1 /* 2 * Copyright (c) 1997, 2014, 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 "oops/method.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/idealGraphPrinter.hpp" 31 #include "opto/locknode.hpp" 32 #include "opto/memnode.hpp" 33 #include "opto/parse.hpp" 34 #include "opto/rootnode.hpp" 35 #include "opto/runtime.hpp" 36 #include "runtime/arguments.hpp" 37 #include "runtime/handles.inline.hpp" 38 #include "runtime/sharedRuntime.hpp" 39 #include "utilities/copy.hpp" 40 41 // Static array so we can figure out which bytecodes stop us from compiling 42 // the most. Some of the non-static variables are needed in bytecodeInfo.cpp 43 // and eventually should be encapsulated in a proper class (gri 8/18/98). 44 45 int nodes_created = 0; 46 int methods_parsed = 0; 47 int methods_seen = 0; 48 int blocks_parsed = 0; 49 int blocks_seen = 0; 50 51 int explicit_null_checks_inserted = 0; 52 int explicit_null_checks_elided = 0; 53 int all_null_checks_found = 0, implicit_null_checks = 0; 54 int implicit_null_throws = 0; 55 56 int reclaim_idx = 0; 57 int reclaim_in = 0; 58 int reclaim_node = 0; 59 60 #ifndef PRODUCT 61 bool Parse::BytecodeParseHistogram::_initialized = false; 62 uint Parse::BytecodeParseHistogram::_bytecodes_parsed [Bytecodes::number_of_codes]; 63 uint Parse::BytecodeParseHistogram::_nodes_constructed[Bytecodes::number_of_codes]; 64 uint Parse::BytecodeParseHistogram::_nodes_transformed[Bytecodes::number_of_codes]; 65 uint Parse::BytecodeParseHistogram::_new_values [Bytecodes::number_of_codes]; 66 #endif 67 68 //------------------------------print_statistics------------------------------- 69 #ifndef PRODUCT 70 void Parse::print_statistics() { 71 tty->print_cr("--- Compiler Statistics ---"); 72 tty->print("Methods seen: %d Methods parsed: %d", methods_seen, methods_parsed); 73 tty->print(" Nodes created: %d", nodes_created); 74 tty->cr(); 75 if (methods_seen != methods_parsed) 76 tty->print_cr("Reasons for parse failures (NOT cumulative):"); 77 tty->print_cr("Blocks parsed: %d Blocks seen: %d", blocks_parsed, blocks_seen); 78 79 if( explicit_null_checks_inserted ) 80 tty->print_cr("%d original NULL checks - %d elided (%2d%%); optimizer leaves %d,", explicit_null_checks_inserted, explicit_null_checks_elided, (100*explicit_null_checks_elided)/explicit_null_checks_inserted, all_null_checks_found); 81 if( all_null_checks_found ) 82 tty->print_cr("%d made implicit (%2d%%)", implicit_null_checks, 83 (100*implicit_null_checks)/all_null_checks_found); 84 if( implicit_null_throws ) 85 tty->print_cr("%d implicit null exceptions at runtime", 86 implicit_null_throws); 87 88 if( PrintParseStatistics && BytecodeParseHistogram::initialized() ) { 89 BytecodeParseHistogram::print(); 90 } 91 } 92 #endif 93 94 //------------------------------ON STACK REPLACEMENT--------------------------- 95 96 // Construct a node which can be used to get incoming state for 97 // on stack replacement. 98 Node *Parse::fetch_interpreter_state(int index, 99 BasicType bt, 100 Node *local_addrs, 101 Node *local_addrs_base) { 102 Node *mem = memory(Compile::AliasIdxRaw); 103 Node *adr = basic_plus_adr( local_addrs_base, local_addrs, -index*wordSize ); 104 Node *ctl = control(); 105 106 // Very similar to LoadNode::make, except we handle un-aligned longs and 107 // doubles on Sparc. Intel can handle them just fine directly. 108 Node *l; 109 switch (bt) { // Signature is flattened 110 case T_INT: l = new (C) LoadINode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInt::INT, MemNode::unordered); break; 111 case T_FLOAT: l = new (C) LoadFNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::FLOAT, MemNode::unordered); break; 112 case T_ADDRESS: l = new (C) LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); break; 113 case T_OBJECT: l = new (C) LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInstPtr::BOTTOM, MemNode::unordered); break; 114 case T_LONG: 115 case T_DOUBLE: { 116 // Since arguments are in reverse order, the argument address 'adr' 117 // refers to the back half of the long/double. Recompute adr. 118 adr = basic_plus_adr(local_addrs_base, local_addrs, -(index+1)*wordSize); 119 if (Matcher::misaligned_doubles_ok) { 120 l = (bt == T_DOUBLE) 121 ? (Node*)new (C) LoadDNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::DOUBLE, MemNode::unordered) 122 : (Node*)new (C) LoadLNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeLong::LONG, MemNode::unordered); 123 } else { 124 l = (bt == T_DOUBLE) 125 ? (Node*)new (C) LoadD_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered) 126 : (Node*)new (C) LoadL_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered); 127 } 128 break; 129 } 130 default: ShouldNotReachHere(); 131 } 132 return _gvn.transform(l); 133 } 134 135 // Helper routine to prevent the interpreter from handing 136 // unexpected typestate to an OSR method. 137 // The Node l is a value newly dug out of the interpreter frame. 138 // The type is the type predicted by ciTypeFlow. Note that it is 139 // not a general type, but can only come from Type::get_typeflow_type. 140 // The safepoint is a map which will feed an uncommon trap. 141 Node* Parse::check_interpreter_type(Node* l, const Type* type, 142 SafePointNode* &bad_type_exit) { 143 144 const TypeOopPtr* tp = type->isa_oopptr(); 145 146 // TypeFlow may assert null-ness if a type appears unloaded. 147 if (type == TypePtr::NULL_PTR || 148 (tp != NULL && !tp->klass()->is_loaded())) { 149 // Value must be null, not a real oop. 150 Node* chk = _gvn.transform( new (C) CmpPNode(l, null()) ); 151 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 152 IfNode* iff = create_and_map_if(control(), tst, PROB_MAX, COUNT_UNKNOWN); 153 set_control(_gvn.transform( new (C) IfTrueNode(iff) )); 154 Node* bad_type = _gvn.transform( new (C) IfFalseNode(iff) ); 155 bad_type_exit->control()->add_req(bad_type); 156 l = null(); 157 } 158 159 // Typeflow can also cut off paths from the CFG, based on 160 // types which appear unloaded, or call sites which appear unlinked. 161 // When paths are cut off, values at later merge points can rise 162 // toward more specific classes. Make sure these specific classes 163 // are still in effect. 164 if (tp != NULL && tp->klass() != C->env()->Object_klass()) { 165 // TypeFlow asserted a specific object type. Value must have that type. 166 Node* bad_type_ctrl = NULL; 167 l = gen_checkcast(l, makecon(TypeKlassPtr::make(tp->klass())), &bad_type_ctrl); 168 bad_type_exit->control()->add_req(bad_type_ctrl); 169 } 170 171 BasicType bt_l = _gvn.type(l)->basic_type(); 172 BasicType bt_t = type->basic_type(); 173 assert(_gvn.type(l)->higher_equal(type), "must constrain OSR typestate"); 174 return l; 175 } 176 177 // Helper routine which sets up elements of the initial parser map when 178 // performing a parse for on stack replacement. Add values into map. 179 // The only parameter contains the address of a interpreter arguments. 180 void Parse::load_interpreter_state(Node* osr_buf) { 181 int index; 182 int max_locals = jvms()->loc_size(); 183 int max_stack = jvms()->stk_size(); 184 185 186 // Mismatch between method and jvms can occur since map briefly held 187 // an OSR entry state (which takes up one RawPtr word). 188 assert(max_locals == method()->max_locals(), "sanity"); 189 assert(max_stack >= method()->max_stack(), "sanity"); 190 assert((int)jvms()->endoff() == TypeFunc::Parms + max_locals + max_stack, "sanity"); 191 assert((int)jvms()->endoff() == (int)map()->req(), "sanity"); 192 193 // Find the start block. 194 Block* osr_block = start_block(); 195 assert(osr_block->start() == osr_bci(), "sanity"); 196 197 // Set initial BCI. 198 set_parse_bci(osr_block->start()); 199 200 // Set initial stack depth. 201 set_sp(osr_block->start_sp()); 202 203 // Check bailouts. We currently do not perform on stack replacement 204 // of loops in catch blocks or loops which branch with a non-empty stack. 205 if (sp() != 0) { 206 C->record_method_not_compilable("OSR starts with non-empty stack"); 207 return; 208 } 209 // Do not OSR inside finally clauses: 210 if (osr_block->has_trap_at(osr_block->start())) { 211 C->record_method_not_compilable("OSR starts with an immediate trap"); 212 return; 213 } 214 215 // Commute monitors from interpreter frame to compiler frame. 216 assert(jvms()->monitor_depth() == 0, "should be no active locks at beginning of osr"); 217 int mcnt = osr_block->flow()->monitor_count(); 218 Node *monitors_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals+mcnt*2-1)*wordSize); 219 for (index = 0; index < mcnt; index++) { 220 // Make a BoxLockNode for the monitor. 221 Node *box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 222 223 224 // Displaced headers and locked objects are interleaved in the 225 // temp OSR buffer. We only copy the locked objects out here. 226 // Fetch the locked object from the OSR temp buffer and copy to our fastlock node. 227 Node *lock_object = fetch_interpreter_state(index*2, T_OBJECT, monitors_addr, osr_buf); 228 // Try and copy the displaced header to the BoxNode 229 Node *displaced_hdr = fetch_interpreter_state((index*2) + 1, T_ADDRESS, monitors_addr, osr_buf); 230 231 232 store_to_memory(control(), box, displaced_hdr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 233 234 // Build a bogus FastLockNode (no code will be generated) and push the 235 // monitor into our debug info. 236 const FastLockNode *flock = _gvn.transform(new (C) FastLockNode( 0, lock_object, box ))->as_FastLock(); 237 map()->push_monitor(flock); 238 239 // If the lock is our method synchronization lock, tuck it away in 240 // _sync_lock for return and rethrow exit paths. 241 if (index == 0 && method()->is_synchronized()) { 242 _synch_lock = flock; 243 } 244 } 245 246 // Use the raw liveness computation to make sure that unexpected 247 // values don't propagate into the OSR frame. 248 MethodLivenessResult live_locals = method()->liveness_at_bci(osr_bci()); 249 if (!live_locals.is_valid()) { 250 // Degenerate or breakpointed method. 251 C->record_method_not_compilable("OSR in empty or breakpointed method"); 252 return; 253 } 254 255 // Extract the needed locals from the interpreter frame. 256 Node *locals_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals-1)*wordSize); 257 258 // find all the locals that the interpreter thinks contain live oops 259 const BitMap live_oops = method()->live_local_oops_at_bci(osr_bci()); 260 for (index = 0; index < max_locals; index++) { 261 262 if (!live_locals.at(index)) { 263 continue; 264 } 265 266 const Type *type = osr_block->local_type_at(index); 267 268 if (type->isa_oopptr() != NULL) { 269 270 // 6403625: Verify that the interpreter oopMap thinks that the oop is live 271 // else we might load a stale oop if the MethodLiveness disagrees with the 272 // result of the interpreter. If the interpreter says it is dead we agree 273 // by making the value go to top. 274 // 275 276 if (!live_oops.at(index)) { 277 if (C->log() != NULL) { 278 C->log()->elem("OSR_mismatch local_index='%d'",index); 279 } 280 set_local(index, null()); 281 // and ignore it for the loads 282 continue; 283 } 284 } 285 286 // Filter out TOP, HALF, and BOTTOM. (Cf. ensure_phi.) 287 if (type == Type::TOP || type == Type::HALF) { 288 continue; 289 } 290 // If the type falls to bottom, then this must be a local that 291 // is mixing ints and oops or some such. Forcing it to top 292 // makes it go dead. 293 if (type == Type::BOTTOM) { 294 continue; 295 } 296 // Construct code to access the appropriate local. 297 BasicType bt = type->basic_type(); 298 if (type == TypePtr::NULL_PTR) { 299 // Ptr types are mixed together with T_ADDRESS but NULL is 300 // really for T_OBJECT types so correct it. 301 bt = T_OBJECT; 302 } 303 Node *value = fetch_interpreter_state(index, bt, locals_addr, osr_buf); 304 set_local(index, value); 305 } 306 307 // Extract the needed stack entries from the interpreter frame. 308 for (index = 0; index < sp(); index++) { 309 const Type *type = osr_block->stack_type_at(index); 310 if (type != Type::TOP) { 311 // Currently the compiler bails out when attempting to on stack replace 312 // at a bci with a non-empty stack. We should not reach here. 313 ShouldNotReachHere(); 314 } 315 } 316 317 // End the OSR migration 318 make_runtime_call(RC_LEAF, OptoRuntime::osr_end_Type(), 319 CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_end), 320 "OSR_migration_end", TypeRawPtr::BOTTOM, 321 osr_buf); 322 323 // Now that the interpreter state is loaded, make sure it will match 324 // at execution time what the compiler is expecting now: 325 SafePointNode* bad_type_exit = clone_map(); 326 bad_type_exit->set_control(new (C) RegionNode(1)); 327 328 assert(osr_block->flow()->jsrs()->size() == 0, "should be no jsrs live at osr point"); 329 for (index = 0; index < max_locals; index++) { 330 if (stopped()) break; 331 Node* l = local(index); 332 if (l->is_top()) continue; // nothing here 333 const Type *type = osr_block->local_type_at(index); 334 if (type->isa_oopptr() != NULL) { 335 if (!live_oops.at(index)) { 336 // skip type check for dead oops 337 continue; 338 } 339 } 340 if (osr_block->flow()->local_type_at(index)->is_return_address()) { 341 // In our current system it's illegal for jsr addresses to be 342 // live into an OSR entry point because the compiler performs 343 // inlining of jsrs. ciTypeFlow has a bailout that detect this 344 // case and aborts the compile if addresses are live into an OSR 345 // entry point. Because of that we can assume that any address 346 // locals at the OSR entry point are dead. Method liveness 347 // isn't precise enought to figure out that they are dead in all 348 // cases so simply skip checking address locals all 349 // together. Any type check is guaranteed to fail since the 350 // interpreter type is the result of a load which might have any 351 // value and the expected type is a constant. 352 continue; 353 } 354 set_local(index, check_interpreter_type(l, type, bad_type_exit)); 355 } 356 357 for (index = 0; index < sp(); index++) { 358 if (stopped()) break; 359 Node* l = stack(index); 360 if (l->is_top()) continue; // nothing here 361 const Type *type = osr_block->stack_type_at(index); 362 set_stack(index, check_interpreter_type(l, type, bad_type_exit)); 363 } 364 365 if (bad_type_exit->control()->req() > 1) { 366 // Build an uncommon trap here, if any inputs can be unexpected. 367 bad_type_exit->set_control(_gvn.transform( bad_type_exit->control() )); 368 record_for_igvn(bad_type_exit->control()); 369 SafePointNode* types_are_good = map(); 370 set_map(bad_type_exit); 371 // The unexpected type happens because a new edge is active 372 // in the CFG, which typeflow had previously ignored. 373 // E.g., Object x = coldAtFirst() && notReached()? "str": new Integer(123). 374 // This x will be typed as Integer if notReached is not yet linked. 375 // It could also happen due to a problem in ciTypeFlow analysis. 376 uncommon_trap(Deoptimization::Reason_constraint, 377 Deoptimization::Action_reinterpret); 378 set_map(types_are_good); 379 } 380 } 381 382 //------------------------------Parse------------------------------------------ 383 // Main parser constructor. 384 Parse::Parse(JVMState* caller, ciMethod* parse_method, float expected_uses) 385 : _exits(caller) 386 { 387 // Init some variables 388 _caller = caller; 389 _method = parse_method; 390 _expected_uses = expected_uses; 391 _depth = 1 + (caller->has_method() ? caller->depth() : 0); 392 _wrote_final = false; 393 _wrote_volatile = false; 394 _alloc_with_final = NULL; 395 _entry_bci = InvocationEntryBci; 396 _tf = NULL; 397 _block = NULL; 398 _first_return = true; 399 _replaced_nodes_for_exceptions = false; 400 _new_idx = C->unique(); 401 debug_only(_block_count = -1); 402 debug_only(_blocks = (Block*)-1); 403 #ifndef PRODUCT 404 if (PrintCompilation || PrintOpto) { 405 // Make sure I have an inline tree, so I can print messages about it. 406 JVMState* ilt_caller = is_osr_parse() ? caller->caller() : caller; 407 InlineTree::find_subtree_from_root(C->ilt(), ilt_caller, parse_method); 408 } 409 _max_switch_depth = 0; 410 _est_switch_depth = 0; 411 #endif 412 413 _tf = TypeFunc::make(method()); 414 _iter.reset_to_method(method()); 415 _flow = method()->get_flow_analysis(); 416 if (_flow->failing()) { 417 C->record_method_not_compilable_all_tiers(_flow->failure_reason()); 418 } 419 420 #ifndef PRODUCT 421 if (_flow->has_irreducible_entry()) { 422 C->set_parsed_irreducible_loop(true); 423 } 424 #endif 425 426 if (_expected_uses <= 0) { 427 _prof_factor = 1; 428 } else { 429 float prof_total = parse_method->interpreter_invocation_count(); 430 if (prof_total <= _expected_uses) { 431 _prof_factor = 1; 432 } else { 433 _prof_factor = _expected_uses / prof_total; 434 } 435 } 436 437 CompileLog* log = C->log(); 438 if (log != NULL) { 439 log->begin_head("parse method='%d' uses='%g'", 440 log->identify(parse_method), expected_uses); 441 if (depth() == 1 && C->is_osr_compilation()) { 442 log->print(" osr_bci='%d'", C->entry_bci()); 443 } 444 log->stamp(); 445 log->end_head(); 446 } 447 448 // Accumulate deoptimization counts. 449 // (The range_check and store_check counts are checked elsewhere.) 450 ciMethodData* md = method()->method_data(); 451 for (uint reason = 0; reason < md->trap_reason_limit(); reason++) { 452 uint md_count = md->trap_count(reason); 453 if (md_count != 0) { 454 if (md_count == md->trap_count_limit()) 455 md_count += md->overflow_trap_count(); 456 uint total_count = C->trap_count(reason); 457 uint old_count = total_count; 458 total_count += md_count; 459 // Saturate the add if it overflows. 460 if (total_count < old_count || total_count < md_count) 461 total_count = (uint)-1; 462 C->set_trap_count(reason, total_count); 463 if (log != NULL) 464 log->elem("observe trap='%s' count='%d' total='%d'", 465 Deoptimization::trap_reason_name(reason), 466 md_count, total_count); 467 } 468 } 469 // Accumulate total sum of decompilations, also. 470 C->set_decompile_count(C->decompile_count() + md->decompile_count()); 471 472 _count_invocations = C->do_count_invocations(); 473 _method_data_update = C->do_method_data_update(); 474 475 if (log != NULL && method()->has_exception_handlers()) { 476 log->elem("observe that='has_exception_handlers'"); 477 } 478 479 assert(method()->can_be_compiled(), "Can not parse this method, cutout earlier"); 480 assert(method()->has_balanced_monitors(), "Can not parse unbalanced monitors, cutout earlier"); 481 482 // Always register dependence if JVMTI is enabled, because 483 // either breakpoint setting or hotswapping of methods may 484 // cause deoptimization. 485 if (C->env()->jvmti_can_hotswap_or_post_breakpoint()) { 486 C->dependencies()->assert_evol_method(method()); 487 } 488 489 methods_seen++; 490 491 // Do some special top-level things. 492 if (depth() == 1 && C->is_osr_compilation()) { 493 _entry_bci = C->entry_bci(); 494 _flow = method()->get_osr_flow_analysis(osr_bci()); 495 if (_flow->failing()) { 496 C->record_method_not_compilable(_flow->failure_reason()); 497 #ifndef PRODUCT 498 if (PrintOpto && (Verbose || WizardMode)) { 499 tty->print_cr("OSR @%d type flow bailout: %s", _entry_bci, _flow->failure_reason()); 500 if (Verbose) { 501 method()->print(); 502 method()->print_codes(); 503 _flow->print(); 504 } 505 } 506 #endif 507 } 508 _tf = C->tf(); // the OSR entry type is different 509 } 510 511 #ifdef ASSERT 512 if (depth() == 1) { 513 assert(C->is_osr_compilation() == this->is_osr_parse(), "OSR in sync"); 514 if (C->tf() != tf()) { 515 MutexLockerEx ml(Compile_lock, Mutex::_no_safepoint_check_flag); 516 assert(C->env()->system_dictionary_modification_counter_changed(), 517 "Must invalidate if TypeFuncs differ"); 518 } 519 } else { 520 assert(!this->is_osr_parse(), "no recursive OSR"); 521 } 522 #endif 523 524 methods_parsed++; 525 #ifndef PRODUCT 526 // add method size here to guarantee that inlined methods are added too 527 if (TimeCompiler) 528 _total_bytes_compiled += method()->code_size(); 529 530 show_parse_info(); 531 #endif 532 533 if (failing()) { 534 if (log) log->done("parse"); 535 return; 536 } 537 538 gvn().set_type(root(), root()->bottom_type()); 539 gvn().transform(top()); 540 541 // Import the results of the ciTypeFlow. 542 init_blocks(); 543 544 // Merge point for all normal exits 545 build_exits(); 546 547 // Setup the initial JVM state map. 548 SafePointNode* entry_map = create_entry_map(); 549 550 // Check for bailouts during map initialization 551 if (failing() || entry_map == NULL) { 552 if (log) log->done("parse"); 553 return; 554 } 555 556 Node_Notes* caller_nn = C->default_node_notes(); 557 // Collect debug info for inlined calls unless -XX:-DebugInlinedCalls. 558 if (DebugInlinedCalls || depth() == 1) { 559 C->set_default_node_notes(make_node_notes(caller_nn)); 560 } 561 562 if (is_osr_parse()) { 563 Node* osr_buf = entry_map->in(TypeFunc::Parms+0); 564 entry_map->set_req(TypeFunc::Parms+0, top()); 565 set_map(entry_map); 566 load_interpreter_state(osr_buf); 567 } else { 568 set_map(entry_map); 569 do_method_entry(); 570 } 571 572 if (depth() == 1 && !failing()) { 573 // Add check to deoptimize the nmethod if RTM state was changed 574 rtm_deopt(); 575 } 576 577 // Check for bailouts during method entry or RTM state check setup. 578 if (failing()) { 579 if (log) log->done("parse"); 580 C->set_default_node_notes(caller_nn); 581 return; 582 } 583 584 entry_map = map(); // capture any changes performed by method setup code 585 assert(jvms()->endoff() == map()->req(), "map matches JVMS layout"); 586 587 // We begin parsing as if we have just encountered a jump to the 588 // method entry. 589 Block* entry_block = start_block(); 590 assert(entry_block->start() == (is_osr_parse() ? osr_bci() : 0), ""); 591 set_map_clone(entry_map); 592 merge_common(entry_block, entry_block->next_path_num()); 593 594 #ifndef PRODUCT 595 BytecodeParseHistogram *parse_histogram_obj = new (C->env()->arena()) BytecodeParseHistogram(this, C); 596 set_parse_histogram( parse_histogram_obj ); 597 #endif 598 599 // Parse all the basic blocks. 600 do_all_blocks(); 601 602 C->set_default_node_notes(caller_nn); 603 604 // Check for bailouts during conversion to graph 605 if (failing()) { 606 if (log) log->done("parse"); 607 return; 608 } 609 610 // Fix up all exiting control flow. 611 set_map(entry_map); 612 do_exits(); 613 614 if (log) log->done("parse nodes='%d' live='%d' memory='" SIZE_FORMAT "'", 615 C->unique(), C->live_nodes(), C->node_arena()->used()); 616 } 617 618 //---------------------------do_all_blocks------------------------------------- 619 void Parse::do_all_blocks() { 620 bool has_irreducible = flow()->has_irreducible_entry(); 621 622 // Walk over all blocks in Reverse Post-Order. 623 while (true) { 624 bool progress = false; 625 for (int rpo = 0; rpo < block_count(); rpo++) { 626 Block* block = rpo_at(rpo); 627 628 if (block->is_parsed()) continue; 629 630 if (!block->is_merged()) { 631 // Dead block, no state reaches this block 632 continue; 633 } 634 635 // Prepare to parse this block. 636 load_state_from(block); 637 638 if (stopped()) { 639 // Block is dead. 640 continue; 641 } 642 643 blocks_parsed++; 644 645 progress = true; 646 if (block->is_loop_head() || block->is_handler() || has_irreducible && !block->is_ready()) { 647 // Not all preds have been parsed. We must build phis everywhere. 648 // (Note that dead locals do not get phis built, ever.) 649 ensure_phis_everywhere(); 650 651 if (block->is_SEL_head() && 652 (UseLoopPredicate || LoopLimitCheck)) { 653 // Add predicate to single entry (not irreducible) loop head. 654 assert(!block->has_merged_backedge(), "only entry paths should be merged for now"); 655 // Need correct bci for predicate. 656 // It is fine to set it here since do_one_block() will set it anyway. 657 set_parse_bci(block->start()); 658 add_predicate(); 659 // Add new region for back branches. 660 int edges = block->pred_count() - block->preds_parsed() + 1; // +1 for original region 661 RegionNode *r = new (C) RegionNode(edges+1); 662 _gvn.set_type(r, Type::CONTROL); 663 record_for_igvn(r); 664 r->init_req(edges, control()); 665 set_control(r); 666 // Add new phis. 667 ensure_phis_everywhere(); 668 } 669 670 // Leave behind an undisturbed copy of the map, for future merges. 671 set_map(clone_map()); 672 } 673 674 if (control()->is_Region() && !block->is_loop_head() && !has_irreducible && !block->is_handler()) { 675 // In the absence of irreducible loops, the Region and Phis 676 // associated with a merge that doesn't involve a backedge can 677 // be simplified now since the RPO parsing order guarantees 678 // that any path which was supposed to reach here has already 679 // been parsed or must be dead. 680 Node* c = control(); 681 Node* result = _gvn.transform_no_reclaim(control()); 682 if (c != result && TraceOptoParse) { 683 tty->print_cr("Block #%d replace %d with %d", block->rpo(), c->_idx, result->_idx); 684 } 685 if (result != top()) { 686 record_for_igvn(result); 687 } 688 } 689 690 // Parse the block. 691 do_one_block(); 692 693 // Check for bailouts. 694 if (failing()) return; 695 } 696 697 // with irreducible loops multiple passes might be necessary to parse everything 698 if (!has_irreducible || !progress) { 699 break; 700 } 701 } 702 703 blocks_seen += block_count(); 704 705 #ifndef PRODUCT 706 // Make sure there are no half-processed blocks remaining. 707 // Every remaining unprocessed block is dead and may be ignored now. 708 for (int rpo = 0; rpo < block_count(); rpo++) { 709 Block* block = rpo_at(rpo); 710 if (!block->is_parsed()) { 711 if (TraceOptoParse) { 712 tty->print_cr("Skipped dead block %d at bci:%d", rpo, block->start()); 713 } 714 assert(!block->is_merged(), "no half-processed blocks"); 715 } 716 } 717 #endif 718 } 719 720 //-------------------------------build_exits---------------------------------- 721 // Build normal and exceptional exit merge points. 722 void Parse::build_exits() { 723 // make a clone of caller to prevent sharing of side-effects 724 _exits.set_map(_exits.clone_map()); 725 _exits.clean_stack(_exits.sp()); 726 _exits.sync_jvms(); 727 728 RegionNode* region = new (C) RegionNode(1); 729 record_for_igvn(region); 730 gvn().set_type_bottom(region); 731 _exits.set_control(region); 732 733 // Note: iophi and memphi are not transformed until do_exits. 734 Node* iophi = new (C) PhiNode(region, Type::ABIO); 735 Node* memphi = new (C) PhiNode(region, Type::MEMORY, TypePtr::BOTTOM); 736 gvn().set_type_bottom(iophi); 737 gvn().set_type_bottom(memphi); 738 _exits.set_i_o(iophi); 739 _exits.set_all_memory(memphi); 740 741 // Add a return value to the exit state. (Do not push it yet.) 742 if (tf()->range()->cnt() > TypeFunc::Parms) { 743 const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms); 744 // Don't "bind" an unloaded return klass to the ret_phi. If the klass 745 // becomes loaded during the subsequent parsing, the loaded and unloaded 746 // types will not join when we transform and push in do_exits(). 747 const TypeOopPtr* ret_oop_type = ret_type->isa_oopptr(); 748 if (ret_oop_type && !ret_oop_type->klass()->is_loaded()) { 749 ret_type = TypeOopPtr::BOTTOM; 750 } 751 int ret_size = type2size[ret_type->basic_type()]; 752 Node* ret_phi = new (C) PhiNode(region, ret_type); 753 gvn().set_type_bottom(ret_phi); 754 _exits.ensure_stack(ret_size); 755 assert((int)(tf()->range()->cnt() - TypeFunc::Parms) == ret_size, "good tf range"); 756 assert(method()->return_type()->size() == ret_size, "tf agrees w/ method"); 757 _exits.set_argument(0, ret_phi); // here is where the parser finds it 758 // Note: ret_phi is not yet pushed, until do_exits. 759 } 760 } 761 762 763 //----------------------------build_start_state------------------------------- 764 // Construct a state which contains only the incoming arguments from an 765 // unknown caller. The method & bci will be NULL & InvocationEntryBci. 766 JVMState* Compile::build_start_state(StartNode* start, const TypeFunc* tf) { 767 int arg_size = tf->domain()->cnt(); 768 int max_size = MAX2(arg_size, (int)tf->range()->cnt()); 769 JVMState* jvms = new (this) JVMState(max_size - TypeFunc::Parms); 770 SafePointNode* map = new (this) SafePointNode(max_size, NULL); 771 record_for_igvn(map); 772 assert(arg_size == TypeFunc::Parms + (is_osr_compilation() ? 1 : method()->arg_size()), "correct arg_size"); 773 Node_Notes* old_nn = default_node_notes(); 774 if (old_nn != NULL && has_method()) { 775 Node_Notes* entry_nn = old_nn->clone(this); 776 JVMState* entry_jvms = new(this) JVMState(method(), old_nn->jvms()); 777 entry_jvms->set_offsets(0); 778 entry_jvms->set_bci(entry_bci()); 779 entry_nn->set_jvms(entry_jvms); 780 set_default_node_notes(entry_nn); 781 } 782 uint i; 783 for (i = 0; i < (uint)arg_size; i++) { 784 Node* parm = initial_gvn()->transform(new (this) ParmNode(start, i)); 785 map->init_req(i, parm); 786 // Record all these guys for later GVN. 787 record_for_igvn(parm); 788 } 789 for (; i < map->req(); i++) { 790 map->init_req(i, top()); 791 } 792 assert(jvms->argoff() == TypeFunc::Parms, "parser gets arguments here"); 793 set_default_node_notes(old_nn); 794 map->set_jvms(jvms); 795 jvms->set_map(map); 796 return jvms; 797 } 798 799 //-----------------------------make_node_notes--------------------------------- 800 Node_Notes* Parse::make_node_notes(Node_Notes* caller_nn) { 801 if (caller_nn == NULL) return NULL; 802 Node_Notes* nn = caller_nn->clone(C); 803 JVMState* caller_jvms = nn->jvms(); 804 JVMState* jvms = new (C) JVMState(method(), caller_jvms); 805 jvms->set_offsets(0); 806 jvms->set_bci(_entry_bci); 807 nn->set_jvms(jvms); 808 return nn; 809 } 810 811 812 //--------------------------return_values-------------------------------------- 813 void Compile::return_values(JVMState* jvms) { 814 GraphKit kit(jvms); 815 Node* ret = new (this) ReturnNode(TypeFunc::Parms, 816 kit.control(), 817 kit.i_o(), 818 kit.reset_memory(), 819 kit.frameptr(), 820 kit.returnadr()); 821 // Add zero or 1 return values 822 int ret_size = tf()->range()->cnt() - TypeFunc::Parms; 823 if (ret_size > 0) { 824 kit.inc_sp(-ret_size); // pop the return value(s) 825 kit.sync_jvms(); 826 ret->add_req(kit.argument(0)); 827 // Note: The second dummy edge is not needed by a ReturnNode. 828 } 829 // bind it to root 830 root()->add_req(ret); 831 record_for_igvn(ret); 832 initial_gvn()->transform_no_reclaim(ret); 833 } 834 835 //------------------------rethrow_exceptions----------------------------------- 836 // Bind all exception states in the list into a single RethrowNode. 837 void Compile::rethrow_exceptions(JVMState* jvms) { 838 GraphKit kit(jvms); 839 if (!kit.has_exceptions()) return; // nothing to generate 840 // Load my combined exception state into the kit, with all phis transformed: 841 SafePointNode* ex_map = kit.combine_and_pop_all_exception_states(); 842 Node* ex_oop = kit.use_exception_state(ex_map); 843 RethrowNode* exit = new (this) RethrowNode(kit.control(), 844 kit.i_o(), kit.reset_memory(), 845 kit.frameptr(), kit.returnadr(), 846 // like a return but with exception input 847 ex_oop); 848 // bind to root 849 root()->add_req(exit); 850 record_for_igvn(exit); 851 initial_gvn()->transform_no_reclaim(exit); 852 } 853 854 //---------------------------do_exceptions------------------------------------- 855 // Process exceptions arising from the current bytecode. 856 // Send caught exceptions to the proper handler within this method. 857 // Unhandled exceptions feed into _exit. 858 void Parse::do_exceptions() { 859 if (!has_exceptions()) return; 860 861 if (failing()) { 862 // Pop them all off and throw them away. 863 while (pop_exception_state() != NULL) ; 864 return; 865 } 866 867 PreserveJVMState pjvms(this, false); 868 869 SafePointNode* ex_map; 870 while ((ex_map = pop_exception_state()) != NULL) { 871 if (!method()->has_exception_handlers()) { 872 // Common case: Transfer control outward. 873 // Doing it this early allows the exceptions to common up 874 // even between adjacent method calls. 875 throw_to_exit(ex_map); 876 } else { 877 // Have to look at the exception first. 878 assert(stopped(), "catch_inline_exceptions trashes the map"); 879 catch_inline_exceptions(ex_map); 880 stop_and_kill_map(); // we used up this exception state; kill it 881 } 882 } 883 884 // We now return to our regularly scheduled program: 885 } 886 887 //---------------------------throw_to_exit------------------------------------- 888 // Merge the given map into an exception exit from this method. 889 // The exception exit will handle any unlocking of receiver. 890 // The ex_oop must be saved within the ex_map, unlike merge_exception. 891 void Parse::throw_to_exit(SafePointNode* ex_map) { 892 // Pop the JVMS to (a copy of) the caller. 893 GraphKit caller; 894 caller.set_map_clone(_caller->map()); 895 caller.set_bci(_caller->bci()); 896 caller.set_sp(_caller->sp()); 897 // Copy out the standard machine state: 898 for (uint i = 0; i < TypeFunc::Parms; i++) { 899 caller.map()->set_req(i, ex_map->in(i)); 900 } 901 if (ex_map->has_replaced_nodes()) { 902 _replaced_nodes_for_exceptions = true; 903 } 904 caller.map()->transfer_replaced_nodes_from(ex_map, _new_idx); 905 // ...and the exception: 906 Node* ex_oop = saved_ex_oop(ex_map); 907 SafePointNode* caller_ex_map = caller.make_exception_state(ex_oop); 908 // Finally, collect the new exception state in my exits: 909 _exits.add_exception_state(caller_ex_map); 910 } 911 912 //------------------------------do_exits--------------------------------------- 913 void Parse::do_exits() { 914 set_parse_bci(InvocationEntryBci); 915 916 // Now peephole on the return bits 917 Node* region = _exits.control(); 918 _exits.set_control(gvn().transform(region)); 919 920 Node* iophi = _exits.i_o(); 921 _exits.set_i_o(gvn().transform(iophi)); 922 923 // On PPC64, also add MemBarRelease for constructors which write 924 // volatile fields. As support_IRIW_for_not_multiple_copy_atomic_cpu 925 // is set on PPC64, no sync instruction is issued after volatile 926 // stores. We want to quarantee the same behaviour as on platforms 927 // with total store order, although this is not required by the Java 928 // memory model. So as with finals, we add a barrier here. 929 if (wrote_final() PPC64_ONLY(|| (wrote_volatile() && method()->is_initializer()))) { 930 // This method (which must be a constructor by the rules of Java) 931 // wrote a final. The effects of all initializations must be 932 // committed to memory before any code after the constructor 933 // publishes the reference to the newly constructor object. 934 // Rather than wait for the publication, we simply block the 935 // writes here. Rather than put a barrier on only those writes 936 // which are required to complete, we force all writes to complete. 937 // 938 // "All bets are off" unless the first publication occurs after a 939 // normal return from the constructor. We do not attempt to detect 940 // such unusual early publications. But no barrier is needed on 941 // exceptional returns, since they cannot publish normally. 942 // 943 _exits.insert_mem_bar(Op_MemBarRelease, alloc_with_final()); 944 #ifndef PRODUCT 945 if (PrintOpto && (Verbose || WizardMode)) { 946 method()->print_name(); 947 tty->print_cr(" writes finals and needs a memory barrier"); 948 } 949 #endif 950 } 951 952 for (MergeMemStream mms(_exits.merged_memory()); mms.next_non_empty(); ) { 953 // transform each slice of the original memphi: 954 mms.set_memory(_gvn.transform(mms.memory())); 955 } 956 957 if (tf()->range()->cnt() > TypeFunc::Parms) { 958 const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms); 959 Node* ret_phi = _gvn.transform( _exits.argument(0) ); 960 assert(_exits.control()->is_top() || !_gvn.type(ret_phi)->empty(), "return value must be well defined"); 961 _exits.push_node(ret_type->basic_type(), ret_phi); 962 } 963 964 // Note: Logic for creating and optimizing the ReturnNode is in Compile. 965 966 // Unlock along the exceptional paths. 967 // This is done late so that we can common up equivalent exceptions 968 // (e.g., null checks) arising from multiple points within this method. 969 // See GraphKit::add_exception_state, which performs the commoning. 970 bool do_synch = method()->is_synchronized() && GenerateSynchronizationCode; 971 972 // record exit from a method if compiled while Dtrace is turned on. 973 if (do_synch || C->env()->dtrace_method_probes() || _replaced_nodes_for_exceptions) { 974 // First move the exception list out of _exits: 975 GraphKit kit(_exits.transfer_exceptions_into_jvms()); 976 SafePointNode* normal_map = kit.map(); // keep this guy safe 977 // Now re-collect the exceptions into _exits: 978 SafePointNode* ex_map; 979 while ((ex_map = kit.pop_exception_state()) != NULL) { 980 Node* ex_oop = kit.use_exception_state(ex_map); 981 // Force the exiting JVM state to have this method at InvocationEntryBci. 982 // The exiting JVM state is otherwise a copy of the calling JVMS. 983 JVMState* caller = kit.jvms(); 984 JVMState* ex_jvms = caller->clone_shallow(C); 985 ex_jvms->set_map(kit.clone_map()); 986 ex_jvms->map()->set_jvms(ex_jvms); 987 ex_jvms->set_bci( InvocationEntryBci); 988 kit.set_jvms(ex_jvms); 989 if (do_synch) { 990 // Add on the synchronized-method box/object combo 991 kit.map()->push_monitor(_synch_lock); 992 // Unlock! 993 kit.shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node()); 994 } 995 if (C->env()->dtrace_method_probes()) { 996 kit.make_dtrace_method_exit(method()); 997 } 998 if (_replaced_nodes_for_exceptions) { 999 kit.map()->apply_replaced_nodes(); 1000 } 1001 // Done with exception-path processing. 1002 ex_map = kit.make_exception_state(ex_oop); 1003 assert(ex_jvms->same_calls_as(ex_map->jvms()), "sanity"); 1004 // Pop the last vestige of this method: 1005 ex_map->set_jvms(caller->clone_shallow(C)); 1006 ex_map->jvms()->set_map(ex_map); 1007 _exits.push_exception_state(ex_map); 1008 } 1009 assert(_exits.map() == normal_map, "keep the same return state"); 1010 } 1011 1012 { 1013 // Capture very early exceptions (receiver null checks) from caller JVMS 1014 GraphKit caller(_caller); 1015 SafePointNode* ex_map; 1016 while ((ex_map = caller.pop_exception_state()) != NULL) { 1017 _exits.add_exception_state(ex_map); 1018 } 1019 } 1020 _exits.map()->apply_replaced_nodes(); 1021 } 1022 1023 //-----------------------------create_entry_map------------------------------- 1024 // Initialize our parser map to contain the types at method entry. 1025 // For OSR, the map contains a single RawPtr parameter. 1026 // Initial monitor locking for sync. methods is performed by do_method_entry. 1027 SafePointNode* Parse::create_entry_map() { 1028 // Check for really stupid bail-out cases. 1029 uint len = TypeFunc::Parms + method()->max_locals() + method()->max_stack(); 1030 if (len >= 32760) { 1031 C->record_method_not_compilable_all_tiers("too many local variables"); 1032 return NULL; 1033 } 1034 1035 // clear current replaced nodes that are of no use from here on (map was cloned in build_exits). 1036 _caller->map()->delete_replaced_nodes(); 1037 1038 // If this is an inlined method, we may have to do a receiver null check. 1039 if (_caller->has_method() && is_normal_parse() && !method()->is_static()) { 1040 GraphKit kit(_caller); 1041 kit.null_check_receiver_before_call(method()); 1042 _caller = kit.transfer_exceptions_into_jvms(); 1043 if (kit.stopped()) { 1044 _exits.add_exception_states_from(_caller); 1045 _exits.set_jvms(_caller); 1046 return NULL; 1047 } 1048 } 1049 1050 assert(method() != NULL, "parser must have a method"); 1051 1052 // Create an initial safepoint to hold JVM state during parsing 1053 JVMState* jvms = new (C) JVMState(method(), _caller->has_method() ? _caller : NULL); 1054 set_map(new (C) SafePointNode(len, jvms)); 1055 jvms->set_map(map()); 1056 record_for_igvn(map()); 1057 assert(jvms->endoff() == len, "correct jvms sizing"); 1058 1059 SafePointNode* inmap = _caller->map(); 1060 assert(inmap != NULL, "must have inmap"); 1061 // In case of null check on receiver above 1062 map()->transfer_replaced_nodes_from(inmap, _new_idx); 1063 1064 uint i; 1065 1066 // Pass thru the predefined input parameters. 1067 for (i = 0; i < TypeFunc::Parms; i++) { 1068 map()->init_req(i, inmap->in(i)); 1069 } 1070 1071 if (depth() == 1) { 1072 assert(map()->memory()->Opcode() == Op_Parm, ""); 1073 // Insert the memory aliasing node 1074 set_all_memory(reset_memory()); 1075 } 1076 assert(merged_memory(), ""); 1077 1078 // Now add the locals which are initially bound to arguments: 1079 uint arg_size = tf()->domain()->cnt(); 1080 ensure_stack(arg_size - TypeFunc::Parms); // OSR methods have funny args 1081 for (i = TypeFunc::Parms; i < arg_size; i++) { 1082 map()->init_req(i, inmap->argument(_caller, i - TypeFunc::Parms)); 1083 } 1084 1085 // Clear out the rest of the map (locals and stack) 1086 for (i = arg_size; i < len; i++) { 1087 map()->init_req(i, top()); 1088 } 1089 1090 SafePointNode* entry_map = stop(); 1091 return entry_map; 1092 } 1093 1094 //-----------------------------do_method_entry-------------------------------- 1095 // Emit any code needed in the pseudo-block before BCI zero. 1096 // The main thing to do is lock the receiver of a synchronized method. 1097 void Parse::do_method_entry() { 1098 set_parse_bci(InvocationEntryBci); // Pseudo-BCP 1099 set_sp(0); // Java Stack Pointer 1100 1101 NOT_PRODUCT( count_compiled_calls(true/*at_method_entry*/, false/*is_inline*/); ) 1102 1103 if (C->env()->dtrace_method_probes()) { 1104 make_dtrace_method_entry(method()); 1105 } 1106 1107 // If the method is synchronized, we need to construct a lock node, attach 1108 // it to the Start node, and pin it there. 1109 if (method()->is_synchronized()) { 1110 // Insert a FastLockNode right after the Start which takes as arguments 1111 // the current thread pointer, the "this" pointer & the address of the 1112 // stack slot pair used for the lock. The "this" pointer is a projection 1113 // off the start node, but the locking spot has to be constructed by 1114 // creating a ConLNode of 0, and boxing it with a BoxLockNode. The BoxLockNode 1115 // becomes the second argument to the FastLockNode call. The 1116 // FastLockNode becomes the new control parent to pin it to the start. 1117 1118 // Setup Object Pointer 1119 Node *lock_obj = NULL; 1120 if(method()->is_static()) { 1121 ciInstance* mirror = _method->holder()->java_mirror(); 1122 const TypeInstPtr *t_lock = TypeInstPtr::make(mirror); 1123 lock_obj = makecon(t_lock); 1124 } else { // Else pass the "this" pointer, 1125 lock_obj = local(0); // which is Parm0 from StartNode 1126 } 1127 // Clear out dead values from the debug info. 1128 kill_dead_locals(); 1129 // Build the FastLockNode 1130 _synch_lock = shared_lock(lock_obj); 1131 } 1132 1133 // Feed profiling data for parameters to the type system so it can 1134 // propagate it as speculative types 1135 record_profiled_parameters_for_speculation(); 1136 1137 if (depth() == 1) { 1138 increment_and_test_invocation_counter(Tier2CompileThreshold); 1139 } 1140 } 1141 1142 //------------------------------init_blocks------------------------------------ 1143 // Initialize our parser map to contain the types/monitors at method entry. 1144 void Parse::init_blocks() { 1145 // Create the blocks. 1146 _block_count = flow()->block_count(); 1147 _blocks = NEW_RESOURCE_ARRAY(Block, _block_count); 1148 Copy::zero_to_bytes(_blocks, sizeof(Block)*_block_count); 1149 1150 int rpo; 1151 1152 // Initialize the structs. 1153 for (rpo = 0; rpo < block_count(); rpo++) { 1154 Block* block = rpo_at(rpo); 1155 block->init_node(this, rpo); 1156 } 1157 1158 // Collect predecessor and successor information. 1159 for (rpo = 0; rpo < block_count(); rpo++) { 1160 Block* block = rpo_at(rpo); 1161 block->init_graph(this); 1162 } 1163 } 1164 1165 //-------------------------------init_node------------------------------------- 1166 void Parse::Block::init_node(Parse* outer, int rpo) { 1167 _flow = outer->flow()->rpo_at(rpo); 1168 _pred_count = 0; 1169 _preds_parsed = 0; 1170 _count = 0; 1171 assert(pred_count() == 0 && preds_parsed() == 0, "sanity"); 1172 assert(!(is_merged() || is_parsed() || is_handler() || has_merged_backedge()), "sanity"); 1173 assert(_live_locals.size() == 0, "sanity"); 1174 1175 // entry point has additional predecessor 1176 if (flow()->is_start()) _pred_count++; 1177 assert(flow()->is_start() == (this == outer->start_block()), ""); 1178 } 1179 1180 //-------------------------------init_graph------------------------------------ 1181 void Parse::Block::init_graph(Parse* outer) { 1182 // Create the successor list for this parser block. 1183 GrowableArray<ciTypeFlow::Block*>* tfs = flow()->successors(); 1184 GrowableArray<ciTypeFlow::Block*>* tfe = flow()->exceptions(); 1185 int ns = tfs->length(); 1186 int ne = tfe->length(); 1187 _num_successors = ns; 1188 _all_successors = ns+ne; 1189 _successors = (ns+ne == 0) ? NULL : NEW_RESOURCE_ARRAY(Block*, ns+ne); 1190 int p = 0; 1191 for (int i = 0; i < ns+ne; i++) { 1192 ciTypeFlow::Block* tf2 = (i < ns) ? tfs->at(i) : tfe->at(i-ns); 1193 Block* block2 = outer->rpo_at(tf2->rpo()); 1194 _successors[i] = block2; 1195 1196 // Accumulate pred info for the other block, too. 1197 if (i < ns) { 1198 block2->_pred_count++; 1199 } else { 1200 block2->_is_handler = true; 1201 } 1202 1203 #ifdef ASSERT 1204 // A block's successors must be distinguishable by BCI. 1205 // That is, no bytecode is allowed to branch to two different 1206 // clones of the same code location. 1207 for (int j = 0; j < i; j++) { 1208 Block* block1 = _successors[j]; 1209 if (block1 == block2) continue; // duplicates are OK 1210 assert(block1->start() != block2->start(), "successors have unique bcis"); 1211 } 1212 #endif 1213 } 1214 1215 // Note: We never call next_path_num along exception paths, so they 1216 // never get processed as "ready". Also, the input phis of exception 1217 // handlers get specially processed, so that 1218 } 1219 1220 //---------------------------successor_for_bci--------------------------------- 1221 Parse::Block* Parse::Block::successor_for_bci(int bci) { 1222 for (int i = 0; i < all_successors(); i++) { 1223 Block* block2 = successor_at(i); 1224 if (block2->start() == bci) return block2; 1225 } 1226 // We can actually reach here if ciTypeFlow traps out a block 1227 // due to an unloaded class, and concurrently with compilation the 1228 // class is then loaded, so that a later phase of the parser is 1229 // able to see more of the bytecode CFG. Or, the flow pass and 1230 // the parser can have a minor difference of opinion about executability 1231 // of bytecodes. For example, "obj.field = null" is executable even 1232 // if the field's type is an unloaded class; the flow pass used to 1233 // make a trap for such code. 1234 return NULL; 1235 } 1236 1237 1238 //-----------------------------stack_type_at----------------------------------- 1239 const Type* Parse::Block::stack_type_at(int i) const { 1240 return get_type(flow()->stack_type_at(i)); 1241 } 1242 1243 1244 //-----------------------------local_type_at----------------------------------- 1245 const Type* Parse::Block::local_type_at(int i) const { 1246 // Make dead locals fall to bottom. 1247 if (_live_locals.size() == 0) { 1248 MethodLivenessResult live_locals = flow()->outer()->method()->liveness_at_bci(start()); 1249 // This bitmap can be zero length if we saw a breakpoint. 1250 // In such cases, pretend they are all live. 1251 ((Block*)this)->_live_locals = live_locals; 1252 } 1253 if (_live_locals.size() > 0 && !_live_locals.at(i)) 1254 return Type::BOTTOM; 1255 1256 return get_type(flow()->local_type_at(i)); 1257 } 1258 1259 1260 #ifndef PRODUCT 1261 1262 //----------------------------name_for_bc-------------------------------------- 1263 // helper method for BytecodeParseHistogram 1264 static const char* name_for_bc(int i) { 1265 return Bytecodes::is_defined(i) ? Bytecodes::name(Bytecodes::cast(i)) : "xxxunusedxxx"; 1266 } 1267 1268 //----------------------------BytecodeParseHistogram------------------------------------ 1269 Parse::BytecodeParseHistogram::BytecodeParseHistogram(Parse *p, Compile *c) { 1270 _parser = p; 1271 _compiler = c; 1272 if( ! _initialized ) { _initialized = true; reset(); } 1273 } 1274 1275 //----------------------------current_count------------------------------------ 1276 int Parse::BytecodeParseHistogram::current_count(BPHType bph_type) { 1277 switch( bph_type ) { 1278 case BPH_transforms: { return _parser->gvn().made_progress(); } 1279 case BPH_values: { return _parser->gvn().made_new_values(); } 1280 default: { ShouldNotReachHere(); return 0; } 1281 } 1282 } 1283 1284 //----------------------------initialized-------------------------------------- 1285 bool Parse::BytecodeParseHistogram::initialized() { return _initialized; } 1286 1287 //----------------------------reset-------------------------------------------- 1288 void Parse::BytecodeParseHistogram::reset() { 1289 int i = Bytecodes::number_of_codes; 1290 while (i-- > 0) { _bytecodes_parsed[i] = 0; _nodes_constructed[i] = 0; _nodes_transformed[i] = 0; _new_values[i] = 0; } 1291 } 1292 1293 //----------------------------set_initial_state-------------------------------- 1294 // Record info when starting to parse one bytecode 1295 void Parse::BytecodeParseHistogram::set_initial_state( Bytecodes::Code bc ) { 1296 if( PrintParseStatistics && !_parser->is_osr_parse() ) { 1297 _initial_bytecode = bc; 1298 _initial_node_count = _compiler->unique(); 1299 _initial_transforms = current_count(BPH_transforms); 1300 _initial_values = current_count(BPH_values); 1301 } 1302 } 1303 1304 //----------------------------record_change-------------------------------- 1305 // Record results of parsing one bytecode 1306 void Parse::BytecodeParseHistogram::record_change() { 1307 if( PrintParseStatistics && !_parser->is_osr_parse() ) { 1308 ++_bytecodes_parsed[_initial_bytecode]; 1309 _nodes_constructed [_initial_bytecode] += (_compiler->unique() - _initial_node_count); 1310 _nodes_transformed [_initial_bytecode] += (current_count(BPH_transforms) - _initial_transforms); 1311 _new_values [_initial_bytecode] += (current_count(BPH_values) - _initial_values); 1312 } 1313 } 1314 1315 1316 //----------------------------print-------------------------------------------- 1317 void Parse::BytecodeParseHistogram::print(float cutoff) { 1318 ResourceMark rm; 1319 // print profile 1320 int total = 0; 1321 int i = 0; 1322 for( i = 0; i < Bytecodes::number_of_codes; ++i ) { total += _bytecodes_parsed[i]; } 1323 int abs_sum = 0; 1324 tty->cr(); //0123456789012345678901234567890123456789012345678901234567890123456789 1325 tty->print_cr("Histogram of %d parsed bytecodes:", total); 1326 if( total == 0 ) { return; } 1327 tty->cr(); 1328 tty->print_cr("absolute: count of compiled bytecodes of this type"); 1329 tty->print_cr("relative: percentage contribution to compiled nodes"); 1330 tty->print_cr("nodes : Average number of nodes constructed per bytecode"); 1331 tty->print_cr("rnodes : Significance towards total nodes constructed, (nodes*relative)"); 1332 tty->print_cr("transforms: Average amount of tranform progress per bytecode compiled"); 1333 tty->print_cr("values : Average number of node values improved per bytecode"); 1334 tty->print_cr("name : Bytecode name"); 1335 tty->cr(); 1336 tty->print_cr(" absolute relative nodes rnodes transforms values name"); 1337 tty->print_cr("----------------------------------------------------------------------"); 1338 while (--i > 0) { 1339 int abs = _bytecodes_parsed[i]; 1340 float rel = abs * 100.0F / total; 1341 float nodes = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _nodes_constructed[i])/_bytecodes_parsed[i]; 1342 float rnodes = _bytecodes_parsed[i] == 0 ? 0 : rel * nodes; 1343 float xforms = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _nodes_transformed[i])/_bytecodes_parsed[i]; 1344 float values = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _new_values [i])/_bytecodes_parsed[i]; 1345 if (cutoff <= rel) { 1346 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)); 1347 abs_sum += abs; 1348 } 1349 } 1350 tty->print_cr("----------------------------------------------------------------------"); 1351 float rel_sum = abs_sum * 100.0F / total; 1352 tty->print_cr("%10d %7.2f%% (cutoff = %.2f%%)", abs_sum, rel_sum, cutoff); 1353 tty->print_cr("----------------------------------------------------------------------"); 1354 tty->cr(); 1355 } 1356 #endif 1357 1358 //----------------------------load_state_from---------------------------------- 1359 // Load block/map/sp. But not do not touch iter/bci. 1360 void Parse::load_state_from(Block* block) { 1361 set_block(block); 1362 // load the block's JVM state: 1363 set_map(block->start_map()); 1364 set_sp( block->start_sp()); 1365 } 1366 1367 1368 //-----------------------------record_state------------------------------------ 1369 void Parse::Block::record_state(Parse* p) { 1370 assert(!is_merged(), "can only record state once, on 1st inflow"); 1371 assert(start_sp() == p->sp(), "stack pointer must agree with ciTypeFlow"); 1372 set_start_map(p->stop()); 1373 } 1374 1375 1376 //------------------------------do_one_block----------------------------------- 1377 void Parse::do_one_block() { 1378 if (TraceOptoParse) { 1379 Block *b = block(); 1380 int ns = b->num_successors(); 1381 int nt = b->all_successors(); 1382 1383 tty->print("Parsing block #%d at bci [%d,%d), successors: ", 1384 block()->rpo(), block()->start(), block()->limit()); 1385 for (int i = 0; i < nt; i++) { 1386 tty->print((( i < ns) ? " %d" : " %d(e)"), b->successor_at(i)->rpo()); 1387 } 1388 if (b->is_loop_head()) tty->print(" lphd"); 1389 tty->cr(); 1390 } 1391 1392 assert(block()->is_merged(), "must be merged before being parsed"); 1393 block()->mark_parsed(); 1394 ++_blocks_parsed; 1395 1396 // Set iterator to start of block. 1397 iter().reset_to_bci(block()->start()); 1398 1399 CompileLog* log = C->log(); 1400 1401 // Parse bytecodes 1402 while (!stopped() && !failing()) { 1403 iter().next(); 1404 1405 // Learn the current bci from the iterator: 1406 set_parse_bci(iter().cur_bci()); 1407 1408 if (bci() == block()->limit()) { 1409 // Do not walk into the next block until directed by do_all_blocks. 1410 merge(bci()); 1411 break; 1412 } 1413 assert(bci() < block()->limit(), "bci still in block"); 1414 1415 if (log != NULL) { 1416 // Output an optional context marker, to help place actions 1417 // that occur during parsing of this BC. If there is no log 1418 // output until the next context string, this context string 1419 // will be silently ignored. 1420 log->set_context("bc code='%d' bci='%d'", (int)bc(), bci()); 1421 } 1422 1423 if (block()->has_trap_at(bci())) { 1424 // We must respect the flow pass's traps, because it will refuse 1425 // to produce successors for trapping blocks. 1426 int trap_index = block()->flow()->trap_index(); 1427 assert(trap_index != 0, "trap index must be valid"); 1428 uncommon_trap(trap_index); 1429 break; 1430 } 1431 1432 NOT_PRODUCT( parse_histogram()->set_initial_state(bc()); ); 1433 1434 #ifdef ASSERT 1435 int pre_bc_sp = sp(); 1436 int inputs, depth; 1437 bool have_se = !stopped() && compute_stack_effects(inputs, depth); 1438 assert(!have_se || pre_bc_sp >= inputs, err_msg_res("have enough stack to execute this BC: pre_bc_sp=%d, inputs=%d", pre_bc_sp, inputs)); 1439 #endif //ASSERT 1440 1441 do_one_bytecode(); 1442 1443 assert(!have_se || stopped() || failing() || (sp() - pre_bc_sp) == depth, 1444 err_msg_res("incorrect depth prediction: sp=%d, pre_bc_sp=%d, depth=%d", sp(), pre_bc_sp, depth)); 1445 1446 do_exceptions(); 1447 1448 NOT_PRODUCT( parse_histogram()->record_change(); ); 1449 1450 if (log != NULL) 1451 log->clear_context(); // skip marker if nothing was printed 1452 1453 // Fall into next bytecode. Each bytecode normally has 1 sequential 1454 // successor which is typically made ready by visiting this bytecode. 1455 // If the successor has several predecessors, then it is a merge 1456 // point, starts a new basic block, and is handled like other basic blocks. 1457 } 1458 } 1459 1460 1461 //------------------------------merge------------------------------------------ 1462 void Parse::set_parse_bci(int bci) { 1463 set_bci(bci); 1464 Node_Notes* nn = C->default_node_notes(); 1465 if (nn == NULL) return; 1466 1467 // Collect debug info for inlined calls unless -XX:-DebugInlinedCalls. 1468 if (!DebugInlinedCalls && depth() > 1) { 1469 return; 1470 } 1471 1472 // Update the JVMS annotation, if present. 1473 JVMState* jvms = nn->jvms(); 1474 if (jvms != NULL && jvms->bci() != bci) { 1475 // Update the JVMS. 1476 jvms = jvms->clone_shallow(C); 1477 jvms->set_bci(bci); 1478 nn->set_jvms(jvms); 1479 } 1480 } 1481 1482 //------------------------------merge------------------------------------------ 1483 // Merge the current mapping into the basic block starting at bci 1484 void Parse::merge(int target_bci) { 1485 Block* target = successor_for_bci(target_bci); 1486 if (target == NULL) { handle_missing_successor(target_bci); return; } 1487 assert(!target->is_ready(), "our arrival must be expected"); 1488 int pnum = target->next_path_num(); 1489 merge_common(target, pnum); 1490 } 1491 1492 //-------------------------merge_new_path-------------------------------------- 1493 // Merge the current mapping into the basic block, using a new path 1494 void Parse::merge_new_path(int target_bci) { 1495 Block* target = successor_for_bci(target_bci); 1496 if (target == NULL) { handle_missing_successor(target_bci); return; } 1497 assert(!target->is_ready(), "new path into frozen graph"); 1498 int pnum = target->add_new_path(); 1499 merge_common(target, pnum); 1500 } 1501 1502 //-------------------------merge_exception------------------------------------- 1503 // Merge the current mapping into the basic block starting at bci 1504 // The ex_oop must be pushed on the stack, unlike throw_to_exit. 1505 void Parse::merge_exception(int target_bci) { 1506 assert(sp() == 1, "must have only the throw exception on the stack"); 1507 Block* target = successor_for_bci(target_bci); 1508 if (target == NULL) { handle_missing_successor(target_bci); return; } 1509 assert(target->is_handler(), "exceptions are handled by special blocks"); 1510 int pnum = target->add_new_path(); 1511 merge_common(target, pnum); 1512 } 1513 1514 //--------------------handle_missing_successor--------------------------------- 1515 void Parse::handle_missing_successor(int target_bci) { 1516 #ifndef PRODUCT 1517 Block* b = block(); 1518 int trap_bci = b->flow()->has_trap()? b->flow()->trap_bci(): -1; 1519 tty->print_cr("### Missing successor at bci:%d for block #%d (trap_bci:%d)", target_bci, b->rpo(), trap_bci); 1520 #endif 1521 ShouldNotReachHere(); 1522 } 1523 1524 //--------------------------merge_common--------------------------------------- 1525 void Parse::merge_common(Parse::Block* target, int pnum) { 1526 if (TraceOptoParse) { 1527 tty->print("Merging state at block #%d bci:%d", target->rpo(), target->start()); 1528 } 1529 1530 // Zap extra stack slots to top 1531 assert(sp() == target->start_sp(), ""); 1532 clean_stack(sp()); 1533 1534 if (!target->is_merged()) { // No prior mapping at this bci 1535 if (TraceOptoParse) { tty->print(" with empty state"); } 1536 1537 // If this path is dead, do not bother capturing it as a merge. 1538 // It is "as if" we had 1 fewer predecessors from the beginning. 1539 if (stopped()) { 1540 if (TraceOptoParse) tty->print_cr(", but path is dead and doesn't count"); 1541 return; 1542 } 1543 1544 // Record that a new block has been merged. 1545 ++_blocks_merged; 1546 1547 // Make a region if we know there are multiple or unpredictable inputs. 1548 // (Also, if this is a plain fall-through, we might see another region, 1549 // which must not be allowed into this block's map.) 1550 if (pnum > PhiNode::Input // Known multiple inputs. 1551 || target->is_handler() // These have unpredictable inputs. 1552 || target->is_loop_head() // Known multiple inputs 1553 || control()->is_Region()) { // We must hide this guy. 1554 1555 int current_bci = bci(); 1556 set_parse_bci(target->start()); // Set target bci 1557 if (target->is_SEL_head()) { 1558 DEBUG_ONLY( target->mark_merged_backedge(block()); ) 1559 if (target->start() == 0) { 1560 // Add loop predicate for the special case when 1561 // there are backbranches to the method entry. 1562 add_predicate(); 1563 } 1564 } 1565 // Add a Region to start the new basic block. Phis will be added 1566 // later lazily. 1567 int edges = target->pred_count(); 1568 if (edges < pnum) edges = pnum; // might be a new path! 1569 RegionNode *r = new (C) RegionNode(edges+1); 1570 gvn().set_type(r, Type::CONTROL); 1571 record_for_igvn(r); 1572 // zap all inputs to NULL for debugging (done in Node(uint) constructor) 1573 // for (int j = 1; j < edges+1; j++) { r->init_req(j, NULL); } 1574 r->init_req(pnum, control()); 1575 set_control(r); 1576 set_parse_bci(current_bci); // Restore bci 1577 } 1578 1579 // Convert the existing Parser mapping into a mapping at this bci. 1580 store_state_to(target); 1581 assert(target->is_merged(), "do not come here twice"); 1582 1583 } else { // Prior mapping at this bci 1584 if (TraceOptoParse) { tty->print(" with previous state"); } 1585 #ifdef ASSERT 1586 if (target->is_SEL_head()) { 1587 target->mark_merged_backedge(block()); 1588 } 1589 #endif 1590 // We must not manufacture more phis if the target is already parsed. 1591 bool nophi = target->is_parsed(); 1592 1593 SafePointNode* newin = map();// Hang on to incoming mapping 1594 Block* save_block = block(); // Hang on to incoming block; 1595 load_state_from(target); // Get prior mapping 1596 1597 assert(newin->jvms()->locoff() == jvms()->locoff(), "JVMS layouts agree"); 1598 assert(newin->jvms()->stkoff() == jvms()->stkoff(), "JVMS layouts agree"); 1599 assert(newin->jvms()->monoff() == jvms()->monoff(), "JVMS layouts agree"); 1600 assert(newin->jvms()->endoff() == jvms()->endoff(), "JVMS layouts agree"); 1601 1602 // Iterate over my current mapping and the old mapping. 1603 // Where different, insert Phi functions. 1604 // Use any existing Phi functions. 1605 assert(control()->is_Region(), "must be merging to a region"); 1606 RegionNode* r = control()->as_Region(); 1607 1608 // Compute where to merge into 1609 // Merge incoming control path 1610 r->init_req(pnum, newin->control()); 1611 1612 if (pnum == 1) { // Last merge for this Region? 1613 if (!block()->flow()->is_irreducible_entry()) { 1614 Node* result = _gvn.transform_no_reclaim(r); 1615 if (r != result && TraceOptoParse) { 1616 tty->print_cr("Block #%d replace %d with %d", block()->rpo(), r->_idx, result->_idx); 1617 } 1618 } 1619 record_for_igvn(r); 1620 } 1621 1622 // Update all the non-control inputs to map: 1623 assert(TypeFunc::Parms == newin->jvms()->locoff(), "parser map should contain only youngest jvms"); 1624 bool check_elide_phi = target->is_SEL_backedge(save_block); 1625 for (uint j = 1; j < newin->req(); j++) { 1626 Node* m = map()->in(j); // Current state of target. 1627 Node* n = newin->in(j); // Incoming change to target state. 1628 PhiNode* phi; 1629 if (m->is_Phi() && m->as_Phi()->region() == r) 1630 phi = m->as_Phi(); 1631 else 1632 phi = NULL; 1633 if (m != n) { // Different; must merge 1634 switch (j) { 1635 // Frame pointer and Return Address never changes 1636 case TypeFunc::FramePtr:// Drop m, use the original value 1637 case TypeFunc::ReturnAdr: 1638 break; 1639 case TypeFunc::Memory: // Merge inputs to the MergeMem node 1640 assert(phi == NULL, "the merge contains phis, not vice versa"); 1641 merge_memory_edges(n->as_MergeMem(), pnum, nophi); 1642 continue; 1643 default: // All normal stuff 1644 if (phi == NULL) { 1645 const JVMState* jvms = map()->jvms(); 1646 if (EliminateNestedLocks && 1647 jvms->is_mon(j) && jvms->is_monitor_box(j)) { 1648 // BoxLock nodes are not commoning. 1649 // Use old BoxLock node as merged box. 1650 assert(newin->jvms()->is_monitor_box(j), "sanity"); 1651 // This assert also tests that nodes are BoxLock. 1652 assert(BoxLockNode::same_slot(n, m), "sanity"); 1653 C->gvn_replace_by(n, m); 1654 } else if (!check_elide_phi || !target->can_elide_SEL_phi(j)) { 1655 phi = ensure_phi(j, nophi); 1656 } 1657 } 1658 break; 1659 } 1660 } 1661 // At this point, n might be top if: 1662 // - there is no phi (because TypeFlow detected a conflict), or 1663 // - the corresponding control edges is top (a dead incoming path) 1664 // It is a bug if we create a phi which sees a garbage value on a live path. 1665 1666 if (phi != NULL) { 1667 assert(n != top() || r->in(pnum) == top(), "live value must not be garbage"); 1668 assert(phi->region() == r, ""); 1669 phi->set_req(pnum, n); // Then add 'n' to the merge 1670 if (pnum == PhiNode::Input) { 1671 // Last merge for this Phi. 1672 // So far, Phis have had a reasonable type from ciTypeFlow. 1673 // Now _gvn will join that with the meet of current inputs. 1674 // BOTTOM is never permissible here, 'cause pessimistically 1675 // Phis of pointers cannot lose the basic pointer type. 1676 debug_only(const Type* bt1 = phi->bottom_type()); 1677 assert(bt1 != Type::BOTTOM, "should not be building conflict phis"); 1678 map()->set_req(j, _gvn.transform_no_reclaim(phi)); 1679 debug_only(const Type* bt2 = phi->bottom_type()); 1680 assert(bt2->higher_equal_speculative(bt1), "must be consistent with type-flow"); 1681 record_for_igvn(phi); 1682 } 1683 } 1684 } // End of for all values to be merged 1685 1686 if (pnum == PhiNode::Input && 1687 !r->in(0)) { // The occasional useless Region 1688 assert(control() == r, ""); 1689 set_control(r->nonnull_req()); 1690 } 1691 1692 map()->merge_replaced_nodes_with(newin); 1693 1694 // newin has been subsumed into the lazy merge, and is now dead. 1695 set_block(save_block); 1696 1697 stop(); // done with this guy, for now 1698 } 1699 1700 if (TraceOptoParse) { 1701 tty->print_cr(" on path %d", pnum); 1702 } 1703 1704 // Done with this parser state. 1705 assert(stopped(), ""); 1706 } 1707 1708 1709 //--------------------------merge_memory_edges--------------------------------- 1710 void Parse::merge_memory_edges(MergeMemNode* n, int pnum, bool nophi) { 1711 // (nophi means we must not create phis, because we already parsed here) 1712 assert(n != NULL, ""); 1713 // Merge the inputs to the MergeMems 1714 MergeMemNode* m = merged_memory(); 1715 1716 assert(control()->is_Region(), "must be merging to a region"); 1717 RegionNode* r = control()->as_Region(); 1718 1719 PhiNode* base = NULL; 1720 MergeMemNode* remerge = NULL; 1721 for (MergeMemStream mms(m, n); mms.next_non_empty2(); ) { 1722 Node *p = mms.force_memory(); 1723 Node *q = mms.memory2(); 1724 if (mms.is_empty() && nophi) { 1725 // Trouble: No new splits allowed after a loop body is parsed. 1726 // Instead, wire the new split into a MergeMem on the backedge. 1727 // The optimizer will sort it out, slicing the phi. 1728 if (remerge == NULL) { 1729 assert(base != NULL, ""); 1730 assert(base->in(0) != NULL, "should not be xformed away"); 1731 remerge = MergeMemNode::make(C, base->in(pnum)); 1732 gvn().set_type(remerge, Type::MEMORY); 1733 base->set_req(pnum, remerge); 1734 } 1735 remerge->set_memory_at(mms.alias_idx(), q); 1736 continue; 1737 } 1738 assert(!q->is_MergeMem(), ""); 1739 PhiNode* phi; 1740 if (p != q) { 1741 phi = ensure_memory_phi(mms.alias_idx(), nophi); 1742 } else { 1743 if (p->is_Phi() && p->as_Phi()->region() == r) 1744 phi = p->as_Phi(); 1745 else 1746 phi = NULL; 1747 } 1748 // Insert q into local phi 1749 if (phi != NULL) { 1750 assert(phi->region() == r, ""); 1751 p = phi; 1752 phi->set_req(pnum, q); 1753 if (mms.at_base_memory()) { 1754 base = phi; // delay transforming it 1755 } else if (pnum == 1) { 1756 record_for_igvn(phi); 1757 p = _gvn.transform_no_reclaim(phi); 1758 } 1759 mms.set_memory(p);// store back through the iterator 1760 } 1761 } 1762 // Transform base last, in case we must fiddle with remerging. 1763 if (base != NULL && pnum == 1) { 1764 record_for_igvn(base); 1765 m->set_base_memory( _gvn.transform_no_reclaim(base) ); 1766 } 1767 } 1768 1769 1770 //------------------------ensure_phis_everywhere------------------------------- 1771 void Parse::ensure_phis_everywhere() { 1772 ensure_phi(TypeFunc::I_O); 1773 1774 // Ensure a phi on all currently known memories. 1775 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 1776 ensure_memory_phi(mms.alias_idx()); 1777 debug_only(mms.set_memory()); // keep the iterator happy 1778 } 1779 1780 // Note: This is our only chance to create phis for memory slices. 1781 // If we miss a slice that crops up later, it will have to be 1782 // merged into the base-memory phi that we are building here. 1783 // Later, the optimizer will comb out the knot, and build separate 1784 // phi-loops for each memory slice that matters. 1785 1786 // Monitors must nest nicely and not get confused amongst themselves. 1787 // Phi-ify everything up to the monitors, though. 1788 uint monoff = map()->jvms()->monoff(); 1789 uint nof_monitors = map()->jvms()->nof_monitors(); 1790 1791 assert(TypeFunc::Parms == map()->jvms()->locoff(), "parser map should contain only youngest jvms"); 1792 bool check_elide_phi = block()->is_SEL_head(); 1793 for (uint i = TypeFunc::Parms; i < monoff; i++) { 1794 if (!check_elide_phi || !block()->can_elide_SEL_phi(i)) { 1795 ensure_phi(i); 1796 } 1797 } 1798 1799 // Even monitors need Phis, though they are well-structured. 1800 // This is true for OSR methods, and also for the rare cases where 1801 // a monitor object is the subject of a replace_in_map operation. 1802 // See bugs 4426707 and 5043395. 1803 for (uint m = 0; m < nof_monitors; m++) { 1804 ensure_phi(map()->jvms()->monitor_obj_offset(m)); 1805 } 1806 } 1807 1808 1809 //-----------------------------add_new_path------------------------------------ 1810 // Add a previously unaccounted predecessor to this block. 1811 int Parse::Block::add_new_path() { 1812 // If there is no map, return the lowest unused path number. 1813 if (!is_merged()) return pred_count()+1; // there will be a map shortly 1814 1815 SafePointNode* map = start_map(); 1816 if (!map->control()->is_Region()) 1817 return pred_count()+1; // there may be a region some day 1818 RegionNode* r = map->control()->as_Region(); 1819 1820 // Add new path to the region. 1821 uint pnum = r->req(); 1822 r->add_req(NULL); 1823 1824 for (uint i = 1; i < map->req(); i++) { 1825 Node* n = map->in(i); 1826 if (i == TypeFunc::Memory) { 1827 // Ensure a phi on all currently known memories. 1828 for (MergeMemStream mms(n->as_MergeMem()); mms.next_non_empty(); ) { 1829 Node* phi = mms.memory(); 1830 if (phi->is_Phi() && phi->as_Phi()->region() == r) { 1831 assert(phi->req() == pnum, "must be same size as region"); 1832 phi->add_req(NULL); 1833 } 1834 } 1835 } else { 1836 if (n->is_Phi() && n->as_Phi()->region() == r) { 1837 assert(n->req() == pnum, "must be same size as region"); 1838 n->add_req(NULL); 1839 } 1840 } 1841 } 1842 1843 return pnum; 1844 } 1845 1846 //------------------------------ensure_phi------------------------------------- 1847 // Turn the idx'th entry of the current map into a Phi 1848 PhiNode *Parse::ensure_phi(int idx, bool nocreate) { 1849 SafePointNode* map = this->map(); 1850 Node* region = map->control(); 1851 assert(region->is_Region(), ""); 1852 1853 Node* o = map->in(idx); 1854 assert(o != NULL, ""); 1855 1856 if (o == top()) return NULL; // TOP always merges into TOP 1857 1858 if (o->is_Phi() && o->as_Phi()->region() == region) { 1859 return o->as_Phi(); 1860 } 1861 1862 // Now use a Phi here for merging 1863 assert(!nocreate, "Cannot build a phi for a block already parsed."); 1864 const JVMState* jvms = map->jvms(); 1865 const Type* t; 1866 if (jvms->is_loc(idx)) { 1867 t = block()->local_type_at(idx - jvms->locoff()); 1868 } else if (jvms->is_stk(idx)) { 1869 t = block()->stack_type_at(idx - jvms->stkoff()); 1870 } else if (jvms->is_mon(idx)) { 1871 assert(!jvms->is_monitor_box(idx), "no phis for boxes"); 1872 t = TypeInstPtr::BOTTOM; // this is sufficient for a lock object 1873 } else if ((uint)idx < TypeFunc::Parms) { 1874 t = o->bottom_type(); // Type::RETURN_ADDRESS or such-like. 1875 } else { 1876 assert(false, "no type information for this phi"); 1877 } 1878 1879 // If the type falls to bottom, then this must be a local that 1880 // is mixing ints and oops or some such. Forcing it to top 1881 // makes it go dead. 1882 if (t == Type::BOTTOM) { 1883 map->set_req(idx, top()); 1884 return NULL; 1885 } 1886 1887 // Do not create phis for top either. 1888 // A top on a non-null control flow must be an unused even after the.phi. 1889 if (t == Type::TOP || t == Type::HALF) { 1890 map->set_req(idx, top()); 1891 return NULL; 1892 } 1893 1894 PhiNode* phi = PhiNode::make(region, o, t); 1895 gvn().set_type(phi, t); 1896 if (C->do_escape_analysis()) record_for_igvn(phi); 1897 map->set_req(idx, phi); 1898 return phi; 1899 } 1900 1901 //--------------------------ensure_memory_phi---------------------------------- 1902 // Turn the idx'th slice of the current memory into a Phi 1903 PhiNode *Parse::ensure_memory_phi(int idx, bool nocreate) { 1904 MergeMemNode* mem = merged_memory(); 1905 Node* region = control(); 1906 assert(region->is_Region(), ""); 1907 1908 Node *o = (idx == Compile::AliasIdxBot)? mem->base_memory(): mem->memory_at(idx); 1909 assert(o != NULL && o != top(), ""); 1910 1911 PhiNode* phi; 1912 if (o->is_Phi() && o->as_Phi()->region() == region) { 1913 phi = o->as_Phi(); 1914 if (phi == mem->base_memory() && idx >= Compile::AliasIdxRaw) { 1915 // clone the shared base memory phi to make a new memory split 1916 assert(!nocreate, "Cannot build a phi for a block already parsed."); 1917 const Type* t = phi->bottom_type(); 1918 const TypePtr* adr_type = C->get_adr_type(idx); 1919 phi = phi->slice_memory(adr_type); 1920 gvn().set_type(phi, t); 1921 } 1922 return phi; 1923 } 1924 1925 // Now use a Phi here for merging 1926 assert(!nocreate, "Cannot build a phi for a block already parsed."); 1927 const Type* t = o->bottom_type(); 1928 const TypePtr* adr_type = C->get_adr_type(idx); 1929 phi = PhiNode::make(region, o, t, adr_type); 1930 gvn().set_type(phi, t); 1931 if (idx == Compile::AliasIdxBot) 1932 mem->set_base_memory(phi); 1933 else 1934 mem->set_memory_at(idx, phi); 1935 return phi; 1936 } 1937 1938 //------------------------------call_register_finalizer----------------------- 1939 // Check the klass of the receiver and call register_finalizer if the 1940 // class need finalization. 1941 void Parse::call_register_finalizer() { 1942 Node* receiver = local(0); 1943 assert(receiver != NULL && receiver->bottom_type()->isa_instptr() != NULL, 1944 "must have non-null instance type"); 1945 1946 const TypeInstPtr *tinst = receiver->bottom_type()->isa_instptr(); 1947 if (tinst != NULL && tinst->klass()->is_loaded() && !tinst->klass_is_exact()) { 1948 // The type isn't known exactly so see if CHA tells us anything. 1949 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1950 if (!Dependencies::has_finalizable_subclass(ik)) { 1951 // No finalizable subclasses so skip the dynamic check. 1952 C->dependencies()->assert_has_no_finalizable_subclasses(ik); 1953 return; 1954 } 1955 } 1956 1957 // Insert a dynamic test for whether the instance needs 1958 // finalization. In general this will fold up since the concrete 1959 // class is often visible so the access flags are constant. 1960 Node* klass_addr = basic_plus_adr( receiver, receiver, oopDesc::klass_offset_in_bytes() ); 1961 Node* klass = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), klass_addr, TypeInstPtr::KLASS) ); 1962 1963 Node* access_flags_addr = basic_plus_adr(klass, klass, in_bytes(Klass::access_flags_offset())); 1964 Node* access_flags = make_load(NULL, access_flags_addr, TypeInt::INT, T_INT, MemNode::unordered); 1965 1966 Node* mask = _gvn.transform(new (C) AndINode(access_flags, intcon(JVM_ACC_HAS_FINALIZER))); 1967 Node* check = _gvn.transform(new (C) CmpINode(mask, intcon(0))); 1968 Node* test = _gvn.transform(new (C) BoolNode(check, BoolTest::ne)); 1969 1970 IfNode* iff = create_and_map_if(control(), test, PROB_MAX, COUNT_UNKNOWN); 1971 1972 RegionNode* result_rgn = new (C) RegionNode(3); 1973 record_for_igvn(result_rgn); 1974 1975 Node *skip_register = _gvn.transform(new (C) IfFalseNode(iff)); 1976 result_rgn->init_req(1, skip_register); 1977 1978 Node *needs_register = _gvn.transform(new (C) IfTrueNode(iff)); 1979 set_control(needs_register); 1980 if (stopped()) { 1981 // There is no slow path. 1982 result_rgn->init_req(2, top()); 1983 } else { 1984 Node *call = make_runtime_call(RC_NO_LEAF, 1985 OptoRuntime::register_finalizer_Type(), 1986 OptoRuntime::register_finalizer_Java(), 1987 NULL, TypePtr::BOTTOM, 1988 receiver); 1989 make_slow_call_ex(call, env()->Throwable_klass(), true); 1990 1991 Node* fast_io = call->in(TypeFunc::I_O); 1992 Node* fast_mem = call->in(TypeFunc::Memory); 1993 // These two phis are pre-filled with copies of of the fast IO and Memory 1994 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 1995 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 1996 1997 result_rgn->init_req(2, control()); 1998 io_phi ->init_req(2, i_o()); 1999 mem_phi ->init_req(2, reset_memory()); 2000 2001 set_all_memory( _gvn.transform(mem_phi) ); 2002 set_i_o( _gvn.transform(io_phi) ); 2003 } 2004 2005 set_control( _gvn.transform(result_rgn) ); 2006 } 2007 2008 // Add check to deoptimize if RTM state is not ProfileRTM 2009 void Parse::rtm_deopt() { 2010 #if INCLUDE_RTM_OPT 2011 if (C->profile_rtm()) { 2012 assert(C->method() != NULL, "only for normal compilations"); 2013 assert(!C->method()->method_data()->is_empty(), "MDO is needed to record RTM state"); 2014 assert(depth() == 1, "generate check only for main compiled method"); 2015 2016 // Set starting bci for uncommon trap. 2017 set_parse_bci(is_osr_parse() ? osr_bci() : 0); 2018 2019 // Load the rtm_state from the MethodData. 2020 const TypePtr* adr_type = TypeMetadataPtr::make(C->method()->method_data()); 2021 Node* mdo = makecon(adr_type); 2022 int offset = MethodData::rtm_state_offset_in_bytes(); 2023 Node* adr_node = basic_plus_adr(mdo, mdo, offset); 2024 Node* rtm_state = make_load(control(), adr_node, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 2025 2026 // Separate Load from Cmp by Opaque. 2027 // In expand_macro_nodes() it will be replaced either 2028 // with this load when there are locks in the code 2029 // or with ProfileRTM (cmp->in(2)) otherwise so that 2030 // the check will fold. 2031 Node* profile_state = makecon(TypeInt::make(ProfileRTM)); 2032 Node* opq = _gvn.transform( new (C) Opaque3Node(C, rtm_state, Opaque3Node::RTM_OPT) ); 2033 Node* chk = _gvn.transform( new (C) CmpINode(opq, profile_state) ); 2034 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 2035 // Branch to failure if state was changed 2036 { BuildCutout unless(this, tst, PROB_ALWAYS); 2037 uncommon_trap(Deoptimization::Reason_rtm_state_change, 2038 Deoptimization::Action_make_not_entrant); 2039 } 2040 } 2041 #endif 2042 } 2043 2044 //------------------------------return_current--------------------------------- 2045 // Append current _map to _exit_return 2046 void Parse::return_current(Node* value) { 2047 if (RegisterFinalizersAtInit && 2048 method()->intrinsic_id() == vmIntrinsics::_Object_init) { 2049 call_register_finalizer(); 2050 } 2051 2052 // Do not set_parse_bci, so that return goo is credited to the return insn. 2053 set_bci(InvocationEntryBci); 2054 if (method()->is_synchronized() && GenerateSynchronizationCode) { 2055 shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node()); 2056 } 2057 if (C->env()->dtrace_method_probes()) { 2058 make_dtrace_method_exit(method()); 2059 } 2060 SafePointNode* exit_return = _exits.map(); 2061 exit_return->in( TypeFunc::Control )->add_req( control() ); 2062 exit_return->in( TypeFunc::I_O )->add_req( i_o () ); 2063 Node *mem = exit_return->in( TypeFunc::Memory ); 2064 for (MergeMemStream mms(mem->as_MergeMem(), merged_memory()); mms.next_non_empty2(); ) { 2065 if (mms.is_empty()) { 2066 // get a copy of the base memory, and patch just this one input 2067 const TypePtr* adr_type = mms.adr_type(C); 2068 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 2069 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 2070 gvn().set_type_bottom(phi); 2071 phi->del_req(phi->req()-1); // prepare to re-patch 2072 mms.set_memory(phi); 2073 } 2074 mms.memory()->add_req(mms.memory2()); 2075 } 2076 2077 // frame pointer is always same, already captured 2078 if (value != NULL) { 2079 // If returning oops to an interface-return, there is a silent free 2080 // cast from oop to interface allowed by the Verifier. Make it explicit 2081 // here. 2082 Node* phi = _exits.argument(0); 2083 const TypeInstPtr *tr = phi->bottom_type()->isa_instptr(); 2084 if( tr && tr->klass()->is_loaded() && 2085 tr->klass()->is_interface() ) { 2086 const TypeInstPtr *tp = value->bottom_type()->isa_instptr(); 2087 if (tp && tp->klass()->is_loaded() && 2088 !tp->klass()->is_interface()) { 2089 // sharpen the type eagerly; this eases certain assert checking 2090 if (tp->higher_equal(TypeInstPtr::NOTNULL)) 2091 tr = tr->join_speculative(TypeInstPtr::NOTNULL)->is_instptr(); 2092 value = _gvn.transform(new (C) CheckCastPPNode(0,value,tr)); 2093 } 2094 } 2095 phi->add_req(value); 2096 } 2097 2098 if (_first_return) { 2099 _exits.map()->transfer_replaced_nodes_from(map(), _new_idx); 2100 _first_return = false; 2101 } else { 2102 _exits.map()->merge_replaced_nodes_with(map()); 2103 } 2104 2105 stop_and_kill_map(); // This CFG path dies here 2106 } 2107 2108 2109 //------------------------------add_safepoint---------------------------------- 2110 void Parse::add_safepoint() { 2111 // See if we can avoid this safepoint. No need for a SafePoint immediately 2112 // after a Call (except Leaf Call) or another SafePoint. 2113 Node *proj = control(); 2114 bool add_poll_param = SafePointNode::needs_polling_address_input(); 2115 uint parms = add_poll_param ? TypeFunc::Parms+1 : TypeFunc::Parms; 2116 if( proj->is_Proj() ) { 2117 Node *n0 = proj->in(0); 2118 if( n0->is_Catch() ) { 2119 n0 = n0->in(0)->in(0); 2120 assert( n0->is_Call(), "expect a call here" ); 2121 } 2122 if( n0->is_Call() ) { 2123 if( n0->as_Call()->guaranteed_safepoint() ) 2124 return; 2125 } else if( n0->is_SafePoint() && n0->req() >= parms ) { 2126 return; 2127 } 2128 } 2129 2130 // Clear out dead values from the debug info. 2131 kill_dead_locals(); 2132 2133 // Clone the JVM State 2134 SafePointNode *sfpnt = new (C) SafePointNode(parms, NULL); 2135 2136 // Capture memory state BEFORE a SafePoint. Since we can block at a 2137 // SafePoint we need our GC state to be safe; i.e. we need all our current 2138 // write barriers (card marks) to not float down after the SafePoint so we 2139 // must read raw memory. Likewise we need all oop stores to match the card 2140 // marks. If deopt can happen, we need ALL stores (we need the correct JVM 2141 // state on a deopt). 2142 2143 // We do not need to WRITE the memory state after a SafePoint. The control 2144 // edge will keep card-marks and oop-stores from floating up from below a 2145 // SafePoint and our true dependency added here will keep them from floating 2146 // down below a SafePoint. 2147 2148 // Clone the current memory state 2149 Node* mem = MergeMemNode::make(C, map()->memory()); 2150 2151 mem = _gvn.transform(mem); 2152 2153 // Pass control through the safepoint 2154 sfpnt->init_req(TypeFunc::Control , control()); 2155 // Fix edges normally used by a call 2156 sfpnt->init_req(TypeFunc::I_O , top() ); 2157 sfpnt->init_req(TypeFunc::Memory , mem ); 2158 sfpnt->init_req(TypeFunc::ReturnAdr, top() ); 2159 sfpnt->init_req(TypeFunc::FramePtr , top() ); 2160 2161 // Create a node for the polling address 2162 if( add_poll_param ) { 2163 Node *polladr = ConPNode::make(C, (address)os::get_polling_page()); 2164 sfpnt->init_req(TypeFunc::Parms+0, _gvn.transform(polladr)); 2165 } 2166 2167 // Fix up the JVM State edges 2168 add_safepoint_edges(sfpnt); 2169 Node *transformed_sfpnt = _gvn.transform(sfpnt); 2170 set_control(transformed_sfpnt); 2171 2172 // Provide an edge from root to safepoint. This makes the safepoint 2173 // appear useful until the parse has completed. 2174 if( OptoRemoveUseless && transformed_sfpnt->is_SafePoint() ) { 2175 assert(C->root() != NULL, "Expect parse is still valid"); 2176 C->root()->add_prec(transformed_sfpnt); 2177 } 2178 } 2179 2180 #ifndef PRODUCT 2181 //------------------------show_parse_info-------------------------------------- 2182 void Parse::show_parse_info() { 2183 InlineTree* ilt = NULL; 2184 if (C->ilt() != NULL) { 2185 JVMState* caller_jvms = is_osr_parse() ? caller()->caller() : caller(); 2186 ilt = InlineTree::find_subtree_from_root(C->ilt(), caller_jvms, method()); 2187 } 2188 if (PrintCompilation && Verbose) { 2189 if (depth() == 1) { 2190 if( ilt->count_inlines() ) { 2191 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2192 ilt->count_inline_bcs()); 2193 tty->cr(); 2194 } 2195 } else { 2196 if (method()->is_synchronized()) tty->print("s"); 2197 if (method()->has_exception_handlers()) tty->print("!"); 2198 // Check this is not the final compiled version 2199 if (C->trap_can_recompile()) { 2200 tty->print("-"); 2201 } else { 2202 tty->print(" "); 2203 } 2204 method()->print_short_name(); 2205 if (is_osr_parse()) { 2206 tty->print(" @ %d", osr_bci()); 2207 } 2208 tty->print(" (%d bytes)",method()->code_size()); 2209 if (ilt->count_inlines()) { 2210 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2211 ilt->count_inline_bcs()); 2212 } 2213 tty->cr(); 2214 } 2215 } 2216 if (PrintOpto && (depth() == 1 || PrintOptoInlining)) { 2217 // Print that we succeeded; suppress this message on the first osr parse. 2218 2219 if (method()->is_synchronized()) tty->print("s"); 2220 if (method()->has_exception_handlers()) tty->print("!"); 2221 // Check this is not the final compiled version 2222 if (C->trap_can_recompile() && depth() == 1) { 2223 tty->print("-"); 2224 } else { 2225 tty->print(" "); 2226 } 2227 if( depth() != 1 ) { tty->print(" "); } // missing compile count 2228 for (int i = 1; i < depth(); ++i) { tty->print(" "); } 2229 method()->print_short_name(); 2230 if (is_osr_parse()) { 2231 tty->print(" @ %d", osr_bci()); 2232 } 2233 if (ilt->caller_bci() != -1) { 2234 tty->print(" @ %d", ilt->caller_bci()); 2235 } 2236 tty->print(" (%d bytes)",method()->code_size()); 2237 if (ilt->count_inlines()) { 2238 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2239 ilt->count_inline_bcs()); 2240 } 2241 tty->cr(); 2242 } 2243 } 2244 2245 2246 //------------------------------dump------------------------------------------- 2247 // Dump information associated with the bytecodes of current _method 2248 void Parse::dump() { 2249 if( method() != NULL ) { 2250 // Iterate over bytecodes 2251 ciBytecodeStream iter(method()); 2252 for( Bytecodes::Code bc = iter.next(); bc != ciBytecodeStream::EOBC() ; bc = iter.next() ) { 2253 dump_bci( iter.cur_bci() ); 2254 tty->cr(); 2255 } 2256 } 2257 } 2258 2259 // Dump information associated with a byte code index, 'bci' 2260 void Parse::dump_bci(int bci) { 2261 // Output info on merge-points, cloning, and within _jsr..._ret 2262 // NYI 2263 tty->print(" bci:%d", bci); 2264 } 2265 2266 #endif