req(TypeFunc::Control, control()); 3166 mb->init_req(TypeFunc::Memory, reset_memory()); 3167 Node* membar = _gvn.transform(mb); 3168 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3169 set_all_memory_call(membar); 3170 return membar; 3171 } 3172 3173 //-------------------------insert_mem_bar_volatile---------------------------- 3174 // Memory barrier to avoid floating things around 3175 // The membar serves as a pinch point between both control and memory(alias_idx). 3176 // If you want to make a pinch point on all memory slices, do not use this 3177 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3178 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3179 // When Parse::do_put_xxx updates a volatile field, it appends a series 3180 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3181 // The first membar is on the same memory slice as the field store opcode. 3182 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3183 // All the other membars (for other volatile slices, including AliasIdxBot, 3184 // which stands for all unknown volatile slices) are control-dependent 3185 // on the first membar. This prevents later volatile loads or stores 3186 // from sliding up past the just-emitted store. 3187 3188 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3189 mb->set_req(TypeFunc::Control,control()); 3190 if (alias_idx == Compile::AliasIdxBot) { 3191 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3192 } else { 3193 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3194 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3195 } 3196 Node* membar = _gvn.transform(mb); 3197 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3198 if (alias_idx == Compile::AliasIdxBot) { 3199 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3200 } else { 3201 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3202 } 3203 return membar; 3204 } 3205 3206 //------------------------------shared_lock------------------------------------ 3207 // Emit locking code. 3208 FastLockNode* GraphKit::shared_lock(Node* obj) { 3209 // bci is either a monitorenter bc or InvocationEntryBci 3210 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3211 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3212 3213 if( !GenerateSynchronizationCode ) 3214 return NULL; // Not locking things? 3215 if (stopped()) // Dead monitor? 3216 return NULL; 3217 3218 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3219 3220 // Box the stack location 3221 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3222 Node* mem = reset_memory(); 3223 3224 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3225 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3226 // Create the counters for this fast lock. 3227 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3228 } 3229 3230 // Create the rtm counters for this fast lock if needed. 3231 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3232 3233 // Add monitor to debug info for the slow path. If we block inside the 3234 // slow path and de-opt, we need the monitor hanging around 3235 map()->push_monitor( flock ); 3236 3237 const TypeFunc *tf = LockNode::lock_type(); 3238 LockNode *lock = new LockNode(C, tf); 3239 3240 lock->init_req( TypeFunc::Control, control() ); 3241 lock->init_req( TypeFunc::Memory , mem ); 3242 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3243 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3244 lock->init_req( TypeFunc::ReturnAdr, top() ); 3245 3246 lock->init_req(TypeFunc::Parms + 0, obj); 3247 lock->init_req(TypeFunc::Parms + 1, box); 3248 lock->init_req(TypeFunc::Parms + 2, flock); 3249 add_safepoint_edges(lock); 3250 3251 lock = _gvn.transform( lock )->as_Lock(); 3252 3253 // lock has no side-effects, sets few values 3254 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3255 3256 insert_mem_bar(Op_MemBarAcquireLock); 3257 3258 // Add this to the worklist so that the lock can be eliminated 3259 record_for_igvn(lock); 3260 3261 #ifndef PRODUCT 3262 if (PrintLockStatistics) { 3263 // Update the counter for this lock. Don't bother using an atomic 3264 // operation since we don't require absolute accuracy. 3265 lock->create_lock_counter(map()->jvms()); 3266 increment_counter(lock->counter()->addr()); 3267 } 3268 #endif 3269 3270 return flock; 3271 } 3272 3273 3274 //------------------------------shared_unlock---------------------------------- 3275 // Emit unlocking code. 3276 void GraphKit::shared_unlock(Node* box, Node* obj) { 3277 // bci is either a monitorenter bc or InvocationEntryBci 3278 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3279 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3280 3281 if( !GenerateSynchronizationCode ) 3282 return; 3283 if (stopped()) { // Dead monitor? 3284 map()->pop_monitor(); // Kill monitor from debug info 3285 return; 3286 } 3287 3288 // Memory barrier to avoid floating things down past the locked region 3289 insert_mem_bar(Op_MemBarReleaseLock); 3290 3291 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3292 UnlockNode *unlock = new UnlockNode(C, tf); 3293 #ifdef ASSERT 3294 unlock->set_dbg_jvms(sync_jvms()); 3295 #endif 3296 uint raw_idx = Compile::AliasIdxRaw; 3297 unlock->init_req( TypeFunc::Control, control() ); 3298 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3299 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3300 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3301 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3302 3303 unlock->init_req(TypeFunc::Parms + 0, obj); 3304 unlock->init_req(TypeFunc::Parms + 1, box); 3305 unlock = _gvn.transform(unlock)->as_Unlock(); 3306 3307 Node* mem = reset_memory(); 3308 3309 // unlock has no side-effects, sets few values 3310 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3311 3312 // Kill monitor from debug info 3313 map()->pop_monitor( ); 3314 } 3315 3316 //-------------------------------get_layout_helper----------------------------- 3317 // If the given klass is a constant or known to be an array, 3318 // fetch the constant layout helper value into constant_value 3319 // and return (Node*)NULL. Otherwise, load the non-constant 3320 // layout helper value, and return the node which represents it. 3321 // This two-faced routine is useful because allocation sites 3322 // almost always feature constant types. 3323 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3324 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3325 if (!StressReflectiveCode && inst_klass != NULL) { 3326 ciKlass* klass = inst_klass->klass(); 3327 bool xklass = inst_klass->klass_is_exact(); 3328 if (xklass || klass->is_array_klass()) { 3329 jint lhelper = klass->layout_helper(); 3330 if (lhelper != Klass::_lh_neutral_value) { 3331 constant_value = lhelper; 3332 return (Node*) NULL; 3333 } 3334 } 3335 } 3336 constant_value = Klass::_lh_neutral_value; // put in a known value 3337 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3338 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3339 } 3340 3341 // We just put in an allocate/initialize with a big raw-memory effect. 3342 // Hook selected additional alias categories on the initialization. 3343 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3344 MergeMemNode* init_in_merge, 3345 Node* init_out_raw) { 3346 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3347 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3348 3349 Node* prevmem = kit.memory(alias_idx); 3350 init_in_merge->set_memory_at(alias_idx, prevmem); 3351 kit.set_memory(init_out_raw, alias_idx); 3352 } 3353 3354 //---------------------------set_output_for_allocation------------------------- 3355 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3356 const TypeOopPtr* oop_type, 3357 bool deoptimize_on_exception) { 3358 int rawidx = Compile::AliasIdxRaw; 3359 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3360 add_safepoint_edges(alloc); 3361 Node* allocx = _gvn.transform(alloc); 3362 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3363 // create memory projection for i_o 3364 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3365 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3366 3367 // create a memory projection as for the normal control path 3368 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3369 set_memory(malloc, rawidx); 3370 3371 // a normal slow-call doesn't change i_o, but an allocation does 3372 // we create a separate i_o projection for the normal control path 3373 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3374 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3375 3376 // put in an initialization barrier 3377 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3378 rawoop)->as_Initialize(); 3379 assert(alloc->initialization() == init, "2-way macro link must work"); 3380 assert(init ->allocation() == alloc, "2-way macro link must work"); 3381 { 3382 // Extract memory strands which may participate in the new object's 3383 // initialization, and source them from the new InitializeNode. 3384 // This will allow us to observe initializations when they occur, 3385 // and link them properly (as a group) to the InitializeNode. 3386 assert(init->in(InitializeNode::Memory) == malloc, ""); 3387 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3388 init->set_req(InitializeNode::Memory, minit_in); 3389 record_for_igvn(minit_in); // fold it up later, if possible 3390 Node* minit_out = memory(rawidx); 3391 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3392 if (oop_type->isa_aryptr()) { 3393 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3394 int elemidx = C->get_alias_index(telemref); 3395 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3396 } else if (oop_type->isa_instptr()) { 3397 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3398 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3399 ciField* field = ik->nonstatic_field_at(i); 3400 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3401 continue; // do not bother to track really large numbers of fields 3402 // Find (or create) the alias category for this field: 3403 int fieldidx = C->alias_type(field)->index(); 3404 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3405 } 3406 } 3407 } 3408 3409 // Cast raw oop to the real thing... 3410 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3411 javaoop = _gvn.transform(javaoop); 3412 C->set_recent_alloc(control(), javaoop); 3413 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3414 3415 #ifdef ASSERT 3416 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3417 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3418 "Ideal_allocation works"); 3419 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3420 "Ideal_allocation works"); 3421 if (alloc->is_AllocateArray()) { 3422 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3423 "Ideal_allocation works"); 3424 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3425 "Ideal_allocation works"); 3426 } else { 3427 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3428 } 3429 } 3430 #endif //ASSERT 3431 3432 return javaoop; 3433 } 3434 3435 //---------------------------new_instance-------------------------------------- 3436 // This routine takes a klass_node which may be constant (for a static type) 3437 // or may be non-constant (for reflective code). It will work equally well 3438 // for either, and the graph will fold nicely if the optimizer later reduces 3439 // the type to a constant. 3440 // The optional arguments are for specialized use by intrinsics: 3441 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3442 // - If 'return_size_val', report the the total object size to the caller. 3443 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3444 Node* GraphKit::new_instance(Node* klass_node, 3445 Node* extra_slow_test, 3446 Node* *return_size_val, 3447 bool deoptimize_on_exception) { 3448 // Compute size in doublewords 3449 // The size is always an integral number of doublewords, represented 3450 // as a positive bytewise size stored in the klass's layout_helper. 3451 // The layout_helper also encodes (in a low bit) the need for a slow path. 3452 jint layout_con = Klass::_lh_neutral_value; 3453 Node* layout_val = get_layout_helper(klass_node, layout_con); 3454 int layout_is_con = (layout_val == NULL); 3455 3456 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3457 // Generate the initial go-slow test. It's either ALWAYS (return a 3458 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3459 // case) a computed value derived from the layout_helper. 3460 Node* initial_slow_test = NULL; 3461 if (layout_is_con) { 3462 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3463 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3464 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3465 } else { // reflective case 3466 // This reflective path is used by Unsafe.allocateInstance. 3467 // (It may be stress-tested by specifying StressReflectiveCode.) 3468 // Basically, we want to get into the VM is there's an illegal argument. 3469 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3470 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3471 if (extra_slow_test != intcon(0)) { 3472 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3473 } 3474 // (Macro-expander will further convert this to a Bool, if necessary.) 3475 } 3476 3477 // Find the size in bytes. This is easy; it's the layout_helper. 3478 // The size value must be valid even if the slow path is taken. 3479 Node* size = NULL; 3480 if (layout_is_con) { 3481 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3482 } else { // reflective case 3483 // This reflective path is used by clone and Unsafe.allocateInstance. 3484 size = ConvI2X(layout_val); 3485 3486 // Clear the low bits to extract layout_helper_size_in_bytes: 3487 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3488 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3489 size = _gvn.transform( new AndXNode(size, mask) ); 3490 } 3491 if (return_size_val != NULL) { 3492 (*return_size_val) = size; 3493 } 3494 3495 // This is a precise notnull oop of the klass. 3496 // (Actually, it need not be precise if this is a reflective allocation.) 3497 // It's what we cast the result to. 3498 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3499 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3500 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3501 3502 // Now generate allocation code 3503 3504 // The entire memory state is needed for slow path of the allocation 3505 // since GC and deoptimization can happened. 3506 Node *mem = reset_memory(); 3507 set_all_memory(mem); // Create new memory state 3508 3509 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3510 control(), mem, i_o(), 3511 size, klass_node, 3512 initial_slow_test); 3513 3514 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3515 } 3516 3517 //-------------------------------new_array------------------------------------- 3518 // helper for both newarray and anewarray 3519 // The 'length' parameter is (obviously) the length of the array. 3520 // See comments on new_instance for the meaning of the other arguments. 3521 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3522 Node* length, // number of array elements 3523 int nargs, // number of arguments to push back for uncommon trap 3524 Node* *return_size_val, 3525 bool deoptimize_on_exception) { 3526 jint layout_con = Klass::_lh_neutral_value; 3527 Node* layout_val = get_layout_helper(klass_node, layout_con); 3528 int layout_is_con = (layout_val == NULL); 3529 3530 if (!layout_is_con && !StressReflectiveCode && 3531 !too_many_traps(Deoptimization::Reason_class_check)) { 3532 // This is a reflective array creation site. 3533 // Optimistically assume that it is a subtype of Object[], 3534 // so that we can fold up all the address arithmetic. 3535 layout_con = Klass::array_layout_helper(T_OBJECT); 3536 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3537 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3538 { BuildCutout unless(this, bol_lh, PROB_MAX); 3539 inc_sp(nargs); 3540 uncommon_trap(Deoptimization::Reason_class_check, 3541 Deoptimization::Action_maybe_recompile); 3542 } 3543 layout_val = NULL; 3544 layout_is_con = true; 3545 } 3546 3547 // Generate the initial go-slow test. Make sure we do not overflow 3548 // if length is huge (near 2Gig) or negative! We do not need 3549 // exact double-words here, just a close approximation of needed 3550 // double-words. We can't add any offset or rounding bits, lest we 3551 // take a size -1 of bytes and make it positive. Use an unsigned 3552 // compare, so negative sizes look hugely positive. 3553 int fast_size_limit = FastAllocateSizeLimit; 3554 if (layout_is_con) { 3555 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3556 // Increase the size limit if we have exact knowledge of array type. 3557 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3558 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3559 } 3560 3561 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3562 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3563 3564 // --- Size Computation --- 3565 // array_size = round_to_heap(array_header + (length << elem_shift)); 3566 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes) 3567 // and align_to(x, y) == ((x + y-1) & ~(y-1)) 3568 // The rounding mask is strength-reduced, if possible. 3569 int round_mask = MinObjAlignmentInBytes - 1; 3570 Node* header_size = NULL; 3571 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3572 // (T_BYTE has the weakest alignment and size restrictions...) 3573 if (layout_is_con) { 3574 int hsize = Klass::layout_helper_header_size(layout_con); 3575 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3576 BasicType etype = Klass::layout_helper_element_type(layout_con); 3577 if ((round_mask & ~right_n_bits(eshift)) == 0) 3578 round_mask = 0; // strength-reduce it if it goes away completely 3579 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3580 assert(header_size_min <= hsize, "generic minimum is smallest"); 3581 header_size_min = hsize; 3582 header_size = intcon(hsize + round_mask); 3583 } else { 3584 Node* hss = intcon(Klass::_lh_header_size_shift); 3585 Node* hsm = intcon(Klass::_lh_header_size_mask); 3586 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3587 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3588 Node* mask = intcon(round_mask); 3589 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3590 } 3591 3592 Node* elem_shift = NULL; 3593 if (layout_is_con) { 3594 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3595 if (eshift != 0) 3596 elem_shift = intcon(eshift); 3597 } else { 3598 // There is no need to mask or shift this value. 3599 // The semantics of LShiftINode include an implicit mask to 0x1F. 3600 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3601 elem_shift = layout_val; 3602 } 3603 3604 // Transition to native address size for all offset calculations: 3605 Node* lengthx = ConvI2X(length); 3606 Node* headerx = ConvI2X(header_size); 3607 #ifdef _LP64 3608 { const TypeInt* tilen = _gvn.find_int_type(length); 3609 if (tilen != NULL && tilen->_lo < 0) { 3610 // Add a manual constraint to a positive range. Cf. array_element_address. 3611 jint size_max = fast_size_limit; 3612 if (size_max > tilen->_hi) size_max = tilen->_hi; 3613 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3614 3615 // Only do a narrow I2L conversion if the range check passed. 3616 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3617 _gvn.transform(iff); 3618 RegionNode* region = new RegionNode(3); 3619 _gvn.set_type(region, Type::CONTROL); 3620 lengthx = new PhiNode(region, TypeLong::LONG); 3621 _gvn.set_type(lengthx, TypeLong::LONG); 3622 3623 // Range check passed. Use ConvI2L node with narrow type. 3624 Node* passed = IfFalse(iff); 3625 region->init_req(1, passed); 3626 // Make I2L conversion control dependent to prevent it from 3627 // floating above the range check during loop optimizations. 3628 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3629 3630 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3631 region->init_req(2, IfTrue(iff)); 3632 lengthx->init_req(2, ConvI2X(length)); 3633 3634 set_control(region); 3635 record_for_igvn(region); 3636 record_for_igvn(lengthx); 3637 } 3638 } 3639 #endif 3640 3641 // Combine header size (plus rounding) and body size. Then round down. 3642 // This computation cannot overflow, because it is used only in two 3643 // places, one where the length is sharply limited, and the other 3644 // after a successful allocation. 3645 Node* abody = lengthx; 3646 if (elem_shift != NULL) 3647 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3648 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3649 if (round_mask != 0) { 3650 Node* mask = MakeConX(~round_mask); 3651 size = _gvn.transform( new AndXNode(size, mask) ); 3652 } 3653 // else if round_mask == 0, the size computation is self-rounding 3654 3655 if (return_size_val != NULL) { 3656 // This is the size 3657 (*return_size_val) = size; 3658 } 3659 3660 // Now generate allocation code 3661 3662 // The entire memory state is needed for slow path of the allocation 3663 // since GC and deoptimization can happened. 3664 Node *mem = reset_memory(); 3665 set_all_memory(mem); // Create new memory state 3666 3667 if (initial_slow_test->is_Bool()) { 3668 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3669 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3670 } 3671 3672 // Create the AllocateArrayNode and its result projections 3673 AllocateArrayNode* alloc 3674 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3675 control(), mem, i_o(), 3676 size, klass_node, 3677 initial_slow_test, 3678 length); 3679 3680 // Cast to correct type. Note that the klass_node may be constant or not, 3681 // and in the latter case the actual array type will be inexact also. 3682 // (This happens via a non-constant argument to inline_native_newArray.) 3683 // In any case, the value of klass_node provides the desired array type. 3684 const TypeInt* length_type = _gvn.find_int_type(length); 3685 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3686 if (ary_type->isa_aryptr() && length_type != NULL) { 3687 // Try to get a better type than POS for the size 3688 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3689 } 3690 3691 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3692 3693 // Cast length on remaining path to be as narrow as possible 3694 if (map()->find_edge(length) >= 0) { 3695 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3696 if (ccast != length) { 3697 _gvn.set_type_bottom(ccast); 3698 record_for_igvn(ccast); 3699 replace_in_map(length, ccast); 3700 } 3701 } 3702 3703 return javaoop; 3704 } 3705 3706 // The following "Ideal_foo" functions are placed here because they recognize 3707 // the graph shapes created by the functions immediately above. 3708 3709 //---------------------------Ideal_allocation---------------------------------- 3710 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3711 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3712 if (ptr == NULL) { // reduce dumb test in callers 3713 return NULL; 3714 } 3715 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3716 ptr = ptr->in(1); 3717 if (ptr == NULL) return NULL; 3718 } 3719 // Return NULL for allocations with several casts: 3720 // j.l.reflect.Array.newInstance(jobject, jint) 3721 // Object.clone() 3722 // to keep more precise type from last cast. 3723 if (ptr->is_Proj()) { 3724 Node* allo = ptr->in(0); 3725 if (allo != NULL && allo->is_Allocate()) { 3726 return allo->as_Allocate(); 3727 } 3728 } 3729 // Report failure to match. 3730 return NULL; 3731 } 3732 3733 // Fancy version which also strips off an offset (and reports it to caller). 3734 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3735 intptr_t& offset) { 3736 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3737 if (base == NULL) return NULL; 3738 return Ideal_allocation(base, phase); 3739 } 3740 3741 // Trace Initialize <- Proj[Parm] <- Allocate 3742 AllocateNode* InitializeNode::allocation() { 3743 Node* rawoop = in(InitializeNode::RawAddress); 3744 if (rawoop->is_Proj()) { 3745 Node* alloc = rawoop->in(0); 3746 if (alloc->is_Allocate()) { 3747 return alloc->as_Allocate(); 3748 } 3749 } 3750 return NULL; 3751 } 3752 3753 // Trace Allocate -> Proj[Parm] -> Initialize 3754 InitializeNode* AllocateNode::initialization() { 3755 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress); 3756 if (rawoop == NULL) return NULL; 3757 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3758 Node* init = rawoop->fast_out(i); 3759 if (init->is_Initialize()) { 3760 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3761 return init->as_Initialize(); 3762 } 3763 } 3764 return NULL; 3765 } 3766 3767 //----------------------------- loop predicates --------------------------- 3768 3769 //------------------------------add_predicate_impl---------------------------- 3770 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3771 // Too many traps seen? 3772 if (too_many_traps(reason)) { 3773 #ifdef ASSERT 3774 if (TraceLoopPredicate) { 3775 int tc = C->trap_count(reason); 3776 tty->print("too many traps=%s tcount=%d in ", 3777 Deoptimization::trap_reason_name(reason), tc); 3778 method()->print(); // which method has too many predicate traps 3779 tty->cr(); 3780 } 3781 #endif 3782 // We cannot afford to take more traps here, 3783 // do not generate predicate. 3784 return; 3785 } 3786 3787 Node *cont = _gvn.intcon(1); 3788 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3789 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3790 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3791 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3792 C->add_predicate_opaq(opq); 3793 { 3794 PreserveJVMState pjvms(this); 3795 set_control(iffalse); 3796 inc_sp(nargs); 3797 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3798 } 3799 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3800 set_control(iftrue); 3801 } 3802 3803 //------------------------------add_predicate--------------------------------- 3804 void GraphKit::add_predicate(int nargs) { 3805 if (UseLoopPredicate) { 3806 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3807 } 3808 // loop's limit check predicate should be near the loop. 3809 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3810 } 3811 3812 void GraphKit::sync_kit(IdealKit& ideal) { 3813 set_all_memory(ideal.merged_memory()); 3814 set_i_o(ideal.i_o()); 3815 set_control(ideal.ctrl()); 3816 } 3817 3818 void GraphKit::final_sync(IdealKit& ideal) { 3819 // Final sync IdealKit and graphKit. 3820 sync_kit(ideal); 3821 } 3822 3823 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 3824 Node* len = load_array_length(load_String_value(ctrl, str)); 3825 Node* coder = load_String_coder(ctrl, str); 3826 // Divide length by 2 if coder is UTF16 3827 return _gvn.transform(new RShiftINode(len, coder)); 3828 } 3829 3830 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 3831 int value_offset = java_lang_String::value_offset_in_bytes(); 3832 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3833 false, NULL, 0); 3834 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3835 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 3836 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 3837 ciTypeArrayKlass::make(T_BYTE), true, 0); 3838 Node* p = basic_plus_adr(str, str, value_offset); 3839 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, 3840 IN_HEAP | C2_CONTROL_DEPENDENT_LOAD); 3841 // String.value field is known to be @Stable. 3842 if (UseImplicitStableValues) { 3843 load = cast_array_to_stable(load, value_type); 3844 } 3845 return load; 3846 } 3847 3848 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 3849 if (!CompactStrings) { 3850 return intcon(java_lang_String::CODER_UTF16); 3851 } 3852 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3853 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3854 false, NULL, 0); 3855 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3856 int coder_field_idx = C->get_alias_index(coder_field_type); 3857 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 3858 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 3859 } 3860 3861 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 3862 int value_offset = java_lang_String::value_offset_in_bytes(); 3863 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3864 false, NULL, 0); 3865 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3866 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 3867 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP); 3868 } 3869 3870 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 3871 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3872 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3873 false, NULL, 0); 3874 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3875 int coder_field_idx = C->get_alias_index(coder_field_type); 3876 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 3877 value, T_BYTE, coder_field_idx, MemNode::unordered); 3878 } 3879 3880 // Capture src and dst memory state with a MergeMemNode 3881 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 3882 if (src_type == dst_type) { 3883 // Types are equal, we don't need a MergeMemNode 3884 return memory(src_type); 3885 } 3886 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 3887 record_for_igvn(merge); // fold it up later, if possible 3888 int src_idx = C->get_alias_index(src_type); 3889 int dst_idx = C->get_alias_index(dst_type); 3890 merge->set_memory_at(src_idx, memory(src_idx)); 3891 merge->set_memory_at(dst_idx, memory(dst_idx)); 3892 return merge; 3893 } 3894 3895 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 3896 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 3897 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 3898 // If input and output memory types differ, capture both states to preserve 3899 // the dependency between preceding and subsequent loads/stores. 3900 // For example, the following program: 3901 // StoreB 3902 // compress_string 3903 // LoadB 3904 // has this memory graph (use->def): 3905 // LoadB -> compress_string -> CharMem 3906 // ... -> StoreB -> ByteMem 3907 // The intrinsic hides the dependency between LoadB and StoreB, causing 3908 // the load to read from memory not containing the result of the StoreB. 3909 // The correct memory graph should look like this: 3910 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 3911 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 3912 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 3913 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 3914 set_memory(res_mem, TypeAryPtr::BYTES); 3915 return str; 3916 } 3917 3918 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 3919 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 3920 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 3921 // Capture src and dst memory (see comment in 'compress_string'). 3922 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 3923 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 3924 set_memory(_gvn.transform(str), dst_type); 3925 } 3926 3927 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 3928 /** 3929 * int i_char = start; 3930 * for (int i_byte = 0; i_byte < count; i_byte++) { 3931 * dst[i_char++] = (char)(src[i_byte] & 0xff); 3932 * } 3933 */ 3934 add_predicate(); 3935 RegionNode* head = new RegionNode(3); 3936 head->init_req(1, control()); 3937 gvn().set_type(head, Type::CONTROL); 3938 record_for_igvn(head); 3939 3940 Node* i_byte = new PhiNode(head, TypeInt::INT); 3941 i_byte->init_req(1, intcon(0)); 3942 gvn().set_type(i_byte, TypeInt::INT); 3943 record_for_igvn(i_byte); 3944 3945 Node* i_char = new PhiNode(head, TypeInt::INT); 3946 i_char->init_req(1, start); 3947 gvn().set_type(i_char, TypeInt::INT); 3948 record_for_igvn(i_char); 3949 3950 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 3951 gvn().set_type(mem, Type::MEMORY); 3952 record_for_igvn(mem); 3953 set_control(head); 3954 set_memory(mem, TypeAryPtr::BYTES); 3955 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 3956 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 3957 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 3958 false, false, true /* mismatched */); 3959 3960 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 3961 head->init_req(2, IfTrue(iff)); 3962 mem->init_req(2, st); 3963 i_byte->init_req(2, AddI(i_byte, intcon(1))); 3964 i_char->init_req(2, AddI(i_char, intcon(2))); 3965 3966 set_control(IfFalse(iff)); 3967 set_memory(st, TypeAryPtr::BYTES); 3968 } 3969 3970 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 3971 if (!field->is_constant()) { 3972 return NULL; // Field not marked as constant. 3973 } 3974 ciInstance* holder = NULL; 3975 if (!field->is_static()) { 3976 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 3977 if (const_oop != NULL && const_oop->is_instance()) { 3978 holder = const_oop->as_instance(); 3979 } 3980 } 3981 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 3982 /*is_unsigned_load=*/false); 3983 if (con_type != NULL) { 3984 return makecon(con_type); 3985 } 3986 return NULL; 3987 } 3988 3989 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 3990 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 3991 // assumption of CCP analysis. 3992 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 3993 }