1 /* 2 * Copyright (c) 1997, 2023, 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 "interpreter/linkResolver.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "oops/method.hpp" 30 #include "opto/addnode.hpp" 31 #include "opto/c2compiler.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/idealGraphPrinter.hpp" 34 #include "opto/locknode.hpp" 35 #include "opto/memnode.hpp" 36 #include "opto/opaquenode.hpp" 37 #include "opto/parse.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/type.hpp" 41 #include "runtime/handles.inline.hpp" 42 #include "runtime/safepointMechanism.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 #include "utilities/bitMap.inline.hpp" 45 #include "utilities/copy.hpp" 46 47 // Static array so we can figure out which bytecodes stop us from compiling 48 // the most. Some of the non-static variables are needed in bytecodeInfo.cpp 49 // and eventually should be encapsulated in a proper class (gri 8/18/98). 50 51 #ifndef PRODUCT 52 uint nodes_created = 0; 53 uint methods_parsed = 0; 54 uint methods_seen = 0; 55 uint blocks_parsed = 0; 56 uint blocks_seen = 0; 57 58 uint explicit_null_checks_inserted = 0; 59 uint explicit_null_checks_elided = 0; 60 uint all_null_checks_found = 0; 61 uint implicit_null_checks = 0; 62 63 bool Parse::BytecodeParseHistogram::_initialized = false; 64 uint Parse::BytecodeParseHistogram::_bytecodes_parsed [Bytecodes::number_of_codes]; 65 uint Parse::BytecodeParseHistogram::_nodes_constructed[Bytecodes::number_of_codes]; 66 uint Parse::BytecodeParseHistogram::_nodes_transformed[Bytecodes::number_of_codes]; 67 uint Parse::BytecodeParseHistogram::_new_values [Bytecodes::number_of_codes]; 68 69 //------------------------------print_statistics------------------------------- 70 void Parse::print_statistics() { 71 tty->print_cr("--- Compiler Statistics ---"); 72 tty->print("Methods seen: %u Methods parsed: %u", methods_seen, methods_parsed); 73 tty->print(" Nodes created: %u", nodes_created); 74 tty->cr(); 75 if (methods_seen != methods_parsed) { 76 tty->print_cr("Reasons for parse failures (NOT cumulative):"); 77 } 78 tty->print_cr("Blocks parsed: %u Blocks seen: %u", blocks_parsed, blocks_seen); 79 80 if (explicit_null_checks_inserted) { 81 tty->print_cr("%u original null checks - %u elided (%2u%%); optimizer leaves %u,", 82 explicit_null_checks_inserted, explicit_null_checks_elided, 83 (100*explicit_null_checks_elided)/explicit_null_checks_inserted, 84 all_null_checks_found); 85 } 86 if (all_null_checks_found) { 87 tty->print_cr("%u made implicit (%2u%%)", implicit_null_checks, 88 (100*implicit_null_checks)/all_null_checks_found); 89 } 90 if (SharedRuntime::_implicit_null_throws) { 91 tty->print_cr("%u implicit null exceptions at runtime", 92 SharedRuntime::_implicit_null_throws); 93 } 94 95 if (PrintParseStatistics && BytecodeParseHistogram::initialized()) { 96 BytecodeParseHistogram::print(); 97 } 98 99 if (DoPartialEscapeAnalysis) { 100 printPeaStatistics(); 101 } 102 } 103 #endif 104 105 //------------------------------ON STACK REPLACEMENT--------------------------- 106 107 // Construct a node which can be used to get incoming state for 108 // on stack replacement. 109 Node *Parse::fetch_interpreter_state(int index, 110 BasicType bt, 111 Node *local_addrs, 112 Node *local_addrs_base) { 113 Node *mem = memory(Compile::AliasIdxRaw); 114 Node *adr = basic_plus_adr( local_addrs_base, local_addrs, -index*wordSize ); 115 Node *ctl = control(); 116 117 // Very similar to LoadNode::make, except we handle un-aligned longs and 118 // doubles on Sparc. Intel can handle them just fine directly. 119 Node *l = nullptr; 120 switch (bt) { // Signature is flattened 121 case T_INT: l = new LoadINode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInt::INT, MemNode::unordered); break; 122 case T_FLOAT: l = new LoadFNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::FLOAT, MemNode::unordered); break; 123 case T_ADDRESS: l = new LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); break; 124 case T_OBJECT: l = new LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInstPtr::BOTTOM, MemNode::unordered); break; 125 case T_LONG: 126 case T_DOUBLE: { 127 // Since arguments are in reverse order, the argument address 'adr' 128 // refers to the back half of the long/double. Recompute adr. 129 adr = basic_plus_adr(local_addrs_base, local_addrs, -(index+1)*wordSize); 130 if (Matcher::misaligned_doubles_ok) { 131 l = (bt == T_DOUBLE) 132 ? (Node*)new LoadDNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::DOUBLE, MemNode::unordered) 133 : (Node*)new LoadLNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeLong::LONG, MemNode::unordered); 134 } else { 135 l = (bt == T_DOUBLE) 136 ? (Node*)new LoadD_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered) 137 : (Node*)new LoadL_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered); 138 } 139 break; 140 } 141 default: ShouldNotReachHere(); 142 } 143 return _gvn.transform(l); 144 } 145 146 // Helper routine to prevent the interpreter from handing 147 // unexpected typestate to an OSR method. 148 // The Node l is a value newly dug out of the interpreter frame. 149 // The type is the type predicted by ciTypeFlow. Note that it is 150 // not a general type, but can only come from Type::get_typeflow_type. 151 // The safepoint is a map which will feed an uncommon trap. 152 Node* Parse::check_interpreter_type(Node* l, const Type* type, 153 SafePointNode* &bad_type_exit) { 154 155 const TypeOopPtr* tp = type->isa_oopptr(); 156 157 // TypeFlow may assert null-ness if a type appears unloaded. 158 if (type == TypePtr::NULL_PTR || 159 (tp != nullptr && !tp->is_loaded())) { 160 // Value must be null, not a real oop. 161 Node* chk = _gvn.transform( new CmpPNode(l, null()) ); 162 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 163 IfNode* iff = create_and_map_if(control(), tst, PROB_MAX, COUNT_UNKNOWN); 164 set_control(_gvn.transform( new IfTrueNode(iff) )); 165 Node* bad_type = _gvn.transform( new IfFalseNode(iff) ); 166 bad_type_exit->control()->add_req(bad_type); 167 l = null(); 168 } 169 170 // Typeflow can also cut off paths from the CFG, based on 171 // types which appear unloaded, or call sites which appear unlinked. 172 // When paths are cut off, values at later merge points can rise 173 // toward more specific classes. Make sure these specific classes 174 // are still in effect. 175 if (tp != nullptr && !tp->is_same_java_type_as(TypeInstPtr::BOTTOM)) { 176 // TypeFlow asserted a specific object type. Value must have that type. 177 Node* bad_type_ctrl = nullptr; 178 l = gen_checkcast(l, makecon(tp->as_klass_type()->cast_to_exactness(true)), &bad_type_ctrl); 179 bad_type_exit->control()->add_req(bad_type_ctrl); 180 } 181 182 assert(_gvn.type(l)->higher_equal(type), "must constrain OSR typestate"); 183 return l; 184 } 185 186 // Helper routine which sets up elements of the initial parser map when 187 // performing a parse for on stack replacement. Add values into map. 188 // The only parameter contains the address of a interpreter arguments. 189 void Parse::load_interpreter_state(Node* osr_buf) { 190 int index; 191 int max_locals = jvms()->loc_size(); 192 int max_stack = jvms()->stk_size(); 193 194 195 // Mismatch between method and jvms can occur since map briefly held 196 // an OSR entry state (which takes up one RawPtr word). 197 assert(max_locals == method()->max_locals(), "sanity"); 198 assert(max_stack >= method()->max_stack(), "sanity"); 199 assert((int)jvms()->endoff() == TypeFunc::Parms + max_locals + max_stack, "sanity"); 200 assert((int)jvms()->endoff() == (int)map()->req(), "sanity"); 201 202 // Find the start block. 203 Block* osr_block = start_block(); 204 assert(osr_block->start() == osr_bci(), "sanity"); 205 206 // Set initial BCI. 207 set_parse_bci(osr_block->start()); 208 209 // Set initial stack depth. 210 set_sp(osr_block->start_sp()); 211 212 // Check bailouts. We currently do not perform on stack replacement 213 // of loops in catch blocks or loops which branch with a non-empty stack. 214 if (sp() != 0) { 215 C->record_method_not_compilable("OSR starts with non-empty stack"); 216 return; 217 } 218 // Do not OSR inside finally clauses: 219 if (osr_block->has_trap_at(osr_block->start())) { 220 assert(false, "OSR starts with an immediate trap"); 221 C->record_method_not_compilable("OSR starts with an immediate trap"); 222 return; 223 } 224 225 // Commute monitors from interpreter frame to compiler frame. 226 assert(jvms()->monitor_depth() == 0, "should be no active locks at beginning of osr"); 227 int mcnt = osr_block->flow()->monitor_count(); 228 Node *monitors_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals+mcnt*2-1)*wordSize); 229 for (index = 0; index < mcnt; index++) { 230 // Make a BoxLockNode for the monitor. 231 Node *box = _gvn.transform(new BoxLockNode(next_monitor())); 232 233 234 // Displaced headers and locked objects are interleaved in the 235 // temp OSR buffer. We only copy the locked objects out here. 236 // Fetch the locked object from the OSR temp buffer and copy to our fastlock node. 237 Node *lock_object = fetch_interpreter_state(index*2, T_OBJECT, monitors_addr, osr_buf); 238 // Try and copy the displaced header to the BoxNode 239 Node *displaced_hdr = fetch_interpreter_state((index*2) + 1, T_ADDRESS, monitors_addr, osr_buf); 240 241 242 store_to_memory(control(), box, displaced_hdr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 243 244 // Build a bogus FastLockNode (no code will be generated) and push the 245 // monitor into our debug info. 246 const FastLockNode *flock = _gvn.transform(new FastLockNode( 0, lock_object, box ))->as_FastLock(); 247 map()->push_monitor(flock); 248 249 // If the lock is our method synchronization lock, tuck it away in 250 // _sync_lock for return and rethrow exit paths. 251 if (index == 0 && method()->is_synchronized()) { 252 _synch_lock = flock; 253 } 254 } 255 256 // Use the raw liveness computation to make sure that unexpected 257 // values don't propagate into the OSR frame. 258 MethodLivenessResult live_locals = method()->liveness_at_bci(osr_bci()); 259 if (!live_locals.is_valid()) { 260 // Degenerate or breakpointed method. 261 assert(false, "OSR in empty or breakpointed method"); 262 C->record_method_not_compilable("OSR in empty or breakpointed method"); 263 return; 264 } 265 266 // Extract the needed locals from the interpreter frame. 267 Node *locals_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals-1)*wordSize); 268 269 // find all the locals that the interpreter thinks contain live oops 270 const ResourceBitMap live_oops = method()->live_local_oops_at_bci(osr_bci()); 271 for (index = 0; index < max_locals; index++) { 272 273 if (!live_locals.at(index)) { 274 continue; 275 } 276 277 const Type *type = osr_block->local_type_at(index); 278 279 if (type->isa_oopptr() != nullptr) { 280 281 // 6403625: Verify that the interpreter oopMap thinks that the oop is live 282 // else we might load a stale oop if the MethodLiveness disagrees with the 283 // result of the interpreter. If the interpreter says it is dead we agree 284 // by making the value go to top. 285 // 286 287 if (!live_oops.at(index)) { 288 if (C->log() != nullptr) { 289 C->log()->elem("OSR_mismatch local_index='%d'",index); 290 } 291 set_local(index, null()); 292 // and ignore it for the loads 293 continue; 294 } 295 } 296 297 // Filter out TOP, HALF, and BOTTOM. (Cf. ensure_phi.) 298 if (type == Type::TOP || type == Type::HALF) { 299 continue; 300 } 301 // If the type falls to bottom, then this must be a local that 302 // is mixing ints and oops or some such. Forcing it to top 303 // makes it go dead. 304 if (type == Type::BOTTOM) { 305 continue; 306 } 307 // Construct code to access the appropriate local. 308 BasicType bt = type->basic_type(); 309 if (type == TypePtr::NULL_PTR) { 310 // Ptr types are mixed together with T_ADDRESS but null is 311 // really for T_OBJECT types so correct it. 312 bt = T_OBJECT; 313 } 314 Node *value = fetch_interpreter_state(index, bt, locals_addr, osr_buf); 315 set_local(index, value); 316 } 317 318 // Extract the needed stack entries from the interpreter frame. 319 for (index = 0; index < sp(); index++) { 320 const Type *type = osr_block->stack_type_at(index); 321 if (type != Type::TOP) { 322 // Currently the compiler bails out when attempting to on stack replace 323 // at a bci with a non-empty stack. We should not reach here. 324 ShouldNotReachHere(); 325 } 326 } 327 328 // End the OSR migration 329 make_runtime_call(RC_LEAF, OptoRuntime::osr_end_Type(), 330 CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_end), 331 "OSR_migration_end", TypeRawPtr::BOTTOM, 332 osr_buf); 333 334 // Now that the interpreter state is loaded, make sure it will match 335 // at execution time what the compiler is expecting now: 336 SafePointNode* bad_type_exit = clone_map(); 337 bad_type_exit->set_control(new RegionNode(1)); 338 339 assert(osr_block->flow()->jsrs()->size() == 0, "should be no jsrs live at osr point"); 340 for (index = 0; index < max_locals; index++) { 341 if (stopped()) break; 342 Node* l = local(index); 343 if (l->is_top()) continue; // nothing here 344 const Type *type = osr_block->local_type_at(index); 345 if (type->isa_oopptr() != nullptr) { 346 if (!live_oops.at(index)) { 347 // skip type check for dead oops 348 continue; 349 } 350 } 351 if (osr_block->flow()->local_type_at(index)->is_return_address()) { 352 // In our current system it's illegal for jsr addresses to be 353 // live into an OSR entry point because the compiler performs 354 // inlining of jsrs. ciTypeFlow has a bailout that detect this 355 // case and aborts the compile if addresses are live into an OSR 356 // entry point. Because of that we can assume that any address 357 // locals at the OSR entry point are dead. Method liveness 358 // isn't precise enough to figure out that they are dead in all 359 // cases so simply skip checking address locals all 360 // together. Any type check is guaranteed to fail since the 361 // interpreter type is the result of a load which might have any 362 // value and the expected type is a constant. 363 continue; 364 } 365 set_local(index, check_interpreter_type(l, type, bad_type_exit)); 366 } 367 368 for (index = 0; index < sp(); index++) { 369 if (stopped()) break; 370 Node* l = stack(index); 371 if (l->is_top()) continue; // nothing here 372 const Type *type = osr_block->stack_type_at(index); 373 set_stack(index, check_interpreter_type(l, type, bad_type_exit)); 374 } 375 376 if (bad_type_exit->control()->req() > 1) { 377 // Build an uncommon trap here, if any inputs can be unexpected. 378 bad_type_exit->set_control(_gvn.transform( bad_type_exit->control() )); 379 record_for_igvn(bad_type_exit->control()); 380 SafePointNode* types_are_good = map(); 381 set_map(bad_type_exit); 382 // The unexpected type happens because a new edge is active 383 // in the CFG, which typeflow had previously ignored. 384 // E.g., Object x = coldAtFirst() && notReached()? "str": new Integer(123). 385 // This x will be typed as Integer if notReached is not yet linked. 386 // It could also happen due to a problem in ciTypeFlow analysis. 387 uncommon_trap(Deoptimization::Reason_constraint, 388 Deoptimization::Action_reinterpret); 389 set_map(types_are_good); 390 } 391 } 392 393 //------------------------------Parse------------------------------------------ 394 // Main parser constructor. 395 Parse::Parse(JVMState* caller, ciMethod* parse_method, float expected_uses, PEAState* caller_state) 396 : _exits(caller) 397 { 398 // Init some variables 399 _caller = caller; 400 _caller_state = caller_state; 401 _method = parse_method; 402 _expected_uses = expected_uses; 403 _depth = 1 + (caller->has_method() ? caller->depth() : 0); 404 _wrote_final = false; 405 _wrote_volatile = false; 406 _wrote_stable = false; 407 _wrote_fields = false; 408 _alloc_with_final = nullptr; 409 _block = nullptr; 410 _first_return = 0; 411 _replaced_nodes_for_exceptions = false; 412 _new_idx = C->unique(); 413 DEBUG_ONLY(_entry_bci = UnknownBci); 414 DEBUG_ONLY(_block_count = -1); 415 DEBUG_ONLY(_blocks = (Block*)-1); 416 #ifndef PRODUCT 417 if (PrintCompilation || PrintOpto) { 418 // Make sure I have an inline tree, so I can print messages about it. 419 InlineTree::find_subtree_from_root(C->ilt(), caller, parse_method); 420 } 421 _max_switch_depth = 0; 422 _est_switch_depth = 0; 423 424 if (TraceOptoParse) { 425 tty->print_raw("Parsing method "); 426 parse_method->print_name(tty); 427 tty->print_cr(" {"); 428 } 429 #endif 430 431 if (parse_method->has_reserved_stack_access()) { 432 C->set_has_reserved_stack_access(true); 433 } 434 435 if (parse_method->is_synchronized() || parse_method->has_monitor_bytecodes()) { 436 C->set_has_monitors(true); 437 } 438 439 _iter.reset_to_method(method()); 440 C->set_has_loops(C->has_loops() || method()->has_loops()); 441 442 if (_expected_uses <= 0) { 443 _prof_factor = 1; 444 } else { 445 float prof_total = parse_method->interpreter_invocation_count(); 446 if (prof_total <= _expected_uses) { 447 _prof_factor = 1; 448 } else { 449 _prof_factor = _expected_uses / prof_total; 450 } 451 } 452 453 CompileLog* log = C->log(); 454 if (log != nullptr) { 455 log->begin_head("parse method='%d' uses='%f'", 456 log->identify(parse_method), expected_uses); 457 if (depth() == 1 && C->is_osr_compilation()) { 458 log->print(" osr_bci='%d'", C->entry_bci()); 459 } 460 log->stamp(); 461 log->end_head(); 462 } 463 464 // Accumulate deoptimization counts. 465 // (The range_check and store_check counts are checked elsewhere.) 466 ciMethodData* md = method()->method_data(); 467 for (uint reason = 0; reason < md->trap_reason_limit(); reason++) { 468 uint md_count = md->trap_count(reason); 469 if (md_count != 0) { 470 if (md_count >= md->trap_count_limit()) { 471 md_count = md->trap_count_limit() + md->overflow_trap_count(); 472 } 473 uint total_count = C->trap_count(reason); 474 uint old_count = total_count; 475 total_count += md_count; 476 // Saturate the add if it overflows. 477 if (total_count < old_count || total_count < md_count) 478 total_count = (uint)-1; 479 C->set_trap_count(reason, total_count); 480 if (log != nullptr) 481 log->elem("observe trap='%s' count='%d' total='%d'", 482 Deoptimization::trap_reason_name(reason), 483 md_count, total_count); 484 } 485 } 486 // Accumulate total sum of decompilations, also. 487 C->set_decompile_count(C->decompile_count() + md->decompile_count()); 488 489 if (log != nullptr && method()->has_exception_handlers()) { 490 log->elem("observe that='has_exception_handlers'"); 491 } 492 493 assert(InlineTree::check_can_parse(method()) == nullptr, "Can not parse this method, cutout earlier"); 494 assert(method()->has_balanced_monitors(), "Can not parse unbalanced monitors, cutout earlier"); 495 496 // Always register dependence if JVMTI is enabled, because 497 // either breakpoint setting or hotswapping of methods may 498 // cause deoptimization. 499 if (C->env()->jvmti_can_hotswap_or_post_breakpoint()) { 500 C->dependencies()->assert_evol_method(method()); 501 } 502 503 NOT_PRODUCT(methods_seen++); 504 505 // Do some special top-level things. 506 if (depth() == 1 && C->is_osr_compilation()) { 507 _tf = C->tf(); // the OSR entry type is different 508 _entry_bci = C->entry_bci(); 509 _flow = method()->get_osr_flow_analysis(osr_bci()); 510 } else { 511 _tf = TypeFunc::make(method()); 512 _entry_bci = InvocationEntryBci; 513 _flow = method()->get_flow_analysis(); 514 } 515 516 if (_flow->failing()) { 517 assert(false, "type flow analysis failed during parsing"); 518 C->record_method_not_compilable(_flow->failure_reason()); 519 #ifndef PRODUCT 520 if (PrintOpto && (Verbose || WizardMode)) { 521 if (is_osr_parse()) { 522 tty->print_cr("OSR @%d type flow bailout: %s", _entry_bci, _flow->failure_reason()); 523 } else { 524 tty->print_cr("type flow bailout: %s", _flow->failure_reason()); 525 } 526 if (Verbose) { 527 method()->print(); 528 method()->print_codes(); 529 _flow->print(); 530 } 531 } 532 #endif 533 } 534 535 #ifdef ASSERT 536 if (depth() == 1) { 537 assert(C->is_osr_compilation() == this->is_osr_parse(), "OSR in sync"); 538 } else { 539 assert(!this->is_osr_parse(), "no recursive OSR"); 540 } 541 #endif 542 543 #ifndef PRODUCT 544 // Dump CFG in RPO order before Parsing. 545 if (Verbose && !CITraceTypeFlow) { 546 _flow->rpo_print_on(tty); 547 } 548 549 if (_flow->has_irreducible_entry()) { 550 C->set_parsed_irreducible_loop(true); 551 } 552 553 methods_parsed++; 554 // add method size here to guarantee that inlined methods are added too 555 if (CITime) 556 _total_bytes_compiled += method()->code_size(); 557 558 show_parse_info(); 559 #endif 560 561 if (failing()) { 562 if (log) log->done("parse"); 563 return; 564 } 565 566 gvn().transform(top()); 567 568 // Import the results of the ciTypeFlow. 569 init_blocks(); 570 571 // Merge point for all normal exits 572 build_exits(); 573 574 // Setup the initial JVM state map. 575 SafePointNode* entry_map = create_entry_map(); 576 577 // Check for bailouts during map initialization 578 if (failing() || entry_map == nullptr) { 579 if (log) log->done("parse"); 580 return; 581 } 582 583 Node_Notes* caller_nn = C->default_node_notes(); 584 // Collect debug info for inlined calls unless -XX:-DebugInlinedCalls. 585 if (DebugInlinedCalls || depth() == 1) { 586 C->set_default_node_notes(make_node_notes(caller_nn)); 587 } 588 589 if (is_osr_parse()) { 590 Node* osr_buf = entry_map->in(TypeFunc::Parms+0); 591 entry_map->set_req(TypeFunc::Parms+0, top()); 592 set_map(entry_map); 593 load_interpreter_state(osr_buf); 594 } else { 595 set_map(entry_map); 596 do_method_entry(); 597 } 598 599 if (depth() == 1 && !failing()) { 600 if (C->clinit_barrier_on_entry()) { 601 // Add check to deoptimize the nmethod once the holder class is fully initialized 602 clinit_deopt(); 603 } 604 605 // Add check to deoptimize the nmethod if RTM state was changed 606 rtm_deopt(); 607 } 608 609 // Check for bailouts during method entry or RTM state check setup. 610 if (failing()) { 611 if (log) log->done("parse"); 612 C->set_default_node_notes(caller_nn); 613 return; 614 } 615 616 entry_map = map(); // capture any changes performed by method setup code 617 assert(jvms()->endoff() == map()->req(), "map matches JVMS layout"); 618 619 // We begin parsing as if we have just encountered a jump to the 620 // method entry. 621 Block* entry_block = start_block(); 622 assert(entry_block->start() == (is_osr_parse() ? osr_bci() : 0), ""); 623 set_map_clone(entry_map); 624 625 merge_common(entry_block, entry_block->next_path_num()); 626 #ifndef PRODUCT 627 BytecodeParseHistogram *parse_histogram_obj = new (C->env()->arena()) BytecodeParseHistogram(this, C); 628 set_parse_histogram( parse_histogram_obj ); 629 #endif 630 631 // Parse all the basic blocks. 632 do_all_blocks(); 633 634 // Check for bailouts during conversion to graph 635 if (failing()) { 636 if (log) log->done("parse"); 637 return; 638 } 639 640 // Fix up all exiting control flow. 641 set_map(entry_map); 642 do_exits(); 643 644 // Only reset this now, to make sure that debug information emitted 645 // for exiting control flow still refers to the inlined method. 646 C->set_default_node_notes(caller_nn); 647 648 if (log) log->done("parse nodes='%d' live='%d' memory='" SIZE_FORMAT "'", 649 C->unique(), C->live_nodes(), C->node_arena()->used()); 650 } 651 652 #ifndef PRODUCT 653 Parse::~Parse() { 654 if (TraceOptoParse) { 655 tty->print("} // "); 656 method()->print_short_name(tty); 657 tty->cr(); 658 } 659 660 if (DoPartialEscapeAnalysis && PEAVerbose) { 661 PEAState& as = _exits.jvms()->alloc_state(); 662 auto objs = PEA()->all_objects(); 663 for (int i = 0; i < objs.length(); ++i) { 664 ObjID obj = objs.at(i); 665 666 if (as.contains(obj)) { 667 ObjectState* os = as.get_object_state(obj); 668 tty->print("%4d | Obj%d\t", i, obj->_idx); 669 670 if (os->is_virtual()) { 671 tty->print_cr("V"); 672 } else { 673 tty->print_cr("M"); 674 } 675 } 676 } 677 } 678 } 679 #endif 680 //---------------------------do_all_blocks------------------------------------- 681 void Parse::do_all_blocks() { 682 bool has_irreducible = flow()->has_irreducible_entry(); 683 684 // Walk over all blocks in Reverse Post-Order. 685 while (true) { 686 bool progress = false; 687 for (int rpo = 0; rpo < block_count(); rpo++) { 688 Block* block = rpo_at(rpo); 689 690 if (block->is_parsed()) continue; 691 692 if (!block->is_merged()) { 693 // Dead block, no state reaches this block 694 continue; 695 } 696 697 // Prepare to parse this block. 698 load_state_from(block); 699 700 if (stopped()) { 701 // Block is dead. 702 continue; 703 } 704 705 NOT_PRODUCT(blocks_parsed++); 706 707 progress = true; 708 if (block->is_loop_head() || block->is_handler() || (has_irreducible && !block->is_ready())) { 709 // mark live objects 'Escaped' in map before mounting phi nodes. 710 if (DoPartialEscapeAnalysis && block->is_loop_head()) { 711 PEAState& as = jvms()->alloc_state(); 712 as.mark_all_live_objects_escaped(PEA(), map()); 713 } 714 // Not all preds have been parsed. We must build phis everywhere. 715 // (Note that dead locals do not get phis built, ever.) 716 ensure_phis_everywhere(); 717 718 if (block->is_SEL_head()) { 719 // Add predicate to single entry (not irreducible) loop head. 720 assert(!block->has_merged_backedge(), "only entry paths should be merged for now"); 721 // Predicates may have been added after a dominating if 722 if (!block->has_predicates()) { 723 // Need correct bci for predicate. 724 // It is fine to set it here since do_one_block() will set it anyway. 725 set_parse_bci(block->start()); 726 add_parse_predicates(); 727 } 728 // Add new region for back branches. 729 int edges = block->pred_count() - block->preds_parsed() + 1; // +1 for original region 730 RegionNode *r = new RegionNode(edges+1); 731 _gvn.set_type(r, Type::CONTROL); 732 record_for_igvn(r); 733 r->init_req(edges, control()); 734 set_control(r); 735 block->copy_irreducible_status_to(r, jvms()); 736 // Add new phis. 737 ensure_phis_everywhere(); 738 } 739 740 // Leave behind an undisturbed copy of the map, for future merges. 741 set_map(clone_map()); 742 } 743 744 if (control()->is_Region() && !block->is_loop_head() && !has_irreducible && !block->is_handler()) { 745 // In the absence of irreducible loops, the Region and Phis 746 // associated with a merge that doesn't involve a backedge can 747 // be simplified now since the RPO parsing order guarantees 748 // that any path which was supposed to reach here has already 749 // been parsed or must be dead. 750 Node* c = control(); 751 Node* result = _gvn.transform(control()); 752 if (c != result && TraceOptoParse) { 753 tty->print_cr("Block #%d replace %d with %d", block->rpo(), c->_idx, result->_idx); 754 } 755 if (result != top()) { 756 record_for_igvn(result); 757 } 758 } 759 760 // Parse the block. 761 do_one_block(); 762 763 // Check for bailouts. 764 if (failing()) return; 765 } 766 767 // with irreducible loops multiple passes might be necessary to parse everything 768 if (!has_irreducible || !progress) { 769 break; 770 } 771 } 772 773 #ifndef PRODUCT 774 blocks_seen += block_count(); 775 776 // Make sure there are no half-processed blocks remaining. 777 // Every remaining unprocessed block is dead and may be ignored now. 778 for (int rpo = 0; rpo < block_count(); rpo++) { 779 Block* block = rpo_at(rpo); 780 if (!block->is_parsed()) { 781 if (TraceOptoParse) { 782 tty->print_cr("Skipped dead block %d at bci:%d", rpo, block->start()); 783 } 784 assert(!block->is_merged(), "no half-processed blocks"); 785 } 786 } 787 #endif 788 } 789 790 static Node* mask_int_value(Node* v, BasicType bt, PhaseGVN* gvn) { 791 switch (bt) { 792 case T_BYTE: 793 v = gvn->transform(new LShiftINode(v, gvn->intcon(24))); 794 v = gvn->transform(new RShiftINode(v, gvn->intcon(24))); 795 break; 796 case T_SHORT: 797 v = gvn->transform(new LShiftINode(v, gvn->intcon(16))); 798 v = gvn->transform(new RShiftINode(v, gvn->intcon(16))); 799 break; 800 case T_CHAR: 801 v = gvn->transform(new AndINode(v, gvn->intcon(0xFFFF))); 802 break; 803 case T_BOOLEAN: 804 v = gvn->transform(new AndINode(v, gvn->intcon(0x1))); 805 break; 806 default: 807 break; 808 } 809 return v; 810 } 811 812 //-------------------------------build_exits---------------------------------- 813 // Build normal and exceptional exit merge points. 814 void Parse::build_exits() { 815 // make a clone of caller to prevent sharing of side-effects 816 _exits.set_map(_exits.clone_map()); 817 _exits.clean_stack(_exits.sp()); 818 _exits.sync_jvms(); 819 820 RegionNode* region = new RegionNode(1); 821 record_for_igvn(region); 822 gvn().set_type_bottom(region); 823 _exits.set_control(region); 824 825 // Note: iophi and memphi are not transformed until do_exits. 826 Node* iophi = new PhiNode(region, Type::ABIO); 827 Node* memphi = new PhiNode(region, Type::MEMORY, TypePtr::BOTTOM); 828 gvn().set_type_bottom(iophi); 829 gvn().set_type_bottom(memphi); 830 _exits.set_i_o(iophi); 831 _exits.set_all_memory(memphi); 832 833 // Add a return value to the exit state. (Do not push it yet.) 834 if (tf()->range()->cnt() > TypeFunc::Parms) { 835 const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms); 836 if (ret_type->isa_int()) { 837 BasicType ret_bt = method()->return_type()->basic_type(); 838 if (ret_bt == T_BOOLEAN || 839 ret_bt == T_CHAR || 840 ret_bt == T_BYTE || 841 ret_bt == T_SHORT) { 842 ret_type = TypeInt::INT; 843 } 844 } 845 846 // Don't "bind" an unloaded return klass to the ret_phi. If the klass 847 // becomes loaded during the subsequent parsing, the loaded and unloaded 848 // types will not join when we transform and push in do_exits(). 849 const TypeOopPtr* ret_oop_type = ret_type->isa_oopptr(); 850 if (ret_oop_type && !ret_oop_type->is_loaded()) { 851 ret_type = TypeOopPtr::BOTTOM; 852 } 853 int ret_size = type2size[ret_type->basic_type()]; 854 Node* ret_phi = new PhiNode(region, ret_type); 855 gvn().set_type_bottom(ret_phi); 856 _exits.ensure_stack(ret_size); 857 assert((int)(tf()->range()->cnt() - TypeFunc::Parms) == ret_size, "good tf range"); 858 assert(method()->return_type()->size() == ret_size, "tf agrees w/ method"); 859 _exits.set_argument(0, ret_phi); // here is where the parser finds it 860 // Note: ret_phi is not yet pushed, until do_exits. 861 } 862 } 863 864 865 //----------------------------build_start_state------------------------------- 866 // Construct a state which contains only the incoming arguments from an 867 // unknown caller. The method & bci will be null & InvocationEntryBci. 868 JVMState* Compile::build_start_state(StartNode* start, const TypeFunc* tf) { 869 int arg_size = tf->domain()->cnt(); 870 int max_size = MAX2(arg_size, (int)tf->range()->cnt()); 871 JVMState* jvms = new (this) JVMState(max_size - TypeFunc::Parms); 872 SafePointNode* map = new SafePointNode(max_size, jvms); 873 record_for_igvn(map); 874 assert(arg_size == TypeFunc::Parms + (is_osr_compilation() ? 1 : method()->arg_size()), "correct arg_size"); 875 Node_Notes* old_nn = default_node_notes(); 876 if (old_nn != nullptr && has_method()) { 877 Node_Notes* entry_nn = old_nn->clone(this); 878 JVMState* entry_jvms = new(this) JVMState(method(), old_nn->jvms()); 879 entry_jvms->set_offsets(0); 880 entry_jvms->set_bci(entry_bci()); 881 entry_nn->set_jvms(entry_jvms); 882 set_default_node_notes(entry_nn); 883 } 884 uint i; 885 for (i = 0; i < (uint)arg_size; i++) { 886 Node* parm = initial_gvn()->transform(new ParmNode(start, i)); 887 map->init_req(i, parm); 888 // Record all these guys for later GVN. 889 record_for_igvn(parm); 890 } 891 for (; i < map->req(); i++) { 892 map->init_req(i, top()); 893 } 894 assert(jvms->argoff() == TypeFunc::Parms, "parser gets arguments here"); 895 set_default_node_notes(old_nn); 896 jvms->set_map(map); 897 return jvms; 898 } 899 900 //-----------------------------make_node_notes--------------------------------- 901 Node_Notes* Parse::make_node_notes(Node_Notes* caller_nn) { 902 if (caller_nn == nullptr) return nullptr; 903 Node_Notes* nn = caller_nn->clone(C); 904 JVMState* caller_jvms = nn->jvms(); 905 JVMState* jvms = new (C) JVMState(method(), caller_jvms); 906 jvms->set_offsets(0); 907 jvms->set_bci(_entry_bci); 908 nn->set_jvms(jvms); 909 return nn; 910 } 911 912 913 //--------------------------return_values-------------------------------------- 914 void Compile::return_values(JVMState* jvms) { 915 GraphKit kit(jvms); 916 Node* ret = new ReturnNode(TypeFunc::Parms, 917 kit.control(), 918 kit.i_o(), 919 kit.reset_memory(), 920 kit.frameptr(), 921 kit.returnadr()); 922 // Add zero or 1 return values 923 int ret_size = tf()->range()->cnt() - TypeFunc::Parms; 924 if (ret_size > 0) { 925 kit.inc_sp(-ret_size); // pop the return value(s) 926 kit.sync_jvms(); 927 ret->add_req(kit.argument(0)); 928 // Note: The second dummy edge is not needed by a ReturnNode. 929 } 930 // bind it to root 931 root()->add_req(ret); 932 record_for_igvn(ret); 933 initial_gvn()->transform(ret); 934 } 935 936 //------------------------rethrow_exceptions----------------------------------- 937 // Bind all exception states in the list into a single RethrowNode. 938 void Compile::rethrow_exceptions(JVMState* jvms) { 939 GraphKit kit(jvms); 940 if (!kit.has_exceptions()) return; // nothing to generate 941 // Load my combined exception state into the kit, with all phis transformed: 942 SafePointNode* ex_map = kit.combine_and_pop_all_exception_states(); 943 Node* ex_oop = kit.use_exception_state(ex_map); 944 RethrowNode* exit = new RethrowNode(kit.control(), 945 kit.i_o(), kit.reset_memory(), 946 kit.frameptr(), kit.returnadr(), 947 // like a return but with exception input 948 ex_oop); 949 // bind to root 950 root()->add_req(exit); 951 record_for_igvn(exit); 952 initial_gvn()->transform(exit); 953 } 954 955 //---------------------------do_exceptions------------------------------------- 956 // Process exceptions arising from the current bytecode. 957 // Send caught exceptions to the proper handler within this method. 958 // Unhandled exceptions feed into _exit. 959 void Parse::do_exceptions() { 960 if (!has_exceptions()) return; 961 962 if (failing()) { 963 // Pop them all off and throw them away. 964 while (pop_exception_state() != nullptr) ; 965 return; 966 } 967 968 PreserveJVMState pjvms(this, false); 969 970 SafePointNode* ex_map; 971 while ((ex_map = pop_exception_state()) != nullptr) { 972 if (!method()->has_exception_handlers()) { 973 // Common case: Transfer control outward. 974 // Doing it this early allows the exceptions to common up 975 // even between adjacent method calls. 976 throw_to_exit(ex_map); 977 } else { 978 // Have to look at the exception first. 979 assert(stopped(), "catch_inline_exceptions trashes the map"); 980 catch_inline_exceptions(ex_map); 981 stop_and_kill_map(); // we used up this exception state; kill it 982 } 983 } 984 985 // We now return to our regularly scheduled program: 986 } 987 988 //---------------------------throw_to_exit------------------------------------- 989 // Merge the given map into an exception exit from this method. 990 // The exception exit will handle any unlocking of receiver. 991 // The ex_oop must be saved within the ex_map, unlike merge_exception. 992 void Parse::throw_to_exit(SafePointNode* ex_map) { 993 // Pop the JVMS to (a copy of) the caller. 994 GraphKit caller; 995 caller.set_map_clone(_caller->map()); 996 caller.set_bci(_caller->bci()); 997 caller.set_sp(_caller->sp()); 998 // Copy out the standard machine state: 999 for (uint i = 0; i < TypeFunc::Parms; i++) { 1000 caller.map()->set_req(i, ex_map->in(i)); 1001 } 1002 if (ex_map->has_replaced_nodes()) { 1003 _replaced_nodes_for_exceptions = true; 1004 } 1005 caller.map()->transfer_replaced_nodes_from(ex_map, _new_idx); 1006 // ...and the exception: 1007 Node* ex_oop = saved_ex_oop(ex_map); 1008 SafePointNode* caller_ex_map = caller.make_exception_state(ex_oop); 1009 // Finally, collect the new exception state in my exits: 1010 _exits.add_exception_state(caller_ex_map); 1011 } 1012 1013 //------------------------------do_exits--------------------------------------- 1014 void Parse::do_exits() { 1015 set_parse_bci(InvocationEntryBci); 1016 1017 // Now peephole on the return bits 1018 Node* region = _exits.control(); 1019 _exits.set_control(gvn().transform(region)); 1020 1021 Node* iophi = _exits.i_o(); 1022 _exits.set_i_o(gvn().transform(iophi)); 1023 1024 // Figure out if we need to emit the trailing barrier. The barrier is only 1025 // needed in the constructors, and only in three cases: 1026 // 1027 // 1. The constructor wrote a final. The effects of all initializations 1028 // must be committed to memory before any code after the constructor 1029 // publishes the reference to the newly constructed object. Rather 1030 // than wait for the publication, we simply block the writes here. 1031 // Rather than put a barrier on only those writes which are required 1032 // to complete, we force all writes to complete. 1033 // 1034 // 2. Experimental VM option is used to force the barrier if any field 1035 // was written out in the constructor. 1036 // 1037 // 3. On processors which are not CPU_MULTI_COPY_ATOMIC (e.g. PPC64), 1038 // support_IRIW_for_not_multiple_copy_atomic_cpu selects that 1039 // MemBarVolatile is used before volatile load instead of after volatile 1040 // store, so there's no barrier after the store. 1041 // We want to guarantee the same behavior as on platforms with total store 1042 // order, although this is not required by the Java memory model. 1043 // In this case, we want to enforce visibility of volatile field 1044 // initializations which are performed in constructors. 1045 // So as with finals, we add a barrier here. 1046 // 1047 // "All bets are off" unless the first publication occurs after a 1048 // normal return from the constructor. We do not attempt to detect 1049 // such unusual early publications. But no barrier is needed on 1050 // exceptional returns, since they cannot publish normally. 1051 // 1052 if (method()->is_initializer() && 1053 (wrote_final() || 1054 (AlwaysSafeConstructors && wrote_fields()) || 1055 (support_IRIW_for_not_multiple_copy_atomic_cpu && wrote_volatile()))) { 1056 if (!DoPartialEscapeAnalysis) { 1057 // If Memory barrier is created for final fields write 1058 // and allocation node does not escape the initialize method, 1059 // then barrier introduced by allocation node can be removed. 1060 if (DoEscapeAnalysis && alloc_with_final()) { 1061 AllocateNode *alloc = AllocateNode::Ideal_allocation(alloc_with_final()); 1062 alloc->compute_MemBar_redundancy(method()); 1063 } 1064 } else { 1065 // in PEA, alloc_with_final stores ObjID 1066 AllocateNode* alloc = (ObjID)alloc_with_final(); 1067 1068 if (DoEscapeAnalysis && alloc != nullptr) { 1069 Node* obj = _exits.jvms()->alloc_state().get_java_oop(alloc); 1070 _exits.insert_mem_bar(Op_MemBarRelease, obj); 1071 alloc->compute_MemBar_redundancy(method()); 1072 } 1073 } 1074 1075 if (PrintOpto && (Verbose || WizardMode)) { 1076 method()->print_name(); 1077 tty->print_cr(" writes finals and needs a memory barrier"); 1078 } 1079 } 1080 1081 // Any method can write a @Stable field; insert memory barriers 1082 // after those also. Can't bind predecessor allocation node (if any) 1083 // with barrier because allocation doesn't always dominate 1084 // MemBarRelease. 1085 if (wrote_stable()) { 1086 _exits.insert_mem_bar(Op_MemBarRelease); 1087 if (PrintOpto && (Verbose || WizardMode)) { 1088 method()->print_name(); 1089 tty->print_cr(" writes @Stable and needs a memory barrier"); 1090 } 1091 } 1092 1093 for (MergeMemStream mms(_exits.merged_memory()); mms.next_non_empty(); ) { 1094 // transform each slice of the original memphi: 1095 mms.set_memory(_gvn.transform(mms.memory())); 1096 } 1097 // Clean up input MergeMems created by transforming the slices 1098 _gvn.transform(_exits.merged_memory()); 1099 1100 if (tf()->range()->cnt() > TypeFunc::Parms) { 1101 const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms); 1102 Node* const ret_phi_old = _exits.argument(0); 1103 Node* ret_phi = _gvn.transform(ret_phi_old); 1104 if (DoPartialEscapeAnalysis && ret_phi_old != ret_phi) { 1105 PEAState& as = _exits.jvms()->alloc_state(); 1106 EscapedState* es = as.as_escaped(PEA(), ret_phi_old); 1107 1108 if (es != nullptr && es->merged_value() == ret_phi_old) { 1109 es->update(ret_phi); 1110 ObjID obj = PEA()->is_alias(ret_phi_old); 1111 PEA()->add_alias(obj, ret_phi); 1112 } 1113 } 1114 if (!_exits.control()->is_top() && _gvn.type(ret_phi)->empty()) { 1115 // If the type we set for the ret_phi in build_exits() is too optimistic and 1116 // the ret_phi is top now, there's an extremely small chance that it may be due to class 1117 // loading. It could also be due to an error, so mark this method as not compilable because 1118 // otherwise this could lead to an infinite compile loop. 1119 // In any case, this code path is rarely (and never in my testing) reached. 1120 #ifdef ASSERT 1121 tty->print_cr("# Can't determine return type."); 1122 tty->print_cr("# exit control"); 1123 _exits.control()->dump(2); 1124 tty->print_cr("# ret phi type"); 1125 _gvn.type(ret_phi)->dump(); 1126 tty->print_cr("# ret phi"); 1127 ret_phi->dump(2); 1128 #endif // ASSERT 1129 assert(false, "Can't determine return type."); 1130 C->record_method_not_compilable("Can't determine return type."); 1131 return; 1132 } 1133 if (ret_type->isa_int()) { 1134 BasicType ret_bt = method()->return_type()->basic_type(); 1135 ret_phi = mask_int_value(ret_phi, ret_bt, &_gvn); 1136 } 1137 _exits.push_node(ret_type->basic_type(), ret_phi); 1138 } 1139 1140 // Note: Logic for creating and optimizing the ReturnNode is in Compile. 1141 1142 // Unlock along the exceptional paths. 1143 // This is done late so that we can common up equivalent exceptions 1144 // (e.g., null checks) arising from multiple points within this method. 1145 // See GraphKit::add_exception_state, which performs the commoning. 1146 bool do_synch = method()->is_synchronized() && GenerateSynchronizationCode; 1147 1148 // record exit from a method if compiled while Dtrace is turned on. 1149 if (do_synch || C->env()->dtrace_method_probes() || _replaced_nodes_for_exceptions) { 1150 // First move the exception list out of _exits: 1151 GraphKit kit(_exits.transfer_exceptions_into_jvms()); 1152 SafePointNode* normal_map = kit.map(); // keep this guy safe 1153 // Now re-collect the exceptions into _exits: 1154 SafePointNode* ex_map; 1155 while ((ex_map = kit.pop_exception_state()) != nullptr) { 1156 Node* ex_oop = kit.use_exception_state(ex_map); 1157 // Force the exiting JVM state to have this method at InvocationEntryBci. 1158 // The exiting JVM state is otherwise a copy of the calling JVMS. 1159 JVMState* caller = kit.jvms(); 1160 JVMState* ex_jvms = caller->clone_shallow(C); 1161 ex_jvms->bind_map(kit.clone_map()); 1162 ex_jvms->set_bci( InvocationEntryBci); 1163 kit.set_jvms(ex_jvms); 1164 if (do_synch) { 1165 // Add on the synchronized-method box/object combo 1166 kit.map()->push_monitor(_synch_lock); 1167 // Unlock! 1168 kit.shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node()); 1169 } 1170 if (C->env()->dtrace_method_probes()) { 1171 kit.make_dtrace_method_exit(method()); 1172 } 1173 if (_replaced_nodes_for_exceptions) { 1174 kit.map()->apply_replaced_nodes(_new_idx); 1175 } 1176 // Done with exception-path processing. 1177 ex_map = kit.make_exception_state(ex_oop); 1178 assert(ex_jvms->same_calls_as(ex_map->jvms()), "sanity"); 1179 // Pop the last vestige of this method: 1180 caller->clone_shallow(C)->bind_map(ex_map); 1181 _exits.push_exception_state(ex_map); 1182 } 1183 assert(_exits.map() == normal_map, "keep the same return state"); 1184 } 1185 1186 { 1187 // Capture very early exceptions (receiver null checks) from caller JVMS 1188 GraphKit caller(_caller); 1189 SafePointNode* ex_map; 1190 while ((ex_map = caller.pop_exception_state()) != nullptr) { 1191 _exits.add_exception_state(ex_map); 1192 } 1193 } 1194 1195 _exits.map()->apply_replaced_nodes(_new_idx); 1196 // don't trust replace list. return_current() may mess it up. 1197 // use AllocationState to update it. 1198 if (DoPartialEscapeAnalysis) { 1199 PEAState& as = _exits.jvms()->alloc_state(); 1200 SafePointNode* map = _exits.map(); 1201 backfill_materialized(map, TypeFunc::Parms, map->req(), as); 1202 } 1203 } 1204 1205 //-----------------------------create_entry_map------------------------------- 1206 // Initialize our parser map to contain the types at method entry. 1207 // For OSR, the map contains a single RawPtr parameter. 1208 // Initial monitor locking for sync. methods is performed by do_method_entry. 1209 SafePointNode* Parse::create_entry_map() { 1210 // Check for really stupid bail-out cases. 1211 uint len = TypeFunc::Parms + method()->max_locals() + method()->max_stack(); 1212 if (len >= 32760) { 1213 // Bailout expected, this is a very rare edge case. 1214 C->record_method_not_compilable("too many local variables"); 1215 return nullptr; 1216 } 1217 1218 // clear current replaced nodes that are of no use from here on (map was cloned in build_exits). 1219 _caller->map()->delete_replaced_nodes(); 1220 1221 // If this is an inlined method, we may have to do a receiver null check. 1222 if (_caller->has_method() && is_normal_parse() && !method()->is_static()) { 1223 GraphKit kit(_caller); 1224 kit.null_check_receiver_before_call(method()); 1225 _caller = kit.transfer_exceptions_into_jvms(); 1226 if (kit.stopped()) { 1227 _exits.add_exception_states_from(_caller); 1228 _exits.set_jvms(_caller); 1229 return nullptr; 1230 } 1231 } 1232 1233 assert(method() != nullptr, "parser must have a method"); 1234 1235 // Create an initial safepoint to hold JVM state during parsing 1236 JVMState* jvms = new (C) JVMState(method(), _caller->has_method() ? _caller : nullptr); 1237 if (_caller != nullptr && DoPartialEscapeAnalysis) { 1238 jvms->alloc_state() = _caller->alloc_state(); 1239 } 1240 1241 set_map(new SafePointNode(len, jvms)); 1242 jvms->set_map(map()); 1243 record_for_igvn(map()); 1244 assert(jvms->endoff() == len, "correct jvms sizing"); 1245 1246 SafePointNode* inmap = _caller->map(); 1247 assert(inmap != nullptr, "must have inmap"); 1248 // In case of null check on receiver above 1249 map()->transfer_replaced_nodes_from(inmap, _new_idx); 1250 1251 uint i; 1252 1253 // Pass thru the predefined input parameters. 1254 for (i = 0; i < TypeFunc::Parms; i++) { 1255 map()->init_req(i, inmap->in(i)); 1256 } 1257 1258 if (depth() == 1) { 1259 assert(map()->memory()->Opcode() == Op_Parm, ""); 1260 // Insert the memory aliasing node 1261 set_all_memory(reset_memory()); 1262 } 1263 assert(merged_memory(), ""); 1264 1265 // Now add the locals which are initially bound to arguments: 1266 uint arg_size = tf()->domain()->cnt(); 1267 ensure_stack(arg_size - TypeFunc::Parms); // OSR methods have funny args 1268 for (i = TypeFunc::Parms; i < arg_size; i++) { 1269 map()->init_req(i, inmap->argument(_caller, i - TypeFunc::Parms)); 1270 } 1271 1272 // Clear out the rest of the map (locals and stack) 1273 for (i = arg_size; i < len; i++) { 1274 map()->init_req(i, top()); 1275 } 1276 1277 SafePointNode* entry_map = stop(); 1278 return entry_map; 1279 } 1280 1281 //-----------------------------do_method_entry-------------------------------- 1282 // Emit any code needed in the pseudo-block before BCI zero. 1283 // The main thing to do is lock the receiver of a synchronized method. 1284 void Parse::do_method_entry() { 1285 set_parse_bci(InvocationEntryBci); // Pseudo-BCP 1286 set_sp(0); // Java Stack Pointer 1287 1288 NOT_PRODUCT( count_compiled_calls(true/*at_method_entry*/, false/*is_inline*/); ) 1289 1290 if (C->env()->dtrace_method_probes()) { 1291 make_dtrace_method_entry(method()); 1292 } 1293 1294 #ifdef ASSERT 1295 // Narrow receiver type when it is too broad for the method being parsed. 1296 if (!method()->is_static()) { 1297 ciInstanceKlass* callee_holder = method()->holder(); 1298 const Type* holder_type = TypeInstPtr::make(TypePtr::BotPTR, callee_holder, Type::trust_interfaces); 1299 1300 Node* receiver_obj = local(0); 1301 const TypeInstPtr* receiver_type = _gvn.type(receiver_obj)->isa_instptr(); 1302 1303 if (receiver_type != nullptr && !receiver_type->higher_equal(holder_type)) { 1304 // Receiver should always be a subtype of callee holder. 1305 // But, since C2 type system doesn't properly track interfaces, 1306 // the invariant can't be expressed in the type system for default methods. 1307 // Example: for unrelated C <: I and D <: I, (C `meet` D) = Object </: I. 1308 assert(callee_holder->is_interface(), "missing subtype check"); 1309 1310 // Perform dynamic receiver subtype check against callee holder class w/ a halt on failure. 1311 Node* holder_klass = _gvn.makecon(TypeKlassPtr::make(callee_holder, Type::trust_interfaces)); 1312 Node* not_subtype_ctrl = gen_subtype_check(receiver_obj, holder_klass); 1313 assert(!stopped(), "not a subtype"); 1314 1315 Node* halt = _gvn.transform(new HaltNode(not_subtype_ctrl, frameptr(), "failed receiver subtype check")); 1316 C->root()->add_req(halt); 1317 } 1318 } 1319 #endif // ASSERT 1320 1321 // If the method is synchronized, we need to construct a lock node, attach 1322 // it to the Start node, and pin it there. 1323 if (method()->is_synchronized()) { 1324 // Insert a FastLockNode right after the Start which takes as arguments 1325 // the current thread pointer, the "this" pointer & the address of the 1326 // stack slot pair used for the lock. The "this" pointer is a projection 1327 // off the start node, but the locking spot has to be constructed by 1328 // creating a ConLNode of 0, and boxing it with a BoxLockNode. The BoxLockNode 1329 // becomes the second argument to the FastLockNode call. The 1330 // FastLockNode becomes the new control parent to pin it to the start. 1331 1332 // Setup Object Pointer 1333 Node *lock_obj = nullptr; 1334 if (method()->is_static()) { 1335 ciInstance* mirror = _method->holder()->java_mirror(); 1336 const TypeInstPtr *t_lock = TypeInstPtr::make(mirror); 1337 lock_obj = makecon(t_lock); 1338 } else { // Else pass the "this" pointer, 1339 lock_obj = local(0); // which is Parm0 from StartNode 1340 } 1341 // Clear out dead values from the debug info. 1342 kill_dead_locals(); 1343 // Build the FastLockNode 1344 _synch_lock = shared_lock(lock_obj); 1345 } 1346 1347 // Feed profiling data for parameters to the type system so it can 1348 // propagate it as speculative types 1349 record_profiled_parameters_for_speculation(); 1350 } 1351 1352 //------------------------------init_blocks------------------------------------ 1353 // Initialize our parser map to contain the types/monitors at method entry. 1354 void Parse::init_blocks() { 1355 // Create the blocks. 1356 _block_count = flow()->block_count(); 1357 _blocks = NEW_RESOURCE_ARRAY(Block, _block_count); 1358 1359 // Initialize the structs. 1360 for (int rpo = 0; rpo < block_count(); rpo++) { 1361 Block* block = rpo_at(rpo); 1362 new(block) Block(this, rpo); 1363 } 1364 1365 // Collect predecessor and successor information. 1366 for (int rpo = 0; rpo < block_count(); rpo++) { 1367 Block* block = rpo_at(rpo); 1368 block->init_graph(this); 1369 } 1370 } 1371 1372 //-------------------------------init_node------------------------------------- 1373 Parse::Block::Block(Parse* outer, int rpo) : _live_locals() { 1374 _flow = outer->flow()->rpo_at(rpo); 1375 _pred_count = 0; 1376 _preds_parsed = 0; 1377 _count = 0; 1378 _is_parsed = false; 1379 _is_handler = false; 1380 _has_merged_backedge = false; 1381 _start_map = nullptr; 1382 _has_predicates = false; 1383 _num_successors = 0; 1384 _all_successors = 0; 1385 _successors = nullptr; 1386 assert(pred_count() == 0 && preds_parsed() == 0, "sanity"); 1387 assert(!(is_merged() || is_parsed() || is_handler() || has_merged_backedge()), "sanity"); 1388 assert(!_live_locals.is_valid(), "sanity"); 1389 1390 // entry point has additional predecessor 1391 if (flow()->is_start()) _pred_count++; 1392 assert(flow()->is_start() == (this == outer->start_block()), ""); 1393 } 1394 1395 //-------------------------------init_graph------------------------------------ 1396 void Parse::Block::init_graph(Parse* outer) { 1397 // Create the successor list for this parser block. 1398 GrowableArray<ciTypeFlow::Block*>* tfs = flow()->successors(); 1399 GrowableArray<ciTypeFlow::Block*>* tfe = flow()->exceptions(); 1400 int ns = tfs->length(); 1401 int ne = tfe->length(); 1402 _num_successors = ns; 1403 _all_successors = ns+ne; 1404 _successors = (ns+ne == 0) ? nullptr: NEW_RESOURCE_ARRAY(Block*, ns+ne); 1405 for (int i = 0; i < ns+ne; i++) { 1406 ciTypeFlow::Block* tf2 = (i < ns) ? tfs->at(i) : tfe->at(i-ns); 1407 Block* block2 = outer->rpo_at(tf2->rpo()); 1408 _successors[i] = block2; 1409 1410 // Accumulate pred info for the other block, too. 1411 // Note: We also need to set _pred_count for exception blocks since they could 1412 // also have normal predecessors (reached without athrow by an explicit jump). 1413 // This also means that next_path_num can be called along exception paths. 1414 block2->_pred_count++; 1415 if (i >= ns) { 1416 block2->_is_handler = true; 1417 } 1418 1419 #ifdef ASSERT 1420 // A block's successors must be distinguishable by BCI. 1421 // That is, no bytecode is allowed to branch to two different 1422 // clones of the same code location. 1423 for (int j = 0; j < i; j++) { 1424 Block* block1 = _successors[j]; 1425 if (block1 == block2) continue; // duplicates are OK 1426 assert(block1->start() != block2->start(), "successors have unique bcis"); 1427 } 1428 #endif 1429 } 1430 1431 if (DoPartialEscapeAnalysis) { 1432 GrowableArray<ciTypeFlow::Block*>* tfp = flow()->predecessors(); 1433 int np = tfp->length(); 1434 _predecessors = np > 0 ? NEW_RESOURCE_ARRAY(Block*, np) : nullptr; 1435 for (int i = 0; i < np; ++i) { 1436 ciTypeFlow::Block* tf2 = tfp->at(i); 1437 Block* block2 = outer->rpo_at(tf2->rpo()); 1438 _predecessors[i] = block2; 1439 } 1440 } 1441 } 1442 1443 //---------------------------successor_for_bci--------------------------------- 1444 Parse::Block* Parse::Block::successor_for_bci(int bci) { 1445 for (int i = 0; i < all_successors(); i++) { 1446 Block* block2 = successor_at(i); 1447 if (block2->start() == bci) return block2; 1448 } 1449 // We can actually reach here if ciTypeFlow traps out a block 1450 // due to an unloaded class, and concurrently with compilation the 1451 // class is then loaded, so that a later phase of the parser is 1452 // able to see more of the bytecode CFG. Or, the flow pass and 1453 // the parser can have a minor difference of opinion about executability 1454 // of bytecodes. For example, "obj.field = null" is executable even 1455 // if the field's type is an unloaded class; the flow pass used to 1456 // make a trap for such code. 1457 return nullptr; 1458 } 1459 1460 1461 //-----------------------------stack_type_at----------------------------------- 1462 const Type* Parse::Block::stack_type_at(int i) const { 1463 return get_type(flow()->stack_type_at(i)); 1464 } 1465 1466 1467 //-----------------------------local_type_at----------------------------------- 1468 const Type* Parse::Block::local_type_at(int i) const { 1469 // This bitmap can be zero length if we saw a breakpoint. 1470 // In such cases, pretend they are all live. 1471 auto live_locals = liveness(); 1472 if (live_locals.size() > 0 && !live_locals.at(i)) 1473 return Type::BOTTOM; 1474 1475 return get_type(flow()->local_type_at(i)); 1476 } 1477 1478 1479 #ifndef PRODUCT 1480 1481 //----------------------------name_for_bc-------------------------------------- 1482 // helper method for BytecodeParseHistogram 1483 static const char* name_for_bc(int i) { 1484 return Bytecodes::is_defined(i) ? Bytecodes::name(Bytecodes::cast(i)) : "xxxunusedxxx"; 1485 } 1486 1487 //----------------------------BytecodeParseHistogram------------------------------------ 1488 Parse::BytecodeParseHistogram::BytecodeParseHistogram(Parse *p, Compile *c) { 1489 _parser = p; 1490 _compiler = c; 1491 if( ! _initialized ) { _initialized = true; reset(); } 1492 } 1493 1494 //----------------------------current_count------------------------------------ 1495 int Parse::BytecodeParseHistogram::current_count(BPHType bph_type) { 1496 switch( bph_type ) { 1497 case BPH_transforms: { return _parser->gvn().made_progress(); } 1498 case BPH_values: { return _parser->gvn().made_new_values(); } 1499 default: { ShouldNotReachHere(); return 0; } 1500 } 1501 } 1502 1503 //----------------------------initialized-------------------------------------- 1504 bool Parse::BytecodeParseHistogram::initialized() { return _initialized; } 1505 1506 //----------------------------reset-------------------------------------------- 1507 void Parse::BytecodeParseHistogram::reset() { 1508 int i = Bytecodes::number_of_codes; 1509 while (i-- > 0) { _bytecodes_parsed[i] = 0; _nodes_constructed[i] = 0; _nodes_transformed[i] = 0; _new_values[i] = 0; } 1510 } 1511 1512 //----------------------------set_initial_state-------------------------------- 1513 // Record info when starting to parse one bytecode 1514 void Parse::BytecodeParseHistogram::set_initial_state( Bytecodes::Code bc ) { 1515 if( PrintParseStatistics && !_parser->is_osr_parse() ) { 1516 _initial_bytecode = bc; 1517 _initial_node_count = _compiler->unique(); 1518 _initial_transforms = current_count(BPH_transforms); 1519 _initial_values = current_count(BPH_values); 1520 } 1521 } 1522 1523 //----------------------------record_change-------------------------------- 1524 // Record results of parsing one bytecode 1525 void Parse::BytecodeParseHistogram::record_change() { 1526 if( PrintParseStatistics && !_parser->is_osr_parse() ) { 1527 ++_bytecodes_parsed[_initial_bytecode]; 1528 _nodes_constructed [_initial_bytecode] += (_compiler->unique() - _initial_node_count); 1529 _nodes_transformed [_initial_bytecode] += (current_count(BPH_transforms) - _initial_transforms); 1530 _new_values [_initial_bytecode] += (current_count(BPH_values) - _initial_values); 1531 } 1532 } 1533 1534 1535 //----------------------------print-------------------------------------------- 1536 void Parse::BytecodeParseHistogram::print(float cutoff) { 1537 ResourceMark rm; 1538 // print profile 1539 int total = 0; 1540 int i = 0; 1541 for( i = 0; i < Bytecodes::number_of_codes; ++i ) { total += _bytecodes_parsed[i]; } 1542 int abs_sum = 0; 1543 tty->cr(); //0123456789012345678901234567890123456789012345678901234567890123456789 1544 tty->print_cr("Histogram of %d parsed bytecodes:", total); 1545 if( total == 0 ) { return; } 1546 tty->cr(); 1547 tty->print_cr("absolute: count of compiled bytecodes of this type"); 1548 tty->print_cr("relative: percentage contribution to compiled nodes"); 1549 tty->print_cr("nodes : Average number of nodes constructed per bytecode"); 1550 tty->print_cr("rnodes : Significance towards total nodes constructed, (nodes*relative)"); 1551 tty->print_cr("transforms: Average amount of transform progress per bytecode compiled"); 1552 tty->print_cr("values : Average number of node values improved per bytecode"); 1553 tty->print_cr("name : Bytecode name"); 1554 tty->cr(); 1555 tty->print_cr(" absolute relative nodes rnodes transforms values name"); 1556 tty->print_cr("----------------------------------------------------------------------"); 1557 while (--i > 0) { 1558 int abs = _bytecodes_parsed[i]; 1559 float rel = abs * 100.0F / total; 1560 float nodes = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _nodes_constructed[i])/_bytecodes_parsed[i]; 1561 float rnodes = _bytecodes_parsed[i] == 0 ? 0 : rel * nodes; 1562 float xforms = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _nodes_transformed[i])/_bytecodes_parsed[i]; 1563 float values = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _new_values [i])/_bytecodes_parsed[i]; 1564 if (cutoff <= rel) { 1565 tty->print_cr("%10d %7.2f%% %6.1f %6.2f %6.1f %6.1f %s", abs, rel, nodes, rnodes, xforms, values, name_for_bc(i)); 1566 abs_sum += abs; 1567 } 1568 } 1569 tty->print_cr("----------------------------------------------------------------------"); 1570 float rel_sum = abs_sum * 100.0F / total; 1571 tty->print_cr("%10d %7.2f%% (cutoff = %.2f%%)", abs_sum, rel_sum, cutoff); 1572 tty->print_cr("----------------------------------------------------------------------"); 1573 tty->cr(); 1574 } 1575 #endif 1576 1577 //----------------------------load_state_from---------------------------------- 1578 // Load block/map/sp. But not do not touch iter/bci. 1579 void Parse::load_state_from(Block* block) { 1580 set_block(block); 1581 // load the block's JVM state: 1582 set_map(block->start_map()); 1583 set_sp( block->start_sp()); 1584 } 1585 1586 1587 //-----------------------------record_state------------------------------------ 1588 void Parse::Block::record_state(Parse* p, int pnum) { 1589 assert(!is_merged(), "can only record state once, on 1st inflow"); 1590 assert(start_sp() == p->sp(), "stack pointer must agree with ciTypeFlow"); 1591 set_start_map(p->stop()); 1592 1593 _from_block = p->block(); 1594 _init_pnum = pnum; 1595 } 1596 1597 1598 //------------------------------do_one_block----------------------------------- 1599 void Parse::do_one_block() { 1600 if (TraceOptoParse) { 1601 Block *b = block(); 1602 int ns = b->num_successors(); 1603 int nt = b->all_successors(); 1604 1605 tty->print("Parsing block #%d at bci [%d,%d), successors:", 1606 block()->rpo(), block()->start(), block()->limit()); 1607 for (int i = 0; i < nt; i++) { 1608 tty->print((( i < ns) ? " %d" : " %d(exception block)"), b->successor_at(i)->rpo()); 1609 } 1610 if (b->is_loop_head()) { 1611 tty->print(" loop head"); 1612 } 1613 if (b->is_irreducible_loop_entry()) { 1614 tty->print(" irreducible"); 1615 } 1616 tty->cr(); 1617 } 1618 1619 #ifndef PRODUCT 1620 if (PEAVerbose) { 1621 PEAState& as = jvms()->alloc_state(); 1622 as.print_on(tty); 1623 } 1624 #endif 1625 assert(block()->is_merged(), "must be merged before being parsed"); 1626 block()->mark_parsed(); 1627 1628 // Set iterator to start of block. 1629 iter().reset_to_bci(block()->start()); 1630 1631 if (ProfileExceptionHandlers && block()->is_handler()) { 1632 ciMethodData* methodData = method()->method_data(); 1633 if (methodData->is_mature()) { 1634 ciBitData data = methodData->exception_handler_bci_to_data(block()->start()); 1635 if (!data.exception_handler_entered() || StressPrunedExceptionHandlers) { 1636 // dead catch block 1637 // Emit an uncommon trap instead of processing the block. 1638 set_parse_bci(block()->start()); 1639 uncommon_trap(Deoptimization::Reason_unreached, 1640 Deoptimization::Action_reinterpret, 1641 nullptr, "dead catch block"); 1642 return; 1643 } 1644 } 1645 } 1646 1647 CompileLog* log = C->log(); 1648 1649 // Parse bytecodes 1650 while (!stopped() && !failing()) { 1651 iter().next(); 1652 1653 // Learn the current bci from the iterator: 1654 set_parse_bci(iter().cur_bci()); 1655 1656 if (bci() == block()->limit()) { 1657 // Do not walk into the next block until directed by do_all_blocks. 1658 merge(bci()); 1659 break; 1660 } 1661 assert(bci() < block()->limit(), "bci still in block"); 1662 1663 if (log != nullptr) { 1664 // Output an optional context marker, to help place actions 1665 // that occur during parsing of this BC. If there is no log 1666 // output until the next context string, this context string 1667 // will be silently ignored. 1668 log->set_context("bc code='%d' bci='%d'", (int)bc(), bci()); 1669 } 1670 1671 if (block()->has_trap_at(bci())) { 1672 // We must respect the flow pass's traps, because it will refuse 1673 // to produce successors for trapping blocks. 1674 int trap_index = block()->flow()->trap_index(); 1675 assert(trap_index != 0, "trap index must be valid"); 1676 uncommon_trap(trap_index); 1677 break; 1678 } 1679 1680 NOT_PRODUCT( parse_histogram()->set_initial_state(bc()); ); 1681 1682 #ifdef ASSERT 1683 int pre_bc_sp = sp(); 1684 int inputs, depth; 1685 bool have_se = !stopped() && compute_stack_effects(inputs, depth); 1686 assert(!have_se || pre_bc_sp >= inputs, "have enough stack to execute this BC: pre_bc_sp=%d, inputs=%d", pre_bc_sp, inputs); 1687 #endif //ASSERT 1688 1689 do_one_bytecode(); 1690 if (failing()) return; 1691 1692 assert(!have_se || stopped() || failing() || (sp() - pre_bc_sp) == depth, 1693 "incorrect depth prediction: sp=%d, pre_bc_sp=%d, depth=%d", sp(), pre_bc_sp, depth); 1694 1695 do_exceptions(); 1696 1697 NOT_PRODUCT( parse_histogram()->record_change(); ); 1698 1699 if (log != nullptr) 1700 log->clear_context(); // skip marker if nothing was printed 1701 1702 // Fall into next bytecode. Each bytecode normally has 1 sequential 1703 // successor which is typically made ready by visiting this bytecode. 1704 // If the successor has several predecessors, then it is a merge 1705 // point, starts a new basic block, and is handled like other basic blocks. 1706 } 1707 } 1708 1709 1710 //------------------------------merge------------------------------------------ 1711 void Parse::set_parse_bci(int bci) { 1712 set_bci(bci); 1713 Node_Notes* nn = C->default_node_notes(); 1714 if (nn == nullptr) return; 1715 1716 // Collect debug info for inlined calls unless -XX:-DebugInlinedCalls. 1717 if (!DebugInlinedCalls && depth() > 1) { 1718 return; 1719 } 1720 1721 // Update the JVMS annotation, if present. 1722 JVMState* jvms = nn->jvms(); 1723 if (jvms != nullptr && jvms->bci() != bci) { 1724 // Update the JVMS. 1725 jvms = jvms->clone_shallow(C); 1726 jvms->set_bci(bci); 1727 nn->set_jvms(jvms); 1728 } 1729 } 1730 1731 //------------------------------merge------------------------------------------ 1732 // Merge the current mapping into the basic block starting at bci 1733 void Parse::merge(int target_bci) { 1734 Block* target = successor_for_bci(target_bci); 1735 if (target == nullptr) { handle_missing_successor(target_bci); return; } 1736 assert(!target->is_ready(), "our arrival must be expected"); 1737 int pnum = target->next_path_num(); 1738 merge_common(target, pnum); 1739 } 1740 1741 //-------------------------merge_new_path-------------------------------------- 1742 // Merge the current mapping into the basic block, using a new path 1743 void Parse::merge_new_path(int target_bci) { 1744 Block* target = successor_for_bci(target_bci); 1745 if (target == nullptr) { handle_missing_successor(target_bci); return; } 1746 assert(!target->is_ready(), "new path into frozen graph"); 1747 int pnum = target->add_new_path(); 1748 merge_common(target, pnum); 1749 } 1750 1751 //-------------------------merge_exception------------------------------------- 1752 // Merge the current mapping into the basic block starting at bci 1753 // The ex_oop must be pushed on the stack, unlike throw_to_exit. 1754 void Parse::merge_exception(int target_bci) { 1755 #ifdef ASSERT 1756 if (target_bci < bci()) { 1757 C->set_exception_backedge(); 1758 } 1759 #endif 1760 assert(sp() == 1, "must have only the throw exception on the stack"); 1761 Block* target = successor_for_bci(target_bci); 1762 if (target == nullptr) { handle_missing_successor(target_bci); return; } 1763 assert(target->is_handler(), "exceptions are handled by special blocks"); 1764 int pnum = target->add_new_path(); 1765 merge_common(target, pnum); 1766 } 1767 1768 //--------------------handle_missing_successor--------------------------------- 1769 void Parse::handle_missing_successor(int target_bci) { 1770 #ifndef PRODUCT 1771 Block* b = block(); 1772 int trap_bci = b->flow()->has_trap()? b->flow()->trap_bci(): -1; 1773 tty->print_cr("### Missing successor at bci:%d for block #%d (trap_bci:%d)", target_bci, b->rpo(), trap_bci); 1774 #endif 1775 ShouldNotReachHere(); 1776 } 1777 1778 //--------------------------merge_common--------------------------------------- 1779 void Parse::merge_common(Parse::Block* target, int pnum) { 1780 if (TraceOptoParse) { 1781 tty->print("Merging state at block #%d bci:%d", target->rpo(), target->start()); 1782 if (!target->is_merged()) { 1783 tty->print(" with empty state"); 1784 } else { 1785 tty->print(" with previous state"); 1786 } 1787 tty->print_cr(" on path %d", pnum); 1788 } 1789 1790 // Zap extra stack slots to top 1791 assert(sp() == target->start_sp(), ""); 1792 clean_stack(sp()); 1793 1794 if (!target->is_merged()) { // No prior mapping at this bci 1795 // If this path is dead, do not bother capturing it as a merge. 1796 // It is "as if" we had 1 fewer predecessors from the beginning. 1797 if (stopped()) { 1798 if (TraceOptoParse) tty->print_cr(", but path is dead and doesn't count"); 1799 return; 1800 } 1801 1802 // Make a region if we know there are multiple or unpredictable inputs. 1803 // (Also, if this is a plain fall-through, we might see another region, 1804 // which must not be allowed into this block's map.) 1805 if (pnum > PhiNode::Input // Known multiple inputs. 1806 || target->is_handler() // These have unpredictable inputs. 1807 || target->is_loop_head() // Known multiple inputs 1808 || control()->is_Region()) { // We must hide this guy. 1809 1810 int current_bci = bci(); 1811 set_parse_bci(target->start()); // Set target bci 1812 if (target->is_SEL_head()) { 1813 DEBUG_ONLY( target->mark_merged_backedge(block()); ) 1814 if (target->start() == 0) { 1815 // Add Parse Predicates for the special case when 1816 // there are backbranches to the method entry. 1817 add_parse_predicates(); 1818 } 1819 } 1820 // Add a Region to start the new basic block. Phis will be added 1821 // later lazily. 1822 int edges = target->pred_count(); 1823 if (edges < pnum) edges = pnum; // might be a new path! 1824 RegionNode *r = new RegionNode(edges+1); 1825 gvn().set_type(r, Type::CONTROL); 1826 record_for_igvn(r); 1827 // zap all inputs to null for debugging (done in Node(uint) constructor) 1828 // for (int j = 1; j < edges+1; j++) { r->init_req(j, nullptr); } 1829 r->init_req(pnum, control()); 1830 set_control(r); 1831 target->copy_irreducible_status_to(r, jvms()); 1832 set_parse_bci(current_bci); // Restore bci 1833 } 1834 1835 // Convert the existing Parser mapping into a mapping at this bci. 1836 store_state_to(target, pnum); 1837 assert(target->is_merged(), "do not come here twice"); 1838 #ifdef ASSERT 1839 target->state().validate(); 1840 #endif 1841 } else { // Prior mapping at this bci 1842 1843 #ifdef ASSERT 1844 if (target->is_SEL_head()) { 1845 target->mark_merged_backedge(block()); 1846 } 1847 #endif 1848 // We must not manufacture more phis if the target is already parsed. 1849 bool nophi = target->is_parsed(); 1850 1851 SafePointNode* newin = map();// Hang on to incoming mapping 1852 Block* save_block = block(); // Hang on to incoming block; 1853 load_state_from(target); // Get prior mapping 1854 1855 assert(newin->jvms()->locoff() == jvms()->locoff(), "JVMS layouts agree"); 1856 assert(newin->jvms()->stkoff() == jvms()->stkoff(), "JVMS layouts agree"); 1857 assert(newin->jvms()->monoff() == jvms()->monoff(), "JVMS layouts agree"); 1858 assert(newin->jvms()->endoff() == jvms()->endoff(), "JVMS layouts agree"); 1859 1860 // Iterate over my current mapping and the old mapping. 1861 // Where different, insert Phi functions. 1862 // Use any existing Phi functions. 1863 assert(control()->is_Region(), "must be merging to a region"); 1864 RegionNode* r = control()->as_Region(); 1865 1866 // Compute where to merge into 1867 // Merge incoming control path 1868 r->init_req(pnum, newin->control()); 1869 1870 if (pnum == 1) { // Last merge for this Region? 1871 if (!block()->flow()->is_irreducible_loop_secondary_entry()) { 1872 Node* result = _gvn.transform(r); 1873 if (r != result && TraceOptoParse) { 1874 tty->print_cr("Block #%d replace %d with %d", block()->rpo(), r->_idx, result->_idx); 1875 } 1876 } 1877 record_for_igvn(r); 1878 } 1879 1880 // Update all the non-control inputs to map: 1881 assert(TypeFunc::Parms == newin->jvms()->locoff(), "parser map should contain only youngest jvms"); 1882 bool check_elide_phi = target->is_SEL_backedge(save_block); 1883 PEAState& pred_as = newin->jvms()->alloc_state(); 1884 PEAState& as = block()->state(); 1885 AllocationStateMerger as_merger(as); 1886 1887 for (uint j = 1; j < newin->req(); ++j) { 1888 Node* m = map()->in(j); // Current state of target. 1889 Node* n = newin->in(j); // Incoming change to target state. 1890 PhiNode* phi; 1891 if (m->is_Phi() && m->as_Phi()->region() == r) 1892 phi = m->as_Phi(); 1893 else 1894 phi = nullptr; 1895 if (m != n) { // Different; must merge 1896 switch (j) { 1897 // Frame pointer and Return Address never changes 1898 case TypeFunc::FramePtr:// Drop m, use the original value 1899 case TypeFunc::ReturnAdr: 1900 break; 1901 case TypeFunc::Memory: // Merge inputs to the MergeMem node 1902 assert(phi == nullptr, "the merge contains phis, not vice versa"); 1903 merge_memory_edges(n->as_MergeMem(), pnum, nophi); 1904 continue; 1905 default: // All normal stuff 1906 if (phi == nullptr) { 1907 const JVMState* jvms = map()->jvms(); 1908 if (EliminateNestedLocks && 1909 jvms->is_mon(j) && jvms->is_monitor_box(j)) { 1910 // BoxLock nodes are not commoning. 1911 // Use old BoxLock node as merged box. 1912 assert(newin->jvms()->is_monitor_box(j), "sanity"); 1913 // This assert also tests that nodes are BoxLock. 1914 assert(BoxLockNode::same_slot(n, m), "sanity"); 1915 C->gvn_replace_by(n, m); 1916 } else if (!check_elide_phi || !target->can_elide_SEL_phi(j)) { 1917 phi = ensure_phi(j, nophi); 1918 1919 // We merges allocation state according to 5.2.4 of the dissertation 1920 // Becase C2 Parse is merging basic blocks, we have to intercept some phi creation or 1921 // PEA MergeProcessor creates duplicated phi nodes. 1922 if (DoPartialEscapeAnalysis && phi != nullptr) { 1923 PartialEscapeAnalysis* pea = PEA(); 1924 as_merger.merge_at_phi_creation(pea, pred_as, phi, m, n); 1925 } // DoPartialEscapeAnalysis 1926 } 1927 } 1928 break; 1929 } 1930 } 1931 // At this point, n might be top if: 1932 // - there is no phi (because TypeFlow detected a conflict), or 1933 // - the corresponding control edges is top (a dead incoming path) 1934 // It is a bug if we create a phi which sees a garbage value on a live path. 1935 1936 if (phi != nullptr) { 1937 assert(n != top() || r->in(pnum) == top(), "live value must not be garbage"); 1938 assert(phi->region() == r, ""); 1939 1940 phi->set_req(pnum, n); // Then add 'n' to the merge 1941 if (pnum == PhiNode::Input) { 1942 // Last merge for this Phi. 1943 // So far, Phis have had a reasonable type from ciTypeFlow. 1944 // Now _gvn will join that with the meet of current inputs. 1945 // BOTTOM is never permissible here, 'cause pessimistically 1946 // Phis of pointers cannot lose the basic pointer type. 1947 debug_only(const Type* bt1 = phi->bottom_type()); 1948 assert(bt1 != Type::BOTTOM, "should not be building conflict phis"); 1949 map()->set_req(j, _gvn.transform(phi)); 1950 debug_only(const Type* bt2 = phi->bottom_type()); 1951 assert(bt2->higher_equal_speculative(bt1), "must be consistent with type-flow"); 1952 record_for_igvn(phi); 1953 } 1954 } 1955 } // End of for all values to be merged 1956 1957 1958 if (DoPartialEscapeAnalysis) { 1959 as_merger.merge(pred_as, this, r, pnum); 1960 } 1961 1962 if (pnum == PhiNode::Input && 1963 !r->in(0)) { // The occasional useless Region 1964 assert(control() == r, ""); 1965 set_control(r->nonnull_req()); 1966 } 1967 1968 map()->merge_replaced_nodes_with(newin); 1969 1970 #ifdef ASSERT 1971 block()->state().validate(); 1972 #endif 1973 // newin has been subsumed into the lazy merge, and is now dead. 1974 set_block(save_block); 1975 1976 stop(); // done with this guy, for now 1977 } 1978 1979 // Done with this parser state. 1980 assert(stopped(), ""); 1981 } 1982 1983 1984 //--------------------------merge_memory_edges--------------------------------- 1985 void Parse::merge_memory_edges(MergeMemNode* n, int pnum, bool nophi) { 1986 // (nophi means we must not create phis, because we already parsed here) 1987 assert(n != nullptr, ""); 1988 // Merge the inputs to the MergeMems 1989 MergeMemNode* m = merged_memory(); 1990 1991 assert(control()->is_Region(), "must be merging to a region"); 1992 RegionNode* r = control()->as_Region(); 1993 1994 PhiNode* base = nullptr; 1995 MergeMemNode* remerge = nullptr; 1996 for (MergeMemStream mms(m, n); mms.next_non_empty2(); ) { 1997 Node *p = mms.force_memory(); 1998 Node *q = mms.memory2(); 1999 if (mms.is_empty() && nophi) { 2000 // Trouble: No new splits allowed after a loop body is parsed. 2001 // Instead, wire the new split into a MergeMem on the backedge. 2002 // The optimizer will sort it out, slicing the phi. 2003 if (remerge == nullptr) { 2004 guarantee(base != nullptr, ""); 2005 assert(base->in(0) != nullptr, "should not be xformed away"); 2006 remerge = MergeMemNode::make(base->in(pnum)); 2007 gvn().set_type(remerge, Type::MEMORY); 2008 base->set_req(pnum, remerge); 2009 } 2010 remerge->set_memory_at(mms.alias_idx(), q); 2011 continue; 2012 } 2013 assert(!q->is_MergeMem(), ""); 2014 PhiNode* phi; 2015 if (p != q) { 2016 phi = ensure_memory_phi(mms.alias_idx(), nophi); 2017 } else { 2018 if (p->is_Phi() && p->as_Phi()->region() == r) 2019 phi = p->as_Phi(); 2020 else 2021 phi = nullptr; 2022 } 2023 // Insert q into local phi 2024 if (phi != nullptr) { 2025 assert(phi->region() == r, ""); 2026 p = phi; 2027 phi->set_req(pnum, q); 2028 if (mms.at_base_memory()) { 2029 base = phi; // delay transforming it 2030 } else if (pnum == 1) { 2031 record_for_igvn(phi); 2032 p = _gvn.transform(phi); 2033 } 2034 mms.set_memory(p);// store back through the iterator 2035 } 2036 } 2037 // Transform base last, in case we must fiddle with remerging. 2038 if (base != nullptr && pnum == 1) { 2039 record_for_igvn(base); 2040 m->set_base_memory(_gvn.transform(base)); 2041 } 2042 } 2043 2044 2045 //------------------------ensure_phis_everywhere------------------------------- 2046 void Parse::ensure_phis_everywhere() { 2047 ensure_phi(TypeFunc::I_O); 2048 2049 // Ensure a phi on all currently known memories. 2050 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 2051 ensure_memory_phi(mms.alias_idx()); 2052 debug_only(mms.set_memory()); // keep the iterator happy 2053 } 2054 2055 // Note: This is our only chance to create phis for memory slices. 2056 // If we miss a slice that crops up later, it will have to be 2057 // merged into the base-memory phi that we are building here. 2058 // Later, the optimizer will comb out the knot, and build separate 2059 // phi-loops for each memory slice that matters. 2060 2061 // Monitors must nest nicely and not get confused amongst themselves. 2062 // Phi-ify everything up to the monitors, though. 2063 uint monoff = map()->jvms()->monoff(); 2064 uint nof_monitors = map()->jvms()->nof_monitors(); 2065 2066 assert(TypeFunc::Parms == map()->jvms()->locoff(), "parser map should contain only youngest jvms"); 2067 bool check_elide_phi = block()->is_SEL_head(); 2068 for (uint i = TypeFunc::Parms; i < monoff; i++) { 2069 if (!check_elide_phi || !block()->can_elide_SEL_phi(i)) { 2070 ensure_phi(i); 2071 } 2072 } 2073 2074 // Even monitors need Phis, though they are well-structured. 2075 // This is true for OSR methods, and also for the rare cases where 2076 // a monitor object is the subject of a replace_in_map operation. 2077 // See bugs 4426707 and 5043395. 2078 for (uint m = 0; m < nof_monitors; m++) { 2079 ensure_phi(map()->jvms()->monitor_obj_offset(m)); 2080 } 2081 } 2082 2083 2084 //-----------------------------add_new_path------------------------------------ 2085 // Add a previously unaccounted predecessor to this block. 2086 int Parse::Block::add_new_path() { 2087 // If there is no map, return the lowest unused path number. 2088 if (!is_merged()) return pred_count()+1; // there will be a map shortly 2089 2090 SafePointNode* map = start_map(); 2091 if (!map->control()->is_Region()) 2092 return pred_count()+1; // there may be a region some day 2093 RegionNode* r = map->control()->as_Region(); 2094 2095 // Add new path to the region. 2096 const uint pnum = r->req(); 2097 r->add_req(nullptr); 2098 2099 for (DUIterator_Fast imax, i = r->fast_outs(imax); i < imax; i++) { 2100 Node* n = r->fast_out(i); 2101 2102 if (n->is_MergeMem()) { 2103 // Ensure a phi on all currently known memories. 2104 for (MergeMemStream mms(n->as_MergeMem()); mms.next_non_empty(); ) { 2105 Node* phi = mms.memory(); 2106 if (phi->is_Phi() && phi->as_Phi()->region() == r && phi->req() <= pnum) { 2107 assert(phi->req() == pnum, "must be same size as region"); 2108 phi->add_req(nullptr); 2109 } 2110 } 2111 } else if (n->is_Phi() && n->as_Phi()->region() == r && n->req() <= pnum) { 2112 assert(n->req() == pnum, "must be same size as region"); 2113 n->add_req(nullptr); 2114 } 2115 } 2116 2117 return pnum; 2118 } 2119 2120 //------------------------------ensure_phi------------------------------------- 2121 // Turn the idx'th entry of the current map into a Phi 2122 PhiNode *Parse::ensure_phi(int idx, bool nocreate) { 2123 SafePointNode* map = this->map(); 2124 Node* region = map->control(); 2125 assert(region->is_Region(), ""); 2126 2127 Node* o = map->in(idx); 2128 assert(o != nullptr, ""); 2129 2130 if (o == top()) return nullptr; // TOP always merges into TOP 2131 2132 if (o->is_Phi() && o->as_Phi()->region() == region) { 2133 return o->as_Phi(); 2134 } 2135 2136 // Now use a Phi here for merging 2137 assert(!nocreate, "Cannot build a phi for a block already parsed."); 2138 const JVMState* jvms = map->jvms(); 2139 const Type* t = nullptr; 2140 if (jvms->is_loc(idx)) { 2141 t = block()->local_type_at(idx - jvms->locoff()); 2142 } else if (jvms->is_stk(idx)) { 2143 t = block()->stack_type_at(idx - jvms->stkoff()); 2144 } else if (jvms->is_mon(idx)) { 2145 assert(!jvms->is_monitor_box(idx), "no phis for boxes"); 2146 t = TypeInstPtr::BOTTOM; // this is sufficient for a lock object 2147 } else if ((uint)idx < TypeFunc::Parms) { 2148 t = o->bottom_type(); // Type::RETURN_ADDRESS or such-like. 2149 } else { 2150 assert(false, "no type information for this phi"); 2151 } 2152 2153 // If the type falls to bottom, then this must be a local that 2154 // is mixing ints and oops or some such. Forcing it to top 2155 // makes it go dead. 2156 if (t == Type::BOTTOM) { 2157 map->set_req(idx, top()); 2158 return nullptr; 2159 } 2160 2161 // Do not create phis for top either. 2162 // A top on a non-null control flow must be an unused even after the.phi. 2163 if (t == Type::TOP || t == Type::HALF) { 2164 map->set_req(idx, top()); 2165 return nullptr; 2166 } 2167 2168 PhiNode* phi = PhiNode::make(region, o, t); 2169 gvn().set_type(phi, t); 2170 if (C->do_escape_analysis()) record_for_igvn(phi); 2171 map->set_req(idx, phi); 2172 2173 return phi; 2174 } 2175 2176 //--------------------------ensure_memory_phi---------------------------------- 2177 // Turn the idx'th slice of the current memory into a Phi 2178 PhiNode *Parse::ensure_memory_phi(int idx, bool nocreate) { 2179 MergeMemNode* mem = merged_memory(); 2180 Node* region = control(); 2181 assert(region->is_Region(), ""); 2182 2183 Node *o = (idx == Compile::AliasIdxBot)? mem->base_memory(): mem->memory_at(idx); 2184 assert(o != nullptr && o != top(), ""); 2185 2186 PhiNode* phi; 2187 if (o->is_Phi() && o->as_Phi()->region() == region) { 2188 phi = o->as_Phi(); 2189 if (phi == mem->base_memory() && idx >= Compile::AliasIdxRaw) { 2190 // clone the shared base memory phi to make a new memory split 2191 assert(!nocreate, "Cannot build a phi for a block already parsed."); 2192 const Type* t = phi->bottom_type(); 2193 const TypePtr* adr_type = C->get_adr_type(idx); 2194 phi = phi->slice_memory(adr_type); 2195 gvn().set_type(phi, t); 2196 } 2197 return phi; 2198 } 2199 2200 // Now use a Phi here for merging 2201 assert(!nocreate, "Cannot build a phi for a block already parsed."); 2202 const Type* t = o->bottom_type(); 2203 const TypePtr* adr_type = C->get_adr_type(idx); 2204 phi = PhiNode::make(region, o, t, adr_type); 2205 gvn().set_type(phi, t); 2206 if (idx == Compile::AliasIdxBot) 2207 mem->set_base_memory(phi); 2208 else 2209 mem->set_memory_at(idx, phi); 2210 return phi; 2211 } 2212 2213 //------------------------------call_register_finalizer----------------------- 2214 // Check the klass of the receiver and call register_finalizer if the 2215 // class need finalization. 2216 void Parse::call_register_finalizer() { 2217 Node* receiver = local(0); 2218 assert(receiver != nullptr && receiver->bottom_type()->isa_instptr() != nullptr, 2219 "must have non-null instance type"); 2220 2221 const TypeInstPtr *tinst = receiver->bottom_type()->isa_instptr(); 2222 if (tinst != nullptr && tinst->is_loaded() && !tinst->klass_is_exact()) { 2223 // The type isn't known exactly so see if CHA tells us anything. 2224 ciInstanceKlass* ik = tinst->instance_klass(); 2225 if (!Dependencies::has_finalizable_subclass(ik)) { 2226 // No finalizable subclasses so skip the dynamic check. 2227 C->dependencies()->assert_has_no_finalizable_subclasses(ik); 2228 return; 2229 } 2230 } 2231 2232 // Insert a dynamic test for whether the instance needs 2233 // finalization. In general this will fold up since the concrete 2234 // class is often visible so the access flags are constant. 2235 Node* klass_addr = basic_plus_adr( receiver, receiver, oopDesc::klass_offset_in_bytes() ); 2236 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), klass_addr, TypeInstPtr::KLASS)); 2237 2238 Node* access_flags_addr = basic_plus_adr(klass, klass, in_bytes(Klass::access_flags_offset())); 2239 Node* access_flags = make_load(nullptr, access_flags_addr, TypeInt::INT, T_INT, MemNode::unordered); 2240 2241 Node* mask = _gvn.transform(new AndINode(access_flags, intcon(JVM_ACC_HAS_FINALIZER))); 2242 Node* check = _gvn.transform(new CmpINode(mask, intcon(0))); 2243 Node* test = _gvn.transform(new BoolNode(check, BoolTest::ne)); 2244 2245 IfNode* iff = create_and_map_if(control(), test, PROB_MAX, COUNT_UNKNOWN); 2246 2247 RegionNode* result_rgn = new RegionNode(3); 2248 record_for_igvn(result_rgn); 2249 2250 Node *skip_register = _gvn.transform(new IfFalseNode(iff)); 2251 result_rgn->init_req(1, skip_register); 2252 2253 Node *needs_register = _gvn.transform(new IfTrueNode(iff)); 2254 set_control(needs_register); 2255 if (stopped()) { 2256 // There is no slow path. 2257 result_rgn->init_req(2, top()); 2258 } else { 2259 Node *call = make_runtime_call(RC_NO_LEAF, 2260 OptoRuntime::register_finalizer_Type(), 2261 OptoRuntime::register_finalizer_Java(), 2262 nullptr, TypePtr::BOTTOM, 2263 receiver); 2264 make_slow_call_ex(call, env()->Throwable_klass(), true); 2265 2266 Node* fast_io = call->in(TypeFunc::I_O); 2267 Node* fast_mem = call->in(TypeFunc::Memory); 2268 // These two phis are pre-filled with copies of of the fast IO and Memory 2269 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 2270 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 2271 2272 result_rgn->init_req(2, control()); 2273 io_phi ->init_req(2, i_o()); 2274 mem_phi ->init_req(2, reset_memory()); 2275 2276 set_all_memory( _gvn.transform(mem_phi) ); 2277 set_i_o( _gvn.transform(io_phi) ); 2278 } 2279 2280 set_control( _gvn.transform(result_rgn) ); 2281 } 2282 2283 // Add check to deoptimize once holder klass is fully initialized. 2284 void Parse::clinit_deopt() { 2285 assert(C->has_method(), "only for normal compilations"); 2286 assert(depth() == 1, "only for main compiled method"); 2287 assert(is_normal_parse(), "no barrier needed on osr entry"); 2288 assert(!method()->holder()->is_not_initialized(), "initialization should have been started"); 2289 2290 set_parse_bci(0); 2291 2292 Node* holder = makecon(TypeKlassPtr::make(method()->holder(), Type::trust_interfaces)); 2293 guard_klass_being_initialized(holder); 2294 } 2295 2296 // Add check to deoptimize if RTM state is not ProfileRTM 2297 void Parse::rtm_deopt() { 2298 #if INCLUDE_RTM_OPT 2299 if (C->profile_rtm()) { 2300 assert(C->has_method(), "only for normal compilations"); 2301 assert(!C->method()->method_data()->is_empty(), "MDO is needed to record RTM state"); 2302 assert(depth() == 1, "generate check only for main compiled method"); 2303 2304 // Set starting bci for uncommon trap. 2305 set_parse_bci(is_osr_parse() ? osr_bci() : 0); 2306 2307 // Load the rtm_state from the MethodData. 2308 const TypePtr* adr_type = TypeMetadataPtr::make(C->method()->method_data()); 2309 Node* mdo = makecon(adr_type); 2310 int offset = in_bytes(MethodData::rtm_state_offset()); 2311 Node* adr_node = basic_plus_adr(mdo, mdo, offset); 2312 Node* rtm_state = make_load(control(), adr_node, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 2313 2314 // Separate Load from Cmp by Opaque. 2315 // In expand_macro_nodes() it will be replaced either 2316 // with this load when there are locks in the code 2317 // or with ProfileRTM (cmp->in(2)) otherwise so that 2318 // the check will fold. 2319 Node* profile_state = makecon(TypeInt::make(ProfileRTM)); 2320 Node* opq = _gvn.transform( new Opaque3Node(C, rtm_state, Opaque3Node::RTM_OPT) ); 2321 Node* chk = _gvn.transform( new CmpINode(opq, profile_state) ); 2322 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) ); 2323 // Branch to failure if state was changed 2324 { BuildCutout unless(this, tst, PROB_ALWAYS); 2325 uncommon_trap(Deoptimization::Reason_rtm_state_change, 2326 Deoptimization::Action_make_not_entrant); 2327 } 2328 } 2329 #endif 2330 } 2331 2332 //------------------------------return_current--------------------------------- 2333 // Append current _map to _exit_return 2334 void Parse::return_current(Node* value) { 2335 if (RegisterFinalizersAtInit && 2336 method()->intrinsic_id() == vmIntrinsics::_Object_init) { 2337 call_register_finalizer(); 2338 } 2339 2340 // Do not set_parse_bci, so that return goo is credited to the return insn. 2341 set_bci(InvocationEntryBci); 2342 if (method()->is_synchronized() && GenerateSynchronizationCode) { 2343 shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node()); 2344 } 2345 if (C->env()->dtrace_method_probes()) { 2346 make_dtrace_method_exit(method()); 2347 } 2348 SafePointNode* exit_return = _exits.map(); 2349 exit_return->in( TypeFunc::Control )->add_req( control() ); 2350 exit_return->in( TypeFunc::I_O )->add_req( i_o () ); 2351 Node *mem = exit_return->in( TypeFunc::Memory ); 2352 for (MergeMemStream mms(mem->as_MergeMem(), merged_memory()); mms.next_non_empty2(); ) { 2353 if (mms.is_empty()) { 2354 // get a copy of the base memory, and patch just this one input 2355 const TypePtr* adr_type = mms.adr_type(C); 2356 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 2357 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 2358 gvn().set_type_bottom(phi); 2359 phi->del_req(phi->req()-1); // prepare to re-patch 2360 mms.set_memory(phi); 2361 } 2362 mms.memory()->add_req(mms.memory2()); 2363 } 2364 2365 // frame pointer is always same, already captured 2366 if (value != nullptr) { 2367 // If returning oops to an interface-return, there is a silent free 2368 // cast from oop to interface allowed by the Verifier. Make it explicit 2369 // here. 2370 Node* phi = _exits.argument(0); 2371 phi->add_req(value); 2372 } 2373 2374 if (_first_return++ == 0) { 2375 _exits.map()->transfer_replaced_nodes_from(map(), _new_idx); 2376 // copy assignment 2377 _exits.jvms()->alloc_state() = jvms()->alloc_state(); 2378 } else { 2379 _exits.map()->merge_replaced_nodes_with(map()); 2380 2381 if (DoPartialEscapeAnalysis) { 2382 PEAState& as =_exits.jvms()->alloc_state(); 2383 PEAState& newin = jvms()->alloc_state(); 2384 AllocationStateMerger mp(as); 2385 // if value is a tracking object and PEA needs to create a phi node to merge it, 2386 // we need to use _exits.argument(0) 2387 ObjID obj = PEA()->is_alias(value); 2388 if (as.contains(obj) && newin.contains(obj)) { 2389 Node* phi = _exits.argument(0); 2390 mp.merge_at_phi_creation(PEA(), newin, phi->as_Phi(), phi->in(_first_return-1), value); 2391 } 2392 mp.merge(newin, &_exits, _exits.control()->as_Region(), _first_return); 2393 } 2394 } 2395 2396 stop_and_kill_map(); // This CFG path dies here 2397 } 2398 2399 2400 //------------------------------add_safepoint---------------------------------- 2401 void Parse::add_safepoint() { 2402 uint parms = TypeFunc::Parms+1; 2403 2404 // Clear out dead values from the debug info. 2405 kill_dead_locals(); 2406 2407 // Clone the JVM State 2408 SafePointNode *sfpnt = new SafePointNode(parms, nullptr); 2409 2410 // Capture memory state BEFORE a SafePoint. Since we can block at a 2411 // SafePoint we need our GC state to be safe; i.e. we need all our current 2412 // write barriers (card marks) to not float down after the SafePoint so we 2413 // must read raw memory. Likewise we need all oop stores to match the card 2414 // marks. If deopt can happen, we need ALL stores (we need the correct JVM 2415 // state on a deopt). 2416 2417 // We do not need to WRITE the memory state after a SafePoint. The control 2418 // edge will keep card-marks and oop-stores from floating up from below a 2419 // SafePoint and our true dependency added here will keep them from floating 2420 // down below a SafePoint. 2421 2422 // Clone the current memory state 2423 Node* mem = MergeMemNode::make(map()->memory()); 2424 2425 mem = _gvn.transform(mem); 2426 2427 // Pass control through the safepoint 2428 sfpnt->init_req(TypeFunc::Control , control()); 2429 // Fix edges normally used by a call 2430 sfpnt->init_req(TypeFunc::I_O , top() ); 2431 sfpnt->init_req(TypeFunc::Memory , mem ); 2432 sfpnt->init_req(TypeFunc::ReturnAdr, top() ); 2433 sfpnt->init_req(TypeFunc::FramePtr , top() ); 2434 2435 // Create a node for the polling address 2436 Node *polladr; 2437 Node *thread = _gvn.transform(new ThreadLocalNode()); 2438 Node *polling_page_load_addr = _gvn.transform(basic_plus_adr(top(), thread, in_bytes(JavaThread::polling_page_offset()))); 2439 polladr = make_load(control(), polling_page_load_addr, TypeRawPtr::BOTTOM, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 2440 sfpnt->init_req(TypeFunc::Parms+0, _gvn.transform(polladr)); 2441 2442 // Fix up the JVM State edges 2443 add_safepoint_edges(sfpnt); 2444 Node *transformed_sfpnt = _gvn.transform(sfpnt); 2445 set_control(transformed_sfpnt); 2446 2447 // Provide an edge from root to safepoint. This makes the safepoint 2448 // appear useful until the parse has completed. 2449 if (transformed_sfpnt->is_SafePoint()) { 2450 assert(C->root() != nullptr, "Expect parse is still valid"); 2451 C->root()->add_prec(transformed_sfpnt); 2452 } 2453 } 2454 2455 #ifndef PRODUCT 2456 //------------------------show_parse_info-------------------------------------- 2457 void Parse::show_parse_info() { 2458 InlineTree* ilt = nullptr; 2459 if (C->ilt() != nullptr) { 2460 JVMState* caller_jvms = is_osr_parse() ? caller()->caller() : caller(); 2461 ilt = InlineTree::find_subtree_from_root(C->ilt(), caller_jvms, method()); 2462 } 2463 if (PrintCompilation && Verbose) { 2464 if (depth() == 1) { 2465 if( ilt->count_inlines() ) { 2466 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2467 ilt->count_inline_bcs()); 2468 tty->cr(); 2469 } 2470 } else { 2471 if (method()->is_synchronized()) tty->print("s"); 2472 if (method()->has_exception_handlers()) tty->print("!"); 2473 // Check this is not the final compiled version 2474 if (C->trap_can_recompile()) { 2475 tty->print("-"); 2476 } else { 2477 tty->print(" "); 2478 } 2479 method()->print_short_name(); 2480 if (is_osr_parse()) { 2481 tty->print(" @ %d", osr_bci()); 2482 } 2483 tty->print(" (%d bytes)",method()->code_size()); 2484 if (ilt->count_inlines()) { 2485 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2486 ilt->count_inline_bcs()); 2487 } 2488 tty->cr(); 2489 } 2490 } 2491 if (PrintOpto && (depth() == 1 || PrintOptoInlining)) { 2492 // Print that we succeeded; suppress this message on the first osr parse. 2493 2494 if (method()->is_synchronized()) tty->print("s"); 2495 if (method()->has_exception_handlers()) tty->print("!"); 2496 // Check this is not the final compiled version 2497 if (C->trap_can_recompile() && depth() == 1) { 2498 tty->print("-"); 2499 } else { 2500 tty->print(" "); 2501 } 2502 if( depth() != 1 ) { tty->print(" "); } // missing compile count 2503 for (int i = 1; i < depth(); ++i) { tty->print(" "); } 2504 method()->print_short_name(); 2505 if (is_osr_parse()) { 2506 tty->print(" @ %d", osr_bci()); 2507 } 2508 if (ilt->caller_bci() != -1) { 2509 tty->print(" @ %d", ilt->caller_bci()); 2510 } 2511 tty->print(" (%d bytes)",method()->code_size()); 2512 if (ilt->count_inlines()) { 2513 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2514 ilt->count_inline_bcs()); 2515 } 2516 tty->cr(); 2517 } 2518 } 2519 2520 2521 //------------------------------dump------------------------------------------- 2522 // Dump information associated with the bytecodes of current _method 2523 void Parse::dump() { 2524 if( method() != nullptr ) { 2525 // Iterate over bytecodes 2526 ciBytecodeStream iter(method()); 2527 for( Bytecodes::Code bc = iter.next(); bc != ciBytecodeStream::EOBC() ; bc = iter.next() ) { 2528 dump_bci( iter.cur_bci() ); 2529 tty->cr(); 2530 } 2531 } 2532 } 2533 2534 // Dump information associated with a byte code index, 'bci' 2535 void Parse::dump_bci(int bci) { 2536 // Output info on merge-points, cloning, and within _jsr..._ret 2537 // NYI 2538 tty->print(" bci:%d", bci); 2539 } 2540 2541 #endif