1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "gc/g1/heapRegion.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/cardTableModRefBS.hpp" 30 #include "gc/shared/collectedHeap.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/graphKit.hpp" 36 #include "opto/idealKit.hpp" 37 #include "opto/intrinsicnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/rootnode.hpp" 43 #include "opto/runtime.hpp" 44 #include "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 47 //----------------------------GraphKit----------------------------------------- 48 // Main utility constructor. 49 GraphKit::GraphKit(JVMState* jvms) 50 : Phase(Phase::Parser), 51 _env(C->env()), 52 _gvn(*C->initial_gvn()) 53 { 54 _exceptions = jvms->map()->next_exception(); 55 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL); 56 set_jvms(jvms); 57 } 58 59 // Private constructor for parser. 60 GraphKit::GraphKit() 61 : Phase(Phase::Parser), 62 _env(C->env()), 63 _gvn(*C->initial_gvn()) 64 { 65 _exceptions = NULL; 66 set_map(NULL); 67 debug_only(_sp = -99); 68 debug_only(set_bci(-99)); 69 } 70 71 72 73 //---------------------------clean_stack--------------------------------------- 74 // Clear away rubbish from the stack area of the JVM state. 75 // This destroys any arguments that may be waiting on the stack. 76 void GraphKit::clean_stack(int from_sp) { 77 SafePointNode* map = this->map(); 78 JVMState* jvms = this->jvms(); 79 int stk_size = jvms->stk_size(); 80 int stkoff = jvms->stkoff(); 81 Node* top = this->top(); 82 for (int i = from_sp; i < stk_size; i++) { 83 if (map->in(stkoff + i) != top) { 84 map->set_req(stkoff + i, top); 85 } 86 } 87 } 88 89 90 //--------------------------------sync_jvms----------------------------------- 91 // Make sure our current jvms agrees with our parse state. 92 JVMState* GraphKit::sync_jvms() const { 93 JVMState* jvms = this->jvms(); 94 jvms->set_bci(bci()); // Record the new bci in the JVMState 95 jvms->set_sp(sp()); // Record the new sp in the JVMState 96 assert(jvms_in_sync(), "jvms is now in sync"); 97 return jvms; 98 } 99 100 //--------------------------------sync_jvms_for_reexecute--------------------- 101 // Make sure our current jvms agrees with our parse state. This version 102 // uses the reexecute_sp for reexecuting bytecodes. 103 JVMState* GraphKit::sync_jvms_for_reexecute() { 104 JVMState* jvms = this->jvms(); 105 jvms->set_bci(bci()); // Record the new bci in the JVMState 106 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState 107 return jvms; 108 } 109 110 #ifdef ASSERT 111 bool GraphKit::jvms_in_sync() const { 112 Parse* parse = is_Parse(); 113 if (parse == NULL) { 114 if (bci() != jvms()->bci()) return false; 115 if (sp() != (int)jvms()->sp()) return false; 116 return true; 117 } 118 if (jvms()->method() != parse->method()) return false; 119 if (jvms()->bci() != parse->bci()) return false; 120 int jvms_sp = jvms()->sp(); 121 if (jvms_sp != parse->sp()) return false; 122 int jvms_depth = jvms()->depth(); 123 if (jvms_depth != parse->depth()) return false; 124 return true; 125 } 126 127 // Local helper checks for special internal merge points 128 // used to accumulate and merge exception states. 129 // They are marked by the region's in(0) edge being the map itself. 130 // Such merge points must never "escape" into the parser at large, 131 // until they have been handed to gvn.transform. 132 static bool is_hidden_merge(Node* reg) { 133 if (reg == NULL) return false; 134 if (reg->is_Phi()) { 135 reg = reg->in(0); 136 if (reg == NULL) return false; 137 } 138 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root(); 139 } 140 141 void GraphKit::verify_map() const { 142 if (map() == NULL) return; // null map is OK 143 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map"); 144 assert(!map()->has_exceptions(), "call add_exception_states_from 1st"); 145 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map"); 146 } 147 148 void GraphKit::verify_exception_state(SafePointNode* ex_map) { 149 assert(ex_map->next_exception() == NULL, "not already part of a chain"); 150 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop"); 151 } 152 #endif 153 154 //---------------------------stop_and_kill_map--------------------------------- 155 // Set _map to NULL, signalling a stop to further bytecode execution. 156 // First smash the current map's control to a constant, to mark it dead. 157 void GraphKit::stop_and_kill_map() { 158 SafePointNode* dead_map = stop(); 159 if (dead_map != NULL) { 160 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed. 161 assert(dead_map->is_killed(), "must be so marked"); 162 } 163 } 164 165 166 //--------------------------------stopped-------------------------------------- 167 // Tell if _map is NULL, or control is top. 168 bool GraphKit::stopped() { 169 if (map() == NULL) return true; 170 else if (control() == top()) return true; 171 else return false; 172 } 173 174 175 //-----------------------------has_ex_handler---------------------------------- 176 // Tell if this method or any caller method has exception handlers. 177 bool GraphKit::has_ex_handler() { 178 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) { 179 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) { 180 return true; 181 } 182 } 183 return false; 184 } 185 186 //------------------------------save_ex_oop------------------------------------ 187 // Save an exception without blowing stack contents or other JVM state. 188 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) { 189 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again"); 190 ex_map->add_req(ex_oop); 191 debug_only(verify_exception_state(ex_map)); 192 } 193 194 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) { 195 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there"); 196 Node* ex_oop = ex_map->in(ex_map->req()-1); 197 if (clear_it) ex_map->del_req(ex_map->req()-1); 198 return ex_oop; 199 } 200 201 //-----------------------------saved_ex_oop------------------------------------ 202 // Recover a saved exception from its map. 203 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) { 204 return common_saved_ex_oop(ex_map, false); 205 } 206 207 //--------------------------clear_saved_ex_oop--------------------------------- 208 // Erase a previously saved exception from its map. 209 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) { 210 return common_saved_ex_oop(ex_map, true); 211 } 212 213 #ifdef ASSERT 214 //---------------------------has_saved_ex_oop---------------------------------- 215 // Erase a previously saved exception from its map. 216 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) { 217 return ex_map->req() == ex_map->jvms()->endoff()+1; 218 } 219 #endif 220 221 //-------------------------make_exception_state-------------------------------- 222 // Turn the current JVM state into an exception state, appending the ex_oop. 223 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) { 224 sync_jvms(); 225 SafePointNode* ex_map = stop(); // do not manipulate this map any more 226 set_saved_ex_oop(ex_map, ex_oop); 227 return ex_map; 228 } 229 230 231 //--------------------------add_exception_state-------------------------------- 232 // Add an exception to my list of exceptions. 233 void GraphKit::add_exception_state(SafePointNode* ex_map) { 234 if (ex_map == NULL || ex_map->control() == top()) { 235 return; 236 } 237 #ifdef ASSERT 238 verify_exception_state(ex_map); 239 if (has_exceptions()) { 240 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place"); 241 } 242 #endif 243 244 // If there is already an exception of exactly this type, merge with it. 245 // In particular, null-checks and other low-level exceptions common up here. 246 Node* ex_oop = saved_ex_oop(ex_map); 247 const Type* ex_type = _gvn.type(ex_oop); 248 if (ex_oop == top()) { 249 // No action needed. 250 return; 251 } 252 assert(ex_type->isa_instptr(), "exception must be an instance"); 253 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) { 254 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2)); 255 // We check sp also because call bytecodes can generate exceptions 256 // both before and after arguments are popped! 257 if (ex_type2 == ex_type 258 && e2->_jvms->sp() == ex_map->_jvms->sp()) { 259 combine_exception_states(ex_map, e2); 260 return; 261 } 262 } 263 264 // No pre-existing exception of the same type. Chain it on the list. 265 push_exception_state(ex_map); 266 } 267 268 //-----------------------add_exception_states_from----------------------------- 269 void GraphKit::add_exception_states_from(JVMState* jvms) { 270 SafePointNode* ex_map = jvms->map()->next_exception(); 271 if (ex_map != NULL) { 272 jvms->map()->set_next_exception(NULL); 273 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) { 274 next_map = ex_map->next_exception(); 275 ex_map->set_next_exception(NULL); 276 add_exception_state(ex_map); 277 } 278 } 279 } 280 281 //-----------------------transfer_exceptions_into_jvms------------------------- 282 JVMState* GraphKit::transfer_exceptions_into_jvms() { 283 if (map() == NULL) { 284 // We need a JVMS to carry the exceptions, but the map has gone away. 285 // Create a scratch JVMS, cloned from any of the exception states... 286 if (has_exceptions()) { 287 _map = _exceptions; 288 _map = clone_map(); 289 _map->set_next_exception(NULL); 290 clear_saved_ex_oop(_map); 291 debug_only(verify_map()); 292 } else { 293 // ...or created from scratch 294 JVMState* jvms = new (C) JVMState(_method, NULL); 295 jvms->set_bci(_bci); 296 jvms->set_sp(_sp); 297 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms)); 298 set_jvms(jvms); 299 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top()); 300 set_all_memory(top()); 301 while (map()->req() < jvms->endoff()) map()->add_req(top()); 302 } 303 // (This is a kludge, in case you didn't notice.) 304 set_control(top()); 305 } 306 JVMState* jvms = sync_jvms(); 307 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet"); 308 jvms->map()->set_next_exception(_exceptions); 309 _exceptions = NULL; // done with this set of exceptions 310 return jvms; 311 } 312 313 static inline void add_n_reqs(Node* dstphi, Node* srcphi) { 314 assert(is_hidden_merge(dstphi), "must be a special merge node"); 315 assert(is_hidden_merge(srcphi), "must be a special merge node"); 316 uint limit = srcphi->req(); 317 for (uint i = PhiNode::Input; i < limit; i++) { 318 dstphi->add_req(srcphi->in(i)); 319 } 320 } 321 static inline void add_one_req(Node* dstphi, Node* src) { 322 assert(is_hidden_merge(dstphi), "must be a special merge node"); 323 assert(!is_hidden_merge(src), "must not be a special merge node"); 324 dstphi->add_req(src); 325 } 326 327 //-----------------------combine_exception_states------------------------------ 328 // This helper function combines exception states by building phis on a 329 // specially marked state-merging region. These regions and phis are 330 // untransformed, and can build up gradually. The region is marked by 331 // having a control input of its exception map, rather than NULL. Such 332 // regions do not appear except in this function, and in use_exception_state. 333 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) { 334 if (failing()) return; // dying anyway... 335 JVMState* ex_jvms = ex_map->_jvms; 336 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains"); 337 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals"); 338 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes"); 339 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS"); 340 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects"); 341 assert(ex_map->req() == phi_map->req(), "matching maps"); 342 uint tos = ex_jvms->stkoff() + ex_jvms->sp(); 343 Node* hidden_merge_mark = root(); 344 Node* region = phi_map->control(); 345 MergeMemNode* phi_mem = phi_map->merged_memory(); 346 MergeMemNode* ex_mem = ex_map->merged_memory(); 347 if (region->in(0) != hidden_merge_mark) { 348 // The control input is not (yet) a specially-marked region in phi_map. 349 // Make it so, and build some phis. 350 region = new RegionNode(2); 351 _gvn.set_type(region, Type::CONTROL); 352 region->set_req(0, hidden_merge_mark); // marks an internal ex-state 353 region->init_req(1, phi_map->control()); 354 phi_map->set_control(region); 355 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO); 356 record_for_igvn(io_phi); 357 _gvn.set_type(io_phi, Type::ABIO); 358 phi_map->set_i_o(io_phi); 359 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) { 360 Node* m = mms.memory(); 361 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C)); 362 record_for_igvn(m_phi); 363 _gvn.set_type(m_phi, Type::MEMORY); 364 mms.set_memory(m_phi); 365 } 366 } 367 368 // Either or both of phi_map and ex_map might already be converted into phis. 369 Node* ex_control = ex_map->control(); 370 // if there is special marking on ex_map also, we add multiple edges from src 371 bool add_multiple = (ex_control->in(0) == hidden_merge_mark); 372 // how wide was the destination phi_map, originally? 373 uint orig_width = region->req(); 374 375 if (add_multiple) { 376 add_n_reqs(region, ex_control); 377 add_n_reqs(phi_map->i_o(), ex_map->i_o()); 378 } else { 379 // ex_map has no merges, so we just add single edges everywhere 380 add_one_req(region, ex_control); 381 add_one_req(phi_map->i_o(), ex_map->i_o()); 382 } 383 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) { 384 if (mms.is_empty()) { 385 // get a copy of the base memory, and patch some inputs into it 386 const TypePtr* adr_type = mms.adr_type(C); 387 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 388 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 389 mms.set_memory(phi); 390 // Prepare to append interesting stuff onto the newly sliced phi: 391 while (phi->req() > orig_width) phi->del_req(phi->req()-1); 392 } 393 // Append stuff from ex_map: 394 if (add_multiple) { 395 add_n_reqs(mms.memory(), mms.memory2()); 396 } else { 397 add_one_req(mms.memory(), mms.memory2()); 398 } 399 } 400 uint limit = ex_map->req(); 401 for (uint i = TypeFunc::Parms; i < limit; i++) { 402 // Skip everything in the JVMS after tos. (The ex_oop follows.) 403 if (i == tos) i = ex_jvms->monoff(); 404 Node* src = ex_map->in(i); 405 Node* dst = phi_map->in(i); 406 if (src != dst) { 407 PhiNode* phi; 408 if (dst->in(0) != region) { 409 dst = phi = PhiNode::make(region, dst, _gvn.type(dst)); 410 record_for_igvn(phi); 411 _gvn.set_type(phi, phi->type()); 412 phi_map->set_req(i, dst); 413 // Prepare to append interesting stuff onto the new phi: 414 while (dst->req() > orig_width) dst->del_req(dst->req()-1); 415 } else { 416 assert(dst->is_Phi(), "nobody else uses a hidden region"); 417 phi = dst->as_Phi(); 418 } 419 if (add_multiple && src->in(0) == ex_control) { 420 // Both are phis. 421 add_n_reqs(dst, src); 422 } else { 423 while (dst->req() < region->req()) add_one_req(dst, src); 424 } 425 const Type* srctype = _gvn.type(src); 426 if (phi->type() != srctype) { 427 const Type* dsttype = phi->type()->meet_speculative(srctype); 428 if (phi->type() != dsttype) { 429 phi->set_type(dsttype); 430 _gvn.set_type(phi, dsttype); 431 } 432 } 433 } 434 } 435 phi_map->merge_replaced_nodes_with(ex_map); 436 } 437 438 //--------------------------use_exception_state-------------------------------- 439 Node* GraphKit::use_exception_state(SafePointNode* phi_map) { 440 if (failing()) { stop(); return top(); } 441 Node* region = phi_map->control(); 442 Node* hidden_merge_mark = root(); 443 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation"); 444 Node* ex_oop = clear_saved_ex_oop(phi_map); 445 if (region->in(0) == hidden_merge_mark) { 446 // Special marking for internal ex-states. Process the phis now. 447 region->set_req(0, region); // now it's an ordinary region 448 set_jvms(phi_map->jvms()); // ...so now we can use it as a map 449 // Note: Setting the jvms also sets the bci and sp. 450 set_control(_gvn.transform(region)); 451 uint tos = jvms()->stkoff() + sp(); 452 for (uint i = 1; i < tos; i++) { 453 Node* x = phi_map->in(i); 454 if (x->in(0) == region) { 455 assert(x->is_Phi(), "expected a special phi"); 456 phi_map->set_req(i, _gvn.transform(x)); 457 } 458 } 459 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 460 Node* x = mms.memory(); 461 if (x->in(0) == region) { 462 assert(x->is_Phi(), "nobody else uses a hidden region"); 463 mms.set_memory(_gvn.transform(x)); 464 } 465 } 466 if (ex_oop->in(0) == region) { 467 assert(ex_oop->is_Phi(), "expected a special phi"); 468 ex_oop = _gvn.transform(ex_oop); 469 } 470 } else { 471 set_jvms(phi_map->jvms()); 472 } 473 474 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared"); 475 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared"); 476 return ex_oop; 477 } 478 479 //---------------------------------java_bc------------------------------------- 480 Bytecodes::Code GraphKit::java_bc() const { 481 ciMethod* method = this->method(); 482 int bci = this->bci(); 483 if (method != NULL && bci != InvocationEntryBci) 484 return method->java_code_at_bci(bci); 485 else 486 return Bytecodes::_illegal; 487 } 488 489 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason, 490 bool must_throw) { 491 // if the exception capability is set, then we will generate code 492 // to check the JavaThread.should_post_on_exceptions flag to see 493 // if we actually need to report exception events (for this 494 // thread). If we don't need to report exception events, we will 495 // take the normal fast path provided by add_exception_events. If 496 // exception event reporting is enabled for this thread, we will 497 // take the uncommon_trap in the BuildCutout below. 498 499 // first must access the should_post_on_exceptions_flag in this thread's JavaThread 500 Node* jthread = _gvn.transform(new ThreadLocalNode()); 501 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset())); 502 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered); 503 504 // Test the should_post_on_exceptions_flag vs. 0 505 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) ); 506 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 507 508 // Branch to slow_path if should_post_on_exceptions_flag was true 509 { BuildCutout unless(this, tst, PROB_MAX); 510 // Do not try anything fancy if we're notifying the VM on every throw. 511 // Cf. case Bytecodes::_athrow in parse2.cpp. 512 uncommon_trap(reason, Deoptimization::Action_none, 513 (ciKlass*)NULL, (char*)NULL, must_throw); 514 } 515 516 } 517 518 //------------------------------builtin_throw---------------------------------- 519 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) { 520 bool must_throw = true; 521 522 if (env()->jvmti_can_post_on_exceptions()) { 523 // check if we must post exception events, take uncommon trap if so 524 uncommon_trap_if_should_post_on_exceptions(reason, must_throw); 525 // here if should_post_on_exceptions is false 526 // continue on with the normal codegen 527 } 528 529 // If this particular condition has not yet happened at this 530 // bytecode, then use the uncommon trap mechanism, and allow for 531 // a future recompilation if several traps occur here. 532 // If the throw is hot, try to use a more complicated inline mechanism 533 // which keeps execution inside the compiled code. 534 bool treat_throw_as_hot = false; 535 ciMethodData* md = method()->method_data(); 536 537 if (ProfileTraps) { 538 if (too_many_traps(reason)) { 539 treat_throw_as_hot = true; 540 } 541 // (If there is no MDO at all, assume it is early in 542 // execution, and that any deopts are part of the 543 // startup transient, and don't need to be remembered.) 544 545 // Also, if there is a local exception handler, treat all throws 546 // as hot if there has been at least one in this method. 547 if (C->trap_count(reason) != 0 548 && method()->method_data()->trap_count(reason) != 0 549 && has_ex_handler()) { 550 treat_throw_as_hot = true; 551 } 552 } 553 554 // If this throw happens frequently, an uncommon trap might cause 555 // a performance pothole. If there is a local exception handler, 556 // and if this particular bytecode appears to be deoptimizing often, 557 // let us handle the throw inline, with a preconstructed instance. 558 // Note: If the deopt count has blown up, the uncommon trap 559 // runtime is going to flush this nmethod, not matter what. 560 if (treat_throw_as_hot 561 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) { 562 // If the throw is local, we use a pre-existing instance and 563 // punt on the backtrace. This would lead to a missing backtrace 564 // (a repeat of 4292742) if the backtrace object is ever asked 565 // for its backtrace. 566 // Fixing this remaining case of 4292742 requires some flavor of 567 // escape analysis. Leave that for the future. 568 ciInstance* ex_obj = NULL; 569 switch (reason) { 570 case Deoptimization::Reason_null_check: 571 ex_obj = env()->NullPointerException_instance(); 572 break; 573 case Deoptimization::Reason_div0_check: 574 ex_obj = env()->ArithmeticException_instance(); 575 break; 576 case Deoptimization::Reason_range_check: 577 ex_obj = env()->ArrayIndexOutOfBoundsException_instance(); 578 break; 579 case Deoptimization::Reason_class_check: 580 if (java_bc() == Bytecodes::_aastore) { 581 ex_obj = env()->ArrayStoreException_instance(); 582 } else { 583 ex_obj = env()->ClassCastException_instance(); 584 } 585 break; 586 } 587 if (failing()) { stop(); return; } // exception allocation might fail 588 if (ex_obj != NULL) { 589 // Cheat with a preallocated exception object. 590 if (C->log() != NULL) 591 C->log()->elem("hot_throw preallocated='1' reason='%s'", 592 Deoptimization::trap_reason_name(reason)); 593 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj); 594 Node* ex_node = _gvn.transform(ConNode::make(ex_con)); 595 596 // Clear the detail message of the preallocated exception object. 597 // Weblogic sometimes mutates the detail message of exceptions 598 // using reflection. 599 int offset = java_lang_Throwable::get_detailMessage_offset(); 600 const TypePtr* adr_typ = ex_con->add_offset(offset); 601 602 Node *adr = basic_plus_adr(ex_node, ex_node, offset); 603 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass()); 604 Node *store = access_store_at(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, C2_ACCESS_ON_HEAP); 605 606 add_exception_state(make_exception_state(ex_node)); 607 return; 608 } 609 } 610 611 // %%% Maybe add entry to OptoRuntime which directly throws the exc.? 612 // It won't be much cheaper than bailing to the interp., since we'll 613 // have to pass up all the debug-info, and the runtime will have to 614 // create the stack trace. 615 616 // Usual case: Bail to interpreter. 617 // Reserve the right to recompile if we haven't seen anything yet. 618 619 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL; 620 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile; 621 if (treat_throw_as_hot 622 && (method()->method_data()->trap_recompiled_at(bci(), m) 623 || C->too_many_traps(reason))) { 624 // We cannot afford to take more traps here. Suffer in the interpreter. 625 if (C->log() != NULL) 626 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'", 627 Deoptimization::trap_reason_name(reason), 628 C->trap_count(reason)); 629 action = Deoptimization::Action_none; 630 } 631 632 // "must_throw" prunes the JVM state to include only the stack, if there 633 // are no local exception handlers. This should cut down on register 634 // allocation time and code size, by drastically reducing the number 635 // of in-edges on the call to the uncommon trap. 636 637 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw); 638 } 639 640 641 //----------------------------PreserveJVMState--------------------------------- 642 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) { 643 debug_only(kit->verify_map()); 644 _kit = kit; 645 _map = kit->map(); // preserve the map 646 _sp = kit->sp(); 647 kit->set_map(clone_map ? kit->clone_map() : NULL); 648 #ifdef ASSERT 649 _bci = kit->bci(); 650 Parse* parser = kit->is_Parse(); 651 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 652 _block = block; 653 #endif 654 } 655 PreserveJVMState::~PreserveJVMState() { 656 GraphKit* kit = _kit; 657 #ifdef ASSERT 658 assert(kit->bci() == _bci, "bci must not shift"); 659 Parse* parser = kit->is_Parse(); 660 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo(); 661 assert(block == _block, "block must not shift"); 662 #endif 663 kit->set_map(_map); 664 kit->set_sp(_sp); 665 } 666 667 668 //-----------------------------BuildCutout------------------------------------- 669 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt) 670 : PreserveJVMState(kit) 671 { 672 assert(p->is_Con() || p->is_Bool(), "test must be a bool"); 673 SafePointNode* outer_map = _map; // preserved map is caller's 674 SafePointNode* inner_map = kit->map(); 675 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt); 676 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) )); 677 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) )); 678 } 679 BuildCutout::~BuildCutout() { 680 GraphKit* kit = _kit; 681 assert(kit->stopped(), "cutout code must stop, throw, return, etc."); 682 } 683 684 //---------------------------PreserveReexecuteState---------------------------- 685 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) { 686 assert(!kit->stopped(), "must call stopped() before"); 687 _kit = kit; 688 _sp = kit->sp(); 689 _reexecute = kit->jvms()->_reexecute; 690 } 691 PreserveReexecuteState::~PreserveReexecuteState() { 692 if (_kit->stopped()) return; 693 _kit->jvms()->_reexecute = _reexecute; 694 _kit->set_sp(_sp); 695 } 696 697 //------------------------------clone_map-------------------------------------- 698 // Implementation of PreserveJVMState 699 // 700 // Only clone_map(...) here. If this function is only used in the 701 // PreserveJVMState class we may want to get rid of this extra 702 // function eventually and do it all there. 703 704 SafePointNode* GraphKit::clone_map() { 705 if (map() == NULL) return NULL; 706 707 // Clone the memory edge first 708 Node* mem = MergeMemNode::make(map()->memory()); 709 gvn().set_type_bottom(mem); 710 711 SafePointNode *clonemap = (SafePointNode*)map()->clone(); 712 JVMState* jvms = this->jvms(); 713 JVMState* clonejvms = jvms->clone_shallow(C); 714 clonemap->set_memory(mem); 715 clonemap->set_jvms(clonejvms); 716 clonejvms->set_map(clonemap); 717 record_for_igvn(clonemap); 718 gvn().set_type_bottom(clonemap); 719 return clonemap; 720 } 721 722 723 //-----------------------------set_map_clone----------------------------------- 724 void GraphKit::set_map_clone(SafePointNode* m) { 725 _map = m; 726 _map = clone_map(); 727 _map->set_next_exception(NULL); 728 debug_only(verify_map()); 729 } 730 731 732 //----------------------------kill_dead_locals--------------------------------- 733 // Detect any locals which are known to be dead, and force them to top. 734 void GraphKit::kill_dead_locals() { 735 // Consult the liveness information for the locals. If any 736 // of them are unused, then they can be replaced by top(). This 737 // should help register allocation time and cut down on the size 738 // of the deoptimization information. 739 740 // This call is made from many of the bytecode handling 741 // subroutines called from the Big Switch in do_one_bytecode. 742 // Every bytecode which might include a slow path is responsible 743 // for killing its dead locals. The more consistent we 744 // are about killing deads, the fewer useless phis will be 745 // constructed for them at various merge points. 746 747 // bci can be -1 (InvocationEntryBci). We return the entry 748 // liveness for the method. 749 750 if (method() == NULL || method()->code_size() == 0) { 751 // We are building a graph for a call to a native method. 752 // All locals are live. 753 return; 754 } 755 756 ResourceMark rm; 757 758 // Consult the liveness information for the locals. If any 759 // of them are unused, then they can be replaced by top(). This 760 // should help register allocation time and cut down on the size 761 // of the deoptimization information. 762 MethodLivenessResult live_locals = method()->liveness_at_bci(bci()); 763 764 int len = (int)live_locals.size(); 765 assert(len <= jvms()->loc_size(), "too many live locals"); 766 for (int local = 0; local < len; local++) { 767 if (!live_locals.at(local)) { 768 set_local(local, top()); 769 } 770 } 771 } 772 773 #ifdef ASSERT 774 //-------------------------dead_locals_are_killed------------------------------ 775 // Return true if all dead locals are set to top in the map. 776 // Used to assert "clean" debug info at various points. 777 bool GraphKit::dead_locals_are_killed() { 778 if (method() == NULL || method()->code_size() == 0) { 779 // No locals need to be dead, so all is as it should be. 780 return true; 781 } 782 783 // Make sure somebody called kill_dead_locals upstream. 784 ResourceMark rm; 785 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 786 if (jvms->loc_size() == 0) continue; // no locals to consult 787 SafePointNode* map = jvms->map(); 788 ciMethod* method = jvms->method(); 789 int bci = jvms->bci(); 790 if (jvms == this->jvms()) { 791 bci = this->bci(); // it might not yet be synched 792 } 793 MethodLivenessResult live_locals = method->liveness_at_bci(bci); 794 int len = (int)live_locals.size(); 795 if (!live_locals.is_valid() || len == 0) 796 // This method is trivial, or is poisoned by a breakpoint. 797 return true; 798 assert(len == jvms->loc_size(), "live map consistent with locals map"); 799 for (int local = 0; local < len; local++) { 800 if (!live_locals.at(local) && map->local(jvms, local) != top()) { 801 if (PrintMiscellaneous && (Verbose || WizardMode)) { 802 tty->print_cr("Zombie local %d: ", local); 803 jvms->dump(); 804 } 805 return false; 806 } 807 } 808 } 809 return true; 810 } 811 812 #endif //ASSERT 813 814 // Helper function for enforcing certain bytecodes to reexecute if 815 // deoptimization happens 816 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 817 ciMethod* cur_method = jvms->method(); 818 int cur_bci = jvms->bci(); 819 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 820 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 821 return Interpreter::bytecode_should_reexecute(code) || 822 is_anewarray && code == Bytecodes::_multianewarray; 823 // Reexecute _multianewarray bytecode which was replaced with 824 // sequence of [a]newarray. See Parse::do_multianewarray(). 825 // 826 // Note: interpreter should not have it set since this optimization 827 // is limited by dimensions and guarded by flag so in some cases 828 // multianewarray() runtime calls will be generated and 829 // the bytecode should not be reexecutes (stack will not be reset). 830 } else 831 return false; 832 } 833 834 // Helper function for adding JVMState and debug information to node 835 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) { 836 // Add the safepoint edges to the call (or other safepoint). 837 838 // Make sure dead locals are set to top. This 839 // should help register allocation time and cut down on the size 840 // of the deoptimization information. 841 assert(dead_locals_are_killed(), "garbage in debug info before safepoint"); 842 843 // Walk the inline list to fill in the correct set of JVMState's 844 // Also fill in the associated edges for each JVMState. 845 846 // If the bytecode needs to be reexecuted we need to put 847 // the arguments back on the stack. 848 const bool should_reexecute = jvms()->should_reexecute(); 849 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms(); 850 851 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to 852 // undefined if the bci is different. This is normal for Parse but it 853 // should not happen for LibraryCallKit because only one bci is processed. 854 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute), 855 "in LibraryCallKit the reexecute bit should not change"); 856 857 // If we are guaranteed to throw, we can prune everything but the 858 // input to the current bytecode. 859 bool can_prune_locals = false; 860 uint stack_slots_not_pruned = 0; 861 int inputs = 0, depth = 0; 862 if (must_throw) { 863 assert(method() == youngest_jvms->method(), "sanity"); 864 if (compute_stack_effects(inputs, depth)) { 865 can_prune_locals = true; 866 stack_slots_not_pruned = inputs; 867 } 868 } 869 870 if (env()->should_retain_local_variables()) { 871 // At any safepoint, this method can get breakpointed, which would 872 // then require an immediate deoptimization. 873 can_prune_locals = false; // do not prune locals 874 stack_slots_not_pruned = 0; 875 } 876 877 // do not scribble on the input jvms 878 JVMState* out_jvms = youngest_jvms->clone_deep(C); 879 call->set_jvms(out_jvms); // Start jvms list for call node 880 881 // For a known set of bytecodes, the interpreter should reexecute them if 882 // deoptimization happens. We set the reexecute state for them here 883 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 884 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 885 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 886 } 887 888 // Presize the call: 889 DEBUG_ONLY(uint non_debug_edges = call->req()); 890 call->add_req_batch(top(), youngest_jvms->debug_depth()); 891 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 892 893 // Set up edges so that the call looks like this: 894 // Call [state:] ctl io mem fptr retadr 895 // [parms:] parm0 ... parmN 896 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 897 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 898 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 899 // Note that caller debug info precedes callee debug info. 900 901 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 902 uint debug_ptr = call->req(); 903 904 // Loop over the map input edges associated with jvms, add them 905 // to the call node, & reset all offsets to match call node array. 906 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 907 uint debug_end = debug_ptr; 908 uint debug_start = debug_ptr - in_jvms->debug_size(); 909 debug_ptr = debug_start; // back up the ptr 910 911 uint p = debug_start; // walks forward in [debug_start, debug_end) 912 uint j, k, l; 913 SafePointNode* in_map = in_jvms->map(); 914 out_jvms->set_map(call); 915 916 if (can_prune_locals) { 917 assert(in_jvms->method() == out_jvms->method(), "sanity"); 918 // If the current throw can reach an exception handler in this JVMS, 919 // then we must keep everything live that can reach that handler. 920 // As a quick and dirty approximation, we look for any handlers at all. 921 if (in_jvms->method()->has_exception_handlers()) { 922 can_prune_locals = false; 923 } 924 } 925 926 // Add the Locals 927 k = in_jvms->locoff(); 928 l = in_jvms->loc_size(); 929 out_jvms->set_locoff(p); 930 if (!can_prune_locals) { 931 for (j = 0; j < l; j++) 932 call->set_req(p++, in_map->in(k+j)); 933 } else { 934 p += l; // already set to top above by add_req_batch 935 } 936 937 // Add the Expression Stack 938 k = in_jvms->stkoff(); 939 l = in_jvms->sp(); 940 out_jvms->set_stkoff(p); 941 if (!can_prune_locals) { 942 for (j = 0; j < l; j++) 943 call->set_req(p++, in_map->in(k+j)); 944 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 945 // Divide stack into {S0,...,S1}, where S0 is set to top. 946 uint s1 = stack_slots_not_pruned; 947 stack_slots_not_pruned = 0; // for next iteration 948 if (s1 > l) s1 = l; 949 uint s0 = l - s1; 950 p += s0; // skip the tops preinstalled by add_req_batch 951 for (j = s0; j < l; j++) 952 call->set_req(p++, in_map->in(k+j)); 953 } else { 954 p += l; // already set to top above by add_req_batch 955 } 956 957 // Add the Monitors 958 k = in_jvms->monoff(); 959 l = in_jvms->mon_size(); 960 out_jvms->set_monoff(p); 961 for (j = 0; j < l; j++) 962 call->set_req(p++, in_map->in(k+j)); 963 964 // Copy any scalar object fields. 965 k = in_jvms->scloff(); 966 l = in_jvms->scl_size(); 967 out_jvms->set_scloff(p); 968 for (j = 0; j < l; j++) 969 call->set_req(p++, in_map->in(k+j)); 970 971 // Finish the new jvms. 972 out_jvms->set_endoff(p); 973 974 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 975 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 976 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 977 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 978 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 979 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 980 981 // Update the two tail pointers in parallel. 982 out_jvms = out_jvms->caller(); 983 in_jvms = in_jvms->caller(); 984 } 985 986 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 987 988 // Test the correctness of JVMState::debug_xxx accessors: 989 assert(call->jvms()->debug_start() == non_debug_edges, ""); 990 assert(call->jvms()->debug_end() == call->req(), ""); 991 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 992 } 993 994 bool GraphKit::compute_stack_effects(int& inputs, int& depth) { 995 Bytecodes::Code code = java_bc(); 996 if (code == Bytecodes::_wide) { 997 code = method()->java_code_at_bci(bci() + 1); 998 } 999 1000 BasicType rtype = T_ILLEGAL; 1001 int rsize = 0; 1002 1003 if (code != Bytecodes::_illegal) { 1004 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1 1005 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V 1006 if (rtype < T_CONFLICT) 1007 rsize = type2size[rtype]; 1008 } 1009 1010 switch (code) { 1011 case Bytecodes::_illegal: 1012 return false; 1013 1014 case Bytecodes::_ldc: 1015 case Bytecodes::_ldc_w: 1016 case Bytecodes::_ldc2_w: 1017 inputs = 0; 1018 break; 1019 1020 case Bytecodes::_dup: inputs = 1; break; 1021 case Bytecodes::_dup_x1: inputs = 2; break; 1022 case Bytecodes::_dup_x2: inputs = 3; break; 1023 case Bytecodes::_dup2: inputs = 2; break; 1024 case Bytecodes::_dup2_x1: inputs = 3; break; 1025 case Bytecodes::_dup2_x2: inputs = 4; break; 1026 case Bytecodes::_swap: inputs = 2; break; 1027 case Bytecodes::_arraylength: inputs = 1; break; 1028 1029 case Bytecodes::_getstatic: 1030 case Bytecodes::_putstatic: 1031 case Bytecodes::_getfield: 1032 case Bytecodes::_putfield: 1033 { 1034 bool ignored_will_link; 1035 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link); 1036 int size = field->type()->size(); 1037 bool is_get = (depth >= 0), is_static = (depth & 1); 1038 inputs = (is_static ? 0 : 1); 1039 if (is_get) { 1040 depth = size - inputs; 1041 } else { 1042 inputs += size; // putxxx pops the value from the stack 1043 depth = - inputs; 1044 } 1045 } 1046 break; 1047 1048 case Bytecodes::_invokevirtual: 1049 case Bytecodes::_invokespecial: 1050 case Bytecodes::_invokestatic: 1051 case Bytecodes::_invokedynamic: 1052 case Bytecodes::_invokeinterface: 1053 { 1054 bool ignored_will_link; 1055 ciSignature* declared_signature = NULL; 1056 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 1057 assert(declared_signature != NULL, "cannot be null"); 1058 inputs = declared_signature->arg_size_for_bc(code); 1059 int size = declared_signature->return_type()->size(); 1060 depth = size - inputs; 1061 } 1062 break; 1063 1064 case Bytecodes::_multianewarray: 1065 { 1066 ciBytecodeStream iter(method()); 1067 iter.reset_to_bci(bci()); 1068 iter.next(); 1069 inputs = iter.get_dimensions(); 1070 assert(rsize == 1, ""); 1071 depth = rsize - inputs; 1072 } 1073 break; 1074 1075 case Bytecodes::_ireturn: 1076 case Bytecodes::_lreturn: 1077 case Bytecodes::_freturn: 1078 case Bytecodes::_dreturn: 1079 case Bytecodes::_areturn: 1080 assert(rsize == -depth, ""); 1081 inputs = rsize; 1082 break; 1083 1084 case Bytecodes::_jsr: 1085 case Bytecodes::_jsr_w: 1086 inputs = 0; 1087 depth = 1; // S.B. depth=1, not zero 1088 break; 1089 1090 default: 1091 // bytecode produces a typed result 1092 inputs = rsize - depth; 1093 assert(inputs >= 0, ""); 1094 break; 1095 } 1096 1097 #ifdef ASSERT 1098 // spot check 1099 int outputs = depth + inputs; 1100 assert(outputs >= 0, "sanity"); 1101 switch (code) { 1102 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break; 1103 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break; 1104 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break; 1105 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break; 1106 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break; 1107 } 1108 #endif //ASSERT 1109 1110 return true; 1111 } 1112 1113 1114 1115 //------------------------------basic_plus_adr--------------------------------- 1116 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) { 1117 // short-circuit a common case 1118 if (offset == intcon(0)) return ptr; 1119 return _gvn.transform( new AddPNode(base, ptr, offset) ); 1120 } 1121 1122 Node* GraphKit::ConvI2L(Node* offset) { 1123 // short-circuit a common case 1124 jint offset_con = find_int_con(offset, Type::OffsetBot); 1125 if (offset_con != Type::OffsetBot) { 1126 return longcon((jlong) offset_con); 1127 } 1128 return _gvn.transform( new ConvI2LNode(offset)); 1129 } 1130 1131 Node* GraphKit::ConvI2UL(Node* offset) { 1132 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot); 1133 if (offset_con != (juint) Type::OffsetBot) { 1134 return longcon((julong) offset_con); 1135 } 1136 Node* conv = _gvn.transform( new ConvI2LNode(offset)); 1137 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint)); 1138 return _gvn.transform( new AndLNode(conv, mask) ); 1139 } 1140 1141 Node* GraphKit::ConvL2I(Node* offset) { 1142 // short-circuit a common case 1143 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot); 1144 if (offset_con != (jlong)Type::OffsetBot) { 1145 return intcon((int) offset_con); 1146 } 1147 return _gvn.transform( new ConvL2INode(offset)); 1148 } 1149 1150 //-------------------------load_object_klass----------------------------------- 1151 Node* GraphKit::load_object_klass(Node* obj) { 1152 // Special-case a fresh allocation to avoid building nodes: 1153 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn); 1154 if (akls != NULL) return akls; 1155 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); 1156 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 1157 } 1158 1159 //-------------------------load_array_length----------------------------------- 1160 Node* GraphKit::load_array_length(Node* array) { 1161 // Special-case a fresh allocation to avoid building nodes: 1162 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn); 1163 Node *alen; 1164 if (alloc == NULL) { 1165 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes()); 1166 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS)); 1167 } else { 1168 alen = alloc->Ideal_length(); 1169 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn); 1170 if (ccast != alen) { 1171 alen = _gvn.transform(ccast); 1172 } 1173 } 1174 return alen; 1175 } 1176 1177 //------------------------------do_null_check---------------------------------- 1178 // Helper function to do a NULL pointer check. Returned value is 1179 // the incoming address with NULL casted away. You are allowed to use the 1180 // not-null value only if you are control dependent on the test. 1181 #ifndef PRODUCT 1182 extern int explicit_null_checks_inserted, 1183 explicit_null_checks_elided; 1184 #endif 1185 Node* GraphKit::null_check_common(Node* value, BasicType type, 1186 // optional arguments for variations: 1187 bool assert_null, 1188 Node* *null_control, 1189 bool speculative) { 1190 assert(!assert_null || null_control == NULL, "not both at once"); 1191 if (stopped()) return top(); 1192 NOT_PRODUCT(explicit_null_checks_inserted++); 1193 1194 // Construct NULL check 1195 Node *chk = NULL; 1196 switch(type) { 1197 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break; 1198 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break; 1199 case T_ARRAY : // fall through 1200 type = T_OBJECT; // simplify further tests 1201 case T_OBJECT : { 1202 const Type *t = _gvn.type( value ); 1203 1204 const TypeOopPtr* tp = t->isa_oopptr(); 1205 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded() 1206 // Only for do_null_check, not any of its siblings: 1207 && !assert_null && null_control == NULL) { 1208 // Usually, any field access or invocation on an unloaded oop type 1209 // will simply fail to link, since the statically linked class is 1210 // likely also to be unloaded. However, in -Xcomp mode, sometimes 1211 // the static class is loaded but the sharper oop type is not. 1212 // Rather than checking for this obscure case in lots of places, 1213 // we simply observe that a null check on an unloaded class 1214 // will always be followed by a nonsense operation, so we 1215 // can just issue the uncommon trap here. 1216 // Our access to the unloaded class will only be correct 1217 // after it has been loaded and initialized, which requires 1218 // a trip through the interpreter. 1219 #ifndef PRODUCT 1220 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); } 1221 #endif 1222 uncommon_trap(Deoptimization::Reason_unloaded, 1223 Deoptimization::Action_reinterpret, 1224 tp->klass(), "!loaded"); 1225 return top(); 1226 } 1227 1228 if (assert_null) { 1229 // See if the type is contained in NULL_PTR. 1230 // If so, then the value is already null. 1231 if (t->higher_equal(TypePtr::NULL_PTR)) { 1232 NOT_PRODUCT(explicit_null_checks_elided++); 1233 return value; // Elided null assert quickly! 1234 } 1235 } else { 1236 // See if mixing in the NULL pointer changes type. 1237 // If so, then the NULL pointer was not allowed in the original 1238 // type. In other words, "value" was not-null. 1239 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) { 1240 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ... 1241 NOT_PRODUCT(explicit_null_checks_elided++); 1242 return value; // Elided null check quickly! 1243 } 1244 } 1245 chk = new CmpPNode( value, null() ); 1246 break; 1247 } 1248 1249 default: 1250 fatal("unexpected type: %s", type2name(type)); 1251 } 1252 assert(chk != NULL, "sanity check"); 1253 chk = _gvn.transform(chk); 1254 1255 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne; 1256 BoolNode *btst = new BoolNode( chk, btest); 1257 Node *tst = _gvn.transform( btst ); 1258 1259 //----------- 1260 // if peephole optimizations occurred, a prior test existed. 1261 // If a prior test existed, maybe it dominates as we can avoid this test. 1262 if (tst != btst && type == T_OBJECT) { 1263 // At this point we want to scan up the CFG to see if we can 1264 // find an identical test (and so avoid this test altogether). 1265 Node *cfg = control(); 1266 int depth = 0; 1267 while( depth < 16 ) { // Limit search depth for speed 1268 if( cfg->Opcode() == Op_IfTrue && 1269 cfg->in(0)->in(1) == tst ) { 1270 // Found prior test. Use "cast_not_null" to construct an identical 1271 // CastPP (and hence hash to) as already exists for the prior test. 1272 // Return that casted value. 1273 if (assert_null) { 1274 replace_in_map(value, null()); 1275 return null(); // do not issue the redundant test 1276 } 1277 Node *oldcontrol = control(); 1278 set_control(cfg); 1279 Node *res = cast_not_null(value); 1280 set_control(oldcontrol); 1281 NOT_PRODUCT(explicit_null_checks_elided++); 1282 return res; 1283 } 1284 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true); 1285 if (cfg == NULL) break; // Quit at region nodes 1286 depth++; 1287 } 1288 } 1289 1290 //----------- 1291 // Branch to failure if null 1292 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen 1293 Deoptimization::DeoptReason reason; 1294 if (assert_null) { 1295 reason = Deoptimization::Reason_null_assert; 1296 } else if (type == T_OBJECT) { 1297 reason = Deoptimization::reason_null_check(speculative); 1298 } else { 1299 reason = Deoptimization::Reason_div0_check; 1300 } 1301 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis, 1302 // ciMethodData::has_trap_at will return a conservative -1 if any 1303 // must-be-null assertion has failed. This could cause performance 1304 // problems for a method after its first do_null_assert failure. 1305 // Consider using 'Reason_class_check' instead? 1306 1307 // To cause an implicit null check, we set the not-null probability 1308 // to the maximum (PROB_MAX). For an explicit check the probability 1309 // is set to a smaller value. 1310 if (null_control != NULL || too_many_traps(reason)) { 1311 // probability is less likely 1312 ok_prob = PROB_LIKELY_MAG(3); 1313 } else if (!assert_null && 1314 (ImplicitNullCheckThreshold > 0) && 1315 method() != NULL && 1316 (method()->method_data()->trap_count(reason) 1317 >= (uint)ImplicitNullCheckThreshold)) { 1318 ok_prob = PROB_LIKELY_MAG(3); 1319 } 1320 1321 if (null_control != NULL) { 1322 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN); 1323 Node* null_true = _gvn.transform( new IfFalseNode(iff)); 1324 set_control( _gvn.transform( new IfTrueNode(iff))); 1325 #ifndef PRODUCT 1326 if (null_true == top()) { 1327 explicit_null_checks_elided++; 1328 } 1329 #endif 1330 (*null_control) = null_true; 1331 } else { 1332 BuildCutout unless(this, tst, ok_prob); 1333 // Check for optimizer eliding test at parse time 1334 if (stopped()) { 1335 // Failure not possible; do not bother making uncommon trap. 1336 NOT_PRODUCT(explicit_null_checks_elided++); 1337 } else if (assert_null) { 1338 uncommon_trap(reason, 1339 Deoptimization::Action_make_not_entrant, 1340 NULL, "assert_null"); 1341 } else { 1342 replace_in_map(value, zerocon(type)); 1343 builtin_throw(reason); 1344 } 1345 } 1346 1347 // Must throw exception, fall-thru not possible? 1348 if (stopped()) { 1349 return top(); // No result 1350 } 1351 1352 if (assert_null) { 1353 // Cast obj to null on this path. 1354 replace_in_map(value, zerocon(type)); 1355 return zerocon(type); 1356 } 1357 1358 // Cast obj to not-null on this path, if there is no null_control. 1359 // (If there is a null_control, a non-null value may come back to haunt us.) 1360 if (type == T_OBJECT) { 1361 Node* cast = cast_not_null(value, false); 1362 if (null_control == NULL || (*null_control) == top()) 1363 replace_in_map(value, cast); 1364 value = cast; 1365 } 1366 1367 return value; 1368 } 1369 1370 1371 //------------------------------cast_not_null---------------------------------- 1372 // Cast obj to not-null on this path 1373 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) { 1374 const Type *t = _gvn.type(obj); 1375 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL); 1376 // Object is already not-null? 1377 if( t == t_not_null ) return obj; 1378 1379 Node *cast = new CastPPNode(obj,t_not_null); 1380 cast->init_req(0, control()); 1381 cast = _gvn.transform( cast ); 1382 1383 // Scan for instances of 'obj' in the current JVM mapping. 1384 // These instances are known to be not-null after the test. 1385 if (do_replace_in_map) 1386 replace_in_map(obj, cast); 1387 1388 return cast; // Return casted value 1389 } 1390 1391 1392 //--------------------------replace_in_map------------------------------------- 1393 void GraphKit::replace_in_map(Node* old, Node* neww) { 1394 if (old == neww) { 1395 return; 1396 } 1397 1398 map()->replace_edge(old, neww); 1399 1400 // Note: This operation potentially replaces any edge 1401 // on the map. This includes locals, stack, and monitors 1402 // of the current (innermost) JVM state. 1403 1404 // don't let inconsistent types from profiling escape this 1405 // method 1406 1407 const Type* told = _gvn.type(old); 1408 const Type* tnew = _gvn.type(neww); 1409 1410 if (!tnew->higher_equal(told)) { 1411 return; 1412 } 1413 1414 map()->record_replaced_node(old, neww); 1415 } 1416 1417 1418 //============================================================================= 1419 //--------------------------------memory--------------------------------------- 1420 Node* GraphKit::memory(uint alias_idx) { 1421 MergeMemNode* mem = merged_memory(); 1422 Node* p = mem->memory_at(alias_idx); 1423 _gvn.set_type(p, Type::MEMORY); // must be mapped 1424 return p; 1425 } 1426 1427 //-----------------------------reset_memory------------------------------------ 1428 Node* GraphKit::reset_memory() { 1429 Node* mem = map()->memory(); 1430 // do not use this node for any more parsing! 1431 debug_only( map()->set_memory((Node*)NULL) ); 1432 return _gvn.transform( mem ); 1433 } 1434 1435 //------------------------------set_all_memory--------------------------------- 1436 void GraphKit::set_all_memory(Node* newmem) { 1437 Node* mergemem = MergeMemNode::make(newmem); 1438 gvn().set_type_bottom(mergemem); 1439 map()->set_memory(mergemem); 1440 } 1441 1442 //------------------------------set_all_memory_call---------------------------- 1443 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) { 1444 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) ); 1445 set_all_memory(newmem); 1446 } 1447 1448 //============================================================================= 1449 // 1450 // parser factory methods for MemNodes 1451 // 1452 // These are layered on top of the factory methods in LoadNode and StoreNode, 1453 // and integrate with the parser's memory state and _gvn engine. 1454 // 1455 1456 // factory methods in "int adr_idx" 1457 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt, 1458 int adr_idx, 1459 MemNode::MemOrd mo, 1460 LoadNode::ControlDependency control_dependency, 1461 bool require_atomic_access, 1462 bool unaligned, 1463 bool mismatched) { 1464 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" ); 1465 const TypePtr* adr_type = NULL; // debug-mode-only argument 1466 debug_only(adr_type = C->get_adr_type(adr_idx)); 1467 Node* mem = memory(adr_idx); 1468 Node* ld; 1469 if (require_atomic_access && bt == T_LONG) { 1470 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1471 } else if (require_atomic_access && bt == T_DOUBLE) { 1472 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched); 1473 } else { 1474 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched); 1475 } 1476 ld = _gvn.transform(ld); 1477 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) { 1478 // Improve graph before escape analysis and boxing elimination. 1479 record_for_igvn(ld); 1480 } 1481 return ld; 1482 } 1483 1484 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt, 1485 int adr_idx, 1486 MemNode::MemOrd mo, 1487 bool require_atomic_access, 1488 bool unaligned, 1489 bool mismatched) { 1490 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" ); 1491 const TypePtr* adr_type = NULL; 1492 debug_only(adr_type = C->get_adr_type(adr_idx)); 1493 Node *mem = memory(adr_idx); 1494 Node* st; 1495 if (require_atomic_access && bt == T_LONG) { 1496 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1497 } else if (require_atomic_access && bt == T_DOUBLE) { 1498 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo); 1499 } else { 1500 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo); 1501 } 1502 if (unaligned) { 1503 st->as_Store()->set_unaligned_access(); 1504 } 1505 if (mismatched) { 1506 st->as_Store()->set_mismatched_access(); 1507 } 1508 st = _gvn.transform(st); 1509 set_memory(st, adr_idx); 1510 // Back-to-back stores can only remove intermediate store with DU info 1511 // so push on worklist for optimizer. 1512 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address)) 1513 record_for_igvn(st); 1514 1515 return st; 1516 } 1517 1518 Node* GraphKit::access_store_at(Node* ctl, 1519 Node* obj, 1520 Node* adr, 1521 const TypePtr* adr_type, 1522 Node* val, 1523 const Type* val_type, 1524 BasicType bt, 1525 C2DecoratorSet decorators) { 1526 // Transformation of a value which could be NULL pointer (CastPP #NULL) 1527 // could be delayed during Parse (for example, in adjust_map_after_if()). 1528 // Execute transformation here to avoid barrier generation in such case. 1529 if (_gvn.type(val) == TypePtr::NULL_PTR) { 1530 val = _gvn.makecon(TypePtr::NULL_PTR); 1531 } 1532 1533 set_control(ctl); 1534 if (stopped()) { 1535 return top(); // Dead path ? 1536 } 1537 1538 assert(val != NULL, "not dead path"); 1539 1540 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1541 return code_gen->store_at(this, obj, adr, adr_type, val, val_type, bt, decorators); 1542 } 1543 1544 Node* GraphKit::access_load_at(Node* obj, // containing obj 1545 Node* adr, // actual adress to store val at 1546 const TypePtr* adr_type, 1547 const Type* val_type, 1548 BasicType bt, 1549 C2DecoratorSet decorators) { 1550 if (stopped()) { 1551 return top(); // Dead path ? 1552 } 1553 1554 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1555 return code_gen->load_at(this, obj, adr, adr_type, val_type, bt, decorators); 1556 } 1557 1558 Node* GraphKit::access_cas_val_at(Node* ctl, 1559 Node* obj, 1560 Node* adr, 1561 const TypePtr* adr_type, 1562 int alias_idx, 1563 Node* expected_val, 1564 Node* new_val, 1565 const Type* value_type, 1566 BasicType bt, 1567 C2DecoratorSet decorators) { 1568 set_control(ctl); 1569 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1570 return code_gen->cas_val_at(this, obj, adr, adr_type, alias_idx, expected_val, new_val, value_type, bt, decorators); 1571 } 1572 1573 Node* GraphKit::access_cas_bool_at(Node* ctl, 1574 Node* obj, 1575 Node* adr, 1576 const TypePtr* adr_type, 1577 int alias_idx, 1578 Node* expected_val, 1579 Node* new_val, 1580 const Type* value_type, 1581 BasicType bt, 1582 C2DecoratorSet decorators) { 1583 set_control(ctl); 1584 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1585 return code_gen->cas_bool_at(this, obj, adr, adr_type, alias_idx, expected_val, new_val, value_type, bt, decorators); 1586 } 1587 1588 Node* GraphKit::access_swap_at(Node* ctl, 1589 Node* obj, 1590 Node* adr, 1591 const TypePtr* adr_type, 1592 int alias_idx, 1593 Node* new_val, 1594 const Type* value_type, 1595 BasicType bt, 1596 C2DecoratorSet decorators) { 1597 set_control(ctl); 1598 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1599 return code_gen->swap_at(this, obj, adr, adr_type, alias_idx, new_val, value_type, bt, decorators); 1600 } 1601 1602 Node* GraphKit::access_fetch_and_add_at(Node* ctl, 1603 Node* obj, 1604 Node* adr, 1605 const TypePtr* adr_type, 1606 int alias_idx, 1607 Node* new_val, 1608 const Type* value_type, 1609 BasicType bt, 1610 C2DecoratorSet decorators) { 1611 set_control(ctl); 1612 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1613 return code_gen->fetch_and_add_at(this, obj, adr, adr_type, alias_idx, new_val, value_type, bt, decorators); 1614 } 1615 1616 void GraphKit::access_clone(Node* ctl, Node* src, Node* dst, Node* size, bool is_array) { 1617 set_control(ctl); 1618 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen(); 1619 return code_gen->clone(this, src, dst, size, is_array); 1620 } 1621 1622 //-------------------------array_element_address------------------------- 1623 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt, 1624 const TypeInt* sizetype, Node* ctrl) { 1625 uint shift = exact_log2(type2aelembytes(elembt)); 1626 uint header = arrayOopDesc::base_offset_in_bytes(elembt); 1627 1628 // short-circuit a common case (saves lots of confusing waste motion) 1629 jint idx_con = find_int_con(idx, -1); 1630 if (idx_con >= 0) { 1631 intptr_t offset = header + ((intptr_t)idx_con << shift); 1632 return basic_plus_adr(ary, offset); 1633 } 1634 1635 // must be correct type for alignment purposes 1636 Node* base = basic_plus_adr(ary, header); 1637 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl); 1638 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) ); 1639 return basic_plus_adr(ary, base, scale); 1640 } 1641 1642 //-------------------------load_array_element------------------------- 1643 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) { 1644 const Type* elemtype = arytype->elem(); 1645 BasicType elembt = elemtype->array_element_basic_type(); 1646 Node* adr = array_element_address(ary, idx, elembt, arytype->size()); 1647 if (elembt == T_NARROWOOP) { 1648 elembt = T_OBJECT; // To satisfy switch in LoadNode::make() 1649 } 1650 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered); 1651 return ld; 1652 } 1653 1654 //-------------------------set_arguments_for_java_call------------------------- 1655 // Arguments (pre-popped from the stack) are taken from the JVMS. 1656 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) { 1657 // Add the call arguments: 1658 uint nargs = call->method()->arg_size(); 1659 for (uint i = 0; i < nargs; i++) { 1660 Node* arg = argument(i); 1661 call->init_req(i + TypeFunc::Parms, arg); 1662 } 1663 } 1664 1665 //---------------------------set_edges_for_java_call--------------------------- 1666 // Connect a newly created call into the current JVMS. 1667 // A return value node (if any) is returned from set_edges_for_java_call. 1668 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) { 1669 1670 // Add the predefined inputs: 1671 call->init_req( TypeFunc::Control, control() ); 1672 call->init_req( TypeFunc::I_O , i_o() ); 1673 call->init_req( TypeFunc::Memory , reset_memory() ); 1674 call->init_req( TypeFunc::FramePtr, frameptr() ); 1675 call->init_req( TypeFunc::ReturnAdr, top() ); 1676 1677 add_safepoint_edges(call, must_throw); 1678 1679 Node* xcall = _gvn.transform(call); 1680 1681 if (xcall == top()) { 1682 set_control(top()); 1683 return; 1684 } 1685 assert(xcall == call, "call identity is stable"); 1686 1687 // Re-use the current map to produce the result. 1688 1689 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control))); 1690 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj))); 1691 set_all_memory_call(xcall, separate_io_proj); 1692 1693 //return xcall; // no need, caller already has it 1694 } 1695 1696 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) { 1697 if (stopped()) return top(); // maybe the call folded up? 1698 1699 // Capture the return value, if any. 1700 Node* ret; 1701 if (call->method() == NULL || 1702 call->method()->return_type()->basic_type() == T_VOID) 1703 ret = top(); 1704 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1705 1706 // Note: Since any out-of-line call can produce an exception, 1707 // we always insert an I_O projection from the call into the result. 1708 1709 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj); 1710 1711 if (separate_io_proj) { 1712 // The caller requested separate projections be used by the fall 1713 // through and exceptional paths, so replace the projections for 1714 // the fall through path. 1715 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) )); 1716 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) )); 1717 } 1718 return ret; 1719 } 1720 1721 //--------------------set_predefined_input_for_runtime_call-------------------- 1722 // Reading and setting the memory state is way conservative here. 1723 // The real problem is that I am not doing real Type analysis on memory, 1724 // so I cannot distinguish card mark stores from other stores. Across a GC 1725 // point the Store Barrier and the card mark memory has to agree. I cannot 1726 // have a card mark store and its barrier split across the GC point from 1727 // either above or below. Here I get that to happen by reading ALL of memory. 1728 // A better answer would be to separate out card marks from other memory. 1729 // For now, return the input memory state, so that it can be reused 1730 // after the call, if this call has restricted memory effects. 1731 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) { 1732 // Set fixed predefined input arguments 1733 Node* memory = reset_memory(); 1734 call->init_req( TypeFunc::Control, control() ); 1735 call->init_req( TypeFunc::I_O, top() ); // does no i/o 1736 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs 1737 call->init_req( TypeFunc::FramePtr, frameptr() ); 1738 call->init_req( TypeFunc::ReturnAdr, top() ); 1739 return memory; 1740 } 1741 1742 //-------------------set_predefined_output_for_runtime_call-------------------- 1743 // Set control and memory (not i_o) from the call. 1744 // If keep_mem is not NULL, use it for the output state, 1745 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM. 1746 // If hook_mem is NULL, this call produces no memory effects at all. 1747 // If hook_mem is a Java-visible memory slice (such as arraycopy operands), 1748 // then only that memory slice is taken from the call. 1749 // In the last case, we must put an appropriate memory barrier before 1750 // the call, so as to create the correct anti-dependencies on loads 1751 // preceding the call. 1752 void GraphKit::set_predefined_output_for_runtime_call(Node* call, 1753 Node* keep_mem, 1754 const TypePtr* hook_mem) { 1755 // no i/o 1756 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) )); 1757 if (keep_mem) { 1758 // First clone the existing memory state 1759 set_all_memory(keep_mem); 1760 if (hook_mem != NULL) { 1761 // Make memory for the call 1762 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) ); 1763 // Set the RawPtr memory state only. This covers all the heap top/GC stuff 1764 // We also use hook_mem to extract specific effects from arraycopy stubs. 1765 set_memory(mem, hook_mem); 1766 } 1767 // ...else the call has NO memory effects. 1768 1769 // Make sure the call advertises its memory effects precisely. 1770 // This lets us build accurate anti-dependences in gcm.cpp. 1771 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem), 1772 "call node must be constructed correctly"); 1773 } else { 1774 assert(hook_mem == NULL, ""); 1775 // This is not a "slow path" call; all memory comes from the call. 1776 set_all_memory_call(call); 1777 } 1778 } 1779 1780 1781 // Replace the call with the current state of the kit. 1782 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) { 1783 JVMState* ejvms = NULL; 1784 if (has_exceptions()) { 1785 ejvms = transfer_exceptions_into_jvms(); 1786 } 1787 1788 ReplacedNodes replaced_nodes = map()->replaced_nodes(); 1789 ReplacedNodes replaced_nodes_exception; 1790 Node* ex_ctl = top(); 1791 1792 SafePointNode* final_state = stop(); 1793 1794 // Find all the needed outputs of this call 1795 CallProjections callprojs; 1796 call->extract_projections(&callprojs, true); 1797 1798 Node* init_mem = call->in(TypeFunc::Memory); 1799 Node* final_mem = final_state->in(TypeFunc::Memory); 1800 Node* final_ctl = final_state->in(TypeFunc::Control); 1801 Node* final_io = final_state->in(TypeFunc::I_O); 1802 1803 // Replace all the old call edges with the edges from the inlining result 1804 if (callprojs.fallthrough_catchproj != NULL) { 1805 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl); 1806 } 1807 if (callprojs.fallthrough_memproj != NULL) { 1808 if (final_mem->is_MergeMem()) { 1809 // Parser's exits MergeMem was not transformed but may be optimized 1810 final_mem = _gvn.transform(final_mem); 1811 } 1812 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem); 1813 } 1814 if (callprojs.fallthrough_ioproj != NULL) { 1815 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io); 1816 } 1817 1818 // Replace the result with the new result if it exists and is used 1819 if (callprojs.resproj != NULL && result != NULL) { 1820 C->gvn_replace_by(callprojs.resproj, result); 1821 } 1822 1823 if (ejvms == NULL) { 1824 // No exception edges to simply kill off those paths 1825 if (callprojs.catchall_catchproj != NULL) { 1826 C->gvn_replace_by(callprojs.catchall_catchproj, C->top()); 1827 } 1828 if (callprojs.catchall_memproj != NULL) { 1829 C->gvn_replace_by(callprojs.catchall_memproj, C->top()); 1830 } 1831 if (callprojs.catchall_ioproj != NULL) { 1832 C->gvn_replace_by(callprojs.catchall_ioproj, C->top()); 1833 } 1834 // Replace the old exception object with top 1835 if (callprojs.exobj != NULL) { 1836 C->gvn_replace_by(callprojs.exobj, C->top()); 1837 } 1838 } else { 1839 GraphKit ekit(ejvms); 1840 1841 // Load my combined exception state into the kit, with all phis transformed: 1842 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states(); 1843 replaced_nodes_exception = ex_map->replaced_nodes(); 1844 1845 Node* ex_oop = ekit.use_exception_state(ex_map); 1846 1847 if (callprojs.catchall_catchproj != NULL) { 1848 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control()); 1849 ex_ctl = ekit.control(); 1850 } 1851 if (callprojs.catchall_memproj != NULL) { 1852 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory()); 1853 } 1854 if (callprojs.catchall_ioproj != NULL) { 1855 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o()); 1856 } 1857 1858 // Replace the old exception object with the newly created one 1859 if (callprojs.exobj != NULL) { 1860 C->gvn_replace_by(callprojs.exobj, ex_oop); 1861 } 1862 } 1863 1864 // Disconnect the call from the graph 1865 call->disconnect_inputs(NULL, C); 1866 C->gvn_replace_by(call, C->top()); 1867 1868 // Clean up any MergeMems that feed other MergeMems since the 1869 // optimizer doesn't like that. 1870 if (final_mem->is_MergeMem()) { 1871 Node_List wl; 1872 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) { 1873 Node* m = i.get(); 1874 if (m->is_MergeMem() && !wl.contains(m)) { 1875 wl.push(m); 1876 } 1877 } 1878 while (wl.size() > 0) { 1879 _gvn.transform(wl.pop()); 1880 } 1881 } 1882 1883 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) { 1884 replaced_nodes.apply(C, final_ctl); 1885 } 1886 if (!ex_ctl->is_top() && do_replaced_nodes) { 1887 replaced_nodes_exception.apply(C, ex_ctl); 1888 } 1889 } 1890 1891 1892 //------------------------------increment_counter------------------------------ 1893 // for statistics: increment a VM counter by 1 1894 1895 void GraphKit::increment_counter(address counter_addr) { 1896 Node* adr1 = makecon(TypeRawPtr::make(counter_addr)); 1897 increment_counter(adr1); 1898 } 1899 1900 void GraphKit::increment_counter(Node* counter_addr) { 1901 int adr_type = Compile::AliasIdxRaw; 1902 Node* ctrl = control(); 1903 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 1904 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1))); 1905 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered); 1906 } 1907 1908 1909 //------------------------------uncommon_trap---------------------------------- 1910 // Bail out to the interpreter in mid-method. Implemented by calling the 1911 // uncommon_trap blob. This helper function inserts a runtime call with the 1912 // right debug info. 1913 void GraphKit::uncommon_trap(int trap_request, 1914 ciKlass* klass, const char* comment, 1915 bool must_throw, 1916 bool keep_exact_action) { 1917 if (failing()) stop(); 1918 if (stopped()) return; // trap reachable? 1919 1920 // Note: If ProfileTraps is true, and if a deopt. actually 1921 // occurs here, the runtime will make sure an MDO exists. There is 1922 // no need to call method()->ensure_method_data() at this point. 1923 1924 // Set the stack pointer to the right value for reexecution: 1925 set_sp(reexecute_sp()); 1926 1927 #ifdef ASSERT 1928 if (!must_throw) { 1929 // Make sure the stack has at least enough depth to execute 1930 // the current bytecode. 1931 int inputs, ignored_depth; 1932 if (compute_stack_effects(inputs, ignored_depth)) { 1933 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d", 1934 Bytecodes::name(java_bc()), sp(), inputs); 1935 } 1936 } 1937 #endif 1938 1939 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request); 1940 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request); 1941 1942 switch (action) { 1943 case Deoptimization::Action_maybe_recompile: 1944 case Deoptimization::Action_reinterpret: 1945 // Temporary fix for 6529811 to allow virtual calls to be sure they 1946 // get the chance to go from mono->bi->mega 1947 if (!keep_exact_action && 1948 Deoptimization::trap_request_index(trap_request) < 0 && 1949 too_many_recompiles(reason)) { 1950 // This BCI is causing too many recompilations. 1951 if (C->log() != NULL) { 1952 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'", 1953 Deoptimization::trap_reason_name(reason), 1954 Deoptimization::trap_action_name(action)); 1955 } 1956 action = Deoptimization::Action_none; 1957 trap_request = Deoptimization::make_trap_request(reason, action); 1958 } else { 1959 C->set_trap_can_recompile(true); 1960 } 1961 break; 1962 case Deoptimization::Action_make_not_entrant: 1963 C->set_trap_can_recompile(true); 1964 break; 1965 #ifdef ASSERT 1966 case Deoptimization::Action_none: 1967 case Deoptimization::Action_make_not_compilable: 1968 break; 1969 default: 1970 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action)); 1971 break; 1972 #endif 1973 } 1974 1975 if (TraceOptoParse) { 1976 char buf[100]; 1977 tty->print_cr("Uncommon trap %s at bci:%d", 1978 Deoptimization::format_trap_request(buf, sizeof(buf), 1979 trap_request), bci()); 1980 } 1981 1982 CompileLog* log = C->log(); 1983 if (log != NULL) { 1984 int kid = (klass == NULL)? -1: log->identify(klass); 1985 log->begin_elem("uncommon_trap bci='%d'", bci()); 1986 char buf[100]; 1987 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf), 1988 trap_request)); 1989 if (kid >= 0) log->print(" klass='%d'", kid); 1990 if (comment != NULL) log->print(" comment='%s'", comment); 1991 log->end_elem(); 1992 } 1993 1994 // Make sure any guarding test views this path as very unlikely 1995 Node *i0 = control()->in(0); 1996 if (i0 != NULL && i0->is_If()) { // Found a guarding if test? 1997 IfNode *iff = i0->as_If(); 1998 float f = iff->_prob; // Get prob 1999 if (control()->Opcode() == Op_IfTrue) { 2000 if (f > PROB_UNLIKELY_MAG(4)) 2001 iff->_prob = PROB_MIN; 2002 } else { 2003 if (f < PROB_LIKELY_MAG(4)) 2004 iff->_prob = PROB_MAX; 2005 } 2006 } 2007 2008 // Clear out dead values from the debug info. 2009 kill_dead_locals(); 2010 2011 // Now insert the uncommon trap subroutine call 2012 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point(); 2013 const TypePtr* no_memory_effects = NULL; 2014 // Pass the index of the class to be loaded 2015 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON | 2016 (must_throw ? RC_MUST_THROW : 0), 2017 OptoRuntime::uncommon_trap_Type(), 2018 call_addr, "uncommon_trap", no_memory_effects, 2019 intcon(trap_request)); 2020 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request, 2021 "must extract request correctly from the graph"); 2022 assert(trap_request != 0, "zero value reserved by uncommon_trap_request"); 2023 2024 call->set_req(TypeFunc::ReturnAdr, returnadr()); 2025 // The debug info is the only real input to this call. 2026 2027 // Halt-and-catch fire here. The above call should never return! 2028 HaltNode* halt = new HaltNode(control(), frameptr()); 2029 _gvn.set_type_bottom(halt); 2030 root()->add_req(halt); 2031 2032 stop_and_kill_map(); 2033 } 2034 2035 2036 //--------------------------just_allocated_object------------------------------ 2037 // Report the object that was just allocated. 2038 // It must be the case that there are no intervening safepoints. 2039 // We use this to determine if an object is so "fresh" that 2040 // it does not require card marks. 2041 Node* GraphKit::just_allocated_object(Node* current_control) { 2042 if (C->recent_alloc_ctl() == current_control) 2043 return C->recent_alloc_obj(); 2044 return NULL; 2045 } 2046 2047 2048 void GraphKit::round_double_arguments(ciMethod* dest_method) { 2049 // (Note: TypeFunc::make has a cache that makes this fast.) 2050 const TypeFunc* tf = TypeFunc::make(dest_method); 2051 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2052 for (int j = 0; j < nargs; j++) { 2053 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2054 if( targ->basic_type() == T_DOUBLE ) { 2055 // If any parameters are doubles, they must be rounded before 2056 // the call, dstore_rounding does gvn.transform 2057 Node *arg = argument(j); 2058 arg = dstore_rounding(arg); 2059 set_argument(j, arg); 2060 } 2061 } 2062 } 2063 2064 /** 2065 * Record profiling data exact_kls for Node n with the type system so 2066 * that it can propagate it (speculation) 2067 * 2068 * @param n node that the type applies to 2069 * @param exact_kls type from profiling 2070 * @param maybe_null did profiling see null? 2071 * 2072 * @return node with improved type 2073 */ 2074 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, bool maybe_null) { 2075 const Type* current_type = _gvn.type(n); 2076 assert(UseTypeSpeculation, "type speculation must be on"); 2077 2078 const TypePtr* speculative = current_type->speculative(); 2079 2080 // Should the klass from the profile be recorded in the speculative type? 2081 if (current_type->would_improve_type(exact_kls, jvms()->depth())) { 2082 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls); 2083 const TypeOopPtr* xtype = tklass->as_instance_type(); 2084 assert(xtype->klass_is_exact(), "Should be exact"); 2085 // Any reason to believe n is not null (from this profiling or a previous one)? 2086 const TypePtr* ptr = (maybe_null && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL; 2087 // record the new speculative type's depth 2088 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2089 speculative = speculative->with_inline_depth(jvms()->depth()); 2090 } else if (current_type->would_improve_ptr(maybe_null)) { 2091 // Profiling report that null was never seen so we can change the 2092 // speculative type to non null ptr. 2093 assert(!maybe_null, "nothing to improve"); 2094 if (speculative == NULL) { 2095 speculative = TypePtr::NOTNULL; 2096 } else { 2097 const TypePtr* ptr = TypePtr::NOTNULL; 2098 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr(); 2099 } 2100 } 2101 2102 if (speculative != current_type->speculative()) { 2103 // Build a type with a speculative type (what we think we know 2104 // about the type but will need a guard when we use it) 2105 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative); 2106 // We're changing the type, we need a new CheckCast node to carry 2107 // the new type. The new type depends on the control: what 2108 // profiling tells us is only valid from here as far as we can 2109 // tell. 2110 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type)); 2111 cast = _gvn.transform(cast); 2112 replace_in_map(n, cast); 2113 n = cast; 2114 } 2115 2116 return n; 2117 } 2118 2119 /** 2120 * Record profiling data from receiver profiling at an invoke with the 2121 * type system so that it can propagate it (speculation) 2122 * 2123 * @param n receiver node 2124 * 2125 * @return node with improved type 2126 */ 2127 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) { 2128 if (!UseTypeSpeculation) { 2129 return n; 2130 } 2131 ciKlass* exact_kls = profile_has_unique_klass(); 2132 bool maybe_null = true; 2133 if (java_bc() == Bytecodes::_checkcast || 2134 java_bc() == Bytecodes::_instanceof || 2135 java_bc() == Bytecodes::_aastore) { 2136 ciProfileData* data = method()->method_data()->bci_to_data(bci()); 2137 maybe_null = data == NULL ? true : data->as_BitData()->null_seen(); 2138 } 2139 return record_profile_for_speculation(n, exact_kls, maybe_null); 2140 } 2141 2142 /** 2143 * Record profiling data from argument profiling at an invoke with the 2144 * type system so that it can propagate it (speculation) 2145 * 2146 * @param dest_method target method for the call 2147 * @param bc what invoke bytecode is this? 2148 */ 2149 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) { 2150 if (!UseTypeSpeculation) { 2151 return; 2152 } 2153 const TypeFunc* tf = TypeFunc::make(dest_method); 2154 int nargs = tf->domain()->cnt() - TypeFunc::Parms; 2155 int skip = Bytecodes::has_receiver(bc) ? 1 : 0; 2156 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) { 2157 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms); 2158 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) { 2159 bool maybe_null = true; 2160 ciKlass* better_type = NULL; 2161 if (method()->argument_profiled_type(bci(), i, better_type, maybe_null)) { 2162 record_profile_for_speculation(argument(j), better_type, maybe_null); 2163 } 2164 i++; 2165 } 2166 } 2167 } 2168 2169 /** 2170 * Record profiling data from parameter profiling at an invoke with 2171 * the type system so that it can propagate it (speculation) 2172 */ 2173 void GraphKit::record_profiled_parameters_for_speculation() { 2174 if (!UseTypeSpeculation) { 2175 return; 2176 } 2177 for (int i = 0, j = 0; i < method()->arg_size() ; i++) { 2178 if (_gvn.type(local(i))->isa_oopptr()) { 2179 bool maybe_null = true; 2180 ciKlass* better_type = NULL; 2181 if (method()->parameter_profiled_type(j, better_type, maybe_null)) { 2182 record_profile_for_speculation(local(i), better_type, maybe_null); 2183 } 2184 j++; 2185 } 2186 } 2187 } 2188 2189 /** 2190 * Record profiling data from return value profiling at an invoke with 2191 * the type system so that it can propagate it (speculation) 2192 */ 2193 void GraphKit::record_profiled_return_for_speculation() { 2194 if (!UseTypeSpeculation) { 2195 return; 2196 } 2197 bool maybe_null = true; 2198 ciKlass* better_type = NULL; 2199 if (method()->return_profiled_type(bci(), better_type, maybe_null)) { 2200 // If profiling reports a single type for the return value, 2201 // feed it to the type system so it can propagate it as a 2202 // speculative type 2203 record_profile_for_speculation(stack(sp()-1), better_type, maybe_null); 2204 } 2205 } 2206 2207 void GraphKit::round_double_result(ciMethod* dest_method) { 2208 // A non-strict method may return a double value which has an extended 2209 // exponent, but this must not be visible in a caller which is 'strict' 2210 // If a strict caller invokes a non-strict callee, round a double result 2211 2212 BasicType result_type = dest_method->return_type()->basic_type(); 2213 assert( method() != NULL, "must have caller context"); 2214 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) { 2215 // Destination method's return value is on top of stack 2216 // dstore_rounding() does gvn.transform 2217 Node *result = pop_pair(); 2218 result = dstore_rounding(result); 2219 push_pair(result); 2220 } 2221 } 2222 2223 // rounding for strict float precision conformance 2224 Node* GraphKit::precision_rounding(Node* n) { 2225 return UseStrictFP && _method->flags().is_strict() 2226 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding 2227 ? _gvn.transform( new RoundFloatNode(0, n) ) 2228 : n; 2229 } 2230 2231 // rounding for strict double precision conformance 2232 Node* GraphKit::dprecision_rounding(Node *n) { 2233 return UseStrictFP && _method->flags().is_strict() 2234 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding 2235 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2236 : n; 2237 } 2238 2239 // rounding for non-strict double stores 2240 Node* GraphKit::dstore_rounding(Node* n) { 2241 return Matcher::strict_fp_requires_explicit_rounding 2242 && UseSSE <= 1 2243 ? _gvn.transform( new RoundDoubleNode(0, n) ) 2244 : n; 2245 } 2246 2247 //============================================================================= 2248 // Generate a fast path/slow path idiom. Graph looks like: 2249 // [foo] indicates that 'foo' is a parameter 2250 // 2251 // [in] NULL 2252 // \ / 2253 // CmpP 2254 // Bool ne 2255 // If 2256 // / \ 2257 // True False-<2> 2258 // / | 2259 // / cast_not_null 2260 // Load | | ^ 2261 // [fast_test] | | 2262 // gvn to opt_test | | 2263 // / \ | <1> 2264 // True False | 2265 // | \\ | 2266 // [slow_call] \[fast_result] 2267 // Ctl Val \ \ 2268 // | \ \ 2269 // Catch <1> \ \ 2270 // / \ ^ \ \ 2271 // Ex No_Ex | \ \ 2272 // | \ \ | \ <2> \ 2273 // ... \ [slow_res] | | \ [null_result] 2274 // \ \--+--+--- | | 2275 // \ | / \ | / 2276 // --------Region Phi 2277 // 2278 //============================================================================= 2279 // Code is structured as a series of driver functions all called 'do_XXX' that 2280 // call a set of helper functions. Helper functions first, then drivers. 2281 2282 //------------------------------null_check_oop--------------------------------- 2283 // Null check oop. Set null-path control into Region in slot 3. 2284 // Make a cast-not-nullness use the other not-null control. Return cast. 2285 Node* GraphKit::null_check_oop(Node* value, Node* *null_control, 2286 bool never_see_null, 2287 bool safe_for_replace, 2288 bool speculative) { 2289 // Initial NULL check taken path 2290 (*null_control) = top(); 2291 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative); 2292 2293 // Generate uncommon_trap: 2294 if (never_see_null && (*null_control) != top()) { 2295 // If we see an unexpected null at a check-cast we record it and force a 2296 // recompile; the offending check-cast will be compiled to handle NULLs. 2297 // If we see more than one offending BCI, then all checkcasts in the 2298 // method will be compiled to handle NULLs. 2299 PreserveJVMState pjvms(this); 2300 set_control(*null_control); 2301 replace_in_map(value, null()); 2302 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative); 2303 uncommon_trap(reason, 2304 Deoptimization::Action_make_not_entrant); 2305 (*null_control) = top(); // NULL path is dead 2306 } 2307 if ((*null_control) == top() && safe_for_replace) { 2308 replace_in_map(value, cast); 2309 } 2310 2311 // Cast away null-ness on the result 2312 return cast; 2313 } 2314 2315 //------------------------------opt_iff---------------------------------------- 2316 // Optimize the fast-check IfNode. Set the fast-path region slot 2. 2317 // Return slow-path control. 2318 Node* GraphKit::opt_iff(Node* region, Node* iff) { 2319 IfNode *opt_iff = _gvn.transform(iff)->as_If(); 2320 2321 // Fast path taken; set region slot 2 2322 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) ); 2323 region->init_req(2,fast_taken); // Capture fast-control 2324 2325 // Fast path not-taken, i.e. slow path 2326 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) ); 2327 return slow_taken; 2328 } 2329 2330 //-----------------------------make_runtime_call------------------------------- 2331 Node* GraphKit::make_runtime_call(int flags, 2332 const TypeFunc* call_type, address call_addr, 2333 const char* call_name, 2334 const TypePtr* adr_type, 2335 // The following parms are all optional. 2336 // The first NULL ends the list. 2337 Node* parm0, Node* parm1, 2338 Node* parm2, Node* parm3, 2339 Node* parm4, Node* parm5, 2340 Node* parm6, Node* parm7) { 2341 // Slow-path call 2342 bool is_leaf = !(flags & RC_NO_LEAF); 2343 bool has_io = (!is_leaf && !(flags & RC_NO_IO)); 2344 if (call_name == NULL) { 2345 assert(!is_leaf, "must supply name for leaf"); 2346 call_name = OptoRuntime::stub_name(call_addr); 2347 } 2348 CallNode* call; 2349 if (!is_leaf) { 2350 call = new CallStaticJavaNode(call_type, call_addr, call_name, 2351 bci(), adr_type); 2352 } else if (flags & RC_NO_FP) { 2353 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type); 2354 } else { 2355 call = new CallLeafNode(call_type, call_addr, call_name, adr_type); 2356 } 2357 2358 // The following is similar to set_edges_for_java_call, 2359 // except that the memory effects of the call are restricted to AliasIdxRaw. 2360 2361 // Slow path call has no side-effects, uses few values 2362 bool wide_in = !(flags & RC_NARROW_MEM); 2363 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot); 2364 2365 Node* prev_mem = NULL; 2366 if (wide_in) { 2367 prev_mem = set_predefined_input_for_runtime_call(call); 2368 } else { 2369 assert(!wide_out, "narrow in => narrow out"); 2370 Node* narrow_mem = memory(adr_type); 2371 prev_mem = reset_memory(); 2372 map()->set_memory(narrow_mem); 2373 set_predefined_input_for_runtime_call(call); 2374 } 2375 2376 // Hook each parm in order. Stop looking at the first NULL. 2377 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0); 2378 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1); 2379 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2); 2380 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3); 2381 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4); 2382 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5); 2383 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6); 2384 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7); 2385 /* close each nested if ===> */ } } } } } } } } 2386 assert(call->in(call->req()-1) != NULL, "must initialize all parms"); 2387 2388 if (!is_leaf) { 2389 // Non-leaves can block and take safepoints: 2390 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0)); 2391 } 2392 // Non-leaves can throw exceptions: 2393 if (has_io) { 2394 call->set_req(TypeFunc::I_O, i_o()); 2395 } 2396 2397 if (flags & RC_UNCOMMON) { 2398 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency. 2399 // (An "if" probability corresponds roughly to an unconditional count. 2400 // Sort of.) 2401 call->set_cnt(PROB_UNLIKELY_MAG(4)); 2402 } 2403 2404 Node* c = _gvn.transform(call); 2405 assert(c == call, "cannot disappear"); 2406 2407 if (wide_out) { 2408 // Slow path call has full side-effects. 2409 set_predefined_output_for_runtime_call(call); 2410 } else { 2411 // Slow path call has few side-effects, and/or sets few values. 2412 set_predefined_output_for_runtime_call(call, prev_mem, adr_type); 2413 } 2414 2415 if (has_io) { 2416 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O))); 2417 } 2418 return call; 2419 2420 } 2421 2422 //------------------------------merge_memory----------------------------------- 2423 // Merge memory from one path into the current memory state. 2424 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) { 2425 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) { 2426 Node* old_slice = mms.force_memory(); 2427 Node* new_slice = mms.memory2(); 2428 if (old_slice != new_slice) { 2429 PhiNode* phi; 2430 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) { 2431 if (mms.is_empty()) { 2432 // clone base memory Phi's inputs for this memory slice 2433 assert(old_slice == mms.base_memory(), "sanity"); 2434 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C)); 2435 _gvn.set_type(phi, Type::MEMORY); 2436 for (uint i = 1; i < phi->req(); i++) { 2437 phi->init_req(i, old_slice->in(i)); 2438 } 2439 } else { 2440 phi = old_slice->as_Phi(); // Phi was generated already 2441 } 2442 } else { 2443 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C)); 2444 _gvn.set_type(phi, Type::MEMORY); 2445 } 2446 phi->set_req(new_path, new_slice); 2447 mms.set_memory(phi); 2448 } 2449 } 2450 } 2451 2452 //------------------------------make_slow_call_ex------------------------------ 2453 // Make the exception handler hookups for the slow call 2454 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) { 2455 if (stopped()) return; 2456 2457 // Make a catch node with just two handlers: fall-through and catch-all 2458 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) ); 2459 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) ); 2460 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) ); 2461 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) ); 2462 2463 { PreserveJVMState pjvms(this); 2464 set_control(excp); 2465 set_i_o(i_o); 2466 2467 if (excp != top()) { 2468 if (deoptimize) { 2469 // Deoptimize if an exception is caught. Don't construct exception state in this case. 2470 uncommon_trap(Deoptimization::Reason_unhandled, 2471 Deoptimization::Action_none); 2472 } else { 2473 // Create an exception state also. 2474 // Use an exact type if the caller has specified a specific exception. 2475 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull); 2476 Node* ex_oop = new CreateExNode(ex_type, control(), i_o); 2477 add_exception_state(make_exception_state(_gvn.transform(ex_oop))); 2478 } 2479 } 2480 } 2481 2482 // Get the no-exception control from the CatchNode. 2483 set_control(norm); 2484 } 2485 2486 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) { 2487 Node* cmp = NULL; 2488 switch(bt) { 2489 case T_INT: cmp = new CmpINode(in1, in2); break; 2490 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break; 2491 default: fatal("unexpected comparison type %s", type2name(bt)); 2492 } 2493 gvn->transform(cmp); 2494 Node* bol = gvn->transform(new BoolNode(cmp, test)); 2495 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN); 2496 gvn->transform(iff); 2497 if (!bol->is_Con()) gvn->record_for_igvn(iff); 2498 return iff; 2499 } 2500 2501 2502 //-------------------------------gen_subtype_check----------------------------- 2503 // Generate a subtyping check. Takes as input the subtype and supertype. 2504 // Returns 2 values: sets the default control() to the true path and returns 2505 // the false path. Only reads invariant memory; sets no (visible) memory. 2506 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding 2507 // but that's not exposed to the optimizer. This call also doesn't take in an 2508 // Object; if you wish to check an Object you need to load the Object's class 2509 // prior to coming here. 2510 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) { 2511 Compile* C = gvn->C; 2512 2513 if ((*ctrl)->is_top()) { 2514 return C->top(); 2515 } 2516 2517 // Fast check for identical types, perhaps identical constants. 2518 // The types can even be identical non-constants, in cases 2519 // involving Array.newInstance, Object.clone, etc. 2520 if (subklass == superklass) 2521 return C->top(); // false path is dead; no test needed. 2522 2523 if (gvn->type(superklass)->singleton()) { 2524 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass(); 2525 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass(); 2526 2527 // In the common case of an exact superklass, try to fold up the 2528 // test before generating code. You may ask, why not just generate 2529 // the code and then let it fold up? The answer is that the generated 2530 // code will necessarily include null checks, which do not always 2531 // completely fold away. If they are also needless, then they turn 2532 // into a performance loss. Example: 2533 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x; 2534 // Here, the type of 'fa' is often exact, so the store check 2535 // of fa[1]=x will fold up, without testing the nullness of x. 2536 switch (C->static_subtype_check(superk, subk)) { 2537 case Compile::SSC_always_false: 2538 { 2539 Node* always_fail = *ctrl; 2540 *ctrl = gvn->C->top(); 2541 return always_fail; 2542 } 2543 case Compile::SSC_always_true: 2544 return C->top(); 2545 case Compile::SSC_easy_test: 2546 { 2547 // Just do a direct pointer compare and be done. 2548 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS); 2549 *ctrl = gvn->transform(new IfTrueNode(iff)); 2550 return gvn->transform(new IfFalseNode(iff)); 2551 } 2552 case Compile::SSC_full_test: 2553 break; 2554 default: 2555 ShouldNotReachHere(); 2556 } 2557 } 2558 2559 // %%% Possible further optimization: Even if the superklass is not exact, 2560 // if the subklass is the unique subtype of the superklass, the check 2561 // will always succeed. We could leave a dependency behind to ensure this. 2562 2563 // First load the super-klass's check-offset 2564 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset())))); 2565 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr())); 2566 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered)); 2567 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset()); 2568 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con); 2569 2570 // Load from the sub-klass's super-class display list, or a 1-word cache of 2571 // the secondary superclass list, or a failing value with a sentinel offset 2572 // if the super-klass is an interface or exceptionally deep in the Java 2573 // hierarchy and we have to scan the secondary superclass list the hard way. 2574 // Worst-case type is a little odd: NULL is allowed as a result (usually 2575 // klass loads can never produce a NULL). 2576 Node *chk_off_X = chk_off; 2577 #ifdef _LP64 2578 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X)); 2579 #endif 2580 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X)); 2581 // For some types like interfaces the following loadKlass is from a 1-word 2582 // cache which is mutable so can't use immutable memory. Other 2583 // types load from the super-class display table which is immutable. 2584 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr())); 2585 Node *kmem = might_be_cache ? m : C->immutable_memory(); 2586 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL)); 2587 2588 // Compile speed common case: ARE a subtype and we canNOT fail 2589 if( superklass == nkls ) 2590 return C->top(); // false path is dead; no test needed. 2591 2592 // See if we get an immediate positive hit. Happens roughly 83% of the 2593 // time. Test to see if the value loaded just previously from the subklass 2594 // is exactly the superklass. 2595 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS); 2596 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1)); 2597 *ctrl = gvn->transform(new IfFalseNode(iff1)); 2598 2599 // Compile speed common case: Check for being deterministic right now. If 2600 // chk_off is a constant and not equal to cacheoff then we are NOT a 2601 // subklass. In this case we need exactly the 1 test above and we can 2602 // return those results immediately. 2603 if (!might_be_cache) { 2604 Node* not_subtype_ctrl = *ctrl; 2605 *ctrl = iftrue1; // We need exactly the 1 test above 2606 return not_subtype_ctrl; 2607 } 2608 2609 // Gather the various success & failures here 2610 RegionNode *r_ok_subtype = new RegionNode(4); 2611 gvn->record_for_igvn(r_ok_subtype); 2612 RegionNode *r_not_subtype = new RegionNode(3); 2613 gvn->record_for_igvn(r_not_subtype); 2614 2615 r_ok_subtype->init_req(1, iftrue1); 2616 2617 // Check for immediate negative hit. Happens roughly 11% of the time (which 2618 // is roughly 63% of the remaining cases). Test to see if the loaded 2619 // check-offset points into the subklass display list or the 1-element 2620 // cache. If it points to the display (and NOT the cache) and the display 2621 // missed then it's not a subtype. 2622 Node *cacheoff = gvn->intcon(cacheoff_con); 2623 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT); 2624 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2))); 2625 *ctrl = gvn->transform(new IfFalseNode(iff2)); 2626 2627 // Check for self. Very rare to get here, but it is taken 1/3 the time. 2628 // No performance impact (too rare) but allows sharing of secondary arrays 2629 // which has some footprint reduction. 2630 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS); 2631 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3))); 2632 *ctrl = gvn->transform(new IfFalseNode(iff3)); 2633 2634 // -- Roads not taken here: -- 2635 // We could also have chosen to perform the self-check at the beginning 2636 // of this code sequence, as the assembler does. This would not pay off 2637 // the same way, since the optimizer, unlike the assembler, can perform 2638 // static type analysis to fold away many successful self-checks. 2639 // Non-foldable self checks work better here in second position, because 2640 // the initial primary superclass check subsumes a self-check for most 2641 // types. An exception would be a secondary type like array-of-interface, 2642 // which does not appear in its own primary supertype display. 2643 // Finally, we could have chosen to move the self-check into the 2644 // PartialSubtypeCheckNode, and from there out-of-line in a platform 2645 // dependent manner. But it is worthwhile to have the check here, 2646 // where it can be perhaps be optimized. The cost in code space is 2647 // small (register compare, branch). 2648 2649 // Now do a linear scan of the secondary super-klass array. Again, no real 2650 // performance impact (too rare) but it's gotta be done. 2651 // Since the code is rarely used, there is no penalty for moving it 2652 // out of line, and it can only improve I-cache density. 2653 // The decision to inline or out-of-line this final check is platform 2654 // dependent, and is found in the AD file definition of PartialSubtypeCheck. 2655 Node* psc = gvn->transform( 2656 new PartialSubtypeCheckNode(*ctrl, subklass, superklass)); 2657 2658 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS); 2659 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4))); 2660 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4))); 2661 2662 // Return false path; set default control to true path. 2663 *ctrl = gvn->transform(r_ok_subtype); 2664 return gvn->transform(r_not_subtype); 2665 } 2666 2667 // Profile-driven exact type check: 2668 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass, 2669 float prob, 2670 Node* *casted_receiver) { 2671 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass); 2672 Node* recv_klass = load_object_klass(receiver); 2673 Node* want_klass = makecon(tklass); 2674 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) ); 2675 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) ); 2676 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN); 2677 set_control( _gvn.transform( new IfTrueNode (iff) )); 2678 Node* fail = _gvn.transform( new IfFalseNode(iff) ); 2679 2680 const TypeOopPtr* recv_xtype = tklass->as_instance_type(); 2681 assert(recv_xtype->klass_is_exact(), ""); 2682 2683 // Subsume downstream occurrences of receiver with a cast to 2684 // recv_xtype, since now we know what the type will be. 2685 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype); 2686 (*casted_receiver) = _gvn.transform(cast); 2687 // (User must make the replace_in_map call.) 2688 2689 return fail; 2690 } 2691 2692 2693 //------------------------------seems_never_null------------------------------- 2694 // Use null_seen information if it is available from the profile. 2695 // If we see an unexpected null at a type check we record it and force a 2696 // recompile; the offending check will be recompiled to handle NULLs. 2697 // If we see several offending BCIs, then all checks in the 2698 // method will be recompiled. 2699 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) { 2700 speculating = !_gvn.type(obj)->speculative_maybe_null(); 2701 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating); 2702 if (UncommonNullCast // Cutout for this technique 2703 && obj != null() // And not the -Xcomp stupid case? 2704 && !too_many_traps(reason) 2705 ) { 2706 if (speculating) { 2707 return true; 2708 } 2709 if (data == NULL) 2710 // Edge case: no mature data. Be optimistic here. 2711 return true; 2712 // If the profile has not seen a null, assume it won't happen. 2713 assert(java_bc() == Bytecodes::_checkcast || 2714 java_bc() == Bytecodes::_instanceof || 2715 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here"); 2716 return !data->as_BitData()->null_seen(); 2717 } 2718 speculating = false; 2719 return false; 2720 } 2721 2722 //------------------------maybe_cast_profiled_receiver------------------------- 2723 // If the profile has seen exactly one type, narrow to exactly that type. 2724 // Subsequent type checks will always fold up. 2725 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj, 2726 ciKlass* require_klass, 2727 ciKlass* spec_klass, 2728 bool safe_for_replace) { 2729 if (!UseTypeProfile || !TypeProfileCasts) return NULL; 2730 2731 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL); 2732 2733 // Make sure we haven't already deoptimized from this tactic. 2734 if (too_many_traps(reason) || too_many_recompiles(reason)) 2735 return NULL; 2736 2737 // (No, this isn't a call, but it's enough like a virtual call 2738 // to use the same ciMethod accessor to get the profile info...) 2739 // If we have a speculative type use it instead of profiling (which 2740 // may not help us) 2741 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass; 2742 if (exact_kls != NULL) {// no cast failures here 2743 if (require_klass == NULL || 2744 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) { 2745 // If we narrow the type to match what the type profile sees or 2746 // the speculative type, we can then remove the rest of the 2747 // cast. 2748 // This is a win, even if the exact_kls is very specific, 2749 // because downstream operations, such as method calls, 2750 // will often benefit from the sharper type. 2751 Node* exact_obj = not_null_obj; // will get updated in place... 2752 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2753 &exact_obj); 2754 { PreserveJVMState pjvms(this); 2755 set_control(slow_ctl); 2756 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2757 } 2758 if (safe_for_replace) { 2759 replace_in_map(not_null_obj, exact_obj); 2760 } 2761 return exact_obj; 2762 } 2763 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us. 2764 } 2765 2766 return NULL; 2767 } 2768 2769 /** 2770 * Cast obj to type and emit guard unless we had too many traps here 2771 * already 2772 * 2773 * @param obj node being casted 2774 * @param type type to cast the node to 2775 * @param not_null true if we know node cannot be null 2776 */ 2777 Node* GraphKit::maybe_cast_profiled_obj(Node* obj, 2778 ciKlass* type, 2779 bool not_null) { 2780 if (stopped()) { 2781 return obj; 2782 } 2783 2784 // type == NULL if profiling tells us this object is always null 2785 if (type != NULL) { 2786 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check; 2787 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check; 2788 2789 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) && 2790 !too_many_traps(class_reason) && 2791 !too_many_recompiles(class_reason)) { 2792 Node* not_null_obj = NULL; 2793 // not_null is true if we know the object is not null and 2794 // there's no need for a null check 2795 if (!not_null) { 2796 Node* null_ctl = top(); 2797 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true); 2798 assert(null_ctl->is_top(), "no null control here"); 2799 } else { 2800 not_null_obj = obj; 2801 } 2802 2803 Node* exact_obj = not_null_obj; 2804 ciKlass* exact_kls = type; 2805 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0, 2806 &exact_obj); 2807 { 2808 PreserveJVMState pjvms(this); 2809 set_control(slow_ctl); 2810 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile); 2811 } 2812 replace_in_map(not_null_obj, exact_obj); 2813 obj = exact_obj; 2814 } 2815 } else { 2816 if (!too_many_traps(Deoptimization::Reason_null_assert) && 2817 !too_many_recompiles(Deoptimization::Reason_null_assert)) { 2818 Node* exact_obj = null_assert(obj); 2819 replace_in_map(obj, exact_obj); 2820 obj = exact_obj; 2821 } 2822 } 2823 return obj; 2824 } 2825 2826 //-------------------------------gen_instanceof-------------------------------- 2827 // Generate an instance-of idiom. Used by both the instance-of bytecode 2828 // and the reflective instance-of call. 2829 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) { 2830 kill_dead_locals(); // Benefit all the uncommon traps 2831 assert( !stopped(), "dead parse path should be checked in callers" ); 2832 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()), 2833 "must check for not-null not-dead klass in callers"); 2834 2835 // Make the merge point 2836 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT }; 2837 RegionNode* region = new RegionNode(PATH_LIMIT); 2838 Node* phi = new PhiNode(region, TypeInt::BOOL); 2839 C->set_has_split_ifs(true); // Has chance for split-if optimization 2840 2841 ciProfileData* data = NULL; 2842 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode 2843 data = method()->method_data()->bci_to_data(bci()); 2844 } 2845 bool speculative_not_null = false; 2846 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile 2847 && seems_never_null(obj, data, speculative_not_null)); 2848 2849 // Null check; get casted pointer; set region slot 3 2850 Node* null_ctl = top(); 2851 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2852 2853 // If not_null_obj is dead, only null-path is taken 2854 if (stopped()) { // Doing instance-of on a NULL? 2855 set_control(null_ctl); 2856 return intcon(0); 2857 } 2858 region->init_req(_null_path, null_ctl); 2859 phi ->init_req(_null_path, intcon(0)); // Set null path value 2860 if (null_ctl == top()) { 2861 // Do this eagerly, so that pattern matches like is_diamond_phi 2862 // will work even during parsing. 2863 assert(_null_path == PATH_LIMIT-1, "delete last"); 2864 region->del_req(_null_path); 2865 phi ->del_req(_null_path); 2866 } 2867 2868 // Do we know the type check always succeed? 2869 bool known_statically = false; 2870 if (_gvn.type(superklass)->singleton()) { 2871 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass(); 2872 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass(); 2873 if (subk != NULL && subk->is_loaded()) { 2874 int static_res = C->static_subtype_check(superk, subk); 2875 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false); 2876 } 2877 } 2878 2879 if (known_statically && UseTypeSpeculation) { 2880 // If we know the type check always succeeds then we don't use the 2881 // profiling data at this bytecode. Don't lose it, feed it to the 2882 // type system as a speculative type. 2883 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj); 2884 } else { 2885 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2886 // We may not have profiling here or it may not help us. If we 2887 // have a speculative type use it to perform an exact cast. 2888 ciKlass* spec_obj_type = obj_type->speculative_type(); 2889 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) { 2890 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace); 2891 if (stopped()) { // Profile disagrees with this path. 2892 set_control(null_ctl); // Null is the only remaining possibility. 2893 return intcon(0); 2894 } 2895 if (cast_obj != NULL) { 2896 not_null_obj = cast_obj; 2897 } 2898 } 2899 } 2900 2901 // Load the object's klass 2902 Node* obj_klass = load_object_klass(not_null_obj); 2903 2904 // Generate the subtype check 2905 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass); 2906 2907 // Plug in the success path to the general merge in slot 1. 2908 region->init_req(_obj_path, control()); 2909 phi ->init_req(_obj_path, intcon(1)); 2910 2911 // Plug in the failing path to the general merge in slot 2. 2912 region->init_req(_fail_path, not_subtype_ctrl); 2913 phi ->init_req(_fail_path, intcon(0)); 2914 2915 // Return final merged results 2916 set_control( _gvn.transform(region) ); 2917 record_for_igvn(region); 2918 return _gvn.transform(phi); 2919 } 2920 2921 //-------------------------------gen_checkcast--------------------------------- 2922 // Generate a checkcast idiom. Used by both the checkcast bytecode and the 2923 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the 2924 // uncommon-trap paths work. Adjust stack after this call. 2925 // If failure_control is supplied and not null, it is filled in with 2926 // the control edge for the cast failure. Otherwise, an appropriate 2927 // uncommon trap or exception is thrown. 2928 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, 2929 Node* *failure_control) { 2930 kill_dead_locals(); // Benefit all the uncommon traps 2931 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr(); 2932 const Type *toop = TypeOopPtr::make_from_klass(tk->klass()); 2933 2934 // Fast cutout: Check the case that the cast is vacuously true. 2935 // This detects the common cases where the test will short-circuit 2936 // away completely. We do this before we perform the null check, 2937 // because if the test is going to turn into zero code, we don't 2938 // want a residual null check left around. (Causes a slowdown, 2939 // for example, in some objArray manipulations, such as a[i]=a[j].) 2940 if (tk->singleton()) { 2941 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr(); 2942 if (objtp != NULL && objtp->klass() != NULL) { 2943 switch (C->static_subtype_check(tk->klass(), objtp->klass())) { 2944 case Compile::SSC_always_true: 2945 // If we know the type check always succeed then we don't use 2946 // the profiling data at this bytecode. Don't lose it, feed it 2947 // to the type system as a speculative type. 2948 return record_profiled_receiver_for_speculation(obj); 2949 case Compile::SSC_always_false: 2950 // It needs a null check because a null will *pass* the cast check. 2951 // A non-null value will always produce an exception. 2952 return null_assert(obj); 2953 } 2954 } 2955 } 2956 2957 ciProfileData* data = NULL; 2958 bool safe_for_replace = false; 2959 if (failure_control == NULL) { // use MDO in regular case only 2960 assert(java_bc() == Bytecodes::_aastore || 2961 java_bc() == Bytecodes::_checkcast, 2962 "interpreter profiles type checks only for these BCs"); 2963 data = method()->method_data()->bci_to_data(bci()); 2964 safe_for_replace = true; 2965 } 2966 2967 // Make the merge point 2968 enum { _obj_path = 1, _null_path, PATH_LIMIT }; 2969 RegionNode* region = new RegionNode(PATH_LIMIT); 2970 Node* phi = new PhiNode(region, toop); 2971 C->set_has_split_ifs(true); // Has chance for split-if optimization 2972 2973 // Use null-cast information if it is available 2974 bool speculative_not_null = false; 2975 bool never_see_null = ((failure_control == NULL) // regular case only 2976 && seems_never_null(obj, data, speculative_not_null)); 2977 2978 // Null check; get casted pointer; set region slot 3 2979 Node* null_ctl = top(); 2980 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null); 2981 2982 // If not_null_obj is dead, only null-path is taken 2983 if (stopped()) { // Doing instance-of on a NULL? 2984 set_control(null_ctl); 2985 return null(); 2986 } 2987 region->init_req(_null_path, null_ctl); 2988 phi ->init_req(_null_path, null()); // Set null path value 2989 if (null_ctl == top()) { 2990 // Do this eagerly, so that pattern matches like is_diamond_phi 2991 // will work even during parsing. 2992 assert(_null_path == PATH_LIMIT-1, "delete last"); 2993 region->del_req(_null_path); 2994 phi ->del_req(_null_path); 2995 } 2996 2997 Node* cast_obj = NULL; 2998 if (tk->klass_is_exact()) { 2999 // The following optimization tries to statically cast the speculative type of the object 3000 // (for example obtained during profiling) to the type of the superklass and then do a 3001 // dynamic check that the type of the object is what we expect. To work correctly 3002 // for checkcast and aastore the type of superklass should be exact. 3003 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 3004 // We may not have profiling here or it may not help us. If we have 3005 // a speculative type use it to perform an exact cast. 3006 ciKlass* spec_obj_type = obj_type->speculative_type(); 3007 if (spec_obj_type != NULL || data != NULL) { 3008 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace); 3009 if (cast_obj != NULL) { 3010 if (failure_control != NULL) // failure is now impossible 3011 (*failure_control) = top(); 3012 // adjust the type of the phi to the exact klass: 3013 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR)); 3014 } 3015 } 3016 } 3017 3018 if (cast_obj == NULL) { 3019 // Load the object's klass 3020 Node* obj_klass = load_object_klass(not_null_obj); 3021 3022 // Generate the subtype check 3023 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass ); 3024 3025 // Plug in success path into the merge 3026 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop)); 3027 // Failure path ends in uncommon trap (or may be dead - failure impossible) 3028 if (failure_control == NULL) { 3029 if (not_subtype_ctrl != top()) { // If failure is possible 3030 PreserveJVMState pjvms(this); 3031 set_control(not_subtype_ctrl); 3032 builtin_throw(Deoptimization::Reason_class_check, obj_klass); 3033 } 3034 } else { 3035 (*failure_control) = not_subtype_ctrl; 3036 } 3037 } 3038 3039 region->init_req(_obj_path, control()); 3040 phi ->init_req(_obj_path, cast_obj); 3041 3042 // A merge of NULL or Casted-NotNull obj 3043 Node* res = _gvn.transform(phi); 3044 3045 // Note I do NOT always 'replace_in_map(obj,result)' here. 3046 // if( tk->klass()->can_be_primary_super() ) 3047 // This means that if I successfully store an Object into an array-of-String 3048 // I 'forget' that the Object is really now known to be a String. I have to 3049 // do this because we don't have true union types for interfaces - if I store 3050 // a Baz into an array-of-Interface and then tell the optimizer it's an 3051 // Interface, I forget that it's also a Baz and cannot do Baz-like field 3052 // references to it. FIX THIS WHEN UNION TYPES APPEAR! 3053 // replace_in_map( obj, res ); 3054 3055 // Return final merged results 3056 set_control( _gvn.transform(region) ); 3057 record_for_igvn(region); 3058 return res; 3059 } 3060 3061 //------------------------------next_monitor----------------------------------- 3062 // What number should be given to the next monitor? 3063 int GraphKit::next_monitor() { 3064 int current = jvms()->monitor_depth()* C->sync_stack_slots(); 3065 int next = current + C->sync_stack_slots(); 3066 // Keep the toplevel high water mark current: 3067 if (C->fixed_slots() < next) C->set_fixed_slots(next); 3068 return current; 3069 } 3070 3071 //------------------------------insert_mem_bar--------------------------------- 3072 // Memory barrier to avoid floating things around 3073 // The membar serves as a pinch point between both control and all memory slices. 3074 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) { 3075 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent); 3076 mb->init_req(TypeFunc::Control, control()); 3077 mb->init_req(TypeFunc::Memory, reset_memory()); 3078 Node* membar = _gvn.transform(mb); 3079 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3080 set_all_memory_call(membar); 3081 return membar; 3082 } 3083 3084 //-------------------------insert_mem_bar_volatile---------------------------- 3085 // Memory barrier to avoid floating things around 3086 // The membar serves as a pinch point between both control and memory(alias_idx). 3087 // If you want to make a pinch point on all memory slices, do not use this 3088 // function (even with AliasIdxBot); use insert_mem_bar() instead. 3089 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) { 3090 // When Parse::do_put_xxx updates a volatile field, it appends a series 3091 // of MemBarVolatile nodes, one for *each* volatile field alias category. 3092 // The first membar is on the same memory slice as the field store opcode. 3093 // This forces the membar to follow the store. (Bug 6500685 broke this.) 3094 // All the other membars (for other volatile slices, including AliasIdxBot, 3095 // which stands for all unknown volatile slices) are control-dependent 3096 // on the first membar. This prevents later volatile loads or stores 3097 // from sliding up past the just-emitted store. 3098 3099 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent); 3100 mb->set_req(TypeFunc::Control,control()); 3101 if (alias_idx == Compile::AliasIdxBot) { 3102 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory()); 3103 } else { 3104 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller"); 3105 mb->set_req(TypeFunc::Memory, memory(alias_idx)); 3106 } 3107 Node* membar = _gvn.transform(mb); 3108 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control))); 3109 if (alias_idx == Compile::AliasIdxBot) { 3110 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory))); 3111 } else { 3112 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx); 3113 } 3114 return membar; 3115 } 3116 3117 //------------------------------shared_lock------------------------------------ 3118 // Emit locking code. 3119 FastLockNode* GraphKit::shared_lock(Node* obj) { 3120 // bci is either a monitorenter bc or InvocationEntryBci 3121 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3122 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3123 3124 if( !GenerateSynchronizationCode ) 3125 return NULL; // Not locking things? 3126 if (stopped()) // Dead monitor? 3127 return NULL; 3128 3129 assert(dead_locals_are_killed(), "should kill locals before sync. point"); 3130 3131 // Box the stack location 3132 Node* box = _gvn.transform(new BoxLockNode(next_monitor())); 3133 Node* mem = reset_memory(); 3134 3135 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock(); 3136 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) { 3137 // Create the counters for this fast lock. 3138 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3139 } 3140 3141 // Create the rtm counters for this fast lock if needed. 3142 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci 3143 3144 // Add monitor to debug info for the slow path. If we block inside the 3145 // slow path and de-opt, we need the monitor hanging around 3146 map()->push_monitor( flock ); 3147 3148 const TypeFunc *tf = LockNode::lock_type(); 3149 LockNode *lock = new LockNode(C, tf); 3150 3151 lock->init_req( TypeFunc::Control, control() ); 3152 lock->init_req( TypeFunc::Memory , mem ); 3153 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3154 lock->init_req( TypeFunc::FramePtr, frameptr() ); 3155 lock->init_req( TypeFunc::ReturnAdr, top() ); 3156 3157 lock->init_req(TypeFunc::Parms + 0, obj); 3158 lock->init_req(TypeFunc::Parms + 1, box); 3159 lock->init_req(TypeFunc::Parms + 2, flock); 3160 add_safepoint_edges(lock); 3161 3162 lock = _gvn.transform( lock )->as_Lock(); 3163 3164 // lock has no side-effects, sets few values 3165 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM); 3166 3167 insert_mem_bar(Op_MemBarAcquireLock); 3168 3169 // Add this to the worklist so that the lock can be eliminated 3170 record_for_igvn(lock); 3171 3172 #ifndef PRODUCT 3173 if (PrintLockStatistics) { 3174 // Update the counter for this lock. Don't bother using an atomic 3175 // operation since we don't require absolute accuracy. 3176 lock->create_lock_counter(map()->jvms()); 3177 increment_counter(lock->counter()->addr()); 3178 } 3179 #endif 3180 3181 return flock; 3182 } 3183 3184 3185 //------------------------------shared_unlock---------------------------------- 3186 // Emit unlocking code. 3187 void GraphKit::shared_unlock(Node* box, Node* obj) { 3188 // bci is either a monitorenter bc or InvocationEntryBci 3189 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces 3190 assert(SynchronizationEntryBCI == InvocationEntryBci, ""); 3191 3192 if( !GenerateSynchronizationCode ) 3193 return; 3194 if (stopped()) { // Dead monitor? 3195 map()->pop_monitor(); // Kill monitor from debug info 3196 return; 3197 } 3198 3199 // Memory barrier to avoid floating things down past the locked region 3200 insert_mem_bar(Op_MemBarReleaseLock); 3201 3202 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type(); 3203 UnlockNode *unlock = new UnlockNode(C, tf); 3204 #ifdef ASSERT 3205 unlock->set_dbg_jvms(sync_jvms()); 3206 #endif 3207 uint raw_idx = Compile::AliasIdxRaw; 3208 unlock->init_req( TypeFunc::Control, control() ); 3209 unlock->init_req( TypeFunc::Memory , memory(raw_idx) ); 3210 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o 3211 unlock->init_req( TypeFunc::FramePtr, frameptr() ); 3212 unlock->init_req( TypeFunc::ReturnAdr, top() ); 3213 3214 unlock->init_req(TypeFunc::Parms + 0, obj); 3215 unlock->init_req(TypeFunc::Parms + 1, box); 3216 unlock = _gvn.transform(unlock)->as_Unlock(); 3217 3218 Node* mem = reset_memory(); 3219 3220 // unlock has no side-effects, sets few values 3221 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM); 3222 3223 // Kill monitor from debug info 3224 map()->pop_monitor( ); 3225 } 3226 3227 //-------------------------------get_layout_helper----------------------------- 3228 // If the given klass is a constant or known to be an array, 3229 // fetch the constant layout helper value into constant_value 3230 // and return (Node*)NULL. Otherwise, load the non-constant 3231 // layout helper value, and return the node which represents it. 3232 // This two-faced routine is useful because allocation sites 3233 // almost always feature constant types. 3234 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) { 3235 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr(); 3236 if (!StressReflectiveCode && inst_klass != NULL) { 3237 ciKlass* klass = inst_klass->klass(); 3238 bool xklass = inst_klass->klass_is_exact(); 3239 if (xklass || klass->is_array_klass()) { 3240 jint lhelper = klass->layout_helper(); 3241 if (lhelper != Klass::_lh_neutral_value) { 3242 constant_value = lhelper; 3243 return (Node*) NULL; 3244 } 3245 } 3246 } 3247 constant_value = Klass::_lh_neutral_value; // put in a known value 3248 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset())); 3249 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 3250 } 3251 3252 // We just put in an allocate/initialize with a big raw-memory effect. 3253 // Hook selected additional alias categories on the initialization. 3254 static void hook_memory_on_init(GraphKit& kit, int alias_idx, 3255 MergeMemNode* init_in_merge, 3256 Node* init_out_raw) { 3257 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory()); 3258 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, ""); 3259 3260 Node* prevmem = kit.memory(alias_idx); 3261 init_in_merge->set_memory_at(alias_idx, prevmem); 3262 kit.set_memory(init_out_raw, alias_idx); 3263 } 3264 3265 //---------------------------set_output_for_allocation------------------------- 3266 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc, 3267 const TypeOopPtr* oop_type, 3268 bool deoptimize_on_exception) { 3269 int rawidx = Compile::AliasIdxRaw; 3270 alloc->set_req( TypeFunc::FramePtr, frameptr() ); 3271 add_safepoint_edges(alloc); 3272 Node* allocx = _gvn.transform(alloc); 3273 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) ); 3274 // create memory projection for i_o 3275 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx ); 3276 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception); 3277 3278 // create a memory projection as for the normal control path 3279 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory)); 3280 set_memory(malloc, rawidx); 3281 3282 // a normal slow-call doesn't change i_o, but an allocation does 3283 // we create a separate i_o projection for the normal control path 3284 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) ); 3285 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) ); 3286 3287 // put in an initialization barrier 3288 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx, 3289 rawoop)->as_Initialize(); 3290 assert(alloc->initialization() == init, "2-way macro link must work"); 3291 assert(init ->allocation() == alloc, "2-way macro link must work"); 3292 { 3293 // Extract memory strands which may participate in the new object's 3294 // initialization, and source them from the new InitializeNode. 3295 // This will allow us to observe initializations when they occur, 3296 // and link them properly (as a group) to the InitializeNode. 3297 assert(init->in(InitializeNode::Memory) == malloc, ""); 3298 MergeMemNode* minit_in = MergeMemNode::make(malloc); 3299 init->set_req(InitializeNode::Memory, minit_in); 3300 record_for_igvn(minit_in); // fold it up later, if possible 3301 Node* minit_out = memory(rawidx); 3302 assert(minit_out->is_Proj() && minit_out->in(0) == init, ""); 3303 if (oop_type->isa_aryptr()) { 3304 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot); 3305 int elemidx = C->get_alias_index(telemref); 3306 hook_memory_on_init(*this, elemidx, minit_in, minit_out); 3307 } else if (oop_type->isa_instptr()) { 3308 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass(); 3309 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) { 3310 ciField* field = ik->nonstatic_field_at(i); 3311 if (field->offset() >= TrackedInitializationLimit * HeapWordSize) 3312 continue; // do not bother to track really large numbers of fields 3313 // Find (or create) the alias category for this field: 3314 int fieldidx = C->alias_type(field)->index(); 3315 hook_memory_on_init(*this, fieldidx, minit_in, minit_out); 3316 } 3317 } 3318 } 3319 3320 // Cast raw oop to the real thing... 3321 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type); 3322 javaoop = _gvn.transform(javaoop); 3323 C->set_recent_alloc(control(), javaoop); 3324 assert(just_allocated_object(control()) == javaoop, "just allocated"); 3325 3326 #ifdef ASSERT 3327 { // Verify that the AllocateNode::Ideal_allocation recognizers work: 3328 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc, 3329 "Ideal_allocation works"); 3330 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc, 3331 "Ideal_allocation works"); 3332 if (alloc->is_AllocateArray()) { 3333 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(), 3334 "Ideal_allocation works"); 3335 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(), 3336 "Ideal_allocation works"); 3337 } else { 3338 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please"); 3339 } 3340 } 3341 #endif //ASSERT 3342 3343 return javaoop; 3344 } 3345 3346 //---------------------------new_instance-------------------------------------- 3347 // This routine takes a klass_node which may be constant (for a static type) 3348 // or may be non-constant (for reflective code). It will work equally well 3349 // for either, and the graph will fold nicely if the optimizer later reduces 3350 // the type to a constant. 3351 // The optional arguments are for specialized use by intrinsics: 3352 // - If 'extra_slow_test' if not null is an extra condition for the slow-path. 3353 // - If 'return_size_val', report the the total object size to the caller. 3354 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize) 3355 Node* GraphKit::new_instance(Node* klass_node, 3356 Node* extra_slow_test, 3357 Node* *return_size_val, 3358 bool deoptimize_on_exception) { 3359 // Compute size in doublewords 3360 // The size is always an integral number of doublewords, represented 3361 // as a positive bytewise size stored in the klass's layout_helper. 3362 // The layout_helper also encodes (in a low bit) the need for a slow path. 3363 jint layout_con = Klass::_lh_neutral_value; 3364 Node* layout_val = get_layout_helper(klass_node, layout_con); 3365 int layout_is_con = (layout_val == NULL); 3366 3367 if (extra_slow_test == NULL) extra_slow_test = intcon(0); 3368 // Generate the initial go-slow test. It's either ALWAYS (return a 3369 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective 3370 // case) a computed value derived from the layout_helper. 3371 Node* initial_slow_test = NULL; 3372 if (layout_is_con) { 3373 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3374 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con); 3375 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test; 3376 } else { // reflective case 3377 // This reflective path is used by Unsafe.allocateInstance. 3378 // (It may be stress-tested by specifying StressReflectiveCode.) 3379 // Basically, we want to get into the VM is there's an illegal argument. 3380 Node* bit = intcon(Klass::_lh_instance_slow_path_bit); 3381 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) ); 3382 if (extra_slow_test != intcon(0)) { 3383 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) ); 3384 } 3385 // (Macro-expander will further convert this to a Bool, if necessary.) 3386 } 3387 3388 // Find the size in bytes. This is easy; it's the layout_helper. 3389 // The size value must be valid even if the slow path is taken. 3390 Node* size = NULL; 3391 if (layout_is_con) { 3392 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con)); 3393 } else { // reflective case 3394 // This reflective path is used by clone and Unsafe.allocateInstance. 3395 size = ConvI2X(layout_val); 3396 3397 // Clear the low bits to extract layout_helper_size_in_bytes: 3398 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit"); 3399 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong)); 3400 size = _gvn.transform( new AndXNode(size, mask) ); 3401 } 3402 if (return_size_val != NULL) { 3403 (*return_size_val) = size; 3404 } 3405 3406 // This is a precise notnull oop of the klass. 3407 // (Actually, it need not be precise if this is a reflective allocation.) 3408 // It's what we cast the result to. 3409 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr(); 3410 if (!tklass) tklass = TypeKlassPtr::OBJECT; 3411 const TypeOopPtr* oop_type = tklass->as_instance_type(); 3412 3413 // Now generate allocation code 3414 3415 // The entire memory state is needed for slow path of the allocation 3416 // since GC and deoptimization can happened. 3417 Node *mem = reset_memory(); 3418 set_all_memory(mem); // Create new memory state 3419 3420 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP), 3421 control(), mem, i_o(), 3422 size, klass_node, 3423 initial_slow_test); 3424 3425 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception); 3426 } 3427 3428 //-------------------------------new_array------------------------------------- 3429 // helper for both newarray and anewarray 3430 // The 'length' parameter is (obviously) the length of the array. 3431 // See comments on new_instance for the meaning of the other arguments. 3432 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable) 3433 Node* length, // number of array elements 3434 int nargs, // number of arguments to push back for uncommon trap 3435 Node* *return_size_val, 3436 bool deoptimize_on_exception) { 3437 jint layout_con = Klass::_lh_neutral_value; 3438 Node* layout_val = get_layout_helper(klass_node, layout_con); 3439 int layout_is_con = (layout_val == NULL); 3440 3441 if (!layout_is_con && !StressReflectiveCode && 3442 !too_many_traps(Deoptimization::Reason_class_check)) { 3443 // This is a reflective array creation site. 3444 // Optimistically assume that it is a subtype of Object[], 3445 // so that we can fold up all the address arithmetic. 3446 layout_con = Klass::array_layout_helper(T_OBJECT); 3447 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) ); 3448 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) ); 3449 { BuildCutout unless(this, bol_lh, PROB_MAX); 3450 inc_sp(nargs); 3451 uncommon_trap(Deoptimization::Reason_class_check, 3452 Deoptimization::Action_maybe_recompile); 3453 } 3454 layout_val = NULL; 3455 layout_is_con = true; 3456 } 3457 3458 // Generate the initial go-slow test. Make sure we do not overflow 3459 // if length is huge (near 2Gig) or negative! We do not need 3460 // exact double-words here, just a close approximation of needed 3461 // double-words. We can't add any offset or rounding bits, lest we 3462 // take a size -1 of bytes and make it positive. Use an unsigned 3463 // compare, so negative sizes look hugely positive. 3464 int fast_size_limit = FastAllocateSizeLimit; 3465 if (layout_is_con) { 3466 assert(!StressReflectiveCode, "stress mode does not use these paths"); 3467 // Increase the size limit if we have exact knowledge of array type. 3468 int log2_esize = Klass::layout_helper_log2_element_size(layout_con); 3469 fast_size_limit <<= (LogBytesPerLong - log2_esize); 3470 } 3471 3472 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) ); 3473 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) ); 3474 3475 // --- Size Computation --- 3476 // array_size = round_to_heap(array_header + (length << elem_shift)); 3477 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes) 3478 // and round_to(x, y) == ((x + y-1) & ~(y-1)) 3479 // The rounding mask is strength-reduced, if possible. 3480 int round_mask = MinObjAlignmentInBytes - 1; 3481 Node* header_size = NULL; 3482 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 3483 // (T_BYTE has the weakest alignment and size restrictions...) 3484 if (layout_is_con) { 3485 int hsize = Klass::layout_helper_header_size(layout_con); 3486 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3487 BasicType etype = Klass::layout_helper_element_type(layout_con); 3488 if ((round_mask & ~right_n_bits(eshift)) == 0) 3489 round_mask = 0; // strength-reduce it if it goes away completely 3490 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded"); 3491 assert(header_size_min <= hsize, "generic minimum is smallest"); 3492 header_size_min = hsize; 3493 header_size = intcon(hsize + round_mask); 3494 } else { 3495 Node* hss = intcon(Klass::_lh_header_size_shift); 3496 Node* hsm = intcon(Klass::_lh_header_size_mask); 3497 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) ); 3498 hsize = _gvn.transform( new AndINode(hsize, hsm) ); 3499 Node* mask = intcon(round_mask); 3500 header_size = _gvn.transform( new AddINode(hsize, mask) ); 3501 } 3502 3503 Node* elem_shift = NULL; 3504 if (layout_is_con) { 3505 int eshift = Klass::layout_helper_log2_element_size(layout_con); 3506 if (eshift != 0) 3507 elem_shift = intcon(eshift); 3508 } else { 3509 // There is no need to mask or shift this value. 3510 // The semantics of LShiftINode include an implicit mask to 0x1F. 3511 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 3512 elem_shift = layout_val; 3513 } 3514 3515 // Transition to native address size for all offset calculations: 3516 Node* lengthx = ConvI2X(length); 3517 Node* headerx = ConvI2X(header_size); 3518 #ifdef _LP64 3519 { const TypeInt* tilen = _gvn.find_int_type(length); 3520 if (tilen != NULL && tilen->_lo < 0) { 3521 // Add a manual constraint to a positive range. Cf. array_element_address. 3522 jint size_max = fast_size_limit; 3523 if (size_max > tilen->_hi) size_max = tilen->_hi; 3524 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin); 3525 3526 // Only do a narrow I2L conversion if the range check passed. 3527 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN); 3528 _gvn.transform(iff); 3529 RegionNode* region = new RegionNode(3); 3530 _gvn.set_type(region, Type::CONTROL); 3531 lengthx = new PhiNode(region, TypeLong::LONG); 3532 _gvn.set_type(lengthx, TypeLong::LONG); 3533 3534 // Range check passed. Use ConvI2L node with narrow type. 3535 Node* passed = IfFalse(iff); 3536 region->init_req(1, passed); 3537 // Make I2L conversion control dependent to prevent it from 3538 // floating above the range check during loop optimizations. 3539 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed)); 3540 3541 // Range check failed. Use ConvI2L with wide type because length may be invalid. 3542 region->init_req(2, IfTrue(iff)); 3543 lengthx->init_req(2, ConvI2X(length)); 3544 3545 set_control(region); 3546 record_for_igvn(region); 3547 record_for_igvn(lengthx); 3548 } 3549 } 3550 #endif 3551 3552 // Combine header size (plus rounding) and body size. Then round down. 3553 // This computation cannot overflow, because it is used only in two 3554 // places, one where the length is sharply limited, and the other 3555 // after a successful allocation. 3556 Node* abody = lengthx; 3557 if (elem_shift != NULL) 3558 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) ); 3559 Node* size = _gvn.transform( new AddXNode(headerx, abody) ); 3560 if (round_mask != 0) { 3561 Node* mask = MakeConX(~round_mask); 3562 size = _gvn.transform( new AndXNode(size, mask) ); 3563 } 3564 // else if round_mask == 0, the size computation is self-rounding 3565 3566 if (return_size_val != NULL) { 3567 // This is the size 3568 (*return_size_val) = size; 3569 } 3570 3571 // Now generate allocation code 3572 3573 // The entire memory state is needed for slow path of the allocation 3574 // since GC and deoptimization can happened. 3575 Node *mem = reset_memory(); 3576 set_all_memory(mem); // Create new memory state 3577 3578 if (initial_slow_test->is_Bool()) { 3579 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick. 3580 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn); 3581 } 3582 3583 // Create the AllocateArrayNode and its result projections 3584 AllocateArrayNode* alloc 3585 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT), 3586 control(), mem, i_o(), 3587 size, klass_node, 3588 initial_slow_test, 3589 length); 3590 3591 // Cast to correct type. Note that the klass_node may be constant or not, 3592 // and in the latter case the actual array type will be inexact also. 3593 // (This happens via a non-constant argument to inline_native_newArray.) 3594 // In any case, the value of klass_node provides the desired array type. 3595 const TypeInt* length_type = _gvn.find_int_type(length); 3596 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type(); 3597 if (ary_type->isa_aryptr() && length_type != NULL) { 3598 // Try to get a better type than POS for the size 3599 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 3600 } 3601 3602 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception); 3603 3604 // Cast length on remaining path to be as narrow as possible 3605 if (map()->find_edge(length) >= 0) { 3606 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn); 3607 if (ccast != length) { 3608 _gvn.set_type_bottom(ccast); 3609 record_for_igvn(ccast); 3610 replace_in_map(length, ccast); 3611 } 3612 } 3613 3614 return javaoop; 3615 } 3616 3617 // The following "Ideal_foo" functions are placed here because they recognize 3618 // the graph shapes created by the functions immediately above. 3619 3620 //---------------------------Ideal_allocation---------------------------------- 3621 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode. 3622 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) { 3623 if (ptr == NULL) { // reduce dumb test in callers 3624 return NULL; 3625 } 3626 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast 3627 ptr = ptr->in(1); 3628 if (ptr == NULL) return NULL; 3629 } 3630 // Return NULL for allocations with several casts: 3631 // j.l.reflect.Array.newInstance(jobject, jint) 3632 // Object.clone() 3633 // to keep more precise type from last cast. 3634 if (ptr->is_Proj()) { 3635 Node* allo = ptr->in(0); 3636 if (allo != NULL && allo->is_Allocate()) { 3637 return allo->as_Allocate(); 3638 } 3639 } 3640 // Report failure to match. 3641 return NULL; 3642 } 3643 3644 // Fancy version which also strips off an offset (and reports it to caller). 3645 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase, 3646 intptr_t& offset) { 3647 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset); 3648 if (base == NULL) return NULL; 3649 return Ideal_allocation(base, phase); 3650 } 3651 3652 // Trace Initialize <- Proj[Parm] <- Allocate 3653 AllocateNode* InitializeNode::allocation() { 3654 Node* rawoop = in(InitializeNode::RawAddress); 3655 if (rawoop->is_Proj()) { 3656 Node* alloc = rawoop->in(0); 3657 if (alloc->is_Allocate()) { 3658 return alloc->as_Allocate(); 3659 } 3660 } 3661 return NULL; 3662 } 3663 3664 // Trace Allocate -> Proj[Parm] -> Initialize 3665 InitializeNode* AllocateNode::initialization() { 3666 ProjNode* rawoop = proj_out(AllocateNode::RawAddress); 3667 if (rawoop == NULL) return NULL; 3668 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) { 3669 Node* init = rawoop->fast_out(i); 3670 if (init->is_Initialize()) { 3671 assert(init->as_Initialize()->allocation() == this, "2-way link"); 3672 return init->as_Initialize(); 3673 } 3674 } 3675 return NULL; 3676 } 3677 3678 //----------------------------- loop predicates --------------------------- 3679 3680 //------------------------------add_predicate_impl---------------------------- 3681 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) { 3682 // Too many traps seen? 3683 if (too_many_traps(reason)) { 3684 #ifdef ASSERT 3685 if (TraceLoopPredicate) { 3686 int tc = C->trap_count(reason); 3687 tty->print("too many traps=%s tcount=%d in ", 3688 Deoptimization::trap_reason_name(reason), tc); 3689 method()->print(); // which method has too many predicate traps 3690 tty->cr(); 3691 } 3692 #endif 3693 // We cannot afford to take more traps here, 3694 // do not generate predicate. 3695 return; 3696 } 3697 3698 Node *cont = _gvn.intcon(1); 3699 Node* opq = _gvn.transform(new Opaque1Node(C, cont)); 3700 Node *bol = _gvn.transform(new Conv2BNode(opq)); 3701 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 3702 Node* iffalse = _gvn.transform(new IfFalseNode(iff)); 3703 C->add_predicate_opaq(opq); 3704 { 3705 PreserveJVMState pjvms(this); 3706 set_control(iffalse); 3707 inc_sp(nargs); 3708 uncommon_trap(reason, Deoptimization::Action_maybe_recompile); 3709 } 3710 Node* iftrue = _gvn.transform(new IfTrueNode(iff)); 3711 set_control(iftrue); 3712 } 3713 3714 //------------------------------add_predicate--------------------------------- 3715 void GraphKit::add_predicate(int nargs) { 3716 if (UseLoopPredicate) { 3717 add_predicate_impl(Deoptimization::Reason_predicate, nargs); 3718 } 3719 // loop's limit check predicate should be near the loop. 3720 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs); 3721 } 3722 3723 void GraphKit::sync_kit(IdealKit& ideal) { 3724 set_all_memory(ideal.merged_memory()); 3725 set_i_o(ideal.i_o()); 3726 set_control(ideal.ctrl()); 3727 } 3728 3729 void GraphKit::final_sync(IdealKit& ideal) { 3730 // Final sync IdealKit and graphKit. 3731 sync_kit(ideal); 3732 } 3733 3734 Node* GraphKit::load_String_length(Node* ctrl, Node* str) { 3735 Node* len = load_array_length(load_String_value(ctrl, str)); 3736 Node* coder = load_String_coder(ctrl, str); 3737 // Divide length by 2 if coder is UTF16 3738 return _gvn.transform(new RShiftINode(len, coder)); 3739 } 3740 3741 Node* GraphKit::load_String_value(Node* ctrl, Node* str) { 3742 int value_offset = java_lang_String::value_offset_in_bytes(); 3743 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3744 false, NULL, 0); 3745 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3746 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull, 3747 TypeAry::make(TypeInt::BYTE, TypeInt::POS), 3748 ciTypeArrayKlass::make(T_BYTE), true, 0); 3749 Node* p = basic_plus_adr(str, str, value_offset); 3750 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT, C2_ACCESS_ON_HEAP); 3751 // String.value field is known to be @Stable. 3752 if (UseImplicitStableValues) { 3753 load = cast_array_to_stable(load, value_type); 3754 } 3755 return load; 3756 } 3757 3758 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) { 3759 if (!CompactStrings) { 3760 return intcon(java_lang_String::CODER_UTF16); 3761 } 3762 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3763 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3764 false, NULL, 0); 3765 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3766 int coder_field_idx = C->get_alias_index(coder_field_type); 3767 return make_load(ctrl, basic_plus_adr(str, str, coder_offset), 3768 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered); 3769 } 3770 3771 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) { 3772 int value_offset = java_lang_String::value_offset_in_bytes(); 3773 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3774 false, NULL, 0); 3775 const TypePtr* value_field_type = string_type->add_offset(value_offset); 3776 access_store_at(ctrl, str, basic_plus_adr(str, value_offset), value_field_type, 3777 value, TypeAryPtr::BYTES, T_OBJECT, C2_ACCESS_ON_HEAP); 3778 } 3779 3780 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) { 3781 int coder_offset = java_lang_String::coder_offset_in_bytes(); 3782 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(), 3783 false, NULL, 0); 3784 const TypePtr* coder_field_type = string_type->add_offset(coder_offset); 3785 int coder_field_idx = C->get_alias_index(coder_field_type); 3786 store_to_memory(ctrl, basic_plus_adr(str, coder_offset), 3787 value, T_BYTE, coder_field_idx, MemNode::unordered); 3788 } 3789 3790 // Capture src and dst memory state with a MergeMemNode 3791 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) { 3792 if (src_type == dst_type) { 3793 // Types are equal, we don't need a MergeMemNode 3794 return memory(src_type); 3795 } 3796 MergeMemNode* merge = MergeMemNode::make(map()->memory()); 3797 record_for_igvn(merge); // fold it up later, if possible 3798 int src_idx = C->get_alias_index(src_type); 3799 int dst_idx = C->get_alias_index(dst_type); 3800 merge->set_memory_at(src_idx, memory(src_idx)); 3801 merge->set_memory_at(dst_idx, memory(dst_idx)); 3802 return merge; 3803 } 3804 3805 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) { 3806 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported"); 3807 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type"); 3808 // If input and output memory types differ, capture both states to preserve 3809 // the dependency between preceding and subsequent loads/stores. 3810 // For example, the following program: 3811 // StoreB 3812 // compress_string 3813 // LoadB 3814 // has this memory graph (use->def): 3815 // LoadB -> compress_string -> CharMem 3816 // ... -> StoreB -> ByteMem 3817 // The intrinsic hides the dependency between LoadB and StoreB, causing 3818 // the load to read from memory not containing the result of the StoreB. 3819 // The correct memory graph should look like this: 3820 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem)) 3821 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES); 3822 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count); 3823 Node* res_mem = _gvn.transform(new SCMemProjNode(str)); 3824 set_memory(res_mem, TypeAryPtr::BYTES); 3825 return str; 3826 } 3827 3828 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) { 3829 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported"); 3830 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type"); 3831 // Capture src and dst memory (see comment in 'compress_string'). 3832 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type); 3833 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count); 3834 set_memory(_gvn.transform(str), dst_type); 3835 } 3836 3837 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) { 3838 /** 3839 * int i_char = start; 3840 * for (int i_byte = 0; i_byte < count; i_byte++) { 3841 * dst[i_char++] = (char)(src[i_byte] & 0xff); 3842 * } 3843 */ 3844 add_predicate(); 3845 RegionNode* head = new RegionNode(3); 3846 head->init_req(1, control()); 3847 gvn().set_type(head, Type::CONTROL); 3848 record_for_igvn(head); 3849 3850 Node* i_byte = new PhiNode(head, TypeInt::INT); 3851 i_byte->init_req(1, intcon(0)); 3852 gvn().set_type(i_byte, TypeInt::INT); 3853 record_for_igvn(i_byte); 3854 3855 Node* i_char = new PhiNode(head, TypeInt::INT); 3856 i_char->init_req(1, start); 3857 gvn().set_type(i_char, TypeInt::INT); 3858 record_for_igvn(i_char); 3859 3860 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES); 3861 gvn().set_type(mem, Type::MEMORY); 3862 record_for_igvn(mem); 3863 set_control(head); 3864 set_memory(mem, TypeAryPtr::BYTES); 3865 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES); 3866 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE), 3867 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered, 3868 false, false, true /* mismatched */); 3869 3870 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN); 3871 head->init_req(2, IfTrue(iff)); 3872 mem->init_req(2, st); 3873 i_byte->init_req(2, AddI(i_byte, intcon(1))); 3874 i_char->init_req(2, AddI(i_char, intcon(2))); 3875 3876 set_control(IfFalse(iff)); 3877 set_memory(st, TypeAryPtr::BYTES); 3878 } 3879 3880 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) { 3881 if (!field->is_constant()) { 3882 return NULL; // Field not marked as constant. 3883 } 3884 ciInstance* holder = NULL; 3885 if (!field->is_static()) { 3886 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop(); 3887 if (const_oop != NULL && const_oop->is_instance()) { 3888 holder = const_oop->as_instance(); 3889 } 3890 } 3891 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(), 3892 /*is_unsigned_load=*/false); 3893 if (con_type != NULL) { 3894 return makecon(con_type); 3895 } 3896 return NULL; 3897 } 3898 3899 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) { 3900 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity 3901 // assumption of CCP analysis. 3902 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true))); 3903 }