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