1 /* 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "interpreter/linkResolver.hpp" 28 #include "oops/method.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/idealGraphPrinter.hpp" 32 #include "opto/locknode.hpp" 33 #include "opto/memnode.hpp" 34 #include "opto/parse.hpp" 35 #include "opto/rootnode.hpp" 36 #include "opto/runtime.hpp" 37 #include "runtime/arguments.hpp" 38 #include "runtime/handles.inline.hpp" 39 #include "runtime/sharedRuntime.hpp" 40 #include "utilities/copy.hpp" 41 42 // Static array so we can figure out which bytecodes stop us from compiling 43 // the most. Some of the non-static variables are needed in bytecodeInfo.cpp 44 // and eventually should be encapsulated in a proper class (gri 8/18/98). 45 46 int nodes_created = 0; 47 int methods_parsed = 0; 48 int methods_seen = 0; 49 int blocks_parsed = 0; 50 int blocks_seen = 0; 51 52 int explicit_null_checks_inserted = 0; 53 int explicit_null_checks_elided = 0; 54 int all_null_checks_found = 0, implicit_null_checks = 0; 55 int implicit_null_throws = 0; 56 57 int reclaim_idx = 0; 58 int reclaim_in = 0; 59 int reclaim_node = 0; 60 61 #ifndef PRODUCT 62 bool Parse::BytecodeParseHistogram::_initialized = false; 63 uint Parse::BytecodeParseHistogram::_bytecodes_parsed [Bytecodes::number_of_codes]; 64 uint Parse::BytecodeParseHistogram::_nodes_constructed[Bytecodes::number_of_codes]; 65 uint Parse::BytecodeParseHistogram::_nodes_transformed[Bytecodes::number_of_codes]; 66 uint Parse::BytecodeParseHistogram::_new_values [Bytecodes::number_of_codes]; 67 #endif 68 69 //------------------------------print_statistics------------------------------- 70 #ifndef PRODUCT 71 void Parse::print_statistics() { 72 tty->print_cr("--- Compiler Statistics ---"); 73 tty->print("Methods seen: %d Methods parsed: %d", methods_seen, methods_parsed); 74 tty->print(" Nodes created: %d", nodes_created); 75 tty->cr(); 76 if (methods_seen != methods_parsed) 77 tty->print_cr("Reasons for parse failures (NOT cumulative):"); 78 tty->print_cr("Blocks parsed: %d Blocks seen: %d", blocks_parsed, blocks_seen); 79 80 if( explicit_null_checks_inserted ) 81 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); 82 if( all_null_checks_found ) 83 tty->print_cr("%d made implicit (%2d%%)", implicit_null_checks, 84 (100*implicit_null_checks)/all_null_checks_found); 85 if( implicit_null_throws ) 86 tty->print_cr("%d implicit null exceptions at runtime", 87 implicit_null_throws); 88 89 if( PrintParseStatistics && BytecodeParseHistogram::initialized() ) { 90 BytecodeParseHistogram::print(); 91 } 92 } 93 #endif 94 95 //------------------------------ON STACK REPLACEMENT--------------------------- 96 97 // Construct a node which can be used to get incoming state for 98 // on stack replacement. 99 Node *Parse::fetch_interpreter_state(int index, 100 BasicType bt, 101 Node *local_addrs, 102 Node *local_addrs_base) { 103 Node *mem = memory(Compile::AliasIdxRaw); 104 Node *adr = basic_plus_adr( local_addrs_base, local_addrs, -index*wordSize ); 105 Node *ctl = control(); 106 107 // Very similar to LoadNode::make, except we handle un-aligned longs and 108 // doubles on Sparc. Intel can handle them just fine directly. 109 Node *l; 110 switch (bt) { // Signature is flattened 111 case T_INT: l = new (C) LoadINode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInt::INT, MemNode::unordered); break; 112 case T_FLOAT: l = new (C) LoadFNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::FLOAT, MemNode::unordered); break; 113 case T_ADDRESS: l = new (C) LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered); break; 114 case T_OBJECT: l = new (C) LoadPNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeInstPtr::BOTTOM, MemNode::unordered); break; 115 case T_LONG: 116 case T_DOUBLE: { 117 // Since arguments are in reverse order, the argument address 'adr' 118 // refers to the back half of the long/double. Recompute adr. 119 adr = basic_plus_adr(local_addrs_base, local_addrs, -(index+1)*wordSize); 120 if (Matcher::misaligned_doubles_ok) { 121 l = (bt == T_DOUBLE) 122 ? (Node*)new (C) LoadDNode(ctl, mem, adr, TypeRawPtr::BOTTOM, Type::DOUBLE, MemNode::unordered) 123 : (Node*)new (C) LoadLNode(ctl, mem, adr, TypeRawPtr::BOTTOM, TypeLong::LONG, MemNode::unordered); 124 } else { 125 l = (bt == T_DOUBLE) 126 ? (Node*)new (C) LoadD_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered) 127 : (Node*)new (C) LoadL_unalignedNode(ctl, mem, adr, TypeRawPtr::BOTTOM, MemNode::unordered); 128 } 129 break; 130 } 131 default: ShouldNotReachHere(); 132 } 133 return _gvn.transform(l); 134 } 135 136 // Helper routine to prevent the interpreter from handing 137 // unexpected typestate to an OSR method. 138 // The Node l is a value newly dug out of the interpreter frame. 139 // The type is the type predicted by ciTypeFlow. Note that it is 140 // not a general type, but can only come from Type::get_typeflow_type. 141 // The safepoint is a map which will feed an uncommon trap. 142 Node* Parse::check_interpreter_type(Node* l, const Type* type, 143 SafePointNode* &bad_type_exit) { 144 145 const TypeOopPtr* tp = type->isa_oopptr(); 146 147 // TypeFlow may assert null-ness if a type appears unloaded. 148 if (type == TypePtr::NULL_PTR || 149 (tp != NULL && !tp->klass()->is_loaded())) { 150 // Value must be null, not a real oop. 151 Node* chk = _gvn.transform( new (C) CmpPNode(l, null()) ); 152 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) ); 153 IfNode* iff = create_and_map_if(control(), tst, PROB_MAX, COUNT_UNKNOWN); 154 set_control(_gvn.transform( new (C) IfTrueNode(iff) )); 155 Node* bad_type = _gvn.transform( new (C) IfFalseNode(iff) ); 156 bad_type_exit->control()->add_req(bad_type); 157 l = null(); 158 } 159 160 // Typeflow can also cut off paths from the CFG, based on 161 // types which appear unloaded, or call sites which appear unlinked. 162 // When paths are cut off, values at later merge points can rise 163 // toward more specific classes. Make sure these specific classes 164 // are still in effect. 165 if (tp != NULL && tp->klass() != C->env()->Object_klass()) { 166 // TypeFlow asserted a specific object type. Value must have that type. 167 Node* bad_type_ctrl = NULL; 168 l = gen_checkcast(l, makecon(TypeKlassPtr::make(tp->klass())), &bad_type_ctrl); 169 bad_type_exit->control()->add_req(bad_type_ctrl); 170 } 171 172 BasicType bt_l = _gvn.type(l)->basic_type(); 173 BasicType bt_t = type->basic_type(); 174 assert(_gvn.type(l)->higher_equal(type), "must constrain OSR typestate"); 175 return l; 176 } 177 178 // Helper routine which sets up elements of the initial parser map when 179 // performing a parse for on stack replacement. Add values into map. 180 // The only parameter contains the address of a interpreter arguments. 181 void Parse::load_interpreter_state(Node* osr_buf) { 182 int index; 183 int max_locals = jvms()->loc_size(); 184 int max_stack = jvms()->stk_size(); 185 186 187 // Mismatch between method and jvms can occur since map briefly held 188 // an OSR entry state (which takes up one RawPtr word). 189 assert(max_locals == method()->max_locals(), "sanity"); 190 assert(max_stack >= method()->max_stack(), "sanity"); 191 assert((int)jvms()->endoff() == TypeFunc::Parms + max_locals + max_stack, "sanity"); 192 assert((int)jvms()->endoff() == (int)map()->req(), "sanity"); 193 194 // Find the start block. 195 Block* osr_block = start_block(); 196 assert(osr_block->start() == osr_bci(), "sanity"); 197 198 // Set initial BCI. 199 set_parse_bci(osr_block->start()); 200 201 // Set initial stack depth. 202 set_sp(osr_block->start_sp()); 203 204 // Check bailouts. We currently do not perform on stack replacement 205 // of loops in catch blocks or loops which branch with a non-empty stack. 206 if (sp() != 0) { 207 C->record_method_not_compilable("OSR starts with non-empty stack"); 208 return; 209 } 210 // Do not OSR inside finally clauses: 211 if (osr_block->has_trap_at(osr_block->start())) { 212 C->record_method_not_compilable("OSR starts with an immediate trap"); 213 return; 214 } 215 216 // Commute monitors from interpreter frame to compiler frame. 217 assert(jvms()->monitor_depth() == 0, "should be no active locks at beginning of osr"); 218 int mcnt = osr_block->flow()->monitor_count(); 219 Node *monitors_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals+mcnt*2-1)*wordSize); 220 for (index = 0; index < mcnt; index++) { 221 // Make a BoxLockNode for the monitor. 222 Node *box = _gvn.transform(new (C) BoxLockNode(next_monitor())); 223 224 225 // Displaced headers and locked objects are interleaved in the 226 // temp OSR buffer. We only copy the locked objects out here. 227 // Fetch the locked object from the OSR temp buffer and copy to our fastlock node. 228 Node *lock_object = fetch_interpreter_state(index*2, T_OBJECT, monitors_addr, osr_buf); 229 // Try and copy the displaced header to the BoxNode 230 Node *displaced_hdr = fetch_interpreter_state((index*2) + 1, T_ADDRESS, monitors_addr, osr_buf); 231 232 233 store_to_memory(control(), box, displaced_hdr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered); 234 235 // Build a bogus FastLockNode (no code will be generated) and push the 236 // monitor into our debug info. 237 const FastLockNode *flock = _gvn.transform(new (C) FastLockNode( 0, lock_object, box ))->as_FastLock(); 238 map()->push_monitor(flock); 239 240 // If the lock is our method synchronization lock, tuck it away in 241 // _sync_lock for return and rethrow exit paths. 242 if (index == 0 && method()->is_synchronized()) { 243 _synch_lock = flock; 244 } 245 } 246 247 // Use the raw liveness computation to make sure that unexpected 248 // values don't propagate into the OSR frame. 249 MethodLivenessResult live_locals = method()->liveness_at_bci(osr_bci()); 250 if (!live_locals.is_valid()) { 251 // Degenerate or breakpointed method. 252 C->record_method_not_compilable("OSR in empty or breakpointed method"); 253 return; 254 } 255 256 // Extract the needed locals from the interpreter frame. 257 Node *locals_addr = basic_plus_adr(osr_buf, osr_buf, (max_locals-1)*wordSize); 258 259 // find all the locals that the interpreter thinks contain live oops 260 const BitMap live_oops = method()->live_local_oops_at_bci(osr_bci()); 261 for (index = 0; index < max_locals; index++) { 262 263 if (!live_locals.at(index)) { 264 continue; 265 } 266 267 const Type *type = osr_block->local_type_at(index); 268 269 if (type->isa_oopptr() != NULL) { 270 271 // 6403625: Verify that the interpreter oopMap thinks that the oop is live 272 // else we might load a stale oop if the MethodLiveness disagrees with the 273 // result of the interpreter. If the interpreter says it is dead we agree 274 // by making the value go to top. 275 // 276 277 if (!live_oops.at(index)) { 278 if (C->log() != NULL) { 279 C->log()->elem("OSR_mismatch local_index='%d'",index); 280 } 281 set_local(index, null()); 282 // and ignore it for the loads 283 continue; 284 } 285 } 286 287 // Filter out TOP, HALF, and BOTTOM. (Cf. ensure_phi.) 288 if (type == Type::TOP || type == Type::HALF) { 289 continue; 290 } 291 // If the type falls to bottom, then this must be a local that 292 // is mixing ints and oops or some such. Forcing it to top 293 // makes it go dead. 294 if (type == Type::BOTTOM) { 295 continue; 296 } 297 // Construct code to access the appropriate local. 298 BasicType bt = type->basic_type(); 299 if (type == TypePtr::NULL_PTR) { 300 // Ptr types are mixed together with T_ADDRESS but NULL is 301 // really for T_OBJECT types so correct it. 302 bt = T_OBJECT; 303 } 304 Node *value = fetch_interpreter_state(index, bt, locals_addr, osr_buf); 305 set_local(index, value); 306 } 307 308 // Extract the needed stack entries from the interpreter frame. 309 for (index = 0; index < sp(); index++) { 310 const Type *type = osr_block->stack_type_at(index); 311 if (type != Type::TOP) { 312 // Currently the compiler bails out when attempting to on stack replace 313 // at a bci with a non-empty stack. We should not reach here. 314 ShouldNotReachHere(); 315 } 316 } 317 318 // End the OSR migration 319 make_runtime_call(RC_LEAF, OptoRuntime::osr_end_Type(), 320 CAST_FROM_FN_PTR(address, SharedRuntime::OSR_migration_end), 321 "OSR_migration_end", TypeRawPtr::BOTTOM, 322 osr_buf); 323 324 // Now that the interpreter state is loaded, make sure it will match 325 // at execution time what the compiler is expecting now: 326 SafePointNode* bad_type_exit = clone_map(); 327 bad_type_exit->set_control(new (C) RegionNode(1)); 328 329 assert(osr_block->flow()->jsrs()->size() == 0, "should be no jsrs live at osr point"); 330 for (index = 0; index < max_locals; index++) { 331 if (stopped()) break; 332 Node* l = local(index); 333 if (l->is_top()) continue; // nothing here 334 const Type *type = osr_block->local_type_at(index); 335 if (type->isa_oopptr() != NULL) { 336 if (!live_oops.at(index)) { 337 // skip type check for dead oops 338 continue; 339 } 340 } 341 if (osr_block->flow()->local_type_at(index)->is_return_address()) { 342 // In our current system it's illegal for jsr addresses to be 343 // live into an OSR entry point because the compiler performs 344 // inlining of jsrs. ciTypeFlow has a bailout that detect this 345 // case and aborts the compile if addresses are live into an OSR 346 // entry point. Because of that we can assume that any address 347 // locals at the OSR entry point are dead. Method liveness 348 // isn't precise enought to figure out that they are dead in all 349 // cases so simply skip checking address locals all 350 // together. Any type check is guaranteed to fail since the 351 // interpreter type is the result of a load which might have any 352 // value and the expected type is a constant. 353 continue; 354 } 355 set_local(index, check_interpreter_type(l, type, bad_type_exit)); 356 } 357 358 for (index = 0; index < sp(); index++) { 359 if (stopped()) break; 360 Node* l = stack(index); 361 if (l->is_top()) continue; // nothing here 362 const Type *type = osr_block->stack_type_at(index); 363 set_stack(index, check_interpreter_type(l, type, bad_type_exit)); 364 } 365 366 if (bad_type_exit->control()->req() > 1) { 367 // Build an uncommon trap here, if any inputs can be unexpected. 368 bad_type_exit->set_control(_gvn.transform( bad_type_exit->control() )); 369 record_for_igvn(bad_type_exit->control()); 370 SafePointNode* types_are_good = map(); 371 set_map(bad_type_exit); 372 // The unexpected type happens because a new edge is active 373 // in the CFG, which typeflow had previously ignored. 374 // E.g., Object x = coldAtFirst() && notReached()? "str": new Integer(123). 375 // This x will be typed as Integer if notReached is not yet linked. 376 // It could also happen due to a problem in ciTypeFlow analysis. 377 uncommon_trap(Deoptimization::Reason_constraint, 378 Deoptimization::Action_reinterpret); 379 set_map(types_are_good); 380 } 381 } 382 383 //------------------------------Parse------------------------------------------ 384 // Main parser constructor. 385 Parse::Parse(JVMState* caller, ciMethod* parse_method, float expected_uses, Parse* parent) 386 : _exits(caller), _parent(parent) 387 { 388 // Init some variables 389 _caller = caller; 390 _method = parse_method; 391 _expected_uses = expected_uses; 392 _depth = 1 + (caller->has_method() ? caller->depth() : 0); 393 _wrote_final = false; 394 _wrote_volatile = false; 395 _wrote_stable = false; 396 _wrote_fields = false; 397 _alloc_with_final = NULL; 398 _entry_bci = InvocationEntryBci; 399 _tf = NULL; 400 _block = NULL; 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 // Check for bailouts during method entry. 573 if (failing()) { 574 if (log) log->done("parse"); 575 C->set_default_node_notes(caller_nn); 576 return; 577 } 578 579 entry_map = map(); // capture any changes performed by method setup code 580 assert(jvms()->endoff() == map()->req(), "map matches JVMS layout"); 581 582 // We begin parsing as if we have just encountered a jump to the 583 // method entry. 584 Block* entry_block = start_block(); 585 assert(entry_block->start() == (is_osr_parse() ? osr_bci() : 0), ""); 586 set_map_clone(entry_map); 587 merge_common(entry_block, entry_block->next_path_num()); 588 589 #ifndef PRODUCT 590 BytecodeParseHistogram *parse_histogram_obj = new (C->env()->arena()) BytecodeParseHistogram(this, C); 591 set_parse_histogram( parse_histogram_obj ); 592 #endif 593 594 // Parse all the basic blocks. 595 do_all_blocks(); 596 597 C->set_default_node_notes(caller_nn); 598 599 // Check for bailouts during conversion to graph 600 if (failing()) { 601 if (log) log->done("parse"); 602 return; 603 } 604 605 // Fix up all exiting control flow. 606 set_map(entry_map); 607 do_exits(); 608 609 if (log) log->done("parse nodes='%d' live='%d' memory='%d'", 610 C->unique(), C->live_nodes(), C->node_arena()->used()); 611 } 612 613 //---------------------------do_all_blocks------------------------------------- 614 void Parse::do_all_blocks() { 615 bool has_irreducible = flow()->has_irreducible_entry(); 616 617 // Walk over all blocks in Reverse Post-Order. 618 while (true) { 619 bool progress = false; 620 for (int rpo = 0; rpo < block_count(); rpo++) { 621 Block* block = rpo_at(rpo); 622 623 if (block->is_parsed()) continue; 624 625 if (!block->is_merged()) { 626 // Dead block, no state reaches this block 627 continue; 628 } 629 630 // Prepare to parse this block. 631 load_state_from(block); 632 633 if (stopped()) { 634 // Block is dead. 635 continue; 636 } 637 638 blocks_parsed++; 639 640 progress = true; 641 if (block->is_loop_head() || block->is_handler() || has_irreducible && !block->is_ready()) { 642 // Not all preds have been parsed. We must build phis everywhere. 643 // (Note that dead locals do not get phis built, ever.) 644 ensure_phis_everywhere(); 645 646 if (block->is_SEL_head() && 647 (UseLoopPredicate || LoopLimitCheck)) { 648 // Add predicate to single entry (not irreducible) loop head. 649 assert(!block->has_merged_backedge(), "only entry paths should be merged for now"); 650 // Need correct bci for predicate. 651 // It is fine to set it here since do_one_block() will set it anyway. 652 set_parse_bci(block->start()); 653 add_predicate(); 654 // Add new region for back branches. 655 int edges = block->pred_count() - block->preds_parsed() + 1; // +1 for original region 656 RegionNode *r = new (C) RegionNode(edges+1); 657 _gvn.set_type(r, Type::CONTROL); 658 record_for_igvn(r); 659 r->init_req(edges, control()); 660 set_control(r); 661 // Add new phis. 662 ensure_phis_everywhere(); 663 } 664 665 // Leave behind an undisturbed copy of the map, for future merges. 666 set_map(clone_map()); 667 } 668 669 if (control()->is_Region() && !block->is_loop_head() && !has_irreducible && !block->is_handler()) { 670 // In the absence of irreducible loops, the Region and Phis 671 // associated with a merge that doesn't involve a backedge can 672 // be simplified now since the RPO parsing order guarantees 673 // that any path which was supposed to reach here has already 674 // been parsed or must be dead. 675 Node* c = control(); 676 Node* result = _gvn.transform_no_reclaim(control()); 677 if (c != result && TraceOptoParse) { 678 tty->print_cr("Block #%d replace %d with %d", block->rpo(), c->_idx, result->_idx); 679 } 680 if (result != top()) { 681 record_for_igvn(result); 682 } 683 } 684 685 // Parse the block. 686 do_one_block(); 687 688 // Check for bailouts. 689 if (failing()) return; 690 } 691 692 // with irreducible loops multiple passes might be necessary to parse everything 693 if (!has_irreducible || !progress) { 694 break; 695 } 696 } 697 698 blocks_seen += block_count(); 699 700 #ifndef PRODUCT 701 // Make sure there are no half-processed blocks remaining. 702 // Every remaining unprocessed block is dead and may be ignored now. 703 for (int rpo = 0; rpo < block_count(); rpo++) { 704 Block* block = rpo_at(rpo); 705 if (!block->is_parsed()) { 706 if (TraceOptoParse) { 707 tty->print_cr("Skipped dead block %d at bci:%d", rpo, block->start()); 708 } 709 assert(!block->is_merged(), "no half-processed blocks"); 710 } 711 } 712 #endif 713 } 714 715 //-------------------------------build_exits---------------------------------- 716 // Build normal and exceptional exit merge points. 717 void Parse::build_exits() { 718 // make a clone of caller to prevent sharing of side-effects 719 _exits.set_map(_exits.clone_map()); 720 _exits.clean_stack(_exits.sp()); 721 _exits.sync_jvms(); 722 723 RegionNode* region = new (C) RegionNode(1); 724 record_for_igvn(region); 725 gvn().set_type_bottom(region); 726 _exits.set_control(region); 727 728 // Note: iophi and memphi are not transformed until do_exits. 729 Node* iophi = new (C) PhiNode(region, Type::ABIO); 730 Node* memphi = new (C) PhiNode(region, Type::MEMORY, TypePtr::BOTTOM); 731 gvn().set_type_bottom(iophi); 732 gvn().set_type_bottom(memphi); 733 _exits.set_i_o(iophi); 734 _exits.set_all_memory(memphi); 735 736 // Add a return value to the exit state. (Do not push it yet.) 737 if (tf()->range()->cnt() > TypeFunc::Parms) { 738 const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms); 739 // Don't "bind" an unloaded return klass to the ret_phi. If the klass 740 // becomes loaded during the subsequent parsing, the loaded and unloaded 741 // types will not join when we transform and push in do_exits(). 742 const TypeOopPtr* ret_oop_type = ret_type->isa_oopptr(); 743 if (ret_oop_type && !ret_oop_type->klass()->is_loaded()) { 744 ret_type = TypeOopPtr::BOTTOM; 745 } 746 int ret_size = type2size[ret_type->basic_type()]; 747 Node* ret_phi = new (C) PhiNode(region, ret_type); 748 gvn().set_type_bottom(ret_phi); 749 _exits.ensure_stack(ret_size); 750 assert((int)(tf()->range()->cnt() - TypeFunc::Parms) == ret_size, "good tf range"); 751 assert(method()->return_type()->size() == ret_size, "tf agrees w/ method"); 752 _exits.set_argument(0, ret_phi); // here is where the parser finds it 753 // Note: ret_phi is not yet pushed, until do_exits. 754 } 755 } 756 757 758 //----------------------------build_start_state------------------------------- 759 // Construct a state which contains only the incoming arguments from an 760 // unknown caller. The method & bci will be NULL & InvocationEntryBci. 761 JVMState* Compile::build_start_state(StartNode* start, const TypeFunc* tf) { 762 int arg_size = tf->domain()->cnt(); 763 int max_size = MAX2(arg_size, (int)tf->range()->cnt()); 764 JVMState* jvms = new (this) JVMState(max_size - TypeFunc::Parms); 765 SafePointNode* map = new (this) SafePointNode(max_size, NULL); 766 record_for_igvn(map); 767 assert(arg_size == TypeFunc::Parms + (is_osr_compilation() ? 1 : method()->arg_size()), "correct arg_size"); 768 Node_Notes* old_nn = default_node_notes(); 769 if (old_nn != NULL && has_method()) { 770 Node_Notes* entry_nn = old_nn->clone(this); 771 JVMState* entry_jvms = new(this) JVMState(method(), old_nn->jvms()); 772 entry_jvms->set_offsets(0); 773 entry_jvms->set_bci(entry_bci()); 774 entry_nn->set_jvms(entry_jvms); 775 set_default_node_notes(entry_nn); 776 } 777 uint i; 778 for (i = 0; i < (uint)arg_size; i++) { 779 Node* parm = initial_gvn()->transform(new (this) ParmNode(start, i)); 780 map->init_req(i, parm); 781 // Record all these guys for later GVN. 782 record_for_igvn(parm); 783 } 784 for (; i < map->req(); i++) { 785 map->init_req(i, top()); 786 } 787 assert(jvms->argoff() == TypeFunc::Parms, "parser gets arguments here"); 788 set_default_node_notes(old_nn); 789 map->set_jvms(jvms); 790 jvms->set_map(map); 791 return jvms; 792 } 793 794 //-----------------------------make_node_notes--------------------------------- 795 Node_Notes* Parse::make_node_notes(Node_Notes* caller_nn) { 796 if (caller_nn == NULL) return NULL; 797 Node_Notes* nn = caller_nn->clone(C); 798 JVMState* caller_jvms = nn->jvms(); 799 JVMState* jvms = new (C) JVMState(method(), caller_jvms); 800 jvms->set_offsets(0); 801 jvms->set_bci(_entry_bci); 802 nn->set_jvms(jvms); 803 return nn; 804 } 805 806 807 //--------------------------return_values-------------------------------------- 808 void Compile::return_values(JVMState* jvms) { 809 GraphKit kit(jvms); 810 Node* ret = new (this) ReturnNode(TypeFunc::Parms, 811 kit.control(), 812 kit.i_o(), 813 kit.reset_memory(), 814 kit.frameptr(), 815 kit.returnadr()); 816 // Add zero or 1 return values 817 int ret_size = tf()->range()->cnt() - TypeFunc::Parms; 818 if (ret_size > 0) { 819 kit.inc_sp(-ret_size); // pop the return value(s) 820 kit.sync_jvms(); 821 ret->add_req(kit.argument(0)); 822 // Note: The second dummy edge is not needed by a ReturnNode. 823 } 824 // bind it to root 825 root()->add_req(ret); 826 record_for_igvn(ret); 827 initial_gvn()->transform_no_reclaim(ret); 828 } 829 830 //------------------------rethrow_exceptions----------------------------------- 831 // Bind all exception states in the list into a single RethrowNode. 832 void Compile::rethrow_exceptions(JVMState* jvms) { 833 GraphKit kit(jvms); 834 if (!kit.has_exceptions()) return; // nothing to generate 835 // Load my combined exception state into the kit, with all phis transformed: 836 SafePointNode* ex_map = kit.combine_and_pop_all_exception_states(); 837 Node* ex_oop = kit.use_exception_state(ex_map); 838 RethrowNode* exit = new (this) RethrowNode(kit.control(), 839 kit.i_o(), kit.reset_memory(), 840 kit.frameptr(), kit.returnadr(), 841 // like a return but with exception input 842 ex_oop); 843 // bind to root 844 root()->add_req(exit); 845 record_for_igvn(exit); 846 initial_gvn()->transform_no_reclaim(exit); 847 } 848 849 //---------------------------do_exceptions------------------------------------- 850 // Process exceptions arising from the current bytecode. 851 // Send caught exceptions to the proper handler within this method. 852 // Unhandled exceptions feed into _exit. 853 void Parse::do_exceptions() { 854 if (!has_exceptions()) return; 855 856 if (failing()) { 857 // Pop them all off and throw them away. 858 while (pop_exception_state() != NULL) ; 859 return; 860 } 861 862 PreserveJVMState pjvms(this, false); 863 864 SafePointNode* ex_map; 865 while ((ex_map = pop_exception_state()) != NULL) { 866 if (!method()->has_exception_handlers()) { 867 // Common case: Transfer control outward. 868 // Doing it this early allows the exceptions to common up 869 // even between adjacent method calls. 870 throw_to_exit(ex_map); 871 } else { 872 // Have to look at the exception first. 873 assert(stopped(), "catch_inline_exceptions trashes the map"); 874 catch_inline_exceptions(ex_map); 875 stop_and_kill_map(); // we used up this exception state; kill it 876 } 877 } 878 879 // We now return to our regularly scheduled program: 880 } 881 882 //---------------------------throw_to_exit------------------------------------- 883 // Merge the given map into an exception exit from this method. 884 // The exception exit will handle any unlocking of receiver. 885 // The ex_oop must be saved within the ex_map, unlike merge_exception. 886 void Parse::throw_to_exit(SafePointNode* ex_map) { 887 // Pop the JVMS to (a copy of) the caller. 888 GraphKit caller; 889 caller.set_map_clone(_caller->map()); 890 caller.set_bci(_caller->bci()); 891 caller.set_sp(_caller->sp()); 892 // Copy out the standard machine state: 893 for (uint i = 0; i < TypeFunc::Parms; i++) { 894 caller.map()->set_req(i, ex_map->in(i)); 895 } 896 // ...and the exception: 897 Node* ex_oop = saved_ex_oop(ex_map); 898 SafePointNode* caller_ex_map = caller.make_exception_state(ex_oop); 899 // Finally, collect the new exception state in my exits: 900 _exits.add_exception_state(caller_ex_map); 901 } 902 903 //------------------------------do_exits--------------------------------------- 904 void Parse::do_exits() { 905 set_parse_bci(InvocationEntryBci); 906 907 // Now peephole on the return bits 908 Node* region = _exits.control(); 909 _exits.set_control(gvn().transform(region)); 910 911 Node* iophi = _exits.i_o(); 912 _exits.set_i_o(gvn().transform(iophi)); 913 914 // Figure out if we need to emit the trailing barrier. The barrier is only 915 // needed in the constructors, and only in three cases: 916 // 917 // 1. The constructor wrote a final. The effects of all initializations 918 // must be committed to memory before any code after the constructor 919 // publishes the reference to the newly constructed object. Rather 920 // than wait for the publication, we simply block the writes here. 921 // Rather than put a barrier on only those writes which are required 922 // to complete, we force all writes to complete. 923 // 924 // 2. On PPC64, also add MemBarRelease for constructors which write 925 // volatile fields. As support_IRIW_for_not_multiple_copy_atomic_cpu 926 // is set on PPC64, no sync instruction is issued after volatile 927 // stores. We want to guarantee the same behavior as on platforms 928 // with total store order, although this is not required by the Java 929 // memory model. So as with finals, we add a barrier here. 930 // 931 // 3. Experimental VM option is used to force the barrier if any field 932 // was written out in the constructor. 933 // 934 // "All bets are off" unless the first publication occurs after a 935 // normal return from the constructor. We do not attempt to detect 936 // such unusual early publications. But no barrier is needed on 937 // exceptional returns, since they cannot publish normally. 938 // 939 if (method()->is_initializer() && 940 (wrote_final() || 941 PPC64_ONLY(wrote_volatile() ||) 942 (AlwaysSafeConstructors && wrote_fields()))) { 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 // Any method can write a @Stable field; insert memory barriers after 953 // those also. If there is a predecessor allocation node, bind the 954 // barrier there. 955 if (wrote_stable()) { 956 _exits.insert_mem_bar(Op_MemBarRelease, alloc_with_final()); 957 #ifndef PRODUCT 958 if (PrintOpto && (Verbose || WizardMode)) { 959 method()->print_name(); 960 tty->print_cr(" writes @Stable and needs a memory barrier"); 961 } 962 #endif 963 } 964 965 for (MergeMemStream mms(_exits.merged_memory()); mms.next_non_empty(); ) { 966 // transform each slice of the original memphi: 967 mms.set_memory(_gvn.transform(mms.memory())); 968 } 969 970 if (tf()->range()->cnt() > TypeFunc::Parms) { 971 const Type* ret_type = tf()->range()->field_at(TypeFunc::Parms); 972 Node* ret_phi = _gvn.transform( _exits.argument(0) ); 973 assert(_exits.control()->is_top() || !_gvn.type(ret_phi)->empty(), "return value must be well defined"); 974 _exits.push_node(ret_type->basic_type(), ret_phi); 975 } 976 977 // Note: Logic for creating and optimizing the ReturnNode is in Compile. 978 979 // Unlock along the exceptional paths. 980 // This is done late so that we can common up equivalent exceptions 981 // (e.g., null checks) arising from multiple points within this method. 982 // See GraphKit::add_exception_state, which performs the commoning. 983 bool do_synch = method()->is_synchronized() && GenerateSynchronizationCode; 984 985 // record exit from a method if compiled while Dtrace is turned on. 986 if (do_synch || C->env()->dtrace_method_probes()) { 987 // First move the exception list out of _exits: 988 GraphKit kit(_exits.transfer_exceptions_into_jvms()); 989 SafePointNode* normal_map = kit.map(); // keep this guy safe 990 // Now re-collect the exceptions into _exits: 991 SafePointNode* ex_map; 992 while ((ex_map = kit.pop_exception_state()) != NULL) { 993 Node* ex_oop = kit.use_exception_state(ex_map); 994 // Force the exiting JVM state to have this method at InvocationEntryBci. 995 // The exiting JVM state is otherwise a copy of the calling JVMS. 996 JVMState* caller = kit.jvms(); 997 JVMState* ex_jvms = caller->clone_shallow(C); 998 ex_jvms->set_map(kit.clone_map()); 999 ex_jvms->map()->set_jvms(ex_jvms); 1000 ex_jvms->set_bci( InvocationEntryBci); 1001 kit.set_jvms(ex_jvms); 1002 if (do_synch) { 1003 // Add on the synchronized-method box/object combo 1004 kit.map()->push_monitor(_synch_lock); 1005 // Unlock! 1006 kit.shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node()); 1007 } 1008 if (C->env()->dtrace_method_probes()) { 1009 kit.make_dtrace_method_exit(method()); 1010 } 1011 // Done with exception-path processing. 1012 ex_map = kit.make_exception_state(ex_oop); 1013 assert(ex_jvms->same_calls_as(ex_map->jvms()), "sanity"); 1014 // Pop the last vestige of this method: 1015 ex_map->set_jvms(caller->clone_shallow(C)); 1016 ex_map->jvms()->set_map(ex_map); 1017 _exits.push_exception_state(ex_map); 1018 } 1019 assert(_exits.map() == normal_map, "keep the same return state"); 1020 } 1021 1022 { 1023 // Capture very early exceptions (receiver null checks) from caller JVMS 1024 GraphKit caller(_caller); 1025 SafePointNode* ex_map; 1026 while ((ex_map = caller.pop_exception_state()) != NULL) { 1027 _exits.add_exception_state(ex_map); 1028 } 1029 } 1030 } 1031 1032 //-----------------------------create_entry_map------------------------------- 1033 // Initialize our parser map to contain the types at method entry. 1034 // For OSR, the map contains a single RawPtr parameter. 1035 // Initial monitor locking for sync. methods is performed by do_method_entry. 1036 SafePointNode* Parse::create_entry_map() { 1037 // Check for really stupid bail-out cases. 1038 uint len = TypeFunc::Parms + method()->max_locals() + method()->max_stack(); 1039 if (len >= 32760) { 1040 C->record_method_not_compilable_all_tiers("too many local variables"); 1041 return NULL; 1042 } 1043 1044 // If this is an inlined method, we may have to do a receiver null check. 1045 if (_caller->has_method() && is_normal_parse() && !method()->is_static()) { 1046 GraphKit kit(_caller); 1047 kit.null_check_receiver_before_call(method()); 1048 _caller = kit.transfer_exceptions_into_jvms(); 1049 if (kit.stopped()) { 1050 _exits.add_exception_states_from(_caller); 1051 _exits.set_jvms(_caller); 1052 return NULL; 1053 } 1054 } 1055 1056 assert(method() != NULL, "parser must have a method"); 1057 1058 // Create an initial safepoint to hold JVM state during parsing 1059 JVMState* jvms = new (C) JVMState(method(), _caller->has_method() ? _caller : NULL); 1060 set_map(new (C) SafePointNode(len, jvms)); 1061 jvms->set_map(map()); 1062 record_for_igvn(map()); 1063 assert(jvms->endoff() == len, "correct jvms sizing"); 1064 1065 SafePointNode* inmap = _caller->map(); 1066 assert(inmap != NULL, "must have inmap"); 1067 1068 uint i; 1069 1070 // Pass thru the predefined input parameters. 1071 for (i = 0; i < TypeFunc::Parms; i++) { 1072 map()->init_req(i, inmap->in(i)); 1073 } 1074 1075 if (depth() == 1) { 1076 assert(map()->memory()->Opcode() == Op_Parm, ""); 1077 // Insert the memory aliasing node 1078 set_all_memory(reset_memory()); 1079 } 1080 assert(merged_memory(), ""); 1081 1082 // Now add the locals which are initially bound to arguments: 1083 uint arg_size = tf()->domain()->cnt(); 1084 ensure_stack(arg_size - TypeFunc::Parms); // OSR methods have funny args 1085 for (i = TypeFunc::Parms; i < arg_size; i++) { 1086 map()->init_req(i, inmap->argument(_caller, i - TypeFunc::Parms)); 1087 } 1088 1089 // Clear out the rest of the map (locals and stack) 1090 for (i = arg_size; i < len; i++) { 1091 map()->init_req(i, top()); 1092 } 1093 1094 SafePointNode* entry_map = stop(); 1095 return entry_map; 1096 } 1097 1098 //-----------------------------do_method_entry-------------------------------- 1099 // Emit any code needed in the pseudo-block before BCI zero. 1100 // The main thing to do is lock the receiver of a synchronized method. 1101 void Parse::do_method_entry() { 1102 set_parse_bci(InvocationEntryBci); // Pseudo-BCP 1103 set_sp(0); // Java Stack Pointer 1104 1105 NOT_PRODUCT( count_compiled_calls(true/*at_method_entry*/, false/*is_inline*/); ) 1106 1107 if (C->env()->dtrace_method_probes()) { 1108 make_dtrace_method_entry(method()); 1109 } 1110 1111 // If the method is synchronized, we need to construct a lock node, attach 1112 // it to the Start node, and pin it there. 1113 if (method()->is_synchronized()) { 1114 // Insert a FastLockNode right after the Start which takes as arguments 1115 // the current thread pointer, the "this" pointer & the address of the 1116 // stack slot pair used for the lock. The "this" pointer is a projection 1117 // off the start node, but the locking spot has to be constructed by 1118 // creating a ConLNode of 0, and boxing it with a BoxLockNode. The BoxLockNode 1119 // becomes the second argument to the FastLockNode call. The 1120 // FastLockNode becomes the new control parent to pin it to the start. 1121 1122 // Setup Object Pointer 1123 Node *lock_obj = NULL; 1124 if(method()->is_static()) { 1125 ciInstance* mirror = _method->holder()->java_mirror(); 1126 const TypeInstPtr *t_lock = TypeInstPtr::make(mirror); 1127 lock_obj = makecon(t_lock); 1128 } else { // Else pass the "this" pointer, 1129 lock_obj = local(0); // which is Parm0 from StartNode 1130 } 1131 // Clear out dead values from the debug info. 1132 kill_dead_locals(); 1133 // Build the FastLockNode 1134 _synch_lock = shared_lock(lock_obj); 1135 } 1136 1137 // Feed profiling data for parameters to the type system so it can 1138 // propagate it as speculative types 1139 record_profiled_parameters_for_speculation(); 1140 1141 if (depth() == 1) { 1142 increment_and_test_invocation_counter(Tier2CompileThreshold); 1143 } 1144 } 1145 1146 //------------------------------init_blocks------------------------------------ 1147 // Initialize our parser map to contain the types/monitors at method entry. 1148 void Parse::init_blocks() { 1149 // Create the blocks. 1150 _block_count = flow()->block_count(); 1151 _blocks = NEW_RESOURCE_ARRAY(Block, _block_count); 1152 Copy::zero_to_bytes(_blocks, sizeof(Block)*_block_count); 1153 1154 int rpo; 1155 1156 // Initialize the structs. 1157 for (rpo = 0; rpo < block_count(); rpo++) { 1158 Block* block = rpo_at(rpo); 1159 block->init_node(this, rpo); 1160 } 1161 1162 // Collect predecessor and successor information. 1163 for (rpo = 0; rpo < block_count(); rpo++) { 1164 Block* block = rpo_at(rpo); 1165 block->init_graph(this); 1166 } 1167 } 1168 1169 //-------------------------------init_node------------------------------------- 1170 void Parse::Block::init_node(Parse* outer, int rpo) { 1171 _flow = outer->flow()->rpo_at(rpo); 1172 _pred_count = 0; 1173 _preds_parsed = 0; 1174 _count = 0; 1175 assert(pred_count() == 0 && preds_parsed() == 0, "sanity"); 1176 assert(!(is_merged() || is_parsed() || is_handler() || has_merged_backedge()), "sanity"); 1177 assert(_live_locals.size() == 0, "sanity"); 1178 1179 // entry point has additional predecessor 1180 if (flow()->is_start()) _pred_count++; 1181 assert(flow()->is_start() == (this == outer->start_block()), ""); 1182 } 1183 1184 //-------------------------------init_graph------------------------------------ 1185 void Parse::Block::init_graph(Parse* outer) { 1186 // Create the successor list for this parser block. 1187 GrowableArray<ciTypeFlow::Block*>* tfs = flow()->successors(); 1188 GrowableArray<ciTypeFlow::Block*>* tfe = flow()->exceptions(); 1189 int ns = tfs->length(); 1190 int ne = tfe->length(); 1191 _num_successors = ns; 1192 _all_successors = ns+ne; 1193 _successors = (ns+ne == 0) ? NULL : NEW_RESOURCE_ARRAY(Block*, ns+ne); 1194 int p = 0; 1195 for (int i = 0; i < ns+ne; i++) { 1196 ciTypeFlow::Block* tf2 = (i < ns) ? tfs->at(i) : tfe->at(i-ns); 1197 Block* block2 = outer->rpo_at(tf2->rpo()); 1198 _successors[i] = block2; 1199 1200 // Accumulate pred info for the other block, too. 1201 if (i < ns) { 1202 block2->_pred_count++; 1203 } else { 1204 block2->_is_handler = true; 1205 } 1206 1207 #ifdef ASSERT 1208 // A block's successors must be distinguishable by BCI. 1209 // That is, no bytecode is allowed to branch to two different 1210 // clones of the same code location. 1211 for (int j = 0; j < i; j++) { 1212 Block* block1 = _successors[j]; 1213 if (block1 == block2) continue; // duplicates are OK 1214 assert(block1->start() != block2->start(), "successors have unique bcis"); 1215 } 1216 #endif 1217 } 1218 1219 // Note: We never call next_path_num along exception paths, so they 1220 // never get processed as "ready". Also, the input phis of exception 1221 // handlers get specially processed, so that 1222 } 1223 1224 //---------------------------successor_for_bci--------------------------------- 1225 Parse::Block* Parse::Block::successor_for_bci(int bci) { 1226 for (int i = 0; i < all_successors(); i++) { 1227 Block* block2 = successor_at(i); 1228 if (block2->start() == bci) return block2; 1229 } 1230 // We can actually reach here if ciTypeFlow traps out a block 1231 // due to an unloaded class, and concurrently with compilation the 1232 // class is then loaded, so that a later phase of the parser is 1233 // able to see more of the bytecode CFG. Or, the flow pass and 1234 // the parser can have a minor difference of opinion about executability 1235 // of bytecodes. For example, "obj.field = null" is executable even 1236 // if the field's type is an unloaded class; the flow pass used to 1237 // make a trap for such code. 1238 return NULL; 1239 } 1240 1241 1242 //-----------------------------stack_type_at----------------------------------- 1243 const Type* Parse::Block::stack_type_at(int i) const { 1244 return get_type(flow()->stack_type_at(i)); 1245 } 1246 1247 1248 //-----------------------------local_type_at----------------------------------- 1249 const Type* Parse::Block::local_type_at(int i) const { 1250 // Make dead locals fall to bottom. 1251 if (_live_locals.size() == 0) { 1252 MethodLivenessResult live_locals = flow()->outer()->method()->liveness_at_bci(start()); 1253 // This bitmap can be zero length if we saw a breakpoint. 1254 // In such cases, pretend they are all live. 1255 ((Block*)this)->_live_locals = live_locals; 1256 } 1257 if (_live_locals.size() > 0 && !_live_locals.at(i)) 1258 return Type::BOTTOM; 1259 1260 return get_type(flow()->local_type_at(i)); 1261 } 1262 1263 1264 #ifndef PRODUCT 1265 1266 //----------------------------name_for_bc-------------------------------------- 1267 // helper method for BytecodeParseHistogram 1268 static const char* name_for_bc(int i) { 1269 return Bytecodes::is_defined(i) ? Bytecodes::name(Bytecodes::cast(i)) : "xxxunusedxxx"; 1270 } 1271 1272 //----------------------------BytecodeParseHistogram------------------------------------ 1273 Parse::BytecodeParseHistogram::BytecodeParseHistogram(Parse *p, Compile *c) { 1274 _parser = p; 1275 _compiler = c; 1276 if( ! _initialized ) { _initialized = true; reset(); } 1277 } 1278 1279 //----------------------------current_count------------------------------------ 1280 int Parse::BytecodeParseHistogram::current_count(BPHType bph_type) { 1281 switch( bph_type ) { 1282 case BPH_transforms: { return _parser->gvn().made_progress(); } 1283 case BPH_values: { return _parser->gvn().made_new_values(); } 1284 default: { ShouldNotReachHere(); return 0; } 1285 } 1286 } 1287 1288 //----------------------------initialized-------------------------------------- 1289 bool Parse::BytecodeParseHistogram::initialized() { return _initialized; } 1290 1291 //----------------------------reset-------------------------------------------- 1292 void Parse::BytecodeParseHistogram::reset() { 1293 int i = Bytecodes::number_of_codes; 1294 while (i-- > 0) { _bytecodes_parsed[i] = 0; _nodes_constructed[i] = 0; _nodes_transformed[i] = 0; _new_values[i] = 0; } 1295 } 1296 1297 //----------------------------set_initial_state-------------------------------- 1298 // Record info when starting to parse one bytecode 1299 void Parse::BytecodeParseHistogram::set_initial_state( Bytecodes::Code bc ) { 1300 if( PrintParseStatistics && !_parser->is_osr_parse() ) { 1301 _initial_bytecode = bc; 1302 _initial_node_count = _compiler->unique(); 1303 _initial_transforms = current_count(BPH_transforms); 1304 _initial_values = current_count(BPH_values); 1305 } 1306 } 1307 1308 //----------------------------record_change-------------------------------- 1309 // Record results of parsing one bytecode 1310 void Parse::BytecodeParseHistogram::record_change() { 1311 if( PrintParseStatistics && !_parser->is_osr_parse() ) { 1312 ++_bytecodes_parsed[_initial_bytecode]; 1313 _nodes_constructed [_initial_bytecode] += (_compiler->unique() - _initial_node_count); 1314 _nodes_transformed [_initial_bytecode] += (current_count(BPH_transforms) - _initial_transforms); 1315 _new_values [_initial_bytecode] += (current_count(BPH_values) - _initial_values); 1316 } 1317 } 1318 1319 1320 //----------------------------print-------------------------------------------- 1321 void Parse::BytecodeParseHistogram::print(float cutoff) { 1322 ResourceMark rm; 1323 // print profile 1324 int total = 0; 1325 int i = 0; 1326 for( i = 0; i < Bytecodes::number_of_codes; ++i ) { total += _bytecodes_parsed[i]; } 1327 int abs_sum = 0; 1328 tty->cr(); //0123456789012345678901234567890123456789012345678901234567890123456789 1329 tty->print_cr("Histogram of %d parsed bytecodes:", total); 1330 if( total == 0 ) { return; } 1331 tty->cr(); 1332 tty->print_cr("absolute: count of compiled bytecodes of this type"); 1333 tty->print_cr("relative: percentage contribution to compiled nodes"); 1334 tty->print_cr("nodes : Average number of nodes constructed per bytecode"); 1335 tty->print_cr("rnodes : Significance towards total nodes constructed, (nodes*relative)"); 1336 tty->print_cr("transforms: Average amount of tranform progress per bytecode compiled"); 1337 tty->print_cr("values : Average number of node values improved per bytecode"); 1338 tty->print_cr("name : Bytecode name"); 1339 tty->cr(); 1340 tty->print_cr(" absolute relative nodes rnodes transforms values name"); 1341 tty->print_cr("----------------------------------------------------------------------"); 1342 while (--i > 0) { 1343 int abs = _bytecodes_parsed[i]; 1344 float rel = abs * 100.0F / total; 1345 float nodes = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _nodes_constructed[i])/_bytecodes_parsed[i]; 1346 float rnodes = _bytecodes_parsed[i] == 0 ? 0 : rel * nodes; 1347 float xforms = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _nodes_transformed[i])/_bytecodes_parsed[i]; 1348 float values = _bytecodes_parsed[i] == 0 ? 0 : (1.0F * _new_values [i])/_bytecodes_parsed[i]; 1349 if (cutoff <= rel) { 1350 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)); 1351 abs_sum += abs; 1352 } 1353 } 1354 tty->print_cr("----------------------------------------------------------------------"); 1355 float rel_sum = abs_sum * 100.0F / total; 1356 tty->print_cr("%10d %7.2f%% (cutoff = %.2f%%)", abs_sum, rel_sum, cutoff); 1357 tty->print_cr("----------------------------------------------------------------------"); 1358 tty->cr(); 1359 } 1360 #endif 1361 1362 //----------------------------load_state_from---------------------------------- 1363 // Load block/map/sp. But not do not touch iter/bci. 1364 void Parse::load_state_from(Block* block) { 1365 set_block(block); 1366 // load the block's JVM state: 1367 set_map(block->start_map()); 1368 set_sp( block->start_sp()); 1369 } 1370 1371 1372 //-----------------------------record_state------------------------------------ 1373 void Parse::Block::record_state(Parse* p) { 1374 assert(!is_merged(), "can only record state once, on 1st inflow"); 1375 assert(start_sp() == p->sp(), "stack pointer must agree with ciTypeFlow"); 1376 set_start_map(p->stop()); 1377 } 1378 1379 1380 //------------------------------do_one_block----------------------------------- 1381 void Parse::do_one_block() { 1382 if (TraceOptoParse) { 1383 Block *b = block(); 1384 int ns = b->num_successors(); 1385 int nt = b->all_successors(); 1386 1387 tty->print("Parsing block #%d at bci [%d,%d), successors: ", 1388 block()->rpo(), block()->start(), block()->limit()); 1389 for (int i = 0; i < nt; i++) { 1390 tty->print((( i < ns) ? " %d" : " %d(e)"), b->successor_at(i)->rpo()); 1391 } 1392 if (b->is_loop_head()) tty->print(" lphd"); 1393 tty->print_cr(""); 1394 } 1395 1396 assert(block()->is_merged(), "must be merged before being parsed"); 1397 block()->mark_parsed(); 1398 ++_blocks_parsed; 1399 1400 // Set iterator to start of block. 1401 iter().reset_to_bci(block()->start()); 1402 1403 CompileLog* log = C->log(); 1404 1405 // Parse bytecodes 1406 while (!stopped() && !failing()) { 1407 iter().next(); 1408 1409 // Learn the current bci from the iterator: 1410 set_parse_bci(iter().cur_bci()); 1411 1412 if (bci() == block()->limit()) { 1413 // Do not walk into the next block until directed by do_all_blocks. 1414 merge(bci()); 1415 break; 1416 } 1417 assert(bci() < block()->limit(), "bci still in block"); 1418 1419 if (log != NULL) { 1420 // Output an optional context marker, to help place actions 1421 // that occur during parsing of this BC. If there is no log 1422 // output until the next context string, this context string 1423 // will be silently ignored. 1424 log->set_context("bc code='%d' bci='%d'", (int)bc(), bci()); 1425 } 1426 1427 if (block()->has_trap_at(bci())) { 1428 // We must respect the flow pass's traps, because it will refuse 1429 // to produce successors for trapping blocks. 1430 int trap_index = block()->flow()->trap_index(); 1431 assert(trap_index != 0, "trap index must be valid"); 1432 uncommon_trap(trap_index); 1433 break; 1434 } 1435 1436 NOT_PRODUCT( parse_histogram()->set_initial_state(bc()); ); 1437 1438 #ifdef ASSERT 1439 int pre_bc_sp = sp(); 1440 int inputs, depth; 1441 bool have_se = !stopped() && compute_stack_effects(inputs, depth); 1442 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)); 1443 #endif //ASSERT 1444 1445 do_one_bytecode(); 1446 1447 assert(!have_se || stopped() || failing() || (sp() - pre_bc_sp) == depth, 1448 err_msg_res("incorrect depth prediction: sp=%d, pre_bc_sp=%d, depth=%d", sp(), pre_bc_sp, depth)); 1449 1450 do_exceptions(); 1451 1452 NOT_PRODUCT( parse_histogram()->record_change(); ); 1453 1454 if (log != NULL) 1455 log->clear_context(); // skip marker if nothing was printed 1456 1457 // Fall into next bytecode. Each bytecode normally has 1 sequential 1458 // successor which is typically made ready by visiting this bytecode. 1459 // If the successor has several predecessors, then it is a merge 1460 // point, starts a new basic block, and is handled like other basic blocks. 1461 } 1462 } 1463 1464 1465 //------------------------------merge------------------------------------------ 1466 void Parse::set_parse_bci(int bci) { 1467 set_bci(bci); 1468 Node_Notes* nn = C->default_node_notes(); 1469 if (nn == NULL) return; 1470 1471 // Collect debug info for inlined calls unless -XX:-DebugInlinedCalls. 1472 if (!DebugInlinedCalls && depth() > 1) { 1473 return; 1474 } 1475 1476 // Update the JVMS annotation, if present. 1477 JVMState* jvms = nn->jvms(); 1478 if (jvms != NULL && jvms->bci() != bci) { 1479 // Update the JVMS. 1480 jvms = jvms->clone_shallow(C); 1481 jvms->set_bci(bci); 1482 nn->set_jvms(jvms); 1483 } 1484 } 1485 1486 //------------------------------merge------------------------------------------ 1487 // Merge the current mapping into the basic block starting at bci 1488 void Parse::merge(int target_bci) { 1489 Block* target = successor_for_bci(target_bci); 1490 if (target == NULL) { handle_missing_successor(target_bci); return; } 1491 assert(!target->is_ready(), "our arrival must be expected"); 1492 int pnum = target->next_path_num(); 1493 merge_common(target, pnum); 1494 } 1495 1496 //-------------------------merge_new_path-------------------------------------- 1497 // Merge the current mapping into the basic block, using a new path 1498 void Parse::merge_new_path(int target_bci) { 1499 Block* target = successor_for_bci(target_bci); 1500 if (target == NULL) { handle_missing_successor(target_bci); return; } 1501 assert(!target->is_ready(), "new path into frozen graph"); 1502 int pnum = target->add_new_path(); 1503 merge_common(target, pnum); 1504 } 1505 1506 //-------------------------merge_exception------------------------------------- 1507 // Merge the current mapping into the basic block starting at bci 1508 // The ex_oop must be pushed on the stack, unlike throw_to_exit. 1509 void Parse::merge_exception(int target_bci) { 1510 assert(sp() == 1, "must have only the throw exception on the stack"); 1511 Block* target = successor_for_bci(target_bci); 1512 if (target == NULL) { handle_missing_successor(target_bci); return; } 1513 assert(target->is_handler(), "exceptions are handled by special blocks"); 1514 int pnum = target->add_new_path(); 1515 merge_common(target, pnum); 1516 } 1517 1518 //--------------------handle_missing_successor--------------------------------- 1519 void Parse::handle_missing_successor(int target_bci) { 1520 #ifndef PRODUCT 1521 Block* b = block(); 1522 int trap_bci = b->flow()->has_trap()? b->flow()->trap_bci(): -1; 1523 tty->print_cr("### Missing successor at bci:%d for block #%d (trap_bci:%d)", target_bci, b->rpo(), trap_bci); 1524 #endif 1525 ShouldNotReachHere(); 1526 } 1527 1528 //--------------------------merge_common--------------------------------------- 1529 void Parse::merge_common(Parse::Block* target, int pnum) { 1530 if (TraceOptoParse) { 1531 tty->print("Merging state at block #%d bci:%d", target->rpo(), target->start()); 1532 } 1533 1534 // Zap extra stack slots to top 1535 assert(sp() == target->start_sp(), ""); 1536 clean_stack(sp()); 1537 1538 if (!target->is_merged()) { // No prior mapping at this bci 1539 if (TraceOptoParse) { tty->print(" with empty state"); } 1540 1541 // If this path is dead, do not bother capturing it as a merge. 1542 // It is "as if" we had 1 fewer predecessors from the beginning. 1543 if (stopped()) { 1544 if (TraceOptoParse) tty->print_cr(", but path is dead and doesn't count"); 1545 return; 1546 } 1547 1548 // Record that a new block has been merged. 1549 ++_blocks_merged; 1550 1551 // Make a region if we know there are multiple or unpredictable inputs. 1552 // (Also, if this is a plain fall-through, we might see another region, 1553 // which must not be allowed into this block's map.) 1554 if (pnum > PhiNode::Input // Known multiple inputs. 1555 || target->is_handler() // These have unpredictable inputs. 1556 || target->is_loop_head() // Known multiple inputs 1557 || control()->is_Region()) { // We must hide this guy. 1558 1559 int current_bci = bci(); 1560 set_parse_bci(target->start()); // Set target bci 1561 if (target->is_SEL_head()) { 1562 DEBUG_ONLY( target->mark_merged_backedge(block()); ) 1563 if (target->start() == 0) { 1564 // Add loop predicate for the special case when 1565 // there are backbranches to the method entry. 1566 add_predicate(); 1567 } 1568 } 1569 // Add a Region to start the new basic block. Phis will be added 1570 // later lazily. 1571 int edges = target->pred_count(); 1572 if (edges < pnum) edges = pnum; // might be a new path! 1573 RegionNode *r = new (C) RegionNode(edges+1); 1574 gvn().set_type(r, Type::CONTROL); 1575 record_for_igvn(r); 1576 // zap all inputs to NULL for debugging (done in Node(uint) constructor) 1577 // for (int j = 1; j < edges+1; j++) { r->init_req(j, NULL); } 1578 r->init_req(pnum, control()); 1579 set_control(r); 1580 set_parse_bci(current_bci); // Restore bci 1581 } 1582 1583 // Convert the existing Parser mapping into a mapping at this bci. 1584 store_state_to(target); 1585 assert(target->is_merged(), "do not come here twice"); 1586 1587 } else { // Prior mapping at this bci 1588 if (TraceOptoParse) { tty->print(" with previous state"); } 1589 #ifdef ASSERT 1590 if (target->is_SEL_head()) { 1591 target->mark_merged_backedge(block()); 1592 } 1593 #endif 1594 // We must not manufacture more phis if the target is already parsed. 1595 bool nophi = target->is_parsed(); 1596 1597 SafePointNode* newin = map();// Hang on to incoming mapping 1598 Block* save_block = block(); // Hang on to incoming block; 1599 load_state_from(target); // Get prior mapping 1600 1601 assert(newin->jvms()->locoff() == jvms()->locoff(), "JVMS layouts agree"); 1602 assert(newin->jvms()->stkoff() == jvms()->stkoff(), "JVMS layouts agree"); 1603 assert(newin->jvms()->monoff() == jvms()->monoff(), "JVMS layouts agree"); 1604 assert(newin->jvms()->endoff() == jvms()->endoff(), "JVMS layouts agree"); 1605 1606 // Iterate over my current mapping and the old mapping. 1607 // Where different, insert Phi functions. 1608 // Use any existing Phi functions. 1609 assert(control()->is_Region(), "must be merging to a region"); 1610 RegionNode* r = control()->as_Region(); 1611 1612 // Compute where to merge into 1613 // Merge incoming control path 1614 r->init_req(pnum, newin->control()); 1615 1616 if (pnum == 1) { // Last merge for this Region? 1617 if (!block()->flow()->is_irreducible_entry()) { 1618 Node* result = _gvn.transform_no_reclaim(r); 1619 if (r != result && TraceOptoParse) { 1620 tty->print_cr("Block #%d replace %d with %d", block()->rpo(), r->_idx, result->_idx); 1621 } 1622 } 1623 record_for_igvn(r); 1624 } 1625 1626 // Update all the non-control inputs to map: 1627 assert(TypeFunc::Parms == newin->jvms()->locoff(), "parser map should contain only youngest jvms"); 1628 bool check_elide_phi = target->is_SEL_backedge(save_block); 1629 for (uint j = 1; j < newin->req(); j++) { 1630 Node* m = map()->in(j); // Current state of target. 1631 Node* n = newin->in(j); // Incoming change to target state. 1632 PhiNode* phi; 1633 if (m->is_Phi() && m->as_Phi()->region() == r) 1634 phi = m->as_Phi(); 1635 else 1636 phi = NULL; 1637 if (m != n) { // Different; must merge 1638 switch (j) { 1639 // Frame pointer and Return Address never changes 1640 case TypeFunc::FramePtr:// Drop m, use the original value 1641 case TypeFunc::ReturnAdr: 1642 break; 1643 case TypeFunc::Memory: // Merge inputs to the MergeMem node 1644 assert(phi == NULL, "the merge contains phis, not vice versa"); 1645 merge_memory_edges(n->as_MergeMem(), pnum, nophi); 1646 continue; 1647 default: // All normal stuff 1648 if (phi == NULL) { 1649 const JVMState* jvms = map()->jvms(); 1650 if (EliminateNestedLocks && 1651 jvms->is_mon(j) && jvms->is_monitor_box(j)) { 1652 // BoxLock nodes are not commoning. 1653 // Use old BoxLock node as merged box. 1654 assert(newin->jvms()->is_monitor_box(j), "sanity"); 1655 // This assert also tests that nodes are BoxLock. 1656 assert(BoxLockNode::same_slot(n, m), "sanity"); 1657 C->gvn_replace_by(n, m); 1658 } else if (!check_elide_phi || !target->can_elide_SEL_phi(j)) { 1659 phi = ensure_phi(j, nophi); 1660 } 1661 } 1662 break; 1663 } 1664 } 1665 // At this point, n might be top if: 1666 // - there is no phi (because TypeFlow detected a conflict), or 1667 // - the corresponding control edges is top (a dead incoming path) 1668 // It is a bug if we create a phi which sees a garbage value on a live path. 1669 1670 if (phi != NULL) { 1671 assert(n != top() || r->in(pnum) == top(), "live value must not be garbage"); 1672 assert(phi->region() == r, ""); 1673 phi->set_req(pnum, n); // Then add 'n' to the merge 1674 if (pnum == PhiNode::Input) { 1675 // Last merge for this Phi. 1676 // So far, Phis have had a reasonable type from ciTypeFlow. 1677 // Now _gvn will join that with the meet of current inputs. 1678 // BOTTOM is never permissible here, 'cause pessimistically 1679 // Phis of pointers cannot lose the basic pointer type. 1680 debug_only(const Type* bt1 = phi->bottom_type()); 1681 assert(bt1 != Type::BOTTOM, "should not be building conflict phis"); 1682 map()->set_req(j, _gvn.transform_no_reclaim(phi)); 1683 debug_only(const Type* bt2 = phi->bottom_type()); 1684 assert(bt2->higher_equal_speculative(bt1), "must be consistent with type-flow"); 1685 record_for_igvn(phi); 1686 } 1687 } 1688 } // End of for all values to be merged 1689 1690 if (pnum == PhiNode::Input && 1691 !r->in(0)) { // The occasional useless Region 1692 assert(control() == r, ""); 1693 set_control(r->nonnull_req()); 1694 } 1695 1696 // newin has been subsumed into the lazy merge, and is now dead. 1697 set_block(save_block); 1698 1699 stop(); // done with this guy, for now 1700 } 1701 1702 if (TraceOptoParse) { 1703 tty->print_cr(" on path %d", pnum); 1704 } 1705 1706 // Done with this parser state. 1707 assert(stopped(), ""); 1708 } 1709 1710 1711 //--------------------------merge_memory_edges--------------------------------- 1712 void Parse::merge_memory_edges(MergeMemNode* n, int pnum, bool nophi) { 1713 // (nophi means we must not create phis, because we already parsed here) 1714 assert(n != NULL, ""); 1715 // Merge the inputs to the MergeMems 1716 MergeMemNode* m = merged_memory(); 1717 1718 assert(control()->is_Region(), "must be merging to a region"); 1719 RegionNode* r = control()->as_Region(); 1720 1721 PhiNode* base = NULL; 1722 MergeMemNode* remerge = NULL; 1723 for (MergeMemStream mms(m, n); mms.next_non_empty2(); ) { 1724 Node *p = mms.force_memory(); 1725 Node *q = mms.memory2(); 1726 if (mms.is_empty() && nophi) { 1727 // Trouble: No new splits allowed after a loop body is parsed. 1728 // Instead, wire the new split into a MergeMem on the backedge. 1729 // The optimizer will sort it out, slicing the phi. 1730 if (remerge == NULL) { 1731 assert(base != NULL, ""); 1732 assert(base->in(0) != NULL, "should not be xformed away"); 1733 remerge = MergeMemNode::make(C, base->in(pnum)); 1734 gvn().set_type(remerge, Type::MEMORY); 1735 base->set_req(pnum, remerge); 1736 } 1737 remerge->set_memory_at(mms.alias_idx(), q); 1738 continue; 1739 } 1740 assert(!q->is_MergeMem(), ""); 1741 PhiNode* phi; 1742 if (p != q) { 1743 phi = ensure_memory_phi(mms.alias_idx(), nophi); 1744 } else { 1745 if (p->is_Phi() && p->as_Phi()->region() == r) 1746 phi = p->as_Phi(); 1747 else 1748 phi = NULL; 1749 } 1750 // Insert q into local phi 1751 if (phi != NULL) { 1752 assert(phi->region() == r, ""); 1753 p = phi; 1754 phi->set_req(pnum, q); 1755 if (mms.at_base_memory()) { 1756 base = phi; // delay transforming it 1757 } else if (pnum == 1) { 1758 record_for_igvn(phi); 1759 p = _gvn.transform_no_reclaim(phi); 1760 } 1761 mms.set_memory(p);// store back through the iterator 1762 } 1763 } 1764 // Transform base last, in case we must fiddle with remerging. 1765 if (base != NULL && pnum == 1) { 1766 record_for_igvn(base); 1767 m->set_base_memory( _gvn.transform_no_reclaim(base) ); 1768 } 1769 } 1770 1771 1772 //------------------------ensure_phis_everywhere------------------------------- 1773 void Parse::ensure_phis_everywhere() { 1774 ensure_phi(TypeFunc::I_O); 1775 1776 // Ensure a phi on all currently known memories. 1777 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 1778 ensure_memory_phi(mms.alias_idx()); 1779 debug_only(mms.set_memory()); // keep the iterator happy 1780 } 1781 1782 // Note: This is our only chance to create phis for memory slices. 1783 // If we miss a slice that crops up later, it will have to be 1784 // merged into the base-memory phi that we are building here. 1785 // Later, the optimizer will comb out the knot, and build separate 1786 // phi-loops for each memory slice that matters. 1787 1788 // Monitors must nest nicely and not get confused amongst themselves. 1789 // Phi-ify everything up to the monitors, though. 1790 uint monoff = map()->jvms()->monoff(); 1791 uint nof_monitors = map()->jvms()->nof_monitors(); 1792 1793 assert(TypeFunc::Parms == map()->jvms()->locoff(), "parser map should contain only youngest jvms"); 1794 bool check_elide_phi = block()->is_SEL_head(); 1795 for (uint i = TypeFunc::Parms; i < monoff; i++) { 1796 if (!check_elide_phi || !block()->can_elide_SEL_phi(i)) { 1797 ensure_phi(i); 1798 } 1799 } 1800 1801 // Even monitors need Phis, though they are well-structured. 1802 // This is true for OSR methods, and also for the rare cases where 1803 // a monitor object is the subject of a replace_in_map operation. 1804 // See bugs 4426707 and 5043395. 1805 for (uint m = 0; m < nof_monitors; m++) { 1806 ensure_phi(map()->jvms()->monitor_obj_offset(m)); 1807 } 1808 } 1809 1810 1811 //-----------------------------add_new_path------------------------------------ 1812 // Add a previously unaccounted predecessor to this block. 1813 int Parse::Block::add_new_path() { 1814 // If there is no map, return the lowest unused path number. 1815 if (!is_merged()) return pred_count()+1; // there will be a map shortly 1816 1817 SafePointNode* map = start_map(); 1818 if (!map->control()->is_Region()) 1819 return pred_count()+1; // there may be a region some day 1820 RegionNode* r = map->control()->as_Region(); 1821 1822 // Add new path to the region. 1823 uint pnum = r->req(); 1824 r->add_req(NULL); 1825 1826 for (uint i = 1; i < map->req(); i++) { 1827 Node* n = map->in(i); 1828 if (i == TypeFunc::Memory) { 1829 // Ensure a phi on all currently known memories. 1830 for (MergeMemStream mms(n->as_MergeMem()); mms.next_non_empty(); ) { 1831 Node* phi = mms.memory(); 1832 if (phi->is_Phi() && phi->as_Phi()->region() == r) { 1833 assert(phi->req() == pnum, "must be same size as region"); 1834 phi->add_req(NULL); 1835 } 1836 } 1837 } else { 1838 if (n->is_Phi() && n->as_Phi()->region() == r) { 1839 assert(n->req() == pnum, "must be same size as region"); 1840 n->add_req(NULL); 1841 } 1842 } 1843 } 1844 1845 return pnum; 1846 } 1847 1848 //------------------------------ensure_phi------------------------------------- 1849 // Turn the idx'th entry of the current map into a Phi 1850 PhiNode *Parse::ensure_phi(int idx, bool nocreate) { 1851 SafePointNode* map = this->map(); 1852 Node* region = map->control(); 1853 assert(region->is_Region(), ""); 1854 1855 Node* o = map->in(idx); 1856 assert(o != NULL, ""); 1857 1858 if (o == top()) return NULL; // TOP always merges into TOP 1859 1860 if (o->is_Phi() && o->as_Phi()->region() == region) { 1861 return o->as_Phi(); 1862 } 1863 1864 // Now use a Phi here for merging 1865 assert(!nocreate, "Cannot build a phi for a block already parsed."); 1866 const JVMState* jvms = map->jvms(); 1867 const Type* t; 1868 if (jvms->is_loc(idx)) { 1869 t = block()->local_type_at(idx - jvms->locoff()); 1870 } else if (jvms->is_stk(idx)) { 1871 t = block()->stack_type_at(idx - jvms->stkoff()); 1872 } else if (jvms->is_mon(idx)) { 1873 assert(!jvms->is_monitor_box(idx), "no phis for boxes"); 1874 t = TypeInstPtr::BOTTOM; // this is sufficient for a lock object 1875 } else if ((uint)idx < TypeFunc::Parms) { 1876 t = o->bottom_type(); // Type::RETURN_ADDRESS or such-like. 1877 } else { 1878 assert(false, "no type information for this phi"); 1879 } 1880 1881 // If the type falls to bottom, then this must be a local that 1882 // is mixing ints and oops or some such. Forcing it to top 1883 // makes it go dead. 1884 if (t == Type::BOTTOM) { 1885 map->set_req(idx, top()); 1886 return NULL; 1887 } 1888 1889 // Do not create phis for top either. 1890 // A top on a non-null control flow must be an unused even after the.phi. 1891 if (t == Type::TOP || t == Type::HALF) { 1892 map->set_req(idx, top()); 1893 return NULL; 1894 } 1895 1896 PhiNode* phi = PhiNode::make(region, o, t); 1897 gvn().set_type(phi, t); 1898 if (C->do_escape_analysis()) record_for_igvn(phi); 1899 map->set_req(idx, phi); 1900 return phi; 1901 } 1902 1903 //--------------------------ensure_memory_phi---------------------------------- 1904 // Turn the idx'th slice of the current memory into a Phi 1905 PhiNode *Parse::ensure_memory_phi(int idx, bool nocreate) { 1906 MergeMemNode* mem = merged_memory(); 1907 Node* region = control(); 1908 assert(region->is_Region(), ""); 1909 1910 Node *o = (idx == Compile::AliasIdxBot)? mem->base_memory(): mem->memory_at(idx); 1911 assert(o != NULL && o != top(), ""); 1912 1913 PhiNode* phi; 1914 if (o->is_Phi() && o->as_Phi()->region() == region) { 1915 phi = o->as_Phi(); 1916 if (phi == mem->base_memory() && idx >= Compile::AliasIdxRaw) { 1917 // clone the shared base memory phi to make a new memory split 1918 assert(!nocreate, "Cannot build a phi for a block already parsed."); 1919 const Type* t = phi->bottom_type(); 1920 const TypePtr* adr_type = C->get_adr_type(idx); 1921 phi = phi->slice_memory(adr_type); 1922 gvn().set_type(phi, t); 1923 } 1924 return phi; 1925 } 1926 1927 // Now use a Phi here for merging 1928 assert(!nocreate, "Cannot build a phi for a block already parsed."); 1929 const Type* t = o->bottom_type(); 1930 const TypePtr* adr_type = C->get_adr_type(idx); 1931 phi = PhiNode::make(region, o, t, adr_type); 1932 gvn().set_type(phi, t); 1933 if (idx == Compile::AliasIdxBot) 1934 mem->set_base_memory(phi); 1935 else 1936 mem->set_memory_at(idx, phi); 1937 return phi; 1938 } 1939 1940 //------------------------------call_register_finalizer----------------------- 1941 // Check the klass of the receiver and call register_finalizer if the 1942 // class need finalization. 1943 void Parse::call_register_finalizer() { 1944 Node* receiver = local(0); 1945 assert(receiver != NULL && receiver->bottom_type()->isa_instptr() != NULL, 1946 "must have non-null instance type"); 1947 1948 const TypeInstPtr *tinst = receiver->bottom_type()->isa_instptr(); 1949 if (tinst != NULL && tinst->klass()->is_loaded() && !tinst->klass_is_exact()) { 1950 // The type isn't known exactly so see if CHA tells us anything. 1951 ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); 1952 if (!Dependencies::has_finalizable_subclass(ik)) { 1953 // No finalizable subclasses so skip the dynamic check. 1954 C->dependencies()->assert_has_no_finalizable_subclasses(ik); 1955 return; 1956 } 1957 } 1958 1959 // Insert a dynamic test for whether the instance needs 1960 // finalization. In general this will fold up since the concrete 1961 // class is often visible so the access flags are constant. 1962 Node* klass_addr = basic_plus_adr( receiver, receiver, oopDesc::klass_offset_in_bytes() ); 1963 Node* klass = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), klass_addr, TypeInstPtr::KLASS) ); 1964 1965 Node* access_flags_addr = basic_plus_adr(klass, klass, in_bytes(Klass::access_flags_offset())); 1966 Node* access_flags = make_load(NULL, access_flags_addr, TypeInt::INT, T_INT, MemNode::unordered); 1967 1968 Node* mask = _gvn.transform(new (C) AndINode(access_flags, intcon(JVM_ACC_HAS_FINALIZER))); 1969 Node* check = _gvn.transform(new (C) CmpINode(mask, intcon(0))); 1970 Node* test = _gvn.transform(new (C) BoolNode(check, BoolTest::ne)); 1971 1972 IfNode* iff = create_and_map_if(control(), test, PROB_MAX, COUNT_UNKNOWN); 1973 1974 RegionNode* result_rgn = new (C) RegionNode(3); 1975 record_for_igvn(result_rgn); 1976 1977 Node *skip_register = _gvn.transform(new (C) IfFalseNode(iff)); 1978 result_rgn->init_req(1, skip_register); 1979 1980 Node *needs_register = _gvn.transform(new (C) IfTrueNode(iff)); 1981 set_control(needs_register); 1982 if (stopped()) { 1983 // There is no slow path. 1984 result_rgn->init_req(2, top()); 1985 } else { 1986 Node *call = make_runtime_call(RC_NO_LEAF, 1987 OptoRuntime::register_finalizer_Type(), 1988 OptoRuntime::register_finalizer_Java(), 1989 NULL, TypePtr::BOTTOM, 1990 receiver); 1991 make_slow_call_ex(call, env()->Throwable_klass(), true); 1992 1993 Node* fast_io = call->in(TypeFunc::I_O); 1994 Node* fast_mem = call->in(TypeFunc::Memory); 1995 // These two phis are pre-filled with copies of of the fast IO and Memory 1996 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO); 1997 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 1998 1999 result_rgn->init_req(2, control()); 2000 io_phi ->init_req(2, i_o()); 2001 mem_phi ->init_req(2, reset_memory()); 2002 2003 set_all_memory( _gvn.transform(mem_phi) ); 2004 set_i_o( _gvn.transform(io_phi) ); 2005 } 2006 2007 set_control( _gvn.transform(result_rgn) ); 2008 } 2009 2010 //------------------------------return_current--------------------------------- 2011 // Append current _map to _exit_return 2012 void Parse::return_current(Node* value) { 2013 if (RegisterFinalizersAtInit && 2014 method()->intrinsic_id() == vmIntrinsics::_Object_init) { 2015 call_register_finalizer(); 2016 } 2017 2018 // Do not set_parse_bci, so that return goo is credited to the return insn. 2019 set_bci(InvocationEntryBci); 2020 if (method()->is_synchronized() && GenerateSynchronizationCode) { 2021 shared_unlock(_synch_lock->box_node(), _synch_lock->obj_node()); 2022 } 2023 if (C->env()->dtrace_method_probes()) { 2024 make_dtrace_method_exit(method()); 2025 } 2026 SafePointNode* exit_return = _exits.map(); 2027 exit_return->in( TypeFunc::Control )->add_req( control() ); 2028 exit_return->in( TypeFunc::I_O )->add_req( i_o () ); 2029 Node *mem = exit_return->in( TypeFunc::Memory ); 2030 for (MergeMemStream mms(mem->as_MergeMem(), merged_memory()); mms.next_non_empty2(); ) { 2031 if (mms.is_empty()) { 2032 // get a copy of the base memory, and patch just this one input 2033 const TypePtr* adr_type = mms.adr_type(C); 2034 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type); 2035 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), ""); 2036 gvn().set_type_bottom(phi); 2037 phi->del_req(phi->req()-1); // prepare to re-patch 2038 mms.set_memory(phi); 2039 } 2040 mms.memory()->add_req(mms.memory2()); 2041 } 2042 2043 // frame pointer is always same, already captured 2044 if (value != NULL) { 2045 // If returning oops to an interface-return, there is a silent free 2046 // cast from oop to interface allowed by the Verifier. Make it explicit 2047 // here. 2048 Node* phi = _exits.argument(0); 2049 const TypeInstPtr *tr = phi->bottom_type()->isa_instptr(); 2050 if( tr && tr->klass()->is_loaded() && 2051 tr->klass()->is_interface() ) { 2052 const TypeInstPtr *tp = value->bottom_type()->isa_instptr(); 2053 if (tp && tp->klass()->is_loaded() && 2054 !tp->klass()->is_interface()) { 2055 // sharpen the type eagerly; this eases certain assert checking 2056 if (tp->higher_equal(TypeInstPtr::NOTNULL)) 2057 tr = tr->join_speculative(TypeInstPtr::NOTNULL)->is_instptr(); 2058 value = _gvn.transform(new (C) CheckCastPPNode(0,value,tr)); 2059 } 2060 } 2061 phi->add_req(value); 2062 } 2063 2064 stop_and_kill_map(); // This CFG path dies here 2065 } 2066 2067 2068 //------------------------------add_safepoint---------------------------------- 2069 void Parse::add_safepoint() { 2070 // See if we can avoid this safepoint. No need for a SafePoint immediately 2071 // after a Call (except Leaf Call) or another SafePoint. 2072 Node *proj = control(); 2073 bool add_poll_param = SafePointNode::needs_polling_address_input(); 2074 uint parms = add_poll_param ? TypeFunc::Parms+1 : TypeFunc::Parms; 2075 if( proj->is_Proj() ) { 2076 Node *n0 = proj->in(0); 2077 if( n0->is_Catch() ) { 2078 n0 = n0->in(0)->in(0); 2079 assert( n0->is_Call(), "expect a call here" ); 2080 } 2081 if( n0->is_Call() ) { 2082 if( n0->as_Call()->guaranteed_safepoint() ) 2083 return; 2084 } else if( n0->is_SafePoint() && n0->req() >= parms ) { 2085 return; 2086 } 2087 } 2088 2089 // Clear out dead values from the debug info. 2090 kill_dead_locals(); 2091 2092 // Clone the JVM State 2093 SafePointNode *sfpnt = new (C) SafePointNode(parms, NULL); 2094 2095 // Capture memory state BEFORE a SafePoint. Since we can block at a 2096 // SafePoint we need our GC state to be safe; i.e. we need all our current 2097 // write barriers (card marks) to not float down after the SafePoint so we 2098 // must read raw memory. Likewise we need all oop stores to match the card 2099 // marks. If deopt can happen, we need ALL stores (we need the correct JVM 2100 // state on a deopt). 2101 2102 // We do not need to WRITE the memory state after a SafePoint. The control 2103 // edge will keep card-marks and oop-stores from floating up from below a 2104 // SafePoint and our true dependency added here will keep them from floating 2105 // down below a SafePoint. 2106 2107 // Clone the current memory state 2108 Node* mem = MergeMemNode::make(C, map()->memory()); 2109 2110 mem = _gvn.transform(mem); 2111 2112 // Pass control through the safepoint 2113 sfpnt->init_req(TypeFunc::Control , control()); 2114 // Fix edges normally used by a call 2115 sfpnt->init_req(TypeFunc::I_O , top() ); 2116 sfpnt->init_req(TypeFunc::Memory , mem ); 2117 sfpnt->init_req(TypeFunc::ReturnAdr, top() ); 2118 sfpnt->init_req(TypeFunc::FramePtr , top() ); 2119 2120 // Create a node for the polling address 2121 if( add_poll_param ) { 2122 Node *polladr = ConPNode::make(C, (address)os::get_polling_page()); 2123 sfpnt->init_req(TypeFunc::Parms+0, _gvn.transform(polladr)); 2124 } 2125 2126 // Fix up the JVM State edges 2127 add_safepoint_edges(sfpnt); 2128 Node *transformed_sfpnt = _gvn.transform(sfpnt); 2129 set_control(transformed_sfpnt); 2130 2131 // Provide an edge from root to safepoint. This makes the safepoint 2132 // appear useful until the parse has completed. 2133 if( OptoRemoveUseless && transformed_sfpnt->is_SafePoint() ) { 2134 assert(C->root() != NULL, "Expect parse is still valid"); 2135 C->root()->add_prec(transformed_sfpnt); 2136 } 2137 } 2138 2139 #ifndef PRODUCT 2140 //------------------------show_parse_info-------------------------------------- 2141 void Parse::show_parse_info() { 2142 InlineTree* ilt = NULL; 2143 if (C->ilt() != NULL) { 2144 JVMState* caller_jvms = is_osr_parse() ? caller()->caller() : caller(); 2145 ilt = InlineTree::find_subtree_from_root(C->ilt(), caller_jvms, method()); 2146 } 2147 if (PrintCompilation && Verbose) { 2148 if (depth() == 1) { 2149 if( ilt->count_inlines() ) { 2150 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2151 ilt->count_inline_bcs()); 2152 tty->cr(); 2153 } 2154 } else { 2155 if (method()->is_synchronized()) tty->print("s"); 2156 if (method()->has_exception_handlers()) tty->print("!"); 2157 // Check this is not the final compiled version 2158 if (C->trap_can_recompile()) { 2159 tty->print("-"); 2160 } else { 2161 tty->print(" "); 2162 } 2163 method()->print_short_name(); 2164 if (is_osr_parse()) { 2165 tty->print(" @ %d", osr_bci()); 2166 } 2167 tty->print(" (%d bytes)",method()->code_size()); 2168 if (ilt->count_inlines()) { 2169 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2170 ilt->count_inline_bcs()); 2171 } 2172 tty->cr(); 2173 } 2174 } 2175 if (PrintOpto && (depth() == 1 || PrintOptoInlining)) { 2176 // Print that we succeeded; suppress this message on the first osr parse. 2177 2178 if (method()->is_synchronized()) tty->print("s"); 2179 if (method()->has_exception_handlers()) tty->print("!"); 2180 // Check this is not the final compiled version 2181 if (C->trap_can_recompile() && depth() == 1) { 2182 tty->print("-"); 2183 } else { 2184 tty->print(" "); 2185 } 2186 if( depth() != 1 ) { tty->print(" "); } // missing compile count 2187 for (int i = 1; i < depth(); ++i) { tty->print(" "); } 2188 method()->print_short_name(); 2189 if (is_osr_parse()) { 2190 tty->print(" @ %d", osr_bci()); 2191 } 2192 if (ilt->caller_bci() != -1) { 2193 tty->print(" @ %d", ilt->caller_bci()); 2194 } 2195 tty->print(" (%d bytes)",method()->code_size()); 2196 if (ilt->count_inlines()) { 2197 tty->print(" __inlined %d (%d bytes)", ilt->count_inlines(), 2198 ilt->count_inline_bcs()); 2199 } 2200 tty->cr(); 2201 } 2202 } 2203 2204 2205 //------------------------------dump------------------------------------------- 2206 // Dump information associated with the bytecodes of current _method 2207 void Parse::dump() { 2208 if( method() != NULL ) { 2209 // Iterate over bytecodes 2210 ciBytecodeStream iter(method()); 2211 for( Bytecodes::Code bc = iter.next(); bc != ciBytecodeStream::EOBC() ; bc = iter.next() ) { 2212 dump_bci( iter.cur_bci() ); 2213 tty->cr(); 2214 } 2215 } 2216 } 2217 2218 // Dump information associated with a byte code index, 'bci' 2219 void Parse::dump_bci(int bci) { 2220 // Output info on merge-points, cloning, and within _jsr..._ret 2221 // NYI 2222 tty->print(" bci:%d", bci); 2223 } 2224 2225 #endif